The Next Generation of Electric Power Unit Commitment Models PDF

THE NEXT GENERATION OF
ELECTRIC POWER
UNIT COMMITMENT MODELS
INTERNATIONAL SERIES IN
OPERATIONS RESEARCH & MANAGEMENT SCIENCE
Frederick S. Hillier, Series Editor
Stanford University
Saigal, R. / LINEAR PROGRAMMING: A Modern Integrated Analysis

Nagurney, A. & Zhang, D. / PROJECTED DYNAMICAL SYSTEMS AND
VARIATIONAL INEQUALITIES WITH APPLICATIONS
Padberg, M. & Rijal, M. / LOCATION, SCHEDULING, DESIGN AND
INTEGER PROGRAMMING
Vanderbei, R. / LINEAR PROGRAMMING: Foundations and Extensions
Jaiswal, N.K. / MILITARY OPERATIONS RESEARCH: Quantitative Decision Making
Gal, T. & Greenberg, H. / ADVANCES IN SENSITIVITY ANALYSIS AND
PARAMETRIC PROGRAMMING
Prabhu, N.U. / FOUNDATIONS OF QUEUEING THEORY
Fang, S.-C., Rajasekera, J.R. & Tsao, H.-S.J. / ENTROPY OPTIMIZATION
AND MATHEMATICAL PROGRAMMING
Yu, G. / OPERATIONS RESEARCH IN THE AIRLINE INDUSTRY
Ho, T.-H. & Tang, C. S. / PRODUCT VARIETY MANAGEMENT
El-Taha, M. & Stidham , S. / SAMPLE-PATH ANALYSIS OF QUEUEING SYSTEMS
Miettinen, K. M. / NONLINEAR MULTIOBJECTIVE OPTIMIZATION
Chao, H. & Huntington, H. G. / DESIGNING COMPETITIVE ELECTRICITY MARKETS
Weglarz, J. / PROJECT SCHEDULING: Recent Models, Algorithms & Applications
Sahin, I. & Polatoglu, H. / QUALITY, WARRANTY AND PREVENTIVE MAINTENANCE
Tavares, L. V. / ADVANCED MODELS FOR PROJECT MANAGEMENT
Tayur, S., Ganeshan, R. & Magazine, M. / QUANTITATIVE MODELING FOR SUPPLY
CHAIN MANAGEMENT
Weyant, J./ ENERGY AND ENVIRONMENTAL POLICY MODELING
Shanthikumar, J.G. & Sumita, U./APPLIED PROBABILITY AND STOCHASTIC PROCESSES
Liu, B. & Esogbue, A.O. / DECISION CRITERIA AND OPTIMAL INVENTORY PROCESSES
Gal, T., Stewart, T.J., Hanne, T./ MULTICRITERIA DECISION MAKING: Advances in MCDM
Models, Algorithms, Theory, and Applications
Fox, B. L./ STRATEGIES FOR QUASI-MONTE CARLO
Hall, R.W. / HANDBOOK OF TRANSPORTATION SCIENCE
Grassman, W.K./ COMPUTATIONAL PROBABILITY
Pomerol, J-C. & Barba-Romero, S. / MULT1CRITERION DECISION IN MANAGEMENT
Axsäter, S. / INVENTORY CONTROL
Wolkowicz, H., Saigal, R., Vandenberghe, L./ HANDBOOK OF SEMI-DEFINITE
PROGRAMMING: Theory, Algorithms, and Applications
Hobbs, B. F. & Meier, P. / ENERGY DECISIONS AND THE ENVIRONMENT: A Guide
to the Use of Multicriteria Methods
Dar-El, E./ HUMAN LEARNING: From Learning Curves to Learning Organizations
Armstrong, J. S./ PRINCIPLES OF FORECASTING: A Handbook for Researchers and
Practitioners
Balsamo, S., Personé, V., Onvural, R./ ANALYSIS OF QUEUEING NETWORKS WITH BLOCKING
Bouyssou, D. et al/ EVALUATION AND DECISION MODELS: A Critical Perspective
Hanne, T./ INTELLIGENT STRATEGIES FOR META MULTIPLE CRITERIA DECISION MAKING
Saaty, T. & Vargas, L./ MODELS, METHODS, CONCEPTS & APPLICATIONS OF THE ANALYTIC
HIERARCHY PROCESS
Chatterjee, K. & Samuelson, W./ GAME THEORY AND BUSINESS APPLICATIONS
THE NEXT GENERATION OF
ELECTRIC POWER
UNIT COMMITMENT MODELS
Editors
Benjamin F. Hobbs
The Johns Hopkins University
Michael H. Rothkopf
Rutgers University
Richard P. O’Neill
Federal Energy Regulatory Commission
Hung-po Chao
Electric Power Research Institute
KLUWER ACADEMIC PUBLISHERS

NEW YORK, BOSTON, DORDRECHT, LONDON, MOSCOW
eBook ISBN: 0-306-47663-0
Print ISBN: 0-7923-7334-0
©2002 Kluwer Academic Publishers

New York, Boston, Dordrecht, London, Moscow
Print ©2001 Kluwer Academic Publishers

Dordrecht
All rights reserved
No part of this eBook may be reproduced or transmitted in any form or by any means, electronic,
mechanical, recording, or otherwise, without written consent from the Publisher
Created in the United States of America
Visit Kluwer Online at: http://kluweronline.com

and Kluwer's eBookstore at: http://ebooks.kluweronline.com
CONTENTS
Acknowledgments vii
I. The Evolving Context for Unit Commitment Decisions
1. Why This Book?: New Capabilities and New Needs for Unit
Commitment Modeling
B. F. Hobbs, W. R. Stewart Jr., R. E. Bixby, M. H. Rothkopf, R. P.
O'Neill, H.-p. Chao 1
2. Regulatory Evolution, Market Design and Unit Commitment

R. P. O'Neill, U. Helman, P. M. Sotkiewicz, M. H. Rothkopf, W. R.
Stewart Jr. 15
3. Development of an Electric Energy Market Simulator

A. Debs, C. Hansen, Y.-C. Wu 39
II. New Features in Unit Commitment Models
4. Auctions with Explicit Demand-Side Bidding in Competitive

Electricity Markets
A. Borghetti, G. Gross, C. A. Nucci 53
5. Thermal Unit Commitment with a Nonlinear AC Power Flow

Network Model
C. E. Murillo-Sánchez, R. J. Thomas 75
6. Optimal Self-Commitment under Uncertain Energy and Reserve

Prices
R. Rajaraman, L. Kirsch, F. L. Alvarado, C. Clark 93
7. A Stochastic Model for a Price-Based Unit Commitment Problem

and Its Application to Short-Term Generation Asset Valuation
C.-L. Tseng 117
8. Probabilistic Unit Commitment under a Deregulated Market

J. Valenzuela, M. Mazumdar 139
III. Algorithmic Advances
9. Solving Hard Mixed-Integer Programs for Electricity Generation

S. Ceria 153
vi The Next Generation of Unit Commitment Models
10. An Interior-Point/Cutting-Plane Algorithm to Solve the Dual Unit

Commitment Problem -- On Dual Variables, Duality Gap, and
Cost Recovery
M. Madrigal, V. H. Quintana 167
11. Building and Evaluating Genco Bidding Strategies and Unit

Commitment Schedules with Genetic Algorithms
C. W. Richter, Jr., G. B. Sheblé 185
12. An Equivalencing Technique for Solving the Large-Scale Ther-

mal Unit Commitment Problem
S. Sen, D.P. Kothari 211
IV. Decentralized Decision Making
13. Strategic Unit Commitment for Generation in Deregulated Elec-

tricity Markets
A. Baíllo, M. Ventosa, A. Ramos, M. Rivier, A. Canseco 227
14. Optimization-Based Bidding Strategies for Deregulated Electric

Power Markets
X. Guan, E. Ni, P. B. Luh, Y.-C. Ho 249
15. Decentralized Nodal-Price Self-Dispatch and Unit Commitment

F. D. Galiana. A. L. Motto, A. J. Conejo, M. Huneault 271
16. Decentralized Unit Commitment in Competitive Energy Markets

J. Xu, R. D. Christie 293
Index 317
ACKNOWLEDGMENTS
This volume contains papers that were presented at a workshop entitled
"The Next Generation of Unit Commitment Models", held September 27-
28, 1999 at the Center for Discrete Mathematics and Theoretical Computer
Science (DIMACS), Rutgers University, Piscataway, NJ. The editors grate-
fully acknowledge the financial support of the co-sponsors of the workshop:
DIMACS (funded by the National Science Foundation under grant NSF
STC 91-19999); and the Electric Power Research Institute (EPRI), which
supported the publication and distribution of this book. The editors also
thank Sarah Donnelly of DIMACS for her organizational support of the
workshop and the subsequent book.
The editors would also like the many speakers and other participants of
the workshop for their ideas and hard work. All the papers were subjected
to anonymous peer review by at least two referees and two editors. The
referees included authors of other papers in this volume, along with Paul
Sotkiewicz and Judith Cardell of the Office of Economic Policy of the Fed-
eral Energy Regulatory Commission.
Technical editing for the volume was ably provided by Debi Rager of
The White Cottage Company. Liz Austin of Johns Hopkins Unviersity
compiled the index. Funding for B. Hobbs' involvement in the workshop
and book came from NSF Grants ECS 96-96014 and 00-80577. Partial
support for M. Rothkopf was provided by NSF grant SBR 97-09861.
This page intentionally left blank
Chapter 1
WHY THIS BOOK?

NEW CAPABILITIES AND NEW NEEDS FOR
UNIT COMMITMENT MODELING
Benjamin F. Hobbs
The Johns Hopkins University
William R. Stewart Jr.

The College of William & Mary
Robert E. Bixby
Rice University and ILOG
Michael H. Rothkopf
Rutgers University
Richard P. O'Neill
Hung-po Chao
Abstract: This book presents recent developments in the functionality of generation unit
commitment (UC) models and algorithms for solving those models. These de-
velopments, the subject of a September 1999 workshop, are driven by institu-
tional changes that increase the importance of efficient and market responsive
operation. We illustrate these developments by demonstrating the use of
mixed integer programming (MIP) to solve a UC problem. The dramatically
lower solution times of modem MIP software indicates that it is now a practi-
cal algorithm for UC. Participants in the workshop also prioritized the features
that need to be considered by UC models, along with topics for research and
development. Among the highest research priorities are: market simulation;
bid selection; reliability and reserve constraints; and fair processes for choos-
ing from alternative near-optimal solutions. The chapter closes with an over-
view of the contributions of the other chapters.
2 The Next Generation of Unit Commitment Models
1. PURPOSE OF THE BOOK

The unit commitment problem can be defined as the scheduling of pro-
duction of electric power generating units over a daily to weekly time hori-
zon in order to accomplish some objective. The problem solution must re-
spect both generator constraints (such as ramp rate limits and minimum up or
down times) and system constraints (reserve and energy requirements and,
potentially, transmission constraints). The objective function should account
for costs associated with energy production, ramping, start-ups and shut-
downs decisions, along with possible effects upon revenues or customer
costs of those decisions. The resulting problem is a large scale nonlinear
mixed integer program.
For many years, the electric power industry has been using optimization
methods to help them solve the unit commitment problem. The result has
been savings of tens and perhaps hundreds of millions of dollars in fuel
costs. Things are changing, however. Optimization technology is improv-
ing, and the industry is undergoing radical restructuring. Consequently, the
role of commitment models is changing, and the value of the improved solu-
tions that better algorithms might yield is increasing. The purpose of this
book is to explore the technology and needs of the next generation of com-
puter models for aiding unit commitment decisions.
Because of the unit commitment problem's size and complexity and be-
cause of the large economic benefits that could result from its improved
solution, considerable attention has been devoted to algorithm development
[1,2]. Heuristics such as priority lists have long been used by industry; but in
the last three decades, more systematic procedures based on a variety of al-
gorithms have been proposed and tested. These techniques have include dy-
namic programming, branch-and-bound mixed integer programming (MIP),
linear and network programming approaches, and Benders decomposition
methods, among others. Recently, metaheuristic methods have been tested,
such as genetic programming and simulated annealing, along with expert
systems and neural networks.
The solution approach that has been most successful, and which is most
widely used at present, is Lagrangian relaxation. This procedure decomposes
the problem by multiplying constraints that couple different generators (such
as energy demand and reserve constraints) by Lagrange multipliers and plac-
ing them in the objective function. Given a set of multiplier values, the prob-
lem is then separable in the generating units, and a dynamic program of low
dimension can be used to obtain a trial schedule for each unit. A process of
multiplier adjustment is used to search for feasible near-optimal solutions.
Lagrangian relaxation has proven useful for quick development of good,
if not optimal, generator schedules. Recent improvements in integer pro-
Introduction 3
gramming codes and other algorithms suggest, however, that it may be pos-
sible to find better solutions more rapidly. Further, such codes can more
readily incorporate additional coupling constraints, such as transmission lim-
its and emissions caps. Meanwhile, restructuring has sharpened the appetite
of generation owners for more efficient operation. In the past, utilities sold
power on a regulated cost-plus basis and so may not have put as much prior-
ity on squeezing out the last few percent improvements in the objective func-
tion. Furthermore, system operators realized that cost functions were ap-
proximate, so the operators were perhaps more likely to be satisfied with
good solutions or marginal improvements that were technically suboptimal.
Now, with restructuring, we have schedule coordinators making commitment
decisions in a market environment, and independent system operators (ISOs)
dealing with bids. Bids are precise, and small improvements in solutions can
result in significant changes in payments to bidders. Further, the fact that
optimization models are, in some cases, being used to determine which gen-
erators will be operated and thus paid implies that there is a greater incentive
to get exact answers to make the bidding process fair and legitimate (and to
disempower the bid-taker).
In other words, electric markets are changing rapidly, as is the role of
unit commitment models. How UC models are solved and what purposes
they serve deserve reconsideration. The goal of the workshop that led to this
book was to bring together people who understand the problem and people
who know what improvements in algorithms are really possible. The papers
in this book summarize the participants' assessments of industry needs to-
gether with new formulations and computational approaches that promise to
make unit commitment models more responsive to those needs.
In Section 2 of this chapter, we present an example to show how the ca-
pabilities of commercially available integer programming software to solve
large unit commitment problems to optimality have dramatically improved
in recent years. This example illustrates how improvements in software may
make it possible to solve bigger problems in less time, while simultaneously
including more of the complications that users want to represent. Section 3
then summarizes the results of a survey of workshop participants in which
they were asked to identify what issues concerning unit commitment model-
ing are most in need of further research and development. Finally, in Sec-
tion 4 we give an overview of the other chapters in this book. There, we de-
scribe how the papers contribute to our two goals of articulating the emerg-
ing needs of the restructured power industry and describing model develop-
ments that can make unit commitment models more responsive to those
needs.
2. EXAMPLE OF NEW CAPABILITIES: SOLVING

UNIT COMMITMENT PROBLEMS USING MIP
When formulated as a mixed integer program (MIP), the unit commit-
ment (UC) problem is a large and complex mathematical programming prob-
lem. As a result, optimal solutions have been hard to obtain for practically
sized problems due to the exponential behavior of solution times as problems
grow larger. As a result, exact solutions to large UC problems have been
unavailable, and approximate techniques are employed to produce solutions
that are near optimal (within 0.5%-2%). In practice, Lagrangian relaxation
methods have performed well, but the non-convexity and overall size of
most practical problems have prevented solving such problems to optimality
or even providing a bound on the optimal solution.
Approximate solutions have two problems. First, since an approximate
solution will be dispatching units that are slightly more costly than the ones
that would have been dispatched had an optimal solution been available, the
approximate solution almost certainly will be inefficient economically and a
deserving facility may have been passed over by the approximate solution.
The second, related problem consists of the political implications of passing
over cost-effective units in a world in which generating units are owned by
different competing entities, as opposed to the historical situation where all
generation was centrally owned by a single, large regulated utility.
The purpose of this section is to report on how the technology for solv-
ing MIPs to optimality (branch-and-bound, cutting planes, etc.) has im-
proved computational times in the past several years, and the implications of
those improvements for solving UC problems. Historically, MIPs have been
notoriously difficult to solve due to the presence of multiple near-optimal
solutions such as would occur when a system contained several generators
with similar operating characteristics and costs. Traditional branch-and-
bound techniques would have to find all of these optimal solutions explicitly
before an optimal solution could be verified. For this reason and because of
the many near-optimal solutions in a typical large MIP, solution times grow
very quickly with the size of the problem because of the large number of
nodes in the branch-and-bound tree that must be fathomed. However, a
number of theoretical improvements have been incorporated into commercial
MIP codes, and further improvements are anticipated. Consequently, these
codes can now solve to optimality problems thought impossible just a few
years ago.
As a concrete example of these improvements, we summarize the results
of applying several recent generations of a widely used MIP code, CPLEX®,
to a test problem. The problem comes from a paper by Johnson et al. [3]. It
requires the scheduling of 17 generators to meet electricity demand over a
Introduction 5
seven-day (168-hour) period. The generators have different operating char-

acteristics and the resulting MIP contains 25,755 variables (including 2,856
integer variables) and 48,939 constraints. While this may be a small prob-
lem by unit commitment standards, until recently it would have been consid-
ered intractable for direct solution by commercial MIP solvers using a linear
programming/branch-and-bound/cutting plane technology. Recent advances
have substantially accelerated these approaches, however; as a result, solu-
tion times for problems such as the one below are approaching the range
where the MIP approach can be considered viable.
Test UC Mixed Integer Linear Program:
Minimize
subject to:
where:
is the MW of energy produced by generator i in period t,
is a binary variable that is 1 if generator i is dispatched during t,
is 1 if generator i is started at the beginning of period t,
is 1 if generator i is shut down at the beginning of period t,
is the MW of spinning reserves available from generator i in t,
and are the fixed cost of operating ($/period), the cost of
generation ($/MW/period), and the cost of start-ups ($), respec-
tively, for generator i during t,
and are the minimum and maximum MW ca-
pacities of plant i, and its maximum reserve contribution,
respectively, and
and are the maximum ramp rates (in MW/period) for

increasing and decreasing output, respectively, from unit i.
The seven-day MIP problem is non-convex, with one integer variable

for each generator for each hour of the week. The rest of the variables
are continuous. In addition to the formulation as shown in (1)-(10), the ob-
jective function coefficients for some of the generators grow with out-
put level. These are approximated by piecewise linear cost functions. There
is also a set of constraints not shown in (1)-(10) that controls for the mini-
mum number of periods a generator can be up and running and the minimum
number of periods that a generator must be shut down.
The objective of UC is to choose a set of generators and operating levels
so as to minimize the total cost of the dispatch (1) subject to meeting hourly
demand (2), having sufficient spinning reserves (3) (arbitrarily set to 3% of
demand in this problem), generation at each operating unit at or above its
minimum run level and below its maximum output level (4)-(5), and spin-
ning reserves below the maximum level for each generator (6). Constraints
(7)-(8) force hour-to-hour changes in generation to respect ramp rate limits.
Constraint (9) requires that a generator be started if it was not dispatched the
previous period and will be this period, while (10) mandates that a unit be
shut down this period if it is was dispatched the previous period and will not
be used this period.
The results for solving two related test problems are reported in Tables 1
and 2. The first problem is a one-day (24 time periods) problem and the
other is the full seven-day problem. The one-day problem has one-seventh
the number of continuous variables, discrete variables, and constraints that
the week-long problem has. The main computational comparison is how the
solution times for these problems improve as they are solved by later ver-
sions of a specific solver (CPLEX). In addition, we show the number of
times a linear program is solved in the branch-and-bound portion of the algo-
rithm (number of nodes).
Introduction 7
The marked improvement in times seen in Version 6.5 (Table 1) can be

attributed to several factors: the inclusion of cutting planes eliminates a lot of
extraneous fractional solutions that were previously explicitly considered;
later versions of the dual simplex algorithm in CPLEX just solve the linear
program faster (a couple of orders of magnitude faster on the seven day
problem, see below); and the inclusion of pre-solve reductions substantially
tightens the problem before branch-and-cut starts. The improvement of
times from CPLEX 6.0 to CPLEX 6.5 of about an order of magnitude gives
promise that this approach can be useful for solving UCPs in the future.
The seven-day (168 time periods) problem was unsolvable in a reason-
able amount of time in CPLEX 4.0; indeed, it took an hour just to get a solu-
tion to the linear programming relaxation. Yet, as Table 2 shows, that prob-
lem was solved in less than two hours using CPLEX 6.5 and under 30 min-
utes using a developmental version of CPLEX on a 500 MHz DEC Alpha.
These results show that innovations in algorithms and their implementations
will continue to have a strong impact on solution times for MIPs and that
realistic unit commitment problems can be solved to optimality by off-the-
shelf software.
3. NEW NEEDS: RESEARCH PRIORITIES

The potential of improved algorithms was one important focus of the
workshop. The other was the functionality of unit commitment models. The
workshop attendees participated in an afternoon meeting to evaluate the
functions that unit commitment models should perform in restructured markets.
The meeting was structured as a nominal group [4] to maximize the efficiency
of information exchange and to ensure that all forty or so participants had a
chance to voice their ideas. The steps of a nominal group exercise are as
follows: participants silently write down ideas in response to one or more
questions; the ideas are posted without attribution; the ideas are discussed in
round robin fashion; and participants numerically rate the ideas either at the
meeting or in a questionnaire after the meeting.
The workshop participants performed the following tasks:
Identify features, characteristics, or requirements of system operation and

bidding that are desirable to include in unit commitment models, and issues
involved in representing these issues in those models.
Describe the extent to which additional research is needed to incorporate
those features in unit commitment models. This was done by grouping the
features into three categories:
I This feature/characteristic/requirement is modeled by existing
models and software;
II Models have been proposed for capturing this feature/
characteristic/requirement, but require development and
implementation; and
III There remain fundamental disputes or uncertainties as to how
this feature/characteristic/requirement should be modeled. In
this case, fundamental research is needed.
Rate the importance of including each feature/characteristic/requirement on
the following Likert scale:
1 Minor importance
3 Somewhat important
5 Very important
7 Crucial
Approximately 40 of the workshop attendees addressed the first two tasks

during the afternoon session. Eighteen participants returned the follow-up
questionnaire on importance rating. The latter group included nine industry
representatives (primarily consultants, but also including generating company
and EPRI representatives); eight university researchers; and a regulator. From
the results of these tasks, a set of research priorities can be distilled.
The following two tables summarize the results. Table 3 ranks the features
by the mean importance rating given by the group, while Table 4 categorizes
them in terms of whether further research is needed. Where participants
disagreed concerning the category, a range is shown.
Table 3 indicates near unanimity concerning the crucial importance of
including the commodities of energy and ancillary services along with dynamic
constraints and decision sequencing. The next most important issue was
market prediction, interpreted as the projection of prices, and how the market
might respond to alternative bidding strategies.
Some importance ratings were mildly surprising. In particular, transmission
issues were rated relatively low. It would seem that research on transmission is
Introduction 9
Introduction 11
not a high priority with this group, as inclusion of AC representations fell in

category II and received a moderate importance rating, while DC
approximations were in I and received one of the lowest average importance
ratings. Other features that received low ratings included distributed
generation; emissions modeling; and particular stochastic methods for

modeling reliability, uncertainty, and variability.
Table 4 reveals that the group unanimously identified 11 features as re-
quiring fundamental research (category III), of which nine received average
importance ratings of 4.0 or more. These features are extremely diverse,
ranging from risk and reliability modeling to strategic bidding, new product
valuation, and dealing with the fairness and institutional implications of mul-
tiple near-optimal solutions.
Another six features were thought by some but not all group members to
be in category III and so are identified as being in category “II-III.” Several
facets of market prediction were highly rated in that category. These facets
include price prediction, bidding strategy assessment, and inclusion of price-
responsive demand. Meanwhile, in the “requires development” category (II),
there are two very important topics: inclusion of ancillary services in unit
commitment models and explicit accounting for the sequential nature of de-
cisions (such as multi-settlement systems). Research of a more applied na-
ture would seem to be justified in these cases.
4. BOOK OVERVIEW
The chapters of this book are grouped into four sections. The first sec-
tion includes this introduction and two other chapters that summarize the
evolving institutional context that has motivated the functional and algo-
rithmic developments described in the rest of the book. O'Neill, Helman,
Sotkiewicz, Rothkopf, and Stewart (Chapter 2) review the recent history of
short-term electricity markets in the U.S., focussing on alternative market
designs and the implications for unit commitment modeling. They also sug-
gest some principles for designing the next generation of UC market models.
An alternative approach to presenting the evolving context of unit commit-
ment is presented in Chapter 3. There, Debs, Hansen, and Wu present a gen-
eral modeling framework that encompasses all the functions of short-term
energy markets, including commitment, with a focus on whether the market
participant is an ISO/RTO, generating company, market administrator, load
serving entity, or even an energy service company.
The other three sections of the book describe novel applications and fea-
tures in UC models, new algorithms for solving those models, and modeling
approaches that represent decentralized commitment by independent generat-
ing firms. Chapter 4 is the first chapter in the second section; there, Bor-
ghetti, Gross, and Nucci show how demand-side bidding can be included in
Lagrangian-relaxation-based unit commitment models, and how such bid-
ding can dampen price volatility and mitigate market power. Their formula-
Introduction 13
tion represents load recovery, hour-by-hour bids for demand reduction, and
multiple bidders. Chapter 5 by Murillo-Sanchez and Thomas is the second
chapter on new model features. They describe how a nonlinear AC power
flow representation can be incorporated, with both active and reactive power
sources. They also discuss a parallel processing implementation.
The last three chapters of the second section concern the inclusion of
price uncertainty in UC models. All use variations of a stochastic dynamic
programming-based commitment model; when generators cannot individu-
ally influence price, the models neatly decompose into a single optimization
model for each generator representing their self-commitment problem. In
Chapter 6, Rajaraman, Kirsch, Alvarado, and Clark describe how uncertain-
ties in both reserve and energy prices can be considered in such models. In
Chapter 7, Tseng shows how such models can be used to quantify rigorously
the worth of operating flexibility (“option value”) for a single generating
asset. Finally, in Chapter 8, Valenzuela and Mazumdar present a model for
single generator optimization that uses probability distributions of market
prices directly derived from assumptions concerning demand variability and
generator availability in the whole market.
In the third section of the book, we turn our attention to improved algo-
rithms for solving the UC problem. Four distinct approaches are represented
in this section: mixed integer programming, Lagrangian relaxation, genetic
algorithms, and aggregation approaches.
Chapter 9 by Ceria, like Section 3 of this introduction, addresses the use-
fulness of MIP for UC, along with recent developments in MIP technology
that have drastically improved solution times. He also briefly reviews two
actual applications by utilities in Europe. Chapter 10, authored by Madrigal
and Quintana, proposes an interior-point/cutting-plane algorithm to solve the
Lagrangian relaxation problem and demonstrates its computational advan-
tages over subgradient and other methods traditionally used to update La-
grange multipliers. They also offer some observations on several issues in-
volved in using UC models to clear power markets, including duality gaps,
cost recovery, and the existence of multiple solutions. Chapter 11, contrib-
uted by Richter and Sheble, reviews a range of considerations involved in
creating bidding and commitment strategies. They then propose genetic al-
gorithms and finite state automata-based simulations for strategy develop-
ment and testing. Chapter 12 by Sen and Kothari shows how aggregation of
generating units into a few sets of similar units can be exploited to improve
solution times. The algorithm involves three basic steps: aggregation; solu-
tion of the simplified UC problem using dynamic programming or another
optimization method; and disaggregation to create schedules for individual
units.
The fourth and final section of the book contains four chapters that ad-
dress modeling and algorithmic issues associated with decentralized com-
mitment decision processes. Important questions include opportunities for
strategic manipulation of prices by market participants, coordination algo-
rithms, and the ability of decentralized processes to approach optimality.
The first two chapters of the fourth section focus on decision-making by
individual firms. Baillo, Ventosa, Ramos, Rivier, and Canseco devote Chap-
ter 13 to a model for committing a firm’s units while recognizing how com-
mitment and dispatch decisions may affect market prices. Rival firms are
assumed to behave according to price-elastic supply functions, which allows
for derivation of total firm revenue as a function of its output. They use
MIP to solve their model. In Chapter 14, Guan, Ni, Luh, and Ho describe
two general approaches to bid development in decentralized markets. One is
based on “ordinal optimization” for obtaining satisficing bidding strategies,
and a second uses stochastic optimization to self-schedule and manage risks
while considering interactions among different markets.
The last two chapters of the fourth section turn to the issue of coordina-
tion of decentralized decisions. Galiana, Motto, Conejo, and Huneault
(Chapter 15) propose a coordination process in which locational prices are
announced, generating firms self-dispatch to maximize their individual prof-
its, and prices are adjusted to ensure that demands and network constraints
are satisfied. The authors use a Newton algorithm to update prices, and im-
pose a “convexifying rule” to facilitate convergence. Case studies illustrate
the process. In Chapter 16, Xu and Christie consider the interactions of stra-
tegic behavior by individual generating firms with a price-based coordinat-
ing mechanism. Firms optimize bidding strategies with the help of a simple
price prediction model. The combined effects of multiple firms using that
approach is explored with a market simulator, which reveals that conver-
gence, feasibility, and price stability can be difficult to achieve, but that price
cycling can be dampened with alternative price prediction models.
REFERENCES
1. G. Sheblé and G. Fahd. Unit commitment literature synopsis. IEEE Trans. Power Syst.,
9(1): 128-135, 1994.
2. S. Sen and D.P. Kothari. Optimal thermal generating unit commitment: a review. Elec.
Power Energy Syst., 20(7): 443-451, 1998.
3. R.B. Johnson, S.S. Oren, and A.J. Svoboda. Equity and efficiency of unit commitment
in competitive electricity markets. Utilities Policy, 6(1): 9-20, 1997.
4. A. Delbecq, A. Van de Ven, and D. Gustafson. Group Techniques for Program Plan-
ning — A Guide to Nominal Group and Delphi Processes. Glenview, IL: Scott Fores-
man and Co., 1975.
Chapter 2
REGULATORY EVOLUTION, MARKET DESIGN

AND UNIT COMMITMENT
Richard P. O’Neill, Udi Helman, and Paul M. Sotkiewicz

Michael H. Rothkopf
Rutgers University
William R. Stewart Jr.

The College of William and Mary
Abstract: In the context of competitive wholesale electricity markets, the unit commit-
ment problem has shifted from a firm level optimization problem to a market
level problem. Some centralized market designs use it to ensure reliability and
determine day-ahead market prices. This chapter reviews the recent history of
short-term electricity markets in the United States to evaluate the experience
with alternative market designs and the implications for unit commitment mod-
eling. It presents principles for the design of the next generation of unit com-
mitment-based markets.
1. INTRODUCTION
Competitive wholesale electricity markets now operate in several major
U.S. markets, confirming the analysis and recommendations of prescient
economists, electrical engineers, and others over the past two decades.1 Since
generation comprises approximately 75 percent of all electricity costs, com-
1
Seminal contributions on competitive wholesale electricity markets include [1,2]. As of
January 1, 2000, regional markets with centralized wholesale electricity exchanges are opera-
tional in California, the Pennsylvania-New Jersey-Maryland (PJM) interconnection, New
England, and New York.
petition in generation promises large efficiency gains and cost savings to con-
sumers. The unit commitment problem, the traditional method by which regu-
lated utilities and power pools conducted internal scheduling of generation to
meet demand at least cost over a multi-hour to multi-day time frame, is now
embedded, in various ways, in competitive markets. Potentially, unit com-
mitment models will be used by different market participants and institutions:
individual firms, centralized auctioneers, decentralized aggregators of genera-
tion schedules, and transmission system operators. This environment presents
a new set of modeling requirements and market design challenges.2 The mar-
ket level unit commitment problem is typically much larger in scale than the
firm level problem. Speed and accuracy are important if an auctioneer uses
the solution by an auctioneer to determine market prices. Evaluating the char-
acteristics of the solution, such as the presence of duality gaps (implying a
lack of market clearing prices) and alternative optima, becomes of direct fi-
nancial interest to market participants.
The new electricity markets, and hence new applications of the unit com-
mitment problem, are being developed within an evolving regulatory context.
Indeed, an important driver of market designs is the guidance given by the
regulator. The Federal Energy Regulatory Commission (henceforth “the
Commission”) initiated regulatory reform of transmission in 1996, with the
objective of encouraging competitive regional electricity markets that pro-
mote economic efficiency without compromising system reliability. The regu-
latory approach, embodied in a series of orders described below, has been to
provide an open market architecture where alternative market designs are im-
plemented, evaluated, and changed when necessary. Research into the unit
commitment problem has largely been reactive to the new regulatory envi-
ronment and the emerging issues in market design. A more proactive ap-
proach is needed. Among the issues that need consideration and research are
the choices between simultaneous and sequential optimization of several en-
ergy and ancillary service products, alternative bidding rules for different
products, different mechanisms for congestion pricing, and inter-regional co-
ordination. In addition, unit commitment modeling now has to confront the
issue of economic incentives in various market settings, which requires a
more extensive familiarity with economics and game theory.
In response to the regulatory evolution it has set in motion, the regulator
also needs to adapt institutionally and develop its technical capabilities. This
is imperative because the Commission is taking an oversight role in market
design decisions across the United States. Several wholesale markets operate
centralized unit commitment auction markets (e.g., PJM, New England, and
2
The literature on market design in electricity markets is extensive; for a survey, see the arti-
cles in [3].
Regulatory Evolution, Market Design, and Unit Commitment 17
New York), the market design of which is the focus of this chapter. These
markets also allow bilateral trading and types of self-scheduling. Following
approval of the basic design, the Commission provides oversight for a flood
of subsequent adjustments and refinements in the search for well functioning
markets. The underlying unit commitment model is often either an implicit or
explicit matter in these market rule decisions.
The objective of this chapter is to describe regulatory evolution and the
market design challenges for unit commitment modeling. The chapter focuses
on day-ahead markets, but much of the discussion is also applicable to real-
time markets. Section 2 of the chapter describes the key regulatory develop-
ments and the design and recent experience of the major regional wholesale
electricity markets. Section 3 focuses on principles that should guide the de-
sign of day-ahead energy and ancillary service markets. Finally, Section 4
offers conclusions.
2. REGULATORY EVOLUTION AND THE

ORGANIZATION OF ELECTRICITY MARKETS
The recent history of electricity regulatory reform in the United States be-
gan when the Commission issued Orders 888 and 889 in 1996 [4,5]. These
orders required an open access transmission regime, based on non-
discriminatory transmission rates and transparent posting of available trans-
mission capacity (ATC). Order 888 also included fairly broad organizational
principles for an independent system operator (ISO), an institution which
separates ownership from control of the grid and can perform market func-
tions. Between 1997 and 2000, ISOs and power exchanges (PXs) were formed
in California and in the three tight power pools of the eastern United States.
These ISOs established day-ahead and real-time markets for energy, ancillary
services, and transmission (in California, the day-ahead energy market is con-
ducted by several separate scheduling coordinators, including the California
Power Exchange). An ISO has also been established in the Midwest, but is not
yet operational. Other regions of the country have been less successful or un-
willing to centralize grid operations, and electricity trading remains bilateral,
with a vertically integrated utility performing the balancing and reliability
functions.
By early 1999, a certain amount of inertia was evident in the development
of wholesale markets. Electricity traders expressed dissatisfaction with the
traditional methods of transmission grid management still employed in large
parts of the United States. Specifically, there was substantial concern about
frequent curtailments of transactions, justified on the basis of reliability but

often questioned by parties to the transactions.3
On December 15, 1999, the Commission took a step toward clarifying the
appropriate transmission access and market institutions with Order 2000,
which requires the formation of regional transmission organizations (RTOs)
[6]. The order establishes in the RTO many of the features that had emerged
in the ISO markets as well as additional characteristics and functions that ad-
dress unresolved issues in both ISO and non-ISO electricity markets. The
RTO is required to serve a region of sufficient scope and configuration to
provide for a reliable, efficient electricity market. With respect to the unit
commitment problem, some of the important features of the RTO are that it
must have exclusive authority for maintaining short term reliability, act as
provider of last resort of ancillary services, address parallel path flows, pro-
vide real time energy balancing, and ensure development of market mecha-
nisms for congestion management.
In many ways, the functions assigned to the RTO are based on the princi-
ples of market design embodied in the better functioning ISOs that have
emerged (in fact, RTOs will subsume existing ISOs). Section 2.1 surveys
some of these ISO market design lessons; Section 2.2 discusses further the
objectives of Order 2000 and some market design issues.
Before considering the details of open access and market design, an im-
portant question is: Why should the regulator remain involved in the design
and oversight of the emerging competitive markets? Recent experience has
made clear that, in the near term, the Commission has a continuing role for at
least three reasons:
1. There is the ongoing development of open access itself, including the
provision of short-term reliability services linked to transmission and
future expansion of the grid. There are public good aspects to reliabil-
ity and transmission expansion.4
2. Where not currently available, there must be an efficient pricing
mechanism for transmission congestion (i.e., pricing of the external-
ities created by parallel path flows).
3. There must be mitigation of market power in markets for electricity
and ancillary services and in provision of transmission capacity. Mar-
ket power is the ability of firms to raise prices above competitive lev-
3
Such curtailments are supposed to follow the North American Electricity Reliability Coun-
cil’s (NERC) Transmission Loading Relief (TLR) procedures, which provide criteria for the
management of congested transmission facilities.
4
A public good is a good that is non-rivalrous and non-excludable. In the case of reliability, a
load’s or generator’s consumption of reliability in no way prevents others from enjoying the
same reliability, and if all of the loads and generators are interconnected on the same system,
they cannot be prevented from enjoying the benefits of reliability.
els. Firms can exercise market power in electricity markets because of

both structural factors (e.g., firm concentration or transmission con-
straints) and opportunities offered by the market design (see [7,8] for
a survey).
The Commission’s policy heretofore has been to approve implementation
of the markets (through more liberal standards for granting market-based
rates) while at the same time evaluating the markets’ operational experience
and providing guidance on market designs as a means to promote efficiency
and competition. In addition, most of the responsibility for day-to-day moni-
toring and mitigation of market power has shifted to the ISOs and the future
RTOs. The regulatory approach, then, seeks to balance the current reality -
some firms can exercise a degree of market power generally or under certain
system conditions – with the expectation that entry of new firms and more
efficient market designs will substantially mitigate future market power.5
2.1 Experience with ISO and Bilateral Markets

Order 888 outlined principles for, but did not require, a particular struc-
ture for competitive wholesale energy markets. Two broad types of market
structures have developed. The first type consists of markets with ISOs, which
may or may not include one or more scheduling coordinators or PXs.6 The
ISO markets with PXs take the form of either a centralized ISO/PX or a de-
centralized ISO and PX(s). The centralized ISO/PX markets use unit com-
mitment models for creating the day-ahead schedule (which incorporates bi-
lateral transactions and self-schedules) while the decentralized ISO markets
require self-commitment by the PXs. The second type of market has no ISO;
rather, the transmission system and much of the generation continues to be
operated by vertically integrated utilities. Power is traded bilaterally.
This section focuses on the various market designs and performance of
the existing ISO markets, as well as offer some conclusions and recommenda-
tions on what designs will work best. It also includes a brief review of the per-
formance of the bilateral, non-ISO markets.
Market Functions and Design of ISO Markets. Competitive wholesale
5
There is a large body of literature on market power due to both structural and market design
characteristics of electricity markets. For analysis of regional energy and ancillary service
markets in the United States, see [8-12].
6
Scheduling coordinators or power exchanges (PXs) are power trading operations functionally
separate from the ISO. These terms will be used interchangeably.
electricity markets can be complex, with multiple interdependent products

sold on different time frames and differentially priced at different locations.
Existing ISO markets can be characterized by the (1) number and types of
different products (energy, ancillary services, capacity), (2) bidding and
scheduling process, (3) relationship of temporal (forward and real-time) mar-
kets, (4) market clearing and settlement rules, and (5) type of congestion
management and transmission rights. Each of the existing ISOs has also es-
tablished market power monitoring and mitigation, but this function will not
be examined here.
The ISO carries out the basic function of assessing the feasibility of pro-
posed generation schedules. The ISO also serves as the buyer, through con-
tracted rates and bid-based auctions, of reliability services, including short-
term ancillary services (voltage support, operating reserves, and automatic
generation control) and possibly longer-term capacity products. The PX facili-
tates and conducts a forward auction market for electric energy. PX functions
can either be carried out by the ISO itself, as in New York, New England, and
PJM, or by one or more separate, unaffiliated PXs, as in California.
With the exception of the California ISO, the ISOs run a unit commitment
model to determine which units will be scheduled to provide energy and an-
cillary services during the following day.
In California, the ISO and PX (which is one of several scheduling coordi-
nators) are separate, a decision intended to keep the transmission system op-
erators (who may have been affiliated with an incumbent utility) functionally
distinct from the market. In the California PX market, generation owners self-
commit their units through scheduling coordinators. This market structure has
experienced certain disadvantages. One problem is that the scheduling coor-
dinators’ submissions can be physically infeasible. The ISO must then engage
in a time consuming iterative process with the scheduling coordinators to re-
solve the infeasibilities. The PX auction algorithm is such, however, that self-
committed units are often asked to start-up and stop, disregarding minimum
run and down times, with potentially adverse results.7 Another problem is that
the separation of the ISO and PX raises the transaction costs for market par-
ticipants. In addition, because the ISO’s decisions about congestion and pur-
chases of ancillary services cannot be remedied by the PX, market participants
may bid strategically into the ISO’s congestion market to ensure that profit-
able transactions are not curtailed.
Bidding and Scheduling. The ISO markets began, largely, with only
supply-side bidding for energy and certain ancillary services. While demand-
side bidding for energy is allowed in some markets (currently the California
ISO and PX and New York, but planned for the other ISOs), the energy de-
7
Conversations with California ISO staff confirm this problem.
mand function is largely inelastic – that is, not price responsive – due to both
technical limitations and historic rate designs. As more price responsiveness is
introduced into demand bid functions (through installation of metering
equipment and technological advances in distributed generation and informa-
tion technology), there should be a reduction in both price volatility and the
potential for exercise of market power, particularly during peak hours.
The structure of bids is another market design issue that has attracted at-
tention. Bids in current ISO energy markets vary in the number of cost com-
ponents and required technical parameters, such as ramp rates, high and low
operating limits, and so on. In the so-called “one-part” incremental energy
bid, the bidder must factor its start-up, no load, and other costs into its day-
ahead energy bid for each megawatt-hour (MWh) offered.8 Even so, genera-
tors face the risk that they may not cover all of their costs in the auction. One-
part bids require generation owners to internalize this risk in some fashion,
which in turn increases their costs (use of the real-time market to make ad-
justments can eliminate part of this risk).9 One-part bids also result in ineffi-
ciency if they are the only costs the dispatcher considers in commitment.
In a three-part bid, the start-up and no-load costs can be separated out,
allowing generators to bid actual operating costs more precisely and allowing
for a more efficient unit commitment. In PJM and New York, generators are
guaranteed to at least recover all of their bid costs if they are committed to
run.10 This mechanism eliminates the uncertainty of whether a generator will
be committed and dispatched only to lose money, and it allows for a more
efficient dispatch.
Market-Clearing and Settlement Systems. Market-clearing rules and
settlement systems are the procedures that determine quantities produced and
consumed, who pays, and who gets paid. As discussed above, ISOs typically
operate multiple markets, including energy, several types of ancillary ser-
vices, and transmission products. There are two basic ways to clear these dif-
ferent markets, sequentially or simultaneously, with variations on each
method. In general, sequential auction markets clear each product separately
8
Currently, the California PX requires one-part energy bids without technical parameters, such
as minimum run times, ramp rates, and so on. This could be called a pure one-part bid. New
England currently requires one-part bids with technical parameters.
9
At its worst, the need to internalize risk in the one-part bidding system could lead to a greater
incentive to internalize via bilateral contract or merger to avoid higher transactions costs.
High market concentrations lead to market power concerns. Further, not being allowed to bid
marginal cost is an easy defense to an inquiry on market power abuse.
10
Generators will receive an “uplift” payment to recover their costs only if the revenues they
receive from the energy and ancillary services markets are less than their total bid costs.
in a sequence, even though several of the products may be alternative (substi-

tute) uses of the same generator. Variations of this approach were adopted
initially in California and in the interim New England market. In contrast, a
simultaneous auction, adopted in PJM and New York, clears the relevant
markets at the same time, minimizing the joint bid cost of providing energy
and ancillary services. This method explicitly takes into account that some
products are substitutes.
One of the clearest market design lessons from ISOs is that sequential
market clearing without product and/or quantity substitution is economically
inefficient and offers opportunities for strategic behavior. In California, en-
ergy, regulation, 10-minute spinning reserves, 30-minute non-spinning re-
serves, and replacement reserves are cleared in the order given.11 It has some-
times been the case that the price for ancillary services with lower production
costs exceeds the price of ancillary services (and energy) with higher produc-
tion costs (for example, providing spinning reserve requires using some fuel,
whereas providing non-spinning reserve requires simply being on stand-by).
The reason for some of these “price inversions” is that generators can strategi-
cally bid high prices in the last markets knowing that there will not be much
capacity remaining after the other markets clear. Hence, the ISO must take
these high price bids. To combat this design flaw, the California ISO has insti-
tuted a pre-processing algorithm, called the Rational Buyer Protocol, which
will allow it to substitute higher quality services for lower quality services if
and only if it reduces its ancillary service procurement cost. New England
also initially used sequential market clearing and experienced problems simi-
lar to California; temporary solutions have included rolling over of bids be-
tween substitute services, as in California, as well as price caps.
Experience with simultaneous market clearing is limited. The simultane-
ous market clearing method will largely, but not entirely, avoid the price in-
versions seen in the California ISO and New England markets because the
software used to clear the markets automatically clears all remaining arbitrage
opportunities. In addition, generators which bid strategically, as in California,
would be far less likely to be selected to provide ancillary services at the
higher price due to the greater substitution possibilities in the simultaneous
market clearing. While more complex computationally, simultaneous market-
clearing appears to be emerging as the better system from an efficiency stand-
point.
11
In California, regulation refers to automatic generation control, but is defined in terms of
whether generation output is increased (regulation up) or decreased (regulation down). Spin-
ning reserves is reserve capacity available in a specified time period from a generator syn-
chronized with the grid. Non-spinning reserves are reserve capacity available in a specified
time period from a generator not synchronized with the grid. Replacement reserves are re-
serves that can be available within 60 minutes.
The settlement systems can be characterized as either single settlement or

multi-settlement based on the number of temporal markets the ISO runs. In
California, there are three temporal markets: day-ahead, hour-ahead, and real-
time with a financial settlement for each. New York and PJM run two tempo-
ral markets: day-ahead and real-time. New England currently has only real-
time markets (bids are due day-ahead but financial settlement only takes place
at real-time prices), but is scheduled to implement day-ahead markets with
financial settlement. A clear market design lesson is that single-settlement
systems, which require generators to submit bids and stand-by day-ahead
while awaiting financial settlement at real-time prices, create problems in
scheduling and often require additional rules to constrain generator incentives
to change their bids. The multi-settlement system has been adopted by all the
ISOs in recognition of the value of the forward market as a financial hedge for
real-time conditions. Also, the forward market should facilitate demand-side
responses by giving demand that has bid to reduce load more time to react to
price signals.
Congestion Management and Pricing. In the ISO context, there are two
general ways to manage congestion: locational pricing and non-locational
pricing. Locational pricing can be sub-divided into approaches defined by the
level of aggregation used to calculate the price. In the typical “nodal pricing”
method, an energy price is calculated at each generation and load bus (node).
The transmission congestion price between any two busses is the difference in
energy prices at the busses.
At higher levels of aggregation, the busses in the system operated by an
ISO can be gathered into one or more congestion zones. Zones are intended to
be indicative of the historical pattern of congestion in the system on the pre-
sumption that congestion will take place between the zones with little or no
congestion within a zone. The price of congestion between zones is the differ-
ence in energy prices between the zones. If congestion occurs within a zone,
the costs of managing it (typically through generator re-dispatch) are shared
by all market participants inside the zone using a system of subsidies.12 This
intrazonal congestion management method could be considered a type of non-
locational pricing.
12
Several ISOs have attempted to operate as single zones (PJM, New England), but have sub-
sequently made the transition to locational pricing. If the single zone system has consistent
congestion between sub-regions (that is, should be at least two zones), this can create oppor-
tunities for generators to leave the spot market and use bilateral contracts to take advantage
of the system price. This was the experience in PJM before it implemented a nodal system;
the ISO was required to adopt administrative measures to curtail the bilateral transactions.
The zonal approach has been adopted in California, which currently has
two (soon to be three) congestion zones but is experiencing congestion that
should trigger new zones (also, each import point effectively creates a new
zone). The California ISO manages inter-zonal congestion through adjustment
bids submitted by generation and load. These bids indicate the price at which
the market participant is willing to be ramped up or down in order to alleviate
congested lines. If this fails to relieve the congested lines, then the ISO must
call on generators with cost-based contracts to relieve congestion.13
California manages intra-zonal congestion by re-dispatch (which incorpo-
rates the transmission constraints into the original, transmission unconstrained
dispatch) with the resulting costs averaged over load in the zone. Persistent
intrazonal congestion indicates that the zones are not properly defined; in ad-
dition, the averaging of congestion costs within the zone is inefficient, since
the congestion costs are also paid by participants not causing the congestion.14
In the long run, zonal pricing as practiced in California can lead to price
signals that distort decisions on siting new generation and transmission as-
sets.15 Neither maintaining fixed zones in the face of intrazonal congestion
nor continuous re-zoning are efficient methods of congestion management.
Zonal market design in California has been instituted in part under the
rationale that it lowers market power. Both in theory and practice this assump-
tions has been proved wrong. Market power cannot be reduced by the declara-
tions of large zones. If this were so, there would be no market power problem.
Transmission constraints and generation costs determine the size of the mar-
ket, not the declaration of zones. The California ISO rules recognize this by
providing for dispatch orders and out-of-market payments to generators in the
same zone separated by constraints.
In contrast, despite opposition from some generators and marketers, nodal
pricing has been adopted in New York and PJM and approved for New Eng-
land (these systems actually use nodal prices for generators and zonal aver-
ages for loads). Nodal pricing eliminates the problem of properly defining
13
These are called Reliability-Must-Run (RMR) contracts. RMR contracts are intended to
ensure that the ISO has sufficient generation capacity to meet various system contingencies,
such as congestion relief and voltage support.
14
The California ISO can create a new congestion management zone if the cost to alleviate
congestion over the previous 12 months exceeds 5 percent of the approximate annual reve-
nue requirement of the transmission operators. In order to be considered an active congestion
zone, the markets on either side of the congested interface must be “workably competitive”
for significant portions of the year.
15
For example, the Commission has rejected a California ISO proposal (Tariff Amendment
No. 19, filed June 23, 1999) that new generators upgrade transmission capacity to alleviate
intrazonal congestion which might arise from their entry on the grounds that it could create
further barriers to entry and market distortions. A similar New England proposal was also
rejected.
zones and the need to average the costs of any intrazonal congestion. In the
short run, load receives the proper price signals about how much to consume,
and the long-run decisions can be made much more easily. Even though loads
pay a zonal price, the nodal price information remains available for decision
making. For financial markets, nodal prices for a region can be aggregated
into fewer “hub” prices, which are weighted averages of the underlying nodal
prices. For example, PJM has two hub prices.
Transmission Rights. Transmission rights have traditionally been used to
reserve access to the transmission system and to ensure that energy transac-
tions would be curtailed only in extreme circumstances. These rights were
physical rights – the right to transmit physically a specific amount of power
over the system for the access charge paid. With the advent of congestion
pricing (whether zonal or nodal), most ISOs have provided both physical
rights and financial rights that can be used as a hedge against congestion costs
(the stochastic nature and potentially high cost of congestion makes financial
hedging necessary).16 In all the markets with locational congestion pricing,
payment of congestion prices is essentially a physical right to transmit be-
tween nodes or zones (although not a right that is bought in advance). On the
other hand, financial rights are typically purely financial mechanisms that
provide revenues but confer no physical priority. They can be traded on a sec-
ondary market.
For example, in New York and PJM, financial transmission rights give
the holder the right to collect congestion rents between a designated point of
injection and point of withdrawal, so that if a transaction incurs congestion
costs, those costs would be offset by the revenues from the financial right.
Auctions for these rights are typically held regularly. The California ISO has
implemented a similar type of zone-to-zone right, but which also confers
some physical priority.17
Performance of ISO Markets. As discussed above, none of the ISO
markets has reached a stable point in terms of market design; some are under-
taking major market re-designs while others are in the process of implement-
ing major components of their market design. There is a convergence in mar-
ket design in many areas: all the ISOs have implemented either sequential
auctions with substitutions or simultaneous auctions for energy and ancillary
services; most ISOs have established multi-settlement systems or will shortly.
Most ISOs offer some form of financial transmission right; in the East coast
16
Transmission rights can take the form of either options or obligations.
17
If the California energy markets fail to clear, the holder of a transmission right usually gets a
better position in the curtailment queue than a generator not holding a right.
ISO markets, nodal pricing is used for generation or is planned for future im-
plementation.
Given these ongoing changes, the preliminary performance of the markets
varies by product and time period. In transmission, the ISOs have recorded
few curtailments. There has been some concern, however, that the number of
bilateral transactions has decreased in nodal congestion management systems
(because point-to-point congestion may be difficult to hedge with the avail-
able transmission rights). The energy markets seem to be functioning fairly
well, although prices under certain system conditions reflect varying levels of
market power [8-13]. Entry of generation, transmission capacity expansion,
and demand-side bidding should lower prices and lessen volatility.
The ancillary service markets have been more problematic. Reserve mar-
kets in particular have experienced price spikes and price inversions, reflect-
ing the greater vulnerability of these markets to market power and to market
design flaws that exacerbate strategic behavior [9,12]. Temporary price and
bid caps and more permanent market re-designs should help solve some of
these problems. Other general market problems include limitations in soft-
ware implementation and technical capabilities (such as using telephone
rather than electronic communications for dispatch), and conflicts that emerge
when system operators depend on rules of thumb to dispatch the system rather
than the outcomes of the auction. In general, however, many market design or
implementation problems are amenable to satisfactory resolution, some
through admittedly short-term “band-aid” solutions, but most with a longer
term fix available. Business confidence is not equally robust in each ISO mar-
ket (PJM appears to be the market with the fewest problems to date), but
should increase as the markets mature.
Performance of Bilateral (Non-ISO) Markets. The largely bilateral
markets, especially those in the Midwest, have experienced many potential
reliability problems as evidenced by the frequency of curtailments under
Transmission Loading Relief (TLR) procedures. These may also be attribut-
able to the lack of independence of the system operator and market partici-
pants.18 Market participants have complained that they could not get access to
the transmission system even when capacity appeared to have been available.
As described below, Order 2000 requires the implementation of more efficient
congestion management practices.
18
The curtailment of transactions in the presence of prices 10 to 100 times the annual average,
due to TLRs and voltage reductions concurrent with power outages, indicate markets are not
working in harmony with reliability constraints. For example, in the summer of 1999 the
ECAR region with bilateral trading called 87 TLRs and the adjacent PJM ISO called three.
For a general review of these complaints, see [6].
2.2 Order 2000 and RTOs

A Regional Transmission Organization (RTO) is a transmission system
operator that is independent of market participants, controls transmission fa-
cilities within a region of appropriate scope and configuration, and is respon-
sible for operating those facilities to provide reliable, efficient, and non-
discriminatory service. All transmission owners must file a proposal to par-
ticipate in a RTO or provide reasons for delaying or avoiding participation.
Order 2000 explicitly notes that the designs for bid-based markets in the four
ISOs operational before the year 2000 should form a basis for the design of
RTO markets. Yet the open architecture adopted in Order 2000 does not pro-
pose a single market model and offers sufficient leeway for further experi-
mentation within the RTO design principles.
With respect to the unit commitment problem, the RTO has certain rele-
vant functions. The RTO must have exclusive authority for maintaining short-
term reliability. To fulfill this function, Order 2000 makes clear that the RTO
requires knowledge of the operational status of generators and load.19 This
includes control over interchange schedules, the authority to require re-
dispatch of generation connected to the grid, and approval over scheduled
outages.
The RTO will determine the required amount of each ancillary service
and the location where the service is to be provided. It will also act as pro-
vider of last resort of ancillary services. That is, market participants can self-
supply or purchase ancillary services from third parties, but the RTO must
have the capability to provide any residual. The RTO or a third party unaffili-
ated with market participants must provide real time energy balancing.
With regard to transmission, the RTO must ensure development of market
mechanisms for congestion management and must develop procedures to take
into account parallel path flows. The RTO will sell physically feasible, short-
and long-term, tradable transmission rights. The RTO may choose to expand
the transmission system and/or invest in advanced technology to increase
capacity.20
19
Such knowledge includes technical information supplied by generators such as ramp rates,
upper and lower operating limits, whether the unit is running or not, start-up times and time
between start-ups. In real-time and for day-ahead planning, the RTO must have information
on generator injections and load withdrawals of energy in order to balance the system.
20
In the comments on Order 2000 [6], various policy suggestions were made regarding increas-
ing transmission capacity, including overbuilding the transmission system (see Joskow com-
ments) and/or investing in the high tech Flexible AC Transmission System (FACTS) and
Wide-Area Measurement System (WAMS) to allow more robust competition to develop.
Finally, the RTO is required to monitor for market power abuses and
market design flaws. It should also evaluate and implement potential effi-
ciency improvements in the markets it operates.
Beyond these requirements and guidelines, specific market designs are
left to the RTO market developers (subject to the proviso that they not limit
the RTO’s ability to improve efficiency further). The remainder of this section
discusses some issues about the conceptualization of the role of the RTO with
respect to financial and physical transactions as well as the relationship of
RTO-operated auction markets in relation to other energy markets. In addi-
tion, some pressing market design issues are reviewed, including pricing of
reserves and inter-regional coordination. Section 3 then draws on the ISO ex-
perience and other sources to outline some principles for the design of day-
ahead RTO markets.
Relationship Between Physical and Financial Transactions. An issue
that has remained contentious in the preliminary design and operation of ISO
markets is the relationship between physical and financial markets –
specifically, the concern that the centralized ISO markets and nodal conges-
tion pricing would inhibit development of the decentralized financial mar-
kets.21 An important principle underlying the future RTO markets is that well-
functioning physical markets promote robust financial markets. For our pur-
poses, physical trades are trades that the RTO has registered as feasible, con-
sidering all other physical trades and required ancillary services. This includes
bids into the ISO markets, bilateral transactions and self-schedules that have
been cleared in the ISO day-ahead schedule (even though these day-ahead
transactions are actually financial contracts until physical delivery). Financial
trades are trades that are not physical trades, but take the form of forward con-
tracts, futures contracts, or options contracts. They are not considered physical
trades until they are confirmed as physically feasible by the RTO. Indeed, the
RTO should be concerned primarily with physical market transactions; it
would not operate purely financial markets and need not be involved in any
financial markets unless the transaction goes to delivery.
Financial markets can and must exist in harmony and equilibrium with
physical reliability markets. If not, the financial markets’ ability to reduce risk
is diminished. Multiple PXs and bilateral trading can fit easily into this market
velop. The latter appears more promising because it promises not only more capacity and
less greenfield construction, but also better system control (see EPRI and EEI comments).
21
One argument has been that uncertainty over nodal congestion prices, calculated hourly in
real-time, increases the risk of bilateral deals concluded prior to the hour. Another is that
some rules regarding three-part bids, in which the start-up payments made by the ISO are
averaged over all electricity load in the system (e.g., in New York), effectively amounts to a
subsidy to generators in the ISO auction market.
framework. Each PX would act as a single scheduler, submitting schedules

and technical information for generation and load to the RTO.
The market design of the physical market should allow full, but optional
interaction with financial markets. To become physical transactions, a market
participant need only self-schedule, that is, submit quantities to the day-ahead
market. If the transaction includes the necessary transmission rights and ancil-
lary services, no charges will be assessed. This allows for fully hedged finan-
cial transactions. In addition, payment would be received for any additional
service provided. Otherwise, the market participant will be billed for conges-
tion, losses, and ancillary services caused by the bilateral transaction.
If not self-supplied, a bilateral trader can place price limits on what it is
willing to pay for transmission and ancillary services. If the price limits are
not met, the transaction will not be scheduled. If all voluntary adjustments are
tried and reliability constraints are still not met, the transaction will be can-
celed in the day-ahead market. This gives ample time for parties to make ad-
justments. This cancellation avoids a potential TLR and the resulting schedule
is very likely to be physically feasible.
Physical markets provide real-time price signals and additional liquidity.
Without good price signals from the physical markets, the financial markets
can become unstable and encourage more speculation and less hedging.
Even though bilateral trading may be highly discriminatory (that is, sell-
ers may charge different prices for the same delivered product to different
buyers), the opportunity for buyers to participate in an efficient,
nondiscriminatory RTO auction market will tend to discipline the bilateral
market. RTO auction markets create options for all buyers and sellers and
thereby allow for a light-handed approach to the regulation of these
transactions. In sum, the benefits of well-designed RTO markets include
lighter-handed regulation of financial markets, more liquidity, less gaming
and risk, more visible prices, lower transactions costs, fewer curtailments, and
compatibility with financial markets.
Reserves Markets. The establishment of efficient ancillary service mar-
kets is an ongoing market design challenge. As the cost of reliability increases
and in the absence of a way to represent willingness to pay for ancillary ser-
vices, the RTO system operator can relax reserve margins and transmission
constraints. This is, in fact, written into the market rules in some ISOs; it re-
mains a contentious issue, largely because it has involved system operator
discretion that results in changes in market prices. The relaxation of these
constraints increases the probability of a system failure; as such, it should be
part of the operational parameters of the auction decided in advance of the
day-ahead auction so that actions of the system operator are not seen as arbi-
trary.
To be effective, reserves must be able to respond to loads that need them.
If transmission is congested between generation and load, reserves on the
generation side are of little help to the load side. Both transmission and gen-
eration can be used to meet reserve requirements. They have substitute
characteristics (strengthening transmission connections within a region can
lower the total generation reserves needed in a region) and complementary
characteristics (if the reserves for a load are not located at the same node,
transmission capacity between the reserves and load will also need to be set
aside). This set-aside is similar to the capacity benefit margin (CBM) concept.
The auction algorithm would set aside transmission capacity to allow reserves
to respond. The modeling of this process is very similar to the modeling of the
energy market itself. Prices for reserves would have a locational component
and the transmission price would reflect the set aside transmission capacity.
Using this locational method of allocating reserves, it could appear that
transmission capacity is being withheld. One method to deal with this is to
have a “use or lose” requirement of transmission rights. The set-aside “use” of
transmission for reserves markets would be considered a “use,” not withhold-
ing, and is under the control of the auction process and the operator. Dealing
with the combination of congestion constraints and soft reserve constraints
simultaneously requires operator independence and transparency in the mar-
ket environment, as a means to promote trust in the market.
Over time, as demand becomes more responsive to price, generation re-
serve margins can decline as offers to reduce demand can substitute for re-
serves. Currently the costs of reserves are averaged among all end users.
Some of these costs can be more directly assigned to specific generating units
and individual customers. For example, a unit with a good reliability record
should be responsible for a lower reserve margin or be charged less for re-
serves based on size and the historic probability of unit failure. Payments or
discounts that differentiate more reliable from less reliable generators should
be handled in the pre-day-ahead markets.
RTO and Expanded Inter-regional Coordination. Spatial boundary
conditions – also called the “seam” or interface problem – are becoming an
important design issue as trading between ISOs increases. ISO coordination
efforts on this matter are in the nascent stages. In terms of system representa-
tion, some ISOs have included a reasonably detailed representation of the in-
terconnection with control areas outside the ISO as part of the boundary con-
ditions. Approaches to the seam problem have included proposed interface
auctions for inter-control area exchanges [16,17]. In day-ahead markets, there
is some time to coordinate these interfaces offline via iterative trading rules.
Research into this problem generally is in its infancy. PJM, New York, New
England, and Ontario are examining a broad set of inter-regional market de-
sign and procedural issues, and these markets have signed a memorandum of
understanding on inter-regional coordination (on the need for such coordina-
tion generally, see [6]). Order 2000 anticipates that RTOs will assist in creat-
ing larger regional markets in which seams issues are resolved.
3. DESIGN PRINCIPLES FOR DAY-AHEAD RTO

AUCTIONS
In the early phases of electricity market design and implementation, the
various disciplines – notably economics and electrical engineering – have not
undertaken adequate inter-disciplinary research or sufficient professional ex-
changes. For example, ISOs have several times misunderstood the incentive
issues in electricity and ancillary service auction designs, as evidenced in the
remedial actions and market re-designs described in Section 2, above. At the
same time, market design has sometimes proceeded with the economics basi-
cally correct but without an adequate consideration of reliability and technical
constraints. As a result, there is sometimes little understanding of what basic
principles ought to underlie these complex markets. This section attempts to
elaborate such principles for day-ahead markets. These principles can be com-
pared with the market design principles in [14].
This section addresses the prospective RTO day-ahead market, which is
defined as the market in which the initial bidding to provide energy and ancil-
lary services for reliability, congestion management, and energy balancing
takes place. As is done currently in ISO markets, this market would be con-
ducted on the day prior to the dispatch day. The dual objectives of the day-
ahead market are to achieve economic efficiency and ensure system reliabil-
ity. The day-ahead market is a physical market where all expected balanc-
ing/ancillary services are scheduled. The design for the day-ahead market is
discussed below and it assumes that there is a real-time market, in which ad-
justments are made to energy and ancillary services reflecting the differences
between day-ahead expectations and real-time conditions.22
22
The real-time market will not be examined in depth here. Efficient market design requires,
however, that most principles be adhered to with respect to the relationship of the real-time
and day-ahead markets. Bids should be submitted separately into the real-time market, and
market prices based just on those bids. Deviations from the day-ahead market should pay the
real-time price unless there is a reliability problem. If a bidder does not deviate from the
day-ahead schedule, there are no additional costs to pay based on the real-time market. Fi-
nally, the market operator needs to keep the system in balance at the nodal level using bids
to the extent possible.
The following is a list of recommended design principles derived from the

foregoing analysis and experience to date, with short explanations and clarifi-
cations.
Principle 1: Maximize economic efficiency. The RTO auction objective is
to maximize economic efficiency through voluntary market bids, bilateral
transactions, and self-scheduling, given the physical and reliability con-
straints.
Not all current ISOs adopt this principle. For example, the California ISO
cannot adjust power schedules submitted to it by the California PX to improve
economic efficiency. This principle requires that if market participants use the
RTO markets, the resulting prices are efficient. The market-clearing proce-
dure will balance the system and the purchase of ancillary services using the
bids it has received. If the financial markets are efficient, there may be few
additional trade gains in this market and this market becomes a reliability
check. Efficiency requires that prices be consistent (e.g., no price inversions
due to market design flaws) and that arbitrage opportunities reflected in the
bids be exhausted.
Principle 2: Ensure physical feasibility of market transactions and system
reliability.
Without physical feasibility, reliability problems cannot be fully ad-
dressed. For ancillary services, the market design should require that genera-
tors committed to provide these services are located so that their capacity is
available when and where they are needed.
Principle 3: Remove disincentives to market participation. Participation in
an RTO market should involve low transaction costs and create minimal
additional risks.
Minimal participation in the market is the submission of generation and
consumption quantities (that is, bids which are taken at any price, or “at mar-
ket”). Any unit dispatched should be guaranteed bid-cost recovery.
Principle 4: Bidding protocols should promote flexibility of participation.

All market participants should be allowed, but not required, to submit
multi-part bids that reflect short-term marginal costs. Market partici-
pants should be allowed to self-schedule, that is, allowed to submit quan-
tity-only bids.
This principle requires that all resources have the option to bid a
reasonable approximation of their short-term marginal cost function, includ-
ing start-up, no load, and energy costs (in addition to technical parameters,
such as minimum and maximum load limits, ramp rates, and minimum shut-
down time). Although a bid function will seldom serve as a perfect match for
actual marginal (incremental or going forward) costs, a good approximation

should be available.
This principle will require changes in some current unit commitment
auctions, which allow only one-part energy bids (of course, in a multi-part
bidding rule, generators can still submit one-part bids, by bidding zero for
start up and no load). As discussed above, there are technical, financial, and
economic reasons for adopting multi-part bidding.
While the multi-part bid allows for more accurate representation of mar-
ginal costs and thus, in the absence of market power, should result in a more
efficient solution, it also results in a non-convex supply function. In turn, this
makes the market equilibrium and prices harder to derive. This complication
can be addressed and made manageable with some simplifying assumptions
about which generators are allowed to bid non-convex costs, whether some
parameters should be fixed for a specified period (such as ramp rates, maxi-
mum and minimum output), and what should be fixed in the bid function.
Demand bid functions are essentially the mirror images of generator bid
functions but will not be discussed in detail here. Consumers need more ex-
plicitly defined contracts to participate properly in the market and should be
allowed bid functions similar to the generators.
Principle 5: Make clear the distinction between financial and physical
commitments. If accepted, bids are financially binding. If needed for reli-
ability, bids are physically binding.
The RTO schedules are physical and financial commitments subject to
liquidated damages. That is, if market participants deviate from their com-
mitments, they are liable for making the affected market participants whole.
Under emergency conditions, the RTO may exact other penalties for non-
performance and/or issue perform-to-contract orders.
Principle 6: Minimize opportunities for arbitrage between different product
markets.
This principle addresses generally the problems that arise when genera-
tors can bid into different product markets (energy and ancillary services) that
are sequentially cleared. As discussed above, simultaneous markets eliminate
opportunities for arbitrage.
If the market design allows opportunities for lower quality products to be
priced higher than high quality products, then there may be cases where gen-
erating units are paid more for not generating than for generating. Preferably,
this should not be the case. The New England ISO attempted to establish this
principle administratively by proposing that the energy price always act as a
cap for operating reserves prices. This proposal was rejected by the Commis-
sion; administrative measures should be at best transitions to market designs

which can efficiently reach the same outcome. As discussed above, the simul-
taneous auction design will automatically allocate energy and reserves effi-
ciently, but cannot guarantee the elimination of price inversions in the pres-
ence of market power.
Principle 7. Prices should be unbundled where possible to minimize averag-
ing (i.e., socializing) of costs.
Averaged prices (whether for congestion, reserves, or other costs) do not
send the correct price signals for the entry of new generation. Uplift charges,
the mechanism used for passing through costs of energy and ancillary services
not covered by market prices, should be coupled with incentives for the RTO
to minimize the use of such charges.
Principle 8. Market-clearing information should be made available as soon
as possible. Fuller information on bids should follow with a suitable de-
lay.23
Market-clearing prices and quantities are the basic results of the bid ac-
ceptance process. They enable market participants and potential future market
participants to assess the market and plan their businesses efficiently. They
also allow market participants to spot and correct obviously erroneous bid
acceptance and rejection decisions.
Disclosure of individual bids should be made eventually, but not immedi-
ately. Such disclosure will allow detection of subtle bid acceptance errors and
it will also allow study of the market by independent analysts and market par-
ticipants. It may lead to the exposure of the exercise of market power. Imme-
diate disclosure of individual bids is undesirable because it might facilitate
collusion by the market participants. Immediate disclosure might reveal in-
formation about market participants who wish to keep their costs confidential.
After 6 months or a year, the information on individual bids has essentially no
value for collusion and discloses little new about any bidder’s current costs,
but the information would have high value for auditing and independent
analysis.
The auction software should be available to the market participant or pub-
lic at a reasonable cost. Improvements to the software are desirable, and the
best way to accomplish this is by making the software available with a set of
test problems.
23
The California ISO, PJM, and New England appear to set the benchmark with updated prices
every five minutes. At a minimum, prices and aggregate quantities should be available be-
fore the next round of bids with enough lead time to allow a reasonable response to the new
information.
Principle 9: Minimize the incentives for market participants to engage in

strategic behavior. The design should not favor market participants with
market power.
Market designs can only imperfectly address structural sources of market
power. An auction does not eliminate the ability of a large firm to withhold
capacity profitably. Auctions have been devised to “pay-off” market power,
but these require significant computation and may not be revenue sufficient.24
Hence, absent a market design solution, the basic problem of structural market
power has to be addressed using both structural remedies (vertical and hori-
zontal dis-integration, encouragement of entry) and regulatory remedies, such
as market power monitoring and mitigation. Of course, structural remedies
may be wrapped up in market design issues. For example, an ISO may seek to
promote rules on transmission interconnection for new generation which ap-
pear to favor incumbent generators.
While market design cannot necessarily mitigate structural market power,
it can certainly exacerbate it; market design can also create opportunities for
strategic behavior by generators other than the obvious large players. An ex-
ample, discussed above, is the sequential clearing of energy and ancillary ser-
vice markets without substitution in both the initial California and New Eng-
land markets. Even a small generator can try to take advantage of shortages
of certain types of reserves in this type of market to raise prices.
4. CONCLUSIONS
The considerable developments in the design of electricity markets over
the past few years have provided the groundwork for the next generation of
short-term markets. This chapter has emphasized that unit commitment mod-
els are now embedded in a variety of market contexts governed by an evolv-
ing regulatory framework that presents new requirements for modeling, in-
cluding incorporation and understanding of market design issues. The open
architecture promoted by the Commission allows for continued experimenta-
tion with RTO market design, but within parameters reflecting lessons learned
heretofore from the ISO markets as well as the non-ISO bilateral markets. The
market design principles presented in the paper are intended to reflect those
lessons.
24
The Vickrey-Clark-Groves (VCG) mechanism, which has been applied to electricity auctions
by [15], is a method to elicit truthful bids from players with market power.
The other primary argument in the chapter is that a well-designed RTO

day-ahead auction market with unit commitment complements “decentral-
ized” financial markets. The computer does the work of resolving reliability
constraints and will ensure that no offered trade gains are missed. Financial
markets are free to create whatever innovative deals they can. The only re-
striction is that if they go to delivery they must satisfy any reliability con-
straints. Also, the opportunity for buyers to participate in an RTO auction
market will tend to discipline the financial market. This will allow for lighter-
handed regulation of financial transactions. The paper identified several other
benefits of well-designed RTO markets: fewer curtailments, more visible
prices, lower transactions costs, and less gaming and risk.
The open architecture also allows for continued progress in efficient pric-
ing, so that causality has financial consequences – prime candidates are trans-
mission rights and reliability. Over time, as price signals are sent and acted on
in real-time, accurate pricing can allow the public good aspects of these mar-
kets to shrink in importance and the private good aspects to grow.
ACKNOWLEDGEMENTS
This paper reflects ongoing discussions among the authors and Carolyn
Berry, Judith Cardell, Benjamin Hobbs, Thanh Luong, David Mead, William
Meroney, and Roland Wentworth.
REFERENCES
1. P. Joskow and R.M. Schmalansee. Markets for Power. Boston: MIT Press, 1983.
2. F.C. Schweppe, M.C. Caramanis, R.E. Tabors, and R.E. Bonn. Spot Pricing of Electricity.
Norwell, MA: Kluwer Academic Press, 1988.
3. H. Chao and H.G. Huntington, eds. Designing Competitive Electricity Markets. Boston:
Kluwer Academic Press, 1998.
4. Federal Energy Regulatory Commission. “Promoting Wholesale Competition Through
Open Access Non-discriminatory Transmission Services by Public Utilities and Recovery
of Stranded Costs by Public Utilities and Transmitting Utilities.” Order No. 888, 61 FR
21, 540, May 10, 1996.
5. Federal Energy Regulatory Commission. “Open Access Same-Time Information System
(formerly Real-Time Information Networks) and Standards of Conduct.” Order No. 889,
61 FR 21,737, May 10, 1996.
6. Federal Energy Regulatory Commission. Regional Transmission Organizations. Order No.
2000, 89 FERC 61,285, December 20, 1999.
7. C.A. Berry, B.F. Hobbs, W.A. Meroney, R.P. O’Neill, and W.R. Stewart. Analyzing Stra-
tegic Bidding Behavior in Transmission Networks. Utilities Policy, 8(3): 139-158, 1999.
8. S. Borenstein, J. Bushnell, and C.R. Knittel. Market Power in Electricity Markets: Be-
yond Concentration Measures. Energy J., 20(4): 65-88, 1999.
9. California ISO. “Annual Report on Market Issues and Performance.” Prepared by the
Market Surveillance Unit, California Independent System Operator, June 6, 1999.
10 . R.E. Bohn, A.K. Klevorick, and C.G. Stalon. “Second Report on Market Issues in the
California Power Exchange Energy Markets.” Prepared for the Federal Energy Regulatory
Commission by The Market Monitoring Committee of the California Power Exchange,
March 9, 1999.
11. S. Borenstein, J. Bushnell, and F. Wolak. “Diagnosing Market Power in California’s De-
regulated Wholesale Electricity Market.” PWP-064, Univ. of Cal. Energy Institute, Berke-
ley, CA, Revised, March 2000.
12. P. Cramton and J. Lien. “Eliminating the Flaws in New England’s Reserve Markets.”
Working Paper, Department of Economics, University of Maryland, College Park, MD,
March 2, 2000.
13. J.E. Bowring, W.T. Flynn, R.E. Gramlich, M.P. Mclaughlin, D.M. Picarelli, and S. Stoft.
“Monitoring the PJM Energy Market: Summer 1999.” PJM Market Monitoring Unit, un-
dated draft.
14. R.D. Wilson. “Design Principles.” In [3].
15. B.F. Hobbs, M. Rothkopf, L. Hyde, and R.P. O’Neill. Evaluation of a truthful revelation
auction in the context of energy markets with non-concave benefits. J. Regulatory Econ.,
18(1): 5-32, 2000.
16 . M.D. Cadwalader, S.M. Harvey, W.W. Hogan, S.L. Pope. “Coordinating Congestion Re-
lief Across Multiple Regions.” PHB Hagler Bailly Inc., Navigant Consulting Inc., and
J.F.K. School of Government, Harvard University, Cambridge, MA, October 7, 1999.
17. B.H. Kim and R. Baldick. Coarse-grained distributed optimal power flow. IEEE Trans.
Power Syst., 12(2): 932-939, 1997.
Chapter 3
DEVELOPMENT OF AN ELECTRIC
ENERGY MARKET SIMULATOR
Atif Debs and Charles Hansen

Decision Systems International
Yu-Chi Wu
National Lien-Ho Institute of Technology and Commerce
Abstract: The paper outlines the development of an electricity market operation

simulator (EMOS) that can be used by competing market participants as
well as independent system operators and independent market operators
such as power exchanges. The uses include training personnel to make
control and operating decisions and to assess bidding strategies by energy
trading entities for energy and ancillary service products. We base the
simulator is based on the power system model used in the operator training
simulator. In the EMOS, market participants use unit commitment to as-
sess bidding strategies and/or to schedule generation in a pool-type elec-
tricity grid. The EMOS provides a framework for users to implement their
own applications with minimal effort and to model their competitors in a
generic manner.
1. INTRODUCTION
This chapter reports on efforts to develop a realistic electricity market op-
eration simulator (EMOS) for the emerging electric energy market. Several
key features of the market are:
(a) An independent system operator (ISO) coordinates supplies and de-
mands with a focus on system security and reliability;
(b) A market clearing entity determines which bids by market partici-
pants are accepted. In some cases this entity is an independent power
exchange (PX), such as the California Power Exchange. In other
cases, the ISO and the market clearing entity are integrated into one
organization (e.g., the PJM Interconnection); and
(c) Market participants (MPs) consist of a mix of unregulated and aver-
age-cost regulated entities. The unregulated entities compete against
each other and submit daily, hourly, and even real-time bids to the
market clearing entities and/or the ISO. They are generally generation
companies (GENCOs), energy service companies (ESCOs), or sched-
uling coordinators (SCs). The regulated entities are mainly distribu-
tion companies whose aim is to purchase the energy through the mar-
ket at competitive prices.
In order to compete, each MP has to follow the rules of the market area
involved and make appropriate “bids” to meets its goals. The bidding is nor-
mally based on the MP’s estimates of market prices and other factors associ-
ated with various risks involved.
The purpose of the EMOS is to simulate the overall market composed of
the entities just mentioned for very short-term operations (one day ahead, up
to real-time). The working tool is the power system model (PSM) of the
EPRI-operator training simulator (OTS) [1-7]. Each MP may have its own
model of the system, but with outputs consisting of its bids into the market.
The market clearing entity generates market clearing prices (MCP) and also
most of the information needed to schedule system operation. The ISO does
the final balancing of the market and the provision of needed “ancillary ser-
vices” for system balancing, security, reliability, and real-time control [8-10].
Through the use of many existing software packages, standard application
program interfaces, and communication protocols, the EMOS can be realized
in a variety of configurations. It is, therefore, usable by any of the entities
mentioned above.
How do unit commitment (UC) and other forms of system optimization fit
into the overall scheme of things? The approach given here uses UC at both
the MP level and the market clearing entity level, whenever applicable. The
EMOS also recognizes the fact that network optimization through the optimal
power flow (OPF) is essential for such market considerations as: congestion
management, reactive power/voltage scheduling, maximization of transfers,
and others.
Section 2 of this chapter provides a background summary of the emerging
electricity market with a focus on short-term scheduling. Section 3 provides a
conceptual framework for the development of MP models. Section 4 describes
the overall market model, and Section 5 contains concluding remarks.
An Electric Energy Market Simulator 41
2. ELEMENTS OF EMERGING ELECTRICITY

MARKETS
2.1 Generation Scheduling in the Competitive Electricity

Market
The traditional approaches of UC scheduling, hydro scheduling (HS), or a
combination thereof (so-called hydro-thermal coordination (HTC)) [11] are
being replaced rapidly by various forms of short-term competitive bidding
schemes [8]. The majority of such schemes are based on a 24-hour, day-ahead
scheduling process – a Forward Day-Ahead Market in the new jargon.
In all the new schemes of electric utility restructuring, the generation
component, as compared with the transmission and distribution components,
is strictly market-driven and competitive. As a result, the central dispatcher
plays the role of implementing the rules of generation scheduling as obtained
through a market auction (bidding) mechanism. The main differences between
the traditional and the “market-driven” approaches for generation scheduling
consist of the following [11-13]:
(a) Market prices are the result of the market design and associated rules
and regulations under consideration. These rules may vary from one
system to another. What complicates the picture is that the “products”
being traded are many – starting with MWH energy and continuing
through a host of ancillary services (A/S) and interconnected opera-
tion services (IOS). Some bidding schemes require full “multi-part”
bids. Others require single-part bids for energy, and either simultane-
ous or sequential bids for the ancillary and IOS services.
(b) Individual for-profit MPs aim to improve their short- and long-term
market positions through various strategies for increasing their prof-
its, controlling their market share, and/or other objectives. In the area
of short-term (e.g., day-ahead and same-day) scheduling their main
focus is on profit maximization. For the long-term strategies, they
tend to use different mixes of tools including financial ones (futures,
options, and swaps), or strategic ones (mergers and acquisitions, in-
vestments in other markets, etc.).
(c) The new markets can become more or less effective through the exis-
tence of catalysts such as a power exchange, scheduling coordinators,
energy brokers, service providers, e-commerce, and independent sys-
tem operators.
2.2 Network and Other “System” Issues in the New Envi-

ronment
Principles that have been followed in restructuring transmission include:
1. In order to allow generating companies – as well as customers and
distribution companies – to compete effectively, an open-access re-
quirement has been suggested and implemented. The functions of
generation scheduling have been “separated” from those of network
access and control.
2. Certain markets (e.g., the PJM Interconnection, New Zealand,
NYISO) use locational marginal prices (LMP). These depend on a so-
lution of both a resource scheduling program (unit commitment, hy-
dro-scheduling, or hydro-thermal coordination) and a security-
constrained optimal power flow (SCOPF). The resulting locational
prices reflect a combination of “charges,” such as transmission losses,
congestion, reactive support, reserves and, others [14,15]. (See Figure
1.)
Even in markets that do not use LMPs (or equivalent), network informa-
tion and the ancillary services protocols can be of great value for all market
players, e.g., the California ISO [8].
In short, different market players will have to adapt to the operating rules
for the systems served by different types of ISOs. In many cases, knowledge
of real-time and forecasted network conditions can be critical in assessing
various risk exposures and in evaluating hedging instruments.
2.3 Market Prediction is a “Learning Process”

Competitive markets do not behave like centrally dispatched systems.
Price discovery is attained through the auction mechanism, which has its own
dynamics [12]. Prices tend to be uncertain with a random component and a
systematic component. The random component stems from external uncer-
tainties (unpredictable weather, forced outage of a unit, system disturbance,
fluctuating fuel prices and interest rates, etc.). The systematic component is
related to physical constraints, market rules, auction clearing mechanisms, etc.
It is also related to “gaming” and market strategies by various players (e.g., to
manipulate prices and increase prices).
Within this systematic component of market price prediction, a “learning
process” becomes necessary to predict the behavior of the competition. Since
energy markets clear prices very frequently, the learning process can be for-
mally analyzed [13]. In most cases, however, it becomes the purview of “ex-
perts,” like skilled traders who earn their income from being very good learn-
ers.
2.4 Rationale of a Simulation-Based Approach

The foregoing discussion points to the fact that various analytical models
may not provide the full answer to all the questions. Given the complexity of
the power system and hence the electric energy market, there is a need in our
opinion for a simulation tool to help in the learning process. Such a tool must
by necessity represent both the systematic and random parts of the process.
From a modeling viewpoint, we have considered the following to be critical
[1-8,13]:
The power system under consideration should be modeled realisti-
cally.
Control and decision mechanisms by the main dispatcher (ISO, TSO,
etc.) should be customized in accordance with the rules of the system
under consideration.
The market bidding and price clearing mechanisms should also be
modeled for each respective market.
The transitions from one market regime to the next (i.e., day ahead to
hour ahead to real-time) should be clearly modeled and accommo-
dated,
The various tools used by individual market participants to play the
market should be modeled. The simulation should have a growing li-
brary of such tools as one learns about the bidding strategies of vari-
ous players.
The simulation should explicitly model random and systematic risk
exposures.
3. MARKET-PARTICIPANT MODELS
The simulation model recognizes that the power system is run as a result
of decisions made by groups of various market participants (MPs). At a
minimum, these MPs would represent the following entities:
Generating companies (GENCOs) who are competing against each
other,
Demand-side entities,
Independent System Operators (ISOs)
Energy Service Providers (ESCOs)
Figure 1 gives the functional structure of MP models. This structure is in
turn specialized for each of the categories of market participants as given in
Figure 2. Figure 3 in Section 3.6 displays the ESCO functional structure.
3.1 Generating Companies (GENCOs)

GENCOs comprise a variety of entities: (a) the generation component of
an integrated utility that is required to have functional separation of genera-
tion trading from other services, (b) a pure generating company with re-
sources scattered over several geographic areas, or (c) an independent power
producer (IPP). All these entities share similar needs and requirements. As a
result, we designed the GENCO model to consist of three basic modules: (1)
input module, (2) bid optimization module and (3) post-mortem analysis
module. These are now presented in some detail as they form the core of the
overall model.
3.1.1 Input Module
The GENCO will not be able to compete successfully unless it is able to

forecast energy and ancillary service prices. Other inputs are related to spe-
cific market rules for the clearing of bids. We discuss these in some detail
below.
The GENCO should be able to forecast market-clearing prices based on
historical information and its relative market position [12]. The approach used
in our modeling is a regression model, which relates market-clearing prices
to: (a) overall system demand and (b) available generation by the GENCO.
Given the system demand forecast, the GENCO can then predict the corre-
sponding price forecast. Several price-forecast scenarios are provided based
on uncertainties in the demand forecast. At a minimum, these are expected,
pessimistic, and optimistic scenarios with associated probabilities of occur-
rence.1 Price forecasts are done for both energy (MWh) prices as well as an-
cillary service prices.
A given GENCO sits somewhere between the two extremes of being (a) a
pure price leader or (b) a pure price follower [12]. Furthermore, the GENCO
may be constrained by socio/economic/regulatory factors, which would limit
its ability to manipulate price, even if it has market dominance.2 The output
1
An alternative to this is fuzzy modeling of the uncertainty.
2
For example, the GENCO may be concerned with anti-monopoly laws or excess profit regula-
tion, which may arise out of aggressive practices.
of this sub-module consists of: (1) limits on market share and (2) sensitivity
of market prices and energy supply to GENCOs bid prices. Both of these are
to be used in the bid optimization module.
3.1.2 Bid Optimization Module (BOM)
The BOM consists of two components, described below.
Component 1: Deterministic bid optimization based on specific input sce-

narios
The optimization is performed for two specific markets: (a) day ahead and
(b) hour ahead markets.3 The market rules vary from market to market. Our
approach models initially two types of market rules – (1) those which are
based on LMP considerations, e.g., the PJM Interconnection, and (2) those
based on bidding by scheduling coordinators4 using a PX, e.g., the California
ISO. Thus, there are four basic sub-modules to consider.
Day Ahead/LMP-Based. The bidding system for LMP-based day-ahead mar-

kets consists of a multi-part bidding process whereby the generation (and
load) bidders are required to submit typically the following:
Energy multi-block price curves which are piecewise constant
No-load price bids
Operating range (min and max MW)
Minimum up-time
Minimum down-time
The ISO control center as a result performs a combination of unit com-
mitment and contingency-constrained optimal power flow computations to
yield an optimal solution. The performance criterion is the minimization of
overall price to meet customer demand. When combined with elastic demand
bids, it is possible to use the same software to maximize social welfare. Ide-
ally, if every generation bidder submits his marginal cost data (plus a speci-
fied mark-up for profit), the result would be almost the same as that of a fully
integrated utility system. The constrained OPF would minimize system losses.
In case the ISO would pay for reactive power production, then the OPF would
attempt to maximize profit from the use of reactive power production.
From the perspective of the GENCO, the objective is to submit a bid that
maximizes profit – at least at this initial deterministic level. Thus for each
3
It is possible to extend the functionality of this module to cover both real-time bids and
longer-term forward market bids.
4
Scheduling coordinators (SCs) are allowed to submit strictly balanced generation/load bids
for the energy part of the market.
forecasted scenario, the control variables consist of the above bid parameters,
as long as they do not violate technical limitations (such as bidding minimum
up-time which is always greater than or equal to the technical minimum up-
time). One of the key issues here is to ensure that the GENCO Market-
Participant is able to model its own cost information, including their non-
convex price incremental cost curves.
Hour-Ahead/LMP-Based. Under this regime, the basic bidding process con-

sists of incremental/decremental bids to allow for adjustments of schedules
based on deviations from the day-ahead schedules. The main tool is strictly
the OPF with an objective function to maximize market surplus. From the
bidder’s perspective, the goal is to maximize profit.
Day-Ahead/SC/PX5-Based. In the SC/PX-based market, the typical bidding

process is decomposed into hourly energy price bid curves and other bids for
congestion management and ancillary services. However, there is a distinction
between a scheduling coordinator (SC) and the power exchange (PX).
In California, the SCs are exposed to network congestion and ancillary
service risks (and opportunities). Their main bid parameters to improve prof-
itability consist of incremental/decremental (INC/DEC) bids for congestion
management and ancillary service bids. In the simulation model the focus is
strictly on congestion management bids.
For GENCOs bidding through the PX, the main objective is to optimize
hourly energy price bid curves. Furthermore, there is a question here whether
portfolio bidding should be allowed. If portfolio bidding is permitted, then
the GENCO has to perform optimization at two levels: one that uses a single
curve for the entire company and a second that provides generation schedules
at the cleared market prices. Again, the GENCO has the opportunity to bid
into the congestion management and ancillary services market.
Hour-Ahead/SC/PX-Based. This is similar to the day-ahead bidding system

with the exception that portfolio bidding may not be allowed.
Component 2: Bid optimization through risk management

For each of the four bidding regimes, the GENCO optimization is per-
formed for a set of scenarios. In the input module, there is a probability asso-
ciated with each scenario. The key output here is a profit probability profile.
The analysis would select a scenario that yields bid prices for the best
profit/risk trade-off.
5
PX = Power exchange, whether independent or part of the central market clearing organiza-
tion.
3.2 Demand-Side Entities

For the purposes of the model under consideration, we distinguish the fol-
lowing demand-side entities:
(a) Major Load Serving Entities (LSE) (e.g., distribution company, or
commercial or industrial loads): An LSE may be able to bid a declin-
ing multi-block purchase price curves based on the composition of its
load and the demand-side management systems in place (typically in-
terruptible loads). The objective here for the LSE is to reduce the cost
of purchases based on its price forecasts (and contracts already in
place).
(b) Pumped Storage Power Plants: These plants try to combine their gen-
eration bids with their demand bids (for pumping). The result is a
composite optimization.
3.3 Independent System Operators (ISOs)

In our model, the ISO represents the component that implements the re-
sults of the bidding process to create a full schedule of system operation that
meets regulatory reliability and security requirements. Normally, the ISO con-
trols only its own area of responsibility (control area). Again, we distinguish
between two types of ISOs: LMP-based and SC/PX-based.
3.3.1 LMP-Based ISO
The model of the LMP-based ISO permits the following functions to be

performed:
Issuance of ISO-demand forecast for the day-ahead and hour-ahead
markets: The demand forecast is a net forecast for all demands to be
met by participants in the day-ahead and hour-ahead markets. In es-
sence, the forecast is adjusted by subtracting any bilateral contracts
and export/import contracts, which are also bilateral in nature.
Hourly optimization of the generating system and the network
through a combination of unit commitment and contingency-
constrained OPF, as described above, based on the totality of the bids
by the market participants. The optimization would yield LMPs,
which are published and released to the market.
Computation of congestion charges and firm transmission rights
(FTR) payments.
Similar functions for (1,2, and 3) the hour-ahead market (no unit
commitment is involved, however).
Real-time system control through automatic generation control
(AGC). The AGC utilizes an economic dispatch component which is
based on the hour-ahead bid prices, or alternative bid prices for sys-
tem regulation as an ancillary service
3.3.2 SC/PX-Based ISO
In this case, the following functions are supported:

Demand forecast similar to the one in the previous section for the
LMP-based ISO
Congestion management for interzonal congestion as well as intra-
zonal congestion for both the day-ahead and hour-ahead markets
Simultaneous and sequential ancillary services bid clearing
Automatic generation control based on ancillary service bids for sys-
tem regulation and load following
3.4 Power Exchange (PX) Model

There is no need for a separate PX model for the LMP-based system,
simply because the ISO in this case clears the market using the various multi-
part bids. The PX model for the SC/PX-based market is quite simple as it just
matches supply and demand curves. The PX, however, may engage in ancil-
lary service markets and options/futures markets and these have to be taken
into account.
3.5 Energy Service Provider (ESCO)

For the purposes of the market simulation model, the ESCO is a combina-
tion of a GENCOs and LSEs. A large ESCO may be bidding in multiple mar-
kets and these have to be modeled separately. The trading performed by the
ESCO may combine a portfolio of contracts/bids to manage risk.
3.6 Overall Market Model

Figure 3 shows the resulting EMOS overall market model. We have de-
signed the model for the following uses:
(a) Studies and training by each market participant entity described above
to improve its performance and meet its own needs.
(b) Studies at regional or national levels to evaluate different market de-
signs.
(c) On-line activities by MPs, including energy bidding strategies, ancil-
lary services bidding, market surveillance, etc., whenever the system
is linked to real-time data sources. Again, a given market participant
should be able to model its own activities using its own proprietary
systems.
3.7 Note on Implementation

The EMOS is implemented using standard software applications devel-
oped primarily by EPRI and other developers. These include:
EPRI-OTS by EPRI for the primary PSM modeling

EPRI-DYNAMICS unit commitment program for various market par-
ticipant modules
EPRI ANN-STLF for short-term load forecasting
DSI-OPF for the multi-area optimal power flow interfaced with
EPRI-DYNAMICS
Other components of the EMOS are still under development. The overall
system uses an effective database system which is compliant with EPRI’s
common information model [1].
4. CONCLUDING REMARKS
The development of an overall market model for the electric energy in-
dustry under restructuring is a major challenge mainly because the electric
physical system is highly complex, changes constantly, and creates a signifi-
cant amount of physical risk to market participants (MPs). The overall market
model consists of an interacting population of MPs. Market rules and designs
are aimed at coordinating all efforts through ISO’s. The model implementa-
tion is based on such developments to allow for effective interoperability and
plug-compatibility among various applications.
We expect that the developments reported here can be used in a variety of
ways by each MP type: GENCOs, LSEs, ISOs, PXs and ESCOs. The main
uses consist of:
Just-In-Time training of personnel – traders, dispatchers, operations
engineers and managers
Enhanced mechanisms for improving competitive positions by traders
Ability of ISOs and market designers to analyze the behavior of mar-
ket participants and test improvements in market rules and overall
market design
Unit commitment as such is used as a tool whenever applicable. The main
users of UC are the GENCOs, ESCOs, and some ISOs with multi-part bidding
market rules.
ACKNOWLEDGEMENTS
The developments reported in this chapter are under the sponsorship of
the National Science Foundation (NSF) under SBIR Phase II Grant, DMI-980
10161. The EPRI-OTS developments and related enabling technologies have
been partially sponsored by EPRI, Palo Alto, California.
REFERENCES
1. A. Debs, Y.-C. Wu, and C. Hansen. “Enhancement of the EPRI-OTS for the Restructured
Electric Utility,” Project Reports, Decision Systems International, Atlanta, Georgia, under
Small Business Innovations Program of the National Science Foundation.
2. A. Debs and C. Hansen. “The Total Power System Simulator: A Comprehensive Tool for
Operation, Control and Planning.” In Proc. Arab Electricity ’97 Conference & Exhibition,
PennWell Europe and DSI, Bahrain, 1997.
3. A. Debs and C. Hansen. “The EPRI-OTS as the Standard for Training and Studies in the
New Era: Strategy for Global Application.” Presentation at the First Asia Pacific Confer-
ence on Operation and Planning Issues in the Emerging Electric Utility Environment,
sponsored by EPRI, Kuala Lumpur, Malaysia, 1997.
4. M.K. Enns, et al. Considerations in designing and using power system operator training
simulators. EPRI EL-3192: 1984.
5. V. Calovic and A. Debs. “The Fully Integrated DTS for ESB, Dublin. ” In Proc. EPRI
First European Workshop on Power System Operation and Planning Tools, Amsterdam,
The Netherlands, 1995.
6. C. Hansen and M. Foley. “Power System Model Enhancement at ESB, Dublin.” In Proc.
EPRI First European Workshop on Power System Operation and Planning Tools, Am-
sterdam, The Netherlands, 1995.
7. S. Lutterodt and A. Debs. “Issues in Use of Operator Training Simulators.” Paper pre-
sented at the CEPSI Conference, Kuala Lumpur, Malaysia, 1996.
8. A. Debs and F. Rahimi. “Modern Power Systems Control and Operation in the Restruc-
tured Environment.” Class notes for intensive short course by Decision Systems Interna-
tional, San Francisco, CA, 1999.
9. A. Debs, P. Gupta, C. Hansen, A. Papalexopoulos, and F. Rahimi. “System Planning in
the Context of Competition and Restructuring.” Class notes for intensive short course by
Decision Systems International, San Francisco, CA, 1996.
10. M. Ilic, F. Graves, L, Fink, and A. DiCaprio. Operating in the open access environment.
Elec. J., 9(3): 61-69,1996.
11. A. Debs. Modern Power Systems Control and Operation. Norwell, MA: Kluwer Aca-
demic Publishers, 1988.
12. B. Sheblé. Computational Auction Mechanisms for Restructured Power Industry Opera-
tions. Boston, MA: Kluwer Academic Publishers, 1999.
13. A. Papalexopoulos. “Design of the Wholesale Market in the USA.” In Proc. EPRI Sec-
ond European Conference – Enabling Technologies and Systems for the Business-Driven
Electric Utility Industry, Vienna, Austria, 1999.
14. Y.C. Wu, A.S. Debs, and R.E. Marsten. A direct nonlinear predictor-corrector primal-dual
interior point algorithm for optimal power flows. IEEE Trans. Power Syst., 9(2): 776-883,
1994.
15. H. Kim and R. Baldick. Coarse-grained distributed optimal power flows. IEEE Trans.
Power Syst., 12(2): 932-939,1997.
16. A. Debs. “The OPF in the Deregulated Environment.” Lecture notes for intensive short
course by Decision Systems International: Modern Power Systems Control and Operation,
San Francisco, CA, 1996.
17. A. Papalexopoulos, et al. Cost/benefits analysis of an optimal power flow: The PG&E
experience. IEEE Trans. Power Syst. 9(2): 796-804, 1994.
18. Y.C. Li and A.K. David. Optimal multi-area wheeling. IEEE Trans. Power Syst., 9(1):
288-294, 1994.
Chapter 4
AUCTIONS WITH EXPLICIT DEMAND-SIDE

BIDDING IN COMPETITIVE ELECTRICITY
MARKETS
A. Borghetti
University of Bologna, Italy
G. Gross
University of Illinois at Urbana-Champaign
C. A. Nucci
University of Bologna, Italy
Abstract: This paper focuses on the development of a model for and the simulation of
electricity auctions with demand-side bidding (DSB) explicitly considered. We
generalize the competitive power pool (CPP) framework developed in [1] to
include DSB. In order to allow customers to play a proactive role in the price
determination process, the DSB provides the opportunity for them to submit
bids for load reductions in specific periods. We study the behavior of DSB in-
clusion in electricity auctions simulations obtained with a specially developed
Lagrangian relaxation scheme that effectively takes advantage of the structure
of the problem. We present some numerical results for a 24-hour simulation on
a small system. This case study is effective in illustrating the various economic
impacts of DSB including system efficiency effects, changes in the system
marginal price and the load recovery effects.
1. INTRODUCTION
This paper deals with the development of a model for the simulation of
electricity auctions with demand-side bidding (DSB) explicitly considered.
The model is a generalization of the competitive power pool (CPP) frame-
work developed in [1] to include DSB. Such a mechanism has been used at
times in the England & Wales Power Pool by large industrial consumers
who can offer their ability to reduce load directly to the Pool and receive a
payment for actually making such a reduction possible [2,3].
In the CPP with only supply-side offers, the point of intersection of the
aggregate supply curve and the fixed forecasted load determines the market-
clearing price. To improve competition, the consumers may be allowed to
participate in the market price definition process by providing them the op-
portunity to submit bids. Without DSB, the system marginal price (SMP) is
determined by constructing the supply curve from the offers submitted by
the supply-side bidders and determining the uniform price paid to all bidders
who are selected to serve the fixed forecasted load demand. The basic idea in
the application of DSB is to use the demand profile to lower SMP by cutting
load during peak periods. DSB may take advantage of large industrial cus-
tomers who can cut load or shift load to benefit from lower electricity prices.
Since industrial customers are not expected to reduce significantly their daily
energy consumption, the overall effectiveness of DSB depends on the how,
the how much, and the when of the energy recovery. Clearly, the more flexi-
bility such a customer has, the more possible it becomes for him to partici-
pate in DSB.
The demand recovery is typically a function of the weather, of the eco-
nomic conditions as well as of the number of reduction periods, and the
amount of load cut in each period. In [4] an analysis has been presented of
data relevant to a four-year period for a regional electricity company (REC)
in the United Kingdom that quantifies the extent of intertemporal substitu-
tion in electricity consumption across pricing periods within the day for five
groups of different industries. Every industry shows electricity substitution
possibilities with its own firm-specific characteristics. We have constructed
an econometric model from the results of this analysis. This model is for the
use of the REC to formulate its demand-side bids and as such takes into ac-
count the specific pricing scheme and rules adopted in the England & Wales
Power Pool. There are additional data available on the characteristics of the
load recovery from experiences gained with the application of commer-
cial/industrial and residential load control programs [5-7]. In this case, there
is electricity substitution only across adjacent load periods, i.e., the energy is
recovered in the periods that follow right after the reduction periods.
In this paper, we simulate and study the inclusion of DSB in electricity
auctions to examine various economic and policy aspects in DSB. The simu-
lation tool implements a Lagrangian relaxation-based scheme, which takes
advantage of the problem structure with the inclusion of DSB. The suffi-
ciently detailed representation of the supply-side bidders allows the model-
ing of the unique physical characteristics of the power generation system. In
particular, it takes into account the main operating considerations including
the operational limits and up- and down-time constraints. The supply-side
bidders are also allowed to include in their bids a separate price for start-up
in addition to the price for the MWh commodity. In the simulation tool, the
impact of DSB is taken into account by incorporating specific energy con-
Auctions with Explicit Demand-Side Bidding 55
straints and load recovery into the Lagrangian relaxation algorithm. Re-
searchers have applied Lagrangian relaxation methods widely in recent years
for the solution of unit commitment problems for large-scale systems due to
their ability to include a more detailed system representation than would be
possible with other techniques [8].
Some aspects of DSB have been analyzed in [9,10]. The studies indicate
that the specific characteristics of the demand-side bidders have a substantial
impact on the system economies. However, additional work is required to
better quantify such impacts. This chapter aims to bring additional insight to
the study of DSB. The principal combinations are the development of a gen-
eralized framework for the analysis of various aspects of DSB and the illus-
tration of the salient characteristics through the simulations obtained with the
tool developed for the implementation of this framework.
The chapter is organized as follows. In section 2, we present the gener-
alization of the CPP structure with the inclusion of DSB into the CPP frame-
work and a description of the Lagrangian relaxation-based algorithm for its
solution. In section 3 we present some numerical results obtained for a 24-
hour simulation on a small system. We compare costs and price signals ob-
tained from DSB in addition to supply-side bidding (SSB) to costs and price
signals obtained from the use of only supply-side bidders. The paper
provides an analysis of the numerical results. Section 4 offers final thoughts
on this research, and the Appendix presents details concerning the imple-
mented code and the power system data.
2. GENERALIZATION OF THE COMPETITIVE

POWER POOL FRAMEWORK FOR INCLUSION
OF DEMAND SIDE BIDDING
We start with the extension of the CPP framework developed in [1] to
include DSB in addition to SSB. With the addition of DSB, the CPP dis-
patcher problem is to select the winning bids and offers from the set of both
supply-side offers and demand-side bids. We use henceforth the term bid to
refer to both the supply-side offer and the demand-side bid and we refer to
each player as a bidder. Each bid has three components:
the bid variable price which describes the per hour price as a
function of MW provided/reduced. The function is here assumed to
be a quadratic or piece-wise linear function;
the bid start-up price which is charged whenever the bidder is
committed; and
the bid offered capacity which is a vector whose
component is the maximum capacity offered by the bidder for use

in period t, expressed as a fraction of the maximum output.
The supply- and demand-side bids have different characteristics.
In particular, for the case of SSB, we have made the assumption that the
operational data1, including the minimum and maximum output and
and minimum up- and down-times and need to be provided by
each bidding plant i.
Moreover, the bid start-up price which is charged whenever unit i
is started in period t, is assumed to be a function of the down time of unit i
at period t-1
where is the thermal time constant of unit i, I is the number of supply side
bids, is a constant related to the cold start charge, and is a constant
related to the fixed operating and maintenance charges [11].
The characteristics of the demand-side schemes require the specification
of the following information for each demand-side bid j,j=1, ..., J:
the bid start-up price that is considered a constant value
the subset of the operator-designated load-reduction periods
that is the subset of the periods in which bidder j may un-
dertake load reduction
the minimum and maximum demand and that can be re-
2
duced by bid j
the subset of the operator-designated load-recovery periods that
is the subset of the periods in which bidder j may under-
take a load recovery
load recovery at period h, that is related to all the load reductions
in each reduction period of by the following expression
1
We do not take ramp and network constraints into account in the model discussed in this
chapter.
2
Note that a requirement is introduced in the problem formulation to ensure that the sum of
the maximum load reductions of all the demand-side bids in a period t cannot exceed the
load forecasted for that period.
where represents the load recovery ratio in period h of the load

reduction in period t.
The above representation of load recovery is similar to that used in [9]. It
allows the representation in a flexible and simple way of the electricity sub-
stitution characteristics of consumers who may participate in load modifica-
tion. It allows, moreover, load recovery to occur both in pre- and post-load
reduction periods. In [9] it is pointed out that the beneficial effects of DSB
are greatly attenuated if some bidders recover their load in the same periods
in which others are reducing it. Therefore, we shall assume that there are no
overlapping periods in which the load reductions and the load recoveries
may cancel each other out.
By taking into account both load reductions and load recoveries, the ac-
tual load in reduction period t is given by
and the actual load in recovery period h is given by
where and are the forecasted loads in period t and h, respectively.

The additions of the relations for the DSB participation do not change
the basic structure of the framework in [1]. Consequently, we can extend the
Lagrangian relaxation framework developed in [1] to the more general
model described here. For this purpose, we use the following notation:
I: the number of supply-side bidders
J: the number of demand-side bidders
T: the number of time periods
D: the T-dimensional vector of the load demands in each period t in the
scheduling horizon
the T-dimensional vector of the zero-one variables indicating
whether supply-side bidder i is committed in period t or not
the T-dimensional vector of the zero-one variables indicating
whether demand-side bidder j is committed in period t or not
the T-dimensional vector of the amount of power that the supply-
side bidder i is producing in period t
the T-dimensional vector of the amount of power that the de-
mand-side bidder j is reducing in period t.
The CPP dispatcher problem determines the most economic commitment

that satisfies the forecasted demands D without violating physical and oper-
ating constraints of the generation equipment and demand specifications.
subject to initial conditions and, by taking into consideration equa-

tions (2) and (3), subject to
We refer to the optimization problem in equations (4)-(7) as P.

While the demand-side bids are defined similarly to supply-side offers,
problem P differs from problem with only supply-side bidding. In fact, the
incorporation of DSB adds the two decision variables w and y that are
characterized by specific constraints specified in (7).
The solution approaches with and without DSB are similar. Also with
the introduction of new DSB decision variables w and y, the Lagrangian re-
laxation approach allows the decomposition of P into I+J sub-problems as-
sociated with each supply-side and demand-side bid. With the inclusion of
DSB, we use the following Lagrangian dual function
or, more explicitly, taking into account (4),

subject to the specified initial conditions and to constraints (6) and (7).
t= l,...,T, are non-negative Lagrange multipliers. The dual function is rear-
ranged as
where, for supply-side bid i
and, for demand-side bid j, by explicitly taking into consideration equations

(2) and (3),
subject to the specified initial conditions and to the constraints (6) and (7).
Problems (10) and (11) are the I + J sub-problems, one for each bidder,
that need to be solved. In order to find the optimal values of the multipliers
to use in equations (10) and (11) the following dual problem is solved
In fact, L * provides a lower bound of the final solution of the primal prob-
lem P [12].
As a by-product of the process of maximizing L we obtain the optimal
Lagrange multipliers and a system schedule {u, w, p, y} resulting from the
solution of the Lagrangian relaxation for In certain cases, this system
schedule does not satisfy the conditions imposed by the demand constraint
and, therefore, a practical approach for computing a near-optimal schedule
has to be implemented [13].
It may be useful to note that the optimal schedule, obtained with the so-
lution of the proposed framework, is the one that minimizes the total “pro-
duction” costs, i.e., the costs paid to satisfy the forecasted demand, on the
basis of the bids of the supply-side and demand-side bidders. The solution
can differ, in general, from the schedule that minimizes the consumer pay-
ments, for the various reasons cited in [14].
3. NUMERICAL RESULTS
We implemented the solution approach based on the extended frame-
work for the inclusion of DSB as a software package. Some of the salient
aspects of this package are summarized in Appendix A. This package has
been used to perform many numerical studies to assess the characteristics
and the impacts of DSB in a competitive power pool.
To illustrate the application of the generalized framework, we present
some typical results of our simulation studies. We use the 10-unit system in
[13] as the supply side of the resources. The operational data and the bid
functions used for the supply-side bidders are given in Appendix B. The be-
havior of the power pool is simulated for the 24-hour period load profile of
Table B.1 in Appendix B. Figure 1 plots the hourly loads. The load reduc-
tions are dispatched on the basis of the demand-side bids that represent the
prices at which the bidders wish to reduce their consumption by the specified
amounts. The bidder that is scheduled to reduce its consumption in a certain
period is also allowed to recover its load partially or totally in other periods.
The model specifies the periods in which load reduction and load recovery
are allowed. These periods have been defined in the previous section as the
operator-designated load-reduction periods and load-recovery periods
To show the influence of such a constraint, the results of a sensitivity
study are presented. The study was carried out by varying only the following
two triggering levels:
the load-reduction triggering level as a fraction of the peak load,
specifies the load level above which the load reductions are allowed;
and
the load-recovery triggering level as a fraction of the difference

between the peak and the base load, specifies the load level below
which the load recoveries are allowed.
For each given and values, the corresponding permitted reduction and
recovery periods, and respectively, are then established. The load
recovery is considered uniform throughout the recovery period.
We discuss the results of several simulations on this system with a single
demand-side bidder in addition to the ten supply-side bidders. The demand-
side bidder limits are and The demand-side bid
is 1 $/MWh with 1 $/h being the no-load bid. These values were selected
specifically to provide a lower cost per unit of energy to the CPP dispatcher
on the demand side than the supply-side. We start out with a case in which,
the DSB is allowed whenever the load exceeds For this case is
set at 0.5. The periods when load reductions are allowed, are
The corresponding recovery periods are
and the recovery is implemented to spread the re-
covery of load in equal amounts in each of the periods in
To assess the impact of DSB, we use the no-DSB case as the reference
basis. We show in Figure 2 the modified loads as a result of DSB under dif-
ferent levels of load recovery. These levels of load recovery are indicated by
the recovery fraction value, that is the total amount of load recovery over the
total amount of reduced load. In Figure 3, we depict the corresponding im-
pact on the system marginal prices (SMP). For the supply-side bidding strat-
egy used, the system marginal prices track closely the modified system
demands Furthermore, DSB, in effect, acts to “smooth” out the price
variations across periods.
We now evaluate the impacts of varying some of the parameters in DSB
implementation. One parameter that has a major impact on deployment of
DSB is the load reduction recovery factor. This parameter is the ratio of the
recovery energy to the cut energy. As this factor decreases from 1 to 0, the
savings increase with respect to the costs for the case without DSB (what we
shall call reference costs). Figure 4 shows results for the test system.
The assessment of DSB on the variability of the SMP is important. DSB
tends to impact prices in different periods lowering the prices in periods with
higher loads and increasing them in periods with lower loads. We evaluate
the hourly average SMP, as well as the volatility impacts in terms of the
standard deviation as a function of recovery fraction. Figure 5 gives these
results with the no-DSB case as reference. While there is a reduction in the
volatility, a reduction in the average value of the SMP is not obtained for all
recovery fractions. The reason is that the schedule is obtained without a

minimization of consumer payments. In the case of DSB without load recov-
ery, the reduction of the volatility is only due to the reduction of the SMP in
the peak periods. Therefore, the reduction in the volatility is lower in this
case than for the case in which there are both the reduction in the price in the
peak periods and the increase of the price in load recovery periods. We note
that at the higher fraction of load recovery, DSB is committed in only a lim-
ited number of periods, and as such both the savings and the reduction in
price volatility are limited.
An important aspect to examine is the impact of different triggering lev-
els for the definition of the operator-designated load-reduction and recovery
periods. For the test system we evaluate the savings with respect to the refer-
ence DSB case for sixteen different triggering levels under the fixed load
recovery fraction of 0.8. Table 1 gives the data, and Figure 6 presents the
plots as a function of the reduction triggering levels for different values
of Due to the fewer number of periods in which the demand-side bidder
may be committed, higher reduction triggering levels result in lower savings.
On the other hand, the effect of the recovery triggering level is less uniform:
as the recovery triggering level increases, the number of recovery periods
increases and the recovery of the load decreases in each recovery period. In
the case of the higher recovery triggering levels, the recovery periods are in
the shoulder load periods resulting in a decrease in the savings compared to
the lower values of
We examine the impact on loads and SMPs for as a function

of Figure 7 shows the load modifications, and Figure 8 depicts the im-
pact on the SMPs. These plots make more concrete the observations in the
previous paragraph on the impacts of different
We next discuss the impacts of a single-supply unit’s bidding strategy on
the market. We focus on the unit 9, and we obtain the results in Figures 2-8
with the bid of unit 9 declaring unavailability in periods 10 to 21. A supply-
side bidder for various reasons or market objectives may select such “strate-
gic” behavior. When unit 9 changes its bid and declares availability for the
periods 10-21, the total costs without DSB change $537,271. This change
represents a saving of $3,186 with respect to the reference case of no-DSB

without the participation of unit 9 for the periods 10-21. From Figure 4, we
can see that such a saving is lower than the saving obtained with DSB recov-
ering the entire energy. We may, therefore, interpret the impacts of DSB to
include its ability to mitigate the market power on the supply-side.
Thus far, we have concentrated on displaying the salient characteristics
of DSB through a single demand-side bidder. The rationale for this was to
provide a good insight into the interaction of the DSB with the electricity
market. We next study the case with multiple DSB bidders. To illustrate the
behavior of DSB with multiple demand-side bidders, we consider two cases:
(i) five DSB bidders and (ii) nine DSB bidders. We compare the behavior in
these two cases to that in the case of the single bidder discussed above. For
these cases, the demand-side bidders have and the load
recovery fraction is 0.9. Recall that for the single DSB bidder, the bid is
1 $/MWh with 1 $/h being the no-load amount and In case
(i) the bids are 1, 2, 3, 4, and 5 $/MWh, respectively, with 1 $/h the no-load
amount for each bidder. In case (ii) the bids are 1, 1.25, 1.5, 1.75, 2., 2.25,
2.5, 2.75, and 3 $/MWh with 1 $/h the no-load amount for each bidder. In
the two cases is 200 MW for each DSB bidder.
Figures 9 and 10 provide plots of the results obtained with the five and
nine demand-side bidders together with the results obtained with the single
DSB bidder. Figure 9 gives the impacts on the load modification, and Figure
10 displays the impacts on the hourly SMPs. The number of demand-side
bidders has been chosen to exceed the number that is committed in the mul-
tiple-bidder cases: in (i) only two are committed, and in (ii) only five are
committed with the fifth being committed only for a single period in the
hour 10.
4. CONCLUSIONS
This chapter focused on the integration of DSB into electricity markets.
We discussed the development of a general framework for the inclusion of
DSB in addition to supply-side bidders. We used the Lagrangian relaxation
based solution approach to examine the impacts of DSB. The tool we devel-
oped for simulation is very useful for investigating a wide spectrum of pol-
icy issues. In our case, we have used this tool to examine the ability of DSB
to mitigate the potential for exercise of market power by the supply-side
bidders. In addition, the impacts of DSB on smoothing system marginal
prices and on mitigating price volatility are important attributes as indicated
by the numerical results.
The numerical studies of the paper provide good insights into various
aspects of DSB. In particular, it is important that load reductions and load
recoveries do not cancel each other out. Therefore, the specification of per-
mitted load reduction periods and permitted load recovery periods has been
added to the model. The results of a sensitivity analysis show that the speci-
fication of appropriate triggering levels, that define the permitted load reduc-
tion periods and permitted load recovery periods, is particularly critical to
make effective use of DSB.
In the chapter, we provide a simple example to study multiple DSB
players. The impacts of competition and possible collusion among DSB
players need to be investigated in more depth. Also, there are many addi-
tional policy issues that remain to be examined. One key issue for further
study is that of DSB remuneration, including the needs for such incentives
and the levels at which to set them. The changes in supply-side bidder be-
havior brought about by the inclusion of DSB are another key issue that re-
quires investigation.
ACKNOWLEDGMENTS
We are grateful for the financial support for the research reported here
from the Italian National Research Council. George Gross received support
from the Power System Engineering Research Center administered by Cor-
nell University.
REFERENCES
1. G. Gross and D.J. Finlay. “An Optimization Framework for Competitive Electricity
Power Pools.” In Proc. PSCC, 815-823, 1996.
2. F.A. Wolak and R.H. Patrick. The impact of market rules and market structure on the
price determination process in the England and Wales electricity market. Working Paper
PWP-047, University of California Energy Institute, 1997.
3. D.W. Bunn. Demand-side participation in the electricity pool of England and Wales.
Decision Technology Centre report, London Business School, 1997.
4. R.H. Patrick and F.A. Wolak. Estimating the customer level demand for electricity under
real time pricing. Working Paper, Department of Economics, Stanford University, 1997.
5. D.V. Stocker. Load management study of simulated control of residential central air
conditioner on the Detroit Edison Company system. IEEE Trans. Power Apparatus Syst.,
4: 1616-1624, 1980.
6. A.I. Cohen, D.H. Oglevee, and L.H. Ayers. An integrated system for residential load
control. IEEE Trans. Power Syst., 3: 645-651, 1987.
7. C.N. Kurucz, D. Brandt, and S. Sim. A linear programming model for reducing system
peak through customer load control programs. IEEE Trans. Power Syst., (11)4: 1817-
1824, 1996.
8. A.J. Svoboda and S.S. Oren. Integrating price-based resources in short-term scheduling
of electric power systems. IEEE Trans. Energy Conversion, (9)4: 760-769, 1994.
9. G. Strbac, E.D. Farmer, and B.J. Cory. Framework for the incorporation of demand-side
in a competitive electricity market. In IEE Proc. – Generation, Transmission and
Distribution, (143)3: 232-237, 1996.
10. G. Strbac and D. Kirschen. Assessing the competitiveness of demand side bidding. IEEE
Trans. Power Syst., (14)1: 120-125, 1999.
11. J. Gruhl, F. Schweppe and M. Ruane. “Unit Commitment Scheduling of Electric Power
Systems.” In System Engineering for Power: Status and Prospects, LH Fink and K.
Carlsen eds., Henniker, NH, 1975.
12. A. Geoffrion. Lagrangian relaxation for integer programming. Math. Prog. Study, (2):
82-114, 1974.
13. J.F. Bard. Short-term scheduling of thermal-electric generators using Lagrangian
relaxation. Oper. Res., (36)5: 756-766, 1988.
14. S. Hao, G.A. Angelidis, H. Singh, and A.D. Papalexopoulos. Consumer payment
minimization in power pool auctions. IEEE Trans. Power Syst., (13)3: 986-991, 1999.
15. R. Fourer, D.M. Gay and B.W. Kernighan. A modeling language for mathematical
programming. Manage. Sci., (36): 519-554, 1990
16. B.A. Murtagh and M.A. Saunders. “MINOS 5.1 User’s Guide.” Technical Report SOL
83-20R, Systems Optimization Laboratory, Department of Operations Research,
Stanford University, 1987.
APPENDICES
A. Development of a Computer Program for Validation

of the Proposed Approach
We have developed a computer program incorporating both supply-side
and demand-side bidding. The program allows multiple demand-side bidders
and supply-side bidders to interact for meeting a load profile over a specified
period. The generality of the program allows us, in particular, to study the
characteristics of the demand-side bidders and the influence of their parame-
ters on the results.
For the Lagrangian relaxation implementation we have followed the pro-
cedure proposed in [13]. To solve subproblems (10) and (11), the values of
are, at first, considered assigned as well as the values of the zero-one deci-
sion variables and It follows that these subproblems, for supply-side
bid i and demand-side bid j, respectively, become
subject to upper/lower limits specified in equation (6) (without the minimum

up/down constraints) and in equation (7).
These problems are piece-wise or quadratic problems (it depends on how
the bid variable price function is defined), and can be solved with a
linear or quadratic programming technique. These optimizations give the
values of power productions and load reductions To find the optimal
values for the decision variables and we solve subproblems (10) and
(11) using the values of and in a forward dynamic programming, taking
into account the start-up prices and the minimum up-/down-time constraints.
In the case of demand-side bidder problem (11) the dynamic program-
ming approach differs from that of supply-side bidders. In particular, to take
into account the load recoveries, we add the following term to the values
corresponding to the states in which demand-side bid j is committed at pe-
riod t
This term represents the contribution to the function of all the load re-
coveries due to the scheduled load reduction at period t that has to be
minimized in equation (11). Because of the presence of this additional term,
the solution of the problem (A.2) and the dynamic programming procedure
are included in a loop that ends when two consecutive solutions result in the
same commitment, i.e., in the same values of the zero-one decision variables
To find the maximum value L * of equation (12), we use a standard sub-

gradient technique [12]. With this technique, the solution process is iterated,
starting from a set of tentative values of multipliers and at each iteration
k the multipliers are updated using
where, taking into consideration equations (2) and (3),

and denote the current solution (iteration k) of the dual problem and
the lowest available solution of the primal one, which represent the lower
and upper bound of the final solution at iteration k, respectively. is a sca-
lar parameter that is progressively reduced whenever the solution of the dual
problem is not improved.
To implement such a technique, we need to find a feasible solution for
the primal problem P at each iteration. This is accomplished by adopting the
priority-list heuristic procedure proposed in [13], based on the calculation of
the marginal costs associated with each supply-side bidder and each period.
The heuristic procedure is implemented in such a way that it does not change
w, i.e,. the commitment of the demand-side bidders. The procedure here de-
scribed is stopped when the relative duality gap (defined as
is lower than a specified value.
We implement the program in the algebraic modeling language for
mathematical programming called AMPL [15]. This language is particularly
useful for the piecewise linear representation of the cost functions. A major
advantage of the modular structure resulting with the AMPL implementation
is the ability to use a library of solvers. In this paper the MINOS [16] solver
has been used to solve the problems (A.1) and (A.2).
B. Case Study System Data

Table B.1 gives the load profile.
Ten generators, considered as supply-side bidders, compose the supply

system. Tables B.2, B.3, and B.4 show the supply-side bidder data and have
been adapted from those used in [13]. Table B.1 shows the values of the
minimum and maximum power outputs the minimum up- and
down-times and the initial status of the units i.e., the number of
periods the unit had been up or down
In Table B.3 the quadratic expression of the cost data of [13] is converted
into piecewise linear form so that the bid variable price is specified by
6 parameters: the no-load price three incremental prices and
two elbow points respectively.
Table B.4 displays the values of the components of the bid start-up price
defined in equation (1).
Chapter 5
THERMAL UNIT COMMITMENT WITH

A NONLINEAR AC POWER FLOW
NETWORK MODEL
Carlos E. Murillo-Sánchez and Robert J. Thomas

Cornell University
Abstract: This chapter presents a formulation of the thermal unit commitment

problem that includes nonlinear power flow constraints, thus allowing a
more accurate representation of the network than is possible with DC
flow models. This also permits potential VAr production to be used as a
criterion for commitment of otherwise expensive generators in strategic
locations. We use a Lagrangian relaxation framework with duplicated
variables for each active and reactive source, permitting the exploitation
of the separable structure of the dual cost. Results for medium-sized
systems in a parallel processing environment are available.
1. INTRODUCTION
The central theme surrounding the use of Lagrangian relaxation for the unit
commitment problem is that of separation. Ever since the early papers [1, 2]
this separability was the objective and for a good reason: the unit commitment
problem, as a mixed-integer mathematical program, suffers from combinatoric
complexity as the number of generators increases. This is what dooms other
algorithms intended for solving it, such as dynamic programming: the combined
state space of several generators in a dynamic program is too large to be able
to attack many realistic problems, even with limited memory schemes.
Classical Lagrangian relaxation permits the decomposition of the problem
into several one-machine problems at each iteration; the coupling to other con-
straints involving more machines is achieved by sharing price information cor-
responding to the relaxed constraints, which is updated from one iteration to
another. The complexity of a given iteration becomes linear in the number of
discrete variables. The price to pay is that of switching to an iterative method
rather than a direct method, as well as giving up certainty of global optimality

if there is a duality gap. This last drawback is not as bad as it seems, however;
very small duality gaps are routinely obtained using Lagrangian relaxation.
The unit commitment problem can be formulated generally as:
where
Length of the planning horizon
Number of generators to schedule
Real power output for generator at time
Reactive power output for generator at time
On/off status (one or zero) for generator at time
The total production cost

The sum of any startup costs
A set of dynamic generator-wise constraints
A set of static instantaneous system-wide constraints
A set of nonseparable constraints.
The production cost function F is assumed to be convex (in fact, quadratic)
and separable over each generator and time period:
For our purposes, the constraints in the problem have been classified into three
kinds. Category groups constraints that pertain to a single generator, but
may span several time periods. These include minimum up or down times
and ramping constraints. Category groups constraints that span the com-
plete system but involve only one time period, such as load/demand matching,
voltage limits, reserve constraints, and generation upper/lower limits. Finally,
category groups constraints that involve more than one generator and more
than one time period. A typical example is the infeasibility of turning on more
than one unit at a time in a given location because of crew constraints.
To show the specific type of separation achieved by means of Lagrangian
relaxation, we take a look at the work of Muckstadt and Koenig [2]. They
considered a lumped one-node network with losses modeled as fixed penalty
factors. They also considered reserve constraints. We write an example formu-
lation including demand and reserve constraints, the relaxation of which yields
the Lagrangian
Thermal Unit Commitment with Nonlinear AC Power Flow 77
where is the real power demand in period is the desired minimum

total committed capacity for the same period, and is the upper operating
limit for the ith generator. Consider the dual objective
and the corresponding dual problem
which can be written explicitly in the following form after collecting terms on
a per generator basis:
Thus, for fixed and evaluation of amounts to solving separate,

single-generator dynamic programs of the form
The dynamic programs can readily accommodate constraints such

as minimal up or down times. Transition costs can easily include cold start or
warm start costs, derived from either banked or total shut-downs. Ramp-rate
constraints can also be introduced by discretizing the generation domain for
the unit, though the dynamic program grows considerably. For a detailed de-
scription of a dynamic programming graph including most of these constraints,
see reference [3].
After confirming that the evaluation of the dual functional is simple, a dual
maximization algorithm can be readily applied to since the sub-gradient
is easy to compute from Equation (3). A method as simple as a Poljak’s sub-
gradient ascent or as sophisticated as a bundle method [4] can be utilized.
A proper algebraic constraint structure is the key to the separability of
the dual functional. Note that the step from Equation (3) to Equation (6)
was possible due to the fact that the constraints are affine in the optimization
variables. We want to separate the variables on a per-generator basis, so, failing
affinity, we need at least separability into sets of variables related to a single
generator each. The unit commitment problem has such separability as long as
the system-wide constraints are separable by generator. Thus, general
nonlinear constraints cannot be readily included in the model, at least not
directly. We will show a way around this later.
2. THE CASE FOR AN AC POWER FLOW MODEL

In the past 20 years, many advances have been made, enhancing the num-
ber and type of constraints that can be treated, addressing convergence issues
when the production costs are not strongly convex, and so on. In particu-
lar, transmission line constraints have been included by using the so-called DC
power flow network models. These constraints, being linear, can be relaxed
with appropriate multipliers and the dual functional will still be separable by
generator. There are, however, genuine engineering considerations for includ-
ing a nonlinear AC flow model. These considerations stem from the fact that
some very important system constraints can only be modeled accurately with
a full-blown AC power flow model. For example, in an actual transmission line
the limit is best expressed as a limit in the current, whereas a transformer’s
capacity is best described by its MVA rating. Both actual MVA loading and
current depend on its orthogonal active and reactive components. The DC flow
model can only predict (and even then, only for relatively small angle deviations
and under a nominal voltage assumption) the active component, and therefore
it is easy to find situations in which it does a poor job of modeling important
constraints. Active branch limits inextricably tie together active and reactive
dispatch restrictions; if the reactive component is large, the dispatch obtained
by the DC flow-based unit commitment algorithm may have to be altered to
comply with actual line limits, or conservative limits may have to be used.
Other limits that cannot be modeled accurately via linear approximations
are voltage limits. Appropriate voltage levels are crucial to the operation of
most electrical apparatus, and it is not uncommon to find so-called “must
run” generators that are needed not so much because of their actual power
generation capacity, but because of their reactive power capacity. In order to
raise the voltage to levels that are adequate for consumers, sometimes large
amounts of reactive power are needed at specific locations in the network, even
if the only sources are generators that are costly to operate.
Granted, common practice to this day is to use linear network models for
constrained economic dispatch, and in many cases accuracy is good enough.
The reason why they work, however is that usually the only branches that are
modeled are tielines, which tend to have controllable reactive compensation
in order to keep power factors near unity. Linear congestion models work
best if unity power factor is maintained. They become inaccurate if used to
approximate thermal limits over broad ranges of voltage and reactive and active
injections, especially if the branch being represented does not have dispatchable
reactive compensation.
In summary, current Lagrangian relaxation algorithms using DC flow-based
network constraints or other linear approximations may do poorly enforcing
voltage limits and true current or MVA limits in grids requiring large reactive
injections. As a result of the limited accuracy of this network representation,
it is quite possible that a solution obtained by such an algorithm is actually
infeasible, requiring additional generators to be committed in order to meet
voltage or congestion constraints.
We find some of the works that have followed the DC flow formulation in their
incorporation of line limits to the dual maximization in [5, 6, 7, 8, 9, 10, 11].
In [12], the authors use linear approximations to include the voltage constraints
in the formulation, but it is not clear that a linear approximation will work
under large excursions in the reactive dispatch. Baldick [8] uses a general for-
mulation that could in principle be used to address AC flow constraints, and he
elaborates a little more in the speculative paper [13], although the authors do
not seem to have actually implemented their scheme. The formulation proposed
in this work was first reported in [14]. It uses fewer multipliers and, by explic-
itly employing an augmented Lagrangian, should have improved convergence
properties.
3. FORMULATION AND ALGORITHM

Our approach has its origins in the variable duplication technique credited to
Guy Cohen in [6] by Batut and Renaud. Baldick [8] later used this technique in
his more general formulation of the unit commitment problem. The principal
achievement of this work is the inclusion of reactive power output variables to
the formulation, so that better loss management may be performed and gener-
ators that are necessary because of their VAr output but not their real power
are actually committed. This is the logical next step in the development of La-
grangian relaxation techniques. At this point, when typical algorithms reduce
the duality gap to figures close to 1% [15, 10], it is important to recognize that
a better handling of the reactive power considerations at the unit commitment
stage may have a payoff that is higher than those last few percentage points in
the duality gap.
The variable duplication technique exploits the fact that the problem
is equivalent to
Notice that the last constraint is linear and therefore amenable to relaxation.
We start by defining two sets of variables. The dynamic constraints will
be posed in terms of the dynamic variables, whereas the static or system-wide
constraints will be posed in terms of the static ones:
Dynamic variables:
Commitment status {0,1} for generator at time
VAr output for generator at time
Static variables:
VAr output for generator at time
We define the optimization problem as:
subject to:
(1.) constraints
(2.) constraints
(3.) and the following additional constraints

where is the minimum combined capacity that is acceptable for the

zone in the period and is the set of indices of generators in the zone.
A word about the representation of the minimum up-time or down-time
constraints is in order. While it is perfectly possible to state these restrictions
in terms of additional state variables obeying state transition rules that reflect
these constraints (see [10]), we have chosen not to include any more variables
because of the already overburdened notation. These constraints can readily
be incorporated in the structure of the dynamic programming graph used to
solve the resulting subproblems, as in [3]. Thus, it will be assumed that we can
enforce the constraints (9–11) on the D variables and the constraints (12–
14) on the S variables, so that we only relax the three last constraints (15–17),
which leads to the following Lagrangian:
where are multipliers on the relaxed equalities of the two kinds of

variables, is the multiplier associated to the zone’s reserve requirement
at the period, and returns the index of the zone to which generator
belongs.
The separation structure of the Lagrangian is obvious upon looking at equa-

tions (19) and (20). It makes it possible to write the dual objective as
By looking again at (19) and (21), we see that the first term can be computed by
solving dynamic programs again; the second term separates into optimal
power flow (OPF) problems with all generators committed but with special cost
curves for generator at time Notice that also has a price.
We assume that the solutions of the dynamic programs meet the constraints
and that the solutions of the optimal power flows meet the constraints.
It would be tempting to apply a dual maximization procedure to the dual
objective as stated, but there are some issues that prevent us from doing that
without some modification of the Lagrangian. The first issue is that the cost
of reflected in the dynamic programs, being linear, is not strongly convex;
this can cause unwanted oscillations in the prescribed by the dynamic pro-
gram [6]. Therefore, we set out to fix this before addressing any other problems
by augmenting the Lagrangian with quadratic functions of the equality con-
straints. This will introduce nonseparable terms. At this stage, we will invoke
the Auxiliary Problem Principle described by Cohen in [16] and [17]. This
principle is a rather general formalism characterizing optimality in an implicit
manner for convex programming, and an algorithm of the fixed point type can
be inferred from it. Convergence for the convex case with affine constraints has
been proved by Cohen. We proceed to write the new augmented Lagrangian as
Within the context of the auxiliary problem principle, it is possible to substitute

the augmentation terms by the following at iteration (see [17] for the general
technique and [6, 7] for the first documented application to unit commitment):
where and are the values obtained at the itera-

tion. Since (22) is separable, we can collect terms of the augmented Lagrangian
on a per-generator basis, so that at the iteration we are faced with
Notice that (24) has the same separation structure of (19).

Now that the separability issue has been resolved, we propose the following
Algorithm: AC Augmented Lagrangian relaxation

Step 0:
Step 1: Initialize to the values of the multipliers on the power flow
equality constraints at generator buses when running an OPF with all
units committed. Initialize to zeros. Initialize a database
of tested commitments to be empty; it will later be filled with the
commitment schedules obtained in the course of the algorithm and
their corresponding feasibility status.
Step 2a: Compute
by solving one-generator dynamic programs.

Step 2b: Compute
by solving OPFs in which all generators are committed, their gen-

eration range has been expanded to include and the special
cost is used. Note: all tasks in steps 2a and 2b can
be solved in parallel.
Step 3: If the commitment schedule Û is not yet in the database of tested
commitments, perform a cheap primal feasibility test. If the results
are not encouraging (i.e., not enough committed capacity), store the
schedule in the database and label it as infeasible, then go to Step 6.
Step 4: Perform a more serious primal feasibility test by actually attempting
to run OPFs with the original constraints. If all OPF’s are
successful in finding a feasible dispatch, store the commitment in the
database, together with the primal cost including startup costs. Else
label the commitment as infeasible, store it in the database, and go to
Step 6.
Step 5: If the mismatch between the two sets of variables is small enough, and
there is already a set of feasible commitment schedules, stop.
Step 6: Update all multipliers using sub-gradient techniques, and
Step 7: Go to Step 2.
The proposed algorithm is very OPF-intensive: the major computational

cost is that of computing OPFs for every iteration in order to solve the static
subproblems, plus extra OPFs in selected iterations when a given commitment
is promising in terms of primal feasibility. Thus, every effort possible must be
made to try to alleviate the burden of OPF computation. The first thing that
can be done is to use as a starting point for the OPF the result of the previous
iteration for the same time period. Most of the times, the only difference in
the data for the OPF would be a small change in the costs (reflected by the
change in from one iteration to another). This, in theory, should result in
fewer iterations needed for the OPF.
Another drawback of the algorithm is that a different set of OPF compu-
tations must be performed to compute the value of the dual objective and to
compute the value of the primal. Thus, before even trying to compute the value
of the primal objective, one should make sure that such a costly computation
is worth doing. Some of the cheap tests include verifying that the reserve con-
straint is met and that the mismatch between the S and the D variables is
small, since feasibility is a given if the mismatch is zero. With respect to the
latter, we have found that if a smaller mismatch should be specified
as requisite to feasibility than if More costly feasibility tests would
involve power flow problems starting from appropriate initial values. Currently
a constrained power flow is being performed at this stage.
4. COMPUTATIONAL RESULTS
An implementation of the algorithm has been written in the MATLAB™
environment. The dynamic subproblems can accommodate minimal up or down
times, warm start and cold start-up costs and are solved using forward dynamic
programming. We solve the static subproblems using a version of MINOS [18]
that has been incorporated in the MATPOWER package [19] by the first au-
thor. It incorporates box constraints on the generator’s active and reactive
output, piecewise-linear or polynomial cost functions for both P and voltage
constraints, line MVA limits, and, of course, the power flow equations. Addi-
tionally, any linear constraint on the optimization variables can be imposed. A
preconditioner for MINOS that performs a constrained power flow is used if
necessary. It implements a Levenberg-Marquardt-like minimization of the sum
of squares of the power flow constraints, with penalty functions on some other
box contraints and constrained variables in the case of voltage limits. Thus,
each iteration involves the solution of a problem rather than a Newton step.
We solve the problem using a MEX-file version of BPMPD [20], an interior
method solver.
The specific sub-gradient technique being used at this point is a simple Poljac
step size schedule, with an iteration-dependent weight that is inversely propor-
tional to iteration number. We chose simplicity initially because convergence
conditions for Poljac’s method are “mild” and well documented: it is advanta-
geous to be able to blame any convergence problems on the formulation itself

and not on the particular kind of sub-gradient update. The obvious drawback
to this decision is that the convergence of the dual iteration process may not
be very fast. More sophisticated gradient updates will be investigated once the
basic features of the algorithm are well understood.
The program was originally tested on a modified IEEE 30-bus system [21]
with six generators and a planning horizon of length six. We modified the test
case so that generator #4, located at bus #27, is needed for voltage support for
many load levels even though it is most uneconomical to operate. For compar-
ison purposes, we also wrote a version of the Lagrangian relaxation algorithm
with DC flow-based relaxed line limits. The AC-based algorithm correctly iden-
tified this unit as a must-run for those time periods, even providing some price
information on the MVArs that this unit produced by means of the correspond-
ing The number of iterations required was usually in the vicinity of one
hundred. In contrast, the DC flow-based algorithm failed to commit unit #4
for any period, producing a commitment schedule that was infeasible in light
of the AC power flow constraints.
The importance of proper selection of the parameters was ap-
parent from the beginning. After several trial runs, we obtained good results
with and Values very different from
these, however, tended to produce somewhat smooth, damped oscillations in
the values of some of the
To highlight one of the new features found in the algorithm, we show the
evolution of versus iteration number for a typical run in Figure 1.
The multipliers with the higher values are all P-type multipliers. Those with
the smaller values correspond to the Most of them settle to zero, indicating
that is essentially free almost always. Yet, a few of them actually have high
prices: these belong to generators and time periods where the OPF tries to use
their MVArs in order to force feasibility or guided by economic considerations,
but the generators are not actually committed. In the course of the algorithm,
these may grow so large that they trigger the respective unit on. Once this
happens, such multipliers tend to approach zero again, since is now plentiful.
In Figure 1 there are two clear examples of this behavior, corresponding to
unit 4 being committed for certain time periods. As the multiplier approaches
zero, the static copy will approach the dynamic
A slightly more ambitious test has been performed using the IEEE 118-bus
system, with 54 generators and a time horizon of 24 periods, corresponding
to two “weekdays” and one day of the weekend, each with 8 three–hour pe-
riods. The total variation of the load relative to the base case is –50% and
+40%. Figure 2 shows the behavior of the norm of the active and reactive
mismatches between the two sets of variables. Figure 3 depicts the evolution
of the multipliers in this case.
5. PARALLEL IMPLEMENTATION
One of the purported advantages of separation by Lagrangian relaxation
is that it allows for simultaneous solution of all of the static and dynamic
subproblems in a given dual iteration, making the computation amenable to
parallelization. Profiling of the algorithm has indicated that more than 95%
of the computation time in this particular implementation is allocated to the
solution of the optimal power flows. In order to test larger systems, a parallel
implementation was deemed not just convenient, but in fact necessary since
running times were already several hours long. Fortunately, MultiMATLAB, a
parallel processing toolbox for MATLAB is being developed at Cornell Univer-
sity by John Zollweg (see [22] and the original work of Trefethen et. al. [23]);
this allowed us to take advantage of much of the existing code.
Although the dynamic programs are also parallelizable, their time-granularity
is much smaller than that of the OPFs, so communications overhead is likely
to reduce the efficiency of their parallelization. We have opted to parallelize
only the OPF computation at this time.
MultiMATLAB works by having several copies of MATLAB, each running
in a different node, communicate by means of a subset of the Message Passing
Interface (MPI) library [24]. The calls to the MPI functions are implemented
by means of MEX (MATLAB-Executable) files. A master node performs the
main algorithm and the dynamic programming subproblems and when a set
of OPFs need to be carried out it sends the OPF data to the other nodes,
working in a master/slaves configuration. The workers’ only task is to receive

OPF input data, run the OPF solver and report the results back to the master.
Two parallel scheduling strategies have been programmed so far. The first
one is a straightforward round-robin scheme in which the master cycles through
the workers, receiving the results of any previously assigned OPF by means of
a blocking receive (MPI’s Recv function), which ties the master until data does
arrive. Then the master sends the worker data for the next OPF and turns its
attention to the next worker. Clearly, having the master wait for the worker is
not optimal, but it has the advantage of employing only MPI’s Send and Recv
calls, whose implementation is stable in MultiMATLAB. A second, more so-
phisticated parallel scheduling strategy makes use of MPI’s Irecv non-blocking
receive function. Immediately after sending data to a worker, the master node
posts a non-blocking receive, setting aside an incoming message reception area.
The master can then turn its attention to other workers and assign further work.
Every now and then the master calls MPI’s Testany function, which informs
the master if any pending Irecv’s have been completed. If so, the master reads
the data from the corresponding buffer and assigns more work if needed. This
strategy promises the most efficient use of the workers. In limited experiments
with up to seven workers, the master is actually free most of the time, testing
for incoming data. This means that it is efficiently keeping the workers busy.
The efficacy of this scheduling strategy can be further improved by performing
the OPFs in order of decreasing expected execution time. Unfortunately, as of
now the Testany implementation is not sufficiently stable to allow running a
complete problem.
The round-robin strategy has been tested using the 118-bus system with
a much longer time horizon (168 periods) and longer start-up and shut-down
times on the generators. Figure 4 shows a plot of the evolution of the mis-
matches.
6. FUTURE WORK
Historically, the best justification for using Lagrangian relaxation has been
the regular achievement of small relative duality gaps in larger problems. So
far, our algorithm seems to behave correctly with proper tuning of parameters,
but there is a need to test much larger (i.e., more than 300 generators) systems
to verify whether small duality gaps are also routine. At that point, comparison
to the results obtained using a DC-flow or linear network model formulation
should also be performed, with special attention to the cost of any commitment
corrections needed to make the schedule generated by the linear network model
algorithm to be AC-feasible. It would be worth it to incorporate other kinds of
constraints and continue improving the robustness of the optimal power flow
subsystem, which has been the weakest link so far.
Three major efforts are being undertaken. First, John Zollweg is implement-
ing several changes that should make the non-blocking receive implementation
in MultiMATLAB more stable, allowing the use of improved parallel scheduling

strategies. Secondly, work is being done to improve the robustness of the OPF
solver. While MINOS has been found to do a good job of finding optima given
a good starting point, the early stages of the Lagrangian relaxation algorithm,
when prices are being adjusted with larger steps, result in OPF problems where
the solution lies far away from the starting point. We have found MINOS not
to behave as well in these cases, especially for larger (i.e., 3000 buses) systems,
even with the constrained power flow preconditioner. So the preconditioner is
being turned into a first-stage optimizer intended to locate binding constraints
and provide a better, feasible starting point. Finally, work is being done on
how to include ramping constraints into the formulation. There are two basic
variations in the literature: the first one involves discretizing the generation
range for the dynamic copy of the active sources and disallowing transitions
that violate ramping constraints in the dynamic program. While relatively
straightforward, this approach would result in much larger dynamic programs.
The second method involves relaxing the linear ramping constraints and adding
them to the Lagrangian. There is, however, some concern about the speed of
convergence of the corresponding multipliers, and special updating strategies
may be part of the solution.
Finally, there remains the question about whether such an algorithm is useful
anymore. Lagrangian relaxation methods have not been used to clear energy
markets that also produce commitment schedules due to fairness problems:
since the cheapest commitment schedule obtained may depend on parameter
tuning, a Lagrangian relaxation solution may lack the clarity and uncontesta-
bility required in a method used to clear a market. In addition, the issue of
computation time also arises, since the algorithm requires the solution of many
complex AC OPFs; a method used to clear a market should be expeditious due

to practical considerations. This second consideration is actually an issue that
will become moot in time, given the exponential growth of typical computing
capacity and the algorithm’s amenability to parallel computation. The first
issue is more serious, and inherent in Lagrangian relaxation methods for unit
commitment. So at this point it is difficult to advocate the use of the algorithm
for market clearing. However, it is still a tool that can be used by ISOs for
location-based market power studies as well as transmission capability assess-
ment and expansion. Lastly, from an experimental economics viewpoint, the
best way to assess the performance of a market is to compare its efficiency to
that of the maximum social welfare solution. This algorithm can provide this
solution (the modifications for elastic demand are trivial) and help to make
market efficiency comparisons.
ACKNOWLEDGEMENTS
We wish to thank Ray Zimmerman and Deqiang Gan for providing us with
their package MATPOWER [19] for the initial tests of the algorithm. We
would also like to thank Csaba Mészáros, whose QP program [20] BPMPD we
use in several of our programs, and finally, John Zollweg, principal developer
of MultiMATLAB.
REFERENCES
1. J.A. Muckstadt and R.C. Wilson. An application of mixed-integer programming
duality to scheduling thermal generating systems. IEEE Trans. Power Apparatus
Syst., 87(12):1968–1978, 1968.
2. J. A. Muckstadt and S.A. Koenig. An application of Lagrange relaxation to
scheduling in power-generation systems. Oper. Res., 25(3):387–403, 1977.
3. G.S. Lauer, D.P. Bertsekas, N.R. Sandell, and T.A. Posbergh. Solution of large–
scale optimal unit commitment problems. IEEE Trans. Power Apparatus Syst.,
101(1):79–85, 1982.
4. C. Lemarechal and J. Zowe. A condensed introduction to bundle methods in non-
smooth optimization, in Algorithms for Continuous Optimization, E. Spedicato,
Ed., Kluwer Academic Pub., 1994.
5. and A new approach for solving extended unit commit-
ment problem. IEEE Trans. Power Syst., 6(l):269–277, 1991.
6. J. Batut and A. Renaud. Daily generation scheduling optimization with trans-
mission constraints: a new class of algorithms. IEEE Trans. Power Syst.,
7(3):982–989, 1992.
7. A. Renaud. Daily generation management at Electricité de France: from planning
towards real time. IEEE Trans. Autom. Cont., 38(7):1080–1093, 1993.
8. R. Baldick. The generalized unit commitment problem. IEEE Trans. Power Syst.,
10(l):465–475, 1995.
9. J.J. Shaw. A direct method for security-constrained unit commitment. IEEE
Trans. Power Syst., 10(3): 1329–1339, 1995.
10. S.J. Wang, S.M. Shahidehpour, D.S. Kirschen, S. Mokhtari, and G.D. Irisarri.
Short-term generation scheduling with transmission and environmental con-
straints using an augmented Lagrangian relaxation. IEEE Trans. Power Syst.,
10(3):1294–1301, 1995.
11. K.H. Abdul-Rahman, S.M. Shahidehpour, M. Aganagic, and S. Mokhtari. A
practical resource scheduling with OPF constraints. IEEE Trans. Power Syst.,
11(1):254–259, 1996.
12. H. Ma and S.M. Shahidehpour. Unit commitment with transmission security
and voltage constraints. IEEE Trans. Power Syst., 14(2):757-764, 1999.
13. X. Guan, R. Baldick, and W.H. Liu. Integrating power system scheduling and
optimal power flow. In Proc. 12th Power Systems Computation Conference, Dres-
den, Germany, August 19–23, 1996, pp. 717–723.
14. C.E. Murillo-Sánchez and R.J. Thomas. Thermal unit commitment including
optimal AC power flow constraints. In Proc. 31st HICSS Conference, Kona,
Hawaii, Jan. 6–9 1998.
15. D.P. Bertsekas, G.S. Lauer, N.R. Sandell, and T.A. Posbergh. Optimal short-
term scheduling of large-scale power systems. IEEE Trans. Autom. Cont.,
28(1):1–11, 1983.
16. G. Cohen. Auxiliary problem principle and decomposition of optimization prob-
lems. Optim. Theory Appl., 32(3):277-305, 1980.
17. G. Cohen and D.L. Zhu. Decomposition coordination methods in large scale
optimization problems: the nondifferentiable case and the use of augmented La-
grangians. In Advances in Large Scale Systems, Vol. 1; J.B. Cruz, Ed., JAI Press
Inc, 1984, pp. 203–266.
18. B.A. Murtagh and M.A. Saunders. MINOS 5.5 user’s guide. Stanford University
Systems Optimization Laboratory Technical Report SOL 83-20R.
19. R. Zimmerman and D. Gan. MATPOWER: A Matlab power system simulation
package. http://www.pserc. Cornell.edu/matpower/.
20. Cs. Mészáros. The Efficient Implementation of Interior point Methods for Linear
Programming and Their Applications. Ph.D. Thesis, Eötvös Loránd University
of Sciences, 1996.
21. O. Alsac and B. Stott. Optimal load flow with steady-state security. IEEE Trans.
Power Apparatus Syst., 93(3):745–751, 1974.
22. J. Zollweg. MultiMATLAB for NT Clusters. Cornell Theory Center Software
Documentation, http://www. tc.Cornell. edu/UserDoc/Intel/Software/
multimatlab/.
23. A.E. Trefethen, V.S. Menon, C.C. Chang, G.J. Czajkowski, C. Myers, and L.N.
Trefethen. MultiMatlab: MATLAB on Multiple Processors. Cornell University
Computer Science Technical Report 96TR239.
24. W. Gropp, E. Lusk, and A. Skjellum. Using MPI: Portable parallel programming
with the Message Passing Interface. MIT Press, 1994.
25. A.I. Cohen and V.R. Sherkat. Optimization-based methods for operations
scheduling. Proc. IEEE, 75(12):1574–1591, 1987.
26. G.B. Sheblé and G.N. Fahd. Unit commitment literature synopsis. IEEE Trans.
Power Syst., 9(1):128–135, 1994.
Chapter 6
OPTIMAL SELF-COMMITMENT UNDER
UNCERTAIN ENERGY AND RESERVE PRICES
R. Rajaraman and L. Kirsch

Laurits R. Christensen Associates
F.L. Alvarado
University of Wisconsin
C. Clark
Abstract: This paper describes and solves the problem of finding the optimal self-commit-
ment policy in the presence of exogenous price uncertainty, inter-product substi-
tution options (energy versus reserves sales), and different markets (real-time
versus day-ahead), while taking into consideration intertemporal effects. The
generator models consider minimum and maximum output levels for energy and
different kinds of reserves, ramping rate limits, minimum up- and down-times,
incremental energy costs and start-up and shut-down costs. Finding the optimal
market-responsive generator commitment and dispatch policy in response to ex-
ogenous uncertain prices for energy and reserves is analogous to exercising a
sequence of financial options. The method can be used to develop bids for en-
ergy and reserve services in competitive power markets. The method can also
be used to determine the optimal policy of physically allocating generating and
reserve output among different markets (e.g., hour-ahead versus day-ahead).
1. INTRODUCTION
Unit commitment refers to the problem of deciding when to start and
when to shut-down generators in anticipation of changing demand [1]. In tra-
ditional utility systems, the problem of unit commitment was formulated and
solved as a multi-period optimization problem. In the traditional problem
formulation, the anticipated demand was an input variable. The problem was
solved for multiple generators, generally owned by the same entity (a utility).
The start-up, shut-down and operating costs of the generators were assumed
known. The standard way to analyze and solve this problem was by dynamic
programming, and within this category of problems, the most popular solution
method in recent years has been the use of Lagrangian relaxation [2,3]. Re-
cently, a new method of decommitment has also been proposed [4,5].
Several things change in a deregulated market. Generators generally have
to self-commit their units optimally. Since in most power pools, no single
merchant owns all the generating assets, the need to meet system load is re-
placed with the need to optimize profits of the merchant’s generating plants
based on the uncertain market prices at the locations where the generators are
located. Of course, the forecasts of markets prices depend upon a number of
factors, the most important of which include demand, system-wide generation
availability and cost characteristics, and transmission constraints.
We pose the problem of finding the profit-maximizing commitment pol-
icy of a generating plant that has elected to self-commit in response to exoge-
nous but uncertain energy and reserve price forecasts. Typically, one genera-
tor’s output does not physically constrain the output of a different generator1,
so this policy can be applied to each generator in the merchant’s portfolio
separately and independently. Therefore, for ease of exposition, we assume
the case of a single generator. Generator characteristics such as start-up and
shut-down costs, minimum and maximum up- and down-times, ramping rates,
etc., of this generator are assumed known. The variation of prices for energy
and reserves in future time frames is known only statistically. In particular,
the prices follow a stochastic rather than deterministic process. We model the
process using a Markov chain. The method is applicable to multiple markets
(e.g., day-ahead, hour-ahead) and multiple products (energy, reserves).
Other researchers have modeled the effect of energy price uncertainty on
generator valuation. In [6], the author models the effect of the spark spread on
short-term generator valuation. In [7], the authors propose mean reverting
price processes and use financial options theory [9] to value a generating
plant. In both [6] and [7], however, the authors neglect the effect of realistic
operating constraints such as minimum start-up and shut-down times. In [8],
the authors improve upon this work to more realistically include the effect of
operating constraints to find the short-term value of a generating asset; yet
they neglect the effect of ramping rates. All these papers make assumptions
about the risk-neutral price process in order to value the power plant.
This chapter focuses not on generator valuation, but on finding the gen-
eral principles for generator self-commitment in the presence of exogenous
price uncertainty and market multiplicity. We derive the basic mathematical
principles from dynamic programming theory [10]. We consider energy and
reserve markets, although the method can be extended to include additional
market choices, such as day-ahead versus real-time markets. Section 2 of this
1 Exceptions include multiple hydroelectric units connected in series and restrictions on aggre-
gate emission levels from multiple generators within an area.
Self-Commitment Under Uncertain Prices 95
chapter defines the problem. Section 3 gives the dynamic programming solu-
tion to the problem, and Sections 4 and 5 illustrate the features of the optimal
commitment policy using simple illustrative examples. Section 6 gives a more
detailed numerical result for a peaking generator, and Section 7 concludes
the paper. Appendix A is a technical section that solves the single-period op-
timal generator dispatch problem given exogenous prices of energy and re-
serves.
2. THE PROBLEM
We begin by describing the exogenous inputs to the problem.
Generator capability and cost characteristics. At any given time t,
generator G is assumed to be in state where is a member of a discrete set
X={state 1, state 2, ..., state K} of possible states. Intertemporal constraints
are represented by state transition rules that specify the possible states that the
generator can move to in time period t+1, given that the generator is in a state
at time t. Generally, there is a cost associated in moving between different
states. In a simple representation, two states are sufficient: “in service” (or
“up”) and “out of service” (or “down”). In general, however, many more
states may be needed to represent the various conditions of the start-up and
shut-down process.
The degree to which a generator can participate in providing reserves de-
pends on its ability to respond to the reserve needs in a timely manner. For
regulating and spinning reserves, the generator must be already be in service;
the amount of MW of reserves that a generator can offer must be consistent
with its ramping rate. Generators that are already at a maximum in terms of
energy provision are unable to also participate in the reserves market. Thus, to
participate in the reserves markets, the generator cannot simultaneously sell
all of its capability in the energy market.
The parameters that describe a generator include:
Minimum and maximum output levels
Ramping rates
Minimum up- and down-times for the generator
Incremental energy costs and no-load costs
Start-up, shut-down, and banking costs.
Generator states and state transitions. Generators can be in any of a
number of several possible UP, DOWN, or transitional states. For example,
for a generator with total capacity of 200 MW, and ramp rate of 100
MW/hour, we could define two UP states: and The state would
cover the operating range [0 MW, 100 MW], while the state would cover
the operating range [101 MW, 200 MW]. Likewise, minimum down times
can be enforced by defining multiple DOWN states. Only certain transitions

among these states are permissible. Furthermore, transition between states
generally involves a cost. For example, going from a cold shutdown to an on-
line state will involve a start-up cost.
Generator dispatch constraints. A generator may have additional dis-
patch constraints that restrict its operation. For example, the generator may
need to be offline (in the “off” state) during certain periods for scheduled re-
pairs. These restrictions are modeled as time-dependent constraints on genera-
tor states.
Exogenous price forecast of energy. A discrete Markov process models
the exogenous price for energy. In each time period, a discrete price state
represents a price range. The price at time is probabilistically related to
the price at time t via a price transition matrix. That is,
is a known quantity. One can think of a price forecast at any time t
to be a baseline price point plus a random uncertainty around the baseline
forecast.
Since the exogenous price forecast is an important input of the problem,
we digress a little to discuss how one may obtain estimates of this input. We
consider two ways in which price forecasts can be made:
One possible method is to use historical data. For example, to obtain a
price forecast for next week, one could use past week data and data
from other weeks with similar load/weather patterns as that predicted
for next week. This would be a statistical data-mining problem.
Another way to forecast prices would be to use numerous Monte Carlo
iterations of structural computer models (such as optimal power flow
models and production cost models) to model the uncertainty of
prices.
The physical spot markets for energy include real-time, hour-ahead, day-
ahead, and possibly week-ahead markets2. Each of these spot markets is a
different market, e.g., energy prices for a particular hour could be different in
day-ahead and real-time markets and could have very different characteristics
in terms of price volatility. A generator will often have a choice as to which
market to use to sell its energy.
Exogenous price forecast of reserves. Operating reserves are distin-
guished by the speed with which they become available and the length of time
that they remain available. In the nomenclature of the Federal Energy Regula-
tory Commission, the primary types of reserves (from fastest to slowest) are
regulating, spinning, supplemental, and backup reserves. For example, regu-
lating reserves need to be available for following moment to moment fluctua-
2
It is doubtful whether forward prices quoted month-ahead (or more) will influence an indi-
vidual generator’s commitment and dispatch decisions.
tions in system demand and can generally be offered by generators on auto-

matic generation control (AGC). As another example, to offer 10 MW of
spinning reserves, a generator must be online and must be capable of produc-
ing 10 MW within 10 minutes. Supplemental reserves and backup reserves
are slower reserve types. For generators who provide reserve services, there
are two types of reserve costs (see [11] for more details). These are reserve
availability costs, which are the costs of making reserves available even if
they are not actually used, and reserve use costs, which are the costs incurred
when the reserves are actually used. Generally, reserve use costs are compen-
sated at the spot price of energy3. Reserve availability costs are the opportu-
nity costs of generators, i.e., they include off-economic dispatch costs, and
costs of starting up or shutting down generators. In California, New York,
New England, and the Pennsylvania-Jersey-Maryland (PJM) system, there is
currently a competitive market-clearing process for setting the reserve avail-
ability costs of some or all of the above reserve types. Reserve availability
prices are modeled similarly to energy prices, i.e., as a discrete Markov chain.
We allow, however, the reserve availability prices to be correlated with the
energy prices. One particularly simple way to model reserve uncertainty is to
assume perfect correlation between energy prices and reserve availability
prices.
Exogenous fuel price forecasts. Fuel prices can be modeled similarly to
reserve availability prices. Since most practical commitment periods are less
than a week and fuel prices typically show much less volatility than energy
prices over this interval, it is a not a bad approximation to keep these inputs
constant. For ease of exposition, therefore, we assume fixed fuel prices
throughout the paper. It is fairly straightforward to also include uncertainty in
fuel prices; see, for example, [8].
Start and end time periods. We will assume that the start time period is
at time 1 and that the end time period is at time T. For most practical prob-
lems, T will be between one day and one week.
Next, we define some notation; explicit functional dependencies are often
omitted for clarity. Given generator state and a vector of energy and re-
serve availability prices, at time t,
1. (or simply defines the commitment policy for time t, i.e., it
signifies a particular valid rule to move the generator from state at
time t to a new state at time t+1.
3
In general, reserve use costs could also include the wear-and-tear costs of ramping up and
down to follow system demand. The current custom is that, in competitive markets for re-
serve services, these costs are not directly compensated and must somehow be internalized
by the generators.
2. (or simply defines the dispatch policy for time t, i.e., it

represents a particular dispatch of energy and reserves for the genera-
tor at time t, given that the generator is in state
3. denotes single-period profits at time i.e., the profits realized
by the dispatch
4. denotes the cost of transitioning (e.g., start-up or shut-down
costs) from state at time t to state at time t+1 due to the
commitment policy used in time period t.
The problem can now be posed as:
[PROBLEM] Find the best commitment and dispatch policy (u*, d*)
that maximizes expected total profits over all possible
commitment policies and all possible dispatch policies
where E denotes the expected value over the uncertain
price forecasts5.
Before we proceed further, it is useful to summarize the essential features
of the problem:
Inter-product substitutability. Market participants have a choice be-
tween sales in the energy versus sales in the reserve markets. These
markets operate simultaneously (though the markets for energy and re-
serves may clear sequentially, as they currently do in California).
Moreover, market participants have a choice of offering their products
in different markets (e.g., day-ahead versus hour-ahead markets).
Price uncertainty. The future prices of energy and of reserves at the
location of the generator of interest are unknown but follow a known
random process. The general characteristics of the random process are
estimated by the generator wishing to self-commit.
Intertemporal ffects. Intertemporal constraints affect the generator’s
operations. This may lead to situations when a market participant can
elect to remain on during certain periods when operation will be at a
loss in return for likely (but not certain) profits in later periods.
We are interested in finding both the optimal commitment and the optimal
dispatch policy. We stress that the problem is complicated by the fact that at
the time the commitment decision is made, future prices are uncertain. The
next section addresses this problem.
4
Current period profits do not include transition costs.
5
This objective function assumes that the generator is risk-neutral. If the generator is risk-
averse, the objective function should reduce the expected outcome according to some meas-
ure of risk. For example, the objective might be to maximize where “V” is
the variance of net profit and “a” is a risk aversion coefficient. As another example, the ob-
jective function could be an exponential utility function with constant relative risk aversion
[12]. Now the objective function would be multiplicative in nature, but we can take natural
logarithms to convert the objective function to the form shown in this chapter.
3. OPTIMAL COMMITMENT AND DISPATCH

POLICY
We now present the optimal commitment and dispatch policy. The opti-
mal dispatch policy is fairly straightforward: given an exogenous price fore-
cast for time period t, the generator takes its current state as given and dis-
patches energy and reserves in an optimal manner for time period t, without
regard to other time periods. Appendix A describes the single-period optimal
dispatch. The profit-maximizing commitment decision for transitioning to the
next time period is more complicated, however, because actions taken now
affect future time periods.
The backward dynamic programming (DP) method for solving this prob-
lem starts at the final time period T and works backward using the following
steps. The backward DP method [3,10] is as follows6:
Step 1. Let over all possible
commitment policies and dispatch policies Let the optimal dispatch
policy be denoted by and the optimal commitment policy by
must be computed for each possible state and each possible price level
Step 2. Let over
all possible commitment policies and dispatch policies The expecta-
tion E is taken over all possible price levels given that the price in time
T–l is is probabilistically related to via the Markov chain. The
state at time T, is related to by the commitment policy Let the op-
timal dispatch policy be and the optimal commitment be
must be computed for each possible state and
each possible price level
Step T. Let over all possi-
ble commitment policies dispatch policies and price levels The ex-
pectation E is taken over all possible price levels given that the price in
time 1 is The state at time 2, is related to the previous state by the
commitment policy i.e., Let the optimal dispatch policy be
and the optimal commitment policy be must be
computed for each possible state and each price level
If the generator is in state and sees price level at time t=l, the
profit-maximizing schedules are represented by the commitment actions
and the dispatch decisions
where for
t=1,2,...,T-l, and is related probabilistically to via the Markov chain,
and the maximum expected profits are Actual profits and actual
6
For ease of exposition, we have ignored the boundary condition at time T+l.
schedules depend upon actual price levels encountered in the different time
periods.
The algorithm for finding the optimal commitment policy is similar to the
problem of determining the value of an option using a tree approach [9].
Therefore, we can characterize the problem of finding an optimal commitment
policy as a generalized tree approach that values and exercises a sequence of
complicated options in each time period. The options involve decisions such
as whether to commit or not, whether to ramp up or ramp down, whether to
participate in the energy or reserves markets, etc.
4. ILLUSTRATIVE EXAMPLE 1
This section illustrates the concept. For simplicity we assume only one
product, energy, and one three-period market, i.e., T=3.
Suppose that the generator parameters are as shown in Table 1. For this
example, the generator at time t can be in one of two states, “UP” or
“DOWN," i.e., X={UP,DOWN}. There are no intertemporal constraints. The
generator can move from any state at time t to any state at time t+1.
A shut-down cost is incurred when the generator moves from an UP state

to the DOWN state. A start-up cost is incurred when the generator moves
from the DOWN state to the UP state. All other transitions result in zero
costs.
In each time period, an exogenous price forecast may be described by
two possibilities: is HIGH or is LOW, each having a specified probability
of occurrence. The HIGH and LOW prices in each time period are allowed to
vary, as Table 2 shows. Further assume that the prices at time t+1 depend
probabilistically upon the prices at time t. The probability of a HIGH-HIGH
transition is and the probability of a LOW-LOW transition is These are
exogenous, with assumed values and
The backward DP algorithm finds the profit-maximizing commitment and

dispatch policy. Tables 3(a)-(c) depict the solution. Columns in these tables
correspond to time periods. The entry in Table 3(a) that corresponds to a
“time period t” column and a “state” row represents the maximum total ex-
pected profits for time periods t to T, given that the state at time t is and the
price is
Similarly, given state and price at time t, the corresponding entries in

Tables 3(b)-(e) show respectively:
1. The optimal dispatch for time t.
2. The optimal commitment policy for time t, i.e., the next state to move
to at time t+1.
3. The maximum profits obtained from the optimal dispatch at time t.
4. The cost of the optimal commitment policy, or the cost to move from
the current state to the new state at time t+1.
Table 3 illustrates the results of a standard Backward DP computation.
These results are obtained, as is standard practice, by solving for the respec-
tive entries in the table from right to left.
Suppose that, at time t=l, the price level is HIGH and the generator is UP.
The maximum expected profits are $415.4 over the three time periods. Since
the current price level is HIGH and the generator is UP, the optimal commit-
ment policy is to stay UP at t=2 (from Table 3(c)), in spite of the possibility of
net losses over the three periods. The actual profits and the commitment poli-
cies at other times would, however, depend upon the actual price levels that
occur in those time periods. For example, if the price level stays HIGH for
both t=2 and t=3, the optimal commitment policy is to stay UP at t=3, realiz-
ing total profits of –5+500+50=$545. If, on the other hand, the price level
becomes LOW for t=2, and LOW for t=3, the optimal commitment policy is
to go DOWN at t=3, realizing profits of –5–10–22= –$37. In other words, the
generator loses money under some price patterns, even with the optimal pol-
icy. The expected profits are maximized, however.
The optimal schedules given by the backward DP are not static. Instead,
they depend upon the exogenous prices in each time period. Thus the DP
method does not merely give an optimal schedule. Rather, it gives an optimal
scheduling policy corresponding to different conditions.
In this section, we describe the “optionality” features of the generator
self-commitment problem, and show that it has features analogous to financial
options. We also make additional three points:
1. Assuming a single average price forecast generally understates the
value of the optionality, and could severely understate expected gen-
erator profits.
2. Running Monte Carlo methods without taking care to ensure that fu-
ture prices are always uncertain (at the time the commitment is made)
generally overstates the value of the optionality and overstates ex-
pected generator profits.
3. Not considering reserve products (and multiple markets) tends to
lower expected generator profits, because these additional products
increase generator optionality.
First, we consider the optionality due to price uncertainty. Assume a sin-
gle time-period horizon and a single product – the energy service. Assume
that the generator for which we want to find the optimal commitment and dis-
patch policy has no start-up or shut-down costs and no intertemporal con-
straints. Assume that generator has a capacity of 100 MW, no minimum gen-
eration constraint, and constant incremental costs of $30/MWh over this
range.
Table 4 shows five different price forecast scenarios. Two possible price
states, HIGH and LOW, each with a 50% probability of occurrence, represent
each price forecast. The expected value of the prices for all price forecasts is
$30/MWh. The optimal policy for the generator is to produce 100 MW when-
ever the energy price exceeds its incremental costs and to produce 0 MW
whenever the energy price is below its incremental costs. For example, for
forecast #3, the generator will produce 0 MW when the price is LOW (or
$20/MWh), and will make no profits. When the price is HIGH ($40/MWh),
the generator will produce 100 MW and make a profit of 100*(40-30) =
$1000. Since both price scenarios are equally likely for this forecast, expected
profits are 0*0.5+1000*0.5 = $500.
Table 4 shows that expected generator profits increase with increasing
price volatility8. This is analogous to the value of a financial option that in-
creases in value when price volatility increases [9]. On the other hand, if one
uses an average price of $30/MWh to find a commitment policy for the gen-
erators, then one will estimate that the generator will make no profit for any
price forecast because the generator incremental costs will not exceed the ex-
pected energy price. Therefore average price forecasts fail to calculate the
value of optionality and understate generator profits.
Next, we examine a potential pitfall associated with Monte Carlo meth-
ods. One possible way of finding the expected generator profits to capture the
optionality value could be to generate a large number of random price scenar-
ios for the time interval [0,T] by Monte Carlo methods. Using the ensemble
7
Price volatility here is defined as the ratio of the standard deviation to the expected price,
expressed in percent (and rounded).
8
All other factors (e.g., expected energy price) remaining constant.
of all generated price scenarios one could then use a deterministic model to
solve for the optimal commitment and dispatch over this period for each
member of the ensemble and then average over different Monte Carlo runs.
This is, however, not always correct. To see this, consider the following ex-
ample. Assume a two-period system, with each period having a 50% chance
of HIGH price ($35/MWh) and a 50% chance of having a LOW price
($10/MWh), regardless of the previous period. The generator has the same
characteristics as the above example, except that there is a minimum genera-
tion limit of 90 MW, and an additional intertemporal constraint that, once
online, the generator has to stay online for two consecutive periods. The
boundary condition is that the generator is offline initially and must be offline
at the end of two periods. It can be verified that the optimal policy is not to
commit the generators regardless of what the period 1 price is9. Therefore, the
expected profit under the optimal commitment policy is zero. On the other
hand, suppose that we first generate all the price scenarios (using a Monte
Carlo method), and then run a deterministic optimal unit commitment on each
possible price sequence. The four equally possible price sequences in the two
periods are {HIGH, HIGH}, {HIGH, LOW}, {LOW, LOW}, and {LOW,
HIGH}. If we make four deterministic unit commitment runs on these four
price sequences, the deterministic unit commitment will only run the genera-
tor at maximum output (100 MW) for both time periods when the price se-
quence is {HIGH, HIGH}. The profit for this price sequence is $1000. For all
other price sequences, the generator will not run, and the profit will be zero.
Hence, expected profits using this method will be 0.25*1000+0.75*0 = $250,
which clearly overstates the expected profits of zero under the true optimal
commitment policy. The reason for this is that in each Monte Carlo run, the
generator “peeped ahead” and “knew” the future prices and therefore chose
the commitment accordingly. Models for commitment based on the traditional
approach are likely to follow a variant of this deterministic optimization
method. This approach would result in overestimation of generator profits.
Monte Carlo methods are very efficient when one needs to simulate a
large number of different random outcomes and find the expected value (or
some other statistic) of some function based on these random outcomes. They
9
If period 1 price is HIGH, and the generator decided to commit, then the generator would
produce 100 MW in period 1 to make a profit of $500. There is a 50/50 chance, however,
that the period 2 price is HIGH or LOW. If the period 2 price is HIGH, the generator’s two-
period profit will be $1000. If period 2 had a LOW price, the generator would produce the
minimum 90 MW and lose 90*30=$2700 in period 2 for a net two-period loss of $2200.
Therefore, if the generator commits to be online when the period 1 price is HIGH, the ex-
pected two-period payoff is 1000*0.5 – $2200*0.5 = ($600), for an expected loss of $600.
are much more complicated to implement10, and prohibitively expensive,

when the value of a function at any given time t itself depends on what may
happen in the future, as in optimal commitment policy problems that have
intertemporal constraints11. For example, using the finance analogy, Monte
Carlo methods are used for European style options and for those other types
of options when one does not have to worry about when it is optimal to exer-
cise the option. American style options, however, are much more difficult to
solve for using Monte Carlo methods [12, pp. 685].
Consider a final example to show the optionality value of multiple prod-
ucts (or multiple markets). Assume a single time-period problem and a single
reserve product. Consider a generator whose incremental cost is $30/MWh,
maximum capacity is 100 MW, minimum capacity is 0 MW, and maximum
reserve capacity of 30 MW. Assume that the price of energy is $45/MWh and
reserve availability costs of $20/MW/h. If the generator offers 100 MW of
energy only, it will make profits of 100*(45–30) = $1500, on revenues of
100*45=$4500. On the other hand, if it maximizes its profits12 and offers 30
MW of reserve and 70 MW of energy, its profits are 30*20+
70*(45–30)=$1650 on revenues of 30*20+70*45=$3750. (Shifting more of
the generator output to reserves increases profits, even though total revenues
decrease relative to the energy-only sale. This is typical.)
In summary, the more optionality that a generator has, the higher its ex-
pected profits will be. Conversely, the more the operational constraints reduce
this optionality (e.g., intertemporal constraints), the lesser will be its expected
profits, all other factors being equal.
We now consider a slightly more realistic example.13 Assume a peaking
generator with the characteristics illustrated in Table 5. The start-up and shut-
down times for the generator are assumed to be zero. Figure 1 shows the base-
10
See reference [8] for the correct way to implement Monte Carlo methods for such problems.
In [8], however, the Monte Carlo method becomes prohibitively expensive as the number of
generator states and price uncertainty states increase. See also Section 6 for how to use
Monte Carlo methods once the optimal self-commitment policy is known.
11
The Monte Carlo method discussed in this section will work correctly on the problem de-
scribed in Table 4, because the optimal self-commitment policy does not have intertemporal
features.
12
Appendix A shows how one may approach the problem when there are more than one re-
serve type. In this example, while there is profit to be made on the sale of both energy and
reserves, the generator sees a higher profit margin in reserves and so maximizes the sale of
reserves (30 MW). The remainder is offered as energy (100-30=70 MW).
13
The results of this section were derived using EPRI’s PROFITMAX model.
line energy price forecast. Figures 2 and 3 illustrate the anticipated baseline
prices for 4 types of reserves: regulating, spinning, supplemental, and
backup.14 The total number of time periods is one week (168 hours).
14
Spinning reserves are defined to be the capability that can be offered in 10 minutes (if
online). Supplemental reserves are the amount of MW available in 20 minutes, and backup
reserves are the amount available in 60 minutes.
We model price uncertainty using a three-node Markov process. We as-

sume that prices can be at one of three levels: HIGH, BASELINE, or LOW.
The BASELINE prices are as illustrated. The HIGH price for energy and re-
serves is 115% of the corresponding baseline price, and the LOW price is
85% of the baseline price. We assume that there is perfect correlation between
energy and reserve availability prices; e.g., when the energy price is HIGH, so
are the reserve availability prices. For simplicity, we assume that the probabil-
ity of transition between any one price state to any other price state is 1/3.
That is, it is equally likely for the price to change states regardless of the pre-
sent state of prices.
Using the backward DP methods described in Section 3, we derived the
optimal commitment and dispatch policy. The optimal commitment and dis-
patch policy at any time t is a function of the state that the generator is in, and
a function of the uncertain price forecast for future time periods as observed at
time t. Future prices are always considered uncertain. We then applied the
optimal policy15 in numerous Monte Carlo runs to simulate different profit
outcomes. From the Monte Carlo prices, we then calculated the actual even-
tual profits and generator outputs. Using these outcomes, we then illustrate the
probability distribution of different generator outputs: generator profits, opti-
mal generator dispatch of energy and reserves, etc.
When we use the optimal policy found by the DP to simulate a number of
possible states, we obtain a distribution of possible outcomes. Figure 4 illus-
trates this distribution. There is no assurance that the distribution will be
neatly “bell-shaped” as in this example. In other examples it is possible to
have distributions that are skewed or bimodal, particularly when we consider
start-up and shut-down costs.
The effects of the optimal commitment policy on the state transitions are
shown in Figure 5. For a large number of scenarios, all transitions between
states that result from the optimal policy are recorded. For some times, the
state is unique (either UP or DOWN), but for other times the system could
end up in either state, depending on the price sequence. This is because it is
not known which state the generator will be at a future time t. Figure 5 shows
that, depending on the prices that are actually realized, the generator could be
in any of the different states at a future time. The ball “size” represents the
probability of ending up in a particular state. The “thickness” of a transition
line indicates the likelihood that the particular transition will take place.
15
We stress that the optimal policy was an input (not an output) in the Monte-Carlo runs.
During the interval depicted in Figure 5, the optimal policy has the option
to take several transitions, depending on the actual price realized. Only hours
93, 94, and 104 to 106 have a certain state (OFF in this case). For other peri-
ods the relative probability of being in either state is represented by the size of
the “ball” and the relative transition probability is represented by the thickness
of the transition “line.” During certain periods there is both an up transition
probability and a down transition probability in the optimal policy.
For comparison, Figure 6 illustrates the optimal commitment policy when

there is no uncertainty, i.e., the baseline price forecast is the perfect forecast.
It can be seen that the effect of price uncertainty recognizes the possibility
that the generator will either begin producing output (periods 104 through
108) or will turn off (periods 94 through 104) during certain periods accord-
ing to what prices are actually realized. That is, a quick-responding generator
has the luxury of producing when the prices are high, and going offline when
the prices are low. Thus high price volatility tends to be beneficial for the ex-
pected profits of the peaking generator. When prices are higher than its in-
cremental costs, the peaking generator will maximize its output (and increase
profits), while when prices are below its incremental costs, the generator will
shut down (and have zero profits). Since profits are bounded from below at
zero, and monotonically increase as a function of price (above the generator’s
incremental costs) the generator’s expected profits will increase.
Figure 7 illustrates this notion. It shows that when expected prices are
held constant among scenarios, expected profits increase as price uncertainty
increases. This again shows that committing and dispatching a generator is
analogous to exercising a financial option. The option value generally in-
creases when uncertainty increases.16 This is because a peaking generator is
able to follow changes more readily than other plant types. Thus, price volatil-
ity is beneficial to peaking units, a result that may be familiar to many read-
ers. In Figure 11, we are able to precisely quantify this benefit.
16
An intermediate or cycling generator has less optionality features because inter-temporal
constraints affect its profits, and it has features similar to Asian options [9]. A baseload gen-
erator has even less optionality features (excepting for the important question of which mar-
ket to sell into), usually because its incremental costs are generally well below the market
prices and is analogous to forward contracts [9].
Figure 8 illustrates the expected sales of energy and reserves as a function

of time. Note the striking fact that the generator in question offers energy only
during certain periods, but derives income from making reserves available
during many more periods.
7. CONCLUSIONS
The contribution of this chapter is to describe the multi-period multi-
market uncertainty framework within which decisions for unit commitment
and dispatch will have to take place for many units that operate in a deregu-
lated market. The chapter applies directly to the problem of optimal generator
self-commitment. It describes a method for finding the most profitable mar-
ket-responsive commitment and dispatch policy that takes into full account
the optionality available to a generator: reserve market opportunities, multiple
markets, price uncertainty, and intertemporal constraints. The model uses
backward dynamic programming, and the algorithm in the model can be
thought of as a generalized tree that values and exercises a sequence of com-
plicated options. This algorithm can be used to obtain optimal power market
bids for energy and reserve services in markets that integrate both these needs.
The method can also be used to profitably allocate output in different physical
forward markets, e.g., hour-ahead versus day-ahead.
ACKNOWLEDGMENTS
We thank EPRI PM&RM for support of this work, with special thanks to
Victor Niemeyer of EPRI for his support and encouragement. In particular,
the work on the PROFITMAX model developed by Christensen Associates
was sponsored by EPRI and initiated in 1997. We thank Blagoy Borissov of
Christensen Associates for his help in simulating some of the results in this
paper. We also thank Fritz Schulz for his help with the original implementa-
tion of PROFITMAX.
REFERENCES
1. A.J. Wood and B.F. Wollenberg. Power Generation Operation and Control, Wiley 1984.
2. A. Merlin and P. Sandrin. A new method for unit commitment at Electricite De France.
IEEE Trans. Power Apparatus Syst., PAS-102(5): 1983.
3. D.P. Bertsekas et al. Optimal short-term scheduling of large-scale power systems. IEEE
Trans. Autom. Cont., AC-28(1): 1983.
4. C.A. Li, R.B. Johnson and A.J. Svoboda. A New Unit Commitment Method. IEEE Trans..
Power Syst., 12(1): 1997.
5. C.L. Tseng, S.S. Oren, A.J. Svoboda, and R.B. Johnson. A unit decommitment method in
power system scheduling. Elec. Power Energy Syst., 19(6): 1997.
6. M. Hsu. Spark spread options are hot! Electricity J., 11(2): 1998.
7. S. Deng, B. Johnson, A. Sogomonian. “Exotic Electricity Options and the Valuation of

Electricity Generation and Transmission Assets.” Research Report #PSERC 98-13, 1998
(also available in the limited access http://www.pserc.wisc.edu/index_publications.html).
8. C.L. Tseng and G. Bartz. “Short Term Generation Asset Valuation.” Research Report
#PSERC 98-20, 1998 (also available in the limited access website http://www.pserc.
wisc.edu/index_publications.html).
9. J.C. Hull. Options, Futures, and Other Derivatives, Prentice Hall, 1997.
10. D.P. Bersekas. Dynamic Programming and Optimal Control, Volumes 1 and 2, Athena
Scientific, June 1995.
11. L. Kirsch and R. Rajaraman. Profiting from operating reserves. Electricity J., March: 1998.
12. H.R. Varian. Microeconomic Analysis, W.W. Norton & Company, Edition, 1992.
13. P. Wilmott. Derivatives: The Theory and Practice of Financial Engineering, John Wiley &
Sons, 1998.
APPENDIX A
This appendix describes how a profit-maximizing generator would dis-
patch energy and reserve availability services given exogenous market prices
for a given hour, and given that it is committed to be online.
Assumptions
Consider the output choice faced by a generator that can offer, in any
given hour, energy and four reserve services. Assume that the energy price is
PE, and that the availability prices for the four reserves are PR1, PR2, PR3,
PR4, respectively. Suppose that the generator has maximum output level ME,
and that, because of ramping limitations, the generator can provide the maxi-
mum quantities of the reserve XR1 for reserve 1, XR2 for reserve 2, XR3 for
reserve 3, and XR4 for reserve 4. Further suppose that the generator’s produc-
tion cost function is:
where a, b, and c are constants and XE is the generator’s quantity of offered

energy. What quantities of energy and reserves should the generator offer if it
is maximizing profits?
The Optimization Problem
The generators problem is to maximize profits17:
subject to the constraints that all energy and reserve quantities must respect
maximum limits:
17
When a generator offers reserves, there is a certain probability of these reserves being called.
When reserves are called to produce energy, they will receive the market energy price. One
can easily include this effect in the objective function (A-2).
For simplicity and without loss of generality, we ignore the constraints

that all reserve quantities must be non-negative, and that the energy quantity
have a minimum limit. Although equations (A-2) and (A-3) make the reserves
all appear to be mathematically identical, they are not because we assume
(reasonably) that reserve prices have a particular order18:
Because production costs (A-l) depend only upon energy output, the gen-
erator will prefer to sell Reserve 1 first and Reserve 4 last.
The Solution
The Lagrangian for the optimization problem is:
The solution to the foregoing problem is:
Note that the shadow value of capacity only if energy and reserves
use the generator to its full capacity, and that only if XRj=Mj for j=1,2,3.
The solution to the above problem is:
18
Theoretically, the reserve availability prices must be highest for the “highest” quality reserve
(regulation) and lowest for the “lowest” quality reserves (backup). The reasoning is that gen-
erators that can offer “higher” quality reserves can always offer “lower” quality reserves, but
not necessarily vice-versa. Therefore, the availability prices for “higher” quality reserves
must be higher than the “lower” quality reserves. Because of market imperfections, however,
this relationship is not always obeyed; e.g. see historical ancillary service prices from the
California ISO website (http://www.caiso.com). Note that in equation (A-4), we do not nec-
essarily assume that the highest quality reserves will have the highest price, i.e., we allow
market imperfections.
If Reserve 1 is (optimally) offered in a positive quantity that is less than

its limit, then But if Reserve 1 is (optimally) offered to its limit, then
In general:
if Reserve j is offered at all, then all reserves <j are at their limits;
if Reserve j is offered in a positive quantity that is less than its limit,
then and all reserves >j are not offered at all.
Chapter 7
A STOCHASTIC MODEL FOR A PRICE-BASED

UNIT COMMITMENT PROBLEM AND ITS
APPLICATION TO SHORT-TERM
GENERATION ASSET VALUATION
Chung-Li Tseng
University of Maryland
Abstract: In this paper, we model the unit commitment problem as a multi-stage stochas-
tic programming problem under price and load uncertainties. We assume that
there are hourly spot markets for both electricity and fuel consumed by the
generators. In each time period, the operator needs to determine which units
are to be scheduled so as to maximize the profit while meeting the demand.
Assuming that the price and load uncertainties can be represented by a sce-
nario tree, we develop a unit decommitment method using dynamic program-
ming to solve this problem. When there is only one unit under consideration,
we show that a scenario tree can be converted to a lattice that allows branch
recombination, which may greatly reduce the size of state space. This one-unit
problem can be used to value a generation asset over a short-term period. In
conclusion, we present our numerical results.
1. INTRODUCTION
Unit commitment is a problem to schedule generating units economi-
cally to meet forecasted demand and operating constraints, such as spinning
reserve requirements, over a short time horizon. The unit commitment deci-
sion determines which units will be used in each time period. It is a mixed-
integer programming problem and is in the class of NP-hard problems (e.g.
[1]). Because of its problem size and the NP-hardness, the optimal solution
of the unit commitment problem is normally difficult to obtain. Many opti-
mization methods have been proposed to solve the unit commitment prob-
lem. These methods include the priority list method [2], the dynamic pro-
gramming approaches (e.g. [3-5]), the branch-and-bound methods [6-8], and
the Lagrangian relaxation methods (e.g. [9-11]). Among them, the Lagran-
gian relaxation methods are the most advanced and widely used approaches.
Although the unit commitment problem has been widely studied during
the past decades, most of the approaches do not consider uncertainties. The
traditional unit commitment problem aims to schedule generation units to
meet the forecasted demand. When the actual demand is not equal to the
forecasted value, the discrepancy can be handled by system spinning reserve
to some extent. In [12], we present a stochastic model for the unit commit-
ment problem in which the demands are not known with certainty. We then
use a scenario tree to capture the demand uncertainty and the apply Lagran-
gian relaxation to decompose the problem into scenario subproblems using
progressive hedging [13]. Carpentier et al. in [14] propose another decompo-
sition scheme using the augmented Lagrangian method. In these two ap-
proaches, the unit commitment decision is modeled as a multi-stage prob-
lem. Carøe and Schultz develop a two-stage stochastic programming model
for the unit commitment problem, in which the authors emphasize the plan-
ning decision over the entire planning horizon rather than multi-stage im-
plementation of the operating decision [15].
With the evolution of deregulation in the electricity industry and the in-
troduction of spot markets for both electricity output and fuel input, incorpo-
rating uncertainties to the unit commitment problem becomes a necessity for
utilities or power generators. In this chapter, from the perspective of a firm
owning generating units, we model the unit commitment problem as a multi-
stage stochastic programming problem under price and load uncertainties.
We assume that there are hourly spot markets for electricity and the fuel
consumed by the generators. Our research uses a scenario tree to represent
the uncertainties as in [12]. However, we develop a new method using unit
decommitment to solve this problem.
Li et al. [16] and Tseng et al. [17] proposed independently the method
of unit decommitment for the traditional unit commitment problem. In [16],
the authors propose a solution procedure, which initially turns on all avail-
able units at all hours and then performs only decommitment. The authors
view their method as a Lagrangian relaxation-like method and take the mul-
tipliers from economic dispatch instead of sub-gradients. On the other hand,
in [17] the authors propose using unit decommitment as a post-processing
tool for existing solution procedures for solving the unit commitment prob-
lem. They consolidate these two approaches as a unified unit commitment
algorithm [18], in which they also conduct extensive numerical tests. Their
results show that the unit decommitment method on average obtains solu-
tions almost as good as the Lagrangian relaxation approaches, but with
much less CPU times.
In this chapter, we extend the unified approach presented in [18] to
solve the stochastic unit commitment problem. Assume that the price and
A Stochastic Model for a Price-Based Unit Commitment Problem 119
load uncertainties can be represented by a scenario tree. Initially, all generat-

ing units are committed in all scenarios and in all time periods, to the best
extent. Each unit’s schedule (in all scenarios) is then tentatively improved in
turn with other units’ schedules fixed. At each iteration, only one unit’s ten-
tative schedule is selected and updated. The iteration proceeds until no im-
provement can be made. Initial numerical tests seem to suggest that the unit
decommitment approach applied to the stochastic unit commitment problem
retains the properties when applied to the traditional deterministic case, as
reported in [18].
When there is only one generating unit under consideration, the problem
can be used to value a generation asset over a short-term period (e.g. [19]).
This problem appraises the flexibility of a power plant’s real options, e.g.
committing or decommitting a unit, in a competitive environment (e.g.
[20]). We will show that in this special case we can convert a scenario tree
to a lattice that allows branch recombination, which may greatly reduce the
size of state space.
In Section 2, we establish the mathematical model for the stochastic unit
commitment problem. Section 3 presents a special case in which the uncer-
tainties of the problem are perfectly known. Through this deterministic spe-
cial case, we derive the method of unit decommitment. Section 4 extends the
unit decommitment method to the stochastic case. We then discuss a short-
term generation asset valuation problem in Section 5. Section 6 gives nu-
merical test results, and we present future directions in Section 7.
2. THE MATHEMATICAL MODEL

In the model development, we introduce the following standard nota-
tion, with additional symbols introduced when necessary.
: index for the number of units
: index for time
: zero-one decision variable indicating whether unit is up or down in
time period
: state variable indicating the length of time that unit has been up or
down in time period
: the minimum number of periods unit must remain on after it has
been turned on
: the minimum number of periods unit must remain off after it has
been turned off
: the number of periods required for the boiler of unit to cool down
: decision variable indicating the amount of power unit is generating

in time period
: minimum rated capacity of unit
: maximum rated capacity of unit
: reserve available from unit in time period
: system demand (MW) in time period
: system reserve (MW) in time period (assumed to be a function of
)
: amount of power transaction in time period Its value can be positive
(power input) or negative (power output).
: electricity price ($/MWh) in time period
: fuel price ($/MMBtu) in time period
: the heat (MMBtu) required for unit to generate (MW) of power
(assumed strictly convex, increasing, and smooth)
: start-up cost associated with turning on unit at state
: shut-down cost associated with turning off unit
2.1 Operational Constraints

The operational constraints to be discussed in this paper can be de-
scribed in terms of state transition. Let be the set of unit commitment
state space for unit is composed of two subsets of the states: for
the on-line states, and for the off-line states.
The operational constraints include:

Minimum up-time/down-time constraints, for and
State transition constraints, for and
Equation (4) also implies the following relation between and
Capacity constraints for and
Initial conditions on for at
2.2 A Multi-Stage Stochastic Programming Formulation

This research considers the problem from the perspective of a firm that
owns generating units. We assume that there are hourly spot markets for
both electricity and fuel consumed by the generators. Each generating unit is
viewed as a cross-commodity instrument. That is, each generator buys fuel
from the fuel market, converts it to electricity, and sells the electricity to the
electricity market. The uncertainties considered in this model are the prices
for both electricity and fuel and the quantity of the electricity sold. The
quantity of electricity sold in the market , called load or demand in this
paper, may be contingent upon time (e.g. peak or non-peak hours) or may be
price-elastic. The firm may also involve option-type transactions, such that
the amount of transactions may depend on the price of electricity. For
example, a customer may tend to call an option of quantity (e.g. swing op-
tions) when the price of electricity is high and vice versa. The firm would
like to maximize its expected profit while meeting its load and transactions.
In the proposed stochastic model, the timing of the event occurrence is
as follows. In time period , the uncertainties are re-
vealed. Based on the states of all units from the previous hour, , the
operator needs to schedule the units to achieve maximum profit while satis-
fying the demand. The operator’s decision is not only based on the informa-
tion obtained in the current period, but also the expectation of future return.
Any commitment decision made in the current period will then become the
initial condition for decision-making in the following hour and will con-
strain the availability or flexibility of the unit commitment in the subsequent
hour(s). In addition, there are start-up and shut-down costs associated with
the commitment decision. In this paper, both costs are generalized in the
following function.
Let be the decision problem to be made in time pe-

riod , given the initial conditions and and the observed elec-
tricity and fuel prices The dynamic programming type recur-
rence equations can be formulated as follows:
where denotes the expectation operator, and the subscript indicates that
the expectation is based on the price information available at time . The
last term in (8) with the expectation operator defines another stochastic pro-
gram to be considered in the subsequent time period, which is also called a
recourse function. Since the fist term on the right-hand side of (8) is a con-
stant that can be pulled out from the maximization, can be decomposed
to two terms as follows.
Equation (10b) is subject to the physical constraints (3) to (6) and the fol-
lowing constraints.
Demand constraints for
Reserve constraints for

where and is the maximum reserve for

unit i.
In (9), accounts for revenue, and represents cost minimization.
The terminal condition of (9) is as follows
subject to (11) and (12) at time The optimal value representing the
maximum expected profit that the operator can make over the period
can be obtained from the last step of the recursive relation as
Finally, note that in the formulation the fuel price directly affects a
generator’s cost characteristics, and the electricity price is correlated to
the transaction amount and the load , which are uncertain per se. All
of these uncertainties influence unit commitments from different perspec-
tives.
2.3 Economic Dispatch

The minimization problem on the right-hand side of (10b) represents
scheduling generating units over a single period with generating costs
minimized. The scheduling decisions include the determination of what
units will be on (i.e. the commitment) and the generation levels for on-line
units. Given a commitment, the economic dispatch problem is to allocate
electricity generation economically to on-line units while satisfying the de-
mand and system reserve constraints. At time given state variable and
for each unit, let be the index set of on-line units at
time (Note that in this chapter a variable with a tilde hat will denote a re-
alization of the variable.) The economic dispatch problem at time denoted
by rced is defined below.
Note that rced stands for “reserve-constrained economic dispatch.” In [18]

the optimality condition of rced is interpreted as follows: the
units in are divided into two sets, one set of units with “cheap reserve”
but “expensive generation” (relatively), and one is the counterpart. The units
with cheap reserve but expensive generation are operated with the same
marginal costs (or as close as possible), which are the Lagrange multipli-
ers corresponding to (15b). The units with expensive reserve but cheap gen-
eration are operated such that their marginal costs for reserve equal (or are
as close as possible) to the Lagrange multipliers corresponding to (15c).
Proposition 1. The solution of rced exists if and only if the

following conditions hold.
3. THE DETERMINISTIC CASE

In this section we discuss a special case with certainty. An algorithm
will be developed, and will be extended to handle stochastics in a later sec-
tion. Assume that the future prices for electricity and fuel as well
as the load and transaction , are fully and perfectly
known. Equation (8) reduces to the following deterministic price-based unit
commitment formulation.
subject to constraints (3) to (6), (11) to (12), and initial conditions on at

for The first term in (17b) is a constant, and the minimiza-
tion term is a traditional unit commitment problem, which can be solved
using the methods reviewed in the introduction of this paper, such as the
Lagrangian relaxation method. In this section, we focus on the unit decom-
mitment method proposed in [17, 18] for solving the cost minimization
problem in (17b). First we review the unit decommitment method.
3.1 Unit Decommitment

The unit decommitment method was proposed in [17] as a post-
processing tool to improve solution quality for methods solving the tradi-
tional unit commitment problem. Given and a feasi-
ble schedule consider the problem of
optimally improving one unit’s schedule by decommitment. That is, in the
time periods when the unit under consideration is already off-line, the unit
remains off-line. The unit may be turned off in some on-line periods only if
doing so is cost-saving and would not cause infeasibility. While we are im-
proving a unit’s schedule, say determining for unit by decommitment,
the commitments of units other than are kept fixed. However, their genera-
tion levels in some time period may change in order to balance demand and
reserve if unit is decommitted in the same time period. The decommitment
rule is summarized below.
Constrained by the decommitment rule, the optimal decommitment problem

for a unit, say is formulated below.
subject to additional constraints such as (3) to (6) and (11) to (12), where
captures the total cost changed from units other than if unit
were turned off in time period The exact value of can be ob-
tained by solving the dispatch problem twice, one with unit committed at
time and the other without.
is a 0-1 integer programming problem, and can be solved using the

following dynamic programming recursive equation.
with boundary condition

where can be interpreted as the minimum cost for op-

erating unit over a period starting from time to with initial state
The optimal solution of is obtained from the last step of the dynamic
programming algorithm as In this research, we call the solu-
tion of the tentative commitment of unit
In the following algorithm, superscript denotes the iteration of
the algorithm. Let be the total cost for unit associated with
the schedule and be the optimal objective
value of solved with respect to (Note the variables in
bold faces are vectors. For example,
The unit decommitment algorithm
Data: Prices for and a feasible solution are

given.
Step 0: Set and evaluate for
Step 1: Solve with respect to and obtain for
Step 2: Select a unit such that If there is no such unit,
stop; otherwise update the commitment of unit in by
the tentative commitment obtained in The resultant unit
commitment is assigned to be
Step 3: Perform dispatch on to obtain and evaluate
for
Step 4: Set go to Step 1.
Although the unit decommitment method starts with an initial feasible

solution of the unit commitment problem, it can be used as a complete algo-
rithm for solving the unit commitment problem. Initially, as many units as
possible are turned on in all hours without violating the minimum downtime
constraints at the initial hours. This resultant unit commitment serves as an
initial commitment, to which the unit decommitment algorithm is then ap-
plied. There are, however, minor modifications required in order to make
this approach work. First, the commitments in the early iterations tend to be
seriously overcommitted such that the first inequality in (16a), the

minimum load condition, is violated. In other words, the dispatch phase in
Step 3 of the unit decommitment algorithm may not be feasible. When the
infeasible situation occurs, the algorithm would dispatch the units to satisfy
the optimality condition to the best extent, in the sense of matching the
units’ marginal costs for both fuel cost and reserve to the corresponding
multipliers and , respectively, as closely as possible (see [18]). By
doing this, as the decommitment procedure proceeds, the commitment ob-
tained eventually satisfies the minimum load condition, and the algorithm
starts to produce feasible schedules.
Instead of obtaining the exact value as in (20), may also be
approximated. In [18] the first order approximation of is derived.
where and are the multipliers associated with rced

With (23) plugged into (19), it is shown in [18] that starting from an eco-
nomically dispatched schedule performing either the unit de-
commitment step with respect to or the Lagrangian relaxation sub-
problem with respect to to any unit would result in the same (tenta-
tive) commitment. However, these two approaches differ in the commitment
updating (the unit decommitment updates one unit at a time, while the La-
grangian relaxation approach updates all units at once) and the multiplier
updating (the unit decommitment performs economic dispatch, while the
Lagrangian relaxation approach uses sub-gradients). In [18], based on their
numerical testing results on randomly generated instances, the authors report
that the error between the solutions obtained by the unit decommitment
method and Lagrangian relaxation approach is within 0.2%, and the unit
decommitment method takes at least 50% less CPU time than the Lagran-
gian relaxation approach on average. In the following section, we extend
and apply the unit decommitment method to the stochastic unit commitment
problem.
4. SOLVING THE MULTI-STAGE STOCHASTIC

MODEL USING UNIT DECOMMITMENT
The multi-stage stochastic model formulated in (8) appears to be intrac-
table generally. A popular approach is to summarize the future realization of
uncertainties by a finite number of possible scenarios. We represent scenar-
ios by the nodes of a tree such that given any node in the tree, there exists a
unique path leading to it from the starting node. A (directed) arc in a tree
connecting two nodes represents the evolution from one scenario to the
other. Consider a scenario tree T(N, A), where N is the set of nodes, and
A is the set of arcs in the tree. Let where is an ordered set

of nodes in the tree corresponding to time period t, and de-
notes the set of the descendents for node at time t such that
Assume the scenario at node is . For each node
and a descendent there is a branching (conditional) prob-
ability associated with this arc denoted by . Given a scenario tree
T(N, A) , the stochastic unit commitment formulated in (10) reduces to the
following equations for a node
and
subject to the demand constraints for
and the reserve constraints for
and (3) to (5) with an appropriate superscript in each variable to denote the
node in the tree. Next, we will show that we can extend the unit decommit-
ment method to solve (25). Given a scenario tree T(N, A) , and a feasible
schedule for similar to the development in
Section 3.1, we discuss what the best strategy is to improve the generating
cost from unit with other units’ strategies fixed in all scenarios (cf. (21)).
with boundary condition (cf. (22))

for Again is subject to the decommitment

rule,
and the minimum up-time and down-time constraints between and

in (3), and the state transition constraint for defined in (4), for
We obtain the optimal solution of from
the last step of the dynamic programming algorithm as
(Note contains only the start node.)
5. SHORT-TERM GENERATION ASSET

VALUATION
In this section, we consider a special case in which there is only one
generating unit (say unit Solving this problem can determine the expected
profit for a generation asset over a short-term period (e.g. [19]). This prob-
lem appraises the flexibility of a power plant’s real options, e.g. committing
or decommitting a unit, in a competitive environment (e.g. [20]). We will
show that, in this special case, a scenario tree can be converted to a lattice
that allows branch recombination, which may greatly reduce the size of state
spaces.
Recall the general formulation of the stochastic unit commitment in (8)
to (10). With employment of a scenario tree, these equations are reduced to
(24) and (25). In (28), the unit subproblem (for unit using unit decommit-
ment is presented. Although each unit subproblem (28) is solved independ-
ently, the commitment status of units other than are also required, which
are implicitly accounted through That is, the value of depends
on the states of all other units. However, at each node to evaluate at
all possible unit commitments is virtually impossible. The approach sug-
gested here is to incorporate a scenario tree such that for each node there
is one corresponding commitment state for each unit i to be determined,
as described in Section 4. In a tree there exists a unique path leading to any

given node from the starting node, and the commitment of all units can be
tracked in each path. By doing this, the tree size is inevitably large, an expo-
nential function of the number of times that the tree branches. For example,
if we would like to design a scenario tree such that each node branches into
two new nodes at each hour, for a 24-hour period the tree would contain
,more than 16 millions paths.
When there is only one generating unit (say j) under consideration, the
situation changes because Equation (28) reduces to
It is now possible to evaluate all possible states at each node since there
are only states of If is evaluated at all possible we
would not rely on the unique path leading to to track the history of unit j’s
commitment. Therefore, we can merge a scenario tree to a lattice that allows
branch recombination without losing any information. The following section
gives numerical results, along with an example of constructing a scenario
lattice.
6. NUMERICAL RESULTS
6.1 General Case with Multiple Units

The presented algorithm has been implemented in Fortran and applied to
a test problem on a Pentium II PC (400MHz). In the test problem, 10 gener-
ating units are considered. The heat requirement of each generator is mod-
eled as a quadratic function.
The start-up cost for is modeled as follows.
Table 1 summarizes the parameters of the generators.

Starting from a current electricity price at 25 ($/MWh), a scenario tree is

used to represent the variability of the electricity prices over the period
[1,24]. This 24-hour period is equally divided into 4 subperiods (1 to 6, 7 to
12, 13 to 18, and 19 to 24). For the first and the fourth subperiods, there are
two price scenarios indicating high (H) and low (L) price cases; for the sec-
ond and the third subperiods, there are three price scenarios indicating high
(H), medium (M) and low (L) price cases. Notations and i= 1,2,3,4 and
i=2,3 are used to represent the high, low, or medium price scenarios in
the i-th subperiod. Figure 1 illustrates the price tree, and Table 2 gives the
data of scenarios. Table 3 presents the conditional probabilities between
scenarios. For simplicity, assume that hourly demand follows the following
relation (including the effect of transaction
where can be considered as some base load, and represents some

“mean” prices for the electricity. When is deviated from , assume
the demand will also deviate from . The load deviation is proportional to
. This may be due to some options the firm has sold, which may be ex-
ercised at different price levels. Table 4 summarizes the data of , ,
and
We apply the proposed unit decommitment method to the test problem

and obtain a near-optimal solution within 2.3 seconds of CPU time. Table 5
reviews the algorithm performance. In all, we performed eight iterations.
Initially, at iteration k=0, all units in all scenarios are turned on at all hours,
to the most extent. The economic dispatch is then performed, and then the
expected profit of each unit is summarized under the column of Note
that at this point the system may be overcommitted, and the dispatch may
not exactly match the demand. In the first iteration (k=1), we perform sub-
problem to each unit and obtain the tentative schedule. One unit m
whose tentative schedule can yield the most improvement on the expected
profit, is selected. This value is recorded under the column
of for each unit. In the first iteration, unit 1 yields the most improvement
for the expected profit. Its tentative commitment that completely shuts down
the unit in all periods replaces its original one. Economic dispatch is per-
formed to the new unit commitment, and the expected profit of each unit is
recorded under The iteration goes on in a similar manner until no im-
provement can be made. We can observe the following:
1. The expected total profit is strictly increasing as the iteration in-
creases.
2. Once a unit has been selected for improvement at some iteration, its
(new) commitment remains “optimal” in the subsequent iterations.
For example, we select unit 1 at the first iteration. The improvement
that can be made through decommitment (under the column of
remains 0 for unit 1 in the subsequent iterations. We can adopt this
observation to improve the algorithm performance, because once a
unit has been selected at some iteration, it can be exempted from
further consideration. Therefore, the number of iterations required
by the algorithm is bounded by the number of units.
3. It can be verified that the following relation:
starts to hold when That is, after iteration 4, the algorithm has
produced a feasible unit schedule. Therefore, solution feasibility is
maintained after the 4-th iteration.
Finally, to verify how good the solution obtained from the proposed method
is, the dual problem using Lagrangian relaxation has been established. We
created a simplistic version of dual optimization to obtain an estimate of the
dual optimum, which can serve as an upper bound of the primal optimum.
Without much effort spent to fine-tune dual iterations, we obtain an ap-
proximate dual optimum of $1,100,339. Compared with the primal optimum
obtained from the proposed method of $1,089,819, the duality gap is within
0.96%. This implies that the obtained solution is fairly close to the true op-
timum.
6.2 Short-Term Generation Asset Valuation

Assume that the electricity and fuel prices follow some Ito processes. As
an example, consider the following two processes advocated in [21].
and
where and are two Wiener processes with instantaneous correlation

In the price models above, captures the seasonal average price. The
parameter captures reverting speed when the price deviates from and
is the price volatility. We characterize the above commodity price models
by mean reversion and log normally distributed, seasonal prices.
Let equivalently We
devised a two-dimensional price lattice to encompass both electricity and gas
prices. Each price node in the plane is designed to branch out into a
3×3 grid of 9 price nodes in the plane corresponding to the following time
period (see Figure 2). To form a lattice, the plane of each time pe-
riod is divided into a predetermined grid such that branching is only allowed
from grid nodes to grid nodes of the following time period. How to select the
nine nodes to which to map, as well as their corresponding (conditional)
probabilities, is discussed in [22]. Basically each branching must match (dis-
crete-time) price movement characteristics (mean, variance, and correlation)
implied by (36a) and (36b). We apply the branching process to each mapped
grid node and repeat the process until time period T is reached. Figure 2
illustrates one such lattice. Note that the size of the lattice grows quadrati-
cally (linearly in each variable) as time increases.
Furthermore, at each node we assume
Equation (37) essentially assumes that the operator would sell to the market
at the amount of electricity that optimizes profit, and that the market has in-
finite capacity. Note that (37) ignores the reserve constraint (27), and the
transaction is set to zero.
Remember that Table 1 give the proposed method as applied to unit 1.
Consider a 7-week (168-hour) period. Assume that the electricity prices and
fuel prices both follow the processes described by (36a) and (36b). We ob-
tain the parameters of the price processes are by fitting the historical price
data series of Nymex natural gas prices and electricity prices from the Cali-
fornia Power Exchange, taking the logarithm of these prices as our basic
data series. For gas we obtain and For electric-

ity we obtain and Refer to [22] for detailed val-
ues. At time 0 suppose prices and are
observed. We assume that the instantaneous correlation coefficient between
electricity and natural gas is
Figures 3 and 4 depict the relation between the expected profit of the
power plant and the length of planning horizon T and the volatility re-
spectively. We see that the expected profit for the power plant increases as
either T or increases. The expected profit is extremely sensitive to the
price volatility. From Figure 4, we can estimate that a 1% increase in the
value of would result in roughly a 1% increase of the expected profit for
the power plant. An intuitive interpretation of these results is to view own-
ing a power plant as holding a series of spark spread call options [20]. The
value of these options increases when (i) the number of options increases (as
T increases) or (ii) the price volatility increases.
7. CONCLUSIONS AND FUTURE DIRECTIONS

In this chapter, we present and model a general formulation of a price-
based unit commitment problem as a multi-stage stochastic programming
problem under price and load uncertainties. We extend the method of unit
decommitment using dynamic programming to solve this problem. The unit
decommitment method is very efficient in computation and does not require
a fine-tuning process for algorithm parameters like in the Lagrangian relaxa-
tion method. Preliminary numerical test results are promising. When we use
a scenario tree to describe the evolution of uncertainties as presented in this
research, it can be fully and efficiently integrated with a dynamic program-
ming approach such as the proposed unit decommitment method. The estab-
lishment of scenario trees remains an open problem. Statistical methods can
be used to measure the correlation of random variables such as prices and
loads at different hours. Also, the scenario trees seem to suffer the curse of
dimensionality. We demonstrate, using the example of short-term generation
asset valuation, that under some problem structure the dimensionality curse
can be relieved. How to model the stochastic unit commitment problem such
that it can provide a nice structure to cope with the dimensionality problem,
and can still reflect real operational restrictions, seems to be a new direction
of research.
REFERENCES
1. C.L. Tseng. On Power System Generation Unit Commitment Problems, Ph.D. Thesis,
University of California at Berkeley, 1996.
2. W.G. Chandler, P.L. Dandeno, A.F. Gilmn, and L.K. Kirchmayer. Short-range operation
of a combined thermal and hydroelectric power system. AIEE Trans., 733: 1057-1065,
1953.
3. C.K. Pang and H.C. Chen. Optimal short-term thermal unit commitment. IEEE Trans.
Power Apparatus Syst., 95(4): 1336-1342, 1976.
4. C.K. Pang, G.B. Sheble, and F. Albuyeh. Evaluation of dynamic programming based
methods and multiple area representation for thermal unit commitments. IEEE Trans.
Power Apparatus Syst., 100(3): 1212-1218, 1981.
5. A.J. Wood and B. F. Wollenberg. Power Generation, Operation and Control. New
York: Wiley, 1984.
6. A. Turgeon. Optimal unit commitment. IEEE Trans. Autom. Control, 23(2): 223-227,
1977.
7. A. Turgeon. Optimal scheduling of thermal generating units. IEEE Trans. Autom. Con-
trol, 23(6): 1000-1005, 1978.
8. T.S. Dillon. Integer programming approach to the problem of optimal unit commitment
with probabilistic reserve determination. IEEE Trans. Power Apparatus Syst., 97(6):
2154-2164, 1978.
9. J.A. Muckstadt and S. A. Koenig. An application of Lagrangian relaxation to scheduling
in power-generation systems. Oper. Res., 25(3): 387-403, 1977.
10. D.P. Bertsekas, G.S. Lauer, N.R. Sandell Jr., and T.A. Posbergh. Optimal short-term
scheduling of large-scale power systems. IEEE Trans. Autom. Control, 28(1): 1-11,
1983.
11. J.F. Bard. Short-term scheduling of thermal-electric generators using Lagrangian relaxa-
tion. Oper. Res., 36(5): 756-766, 1998.
12. S. Takriti, J.R. Birge, and E. Long. A stochastic model for the unit commitment prob-
lem. IEEE Trans. Power Syst., 11(3): 1497-1508, 1996.
13. R.T. Rockafellar and R.J.-B. Wets. Scenarios and policy aggregation in optimization
under uncertainty. Math. Oper. Res., 16(1): 119-147, 1991.
14. P. Carpentier, G. Cohen, J.-C. Culioli, and A. Renaud. Stochastic optimization of unit
commitment: a new decomposition framework. IEEE Trans. Power Syst., 11(2): 1067-
1073, 1996.
15. C.C. Carøe and R. Schultz. A Two-Stage Stochastic Program for Unit Commitment un-
der Uncertainty in a Hydro-Thermal Power System. Working paper, University of Co-
penhagen, Institute of Mathematics, Denmark, 1998.
16. C.A. Li, R.B. Johnson, and A.J. Svoboda. A new unit commitment method. IEEE Trans.
Power Syst., 12(1): 113-119, 1997.
17. C.L. Tseng, S.S. Oren, A.J. Svoboda, and R.B. Johnson. A unit decommitment method
in power system scheduling. Elec. Power Energy Syst., 19(6): 357-365, 1997.
18. C.L. Tseng, C.A. Li, and S.S. Oren. Solving the unit commitment problem by a unit
decommitment method. J. Optimization Theory Appl., 105(3), 2000 (in press).
19. C.L. Tseng and G. Barz. “Short-term generation asset valuation.” In Proc. 32nd Hawaii
International Conf. Syst. Sci. (CD-ROM), Maui, Hawaii, Jan 5-8, 1999.
20. M. Hsu. Spark spread options are hot! Electricity J., 11(2): 28-39, 1998.
21. G. Barz. Stochastic Financial Models for Electricity Derivatives. Ph.D. Thesis, Stanford
University, 1999.
22. C.L. Tseng. Valuing the operational real options of a power plant using discrete-time
price trees. Submitted to Elect. Power Energy Syst.
Chapter 8
PROBABILISTIC UNIT COMMITMENT

UNDER A DEREGULATED MARKET
Jorge Valenzuela and Mainak Mazumdar

University of Pittsburgh
Abstract: In this paper, we propose a new formulation of the unit commitment problem
that is suitable for the deregulated electricity market. Under these conditions, an
electric power generation company will have the option to buy or sell from a
power pool in addition to producing electricity on its own. We express the unit
commitment problem as a stochastic optimization problem in which the objec-
tive is to maximize expected profits and the decisions are required to meet the
standard operating constraints. Under the assumption of competitive market and
price-taking, we show that the unit commitment problem for a collection of M
generation units can be solved by considering each unit separately. The volatil-
ity of the spot market price of electricity is represented by a stochastic model.
We use probabilistic dynamic programming to solve the stochastic optimization
problem pertaining to unit commitment. We show that for a market of 150 units
the proposed unit commitment can be accurately solved in a reasonable time by
using the normal, Edgeworth, or Monte Carlo approximation methods.
1. INTRODUCTION
In the short-term, typically considered to run from twenty-four hours to
one week, the solution of the unit commitment problem (UCP) is used to as-
sist decisions regarding generating unit operations [1]. In a regulated market,
a power generating utility solves the UCP to obtain an optimal schedule of its
units in order to have enough capacity to supply the electricity demanded by
its customers. The optimal schedule is found by minimizing production cost
over a time interval while satisfying demand and operating constraints. Mini-
mization of the production costs assures maximum profits because the power
generating utility has no option but to supply the prevailing load reliably. The
price of electricity over this period is predetermined and unchanging; there-
fore, operating decisions have no effect on the firm’s revenues.
As various regions of the United States implement deregulation [2], the
traditional unit commitment problem continues to remain applicable for the
commitment decisions made by the Independent System Operator (ISO). The

ISO resembles very much the operation of a power generating utility under
regulation. The ISO manages the transmission grid, controls the dispatch of
generation, oversees the reliability of the system, and administers congestion
protocols [3,4,5]. The ISO is a non-profit organization. Its economic objective
is to maximize social welfare, which is obtained by minimizing the costs of
reliably supplying the aggregate load. Under deregulation, the UCP for an
electric power producer will require a new formulation that includes the elec-
tricity market in the model. The main difficulty here is that the spot price of
electricity is no longer predetermined but set by open competition. Thus far,
the hourly spot prices of electricity have shown evidence of being highly
volatile. The unit commitment decisions are now harder and the modeling of
spot prices becomes very important in this new operating environment. Dif-
ferent approaches can be found in the literature in this regard. Takriti et al. [6]
have introduced a stochastic model for the UCP in which the uncertainty in
the load and prices of fuel and electricity are modeled using a set of possible
scenarios. The challenge here is to generate representative scenarios and as-
sign them appropriate probabilities. Allen and Ilic [7] have proposed a sto-
chastic model for the unit commitment of a single generator. They assume
that the hourly prices at which electricity is sold are uncorrelated and nor-
mally distributed. In [8] Tseng uses Ito processes to model the prices of elec-
tricity and fuel in the unit commitment formulation.
The purpose of this chapter is to present a new formulation to the UCP
suitable for an electric power producer in a deregulated market and consider
computationally efficient procedures to solve it. We express the UCP as a sto-
chastic optimization problem in which the objective is to maximize expected
profits and to meet operating constraints such as capacity limits and minimum
up- and down-times. We show that when the spot market of electricity is in-
cluded, the optimal solution of a UCP with M units can be found by solving M
uncoupled sub-problems. We obtain a subproblem by replacing the values of
the Lagrange multipliers by spot power prices. The volatility of the spot mar-
ket price of electricity is taken into account by using a variation of the sto-
chastic model proposed by Ryan and Mazumdar [9], The model, which is re-
ferred to as the probabilistic production-costing model, incorporates the sto-
chastic features of load and generator availabilities. It is often used to obtain
approximate estimates of production costs [10,11,12]. This model ignores the
unit commitment constraints and assumes that a strict predetermined merit
order of loading prevails. This implies that a generator will be dispatched only
when the available unit immediately preceding it in the loading order is work-
ing at its full capacity. We believe that this model provides a good approxima-
tion to the operation of an electricity market, such as the California market, in
which no centralized unit commitment decisions are taken. The model cap-
Probabilistic Unit Commitment 141
tures the fundamental stochastic characteristics of the system. At any moment,

a power producer may not be fully aware of the exact characteristics of the
units comprising the market at that particular time. But it is likely to possess
information about the steady state statistical characteristics of the units par-
ticipating in the market. Ryan and Mazumdar’s probabilistic production cost-
ing model can be used to provide a steady-state picture of the market.
We determine the hourly spot market price of electricity by the market-
clearing prices. The market-clearing price can be shown to be the variable
cost or bid of the last unit used to meet the aggregate load prevailing at a par-
ticular hour. This unit is called the marginal unit. We determine the probabil-
ity distribution of the hourly market-clearing price based on the stochastic
process governing the marginal unit, which depends on the aggregate load and
the generating unit availabilities. We model the aggregate load as a Gauss-
Markov stochastic process and use continuous-time Markov chains to model
the generating unit availabilities [10,13]. We assume that the information is
available on mean time to repair, mean time to failure, capacity, and variable
operating cost of each unit participating in the market required to characterize
these processes. We use probabilistic dynamic programming to solve the sto-
chastic optimization problem pertaining to unit commitment. We also report
results on the accuracy and computational efficiency of several analytical ap-
proximations as compared to Monte Carlo simulation in estimating probabil-
ity distributions of the spot market price for electric power.
2. FORMULATION
We consider the situation in which an electric power producer owns a set
of M generating units and needs to determine an optimal commitment sched-
ule of its units such that the profit over a short period of length T is maxi-
mized. Revenues are obtained from fulfilling bilateral contracts and selling
electric power, at spot market prices, to the power pool. It is assumed that the
electric power company is a price-taker. If at a particular hour the power sup-
plier decides to switch on one of its generating units, it will be willing to take
the price that will prevail at this hour. We also assume that the power supplier
has no control over the market prices and the M generating units will remain
available during the short time interval of interest.
In determining an optimal commitment schedule, there are two decision
variables, denoted and The first variable denotes the amount of power
to be generated by unit at time and the latter is a control variable, whose
value is “1” if the generating unit is committed at hour and “0” otherwise.
The cost of the power produced by the generating unit depends on the
amount of fuel consumed and is given by a known cost function
where is the amount of power generated. The start-up cost, which for ther-
mal units depends on the prevailing temperature of the boilers, is given by a
known function The value of specifies the consecutive time that
the unit has been on (+) or off (-) at the end of the hour In addition, a gener-
ating unit must satisfy operating constraints. The power produced by a gener-
ating unit must be within certain limits. When the generating unit is run-
ning, it must produce an amount of power between and If
the generating unit is off, it must stay off for at least hours, and if it is on,
it must stay on for at least hours.
2.1 Problem Formulation

The objective function is given by the sum over the hours in the interval
[0,T] of the revenue minus the cost. The revenue is obtained from supplying
the bilateral contracts and by selling to the power pool at a price of per
MWH the surplus energy produced in each hour t. The cost in-
cludes the cost of producing the energy, buying shortfalls (if needed) from the
power pool, and the start-up costs. Defining the supply amount stipulated un-
der the bilateral contract by and by R ($/MWH) the price, the objec-
tive function (maximum total profit) is given by:
A positive value of indicates that megawatts hour are bought from

the power pool and a negative value indicates that megawatts hour are sold
to the pool. Since the quantity is a constant, the optimization problem re-
duces to:
subject to the following constraints(for and
Load:
Capacity limits:
Minimum down-time:
Minimum up time:
where
and unrestricted in sign
After substituting in the objective function the value of
obtained from Equation 4, we re-write Equation 3 as follows
which after removing constant terms is equivalent to:
subject to the operating constraints. Because the constraints (5) to (7) refer to
individual units only, the advantage of Equation 9 is that the objective func-
tion is now separable by individual units. The optimal solution can be found
by solving M de-coupled subproblems. Thus, the subproblem for the
unit
subject to operating constraints of the unit. Equation 10 is similar to the

subproblem obtained in the standard version of the UCP [14] excepting that
the value of the Lagrange multipliers are now replaced by the spot market
price of electricity.
2.2 Stochastic Formulation of the Subproblem

We next consider the spot market price of electricity, which is deter-
mined by the market-clearing price, as a random variable. When the optimiza-
tion subproblem is being solved for a particular unit, we assume that the mar-
ket, which includes the M-1 units owned by the power producer solving the
problem, consists of N generating units (N>>M). The generating unit for
which the subproblem is solved is excluded from the market. We assume that
the unit commitment decisions for any one unit have a negligible effect on the
determination of the marginal unit of the market for a given hour.
To model the market-clearing price, we assume that the generators par-
ticipating in the market are brought into operation in an economic merit order
of loading. The unit in the loading order has a capacity (MW), variable
energy cost ($/MWH), and a forced outage rate Under the assumption of
economic merit order of loading, the market-clearing price at a specific hour
is equal to the operating cost (S/MWH) of the last unit used to meet the load
prevailing at this hour. The last unit in the loading order is called the marginal
unit and is denoted by The market-clearing price, is thus equal to
The values of and depend on the prevailing aggregate load and the
operating states of the generating units in the loading order.
We write the objective function of the subproblem for one of the M gen-
erating units as follows:
subject to the operating constraints: capacity limits, minimum up-time, and

minimum down-time.
2.3 Probabilistic Dynamic Programming Solution

The maximum profit over the period T (Equation 11) is a random variable
because the hourly market-clearing price is a random variable. We assume
that at the time of the decision, hour zero, the marginal unit and the load for
all the hours before hour zero are known. We denote the marginal unit at time
zero by and solve the subproblem by maximizing the conditional expected
profit over the period T. We express the objective function as:
This equation is subject to the same operating constraints described ear-

lier. We use probabilistic dynamic programming to solve this optimization
problem. We define the function by the following equation:
This function denotes the maximum profit at hour given that at this hour
the unit is determining the market-clearing price and the generator to be
scheduled is in the operating state We also define the recursive function
to be the optimum expected profit from hour to hour T of operating the
generator that is in state at time Thus, the expression for hour zero is
For hour the expression is given by the following recursive relation:
Setting the expected incoming profit at time T+1 to be zero

we obtain the boundary condition for the last stage
The initial conditions are given by the initial state of the generator and
and the marginal unit at hour zero Consequently, the optimal schedule is
given by the solution of To solve the problem, the following condi-
tional probabilities need to be computed:
Thus, the joint probability distribution of and and the marginal

probability distribution of are needed.
3. STOCHASTIC MODEL FOR THE MARKET-

CLEARING PRICE
The stochastic model of the market-clearing price uses the production-
costing model proposed by Ryan and Mazumdar [9]. This model has been
used in estimating the mean and variance of production cost [12] and mar-
ginal cost [15] of a power generating system.
3.1 Stochastic Model of the Market

The assumptions of a market model with N generating units are:
1. The generators are dispatched at each hour in a fixed, pre-assigned loading
order, which depends only on the load and the availability of the generat-
ing units. Operating constraints such as minimum up- and down-times,
spinning reserve, and scheduled maintenance are not considered.
2. The unit in the loading order has capacity (MW), variable energy
cost (S/MWH), mean time to failure mean time to repair and
forced outage rate,
3. After adjusting for the variations in the ambient temperature and periodic-
ity, the load at time is assumed to follow a Gauss-Markov process
[16,17] with and where and are as-
sumed to be known. (An analysis given in [13] validates this assumption.)
4. The operating state of each generating unit follows a two-state continu-
ous-time Markov chain, with failure rate and repair rate
The forced outage rate equals
5. For and are statistically independent for all values of and
Each is independent of for all values of
3.2 Probability Distribution of the Marginal Unit

To derive an analytical expression for the probability mass function of the
marginal unit at time first note that
and that the events and are equivalent. Thus,
the following equality holds:
Thus, to obtain the probability mass function of the probability that
is greater than zero for all values of needs to be computed.
3.3 Bivariate Probability Distribution of the Marginal Unit

An analytical approximation for the bivariate probability mass function of
and needed for evaluating Equation 17, requires the following devel-
opment. Writing
and observing that the events
are equivalent, we obtain the following equality:
Therefore, to compute the bivariate probability mass function of

and the probability that
needs to be evaluated for all values of and
The computational effort in evaluating equations 18 and 20 depends on
the many values that the expression can take, which in the worst
case is (when Thus, the computational time increases exponentially

as N increases and it would make an exact computational procedure prohibi-
tive for large N. In our numerical examples, we have used three approxima-
tion methods: the normal, Edgeworth, and Monte Carlo approximations. The
Edgeworth approximation [18] is known in the power system literature as the
method of cumulants. We also attempted the use of the large deviation or
equivalently, the saddlepoint approximation method [19], but it turned out to
be prohibitively time-consuming for very large systems.
4. SOLUTION OF THE PROBABILISTIC UNIT

COMMITMENT PROBLEM: A NUMERICAL
EXAMPLE
For our purpose, we assume that a complete description of the electricity
market is given by the data concerning the N power generators that comprise
the market, historical data of the aggregate load, and the hourly temperature
forecast for the day of trading. The description of the power generators in-
cludes the order in which they will be loaded by the ISO, their capacities, en-
ergy costs, mean times to failure, and mean times to repair. The data for the
aggregate load gives the historically forecast ambient temperature and the cor-
responding load for each hour in the region served by the marketplace. In this
example, a data set that gave the actual ambient temperature and the corre-
sponding load for each hour in a region covering the northeast United States
during the calendar years 1995 and 1996 was used. We chose the last day of
this data set, September 20, 1996, as the trading day. Table 1 gives the actual
temperatures on this day, which were assumed to be the forecast temperatures.
Example: The market is described by the aggregate load and a power genera-
tion system consisting of generators participating in the market. Using statisti-
cal time series analysis on the data at hand, we found that an ARIMA
provided a very good description of the actual load ob-
served. The model [13] used is as follows:
where is the average ambient temperature at hour and is defined as:
and
where is a Gaussian white noise process with mean zero and estimated
variance Table 2 provides the estimates of the least-square re-
gression coefficients, relating load to temperature.
We assume the market consists of 150 power generating units. This sys-
tem was obtained by repeating ten times each unit of a 15-unit system, which
in its turn is a smaller version of the IEEE Reliability Test System (RTS) [20].
The load data from [13] was also multiplied by a factor of ten. Defining as
the cumulative capacity of the first i units,
we assume that the variable cost of each unit is given by the function:
This assumption allows for the unit operating costs to increase in order of the
position of the units in the loading order. Table 3 gives the relevant character-
istics of the units, in their loading order.
The problem is to schedule one of the generators of the power producer
for the next 23 hours given the information about the electricity market and
the known initial operating conditions for the generating unit. The characteris-
tics of this generator were taken from [1], and they are reproduced in Table 4.
We modified the fuel-cost function of the unit to be consistent with the range
of the individual units’ energy costs. The objective is to maximize the ex-
pected profit over this period. We assumed the generator to have been on for
eight consecutive hours. As mentioned in Section 2.3, this generator is not
included in the set of generators that comprise the market. We also assumed
that the variable cost of the unit is currently determining the market-
clearing price, which is $19.73 per MWH.
Table 5 summarizes the unit commitment solutions obtained using the dif-
ferent algorithms. The optimal schedule produced by the Monte Carlo simula-
tion (200,000 replicates used in estimating the probability distributions) is to
turn the generating unit off during the first four hours. Then, the unit is turned
back on for the next nineteen hours. The Monte Carlo procedure estimates
that the execution of this schedule will generate an expected profit of $37,509.
The normal and Edgeworth approximation methods provide this schedule as
well. However, they estimate expected profits of $37,483 and $37,351, re-
spectively. Details of these computations are given in [21].
5. DISCUSSION
In this chapter, we have proposed a new formulation of the unit commit-
ment problem that is valid under deregulation. We have shown that, when we
assume a competitive market and price-taking, the unit commitment problem
can be solved separately by each individual generating unit. The solution
method for the new formulation requires the computation of the probability
distribution of the spot market price of electricity. The power generation sys-
tem of the marketplace has been modeled using a variation of the Ryan-
Mazumdar model. This model takes into account the uncertainty on the load
and the generating unit availabilities. The probability distribution of the spot
market price, which is determined by the market-clearing price, is based on
the probability distribution of the marginal unit.
The exact computation of the probability distribution is prohibitive for
large systems. We evaluated three approximation methods. From the compu-
tational experience, it appears that the proposed unit commitment can be ac-
curately solved in a reasonable time by using the normal, Edgeworth, or
Monte Carlo approximations.
ACKNOWLEDGMENTS
The authors are indebted to the editors and the two reviewers for their
very helpful comments. This research was supported by the National Science
Foundation under grant ECS-9632702.
REFERENCES
1. A. Wood and B. Wollenberg. Power Generation Operation and Control. New York:
Wiley & Sons, Inc., 1996.
2. “Promoting Wholesale Competition Through Open Access Non-discriminatory Transmis-

sion Services by Public Utilities.” FERC Order No. 888, 1996.
3. F.D. Galiana and M.D. Ilic. A mathematical framework for the analysis and management
of power transactions under open access. IEEE Trans. Power Syst., 13: 681-687, 1998.
4. “Introduction to the New California Power Market.” Draft/work document posted on the
Internet, http://www.calpx.com, July 1997.
5. “Zonal Market Clearing Prices: A Tutorial California Power Exchange.” Document posted
on the Internet, http://www.calpx.com.
6. S. Takriti, B. Krasenbrink, and L. Wu. “Incorporating Fuel Constraints and Electricity
Spot Prices into the Stochastic Unit Commitment Problem.” IBM Research Report, RC
21066, Subject Area Computer Science/Mathematics, 1997.
7. E.H. Allen and M.D. Ilic. “Stochastic Unit Commitment in a Deregulated Utility indus-
try.” In Proc. North American Power Conference, Laramie, WY: 105-112, 1997.
8. C. Tseng. “A Stochastic Model for the Price-based Unit Commitment Problem.” Paper
presented at the Workshop on Next Generation of Unit Commitment Models, Co-
sponsored by DIMACS and EPRI, September 27-28, 1999.
9. S.M. Ryan and M. Mazumdar. Effect of frequency and duration of generating unit outages
on distribution of system production costs. IEEE Trans. Power Syst., 5: 191-197, 1990.
10. M. Mazumdar and A. Kapoor. Stochastic models for power generation system production
costs. Electric Power Syst. Res., 35: 93-100, 1995.
11. S.M. Ryan and M. Mazumdar. Chronological influences on the variance of electric power
production costs. Oper. Res., 40: S284-S292, 1992.
12. F. Shih, M. Mazumdar, and J.A. Bloom. Asymptotic mean and variance of electric power
generation system production costs via recursive computation of the fundamental matrix
of a Markov chain. Oper. Res., 47: 703-712, 1999.
13. J. Valenzuela and M. Mazumdar. Statistical analysis of electric power production costs.
Accepted for publication in IIE Trans.
14. J.F. Bard. Short-term scheduling of thermal-electric generators using Lagrangian Relaxa-
tion. Oper. Res., 36: 756-766, 1988.
15. F. Shih and M. Mazumdar. An analytical formula for the mean and variance of marginal
costs for a power generation system. IEEE Trans. Power Syst., 13: 731-737, 1998.
16. A.M. Breipohl, F.N. Lee, D. Zhai, and R. Adapa. A Gauss-Markov load model for
application in risk evaluation and production simulation. IEEE Trans. Power Syst., 7:
1493-1499, 1992.
17. J. Valenzuela, M. Mazumdar, and A. Kapoor. Influence of temperature and load forecast
uncertainty on estimates of power generation production costs. Accepted for publication in
IEEE Trans. Power Syst.
18. M. Mazumdar and Y. Wang. On the application of Esscher's approximation to computa-
tion of generating system reliability and production costing indexes. IEEE Trans. Power
Apparatus Syst., 104: 3029-3036, 1985.
19. S. lyengar and M. Mazumdar. A saddle point approximation for certain multivariate tail
probabilities. SIAM J. Scient. Comput, 19: 1234-1244, 1998.
20. APM Subcommittee. IEEE reliability test system. IEEE Trans. Power Apparatus Syst., 98:
2047-2054, 1979.
21. J. Valenzuela. Stochastic Optimization of Electric Power Generation in a Deregulated
Market. Ph.D. Dissertation, School of Engin., University of Pittsburgh, 1999.
Chapter 9
SOLVING HARD MIXED-INTEGER PROGRAMS
FOR ELECTRICITY GENERATION
Sebastian Ceria
Columbia University and Dash Optimization, Inc.
Abstract: In this chapter, we describe the most recent advances in the solution of mixed-
integer programming problems. The last ten years have seen enormous im-
provements in the solution of the most difficult mixed-integer programs. The
trend towards integration of modeling and optimization now makes it possible to
solve the hardest optimization problems arising from electricity generation, such
as the unit commitment problem. We report results with a leading software
package that was used successfully to solve unit commitment problems in two
European utility companies.
1. INTRODUCTION
Electricity generating companies and power systems face the daily prob-
lem of deciding how best to meet the varying demand for electricity. In the
short-term, electric utilities need to optimize the scheduling of generating
units so as to minimize the total fuel cost or to maximize the total profit over a
study period of typically a day, subject to a large number of constraints that
must be satisfied. This problem, which is of crucial importance for electric
generation companies, is referred to as the “unit commitment problem.” This
problem can be best solved by using state-of-the-art optimization technology.
Over the last twenty years, we have been witnesses to a revolution in
computational optimization. The availability of powerful computers and the
improvements in algorithmic development now make it possible to solve
problems that only a few years ago were thought to be unsolvable. Indeed,
modern commercial linear and integer programming codes, such as CPLEX1,
LINDO2, OSL3, and XPRESS-MP4 and research linear and integer program-
1
CPLEX is a trademark of ILOG, Inc.
2
LINDO is a trademark of LINDO Systems, Inc
3
OSL is a trademark of IBM Corporation
4
XPRESS-MP is a trademark of Dash Associates, Ltd.
ming codes, such as MINTO [28, 34], ABACUS [24], MIPO [8], and bc-opt
[9], are several orders of magnitude more powerful than the most efficient
implementations of linear and integer programming algorithms that were
available only ten years ago (when compared the present generation of com-
puters).
In this chapter, we will focus on the solution of basic mixed-integer (bi-
nary) programming problems of the form shown in Figure 1.
In Figure 1, is an A is a matrix of rows, is a vector of com-

ponents, and is an The first variables are constrained to taking
only 0-1 values; thus, the problem in Figure 1 is called a mixed 0-1 program.
Traditionally, integer programming problems have been solved by using
one of two algorithms, branch-and-bound and cutting planes, both based on
linear programming technology (see [23] for a complete survey of the area).
For the last ten years, several researchers have been studying and implement-
ing a combination of the above, which is commonly referred to as branch-and-
cut. The term was originally coined by Padberg and Rinaldi [31, 32] who
were the first researchers to demonstrate the computational efficiency of the
algorithm for the solution of traveling salesman problems.
But Padberg and Rinaldi’s branch-and-cut code goes beyond a simple
combination of branch-and-bound and cutting planes. It adds heuristics to
speed up the identification of feasible (almost optimal) solutions. It uses pre-
processing to reduce the size of the problem and hence improve its solvability.
Finally, Padberg and Rinaldi exploit the algebraic structure of the problem in
order to generate cutting planes. Hence, large-scale traveling salesman prob-
lems, involving thousands of cities, which translates into millions of variables
and an exponential number of constraints, can now be solved efficiently in a
few minutes. The success of their algorithm can be largely attributed to ex-
ploiting the mathematical knowledge of the problem.
Researchers have solved other hard optimization problems using similar
techniques. The caveat of the approach is that the algorithm has to be specifi-
cally tailored for a particular problem, with the added difficulty that any
Solving Hard Mixed-Integer Programs for Electricity Generation 155
changes to the model cannot be easily accommodated. It is for this reason that
commercial software companies have recently concentrated their efforts on
the development of programming environments that facilitate the develop-
ment of such algorithms (see for example [11]). The main idea behind such
products is to break the separation that has existed for decades between mod-
eling environments and optimization algorithms, thus making it easy for the
developer to prototype rapidly sophisticated algorithms for the solution of the
hardest optimization problems. In Section 4 we briefly explore an application
of this technology to the solution of unit commitment problems.
The resurgence of mixed-integer programming as a viable technology for
solving hard optimization problems is due to significant improvements in the
quality of the basic optimizers. Section 2 lists some of these developments
and includes several references. In Section 3, we study integrated tools for
modeling and optimization and discuss this paradigm shift. The integration of
modeling and optimization has been made possible by new software tools
available in the market and has recently been embraced by practitioners.
Finally, let us add that it is not the intent of this chapter to serve as a full
reference for the state-of-the-art in integer programming, nor to provide a
complete list of commercial or research software available for solving such
problems. Interested readers should also consult references, such as [7] and
[23]. Classic references for integer programming include [15], [29], and [35].
2. INTEGER PROGRAMMING: RECENT ADVANCES

Research in the algorithmic solution of mixed-integer programming prob-
lems started in the early sixties with the development of two classes of meth-
ods. The first one is that of cutting plane algorithms. In a cutting plane algo-
rithm, valid inequalities that cut off the current solution of the linear pro-
gramming relaxation are used to tighten the initial formulation until a feasible
integer solution can be found (see Figure 2). Their origins can be traced as far
back as the mid-1950s, with the landmark paper of Dantzig, Fulkerson, and
Johnson [12, 13] where they show how cutting planes can be used to solve a
48-city traveling salesman problem. Gomory [16, 17, 18], was the first one to
propose cutting planes as a general solution procedure. Branch-and-bound is
the second class of methods, which are used to solve mixed-integer programs.
The seminal papers of Land and Doig [25] for general mixed integer pro-
grams and of Balas [1] for the case of pure 0-1 programs introduced the basic
ideas that would later become the foundation for the area. Branch-and-bound
methods are of the enumerative type, since they solve the problem by enu-
merating a certain number, hopefully as small as possible, of feasible solu-
tions. Branch-and-bound has been for long the prevalent way chosen by prac-
titioners to solve mixed 0-1 programs.
In the last 10 to 15 years there have been several important innovations in

the practical solution of mixed-integer programs. The first is one of incorpo-
rating combinatorial cutting planes, as opposed to general cutting planes for
tightening the linear programming relaxation of certain classes of pure integer
programming problems. The main drawback of these methods is that there has
to be an underlying combinatorial structure in order to generate the cutting
planes. Crowder, Johnson, and Padberg [10] propose a method that circum-
vents this difficulty and that applies to pure 0-1 programs. In order to obtain
deep cuts they generate facets of the knapsack polyhedra obtained by consid-
ering each constraint in the constraint set separately. Van Roy and Wolsey
[37, 38] use similar ideas for mixed 0-1 programs. In both cases, after the lin-
ear programming relaxation had been tightened with the newly added cuts, the
authors apply branch-and-boun, and show that a significant reduction in the
number of nodes in the branch-and-bound tree can be obtained.
Recently, researchers have explored other ways of tightening the linear
programming relaxation of an integer program that involves reformulating the
problem into a higher dimensional space (one that uses more variables),
where a more convenient formulation may give a tighter relaxation. Such pro-
cedures are based on the work of Sherali and Adams [36]; Lovász and Schri-
jver [27]; and Balas, Ceria, and Comuejols [2], [3], [4].
Cutting planes can also be used in conjunction with branch-and-bound as
a method for strengthening the formulations at the nodes of the search tree
(see Figure 3). Padberg and Rinaldi [31] were the first to propose such a
method – that they baptized branch-and-cut – for the traveling salesman prob-
lem and obtained impressive results with it. There are two important keys to
the success of their approach. First, the combinatorial cutting planes they gen-
erate are deep cuts, and often facets, of the underlying integer polyhedron.
Second, because of the combinatorial nature of their cutting planes, the cuts
they generate at one part of the tree can be easily made valid throughout the
enumeration tree. There has recently been considerable interest in the devel-
opment of a branch-and-cut method that can be applied to general mixed 0-1
programs. Hoffman and Padberg [21, 22] obtained impressive results by us-
ing the cutting planes of Crowder, Johnson, and Padberg in a branch-and-cut
context for a class of sparse problems arising from applications.
Recent implementations of branch-and-cut have been proven it to be an

effective method for tackling some classes of mixed 0-1 programs, but there is
still an important need for a procedure that can solve general mixed 0-1 pro-
grams, independently of any underlying combinatorial structure. In [2, 3, 4]
we show that a branch-and-cut method based on lift-and-project cutting
planes is able to find optimal solutions to some instances of mixed 0-1 pro-
grams that were previously unsolved, and it is faster than competing methods
on other instances.
Most of these recent advances have been implemented in commercial
software. These improvements have been combined with preprocessing, heu-
ristics, and fancy data structures. A modern MIP code starts by preprocessing
the problem, thus reducing the size of the coefficients and eliminating un-
needed variables; then running a heuristic to improve the quality of the
bounds; and finally applying a combination of cutting planes and branch-and-
bound. It is now possible to solve problems that only a few years ago seemed
unsolvable. Most commercial packages incorporate features that are particu-
larly helpful when modeling some realistic problems in electricity generation.
For example, XPRESS-MP allows the user to define semi-continuous vari-
ables (variables that take either the value 0 or a value between a lower-bound
(greater than 0) and an upper-bound). This feature can be used to model elec-
tricity generation or consumption processes naturally.
We include computational results with XPRESS-MP5 are included in Fig-

ure 5 (for linear programming) and Figure 6 (for mixed-integer program-
5
XPRESS-MP is one of several commercial LP/MIP packages commercially available.
ming). The code is compared with its earlier version that came on the market
only one year before the current release. The significant improvements in so-
lution times are due to improved preprocessing, faster LP engines, and the
utilization of state-of-the-art cutting planes. The data sets are publicly avail-
able through either NETLIB or MIPLIB.
3. INTEGRATED MODELING AND OPTIMIZATION

Traditionally, mathematical programmers that use modeling languages to
express their problems have “artificially” separated modeling from optimiza-
tion. Practitioners and researchers have to sacrifice either performance by
constantly switching between a modeling and optimization environments, or
they have to design specific algorithms and implementations that completely
ignore modeling environments and work directly with an internal representa-
tion of the problem. In the first case, knowledge about the structured algebraic
model either missing and thus difficult to exploit. In the second, it is so em-
bedded into the implementation that small alterations to the model may render
the implementation completely useless.
There have been several attempts to integrate modeling environments and

optimization. The first such environment that has been implemented for the
solution of mixed-integer programming problems is EMOSL6 (Entity Model-
ing Optimization Subroutine Library) [6, 14]. EMOSL is a combined model-
ing and optimization subroutine library that overcomes the separation between
modeling and optimization by allowing the problem to be manipulated using
the notation of the model language. By integrating modeler and optimizer, the
scope and ease of algorithm development is greatly improved, without any
observable degradation of performance. Figure 7 provides an illustration of
the principle.
EMOSL was used in [11] for four applications: a block structure extrac-
tion tool and the implementations of three MIP heuristics and cutting planes
techniques. The result of their research effort is now a commercial product
that is also used in the electricity industry (see Section 4).
This does not mean that researchers have not been integrating modeling
and optimization in the implementation of their mixed-integer programming
algorithms. If this integration is done in the modeling environment then the
drawbacks in terms of speed are so great that these platforms cannot be con-
sidered for serious application of efficient algorithms. On the other hand, if
researchers use a low-level language, the original model becomes very hard to
maintain. For mixed-integer programming, the majority of current algorithmic
6
EMOSL is a trademark of Dash Associates Ltd.
implementations and development work uses conventional optimizer subrou-

tine libraries (for example: ABACUS [24], bc-opt [9], MINTO [28, 34] and
MIPO [8]). EMOSL is essentially a unification of an optimizer subroutine
library and a callable modeler. It provides all of the standard functions of an
optimizer library and many advanced features; we mention several later. It
provides functions to parse a model and to regenerate it after changes have
been made. The key feature is the new functionality to access the attributes of
the variables and constraints of the problem, i.e., functions to inspect and
change the coefficients of the variables in constraints, the sense of constraints,
the upper and lower bounds on variables, the type of variables (continuous,
binary, integer), and functions to inspect the primal and dual solution values
of the variables and constraints.
Other modeling languages, such as AMPL,7 GAMS,8 MPL,9 and OPL,10
are increasingly blurring the dividing line between modeler and optimizer. For
example, MPL has a feature that allows access to entities in the modeling lan-
guage from inside a C program, and OPL allows C/C++ model generation,
effectively allowing the combination of model and algorithm.
7
AMPL is a trademark of AT&T.
8
GAMS is a trademark of Gams Development Corp.
9
MPL is a trademark of Maximal Software, Inc.
10
OPL is a trademark of ILOG, Inc.
4. EXAMPLES IN ELECTRICITY GENERATION

There are several ways of formulating the UC problem. Most formula-
tions make use of mixed-integer and nonlinear mixed-integer models. Unit
commitment problems are intrinsically difficult because in many cases operat-
ing constraints are best represented by logical conditions, whereas in other
cases they are represented by non-convex functions.
The UC problem has been tackled by using a variety of methods. Heuris-
tic procedures were first studied in [20], while most of the current research
has focused on a combination of dynamic programming and Lagrangian re-
laxation [5, 19, 30, 33, 39].
In [26], Lekane and Gheury, from the Belgian electricity company Trac-
tebel, consider a system including nuclear units, fossil-fired units, and
pumped storage plants. Their model takes into account operating considera-
tions such as the operating reserve constraint, the minimum operating power
and ramping rate constraints, the commitment constraints of the fossil-fired
units, the operation of the nuclear plants at constant output levels, and the en-
ergy constraints associated with the pumped storage plants.
Their paper first presents a solution approach based on the Lagrangian re-
laxation method and discusses the difficulties associated with discovering a
primal feasible solution. Then the authors describe a new solution system,
which was investigated in the frame of the ESPRIT project called MEMIPS.
The approach presented there is based on a column generation method that
integrates dynamic programming to generate feasible power profiles for the
resources and linear programming using EMOSL to find a set of profiles
minimizing the total operating cost while respecting the system constraints.
Heuristics are then applied to determine integer solutions. The system heavily
exploits the integration of modeler and optimizer available through EMOSL.
By using EMOSL, the authors have the flexibility to change the model by
adding new constraints without necessarily affecting the algorithm implemen-
tation.
The approach presented in [26] is based on a reformulation of the UC
problem as a problem of determining the set of feasible thermal unit profiles
which minimizes the total operation cost over the studied period while re-
specting the demand and reserve constraints as well as the pumped storage
plant constraints. A feasible profile for a thermal unit is defined by both
power and reserve profiles, which specify its hourly generation and reserve
contribution over time. The principle of the proposed approach consists in
generating a set of feasible thermal profiles and solving the MIP problem over
the specified set of profiles. Since a huge number of thermal profiles exist, an
approximate procedure is used to solve the MIP problem over a limited set of
profiles. This procedure comprises four main steps, namely:
Generation of feasible thermal profiles: An initial set of feasible ther-

mal profiles is generated by using the solutions to the sub-problems found
when solving the dual problem in the Lagrangian relaxation approach.
Column Generation: The LP relaxation of the MIP that contains only
these thermal profiles is solved using column generation.
Relax and Fix: The resulting MIP problem is solved approximately by
using a relax and fix method.
Solution of reduced MIP: The MIP problem obtained by fixing the bi-
nary variables associated with the thermal units at the values found as a
result of the relax and fix heuristic is solved to optimality.
In this case, the advantages of using a framework that combines modeling

and optimization for such an implementation are clear:
The combined nature of the modeling and optimization subroutine library

facilitates the integration.
Access to variables and constraints in the modeler’s language allows for
ease of algorithm implementation.
An implementation that uses both modeling and optimization constructs is
extremely efficient since no overhead is incurred while communicating
with the modeling system.
The integrated modeling and optimization implementation (in EMOSL)
results in easy-to-maintain code.
Another example solving unit commitment problems using mixed-integer

programming comes from Power Optimisation, Ltd11. One version of the unit
commitment software, developed for Northern Ireland Electricity (NIE), has
been used every day since December 1996 to schedule the generating units in
the Northern Ireland power system. The users at NIE report that the schedules
produced by the software are consistently of a very high quality. The software
uses a novel multi-stage solution method, which drastically reduces the com-
puter run-time required to find an excellent feasible schedule to just a few
minutes. A great advantage of using and MIP method is that it has proved to
be easy and quick to introduce new constraints and features into the unit
commitment software. Hence the software can easily be adapted to model the
changes in the plant and operating rules of the power system that occur from
time to time.
11
More information on this application can be found at www.poweroptimisation.com. The
application was implemented using the XPRESS-MP software package.
5. CONCLUSIONS
In this chapter, we reviewed the latest developments in integer program-
ming and the integration of modeling and optimization. We illustrated the
principles in this integration through a several examples in the area of
electricity generation problems. Even though mixed-integer programming
techniques are not yet widely used in the electricity industry, we believe that
the recent developments in theoretical and practical mixed-integer
programming will enable practitioners to use these techniques more
effectively. Undoubtedly, the biggest challenge will be the efficient solution
of large-scale instances of these problems.
REFERENCES
1. E. Balas. Disjunctive programming. Ann. Discrete Math., 5: 3-51, 1979.
2. E. Balas, S. Ceria, and G. Cornuéjols. Mixed 0-1 programming by lift-and-project in a
branch-and-cut framework. Manage. Sci., 42(9): 1229-1246, 1996.
3. E. Balas, S. Ceria, and G. Cornuéjols. A lift-and-project cutting plane algorithm for mixed
0-1 programs. Math. Prog., 58: 295-324, 1993a.
4. E. Balas, S. Ceria, and G. Cornuéjols. “Solving Mixed 0-1 Programs by a Lift-and-Project
Method.” In Proc. Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, 232-
242, 1993b.
5. J. Batut and R.P. Sandrin. “New Software for the Generation Rescheduling in the Future
EDF National Control Center.” In Proc. PSCC, Graz, Austria, 1990.
6. Belvaux, G., Solving lot-sizing problems by branch & cut: Basic guide to use of mg-
EMOSL. Working paper, CORE, Université catholique de Louvain, February 1998.
7. A. Caprara and M. Fischetti. Branch-and-Cut Algorithms, in Annotated Bibliographies in
Combinatorial Optimization. M. Dell’Amico, F. Maffioli, and S. Martello, eds. Chiches-
ter: John Wiley & Sons, Chichester, 45-63, 1997.
8. S. Ceria. “MIPO: A Mixed-Integer Program Optimizer.” Presented at: International
Symposium on Mathematical Programming, Lausanne, August 1996.
9. C. Cordier, H. Marchand, R.S. Laundy, and L.A. Wolsey. “bc-opt: A branch-and-cut code
for mixed integer programs.” Discussion paper DP 9778, CORE, Universite catholique de
Louvain, 1997.
10. H. Crowder, E. Johnson, and M. Padberg. Solving large-scale zero-one linear program-
ming problems. Op. Res., 31: 803-834, 1983.
11. R. Daniel and J. Tebboth. A tightly integrated modelling and optimisation library: A new
framework for rapid algorithm development. Ann. Op. Res., 0: 1-21, 1998.
12. G.B. Dantzig, D.R. Fulkerson, and S.M. Johnson. Solution of a large-scale traveling
salesman problem. Op. Res., 2: 393-410, 1954.
13. G.B. Dantzig, D.R. Fulkerson, and S.M. Johnson. On a linear programming, combinatorial
approach to the traveling salesman problem. Op. Res., 7: 58-66, 1959.
14. Dash Associates. XPRESS-MP EMOSL Reference Manual, Release 10. Blisworth, Nort-
hants NN7 3BX, UK: Blisworth House Church Lane, 1997.
15. R.S. Garfinkel and G.L. Nemhauser. Merger Programming. New York: John Wiley &
Sons, Inc., 1972.
16. R. Gormory. “An Algorithm for the Mixed-Integer Problem. Technical Report RM-2597,
The Rand Corporation, 1960.
17. R. Gormory. 1958. Outline of an algorithm for integer solutions to linear programs. Bull.
Am. Math. Soc., 64: 275-278, 1958.
18. R. Gormory. “Solving Linear Programming Problems in Integers. In Combinatorial
Analysis R.E. Bellman, M. Hall Jr., eds. Providence, RI: American Mathematical Society,
211-216, 1960.
19. X. Guan, P.B. Luh, and L. Zhang. Non-linear approximation method in Lagrangian relaxa-
tion-based algorithms for hydrothermal scheduling. IEEE Trans. Power Syst., 10(2): 772-
778, 1995.
20. H.H. Happ, R.C. Johnson, and W.J. Wright. Large-scale unit commitment method and
results. IEEE Trans. PAS, PAS-90: 1373-1383, 1971.
21. K.L. Hoffman and M. Padberg. Techniques for improving the LP-representation of zero-
one linear programming problems. ORSA J. Computing, 3: 121-134, 1991.
22. K.L. Hoffman and M. Padberg. Solving airline crew scheduling problems by branch-and-
cut. Manage. Sci, 39: 657-682, 1993.
23. E.L. Johnson, G.L. Nemhauser, and M.W.P. Savelsbergh. “Progress in Linear Program-
ming Based Algorithms for Integer Programming: An Exposition.” School of Industrial
and Systems Engineering, Georgia Institute of Technology, Atlanta, Georgia, August
1999.
24. M. Jünger and S. Thienel. Introduction to ABACUS – A branch-and-cut system. Op. Res.
Lett., 22: 83-95, 1998.
25. H.A. Land and A.G. Doig. An automatic method for solving discrete programming prob-
lems. Econometrica 28: 497-520, 1960.
26. T. Lekane and J. Ghuery. “Short Term Operation of an Electric Power System.” Tractebel
Energy Engineering, Brussels, Belgium.
27. L. Lovasz and A. Schrijver. Cones of matrices and set-functions and 0-1 optimization.
SIAM J. Optimization, 1: 166-190, 1990.
28. G.L. Nemhauser, M.W.P. Savelsbergh, and G.C. Sigismondi. MINTO, a Mixed INTeger
Optimizer. Op. Res. Lett., 15: 47-58, 1993.
29. G.L. Nemhauser and L.A. Wolsey. Integer and Combinatorial Optimization. New York:
John Wiley & Sons, 1988.
30. S.O. Orero and M.R. Irving. Large-scale unit commitment using a hybrid genetic algo-
rithm. Elec. Power Energy Syst., 19(1): 45-55, 1997.
31. M.W. Padberg and G. Rinaldi. A branch-and-cut algorithm for the resolution of large-
scale symmetric traveling salesman problems. SIAM Rev., 33: 60-100, 1991.
32. M.W. Padberg and G. Rinaldi. Optimization of a 532 city symmetric traveling salesman
problem by branch-and-cut. Op. Res. Lett., 6: 1-7, 1987.
33. C.K. Pang, G.B. Sheblé, and F. Albuyeh. Evaluation of dynamic programming based
methods and multiple area representation for thermal unit commitments. IEEE Trans.
PAS, PAS 100(3): 1212-1218, 1981.
34. M.W.P. Savelsbergh and G.L. Nemhauser. Functional description of MINTO, a mixed
integer optimizer, version 2.3. School of Industrial and Systems Engineering, Georgia In-
stitute of Technology, Atlanta, Georgia, November 1996.
35. A. Schrijver. Theory of Linear and Integer Programming. Chicester: John Wiley & Sons,
1986.
36. H. Sherali and W. Adams. A hierarchy of relaxations between the continuous and convex
hull representations for zero-one programming problems. SIAM J. Disc. Math., 3: 411-
430, 1990.
37. T.J. Van Roy and L.A. Wolsey. Solving mixed-integer programming problems using
automatic reformulation. Op. Res., 35: 45-57, 1987.
38. T.J. Van Roy and L.A. Wolsey, Solving mixed 0-1 programs by automatic reformulation.
Op. Res. 35: 145-163, 1987.
39. S.J. Wand, S.M. Shahidehpour, D.S. Kirschen, S. Mokhtari, and G.D. Irisarri. Short-term
generation scheduling with transmission and environmental constraints using an aug-
mented Lagrangian relaxation. IEEE Trans. Power Syst., 10(3): 1294-1301, 1995.
Chapter 10
AN INTERIOR-POINT/CUTTING-PLANE
ALGORITHM TO SOLVE THE DUAL UNIT
COMMITMENT PROBLEM – ON DUAL
VARIABLES, DUALITY GAP, AND COST
RECOVERY
Marcelino Madrigal and Victor H. Quintana

University of Waterloo
Abstract: In this chapter, we use an interior-point/cutting-plane (IP/CP) method for non-

differentiable optimization to solve the dual to a unit commitment (UC) prob-
lem. The IP/CP method has two advantages over previous approaches, such as
the sub-gradient and bundle methods: first, it has proved to have better conver-
gence characteristics in an actual implementation; and second, it does not suffer
from the parameter-tuning drawback. The results of performance tests using
systems with up to 104 units confirm the superiority of the IP/CP method over
previous approaches to solve the dual UC problem. We discuss issues that have
influenced whether or not UC models are used as the clearing mechanism in
electricity markets; these issues include duality gap, cost recovery, and the exis-
tence of multiple solutions.
1. INTRODUCTION
Lagrangian relaxation (LR) has become one of the most practical and ac-
cepted approaches to solve unit commitment (UC) problems of real-life di-
mensions [1-3]. The key idea in LR-based approaches is to solve the dual
problem instead of the primal; the dual function has a separable structure, i.e.,
in a per thermal-unit basis, which permits its easy evaluation and, at the same
time, provides a primal solution. The dual function is concave but not differ-
entiable [16]; therefore, non-differentiable optimization techniques are re-
quired to solve the dual problem. Pioneering work on LR-based UC solution
approaches has used sub-gradient (SG) methods as the dual maximization
engine [2, 4]. Despite their bad convergence characteristics, they are still be-
ing used due to its simplicity and low per-iteration computer effort. Several
cutting-plane (CP) variants to solve non-differentiable optimization have been
employed to solve the dual to unit commitment or other power scheduling
problems. For instance, in [5] a penalty-bundle (PB) method is used to solve
the UC problem. In [6], a CP with dynamically adjusted constraints is used to
solve the hydrothermal coordination problem. In [7], a reduced complexity
bundle method is introduced to solve the dual of a power-scheduling problem.
All these cutting-plane variants still have the disadvantage that parameters
need be carefully tuned; these parameters define a stabilization scheme that
prevents unboundedness in the maximization of the dual function and help
improve convergence [8].
Interior-point/cutting-plane (IP/CP) methods have been used to solve non-
differentiable problems in engineering applications, such as the lot sizing [9]
and multi-commodity flow problems [10]. Recently, the authors have pro-
posed the use of an IP/CP method to solve the UC problem [11]. This paper
describes the use of such a non-differentiable optimization technique to solve
the dual to a unit commitment problem. IP/CP methods have two advantages
over previous approaches, such as the sub-gradient and penalty-bundle meth-
ods: first, they have better convergence characteristics; and second, they do
not suffer from the parameter-tuning drawback.
The first mechanism used by an electricity market to select power among
competing generators was a unit commitment model. In this Pool market-
model, generators act as independent entities and the pool operator, based on
the costs submitted by generators and their physical limitations, decides the
schedules for all generators and sets a market price. UC models may naturally
have multiple solutions that, from the Pool point of view, are equally good (all
minimize the cost) but, for generators, it means different schedules and, there-
fore, different revenues that create a clear conflict of interest. LR algorithms
are still dependent on heuristics and are not able to identify or distinguish
multiple solutions [12-13].
The IP/CP method presented in this chapter has eliminated one of the
drawbacks (tuning in the dual maximization) of LR algorithms. We also pre-
sent some other findings that relate duality gap and cost recovery when dual
variables set the market price. It has been shown that UC cost-minimization
models and artificial price definitions, derived from their solutions, do not
necessarily lead to lower prices for consumers [14]. For a simplified UC
model, we show that, in the absence of duality gap, the dual variables are the
minimum uniform market price that recovers participants’ cost.
In Sections 2 and 3, we describe the unit commitment problem and its
dual. Section 4 presents the solution of the dual problem by the IP/CP method.
Section 5 discusses the issues of duality gap, stopping criterion, cost recovery,
Unit Commitment By Interior-Point/Cutting-Planes 169
and multiple solutions. Section 6 contains numerical results on the perform-

ance of the IP/CP method, and we give some conclusions in Section 7.
2. THE UNIT COMMITMENT PROBLEM

The unit commitment problem consists on determining the generators that
need to be committed and the production levels that are required for supplying
the forecasted short-term (24 or, at most, 168 hours) demand and spinning
reserve requirements (2-3); all of this at the minimum cost (1). The operation
of the units is subject to several constraints (4), to be described. The primal
unit commitment (PUC) is a very large, non-linear, mixed-integer problem;
therefore, the non-convex optimization problem:
T is the set of time periods in the optimization horizon (subsequently | T |

denotes the number of periods); I is the set of thermal units. The objective
function is comprised of production, and start-up cost, These costs are
represented, respectively, using the classic [2] quadratic-convex and exponen-
tial models:
The commitment state of unit at time is defined by the binary variable

(1: committed, 0: decommitted); represents the power output
of the unit. Equations (2) and (3) are the constraints to satisfy the system-wide
power, and reserve, requirements. The reserve contribution of every
thermal plant is given by where is the maximum allow-
able power output.
In (4), the vector contains the commitment states and production levels
of the particular unit that is, Each thermal
unit operation constraint set, contains: (i) minimum and maximum power
output constraints; (ii) ramp-constraints; and, (iii) minimum up- and down-
time constraints [2]. In general, the sets contain non-linear and mixed-
integer restrictions. The UC has been proven to be NP-hard [15].
3. THE DUAL UNIT COMMITMENT PROBLEM

The (Lagrangian) dual problem to PUC is
We obtain the dual objective function, from the Lagran-

gian that is formed by adding to the objective function (1), through Lagrange
multipliers, the system-wide demand constraints (2), and the reserve con-
straints (3). The dual variables of the power-balance equality constraints are
not necessarily constrained to be positive, but, since the objective function,
see (5-6), always increases with power production, they will always take posi-
tive values. After arranging some terms, the dual objective function has the
following separable structure:
where
and is the dual variable vector. The dual func-

tion (8), as proven in [16], is concave but not differentiable. The separable
nature of the dual function is exploited in a two-stage solution process (La-
grangian Relaxation) of the DUC problem.
3.1 The Lagrangian Relaxation Algorithm

The LR approach to solve the UC problem is outlined below:
Initialization.
Obtain an initial dual vector and set Obtain a priority-list UC
and afterwards perform a simplified economic dispatch for each time
The dual variables of the power demand constraints are used to initialize
the respective Spinning reserve constraints’ dual variables are initial-
ized to zero, i.e.,
Phase 1. DUC Maximization.
1. Dual objective function evaluation. From (8), evaluate the dual objective
function by solving the | I | individual unit commitment subprob-
lems (9) using a forward dynamic programming approach. A primal vec-
tor, (not necessarily feasible) is obtained after the subproblems are
solved.
2. Convergence test. If a convergence criterion is satisfied, then go to
Phase 2. Otherwise, continue.
3. Dual Maximization. Find an improved dual solution vector, using
a non-differentiable optimization technique. Set go back to 1.
Phase 2. Feasibility search.
Use a heuristic procedure to map the non primal-feasible solution, (ob-
tained in Phase 1) to a feasible one, say Usually, feasibility is achieved
with little modifications on We follow the procedure described in [2]
to perform such feasibility search.
The structure of the non-linear, mixed-integer subproblems (9) is such that

they can be easily solved using a dynamic programming [2-3] or branch and
bound techniques. Semi-definite programming has also been used to solve the
UC subproblems [17].
4. AN INTERIOR-POINT/CUTTING-PLANE METHOD
TO SOLVE THE DUC PROBLEM
Classic methods for non-differentiable optimization, such as penalty-
bundle and Kelley’s cutting plane methods [8], maximize a cutting plane ap-
proximation of the objective function over a set of restrictions (stabilization
region) that encloses the optimal solution and helps improve convergence by
properly setting parameters. Instead of maximizing a cutting plane
approximation over a stabilization region, IP/CP method takes the analytic
center of a localization set as the new improved dual solution [9, 18]. The
localization set, is a convex closed region defined by
The components of are described as follows: (i) a cutting-plane ap-

proximation of the UC dual function; (ii) the known dual variables’ lower
bounds (iii) a box constraint and (iv) a lower bound to the
dual objective function value. Constraints (ii) and (iii) limit the localization
set from the “left” and “right,” respectively; and the constraints (i) and (iv)
limit the localization set from “above” and “below,” respectively. The cutting-
plane approximation of the dual function is constructed using a bundle of in-
formation from all previous iterations of the maximization process. A bun-
dle is a collection of: (i) dual vectors (ii) their corresponding
dual objective function values and (iii) the sub-gradients
A sub-gradient at iteration is readily available by com-
puting the mismatches of the power and reserve constrains,
This vector always belongs to the sub-differential set
Therefore, it qualifies as a sub-gradient for the concave function (8) at

As pointed in [18], the selection of the box constraint, in (10), has a
limited influence on the convergence characteristics of the IP/CP method; in
practice, any large enough number based on knowledge of the problem can be
chosen. For the UC problem, similar convergence characteristic is achieved
with any large value as presented in the results section. This value is ob-
tained based on units’ cost coefficients. A lower bound of the dual objective
function is readily available from previous iterations, as
The localization set in (10) can be rewritten as
where
with and and

Note that has iteration-increasing size. The analytic center
of is defined as the point solving the potential problem
where defines a potential function whose maximum is achieved at

a point that is centered in the localization set. For example, a non-centered
point touching hyperplane has and the associated potential compo-
nent From is taken as an updated dual solution vec-
tor, which is used to evaluate the dual function. If a stopping criterion is not
satisfied, the localization set (10) is updated to by adding a new cut,
and by replacing the dual-function lower
bound approximation by The analytic center of the updated
localization set is obtained and the process in repeated again. We depict the
third, fourth, and fifth iterations of an IP/CP method for a dual function in
schematically in Figure 1. The dot inside each updated localization set
(shaded region) represents the analytic center; the horizontal dotted line repre-
sents the lower bound and the bold curved line represents the dual func-
tion. This figure depicts classical behaviour of the IP/CP method; cuts gener-
ated from the analytic center are deeper and the localization set rapidly
shrinks towards a single point corresponding to the optimal value.
Problem (14) can be efficiently solved using a primal, dual, or primal-

dual interior-point method [19]. In [11], the authors use a semi-definite pro-
gramming formulation to solve the potential problem. In this chapter, we em-

ploy a primal-dual logarithmic-barrier interior-point method to solve the po-
tential problem (14) and describe it briefly in the following paragraphs.
We derive the primal-dual interior-point method from Kojima, Megiddo,
and Mizuno’s primal-dual infeasible-interior-point algorithm [20]. The bar-
rier-Lagrangian to (14) is given by
where is an iteration-decreasing barrier parameter The first-

order necessary conditions for the optimality of (15) are
In these equations, the complementarity primal and dual

residuals are defined. From an initial strictly positive point, the
primal-dual interior-point method generates new points of the form
We compute the search directions by a one-iteration Newton’s method that is

applied to solve the optimality conditions (16)-(18). The x and y Newton’s
directions can be computed by solving the reduced system
The s -search direction is obtained from
The step length in (19) is computed as where

and is a safety factor that prevents variables touching zero. A

barrier parameter reduction scheme, derived from (16), is given by
The algorithm can be stopped when the optimality conditions (16)-(17)

are satisfied up to an specified tolerance.
5. ON DUALITY GAP, COST RECOVERY, AND

MULTIPLE OPTIMAL SOLUTIONS
From well-known results on duality theory [16], we know that, for any
feasible primal dual pair the dual objective value is always a lower
bound to the primal objective value Duality gap is defined as
the difference between the optimal primal and optimal dual objective function
values, i.e.,
If a feasible primal dual pair satisfies then the pair is

optimal, and there is no duality gap. It may happen, as in several non-convex
programming problems, that duality gap exists; that is, the optimal primal and
dual objective function values are not equal, and therefore,
5.1 Duality Gap And Stopping Criterion
At every iteration of the LR algorithm, a primal and dual

feasible solutions can be obtained. The dual solution is directly obtained from
the dual maximization. We achieve a primal solution (not necessarily feasible)
when the dual function is evaluated; a feasibility search (based on heuristics)
can yield a feasible one. Let us define the complementarity and relative com-
plementarity gaps as
Through the iterative process, the values CG or RCG can be used as op-
timality indicators. The LR algorithm can be executed until DG is as small as
possible. If after a number of iterations it is found that CG = 0, then an opti-
mal solution has been found and there is no duality gap. If CG cannot be re-
duced below some value there are two possibilities: (i) a duality gap exists,
and the solution at hand is optimal; or (ii) the LR approach fails to further im-
prove the solution. Experience on solving UC problems [2, 15, 21] shows that
RCG can be reduced to “small” values, about 1-2%, especially for large
problems. These experimental results can be explained by theoretical results
[1, 16] that show that, as the number of separable components (generators) in
the dual function increases, the DG decreases.
5.2 Duality Gap and Cost Recovery: Do Dual Variables

Recover the Participants’ Cost ?
Using a simplified unit commitment problem, the authors have been able
to show that there is a direct relationship between duality gap and partici-
pants’ cost recovery when dual variables are used to set the market price when
the unit commitment is used as an optimization-based auctioning device [22].
The simplified (static, 1-hour) UC problem considers cost functions that con-
tain a linear and a constant start-up cost. The reserve constraint is dropped and
the operation of units is only constrained by the upper production limit; that
is,
Let be the uniform price set by a market to pay participants for

their power output (optimal solution to (24)). The profit (revenue minus
production cost) for each participant is given by
where has been introduced for convenience. The market

price has to be the minimum possible that leads to the recovery of partici-
pants’ cost; i.e., the profits are positive The dual problem to
(24) has the form of (7), and the dual function is given by
The dual function is piece-wise concave, and has slope-changing points at

every where as shown in Figure 2.
The optimal value of the dual function (26) can be found in a closed form;
it just suffices to determine the point at which the dual-function slope be-
comes zero or negative. Without loss of generality, let us assume that the units
are ordered so that An optimal dual solution is given
by where is the smallest index such that In [22], the
conditions under which duality gap exists for this problem are derived. In the
absence of duality gap, if the optimal dual variable (or any of its multiple val-
ues) is used to set the market price, then all the units will recover
their cost. Since for all scheduled units, then If a duality gap
exists, there is a cost that is not recovered and its magnitude equals the gap.
Table 1 shows the data from an illustrative example, where we consider

three different demand conditions and 290 MW). In the first
case, the dual problem has multiple dual solutions, but there is no duality gap;
and, therefore, as Table 3 shows, all the units’ costs are recovered. For this
case, there is only one optimal (global) primal solution. Notice that the profit
of the last scheduled unit is zero; hence, there is no other smaller value (that is
used as an uniform price) which can recover participants’ cost. This means
that in the absence of duality gap, the optimal dual variables from the cost
minimization problem are also the minimum possible prices.
In the second case, the dual problem has a unique solu-
tion. Since the optimal primal objective function is different from the optimal
dual objective function, however, there is duality gap. The cost not recovered
is equal to the duality gap; see Table 3. In the last case, there is also a duality
gap of $3 and, therefore, the same quantity is not recovered. In this case, there
are two multiple primal solution; 40 MW can be supplied either by unit 4 or 5.
Some other results in [22] show that for systems with a larger number of
units, the size of the duality gap dramatically decreases. This suggests that
small (or no) adjustments to the optimal dual variables need to be made in
order to use them to set the market price. Even though the relation between
duality gap and cost recovery has not been proved for the general UC problem
(1)-(4), the results in the next section suggest that the relation may hold.
Multiple solutions are more likely to exist for this type of combinatorial
problem, as happened in the last example. An approach that can identify all
the alternative optimal solutions is, obviously, the use of enumeration. This
would require the enumeration of possible unit commitment combina-

tions; from these combinations, we take the ones that can satisfy the demand
and solve them using a simple linear programming model. If different combi-
nations have the same objective function, then there are multiple (combinato-
rial) solutions. For a particular combination, multiple solutions (continuous
multiple solutions) may also exist. The identification of multiple solutions for
the PUC problem is a much more complex task.
6. RESULTS OF THE IP/CP FOR THE DUC PROBLEM

The LR procedure to solve the UC problem has been coded in C; test runs
have been performed on a Pentium 200Mhz computer under the LINUX op-
erative system. We solve the potential problem (14) using the primal-dual in-
terior-point method, described in Subsection 4.1.1; it is also coded in C. The
Newton’s system (20) is solved using an efficient sparse-matrix processing
and factorization techniques. We use the following four UC test systems: (i) a
26-unit system [23]; (ii) a 32-unit system [15]; (iii) two larger systems with 76
and 104 units, which are created by using the data from the systems in (i) and
(ii). The data for the last two systems is selected in such a way that large, me-
dium, and small units, as well as expensive, moderate, and cheap units coex-
ist. All tests are performed over a | T |= 24 hour optimization horizon.
6.1 Comparative Performance

Table 4 exhibits the number of iterations and time required by each of the
methods to solve the dual problem up to a point where the RCG is less than
2% for the 26-units case, and less than 1% for the other cases. Notice that the
IP/CP method requires far less iterations than the SG and PB methods, and its
solution time is competitive, especially for larger systems.
The results obtained with the SG and PB methods, shown in Table 4, are
achieved after extensive individual parameters tuning for each system;
whereas no parameter tuning is required for the IP/CP method. The box con-
straint is easily set up and left unchanged in all the tests performed. Figure 3
presents the evolution of the dual function for each method. The results shown
correspond to the 104-unit system; we observe a similar behaviour with the
other test systems. In the first four to five iterations, the IP/CP method is able
to reduce the RCG to about 3%.
6.2 Effect of the Box Constraint

We established and left fixed box constraints in all test systems to the
maximum incremental cost (evaluated at maximum power) of a system unit;
that is, It is likely that the incremental
cost at the solution point will never be larger than this value. If the incre-
mental cost in the system at a particular time reaches this value, unfeasibility
is likely to occur since the most expensive unit is being put at maximum gen-
eration. It has been observed that if the box constraint is set up to any other
large value, similar convergence occurs. This claim is supported by the results
shown in Table 5, where the box constraint has been changed to 75%, 200%
and even 300% of the base value Notice that for larger systems the set up
of the box constraint has less effect; i.e., in the UC-104 system, only 10 more
iterations are required for convergence if the set up is 300% of
6.3 Duality Gap and Cost Recovery

As shown in Section 5.2 for a simplified UC problem, the magnitude of
the duality gap is equal to the cost not recovered by the units when the opti-
mal dual variables are used to set the market price. For the PUC (1)-(4) the
same relationship seems to hold as shown by the results presented in Table 6.
For all the UC test cases, the dual variables to the power balance constraints
(obtained in the last iteration of the LR algorithm) are used to set the market
price The profit for each unit is and the
total non-recovered cost is As we can see, the
CG and NRC are very similar; this suggests that the relation between duality
gap and cost recovery may still hold for the UC problem. Small modifications
could be made to the dual variables in order to use them as the market price.
For instance, the revenues of participants whose profits are positive can be
proportionally (to the total positive profits) adjusted (reduced) so that the re-
duction is equal the duality gap and, therefore, is enough to recover all par-
ticipants costs. This is equivalent to define a non-uniform market price; each
supplier received a price computed from the dual variable, but scaled to pro-
portionally recover the duality gap. Modification of prices to achieve desir-
able revenue levels (revenue reconciliation) has been proposed for spot mar-
kets [24].
It is worthwhile to point out that DG is a characteristic of the model and its
parameters; it cannot be reduced by the use of one or other algorithm used to
solve the problem. An algorithm does better than the other if it reduces CG
closer to DG. Changes in the model can lead to different DG, but now dual
variables have different meanings.
7. CONCLUSIONS
Interior-point methods are among the best alternatives to solve a wide
class of very large linear and even non-linear programming problems. Re-
search on combinatorial and non-differentiable optimization has again put
interior-point methods as one of the promising approaches to most efficiently
solve such problems.
In this chapter, we proposed an interior-point/cutting-plane method to
solve the dual to UC problems. Comparisons to previously used approaches,
such as sub-gradient and penalty-bundle methods, demonstrate the advantages
of the IP/CP method. The IP/CP method has better convergence characteris-
tics and completely eliminates the need for parameter tuning.
The chapter also presents results that relate duality gap and cost recovery
if the dual variables are used as market prices. We showed, using a simplified
UC model, that when duality gap does not exist, the dual variable is the
minimum market price that can be used to set a uniform market price that re-
covers participants’ costs. When duality gap exists, the optimal dual variables
do not recover the costs; the un-recovered cost equals the magnitude of the
duality gap. The same relation seems to hold for the general UC problem, as
shown by the numerical results. This suggests that for large unit commitment
problems, small (or none) modification can be made to the dual variables in
order to be used as market prices. More research into this problem and on the
effective detection of multiple optimal solutions may lead to better under-
standing of the failures encountered in pioneering markets, as well as help
design better auctioning optimization models.
ACKNOWLEDGEMENTS
The first author gratefully acknowledges CONACyT and Instituto Tec-
nológico de Morelia in México, for providing financial support to pursue his
Ph.D. studies at University of Waterloo.
REFERENCES
1. D.P. Bertsekas, G.S. Lauer, N.R. Sandell, and T.A. Posberg. Optimal short-term schedul-
ing of large scale power systems. IEEE Trans. Autom. Control, AC-28(1): 1-11, 1983.
2. F. Zhuang and F.D. Galiana. Towards a more rigorous and practical unit commitment by
Lagrangian relaxation. IEEE Trans. Power Syst., 3(2): 763-773, 1988.
3. X. Guang, P.B. Luh, and H. Yan. An optimization-based method for unit commitment.
Electrical Power Energy Syst., 14(1): 9-17, 1992.
4. J.A. Muckstadt and S.A. Koenig. An application of Lagrangian relaxation to scheduling

power-generation systems. Oper. Res., 25(3): 387-403, 1977.
5. F. Pellegrino, A. Renaud, and T. Socroun. “Bundle and augmented Lagrangian methods
for short-term unit commitment.” In Power Systems Computation Conference Proc.,
pp. 730-739, 1996.
6. N. Jiménez and A.J. Conejo. Short-term hydro-thermal coordination by Lagrangian relaxa-
tion: solution to the dual problem. IEEE Trans. Power Syst., 14(1): 89-95, 1999.
7. P.B. Luh, D. Zhang, and R.N. Tomastik. An algorithm for solving the dual problem of
hydrothermal scheduling. In IEEE-PES, Winter Meeting paper PE-333-PWRS-0-12-1997,
New York, 1997.
8. C. Lemarechal and J Zowe. “A Condensed Introduction to Bundle Methods in Nonsmooth
Optimization.” In Algorithms for Continuous Optimization, pp. 357-382, ed. E. Spedicato.
Kluwer Academic Publishers, 1994.
9. O. du Merle, J.L. Goffin, C. Trouiller, and J.P. Vial. A Lagrangian relaxation of the ca-
pacitated multi-item lot sizing problem solved with an interior point cutting plane method.
Technical report, Faculty of Management, McGill University, 1997.
10. J.L. Goffin, J. Gondzio, R. Sarkissian and J.P. Vial. Solving nonlinear multi-commodity
flow problems by the analytic center cutting plane method. Math. Prog., (76): 131-154,
1996.
11. M. Madrigal and V.H. Quintana. “An Interior-point/Cutting-plane Algorithm to Solve
Unit Commitment Problems.” In IEEE-PES Power Industry Computer Applications Con-
ference Proc., pages 179-185, Santa Clara, California, 1999. To appear in IEEE Trans.
Power Syst.
12. R.B. Johnson, S.S. Oren, and A.J. Svodoba. Equity and efficiency of unit commitment in
competitive electricity markets. Technical Report PWP-039, POWER-series, The Univer-
sity of California Energy Institute, 1996.
13. S. Dekrajangpetch, G.B. Sheble, and A.J. Conejo. Auction implementation problems using
14. J.M. Jacobs. Artificial power markets and unintended consequences. IEEE Trans. Power
Syst, 12(2): 968-972, 1997.
15. C.L. Tseng. On Power Systems Generation Unit Commitment Problems. Ph.D. Thesis,
University of California, Berkeley, 1996.
16. D.P. Bertsekas. Nonlinear Programming. Athena Scientific, 1997.
17. M. Madrigal and V.H. Quintana. “Semi-definite Programming Relaxations for {0,1}-
Power Dispatch Problems.” In IEEE-PES, 1999 Summer Meeting Conference Proc., pp.
697-702, Edmonton, Alberta, Canada, 1999.
18. O. Bahn, J.L. Goffin, J.P. Vial, and O. Du Merle. Experimental behaviour of an interior
point cutting plane algorithm for convex programming: an application to geometric pro-
gramming. Discrete Appl. Math, 49: 3-23, 1994.
19. J.L. Goffin and J.P. Vial. Interior point methods for nondifferentiable optimization. Tech-
nical reports, Faculty of Management, McGill University, 1997.
20. M. Kojima, N. Megiddo, and S. Mizuno. A primal-dual infeasible-interior-point algorithm
for linear programming. Math. Prog, (61): 263-280, 1993.
21. S. Sen and D.P. Kothari. Optimal thermal generating unit commitment: a review. Electri-
cal Power Energy Syst., 20(7): 443-451, 1998.
22. M. Madrigal and V.H. Quintana. “Using Optimization Models and Techniques to Imple-
ment Electricity Auctions.” In IEEE-PES, 2000 Winter Meeting Conference Proc., Singa-
pore, 2000.
23. S.J. Wang, S.M. Shahidehpour, D.S. Kirschen, S. Mokhtari, and G.D. Irisarri. Short-term
generation scheduling with transmission and environmental constraints using augmented
24. F.C. Schweppe, M.C. Caramanis, R.D. Tabors and R.E. Bohn. Spot Pricing of Electricity.
Kluwer Academic Publishers, 1987.
Chapter 11
BUILDING AND EVALUATING

GENCO BIDDING STRATEGIES AND
UNIT COMMITMENT SCHEDULES
WITH GENETIC ALGORITHMS
Charles W. Richter Jr. and Gerald B. Sheblé

Iowa State University
Abstract: Far from being an artifact of the past, the unit commitment (UC) algorithm is
essential to making economical decisions in today’s competitive electricity in-
dustry. Increasing competition; decreasing obligations-to-serve; and enhanced
futures, forwards, and spot market trading in electricity and other related mar-
kets make the decision of which units to operate more complex than ever before.
Decentralized auction markets currently being implemented in countries like
Spain use UC-type models, which should encourage researchers to continue
working on finding better and faster solution techniques. UC schedules may be
developed for a generation company, a system operator, etc. The need for many
flavors of UC algorithms, each considering different inputs and objective func-
tions, is growing. Factors such as historical reliability of units should be consid-
ered in designing the UC algorithm. Although a particular schedule may result in
the lowest cost, or highest profit, it may depend on generators that have varying
availabilities. Traditionally, consumers had very reliable electricity whether they
needed it or not. Given a choice in a market-based electricity system, many con-
sumers might choose to pay for a slightly lower level of power availability if it
would result in sufficient savings. As the number of inputs and options grows in
UC problems, the genetic algorithm (GA) becomes an important tool for search-
ing the large solution space. GA times-to-solution often scale up linearly with
the number of units, or hours being considered. Another benefit of using the GA
to generate UC schedules is that an entire population of schedules is developed,
some of which may be well suited to situations that may arise quickly due to un-
expected contingencies.
1. INTRODUCTION
Electric generators and energy service companies around the world are
presently embracing competition. Generation companies and energy service
companies can negotiate profitable electricity contract prices bilaterally or via
auction markets, which will likely play a vital role in price discovery. Auction
market bidding strategies will be important for participants’ profitability and

survival. No longer guaranteed consumers nor a rate-of-return by a public
utilities commission (PUC), it is the responsibility of each generation com-
pany (GENCO) to operate units profitably. Good bidding strategies must al-
low the trader to negotiate profitable contracts, not only in the short-term, but
also in the mid- to long-term.
Providing a basic foundation for effective bidding strategies, the unit
commitment (UC) algorithm will remain one of the central tools in the com-
petitive electricity industry. The business of GENCO is ultimately to generate
electricity. The changes brought about by deregulation (e.g., increases in
competition, decreases in the utility's obligations-to-serve, enhanced futures,
forwards, and spot market trading in electricity and other related markets)
make the decision of which units to operate more complex than ever before.
The need for many varieties of UC algorithms is growing. UC schedules
may be developed and optimized for a particular generation company or for a
system operator. When choosing a UC schedule to the objective function may
consider a host of factors, such as historical availability of units or transmis-
sion lines, the credit risk of dealing with a particular consumer, etc. Although
a particular schedule may yield the lowest cost or greatest profit, considera-
tion of historical generator availability may reduce expected profits or in-
crease expected costs.
Historically, consumers were provided with (and paid for) very reliable
electricity regardless of their requirements. Given a choice in a market-based
electricity system, many consumers would choose to pay for a slightly lower
level of power availability if it resulted in substantial savings. Allowing the
trade of electricity with varying reliability levels in a market environment will
require improvements in allocation algorithms including UC algorithms.
As the number of choices and contract-types grow, the UC schedule
search space becomes large. The genetic algorithm (GA) is an important tool
in searching large solution spaces. GA times-to-solution often scale up line-
arly with the number of units or the number of hours being considered. An-
other strength that GAs have over other solution techniques is that entire
populations of solutions are developed in parallel. While GAs work well for
searching large spaces, there is no proof that they are converging to the opti-
mal solution. Because entire populations are being evolved during each run,
complex fitness evaluations may require much computation.
If a trader is participating in a market, he wants to take as much profit as
he can. The auction market may be viewed as a game for which various
strategies exist. How does the GENCO use its UC schedules and bidding
strategies in a manner that will result in long term profit? The procedure be-
gins by developing a working model of the competition. Since generating unit
characteristics don’t change fundamentally in the move from monopolistic
Using Genetic Algorithms in GENCO Strategies and Schedules 187
operation to competitive operation, information made public during regulated

operation remains valid and can help develop a basic competitor model. Addi-
tional information influencing the competitor is also public, e.g., price of fuel.
These basic competitor models can be used to simulate future (e.g., for T
hours into the future) trading. The models allow study of strategic behavior
that may affect how close the price is to the price achieved under perfect
competition. Any additional useful information should be considered in de-
termining the best forward price and demand forecasts. A price-based UC
algorithm uses these forecasts when searching for the weekly schedule that
maximizes expected profit, subject to some risk level. The schedule(s) should
be analyzed under various contingencies. Candidate UC schedules should be
used as an input to the bidding strategy development process. Figure 1 shows
an outline of this procedure in block diagram form.
This chapter gives a brief introduction to a basic market framework to

experimental economics. (The reader should be cautioned that this is a repre-
sentative market and is not identical to existing markets.) We discuss issues
related to determining competitor models, and we present methods used to
simulate the various markets (spot and forward). Our work investigates the
models’ impact on GENCO operations and planning on daily and weekly op-
erations. We highlight creation of bidding strategies for single and multi-
period auctions and discuss how forwards and futures markets provide a
means to manage risk. We present a price-based unit commitment formula-

tion, followed by a unit commitment genetic algorithm (UC-GA) formulation.
Additional criteria (e.g., considering contingencies such as non-availability of
generators, network problems, unforeseen market disturbances) may highlight
differences in UC schedules that have equivalent cost and profit. The authors
discuss each procedure to compare UC schedules with each other based on
their expected monetary value under likely scenarios or contingencies. Adap-
tive agents modelling buyers and sellers test the developed strategies by trad-
ing electricity in a simulated deregulated electricity market.
2. MARKETS AND EXPERIMENTAL ECONOMICS

2.1 A Basic Market Framework
Many countries and some regions of the United States (e.g., California,
Pennsylvania-Jersey-Maryland (PJM), New England, etc.) have deregulated.
There does not yet appear to be a standardized final market structure that will
be adopted for all areas, but each market developed should be an improve-
ment over those previously developed.
The framework described here is one in which electric energy is produced
by GENCOs, sold to energy service companies (ESCOs), and delivered on
wires owned by distribution companies (DISTCOs) and transmission compa-
nies (TRANSCOs). An entity such as the North American Electric Reliability
Council (NERC) sets the reliability standards. The contract prices are discov-
ered in an auction. Buyers and sellers of electricity make bids and offers that
are matched subject to the approval of the independent contract administrator
(ICA), who ensures that the contracts will result in a system operating safely
within limits. The ICA submits information to an independent system opera-
tor (ISO) for implementation. The ISO is responsible for physically control-
ling the system to maintain its security and reliability. Participants are pro-
vided with average contract price and volume, last accepted bid or offer, and
whether the participant’s previous bid was accepted or rejected.
Most of the market framework assumed here has been developed in pre-
vious publications [1,2,3,4]. It allows for cash (spot and forward), futures, and
planning markets as shown in Figure 2. The spot market is a market where
sellers and buyers negotiate (either bilaterally or multilaterally through an
exchange) a price for a certain number of MWh to be delivered in the near
future (e.g., 30 MWh – 10 MW from 1:00 p.m. to 4:00 p.m. tomorrow). The
buyer and seller must arrange a transmission path for the electrons. The day-
ahead market in California is an example of a spot market. The forward mar-
ket allows trading in a manner similar to the spot market but is further into the
future than is the spot market. In both forward and spot markets, the buyer
and seller intend to exchange the physical good (i.e., the electrical energy). In
contrast, the futures market is primarily financial, allowing traders to reduce
uncertainty by locking in a price for a commodity in some future month. The
provision for physical delivery exists, but since it is not normally intended,
other entities (e.g., TRANSCOs, ICAs, and ISOs) need not be informed of
futures trading. Buying a futures contract is akin to purchasing insurance. It
allows the traders to manage their risk by limiting potential losses or gains.
To ensure sufficient interest for price discovery, futures contracts are gener-
ally standardized such that it is not possible to tell one unit of the good from
another (e.g., 1 MWh of electricity of a certain quality, voltage level, etc.).
Although provisions for delivery exist, they are generally not convenient. The
trader ultimately cancels his position in the futures market either with a gain
or loss. The physical goods are then purchased on the spot market to meet
demand with the profit or loss having been locked-in via the futures contract.
The planning market aids in financing long-term projects like transmission
lines and power plants.
2.2 Experimental Economics

Economic theories are often based on several assumptions that tend to be
true when looking at society as a whole, but the theories may not necessarily
ring true for a particular agent or group of agents. The data used to substanti-
ate these theories often comes from indicators observed in the real economy.
It is hard to isolate the effect of a change in one input on one output in the real
world economy, because many things are changing simultaneously. Experi-
mental economics seeks to advance the theories of economics through ex-
perimentation. Pioneers such as Charles Plott and Vernon Smith popularized

the field through their work using the auction as an allocation mechanism. In
general, researchers study agents in a laboratory setting where their behavior
is monitored as they take part in experiments. Although initial experiments
were with humans, intelligent computerized agents are faster, cheaper, and
more consistent than human agents in the laboratory environment. Drawing
from the field of complex adaptive systems/artificial intelligence, computer-
ized agents can be given the ability to learn and to develop sensible GENCO
bidding strategies. (Figure 1 shows that simulating auction markets is one of
the first steps to validate and fine-tune the forward price and demand curves.)
The simulated markets (described later) are populated with computerized
trading agents (seeded with GENCOs and ESCOs models) to obtain forecasts
that account for the strategic behaviors of the competitors.
3. DETERMINING THE COMPETITOR MODEL

If the auction simulations are going to produce any results that are to be
used in actual GENCO operation, the models of the competing ESCOs and
GENCOs need to be reasonably accurate. Developing competitor models is
an involved process. Under the market framework assumed for this research,
the results of the auctions are public information, similar to the Australian
electric power market. As indicated in Figure 3, if it is available, a database of
auctions from previous periods contains bidding information and can be intel-
ligently mined to determine the general rules that the competitors are using.
This information can be combined with any additional information known
about the system or about the competitors to develop a fine-tuned model of
the competing computerized agents in the simulated markets [5].
The competitor models populate computerized agents participating in
simulated auctions for each period of interest. The competitor models may
represent a single aggregated GENCO competitor and a single aggregated
ESCO, or many independent GENCOs and ESCOs. Important information
(e.g., weather, unit outage information, status of transmission system, and
time of day) that will influence the bidding process is used to fine-tune the
competitor model as shown in Figure 4.
4. SIMULATING AN HOURLY AUCTION MARKET

The procedure described here is loosely based on the auctions seen at the
Chicago Board of Trade. To prevent an infinite loop, we select a maximum
number of cycles per round and maximum number of rounds per period. Buy-
ers and sellers determine how much they would like to buy and sell. They
determine their bids and offers and submit them to the auctioneer. All transac-
tions are subject to the approval of the ICA. Price discovery occurs when
there are a sufficient number of buy bids and sell offers to allow a predeter-
mined portion of the total participants to be satisfied with the resulting con-
tract. Submission of the bids marks the beginning of the bidding cycle. When
the auctioneer reports the results of the auction to the market participants the
cycle is complete. If, after the present cycle, the price has not been discov-
ered, the auctioneer reports that price discovery did not occur and asks for
new bids and offers. Market rules (funnel rules) force subsequent cycles
closer to price discovery by requiring the buyers to increase their bids and
sellers to decrease their offers. The cycles continue until price discovery oc-
curs, or until the auctioneer decides to match whatever valid matches exist
and continue to the next round or hour of bidding. Figure 5 shows this proc-
ess.
5. FORECASTING PRICES AND DEMANDS

For each auction market period, it is possible to determine a curve relat-
ing the price to the quantity demanded and a curve relating the price to the
quantity supplied. In the cost-based world of the past, the forward price curve
for an individual hourly market was simply a horizontal summing of GEN-
COs’ cost curves, which were public information. The forward price curve in
the competitive world will look fairly similar. In the price-based competitive
world, however, GENCOs no longer have to reveal their true cost. This in-
formation can be approximated from the forward and futures trading. Models
for generating unit output costs are relatively consistent from year-to-year; so
much of the information is publicly available from historical filings made
with regulators.
Some uncertainty comes with not knowing the consumer demand, which
is highly dependent on the weather. We can use neural networks and statisti-
cal techniques to obtain demand forecasts. These price and demand forecasts
are given to the unit commitment scheduler, which attempts to find the
schedule of generating units that maximizes utility. For simplification, the UC
algorithm described in the next section uses a single expected price and quan-
tity for each period. A UC schedule is typically developed for each hour of
the following week. Each period the latest forecasts and market price are
given to the algorithm and updated schedule is developed.
6. PRICE-BASED UNIT COMMITMENT

In an environment with vertically integrated monopolies, the scheduling
of generating units to be on, off, or in stand-by/banking mode is done in a
manner that minimizes costs. Consideration must be given to factors like
varying fuel costs, start-up and shut-down parameters/constraints of each
power plant, and crew constraints. In order to determine the cost associated
with a given schedule, an economic dispatch calculation (EDC), in which all
unconstrained operating units are set so that their marginal costs are equal,
must be performed for each hour under consideration. One possible way to
determine the optimal schedule is to do an exhaustive search. Exhaustively
considering all possible ways that units can be switched on or off for a small
system can be done, but for a reasonably sized system the amount of time it
would take is too long. Solving the problem generally involves using methods
like Lagrangian relaxation, dynamic programming, genetic algorithms, or
other methods using heuristic search techniques. Many references for the tra-
ditional UC can be found in Sheblé and Fahd [6] and in Wood and Wollen-
berg [7].
Without competition, demand forecasts of total consumer demand in the
service territory advised power system operators of the amount of power that
needed to be generated. Under competition, bilateral spot and forward con-
tracts make part of the total demand known a priori. Predicting the GENCO’s
share of the remaining demand may be difficult, however, since it is depend-
ent on its prices compare to that of other suppliers, and the condition of the
transmission network, among other factors. Since the number of units

switched on or in banking mode affects the average cost of production, the
UC schedule indirectly affects the price, making it an essential input to any
successful bidding strategy.
Utilities operating in regulated monopoly fashion have an obligation to
serve their customers within a designated service territory. This translates into
a demand constraint that ensures all demand is met. For the UC problem, this
might mean switching on an additional unit to meet a remaining MW or two.
Without an obligation to serve, the competitive GENCO can consider a
schedule that produces less than the predicted demand. It can allow other
suppliers to provide that one or two MWs that might have increased average
costs (they might not have secured that contract for which they would have
had to compete).
Demand forecasts and expected market prices are an important input to the
profit-based UC algorithm; they are used to determine the expected revenue,
which affects the expected profit. If a GENCO comes up with two potential
UC schedules each having different expected costs and different expected
profits, it should take the one that provides for the greatest expected benefit to
cost ratio, which will not necessarily cost least. Since price and demand are so
important in determining the optimal UC schedule, price prediction and de-
mand forecasts become crucial.
Mathematically the traditional cost-based UC problem has been formu-
lated as follows [8]:
Subject to the following constraints:
The cost based UC algorithm for the regulated monopolist has been well
researched. Recently, research has included the use of markets in developing
UC schedules. Takriti, Krasenbrink, and Wu [9] present a good description
and a stochastic solution of the UC problem that considers spot markets.
Their research differs from that presented here in that it minimizes costs
rather than maximizes profits. Here, we redefine the UC problem for the
GENCO operating in the competitive environment by changing the demand
constraint from an equality to less than or equal (assume buyers are required
to purchase reserves per contract) and changing the objective function from
cost minimization to profit maximization:
Subject to:
Where individual terms are defined as follows:

(Capacity limits)
(Ramp rate limits)
up-/down-time status of unit n at time period
power generation of unit n during time period t

load level in time period t
forecasted demand w/ reserves for period t
forecasted price for period t
system reserve requirements in time period t
production cost of unit n in time period t
start-up cost for unit n, time period t
shut-down cost for unit n, time period t
maintenance cost for unit n, time period t
number of units
number of time periods
generation low limit of unit n
generation high limit of unit n
maximum contribution to reserve for unit n
Since a GENCO is not obligated to serve under competition, it may
choose to generate less than the total consumer demand. Therefore, maximiz-
ing the profit is not the same as minimizing the cost. Choosing the amount of
demand to serve allows the GENCO more flexibility in developing the UC
schedules. Because prices fluctuate with supply and demand, engineers often
assume that making the load curve level minimizes cost and should be a rea-
sonable operating strategy. When maximizing profit, the GENCO may find
that during certain conditions greater profit may be possible under a non-level
load curve. If revenue increases more than the cost does, the profit will in-
crease. Figure 6 shows a block diagram of the UC solution process.
EDC is an important part of UC. Formerly used to minimize costs, it is
necessary to redefine EDC for price-based operation. Where the cost-
minimizing EDC ignored transition and fixed costs to adjust the power level
of the units until they each had the same incremental cost
our new EDC attempts to set equal to a pseudo price (i.e., pro-
duce until the marginal cost equal the price). This pseudo price accounts for
transition and fixed costs as shown in the following formula:
Transition costs include start-up, shut-down, and banking costs. Fixed

costs (present for each hour that the unit is on), are represented by the con-
stant term in the typical quadratic cost curve approximation.
During each period, the fixed and transition costs are accounted for by
adding an average value to the incremental cost. Note that this is one method
of allocating the transition and fixed costs, but there are many others that
could be used. For instance, if GENCO A’s generators are able to produce
electricity far less expensively than the competing GENCOs during the night,
but don’t have that advantage during the daytime, GENCO A could shift
some of its daytime costs to be recovered through bids from the night-time
periods.
7. A GENETIC-BASED UC ALGORITHM
7.1 The Basics of Genetic Algorithms

A genetic algorithm is a search algorithm often used in non-linear discrete
optimization problems. A population of data structures appropriate for the
problem solution is initialized randomly, evolves over time, and finds a suit-
able answer (or answers) to the problem. The data structures often consist of
strings of binary numbers, which are mapped onto the solution space for
evaluation. Each solution (often termed a creature) is assigned a fitness - a
heuristic measure of its quality. During the evolutionary process, those crea-
tures having higher fitness are favored in the parent selection process and are
allowed to procreate. The parent selection is essentially a random selection
with a fitness bias. The parent selection method determines the type of fitness
bias. Following the parent selection process, the processes of crossover and
mutation are utilized and new creatures are developed which ideally explore a
different area of the solution space. These new creatures replace less fit crea-
tures from the existing population.
7.2 GA Formulation for Price-Based UC

The algorithm presented here solves the UC problem for the profit-
maximizing GENCO operating in the competitive environment [10]. Various
GA formulations for finding optimal cost-based UC schedules have been pro-
posed by researchers [11, 12]. The profit or price-based algorithm presented
here is a modification of a genetic-based UC algorithm for the cost-
minimizing monopolist that was described by Maifeld and Sheblé [13]. The
fitness function now rewards schedules that maximize profit. The intelligent
mutation operators are preserved in their original form. The schedule format
is the same. Figure 7 shows the algorithm.
The algorithm first reads in the contract demand and prices, the forecast of
remaining demand, and forecasted spot prices. During the initialization step, a
population of UC schedules is randomly initialized (see Figure 8). For each
UC period of each member of the population, EDC is called to set the units’
generation levels. The cost of each schedule is determined from the paramet-
ric generator data and the demand and price data read at the beginning of the
program. Next, the fitness (i.e., the profit) of each schedule in the population
is calculated. “Done?” checks to see whether the algorithm has either cycled
through for the maximum number of generations allowed, or whether other
stopping criteria have been met. If the algorithm is done, the results are writ-
ten to a file; if it is not done, the algorithm proceeds to the reproduction proc-
ess.
During reproduction, the algorithm creates new schedules. The first

step of reproduction is to select parents from the population. After selecting
parents, children are created using two-point crossover as shown in Figure 9.
Following crossover, standard mutation is applied. Standard mutation in-
volves turning a randomly selected unit on or off within a given schedule.
An important feature of this UC-GA is that it spends as little time as pos-

sible doing EDC. After standard mutation, EDC is called to update the profit
only for the mutated hour(s). An hourly profit number is maintained and
stored during the reproduction process, which dramatically reduces the
amount of time required to calculate the profit over what it would be if EDC
had to work from scratch at each fitness evaluation. In addition to the stan-
dard mutation, the algorithm uses two “intelligent” mutation operators that
work by recognizing that, because of transition costs and minimum up-time
and down-time constraints, 101 or 010 combinations are undesirable. The first
of these operators purges these undesirable combinations by randomly chang-

ing 1s to 0s or vice versa. The second of these intelligent mutation operators
purges the undesirable combination by changing 1 to 0 or 0 to 1 based on
which of these is more helpful to the profit objective.
7.3 Price-Based UC-GA Results

We tested the UC-GA on some small systems and the results were com-
pared to the solutions found by exhaustive search. In all of the trials for which
known optima existed, the GA was successful in locating the optimal UC
schedule. Figure 10 shows the costs and average costs (without transition
costs) of the 10 generators, as well as the hourly price and load forecasts for
the 48 hours. We chose the data so that the optimal solution was known a pri-
ori. The dashed line in the load forecast represents the maximum output of the
10 units. In addition to the information shown in the figure, the UC-GA pro-
gram requires the start-up and shut-down costs, the minimum up and down
times, and the cost to bank each generator. The generators are modelled with
a quadratic cost curve (e.g., , where P is the power level of
the unit). Though the GA took 730 seconds to find a population of solutions
containing the best possible solution for a 10-unit, 48-hour case (see Figure
11).
Prior to initiating the UC-GA, the control parameters shown in Table 1
are specified, including the “Number of Units” and Number of Hours” to con-
sider. “Popsize” is the size of the GA population. Increasing the population
size increases the amount of space searched during each generation of the
GA. This must be balanced with the consideration that execution time for
each generation of the algorithm varies approximately linearly with the size
of the popsize. The number of “Generations” indicates the number of times
the GA will undergo the selection and reproduction. “System reserve” is the
percentage of reserves that the buyer must maintain for each contract. “Chil-
dren per generation” tells us how much of the population will be replaced
each generation. Changing this can affect the convergence rate. If there are
multiple optima, faster convergence may trap the GA in local sub-optimal
solution. “UC schedules to keep” indicates the number of evolved schedules
to write to file. There is also a “random number seed” that may be set be-
tween 0 and 1.
The algorithm accurately calculates the cost of schedules in which mini-

mum up- and down-time constraints appear to be violated by considering a
zero surrounded by ones to be a banked unit, and a one surrounded by zeros is
ignored (unit remains off) if it violates the minimum up constraint.
An advantage of using the GA is that its solution time scales up only line-
arly with the number of hours and units, while dynamic programming quickly
becomes too computationally expensive to solve. The existence of additional
valid solutions, which may be only slightly sub-optimal in terms of profit, is

another main advantage of using the GA. It gives the system operator the
flexibility to choose the best schedule from a group of schedules to accom-
modate things like forced maintenance.
8. COMPARING AND SELECTING UC SCHEDULES

Even though a large percentage of the UC schedules encountered by the
genetic algorithm (or other search technique) may satisfy (within some small
tolerance) the primary objective function, they may not be equal. A set of UC
schedules may initially have indistinguishable costs or profits, but when we
consider additional criteria, differences between the schedules may be re-
vealed. A few examples of additional criteria might be:
Impact of units or transmission system availability
Ability to respond to spot-market price fluctuations
Schedule’s profitability during network contingencies
Ability to accommodate maintenance activities
A unit that is unavailable likely reduces GENCO profitability. A unit may
be unavailable due to a unit outage (scheduled maintenance or forced), the
transmission line connecting it to the load may be congested, or other unfore-
seen circumstances may exist that prevent the GENCO from selling electric-
ity. The amount of time that a generating unit is forced out of the market is
often unpredictable and variable. If a unit undergoes a forced outage or is
taken off-line for other reasons, costs such as shut-down (and subsequent
start-up) must be recovered. These unanticipated costs may have a large im-
pact on the profitability of the UC schedule. Although reserves may mitigate
the consequences of a single unit outage, a possibility exists that many units
may be inoperable or at reduced capacity simultaneously, since independent
unit outages may occur, as well as contingencies promoting system-wide dis-
turbances. Therefore, the possibility that more than one unit is forced off-line
in a given period of time must be considered. We can use historical availabil-
ity of generating units and of the transmission system itself to differentiate
between the UC schedules under consideration.
The spot market prices may undergo short-term unanticipated changes
that could be quite profitable for the GENCO having a UC schedule allowing
the amount of power to be increased or decreased easily (i.e., without turning
on or off additional units).
A schedule’s performance under various contingencies may distinguish it
from others. While searching for the optimal UC schedule, certain network
conditions, unit availabilities, load and price forecasts may have been as-
sumed. Contingencies will impact some of the candidate schedules more than
others.
The ability to schedule maintenance activities may be a characteristic of
schedules that distinguishes them from each other. Perhaps two schedules
result in roughly the same amount of profit, but one of the schedules allows
for preventive maintenance activities on some key units.
9. TESTING THE UC SCHEDULE IN SIMULATED

COMPETITION
9.1 Intelligent Bidding Strategies
Once the tentative UC schedule(s) is developed, the GENCO should have
bidding strategies that guide it in placing bids and in taking market positions.
These strategies might be designed to limit risk, maximize profit, a combina-
tion of both, or something entirely different.
Intelligent strategies can detect various market scenarios and respond ap-
propriately (i.e., profitably). Because the number of scenarios that the agent
might encounter is extremely large, discrete, and non-linear, finding an opti-
mal bidding strategy is a problem naturally suited for genetic algorithms. Ge-
netic algorithms and genetic programming have been used to evolve bidding
strategies that maximize profit for individual hourly spot markets [2, 14].
Representing the agent’s strategy in a easy-to-use and easy-to-evolve format
is sometimes the most difficult aspect of GA application. Fixed binary strings
may be the simplest method of encoding a problem. Finite state ma-
chines/automata offer a more powerful means of encoding the strategies, but
might be rather large to encode an independent response for each scenario
encountered. The authors use GP-Automata to circumvent this problem. We

test the evolved bidding strategies are tested by repeatedly using them in
simulated trades against competing strategies. Before we present the basics of
GP-Automata, here is a brief introduction of genetic programming.
9.2 The Basics of Genetic Programming

A sub-class of genetic algorithms, genetic programming (GP) is a new
discipline attributed to John Koza [15, 16]. Although not for the electricity
market, Andrews and Prager published research indicating that GP works for
representing simple double auction market strategies [19]. The evolving data
structures in GP are “parse trees” which allow complex relations to be de-
scribed. Genetic programs (GPs) contain nodes and branches, with branches
connecting the nodes. Nodes can be either operational nodes, having argu-
ments and performing operations involving those arguments, or terminal
nodes. Figure 12 provides examples of randomly generated GP-trees. The tree
on the right side of the figure would return the average of five plus the aver-
age buy bid from the previous round of bidding. The tree on the left would
multiply 10 by the absolute value of the high buy bid, and then (inefficiently)
take the absolute value of the result.
The designer specifies the set of valid operators and terminals suitable to
the problem being investigated. For instance, in developing bidding strategies,
suitable operators and terminals might be those described in Table 2. In de-
signing GPs for the GP-Automata, it is desirable to give the trees an opportu-
nity to return numbers in the range of competitive bids.
Valid GP trees are initialized randomly and then evolved in a standard

genetic algorithm (as described in the previous section) with the following
modifications. The crossover of two parents involves randomly selecting a
node from each parent and swapping the sub-trees rooted at those nodes. Mu-
tation involves randomly selecting a node in the candidate child and throwing
away its sub-tree. In its place a new sub-tree is generated randomly.
9.3 The Basics of GP-Automata

GP-Automata are a combination of finite state automata and GP. They
were first described as such by Ashlock [17] and were used by Ashlock and
Richter [18]. The typical finite state automaton specifies an action and “next
state” transition for a given input or inputs. With only one or two binary in-
puts to work with, it can be fairly simple to develop a finite state diagram to
cover the possible input/output relations. When the number of inputs is large
the task is much harder. The number of transitions needed to cover all possi-
ble combinations of inputs grows exponentially (e.g., 10 inputs, each having
five possible values would require transitions). This is where genetic
programming comes in. The GP-trees perform bandwidth compression for the
GP-Automata by selecting which inputs to consider and performing computa-
tions involving these inputs.
The four-state GP-Automaton in Figure 12 begins by bidding the number
in the “Initial Action” field (in this case, 24). The “Initial State” tells us which
state is used next (in this case, 2). Coupled with each of these states is a GP-
tree, termed a “Decider.” When executed, the decider returns a value in the
valid range of bidding. Following the decider evaluation, one of the following
two things will happen: (a) if the value is even after truncation, the action
listed under “IF EVEN” is taken and the current state becomes the one listed
under the “IF EVEN” next state; or (b) if the returned value is odd after trun-
cation, then the action and next state listed under “IF ODD” is used. The “Ac-
tion” is the number listed in the action field of the automaton, with two
exceptions. The first exception is the “U” that indicates that the value re-
turned by the decider should be taken directly as the action. The second ex-
ception is a “*” indicating further computation is necessary – the GP-
automaton immediately moves to the next state. This gives rise to the possi-
bility of complex (multi-state) computation. To prevent infinite loops, after an
externally specified maximum number of *s have been processed, an action is
selected at random from the valid actions.
A population of GP-Automata bidding strategies evolves in a GA. Fitness

is dependent on the goal of the strategies. Children/offspring are produced
using crossover and mutation. Crossover for the GP-Automata involves se-
lecting (with a uniform probability) a crossover point ranging from zero to the
number of states. Then parentl’s states from zero to the crossover point are
copied to childl and parent2’s states are copied to child2. Following the
crossover point, childl gets parent2’s state information and child2 gets par-
entl’s state information (including the associated decider). Before replacing
less fit members of the population, each child is subjected to mutation. Muta-
tion may be standard GA mutation that selects a state or action at random and
replaces it with a valid entry. Other forms of mutation are acceptable as well.
The goal is to introduce new combinations of genetic material into the
population.
9.4 Auction Bidding with the GP-Automata Strategies

While they are evolving in the genetic algorithm, each GP-Automaton’s
strategy competes against several other GP-Automata every generation (see
Figure 14). This helps to ensure that the resulting GP-Automata strategies will
be robust. Based on the competition, some fitness measure (e.g., the expected
amount of profit that results from using the strategies to win contracts) is as-
signed to each GP-Automaton. Through the natural selection process, the
population evolves, finding strategies that are more likely to achieve good
fitness.
State information is supplied to the GP-Automata via the terminals. The

GP-trees use both the state information accessed by the terminals as well as
constants in the valid bidding range. Bids are taken from the action cell of the
automata, except in the cases where the action is listed as a * or a U, as de-
scribed previously. The bids are submitted, along with the bids from the com-
peting sellers and buyers, to the auctioneer for evaluation. The bids and offers
are matched and a would-be price is reported, completing one cycle of the
auction. The cycles continue until price discovery occurs or until some maxi-
mum number of cycles (maxcycles) has passed. There is a maxcycles parame-
ter, which is selected uniformly over a range to prevent the strategies from
falling in a local optima where the strategies only work well when the number
of cycles never changes over the trials in a given generation.
An evolved GP-Automaton contains trained rules, which may be quite
complex. These rules may be used directly in a real auction just as they were
used in the simulated auction during evolution.
10. SUMMARY
The GENCO’s business is still one of generating electricity, and it must
ultimately determine a profitable schedule to operate its generating units.
Thus, the UC algorithm will continue to be an important tool in the evolving
industry. While the ICA may minimize total costs when matching bids, the
GENCO must maximize its profit. GENCOs must make decisions based on
market projections. Customers and demand may no longer be guaranteed, but
bilateral and multilateral forward contracts will ensure that the GENCO
knows much of its load ahead of time. Accurate forecasts of the quantity de-
manded and prices are crucial when solving the UC problem. If a GENCO’s
market projections are incorrect, the UC schedule may no longer be optimal.
Flexible schedules and bidding strategies are important. With huge potential
losses/profit at stake, UC schedules should be tested before use. Intelligent
agents with evolvable strategies provide realistic competitive behavior and
are thus ideal for robustness testing. GAs have demonstrated an ability to
learn and to build and adapt UC and bidding strategies for given scenarios.
REFERENCES
1. J. Kumar and G. Shebté. “Framework for Energy Brokerage System with Reserve Margin
and Transmission Losses.” In Proc. 1996 IEEE/PES Winter Meeting, 96 WM 190-9
PWRS, NY: IEEE.
2. C. Richter and G. Sheblé. “Genetic Algorithm Evolution of Utility Bidding Strategies for
the Competitive Marketplace.” In Proc. 1997 IEEE/PES Summer Meeting, PE-752-
PWRS-1-05-1997. New York: IEEE.
3. G. Sheblé. “Electric energy in a fully evolved marketplace.” Paper presented at the 1994
North American Power Symposium, Kansas State University, KS, 1994.
4. G. Sheblé. “Priced based operation in an auction market structure.” Paper presented at the
1996 IEEE/PES Winter Meeting. Baltimore, MD, 1996.
5. C. Richter. Profiting from Competition: Financial Tools for Competitive Electric Genera-
tion Companies. Ph.D. dissertation, Iowa State University, Ames, IA, 1998.
6. G. Sheblé and G. Fahd. Unit commitment literature synopsis. IEEE Trans. Power Syst.,
9(1): 128-135, 1994.
7. A. Wood and B. Wollenberg. Power Generation, Operation, and Control. New York:
John Wiley & Sons, 1996.
8. G. Sheblé. Unit Commitment for Operations. Ph.D. Dissertation, Virginia Polytechnic

Institute and State University, 1985.
9. S. Takriti, B. Krasenbrink, and L.S.-Y. Wu. “Incorporating Fuel Constraints and Electric-
ity Spot Prices into the Stochastic Unit Commitment Problem,” IBM Research Report: RC
21066, Mathematical Sciences Department, T.J. Watson Research Center, Yorktown
Heights, NY, 1997.
10. C. Richter and G. Sheblé. “A Price-Based Unit Commitment GA for Uncertain Price and
Demand Forecasts.” In Proc. 1998 North American Power Symposium, 1998.
11. S. Kondragunta. Genetic algorithm unit commitment program, M.S. Thesis, Iowa State
University, Ames, IA, 1997.
12. S. A. Kazarlis, A. G. Bakirtzis, and V. Petridis. “A Genetic Algorithm Solution to the Unit
Commitment Problem.” In Proc. 1995 IEEE/PES Winter Meeting 152-9 PWRS, New
York: IEEE, 1995.
13. T. Maifeld and G. Sheblé. Genetic-based unit commitment. IEEE Trans. Power Syst.,
11(3): 1359, 1996.
14. C. Richter, D. Ashlock, and G. Sheblé. “Effects of Tree Size and State Number on GP-
Automata Bidding Strategies.” In Proc. 1998 Conference on Genetic Programming,
Denver, CO: Morgan Kaufmann, 1998.
15. J. Koza. Genetic Programming. Cambridge, Massachusetts: The MIT Press, 1992.
16. J. Koza. Genetic Programming II. Cambridge, Massachusetts: The MIT Press,1994.
17. D. Ashlock. “GP-automata for Dividing the Dollar.” Mathematics Department, Iowa State
University, Ames, IA 1995.
18. D. Ashlock and C. Richter. “The Effects of Splitting Populations on Bidding Strategies.”
In Proc. 1997 Conference on Genetic Programming, Denver, CO: Morgan Kaufmann,
1997.
19. M. Andrews and R. Prager. “Genetic programming for the acquisition of double auction
market strategies.” In Advances in Genetic Programming, K. Kinnear Jr., ed. Cambridge,
MA: The MIT Press, 1994.
Chapter 12
AN EQUIVALENCING TECHNIQUE FOR

SOLVING THE LARGE-SCALE THERMAL
UNIT COMMITMENT PROBLEM
Subir Sen
Power Grid Corporation of India, Ltd.
D.P. Kothari
Indian Institute of Technology at Delhi
Abstract: This chapter presents a new efficient solution approach for solving the unit
commitment schedule of thermal generation units of a realistic large scale power
system. We base the approach on cardinality reduction by the generator equiva-
lencing concept. This concept reduces the number of units in the large-scale
power system to the lowest possible number based on the units’ fuel/generation
cost and other physical characteristics, such as minimum up and down time, etc.
with units having similar (almost the same) characteristics form one group. The
reduced system consists of only each group of representative units and is first
solved by the modified dynamic programming technique (one of the new solu-
tion methods developed by the authors). Another option is to use any of the stan-
dard unit commitment solution techniques. We obtain the overall solution to the
original unit commitment problem of the entire system by un-crunching the
solved reduced system based on certain rules. This chapter also presents test re-
sults for real-life systems of up to 79 units and comparisons with results obtained
using Lagrangian relaxation and truncated dynamic programming (DP-TC).
1. INTRODUCTION
One of the most important problems in operational scheduling of electrical
power generation is the unit commitment (UC) problem. It involves determin-
ing the start-up and shut-down schedules of thermal units to be used to meet
forecasted demand over a future short term (24-168 hour) period. The objective
is to minimize total production cost while observing a large set of operating
constraints. The unit commitment problem (UCP) is a complex mathematical
optimisation problem having both integer and continuous variables. One ob-
tains the exact solution to the problem by complete enumeration, which cannot
be applied to realistic power systems due to its excessive computation time re-
quirements [1,2]. In solving the unit commitment problem of a large system,
the main cause of difficulty is the involvement of large number of units for
commitment. The problem cannot be solved easily if all units are involved in
the search for the optimal solution, since computational facilities could be ex-
hausted. Research efforts have concentrated, therefore, on efficient, sub-optimal
UC algorithms which can be applied to realistic power systems and have rea-
sonable storage and computation time requirements. The basic UC methods
reported in the literature can broadly be classified in six categories [3]:
Priority list
Dynamic programming
Lagrangian relaxation
Augmented Lagrangian relaxation
Branch-and-Bound
Benders decomposition
Since improved UC schedules may save the electric utilities substantial re-
sources per year in production costs, the search for closer to optimal commit-
ment schedules continues. Recent efforts include application of simulated an-
nealing, expert systems, Hopfield neural networks and genetic algorithms to
solve the UCP. References [4,5] give a survey of various approaches and their
merits and demerits in this field. Some of these methods achieved a reduction
of the computation requirement for large power systems. Researchers have yet
to obtain an optimal solution to the problem for such systems.
There have been some past attempts in other areas such as coal modeling
[6] to reduce a large-scale system to a smaller system. This chapter proposes a
new, efficient solution approach to the UCP of a large-scale power system. The
approach is based on cardinality reduction by generator equivalencing (hereaf-
ter called “equivalencing”), which reduces the number of units in the large-
scale power system to the lowest possible number according to their similar
fuel/generation cost characteristics and minimum up- and down-time character-
istics. We first solve the reduced system using a modified dynamic program-
ming technique [7] and then obtain an overall solution to the original unit
commitment problem of the entire system by un-crunching the reduced solved
system and using certain rules. The Appendix explains the modified dynamic
programming (MDP) technique.
An Equivalencing Method 213
2. NOTATION
N : number of thermal generation units
T : total scheduling period
load demand, in MW
power generation by nth unit, in MW
minimum generation capacity limit for n-th unit
maximum generation capacity limit for n-th unit
cost of power generation by nth unit, in Rs/hour
start-up and shut-down cost for nth unit, in Rs/hour
system reserve requirement in time period t
up-/down-time status of nth unit
unit on; unit off
duration of unit n on and off, in hour
minimum up- and down-time for nth unit, in hour
number of units in group i
number of groups in the system
power output of the equivalent system
minimum and maximum generation limit of group g
3. DESCRIPTION OF THE UNIT COMMITMENT

PROBLEM
The objective of the UC problem is to minimize the total production cost
over the scheduling horizon. The total production cost consists of: fuel costs,
start-up costs, and shut-down costs.
We calculate fuel costs by using unit heat rate and fuel price information.
We express the start-up cost as a function of the number of hours the unit has
been down (exponential when cooling and linear when banking). The shut-
down cost is given by fixed amount for each unit per shut down.
The following must be satisfied during the optimization process:
(a) system power balance (demand plus loss plus exports)
(b) system reserve requirement
(c) unit initial condition
(d) unit high and low MW limits (economic, operating)
(e) unit minimum up time

(f) unit minimum down time
(g) unit status restrictions (must-run, fixed-MW, unavailable, available)
(h) unit rate limits
(i) unit start up ramps
(j) plant crew constraints
Constraints (a) and (b) concern all the units of the system and are called
system, or coupling, constraints. For multi-area unit commitment, the system
constraints must be modified to take into account the interchange schedules and
the tie-line limitations. In general, the system constraints must take into account
possible transmission bottlenecks in the allocation of the demand and the re-
serves to the generating units.
Constraints (c) through (i) concern individual units and are called local
constraints. Plant crew constraints can also be classified along with local con-
straints, but they involve all the units in a plant.
4. THE EQUIVALENCING METHOD

The “equivalencing” method for solving large scale thermal unit commit-
ment problem consider the particular reference to units generation cost (input-
output characteristic), minimum up and down time, and ramp rate criterion. A
large-scale power system consists of a large number of generation units, in
which all the units are not of same size and similar characteristics. Yet there are
many clusters of units throughout the entire power system which have almost
similar cost coefficients and other physical characteristics. Therefore, units in
the large scale system are re-grouped/ partitioned into various groups based on
their fuel cost, having cost coefficients of almost similar value (maximum 1%
variation), and other physical characteristics, such as same capacity, minimum
up and down time, ramp rate, etc. so that identical/similar characteristic units
form one group. Each such group is then represented by any one unit of this
group and is called a representative unit. Consequently, an equivalent smaller
system of units consisting only of representative units is generated. The basic
concept of this method is that the large-scale power system is represented by an
equivalent smaller system with a smaller number of generation units, such that
the unit commitment problem of the equivalent system would be easier to han-
dle and solve than that of the original large size system. Once the solution of
the unit commitment problem of the equivalent small system is obtained, then
the solution to the original unit commitment problem can be determined
accordingly by un-crunching the equivalent problem solution.
5. PROBLEM FORMULATION
The function to be minimized for the unit commitment problem can be ex-
pressed in mathematical form as follows:
subject to the following major constraints:

i) demand constraint:
ii) capacity constraint:
iii) unit’s minimum up- and down-time constraints:
iv) unit’s generation capacity constraint:
Formulate the large-scale unit commitment problem by using the “equiva-

lencing” method as given below:
v) Generate the equivalent system (based on the generator cost and other physi-
cal characteristics) with the lowest possible number of representative units
only.
vi) Determine the number of groups in the system and the number of units in a
group, such that identical units form one group. The total number of units
in the system would be
vii) Represent each group by one representative unit to form equivalent system.
The total output of the equivalent system would be:
viii) Maximum and minimum capacity limit of each group (i.e., representative
unit) would be
ix) Minimum up and down time, ramp rate, etc. of the representative unit of a
group is set as any individual unit’s characteristics of that group.
5.1 Solution Technique

The large-scale unit commitment problem is solved by performing the
commitment (0-1 status of each unit) of the equivalent system using the modi-
fied dynamic programming technique [7] developed by the authors. It can be
solved, however, by any other standard unit commitment solution technique. If,
during the solution of the equivalent system, the minimum output of a particu-
lar representative unit of a group is less than the equivalent minimum run level
of that group, perform the sub-unit commitment of the individual units of that
particular group. Once the reduced system is solved, the units in each group in
the equivalent system are treated according to the status of its representative
unit, based on the following strategy:
If the representative unit of group i at interval t is “OFF,” then all units
of the group should be off.
If the representative unit of group i at interval t is “ON” and operated at
its equivalent minimum/maximum output, all the units in the group
should be operated at its individual minimum/maximum output.
If the representative unit of group i at interval t is “ON” and operated at
some percentage of its maximum/minimum output, carry out further
commitment scheduling within the loop among the units in that group.
Finally, obtain the total cost over the commitment period using Equation (1).
5.2 The New Algorithm

Step 1: Read input data, i.e., unit characteristics, load demand profile, etc.
Step 2: Categorize the units having similar/identical fuel costs, minimum up
and down time, and ramp rate characteristics into different groups.
Step 3: Generate the equivalent smaller system of the representative units.
Step 4: Perform the commitment schedule of the equivalent system by modi-
fied dynamic programming or by any standard technique with certain
rules, i.e., if during the solution of the equivalent system, the minimum
output of a particular representative unit of a group comes out less than
the equivalent minimum run level of that group, perform the sub-unit
commitment of the individual units of the group.
Step 5: Generate the unit commitment schedule of the original system using the
rules as specified below:
if representative unit is “OFF,” all units in that group would be off.
if representative unit is “ON,” and generate up to its equivalent
maximum or minimum capacity output, all units in that group would
generate their minimum/ maximum output.
if the representative unit is “ON” and generate a percentage of its
maximum/minimum capacity output, schedule the units in that
group using modified dynamic programming to generate the optimal
scheduling of the units in that group.
Step 6: Calculate the total cost for scheduling the normal way by Equation (1).
Figure 1 presents the outline of the new algorithm for solving large scale
unit commitment problem.
6. TEST SYSTEMS AND RESULTS

The authors have tested the new “equivalencing” method for solving short
term thermal unit commitment problem solution for two large power systems
consisting of the 26 and 79 thermal units, respectively, with a particular daily
load demand profile [8,9]. We base the system spinning reserve on the capacity
of the largest on-line unit. The 79-unit system represents Eastern Regional grid
of the Indian power system. We implemented the new algorithm in FORTRAN
77 code on a PC-486 computer.
In the equivalent system representation, 8 units represent the complete 26-
unit system, and 19 units represent the 79-unit system. Therefore, we can see
that the size of the problem has been reduced drastically. The comparison of the
proposed method with other traditional techniques like Lagrangian relaxation
and truncated dynamic programming [10,11] are presented in Tables 1 and 2
for 26-unit and 79-unit systems, respectively.
In the Lagrangian relaxation technique, form the Lagrangian dual by ap-

pending the relaxed constraints to the primal objective. Find the minimum of
the primal objective by maximizing the dual objective. Update the Lagrangian
multipliers using a sub-gradient method that drives the solution towards feasi-
bility [12]. In the DP-TC technique, the ordering of the units are made based on
an average full load cost of each unit, and the size of the search range of four
units (16 combinations) has been considered.
7. DISCUSSION
The results given in Tables 1 and 2 clearly show that the new equivalencing
method for solving the large scale unit commitment problem is able to provide
solutions very close to the best solutions found by other approaches. The varia-
tion in cost is within 0.05% and 0.19% respectively for the 26-unit system
compared to truncated dynamic programming and the Lagrangian relaxation
approach, respectively. For the 79-unit system, the cost obtained using the new
method is 0.7% more compared to the DP-TC-based method while it is 1.27%

less compared to Lagrangian relaxation-based method.
The solution time in the new approach, however, is smaller as compared to
other techniques. This is one of the main requirements to solve the unit com-
mitment problem for a large-scale power system. In addition, the computer
space requirement for the new method is comparatively less than other standard
methods of UCP solution. Also, the new method simplifies the unit commit-
ment problem in terms of dimensionality of the problem. As a result, the new
method of solving the unit commitment problem for large-scale systems turns
out to be a promising one.
We tested the algorithm for one sector of Indian power system, however, it
is not presently implemented by Indian power utility. In fact, due to restrictions
on the available generation capacity, presently India uses just merit order
scheduling. Further, the algorithm neglects to consider the transmission con-
straint. If it must be considered, however, then the concept presented in this
work would have to be applied to a cluster of units which are relatively close to
each other, and then the algorithm would not violate the transmission con-
straint.
8. CONCLUSION
In this chapter, we developed the “Equivalencing” method based on cardi-
nality reduction by generator equivalencing for estimating the short-term ther-
mal unit commitment schedule for large scale power systems. In this method,
we reduce a large-scale power system to an equivalent reasonably small-sized
power system. We first solve the equivalent system by using a modified dy-
namic programming approach (or it can be solved by any other standard unit
commitment solution technique). Finally, we return to the original system to
calculate the unit commitment schedule cost and determine the units’ final
status. This method simplifies the unit commitment problem in terms of dimen-
sionality, and consequently, computer space as well as the computer time re-
quired for solution can be reduced remarkably with an acceptable overall solu-
tion of the large-scale power system. We have tested the model on a practical
large and complex Indian power system. The results obtained from the above
model are highly impressive and encouraging for implementation of real-life
large scale power systems having large numbers of units. In short, the proposed
unit commitment model yields a promising approach to solve the short-term
thermal unit commitment problem and offers good performance which provides
fast solutions for large-scale power systems.
ACKNOWLEDGEMENTS
The chapter’s presentation was greatly improved by the comments and sug-
gestions of the editors and three anonymous referees. The authors also wish to
thank B. Hobbs for suggestions, which improved many aspects of this chapter.
REFERENCES
1. A.J. Wood and B.F. Wollenberg. Power Generation Operation and Control. New York:
John Wiley, 1996.
2. I.J. Nagrath and D.P. Kothari. Power System Engineering. New Delhi: Tata McGraw-Hill,
1994.
3. S.A. Kazarils, A.G. Bakirtzis and V. Petridis. A genetic algorithm solution to the unit com-
mitment problem. IEEE Trans. Power Syst., 11(1): 83-92, 1996.
4. G.B. Sheble and G.N. Fahd. Unit commitment literature synopsis. IEEE Trans. Power Syst.,
9(1): 128-135, 1994.
5. S. Sen and D.P. Kothari. Optimal thermal generating unit commitment: A review. Elec.
Power Energy Syst., 20(7): 443-451, 1998.
6. S. Bullard and R.E. Wiggans. Intelligent data compression in a coal model. Oper. Res., 36:
521-531, 1988.
7. D.P. Kothari and S. Sen. Optimal thermal generating unit commitment – a novel approach.
In Proc. International Seminar on Modelling & Simulation, Australia, 331-336, 1997.
8. C. Wang and S.M. Shahidehpour. Effects of ramp-rate limits on unit commitment and eco-
nomic dispatch. IEEE Trans. Power Syst., 8(3): 1341-1350, 1993.
9. S. Sen and D.P. Kothari. Evaluation of benefit of inter-area energy exchange of Indian power
system based on multi area unit commitment approach. Elec. Machines Power Syst., 26(8):
801-813, 1998.
10. S.J. Wang, S.M. Shahidehpour, D.S. Kirschen, and G.D. Irisarri. Short-term generation
scheduling with transmission and environmental constraints using an augmented Lagrangian
relaxation. IEEE Trans. Power Syst., 10(3): 1294-1301, 1995.
11. C.K. Pang, G.B. Sheble and F. Albuyeh. Evaluation of dynamic programming-based meth-
ods and multiple area representation for thermal unit commitments. IEEE Trans. Power
Syst. PAS-100, 3: 1212-1218, 1993.
12. W.L. Peterson and S.R. Brammer. A capacity-based Lagrangian relaxation unit commitment
with ramp rate constraints. IEEE Trans. Power Syst., 10(2): 1077-1084, 1995.
APPENDIX: Modified DP Technique for Solving UCP

Theory
Formation of Unit Commitment Schedule Table
Let the cost function of the n-th unit at a plant be [2]:

For minimization of cost, using well-known dynamic programming, a simple

recursive expression can be obtained as given below:
Where is the cost of generation at the n-th unit with as power dispatch.
is the minimum cost of generating by the remaining (N-
1) units of the plant.
Re-writing equation (10) as
where, For minimum of
or
where
From equations (11) and (12), obtain the composite cost function of units
#1 and #2 for a demand of as shown below:
where
Equation (13) represents the most economical cost of the two units for total
generation allocation over two units with as generation allocation on the sec-
ond unit.
In general form,
where
Equation (14) can also be expressed as
We obtain the critical value of generator loading/generation dispatch by

equating (14) and (15) above, in which the combination of n-number of unit
commitment will be economical as compared to number of units combina-
tion. Therefore, a “loading range” of operation can be obtained for units
for which the cost is minimum as compared to n-units combination. In this
chapter, we use this principle to prepare a sequential economic order of units
for commitment and the economic “loading range” of operation for those com-
binations.
Generation Limits Constraints Fixing
To satisfy the maximum and minimum generation limits during generation

scheduling, we apply the following limits constraint fixing technique.
i) Solve unconstrained generation dispatch problem.
ii) If there is no limit violation, the solution is optimal, otherwise compute

as follows:
for all generation limit upper-bound violation.
for all generation limit lower bound violation.
(iii)If ¨ fix all upper bound violations units to the upper limits, if
fix all lower bound violations units to the lower limits, otherwise
fix both upper and lower bound violations to upper and lower limits re-
spectively.
iv) Determine the new demand which is the original demand minus the sum of
fixed generation levels.
v) Set aside fixed generation level units for scheduling of new demand.
vi) Reschedule new demand among remaining committed units and return to
step (ii).
UP- and DOWN-Time Constraint Fixing
The purpose of this postprocessor algorithm (rules) is to detect the violation

of the minimum up- and down-time constraint of units. Minimum up and down
times are particularly difficult to model and cannot be incorporated directly in
the main program routine. So to detect a violation, we have established certain
rules:
Check Constraint (up time/down time violated)
Condition (unit on/off)
Condition (unit on/off time < min. up/down time)
If a unit is committed, then de-committed, and the duration between on and
off state is less than its minimum up time, then the unit up time is violated.
Therefore, the unit is charged as if it was on stand-by for those hours. In the
original unit commitment schedule, the stand-by hours are set to 1 (unit “on”)
instead of 0 (unit “off”). The additional number of hours needed to satisfy the
minimum up-time constraint are multiplied by the banking cost and then added
to the UC schedule cost.
On the other hand, if a unit is de-committed, then committed, and the dura-
tion between off and on states is less than its minimum down time, then the unit
down time is violated. Therefore, the unit is charged as if it was on stand-by for
the additional number of hours needed to satisfy the constraint. The original
unit commitment schedule has those hours set to 1 instead of 0. UC schedule is
“costed” as if it was banking for those hours.
Algorithm
Step 1: Read number of units, unit parameters, hourly demands, etc.

Step 2: Select any unit as a first unit from the list of the total number of units
and using equations (14) and (15), the composite cost function of the
two units, taking all units one by one is formed.
Step 3: We find the critical loading value for the combination of two units by
equating the cost function of the first unit with the composite cost
function of two units or equivalent two units taking sequentially. The
accepted combination of two units is that having minimum critical
“loading” value
Step 4: Repeat Steps 1 and 2, taking all units sequentially as the first unit and
find the N-number of minimum critical loading are found from each
combination. Among these, the maximum of minimum critical loading
is the best combination and sequential order of the first two
units. If is higher than the sum total of maximum capacity of two
or equivalent two units, then is reset to the sum of the maximum
capacity. For the equivalent two units combination case, the minimum
loading will be the just slightly higher than the maximum loading of
the previous combination.
Step 5: The sequence of the two units formed in Step 3 can be used as a basic
combination as the equivalent first unit to search the third most eco-
nomic unit. Repeat the procedure in Steps 1 to 3 to find the best com-
bination of three units and the corresponding third unit in sequence.
The algorithm uses the combination of three units as basic combina-
tion to search the next economic combination of four units. In this
way, for any addition of units in basic combination of j units, follow
the procedure mentioned in Steps 1 to 3 until all units are considered
to form a sequential order/ combinations of units and a corresponding
order of combination and loading range of operation.
Step 6: Execute limits constraints fixing, up- and down-time violation rules.
Step 7: Based on the general unit commitment schedule table and Step 6, per-
form the unit commitment schedule for the particular load profile.
Chapter 13
STRATEGIC UNIT COMMITMENT FOR GEN-

ERATION IN DEREGULATED ELECTRICITY
MARKETS
A. Baíllo, M. Ventosa, A. Ramos, M. Rivier

Universidad Pontificia Comillas de Madrid.
A. Canseco
IBERDROLA S.A.
Abstract: In this chapter we address some of the new short-term problems that are faced
by a generation company in a deregulated electricity market, and we propose a
decision procedure to address them. Additionally, we propose a strategic unit
commitment model, which deals with the weekly operation of the firm’s gener-
ating facilities. In it we combine traditional cost-evaluation techniques and tech-
nical constraints that grant a feasible schedule with new market-modeling equa-
tions. We suggest strategic constraints that allow the accomplishment of the
firm’s medium-term objectives. We have formulated the model as a mixed-
integer-programming problem and solved it by means of a commercial algo-
rithm, instead of using the traditional Lagrangian relaxation approach. Results of
the application of the method to a numerical example are presented. The proce-
dure is a simplified version of one of several tools currently being used by a
leading Spanish generation company, Iberdrola, for the weekly operation of its
generation assets in the Spanish wholesale electricity market.
1. INTRODUCTION
The electricity industry is in the midst of a profound restructuring process
in an increasing number of countries. These changes are intended to bring
about competition in some of the electricity business activities, so as to pro-
mote a higher level of efficiency in the provision of electric services. Al-
though the details of the deregulated marketplace may vary from one case to
another, it is generally assumed that electricity should be traded in a similar

fashion to other energy commodities.
Generation companies have traditionally been subject to regulatory poli-
cies that guaranteed the full recovery of their costs. In the new framework,
electricity generation is a deregulated activity and firms have to compete to
sell the electric services provided by their facilities. Therefore, generation
companies are now fully responsible not only for the efficient operation of
their units, but also for selling their output.
The particular design of the marketplace where the electric services are
traded is of great importance. Two mechanisms coexist in recently deregu-
lated electric industries throughout the world. The most common is a central-
ized power exchange based on auctions where the price for each electric ser-
vice during a certain time period is determined by the intersection of the ag-
gregate supply and demand bid curves. Additionally, bilateral trading between
buyers and sellers is usually permitted.
The fact that generation companies’ revenues depend on the market forces
leads to a higher degree of uncertainty and risk. New procedures and tools
devoted to the maximization of the firm’s profit, taking into account the dif-
ferent market mechanisms available and keeping an upper bound on the de-
gree of risk exposure, are needed. However, introducing in a model all the
complexity of electricity trading does not necessarily result in a better under-
standing of the environment. A gradual implementation of solutions for the
open issues should lead to a deeper and more solid knowledge of their impli-
cations.
In this chapter, we address some of the new problems that are faced by a
generation company in the short term (one day to a week). Assuming that the
majority of energy is traded in an energy exchange based on 24 day-ahead
hourly uniform-price auctions, we outline a decision procedure. One of the
tools incorporated in this procedure is a strategic unit-commitment model,
which deals with the weekly operation of the firm’s generating facilities. It
includes traditional cost evaluation techniques and takes into account techni-
cal constraints that grant a feasible schedule. Its major contribution lies in the
inclusion of a set of market-modeling equations intended to express the rela-
tionship between the firm’s output and the revenue obtained in the successive
auctions. Additionally, we show the need to consider a set of strategic con-
straints so as to direct the solution towards the firm’s medium-term objec-
tives. We have formulated the model as a MIP optimization problem and
solved it by means of a commercial algorithm instead of using the traditional
Lagrangian relaxation approach. We present results of the application of the
method to a numerical example.
Strategic Unit Commitment for Generation Companies 229
2. DECISION PROCEDURE
In the new competitive framework, a generation company not only has to
determine how to operate its generation facilities in the most efficient manner,
but also must decide on the amount of each electric service that should be
supplied, at which moment it should be produced, at what price it should be
sold, and with which units it should be provided. These new challenges re-
quire decision procedures and tools specifically oriented to the maximization
of the firm’s profit and the hedging of its risk. The short-term decision proce-
dure outlined in this section is represented in Figure 1.
2.1 Medium Term Guidelines

As in the past, generation firms have to decompose the problem of plan-
ning the operation of their units into different time scopes to make it tractable.
The traditional hierarchy used to classify the decision support tools into long-,
medium-, and short-term models is still completely in force. The results ob-
tained from a model with a longer time scope must affect all the inferior mod-
els. Consequently, in a short-term decision procedure the generation firm’s
medium-term goals play an important role.
In its medium term analysis (one month to a year) the generation firm still
has to make traditional decisions related to the maintenance of the groups, the
annual management of water reserves, or the fuel consumption of thermal
plants. New issues include determining the expected equilibrium between
generation companies and estimating the prices that are likely to appear dur-
ing the following months.
Two very important results of the medium term, which the firm must con-
sider in the short term (one day to a week), are the water value assigned to the
water reserves and the medium-term market position that allows an equilib-
rium with the rest of generation firms.
Firms must assign a certain value to their stored energy. Otherwise, short-
term tools will interpret that producing with these reserves has no associated
costs and the result will be that hydro units must permanently produce at their
maximum capacity. Bushnell [1] analyzed the strategic management of hydro
resources in a competitive environment and established a relationship be-
tween the value of the available hydro energy and the marginal cost of ther-
mal units. Scott and Read [2] used Dual Dynamic Programming methods to
build up a weekly curve giving the optimal output for a certain water value.
Similarly, if a short-term model is not aware of the market position de-
fended in the medium term by the firm, it will tend to follow blindly the short-
term signals transmitted by the competitors through their supply curves. We
will analyze in detail how the slope of the bid curve presented both by other
generators and by the demand exerts an influence on the results given by a
short-term model that tries to maximize the short-term profit of a generation
firm. This influence must be limited and controlled so that the firm is able to
keep a steady pace towards its medium-term objectives, which include de-
fending its market position. A firm may lose its market position by systemati-
cally producing less than the market share it should have according to the cost
and size of its generating assets relative to those of its competitors. Several
approaches have been proposed for approximating the medium-term equilib-
rium that a number of generating firms should reach in a competitive electric-
ity market. Some of these consider the generation firms as Cournot agents
whose decision variables are the quantities produced in each time period.
Otero-Novas et al. [3] developed a simulation platform to evaluate the me-
dium-term evolution of the Spanish electricity market. Ramos et al. [4] and
Ventosa et al. [5] successfully combined a detailed representation of the gen-
eration operational costs with the Cournot-equilibrium conditions in a cost-
minimization framework. Rivier et al. [6] used the complementarity problem
approach with satisfying results to determine the expected medium-term equi-
librium reached in the recently deregulated Spanish electricity industry.
Hobbs [7] and Wei and Smeers [8] extended the usage of the complementarity
problem to predict the outcome of an electricity market with significant
transmission constraints. An important feature of all these models is that they
reflect the objective of profit maximization of all generation firms so that the
medium-term market equilibrium is accurately represented.
2.2 Forecasting Techniques

A major challenge for a generation company in the new framework is the
development of forecasting techniques devoted to the estimation of the oppo-
nents’ expected behavior. Information concerning the amount of each service
traded at different prices is extremely relevant. These data can be exploited to
estimate future competitors’ hourly offer curves. The expected level of de-
mand is also a decisive variable, whereas the elasticity of the demand curve is
a parameter whose importance is expected to grow as the agents gain experi-
ence in the new regulatory scheme.
2.3 Short-term Generation Scheduling

Given a certain scenario for the competitors’ offers and the demand bids
for the different electric services markets, the firm has to determine the en-
ergy that it should offer as well as the capacity that should be reserved for
ancillary services. When the firm’s revenue is based on administrative and
centralized decisions, the amount of energy that each unit must produce and
the precise moments when this unit should start up and shut down are tackled
with weekly cost-minimization unit-commitment models. However, deregula-
tion has shifted focus from obligation of supply and cost-minimization to
competition and profit maximization. Therefore, the competitive environment
requires unit-commitment models that take into account the expected price
series for each of the electric services. Moreover, if the firm has a significant
market share, the response of prices to the output of the generation firm must
be considered [9].
Unit-commitment models provide a commitment schedule. However, they
usually give only approximate generation levels for the generators [10]. Once
the commitment decisions are taken, the best hourly output of each individual
generator can be determined with a daily model. This model will include a
more detailed representation of the generating equipment, such as an accurate
description of hydro generation resources.
2.4 Strategic bidding

The final stage of the generation firm’s decision procedure is the design
of the hourly offer curves that must be submitted to the different auctions. The
results of the previous short-term decision support tools include the firm’s
expected hourly optimal productions together with expected hourly prices.
However, the competitors’ sell offers as well as the demand-side buy bids are
estimated with a certain degree of uncertainty, since the short-term volatility
of electricity prices is higher than those of other energy commodities. In this
context, instead of sticking to a single hourly quantity, the firm can submit an
hourly offer curve for each of the electric services. The bigger and more
flexible its generation portfolio is, the wider the variety of hourly offer curves
the firm can design. In this chapter, we assume that offer curves consist of a
set of quantity-price pairs and that no additional information, such as fixed-
costs, is submitted. Given a probability distribution for the last accepted bid
(Figure 2), the generation firm can derive the offer curve that maximizes its
objective function. This can be a combination of the expected short-term
profit and other targets such as market share goals.
3. MODEL DESCRIPTION
In the previous section, we suggested a short-term decision procedure for
a generation firm participating in a day-ahead auction-based energy exchange.
The unit-commitment model that we will develop henceforth, however, is a
simplified part of the complex combination of tools that a generation firm
should use to face the short-term problems that will arise in the new competi-
tive framework. We will only consider the day-ahead hourly energy market.
Our aim is to gain insight into certain modeling features such as the manage-
ment of hydro reserves or the influence of strategic constraints.
3.1 Objective Function
A major difference between the strategic unit commitment model and a

traditional unit commitment model lies in the objective function. The aim of a
traditional model is the minimization of the overall system costs. In spite of

their complexity, cost functions of thermal units have frequently been mod-
eled as convex piecewise linear functions. In contrast, the strategic unit com-
mitment model guides a generation firm to its maximum profit objective,
which is a non-linear (and frequently non-convex) function of the firm’s en-
ergy output.
3.2 Generation System Representation
Thermal units’ costs representation includes fuel costs (which, for sim-
plicity, can be defined as a convex piecewise linear function of the unit’s out-
put), start-up costs and shut-down costs. Thermal units’ most relevant con-
straints are the minimum stable load, maximum output, and upwards and
downwards ramp limits.
Hydro units produce with nearly zero variable costs. Water reserves have
associated opportunity costs, however, as they can be used to substitute ther-
mal units. Therefore, we can define a cost function known as water value. The
water value function gives the hydro energy that must be produced if the mar-
ginal revenue of the firm exceeds a certain value. In the strategic unit-
commitment model, we divide hydro reserves into several reserve levels and
assign a different water value to each one of them. Depending on the market
circumstances, the model will decide to use a certain amount of each one of
these levels. To keep track of the contents of these reserve levels we will in-
clude hourly reserve balance equations. Hydro units also have a minimum and
a maximum power output.
The operation of pumped-storage power plants is subject to the same con-
straints as regular hydro power plants, except that the water balance con-
straints are modified to include the pumping mode of operation.
3.3 Market Representation

The strategies followed by the firm’s competitors are expressed by means
of their expected hourly offer curves (known also as hourly supply functions).
Similarly, instead of using a fixed level of demand for each time period, en-
ergy buyers submit hourly demand curves.
From the generation firm’s point of view, the competitors’ hourly offer
curves and the hourly demand curve exert a very similar influence on the
firm’s profit (see Figure 3). If the firm increases its energy output, lower
prices will result. This is due to the combined effect of a decrease in the com-
petitors’ output and an increase in the energy consumption. Therefore, to a
certain extent, the firm is able to adjust its hourly revenue by varying its en-
ergy output. In microeconomic theory, this is modeled by means of the resid-
ual demand function. This gives the energy the firm is able to sell at each
price. The firm should try to estimate this function for each hour.
Going a little further, a change in the firm’s production also modifies the
competitors’ revenue. It must not be forgotten that a decrease in the firm’s
production may or may not increase its profit, but it will surely benefit the
competitors as they are able to produce more and at a higher price (Figure 4).
The objective function whose maximum is sought is the firm’s profit, de-
fined as the difference between the obtained revenue and the incurred costs.
Owing to the fact that the most powerful available solvers are those designed
for mixed integer linear problems, such as CPLEX and OSL, a linearization
procedure for the firm’s revenue function is of great interest. An intuitive
method is to divide the firm’s hourly revenue function into convex sections
and approximate each one by a piecewise linear function. The slope obtained
for each linear segment is the firm’s marginal revenue at the corresponding
energy output (Figure 5).
A group of consecutive segments with strictly decreasing marginal reve-

nues defines a convex section in the revenue function. When we seek the op-
timum we select a specific convex section by switching its binary variable
from zero to one. Once we have chosen a convex section we fill its segments
with continuous bounded variables. In other words, we obtain the hourly
revenue by integrating the marginal-revenue function.
With this approach, we replace the set of constraints that represent the
competitors’ generation facilities in a traditional unit commitment by a set of
hourly constraints, which define the firm’s hourly revenue as a function of its
energy output. Hourly prices do not appear explicitly in our model.
3.4 Medium-term Guidelines
The results obtained from medium-term models add information to short-

term decision-support tools. Good examples are the water value assigned to
hydro reserves or the position that the firm must defend if it wants to maxi-
mize its profit in the long run.
Our model decides the amount of hydro energy to be used depending on
the water value received from a hydrothermal co-ordination model. It also
distributes this energy along the time scope of the model, trying to obtain a
uniform marginal revenue.
The total hourly energy production that the model suggests not only de-
pends on the expected price and on the firm’s marginal production costs. We
also consider the slope of the residual demand curve, which exerts a major
influence on the clearing price (Figure 6). If this curve is very steep and the
firm’s output is high (on-peak hours) the model will blindly tend to reduce the
firm’s production. This causes a rise of the energy price and an increase of the
firm’s profit. Another result is that competitors are able to produce more at a
higher price. Taking into account that the price of electricity usually behaves
like a mean-reverting process, if the firm gives up its position repeatedly dur-
ing on-peak hours, competitors will increase their market shares and, in the
long run, prices will return to the original level.
An alternative is to define a set of hourly minimum-market-share con-

straints. In this fashion, the short-term model maximises the firm’s short-term
profit while following the right medium-term strategy. We will analyze the
influence of minimum-market-share constraints in the case study.
Many other constraints can be designed ad hoc to fulfill the firm’s re-
quirements such as existing physical or financial contracts. In our model, they
must be expressed as linear equations. The strategic UC solution will be dif-
ferent in each particular case. A great effort must be made to interpret cor-
rectly the influence of each factor both in the short and long term.
4. MATHEMATICAL FORMULATION
4.1 Notation
In this section, we identify the symbols used in this chapter and classify
them according to their use. Table 1 shows the indices and sets considered,
capitals being used for sets and lower-case for indices. Table 2 includes the
decision variables. Table 3 lists the auxiliary variables. Table 4 defines the
information given to the model as fixed data.
4.2 Model formulation

We formulate the model as an MIP optimization problem. The objective
function to be maximized is the firm’s total profit for the scope of the model.
We have classified operating constraints into thermal and hydro constraints.
Additional market constraints model the behavior of the competitors and the
demand side in the day-ahead power market.
4.2.1 Objective Function
The objective function represents the firm’s profit defined as the differ-
ence between the firm’s revenue and the firm’s operating costs for all load
levels within the scope of the model:
4.2.2 Thermal Generation Constraints
Total thermal operating costs include fuel costs, O&M costs, start-up
costs and shut-down costs:
For each committed thermal unit, the maximum generation is less than the
maximum available capacity, and the minimum generation is greater than the
minimum stable load:
The hourly change in the output of each thermal unit is limited by the
ramp rates:
A logical relationship exists between the start-up, shut-down, and com-

mitment variables:
Since the commitment decision variables are binary, both the start-up and
the shut-down decision variables can be continuous but must have upper and
lower bounds:
4.2.3 Hydro Generation Constraints
The water reserves not used by the model have a value for the future:
Each unit has an upper and a lower limit for its power output:
The contents of the reservoirs depend on the energy produced or stored

during each time period and have upper and lower bounds:
4.2.4 Market Constraints
Each segment of the firm’s net hourly energy output is valued at a differ-
ent marginal revenue. The sum of all the segments must equal the sum of the
power produced by thermal and hydro units minus the power consumed by
pumped-storage units:
Each segment has an upper and a lower bound and the convex sections
must be chosen in order:
We calculate the total revenue by valuing the different segments of the net
energy output at their corresponding marginal revenues. In other words, the
revenue is obtained by integrating the marginal revenue function:
The hourly price of energy is not obtained explicitly with this formula-
tion. It must be calculated after the execution of the model. To do so, we sim-
ply divide the firm’s hourly revenue by the firm’s hourly production.
4.2.5 Strategic Constraints
We define a set of hourly minimum-market-share constraints. In the nu-

merical example, we investigate the influence of this constraint on hourly
prices and on the firm’s short-term benefit. In our formulation, we suppose
that demand is perfectly inelastic. Consequently, the only variations of de-
mand we allow are those introduced by pumping:
5. NUMERICAL EXAMPLE
The strategic unit commitment model has been implemented in GAMS
[11]. A case study has been solved with the optimizer CPLEX 6.5.
5.1 Case Study

We include the results of the application of the model to a case study. Our
aim is to highlight the influence of the different modeling decisions previ-
ously explained. Computational and convergence issues are secondary and
should be treated only when the researcher is sure that the model adequately
represents the problem he is trying to solve. Consequently, we will only give
the essential information to define the case study. (The authors may be con-
tacted for details if the reader wishes to reproduce the results.) The firm’s
generating equipment is described in Table 5.
Hydro and reserves have been classified into the levels shown in Table 6
according to the results of a hydrothermal coordination model.
We have estimated a piecewise linear residual demand function for each

hour. Figure 7 shows an example of a residual demand curve with its corre-
sponding revenue function, formed by two convex sections. Each of these
sections has been divided into five segments. We have assigned a constant
marginal revenue to each segment.
In this example, we include a set of hourly minimum-market-share con-

straints to obtain a generation schedule similar to the traditional one. If no
strategic constraints were used, the model would blindly follow all the short-
term opportunities and the resulting operation would require extremely ineffi-
cient dynamic performance of the generating units. The weekly problem is
formed by 19446 equations, 24555 continuous variables and 4692 binary
variables. It is solved in 89 seconds on a PC Pentium II 350 MHz 64 MB.
5.2 Results
In this section, we describe the results of the model when a minimum-
market-share constraint of 29% is used. Table 7 states the differences between
the solution given as optimal by CPLEX and the first feasible solution.
The model decides the hourly power output for each ofthe firm’s generat-
ing units. The hourly energy that each kind of unit should produce to achieve
this hourly market share has been represented in Figure 8.
Each hourly energy output determines an hourly revenue by means of the

estimated marginal revenues. Additionally, the production of that energy
leads to an hourly cost. The difference gives the firm’s expected hourly profit.
These three variables have been represented in Figure 9.
A subproduct of the problem is an estimation of the hourly price of en-

ergy. We calculate it after the execution of the model, dividing the firm’s
hourly revenue by its hourly production. As we observe in Figure 10, due to
the differences among hourly revenue functions, the series of energy prices is
only partially correlated to the series of the firm’s energy outputs.
The model also decides the optimum strategic management of the existing
water reserves. The usage of hydro energy depends on the assigned water val-
ues. The model will use a certain amount of water reserves if their value is
lower than the maximum marginal revenue reached during the week. In this
example, all reserve levels are used except for the one valued at 37.5 $/MWh.
This indicates that the marginal revenue never reaches that value.
Similarly, the pumped-storage unit consumes energy when the value of its
reserve is times higher than the weekly minimum marginal revenue. Con-
versely, this water will be released if the weekly maximum marginal revenue
reaches the water value of the pumped-storage (Figure 11). In this case the
final contents of the pumped-storage reservoir have been set equal to the ini-
tial ones.
5.3 Influence of the Strategic Constraints

We now analyze the influence of introducing a minimum-market-share
constraint. In our case study, the firm faces such a steep residual demand
curve that the incentive to withdraw production from the market is very
strong. Although this leads to high prices in the short term, the firm should
administrate its market power with care to guarantee a solid long-term market
position.
We have solved the case study with ten levels of minimum market share,
ranging from 26% to 30.5%. Figure 12 shows part of the price series obtained
for three of these market-share levels.
As we can see, lower market shares produce higher prices. In this case,
the rise of prices overcomes the market share reduction. Consequently, by
withdrawing, the firm achieves both higher revenues and lower costs (Figure
13).
As stated by Viscusi et al. [12], one pricing strategy is for an incumbent

firm always to set price so as to maximize current profit. Typically, setting
such a high price will cause the fringe to invest in capacity and expand.
Therefore, this can be called myopic pricing. The polar opposite case is for
the incumbent to set price so as to prevent all fringe expansion (limit pricing).
Myopic pricing gives higher profits today, while limit pricing gives higher
profits in the future. Pricing at a level to exclude from the market less effi-
cient competitors is, of course, what competition is supposed to do. Pricing to
exclude equally or more efficient competitors is known as predatory pricing
and constitutes an intent to acquire the monopoly position.
To determine the position that must be defended in the market, a genera-
tion firm can assign a cost to the deviations from its medium-term market-
share objective. The firm must keep an eye on the, say, one-month moving
average of its market share. If its market share remains for more than one
month below the objective, then its competitors may understand this as a
change in the medium-term equilibrium conditions. Therefore, the firm can
expect prices to stay high for a month. After this period, prices will revert to
their medium-term mean. If this happens, the firm will be forced to suffer low
prices to recover the lost position. With this approach, each time the firm’s
one-month market-share moving average lies below the medium-term objec-
tive it accounts for a loss. On the other hand, the firm should increase its mar-
ket share cautiously, as this can lead to a price war. The cost function looks
like the one sketched in Figure 14.
Depending on the short-term market conditions and on the accumulated

market-share, there will be weeks when the model will suggest producing
above or below the market-share objective. An additional consideration is that
the cost of market-share deviations will probably be higher in on-peak hours
than in off-peak hours.
6. CONCLUSION
In the new deregulated electric marketplace, generation companies have
to compete to sell the electric services provided by their facilities. They must
develop new procedures and tools devoted to maximize profit and hedge
risks.
In this chapter, we addressed several new short-term problems that are
faced by a generation company. Stemming from the medium-term objectives
of the firm, our aim has been to determine the optimal combination of offers
and contracts for the supply of electric services through the different market
mechanisms available. We have tried to emphasize the importance of an ade-
quate design and use of the decision-support tools.
In this context, we presented a strategic unit-commitment model specifi-

cally devoted to the profit maximization of an electric generation firm. It in-
corporates new market-modeling equations intended to express the relation-
ship between the firm’s output and the obtained revenue. We have made a
special effort to interpret the effect of these new equations on the management
of hydro reserves. Additionally, we demonstrated how the position that the
firm must defend in the market must guide the short-term operation towards
the medium-term equilibrium.
The procedure is a simplified version of the one currently being used by a
leading Spanish generation company, Iberdrola, for the weekly operation of
its generation assets in the Spanish wholesale electricity market.
REFERENCES
1. J. Bushnell. Water and Power: Hydroelectric Resources in the Era of Competition in the
Western US. POWER Conference on Electricity Restructuring. University of California
Energy Institute, 1998.
2. T.J. Scott, and E.G. Read. Modeling hydro reservoir operation in a deregulated electricity
market Int. Trans. Oper. Res., 3: 243-253, 1996.
3. I. Otero-Novas, C. Meseguer, and J.J. Alba. A Simulation Model for a Competitive Gen-
eration Market. IEEE Power Engr. Soc., Paper PE-380-PWRS-1-09-1998.
4. A. Ramos, M. Ventosa, and M. Rivier. Modeling competition in electric energy markets
by equilibrium constraints. Utilities Policy, 7(4): 233-242, 1998.
5. M. Ventosa, A. Ramos, and M. Rivier. “Modeling Profit Maximization in Deregulated
Power Markets by Equilibrium Constraints.” PSCC Conference, Norway, 1: 231-237,
1999.
6. M. Rivier, M. Ventosa, and A. Ramos. A generation operation planning model in deregu-
lated electricity markets based on the complementarity problem. ICCP99 Conference,
Wisconsin, 1999.
7. B.F. Hobbs. “LCP Models of Nash – Cournot Competition in Bilateral and POOLCO–
Based Power Markets.” In Proc. IEEE Winter Meeting, New York, 1999.
8. J.Y. Wei and Y. Smeers. Spatial oligopolistic electricity models with Cournot generators
and regulated transmission prices. Oper. Res., 47(1): 102-112, 1999.
9. J. Garcia, J. Roman, J. Barquín, and A. Gonzalez. “Strategic Bidding in Deregulated
Power Systems.” PSCC Conference, Norway, 1: 258-264, 1999.
10. R. Baldick. The generalized unit commitment problem. IEEE Trans. Power Syst., 10(1):
465-475, 1995.
11. A. Brooke, D. Kendrick, and A. Meeraus. GAMS A User’s Guide. Boyd and Fraser, 1992.
12. W.K. Viscusi, J.M. Vernon, and J.E. Hamington. In Economics of Regulation and Anti-
trust, ed, The MIT Press, 1998.
Chapter 14
OPTIMIZATION-BASED BIDDING STRATEGIES

FOR DEREGULATED ELECTRIC POWER
MARKETS
Xiaohong Guan
Harvard University, on leave from Xian Jiaotong University, China
Ernan Ni and Peter B. Luh

University of Connecticut
Yu-Chi Ho
Harvard University
Abstract: Deregulation of the electric power industry worldwide raises many challenging
issues. Aiming at these challenging issues and using California and New Eng-
land power markets as background, this chapter focuses on the methodologies
for integrated generation scheduling and bidding strategies for deregulated elec-
tric power markets. We present a systematic bid selection method based on or-
dinal optimization for obtaining “good enough” bidding strategies for generation
suppliers. A stochastic optimization method for integrated bidding and schedul-
ing is developed with consideration of risk management, self-scheduling re-
quirements, and the interaction between different markets.
1. INTRODUCTION
The electric power industry worldwide is experiencing an unprecedented
restructuring. In the United States, California was the first state to establish a
deregulated power market starting in April 1998. Since then, almost every
state has or is deregulating its power industry [1-3]. This chapter summarizes
the methodologies developed by the authors and the results achieved so far in
dealing with some challenging issues on integrated resource bidding and
scheduling in deregulated electric power markets.
In terms of the structure of resource allocation and scheduling, current
power markets can be classified into two types: individual and pool. The Cali-
fornia market belongs to the first kind. It includes a Power Exchange (PX)
managing the “day-ahead” and “hour-ahead” energy markets, an Independent
System Operator (ISO) handling real-time balancing, reserve and other ancil-
lary markets, and various energy and service suppliers and demanders. The
structure and functions of ISO and PX have been described in [3-4]. In the
day-ahead energy market, a power supplier submits to the PX piece-wise lin-
ear and monotonically increasing power-price “supply bid curves” for each
generator or for a portfolio of generating units, one for each hour of the next
day. On the other hand, an energy service company submits to the PX an
hourly power-price “demand bid curve” reflecting its forecasted demand. The
PX aggregates supply and demand bid curves to determine a “Market Clear-
ing Price” (MCP) and “Market Clearing Quantity” (MCQ) as shown in Figure
1. The power to be awarded to each bidder is then determined based on the
individual bid curves and the MCP. All the power awards will be compen-
sated at the MCP. After the auction closes, each supply bidder aggregates its
power awards as its system demand, and performs a traditional unit commit-
ment or hydrothermal scheduling to meet its obligations at minimum cost over
the bidding horizon. The ISO will check if the schedules submitted by suppli-
ers can be implemented through the transmission grid by performing a power
flow calculation and modify the schedules by calling some must-run units as
ancillary service. In this case, the constrained MCPs will be re-determined
and would be different for different areas. As pointed out in [1], suppliers’
bidding decisions are coupled with generation scheduling since generator
characteristics and how they will be used to meet the accepted bids in the fu-
ture have to be considered before bids are submitted. Therefore, bidding deci-
sions must consider the anticipated MCP, generation award and costs, and
competitors’ decisions and other complicating factors, such as transmission
constraints of the power grid.
The power market in the United States New England region is formed
based on the New England Power Pool and belongs to the second type [5-6].
Optimization-Based Bidding Strategies 251
Organizationally, the functions of PX and ISO are combined under the aus-
pices of a single ISO. Unlike the separate and sequential energy and ancillary
service markets in California, the energy bids are integrated with other service
bids such as reserve and AGC. In addition, since the ISO has all the system
operational parameters of each generator, it clears the MCP and other market
prices by performing unit commitment or generation scheduling for the whole
power system in the market based on the power-price bid curves received.
Although the capacity of each unit has to be bid, an energy supplier may
“withhold” or self-schedule some capacity to meet some percentage of its own
load or to fulfill bilateral transactions with other market participators by bid-
ding zero or negative prices. Another difference is that a bid curve in the New
England market is a staircase or piece-wise constant function rather than a
piece-wise linear function.
Many challenging issues arise under the new competitive market struc-
ture. Instead of centralized decision-making in a monopoly environment as in
the past, many parties with different goals are now involved and competing in
the market. The information available to a party may be limited, regulated,
and received with time delay, and decisions made by one party may influence
the decision space and well-being of others. These difficulties are com-
pounded by the underlying uncertainties inherent in the system, such as the
demand for electricity, fuel prices, outages of generators and transmission
lines, tactics by certain market participants, etc. Consequently, the market is
full of uncertainty and risk. The recent experience learned from the many
markets has shown that MCPs are volatile and often out of the range of bid-
ders’ expectation. How to handle MCP volatilities, how to manage uncertain-
ties and risks, and how to allocate the generating capacity into different mar-
kets have become extremely important under the new market environment.
Aiming at these challenging problems and using the California and New
England power markets as background, we focus on the methodologies for
optimizing bidding strategies. Since bidding problems are multi-person, game
theoretic problems generally associated with inherent uncertainties and com-
putational difficulties, it is more desirable to ask which decision is better as
opposed to seeking an optimal solution. In our research, we develop two bid-
ding methods based on the structures and rules of two actual power markets:
the United States California and New England markets. We first concentrate
on a systematic bid selection method based on ordinal optimization to obtain
“good enough bidding” strategies for generation suppliers. We then present a
stochastic optimization method for integrated bidding and scheduling with
self-scheduling constraints. The risks in supply bidding are managed in a sys-
tematic way. We also explore the interactions between energy and other ancil-
lary service markets. Numerical testing shows that the algorithm is efficient
for daily bidding and scheduling.
2. LITERATURE REVIEW
Many approaches have been reported in the literature to address the struc-
tures and mechanism of deregulated power markets. Prior to the deregulation
in the United States, the market structure model discussed most is the “British
Model” [2]. The California model is presented in [3]. The structure of the
New England market is described in [5] and its ISO’s energy and ancillary
service dispatch problem is presented in [6]. Some recent studies on market
mechanism are primarily concerned with market analysis and market power
issues [7-8].
Game theory is a natural platform to model a gaming environment where
each participant is determined to maximize its profit [2, 9-12]. Optimal bid-
ding strategies to maximize a bidder’s profit based on the pool model of Eng-
land and Wales were presented in [2] under the assumption that any particular
bid has no effect on the MCP. For a market where a bid consists of start-up
cost, variable price, and generator capacity, it was demonstrated that profit is
maximized by bidding each generator at its physical cost curve and maximum
capacity. This is done by showing that such a strategy is a “Nash equilibrium”
for the market. The perfect conditions assumed in [2], however, may not be
true. Matrix games have been reported in [11] and [12]. Bidding strategies are
discretized, such as “bidding high,” “bidding low,” or “bidding medium,” and
an ”equilibrium” of the “matrix bidding game” can be obtained. The strategic
gaming behaviors and how the market structure affect the competition is ana-
lyzed in [9]. It is shown that the strategic behavior on electric network may
produce unexpected results from the traditional economic theory. Game the-
ory is used in [13] to minimize the risk in bidding problems.
Various other methods for solving bid selection problems at different lev-
els of the market have also been discussed. In [14], a bid clearing system in
New Zealand is presented. Detailed models are used, including network con-
straints, reserve constraints, and ramp-rate constraints, and linear program-
ming are used to solve the problem. Other approaches addressing various as-
pects of generation and ancillary service bidding can be found in [15-16],
where Lagrangian relaxation, and decision trees were used to analyze and
support the bidding process. For example, a bidding strategy considering
revenue adequacy was presented in [17] based on Lagrangian relaxation and
an iterative bid adjustment process. However this process may not be avail-
able for the current California PX market. A bidding method considering the
uncertainties of other bidders, the ISO’s bid selection process and self-
scheduling in New England power market is presented in [18]. The problem is
solved within a simplified game theoretical framework. The exact “gaming”
phenomenon among bidders however is not captured. Bidding behaviors un-
der a simple auction market are studied in [19]. The results show that power
suppliers would tend to bid above their production costs to hedge against the
possibility of winning on the margin. The factors affecting bidding strategies

in Australia power market are analyzed in [7]. More recent work on bidding
strategies concentrates on the adaptive method, i.e., reacting to inputs, based
on genetic programming and finite state automata [31]. It should be noted that
Lagrangian relaxation is a very successful price based method for hydrother-
mal scheduling [20], and the framework is also very useful in dealing with the
new integrated bidding and scheduling issues [4].
Recently, an intelligent computational method – Ordinal Optimization
(OO) has been developed to solve complicated optimization problems possi-
bly with uncertainties [21-22]. Ordinal optimization is based on the following
two tenets. First, it is much easier to determine “order” than “value.” To de-
termine whether A is larger or smaller than B is a simpler task than to deter-
mine the value of (A-B) especially when uncertainties exist. Second, instead
of asking the “best for sure,” we seek the “good enough with high probabil-
ity.” Softening the goal of optimization should make the problem easier. A
bid selection strategy is developed to generate good enough bids based on the
framework of ordinal optimization [23].
It can be seen from above that tools to support integrated bidding and
scheduling process of deregulated power markets are far from satisfactory in
view of the inherent complexity (multiple participants with their own objec-
tives in a dynamic and uncertain environment) and the sizes of practical prob-
lems (tens or hundreds of generators with various constraints). High quality
and computationally efficient approaches are critically needed to address the
new challenges.
3. ORDINAL OPTIMIZATION-BASED BIDDING

STRATEGIES
Assume the bidding strategy is developed for an energy or generation
supply company E. Suppose there are I generating units in E. Supply bids can
be submitted for individual units or an aggregated bid can be submitted in the
PX market. The bidding objective for E is to select its bid curves
to maximize its profit over a time horizon T, i.e.,
where
Aggregated energy (generation)-price supply bid curve
of E for hour t;
Generation and demand bids of other bidders unknown
to E for hour t;
Generation cost for delivering generation award
Power generation award for E depending on the bid of E

and the bids of other bidders;
T: Bidding time horizon (24 hours for the day-ahead mar-
ket); and
Market clearing price (MCP) determined by the aggre-
gated supply bid curve and the aggregated demand bid
curve of all market participants as shown in Figure 1.
The bid curves of other participants are assumed to be
fixed, and we are only interested in the influence of E’s
own bids on the MCP, which will be modeled later.
According to the PX rules, if a supplier bidder awarded it will
be compensated by the dollar amount regardless of the origi-
nal bid submitted by that supplier. Start-up costs should be accounted for in
bid curves since there is no direct start-up compensation. The problem de-
scribed in (1) is thus an optimization problem to determine the optimal supply
bid curves to maximize the profit subject to relevant operating
constraints, such as the minimum down-/up-time, ramp-rate constraints, etc.
Since the PX rules require bid curves to be piece-wise linear and monotoni-
cally increasing, searching the optimal bidding strategy is to determine the
corner points of the bid curves. Note that MCPs are determined by the bids
submitted by all the bidders, and when submitting the bids, a bidder does not
know the bid curves submitted by others.
Once the PX determines the MCP, the generation award for individual
units are determined as the intersection from their bid curves or for the gen-
eration supply company E from its aggregated bid curve with as the
given price. Although E can submit bids for each individual unit, the PX view
total generation award to E as obligation. Given a total generation award, it is
desirable for E to schedule or to reallocate all generating units across the bid-
ding horizon to deliver its total award at the minimal cost. This can be formu-
lated as a traditional unit commitment or hydrothermal scheduling problem
with the total generation award as system demand as follows:
subject to
and other individual operating constraints such as minimum down-/up-times

described in [23-26], where
C: Total generation cost over the entire bidding horizon;
Cost for generating for unit i at hour t;
Start-up or shut-down cost associated with the up-/down-
state transitions for unit i to generate at hour t;
Power generation of unit i at hour t; and
Generation award for unit i at hour t according to unit’s
individual bid curve.
If we know the market-clearing price (MCP) and consequently know the
generation award, we can re-write the profit calculation of (1) as
To select a good bidding strategy, it is necessary to identify two situa-

tions: 1) the participant is small and has no significant influence on the MCP;
2) the participant has market power and can influence the MCP. In the first
case, it is desirable to find a nominal bid curve that would maximize its profit
at any given MCP. For the second case, it is necessary to establish the rela-
tionship between one’s bidding strategy and the MCP. One approach is to ob-
tain the MCP by simulating market participants’ bidding strategies. By chang-
ing one’s own bidding strategies, it is possible to establish the influence of
bidding strategies on the MCP. Another method is to establish an influence
function by experience or by regression through one’s historical bids and the
MCPs. Either way, assume the following influence functions:
where is the nominal MCP forecasted by the historical

MCPs the forecasted system demand and other forecasted factors
such as fuel prices, etc. The nominal MCP forecasting using an ANN
model has been reported in our recent work [30]. The influence function
is determined by the difference between one’s bid curve and
the nominal bid curve and can also be established by regression method based
on the historical MCPs and records of one’s own bids plus bidders’ experi-
ence. The advantage the ordinal optimization method is that it has no restric-
tion on any model used to describe bidding strategy influence on the MCP. In
fact (5) can be replaced by a Monte Carlo simulation procedure or other

game-theoretic models.
The basic idea of the ordinal optimization method includes the following
steps:
1. Generating a nominal bid curve by using Lagrangian relaxation method
for hydrothermal scheduling;
2. Generating N bid curves by perturbing the nominal bid curve and obtain-
ing the MCPs associated with these bids;
3. Generating a “good enough” select bid set by evaluating these N bids us-
ing a very crude model with little computation effort, and ranking and se-
lecting them based on ordinal optimization; and
4. Evaluating the bids in the select bid set using an accurate model and solv-
ing the computationally time-consuming hydrothermal scheduling prob-
lem (2), and then selecting the best one.
The above four steps are briefly described next (see [23] for details).
It should be noted that reflecting the transmission constraints of the power
grid in bids is very difficult and has not been considered in the above bid se-
lection procedure. In fact, a unit can provide must-run service when the
transmission grid is limited by its transfer capability versus providing energy
and other services. Resource allocation or asset allocation among different
energy and service markets is an important issue that will be partially ad-
dressed in the next section.
3.1 Generating the Nominal Bid Curve

A nominal bid curve should serve two purposes: 1) it is a basis in (5) to
define the relationship between bid curves and the MCP; and 2) it is an opti-
mal bid curve if bidding strategy has no influence on the MCP. That is, if E is
a “price-taker,” the nominal bid curve should maximize E’s profit for any
MCP determined by the market. To achieve this goal, let a set of MCPs be
given as bid prices and the optimal generation for an indi-
vidual unit to maximize its profit can be obtained based on (6):
subject to the individual operating constraints, where k is the point index on a

bid curve. For the given series of (6) is a subprob-
lem within the unit commitment or hydrothermal scheduling context when
solved using the Lagrangian relaxation technique, where can be viewed
as the Lagrange multipliers given by the high level problem. The problem can
thus be efficiently solved by dynamic programming as in our previous work

[24-25]. The results are the optimal generation levels for all units at
each hour. The nominal bid curves for individual units and the aggregated bid
curve for E are thus generated as
and
Based on the procedure where the nominal bid curve is created, we see that if
the MCP determined by the market is equal to the generation award
would maximize the profit of an individual unit as in (6).
3.2 Perturbing the Nominal Bid Curve

The nominal bid curve given above is perturbed to generate N bid curves
as
where is a perturbation function. A simple way to implement

is to keep the power generation of a bidding generation point the
same and uniformly sample in the neighborhood
so that always specifies a monotonically

increasing piece-wise linear bid curve as required by the PX. Based on a per-
turbed bid curve the corresponding estimated by using
(5) can be obtained.
3.3 Selecting Good Bids
The N bid curves obtained in (7) can be evaluated and ranked by ordinal
optimization. The estimated profit of each set of bid curves is calculated as
T I
Note (9) is just a rough profit evaluation with the MCP given but without
solving scheduling problems. It is assumed that the generation award to a unit

based on its bid curves will also be delivered by that unit. This may not be
true since the generation award to a unit may not even satisfy the inter-
temporal individual operating constraints. A generation company will sched-
ule or re-allocate all its units to meet its total generation award with the mini-
mal total generation cost for delivering its total obligation and satisfying every
individual operating constraint of its units. The true profit can be obtained
by solving (2)-(4), and there may be significant error due to the above ap-
proximation
where is error. The advantage of the ordinal optimization method is its

capability to separate the good from bad even with a very crude model,
namely, the performance “order” is relatively immune to large approximation
errors. Even if the rough estimation is used to rank N bid curves, some good
enough bids will be kept within the select set with high probability.
The major task in applying ordinal optimization is to construct the se-
lected subset S containing “good enough” bids with high probability, includ-
ing the determination of its size The quantitative measure of the “good
enough” is the alignment probability defined as
where G is the “good enough” bid set and a is called the alignment level. In-
tuitively the alignment probability is the probability of the event that there are
at least a elements in the good enough set G matched in the select set S.
To select s good ones from the N perturbed bids generated by (7), the
profits are estimated by (9) and ranked. The top s bids are then selected as S
and its size s is determined by a regressed nonlinear equation to satisfy certain
confidence requirement ([27]). The value of s can be estimated by
where is the size of G and and are coefficients or parameters

obtained by nonlinear regression.
Evaluating N bidding strategies using (9) is computationally efficient and
the ordinal optimization method can guarantee that good enough bids will be
among the s selected strategies. The result of [27] tells us how large s should
be. We apply a more accurate, but time-consuming, evaluation that requires
solving generation scheduling problems to evaluate the s selected bids. For
each bid with the generation award, a traditional generation scheduling or unit
commitment problem described by (2) and (3) is solved to calculate E’s profit
using the Lagrangian relaxation-based algorithm in ([24-25]). The best bid is
then selected from by evaluating those bids in the subset S. Since is
much smaller than N, the ordinal optimization method is extremely efficient
in comparison with the brute force method of solving the N scheduling prob-
lems.
3.4 Numerical Testing

We perform numerical testing is performed for the California day-ahead
energy market based on a 10-unit generation company and historical PX mar-
ket clearing prices as the nominal MCPs. For testing purposes, we use the
published MCPs for May 1, 1998, and January 4, 1999, on the California PX
day-ahead energy market as the nominal MCPs in (5). The bidding influence
function in (5) is simplified as a linear function just to dem-
onstrate the effectiveness of the ordinal optimization method. The parameters
for ordinal optimization are selected as follows.
Search space size: N= 1000;
Alignment probability:
Good enough set: G = top 50 bids among N bids, i.e., g = 50;
Alignment level:
To observe “order is relatively immune from error” claimed by the ordi-
nal optimization method, the size of intersection of S and G, is listed
in Table 1. The ordinal optimization method tells us that at least 5 bids in S
should be also be in G. It is observed from Table 1 that sizes of the intersec-
tions of G and S are greater than 5 for all cases. Therefore, the good enough
bids will not “slip away” from select set S because of very crude approxima-
tion and the ordinal optimization method is very effective. The testing results
also show that for the four cases tested the best bid is selected among the N
perturbed bids. The computational time is reduced from 11 hours to about 40
minutes in comparison with the brute-and-force evaluation method. For de-
tails see [14].
4. INTEGRATED BIDDING AND SCHEDULING

WITH RISK MANAGEMENT
This section presents an integrated bidding and scheduling problem for
maximizing the profit of a power supplier while reducing bidding risks. The
method is generic, but the problem considered here is for the New England
market. As mentioned in the introduction, the MCP may be volatile in view of

the underlying uncertainties in the market. It is, therefore, important to deal
with the volatility and to manage the risks. Another issue is how to allocate
the generating capacity among different markets such as the energy and re-
serve markets.
First we present a formulation of integrated bidding and scheduling. We
assume the “perfect market” where the MCP is not affected by any single bid
and assume the MCP series to be a Markov chain. The risk management is
explicitly modeled, and the reserve market and the self-scheduling require-
ments are also considered. The problem is formulated as a mixed-integer sto-
chastic optimization problem with separable structure. To solve this problem,
we develop a Lagrangian relaxation-based method to decompose the problem
into a number of subproblems, one for each unit. Then we apply stochastic
dynamic programming to solve individual unit subproblems, and generate a
set of generation strategies: how much power should each unit provide for
each possible market price value at what probability? We then create the bid
curves based on these generation strategies. Numerical testing base on an 11-
unit system including a large pumped-storage unit shows the algorithm can
generate good energy and reserve bid curves in 4-5 minutes on a P-III/600
PC. The results also demonstrate how a pumped-storage unit affects the bid-
ding of thermal units and how the risk management and the reserve market
affect the bidding in the energy market.
4.1 Problem Formulation

Consider a utility company in New England with I generators of thermal,
hydro, and pumped-storage units. The problem is to select bid curves for indi-
vidual units to maximize the profit while reducing risks. The formulation in-
volves market assumptions, the objective function to be minimized, and the
constraints to be satisfied.
In this section, the “perfect market” assumption is assumed, i.e., the MCP
is not affected by a single market participant. Since there exist various time-
dependent operating constraints such as minimum up-/down-times and ramp
rate limits, the bid curves and thus the MCPs between hours are coupled. To
model time dependence of the MCPs, the MCP series is assumed to be a
Markov chain. This means that the MCP at hour only depends on that of
the previous The time independent MCP series can be consid-
ered as a special case of the Markov chain. In practice, a “perfect market”
may not exist, and the MCP is usually affected by the bids. This issue will
also be discussed in this section.
As mentioned in several previous sections, the bid curve for unit
i at time t is a power-price function indicating how much power the company

wants the unit to generate at any given price. Based on the bid curve, if the
MCP is then unit i will be awarded the amount of power
Therefore, to determine the bid curve for unit i is to decide a generation level
for any possible price
The form of bid curves may be different in different markets. In the New
England market, up to 10 power blocks, each with a price, form a bid. Similar
to the procedure of generating the nominal bid curve presented in section 3.1,
for each hour are given based on the MCP forecast-
ing. The transition probabilities between and are also known
and will not be affected by any particular bidder. The same is also assumed
for the reserve prices We can construct the bid curves by obtaining the
optimal generation for each given (k= 1, 2, ..., K), and
the optimal reserve for each reserve price (j=1, 2, ..., K) for
each unit i and any hour t.
The utility may have its “own load” and the associated reserve re-
quirement R(t) at hour t. We ignore the own load uncertainties since they are
usually accurate to within 2% and are less significant than the MCP uncertain-
ties. As a strategic decision for reducing the risks, the company may want to
cover by itself, on an average, at least a certain percentage of its own load and
reserve requirements. This “self-scheduling requirement” can be formulated
as the following constraints on expected and reserve price
and
where and are the self-scheduling coefficients. The algo-

rithm would be just simpler for the cases without considering the self-
scheduling requirements.
Currently there is no demand market in New England. Therefore, if the
MCP at hour t is and the reserve price is then
is the amount of the energy that the company has to buy (positive) from or sell
(negative) to the market at and the amount of reserve
the company has to buy (positive) or sell (negative) at The profit equals
the revenue from the energy and reserve markets minus the operation costs.
The objective is to maximize the expected profit or equivalently to minimize
negative expected profit, which requires the utility to sell power at hours with
high prices and buy power at times with low prices. With known transition
probabilities of market prices, the MCP variances and reserve price
and reflecting the market uncertainties at hour t can be calculated. Ac-
cording to historical observation, the MCP may jump from its normal value of
20 or 30 dollars to over 1000 dollars per megawatt hour. A company usually
prefers to sell power when the market has large uncertainty and buy power
when the market has low uncertainty so as to avoid risks. Therefore, bidding
risks can be reduced by penalizing the product of price variances and the pur-
chased amount of energy. Combining the above analysis, the objective to be
minimized is a weighted sum of the negative profit (the first part) and the risk
terms (the second part), i.e.,
where is a weight to balance the profit versus risks.

The problem is also subject to individual unit constraints such as mini-
mum up-/down-time and ramp rate limits for thermal units, and reservoir dy-
namics and volume limits for hydro and pumped-storage units. The con-
straints for thermal and hydro units have been presented in [24, 25, 28], and
those for pumped-storage units were presented in [26] and [29].
4.2 Solution Methodology

The above integrated bidding and scheduling problem is a mixed-integer
stochastic optimization problem. It is very difficult to obtain the optimal solu-
tion in view of its computational complexity. One way to solve the problem is
to perform scenario analysis. Since the number of scenarios may be very large,
scenario analysis is inefficient. In view of the separable structure of the prob-
lem, Lagrangian relaxation is a very efficient approach, and we apply it to
solve the problem.
By taking (14) and (15) as the demand and reserve requirements and re-
laxing them using two sets of multipliers and respectively, a
two-level optimization is formed. The high level is to update the multipliers
so as to maximize the dual. It is formulated as
where
and
Given multipliers and the low level problem consists of I

subproblems, one for each generator. The subproblem i is formulated as
min
subject to individual unit constraints.

These two levels iterate until a convergence criterion is achieved. Based
on the subproblem solutions, the bid curves are constructed after the dual pro-
cedure terminates. We briefly describe these steps next. For details, refer to
[29].
4.2.1 Solving Subproblems
The above subproblem (20) is similar to that of a hydrothermal schedul-

ing problem as in [24, 25, 28], where and can be interpreted as the
energy and reserve marginal costs of the system at hour t, respectively. How-
ever, and are random variables depending on market prices and
The subproblem solution is a set of strategies rather than a generation
level for each hour. It establishes a relation of and the reserve
price with a probability to generate a certain quantity.
The subproblems for different type units are different. For example, there
are usually no fuel cost and start-up cost for hydro and pumped-storage unit
subproblems. In view of the combinatorial nature and the large number of
possible MCP realizations, it is hard to solve the subproblem (20). Based on
the transition probabilities of a Markov chain, a stochastic dynamic program-
ming (SDP) is developed, in which a stage corresponds to an hour, and a state
at hour t is associated with a pair of and (k, j= 1, 2, ..., K).
Constraints that couple the decisions at different hours, such as ramp rate and
pond limit constraints, are relaxed by using an additional set of Lagrangian
multipliers so that the problems are stage-wise decomposable. We then obtain

the optimal generation level and reserve level for
each state by optimizing the stage-wise cost function, subject to feasible op-
eration regions. Since the coefficients of the stage-wise cost function are de-
termined by and depends on and and is gen-
erally an increasing function of The probability for each
state is calculated based on the transition probabilities of the Markov chain.
The subproblem solution here is to provide a set of strategies: for each pair of
and each unit should provide amount of power
with probability We will use these strategies in constructing
bid curves.
4.2.2 Solving the Dual Problem
After solving the subproblems, the sub-gradient associated with (14) can
be obtained as
Then the multipliers can be iteratively updated by using a sub-gradient

or bundle method. The updating of is similar.
4.2.3 Generating and Selecting Bid Curves
After the dual procedure converges, the bid curves for each unit are con-
structed based on the strategies obtained in the subproblem solutions. First,
we calculate the amount of power to be submitted for each (k=1, 2, ...,
K) as the expectation of over for j=1, 2, ..., K. Since the
operation regions of a unit may be discontinuous, the result obtained is then
projected into a feasible operation region to get the optimal power amount.
This procedure is
Then we construct the bid curve for unit i at home t. As mentioned, the bid
curve required by the New England market is a staircase function. The size of
its power block equals and the price is as
The reserve bid curves can also be constructed similarly.

In view of the existence of integer decision variables such as the unit
commitment decision, the maximum dual function does not mean the best
primal solution. Therefore, the ordinal optimization described in Section 3 is
applied to sort the bid curves generated at the last several iterations. The best
one is selected as the final set of bid curves.
4.2.4 MCP Influenced by Bidding Strategies
In the above, we model the MCP series as a Markov chain under the “per-
fect market.” As mentioned in section 3, however, the MCP may be influ-
enced by the bids to be determined. To estimate the impact of bidding behav-
iors on the MCP, a neural network model for predicting the MCP is under
development. The idea is that in addition to the market information such as
forecasted market demand and MCP lags [30], the aggregate bid curves are
also part of the inputs to the neural network, and the sensitivity of the MCP
versus the bid curves can thus be analyzed. This model is to be integrated with
the above algorithm and the MCP distribution will be updated based on the
bid curves generated at each iteration.
4.3 Self-scheduling, Market Interactions, and Risk

Management
Since plays the role of marginal energy cost and the marginal re-
serve cost, their values determine the bid curves for each unit. From (18)-(20)
and (22), one can see the impact of self-scheduling constraints and risk man-
agement on bidding curves, and the interaction of the energy and reserve mar-
kets.
The self-scheduling constraints couple the different units via and
and affect the marginal and as in (18) and (19). For exam-
ple, if the self-scheduling requirement at hour t is high, will be large,
causing large Consequently, the units have to bid the same amount of
energy at a lower price to satisfy the self-scheduling constraints.
The tradeoff in allocating the limited generation capacity in the energy
and reserve markets depends on the market prices and In the normal
situation, where is much higher than the profit largely depends on
the MCPs for energy. Occasionally, when the reserve price is relatively
high as observed in many markets, more capacity should be allocated to the
reserve market. This may result in significant changes in the energy bidding
strategies, especially for pumped-storage units. A pumped-storage unit pro-
vides large reserves at either pumping or generating. In view of the reservoir

dynamics and limits, the bidding strategies are coupled across hours. That is,
decisions in one hour will affect results of other hours. Furthermore, the
pumped-storage bidding strategies also affect thermal bidding strategies in
view of self-scheduling constraints. Therefore, the pumped-storage units play
a key role in both the energy and reserve markets.
The risk management affects bidding strategies through and re-
lated to the variances and If is high, uncertainty on is
large and is large. Then the optimal generation level ob-
tained in DP is large, resulting in large as in (22). As a result, the
units will bid the same amount of energy at low prices so that large amounts
of energy could be sold on the market or buying large amounts of power at
potential high MCP could be avoided.
4.4 Numerical Testing

We perform numerical testing for the New England energy and reserve
markets based on a system with 10 thermal units and a 4-identical-generator
pumped-storage system with a large reservoir and the historical market prices.
To investigate the influence of the pumped-storage units on thermal bidding
strategy, we test the pure thermal system. To investigate the influence of the
reserve market on the bidding strategies, we also perform testing where the
reserve market is ignored. The testing results are summarized below. For de-
tailed analysis refer to [29].
The results show that the pumped-storage unit plays important roles in
both the energy and reserve markets and can significantly change the thermal
bid curves. For example, they pump at hours with low MCPs, making the
thermal units bid at a low price to satisfy the high self-scheduling require-
ments and generate at hours with high MCPs, reducing the self-scheduling
requirement for the thermal units.
Risk management proves to be an effective way to avoid buying large
amounts of power from the market at potentially high MCPs and to reduce the
profit variances. In section 4.3, we explained why the bid price should be low
when the MCP variance is high. In this case, it is observed that the generation
award is large.
The significant influence of the reserve market on the bidding strategies is
also observed. Figure 2 shows the comparison of pumped-storage bid curves
for hour eight on July 12, 1999, with and without considering the reserve
market. The reserve market is considered in Case 1, but ignored in Case 2.
Figure 3 gives the MCPs and reserve prices. According to the market rules,
the reserve contribution of a pumped-storage generator is same as the pump-

ing level or the one-line generation capacity minus the current generation
level. In view of the high reserve price at hour eight, each generator of the
pumped-storage system should either pump at its capacity or generate at its
minimum generation limit so as to provide maximum reserve for maximizing
the profit. Figure 2 depicts these two bidding strategies, and the bid curves are
significantly different.
The CPU time of the algorithm is about four to five minutes on a 600 Hz
Pentium III PC. Therefore, the algorithm is efficient for daily bidding and
scheduling.
5. CONCLUSIONS AND ON-GOING WORK

In this research, we developed effective methods for dealing with the new
challenging problems in deregulated electric power markets. We created an
ordinal optimization based bidding strategy to select “good enough” bids for a
power supplier. Numerical results show this method is efficient and can yield
good bids. We developed a stochastic optimization approach for integrated
bidding and scheduling with the consideration of risk management, self-
scheduling requirements, and interaction between energy and reserve markets.
Numerical testing results based on an 11-unit system of the New England
market show that the algorithm is efficient for daily bidding and scheduling.
The bidding strategies of pumped-storage units have significant influence on
the bidding strategies for thermal units and play important roles in both the
energy and reserve markets. On-going work includes simulation and forecast-
ing of the market indicators such as the MCP, game theoretic modeling and
analysis of the MCP price spikes and bidding strategies, and integration of
forecasting and bidding methods.
ACKNOWLEDGEMENTS
The research reported in this chapter is supported in part by EPRI/ARO
Contract WO833-03, National Science Foundation under Grant ECS-
9726577; National Outstanding Young Investigator Grant 6970025; and a
Key Project Grant 59937150, National Science Foundation of China and 863
Project of China. The authors would like to thank Dr. David Pepyne of Har-
vard University for his valuable insight and comments.
REFERENCES
1. W. Dunn, M. Rossi, and B. Avaramovic. Impact of market restructuring on power systems
operation. IEEE Comput. Appl. Power Engr., 8(1): 42-47,1995.
2. G. Gross and D. J. Finlay. “Optimal Bidding Strategies in Competitive Electricity Mar-
kets.” In Proc. Power Systems Computation Conference, Dresden, 1996, pp. 815-822.
3. A. Farrokh and Z. Alaywan. California ISO formation and implementation. IEEE Comput.
Appl. Power Engr., 12(4): 30-34,1999.
4. X. Guan and P.B. Luh. Integrated resource scheduling and bidding in the deregulated elec-
tric power market: New challenges. Special Issue, J. Discrete Event Dyn. Syst. 9(4): 331-
350, 1999.
5. P.A. Fedora. “Development of New England Power Pool’s Proposed Markets.” In Proc.
Hawaii International Conference on System Sciences, Maui, Hawaii. 1999.
6. K.W. Chenug, P. Shamsollahi, and D. Sun. “Energy and Ancillary Service Dispatch for
the Interim ISO New England Electricity Market.” In Proc. International Conference
on Power Industry Computer Applications, Santa Clara, California, 1999, pp. 47-53.
7. W. Mielczarski, G. Michalik, and M. Wildjaja. “Bidding Strategies in Electricity Mar-

kets.” In Proc. 21th International Conference on Power Industry Computer Applications,
Santa Clara, California, 1999, pp. 71-76.
8. T. J. Overbye, J.D. Weber, and K. J. Patten. “Analysis and Visualization of Market Power
in Electric Power Systems.” In Proc. Hawaii International Conference on Systems
Science, Maui, Hawaii, 1999.
9. C.A. Berry, B.F. Hobbs, W.A. Meroney, R.P. O’Neill, and W.R. Stewart. “Analyzing
Strategic Bidding Behavior in Transmission Networks.” In IEEE Tutorial on Game The-
ory Applications in Electric Power Markets, TP-136-0: 7-32, 1999.
10. Owen G. Game Theory, Third Edition, Academic Press, 1995.
11. V. Krishna and V.C. Ramesh. “Intelligent Agents for Negotiations in Market Games, Part
I and Part II.” In Proc. International Conference on Power Industry Computer Appli-
cations, Columbus, Ohio, 1997, pp. 388-399.
12. R.W. Ferrero, S.M. Shahidehpour, and V C. Ramesh. Transaction analysis in deregulated
power systems using game theory. IEEE Trans. Power Syst., 12(3): 1340-1347, 1997.
13. H.Y. Yamin and S.M. Shahidepour. “Risk Management Using Game Theory by
Transmission Constraint Unit Commitment in a Deregulated Power Market.” In IEEE
Tutorial on Game Theory Applications in Electric Power Markets, TP-136-0: 50-60, 1999.
14. T. Alvey, D. Goodwin, X. Ma, D. Streiffert, and D. Sun. A security-constrained bid-
clearing system for the New Zealand wholesale electricity market. IEEE Trans. .Power
Syst., 13(3): 986-991, 1998.
15. S. Dekrajangpetch, G.B. Sheble, and A.J. Conejo. Auction implementation problem using
16. G.B. Sheble. “Decision Analysis Tools for GENCO Dispatchers.” IEEE/PES Summer
Meeting, San Diego, California, 1998, PE-231-PWRS-0-06-1998.
17. C. Li, A. Svoboda, X. Guan, and H. Singh. Revenue adequate bidding strategies in com-
petitive electricity markets. IEEE Trans. Power Syst., 14(2): 492-497, 1999.
18. D. Zhang, Y. Wang, and P.B. Luh, “Optimization-Based Bidding Strategies in the
Deregulated Market.” In Proc. International Conference on Power Industry Computer
Applications, Santa Clara, California, 1999, pp. 63-69.
19. S. Hao. “A Study of Basic Bidding Strategy in Clearing Pricing Auctions.” In Proc.
International Conference on Power Industry Computer Applications, San Clara, Califor-
nia, 1999, pp. 55-60.
20. A. Cohen and V. Sherkat. “Optimization-Based Methods for Operations Scheduling.” In
Proceedings of IEEE, 75(2): 1574-1591,1987.
21. Y. C. Ho. “Soft Optimization for Hard Problems.” Web-distributed copy of a tutorial for
ordinal optimization, http://www.hrl.harvard.edu/people/faculty/ho/DEDS, 1997.
22. Y.C. Ho. On the numerical solution of stochastic optimization problems. IEEE Trans.
Autom. Cont, 42(5): 727-729, 1997.
23. X. Guan, Y.C. Ho, and F. Lai. An ordinal optimization-based bidding strategy for electric
power suppliers in the daily energy market. Submitted, 1999.
24. X. Guan, P.B. Luh, H. Yan, and J.A. Amalfi. An optimization-based method for unit
commitment. Int. J. Electric Power Energy Syst., 14(1): 9-17, 1992.
25. X. Guan, E. Ni, R. Li, and P.B. Luh. An algorithm for scheduling hydrothermal power
systems with cascaded reservoirs and discrete hydro constraints. IEEE Trans. Power Syst.,
12(4): 1775-1780, 1997.
26. X. Guan, P.B. Luh, H. Yan and P.M. Rogan. Optimization-based scheduling of hydro-
thermal power systems with pumped-storage units. IEEE Trans. Power Syst., 9(2): 1023-
1031, 1994.
27. T.W.E. Lau and Y.C. Ho. Universal alignment probabilities and subset selection for ordi-
nal optimization. J. Optim. Theory Appl., 39(3): 455-490, 1997.
28. E. Ni, X. Guan, and R. Li. Scheduling hydrothermal power systems with cascaded and
head-dependent reservoirs. IEEE Trans. Power Syst., 14(3): 1127-1132, 1999.
29. E. Ni and P.B. Luh. “Optimal Integrated Bidding and Scheduling for Hydrothermal Power
Systems with Risk Management and Self-Scheduling Requirements.” To appear in Proc.
World Congress on Intelligent and Control, Hefei, Anhui, China, 2000.
30. F. Gao, X. Guan, X. Cao, and A. Papalexopoulos. “Forecasting Power Market Clearing
Price and Quantity Using a Neural Network Method.” To appear in Proc. 2000 IEEE/PES
Summer Meeting.
31. C. Richter, G. Sheble’, and D. Ashlock. Comprehensive bidding strategies with genetic
programming/finite state automata. IEEE Trans. Power Syst., 14(4): 1207-1212, 1999.
Chapter 15
DECENTRALIZED NODAL-PRICE
SELF-DISPATCH AND UNIT COMMITMENT
Francisco D. Galiana and Alexis L. Motto

McGill University
Antonio J. Conejo
University of Castilla-La-Mancha
Maurice Huneault
Hydro-Quebec Research Institute (IREQ)
Abstract: This chapter sets forth a scheme for self-scheduling independent market partici-
pants in a power pool. The approach, named DNSA for Decentralized Nodal-
Price Self-Scheduling Auction, is proposed as an alternative to centralized Pool
auctions and operation. DNSA exploits the intrinsic parallelism of the dual unit
commitment problem to decentralize the various scheduling and dispatch func-
tions. Each competing participant (GENCO, DISTCO) maximizes its profit for
any set of nodal prices by choosing its level of production or consumption.
Similarly, the TRANSCO independently maximizes its merchandising surplus
within the network security constraints. The price caller, a centralized entity
without access to proprietary cost information, updates prices through an effec-
tive Newton algorithm until the power balance at each bus is satisfied. DNSA
does not assume a perfect market and accounts for the AC load flow model in-
cluding transmission losses and line congestion, in addition to integer variables,
ramping rates, start-up costs, and minimum up and down times. The conver-
gence of DNSA hinges on the notions of profit optimality and the convexifying
market rule. We present several study cases to illustrate the characteristics of
DNSA. We conclude that to achieve fairness of treatment for all competing par-
ticipants, they should be allowed to optimize their profit by self-scheduling.
Therefore, to the extent possible, the next generation of unit commitment mod-
els should include profit optimality.
1. INTRODUCTION
This chapter summarizes a number of research activities conducted at
McGill University on the decentralized self-operation of electricity market
participants, namely, GENCOs, DISTCOs, and TRANSCOs. A main goal of
this research is to demonstrate that, under certain conditions, self-operation
based on maximizing individual profits yields the same electrical and eco-
nomic operating point as centralized maximum welfare operation. This
equivalence is known for simple cases of economic dispatch, but is less obvi-
ous for more complex scheduling problems including unit commitment, in-
tertemporal constraints, nonlinear transmission losses, flow congestion, and
other network constraints. In particular, we will show that the equivalence
between self- and centralized operation is also valid for conditions other than
a perfect market.
The proof of equivalence is based on purely mathematical arguments un-
der a main assumption, here named “profit optimality,” by which we mean
that all competing participants are free to maximize profit subject only to
market prices. Without profit optimality, the centralized solution may require
some individual participants to operate at less than maximum profit, an un-
avoidable consequence of the difficulty of defining a measure of social wel-
fare.1 Therefore, we argue that market rules can be considered unfair if they
can lead to solutions where some but not all participants are dispatched at
maximum profit. Such a possibility must be precluded by the market rules
since equality of treatment is of paramount concern among competing entities.
Accordingly, in our view, profit optimality is not a theoretical issue but a de-
sirable restriction of the next generation of unit commitment problems.
Under profit optimality, the dual and primal solutions of the mixed integer
unit commitment problem are shown to be identical. This then leads to the
proposed decentralized, nodal-price self-dispatch and scheduling auction,
DNSA, the essence of which is to optimize the dual function while simultane-
ously meeting the relaxed constraints of the primal problem. The DNSA com-
prises some key innovations centered on the application of an iterative auc-
tion. These are the introduction of a central price caller that updates and
broadcasts trial prices, the delegation of the self-scheduling tasks to the par-
ticipants who then respond to the trial prices, the use of a fast Newton price-
updating scheme, and the consideration of the full set of non-linear network
constraints.
This chapter is organized as follows: after the nomenclature, we compare
the general features of centralized operation to those of decentralized self-
operation, in particular under the proposed DNSA scheme. Next, the unit
1
See [1] for a compact summary on Arrow’s impossibility theorem. Kenneth J. Arrow was a
1972 Nobel laureate in economic science.
Decentralized Nodal-Price Self-Dispatch and Commitment 273
commitment problem primal and dual forms are formulated and shown to
yield identical solutions under the condition of profit optimality. The DNSA
scheme is then described and compared to the classical centralized dual-based
unit commitment. DNSA requires two conditions to converge, the previously
defined profit optimality, together with the “convexifying” rule imposed to
ensure rational participants’ behavior in response to nodal prices. We then
describe the Newton-based price-updating component of DNSA. Finally, a
number of numerical results illustrate the characteristics of DNSA.
2. NOTATION
For quick reference, we classify below the main mathematical symbols
used throughout this chapter.
feasible space of demand i

feasible space of generation i
minimum down time of generator i
minimum up time of generator i
set of network nodes
average real power injection at bus i during period
average real power of load i during period
average real power of generator i during period
profit of load i over the scheduling horizon
profit of generator i over the scheduling horizon
Benefit function of load i
ramp-down limit of generator i
ramp-up limit of generator i
shut-down cost of generator i occurring at the beginning of period
start-up cost of generator i occurring at the beginning of period
set of time intervals in scheduling horizon
0/1 variable, which is equal to 1 if generator i is committed during
period or 0 otherwise
0/1 variable, which is equal to 1 if load i is committed during period
or 0 otherwise
0/1 variable, which is equal to 1 if generator i is started up at the
beginning of period or 0 otherwise
0/1 variable, which is equal to 1 if generator i is shut down at the

beginning of period or 0 otherwise
average voltage phase angle at node i during period
cumulative up/down time of generator i at the start of period
average nodal price of active power at node i during period
the vector over all time intervals for node i
the vector over all nodes for time
the vector all nodes and all time intervals
3. MOTIVATION FOR SELF-OPERATION
3.1 Current Practices

The emerging restructured power industry is guided towards two generic
trading models: pool and bilateral markets [3,4].2 In the pool context, a central
agent collects the bids from sellers (GENCOs) and buyers (DISTCOs), deter-
mining the winners and the amount of power each is required to sell or buy as
well as the market-clearing prices. Generally, this selection results from a cen-
tralized constrained optimization. In contrast, the bilateral model allows par-
ticipants to arrange contracts among themselves. Though the transmission
monopoly may create unforeseeable challenges to the implementation of bi-
lateral transactions, some authors argue that such contracts offer participants
more freedom, thereby achieving greater decentralization in decision-making
[5]. In view of their widespread use, the scope of this chapter is limited to
pure pool-operated markets.
Figure 1 illustrates the state of the art of today’s Pool operation. The mar-
ket operates with three independent entities, GENCOs, DISTCOs, and the
ISO.3 Under this model, strictly speaking, there is no independent transmis-
sion provider, as the network is under the authority of the ISO. The latter sets
the pool price (or nodal prices) by trying to maximize the overall social wel-
fare, irrespective of the participants’ revenue or profit requirements. Clearly,
the ISO has an important and often coercive decision-making authority, a
practice inherited from the traditional centrally controlled utility industry.
2
The central buying entity for all suppliers of electricity, which in turn is the single agent for
selling power to retail customers and their aggregators [2].
3
We use the term ISO to refer to the combined power exchange and system operator.
3.2 On Decentralizing Pool Operation

Reduced reliance on the decisions of a centralized scheduling and dis-
patching entity is an alternative that some electricity market participants al-
ready actively seek [6,7,8]. Centralized authority has been brought into ques-
tion for both technical and philosophical reasons. For example, it is accepted
that solutions to optimal power flow [9] and unit commitment [10] programs
can be highly sensitive to small variations in the input parameters and to the
algorithm heuristics. Whereas the sensitivity of the objective function to pa-
rameter variations is usually insignificant, that of the individual participants’
responses can be more substantial, a potentially unfair result. Philosophically,
it can also be argued that, in a competitive environment, the decision as to
whether to trade or not and how much to trade must rest solely with each
trader. Furthermore, such a choice should be guided by each participant’s pri-
vate and confidential profitability expectations as well as on the market condi-
tions. For this decision-making independence to function in the context of a
power system, market prices alone must offer sufficient incentives to satisfy
all network security and power balance requirements without the need for
centralized intervention.
One alternative to the centralized pool with the above characteristics is
the DNSA scheme put forward in this chapter. DNSA is an auction mecha-
nism [11,12] that allows each independent participant to self-commit and dis-
patch based on its own profit evaluation (see Figure 2).
It is instructive to compare the dual communication in DNSA between

each participant and the auctioneer – here the Price Caller – to the one-way
relation between participants and the ISO in the centralized auction of Figure
1. Under DNSA, convergence to equilibrium involves repeated interaction of
all participants, with the Price Caller updating trial prices and the independent
entities responding accordingly.
Note that the DNSA differs from other attempts to decentralize the au-
thority of the ISO. For example, in the approach of Griffes [13], GENCOs and
DISTCOs are allowed to resubmit new bids to the ISO if the previous sched-
ule is not to their satisfaction. In that approach, however, for each new set of
bids the ISO has to repeat a fully centralized scheduling operation.
The DNSA, in a similar fashion as an iterative Lagrangian relaxation al-
gorithm [14,15], exploits the decomposability of the unit commitment prob-
lem in its dual form to achieve the goal of decentralization and self-
scheduling. The next section analyzes the conditions under which dual and
primal methods yield the same solutions.
4. THE UNIT COMMITMENT PROBLEM

Traditionally, the unit commitment problem develops the on/off schedule of
generating units (loads) over an operating horizon. Once committed, a genera-
tor (load) is synchronized to the grid and is ready to deliver (consume) power.
The dispatch problem consists of determining the levels of production (con-
sumption) of the committed generators (loads). In practice, however, these
two problems are solved concurrently and often referred to as the unit com-
mitment problem.
4.1 Primal Problem

The primal unit commitment problem [16], abbreviated here as PUC, can be
formulated as the maximization of the social welfare,
subject to the constraints,
The objective function (1) includes generation costs, and con-

sumer benefit functions, The total operation cost, in ($/h)
incurred by Genco(i) can be written as,
where, is the fixed cost, is the variable cost, is the

4
start-up cost, and is the shut-down cost. Note that cost and benefit
data are bids.
Constraints (2) implement the real power balance equation at each net-
work node, line power flow limits are enforced by (3), consumption limita-
tions such as budget constraints are represented by (4), while generator con-
straints are contained in (5). The symbols and represent the feasible
regions of the variables and respectively. The set in-
cludes initial conditions, (7), output limits, (8), minimum up/down times, (11)
–(12), and ramp-up/down limits, (9)–(10), that is,
4
Refer to [17] for a formal discussion of cost and consumer benefit functions.
5
A similar set of constraints could apply to
A few caveats must be made at this point. For the sake of clarity, we
made the decision to simplify some of the models. We have attempted, how-
ever, to introduce the results in such a way that generalizations are apparent.
Thus, the network optimization variables have been limited to the voltage
phase angles, while still retaining the full nonlinear AC load flow. Reserve
constraints and multiple-generator (-loads) GENCOs (DISTCOs) have not yet
been incorporated in DNSA.
The implementation of PUC for a typical power system involves a large
number of 0/1 variables and numerous constraints. This problem is known to
be NP-hard and, unless drastic simplifications are made, seldom tractable.6
Furthermore, the complexity of the models that are solved increases under
competition since profit-driven agents have no choice but to model their sys-
tems with additional variables in order to achieve greater accuracy. In order to
apply integer programming [18], in our simulations all variables are constant
during the discrete time intervals and all functions are approximated by dis-
crete piecewise-linear mappings [19,20,21].
4.2 Dual Problem

In contrast to the primal, experience demonstrates that dual methods such
as Lagrangian Relaxation are quite successful at solving the unit commitment
problem [22,23,15], notwithstanding the heuristic approximations that be-
come necessary whenever the duality gap is non-zero.7
In the dual approach, relaxing the power balance constraints at every bus
and for every time period yields the Lagrangian function,
6
A problem is said to be NP-hard if the zero-one integer programming problem can be mapped
to it in polynomial time.
7
In such cases, the primal solution is sub-optimal and can dispatch some participants at less
than the maximum possible profit for the market prices.
from which the dual unit commitment problem (DUC) can be written as
By the weak duality theorem, the dual value is a lower bound to the
primal value defined by equation (1). Regardless of the structure of the objec-
tive and constraints of the primal problem, the domain of the dual function
is convex and the function is concave over These elegant convexity
properties combined with the problem decomposability allow a solution of the
dual problem consisting of several reduced sub-problems that can be solved in
parallel as outlined next.
4.3 Self-Operation Sub-Problems

The three components in the right-hand-side of equation (18) can be in-
terpreted as monetary objectives to be maximized by each type of entity for
the given prices.
4.3.1 The GENCO Profit Maximization Sub-Problem
Given any set of nodal prices GENCO(i) maximizes its profit while satis-
fying its generation constraints, that is,
The arguments of this sub-problem are functions of the prices and are denoted
by and The above-defined maximum profit is non-negative,
that is, This result stems from the fact that since a GENCO can
control its commitment, it will always turn itself off over the entire time hori-
zon rather than operate at a loss.
4.3.2 The DISTCO Profit Maximization Sub-Problem
Given any set of nodal prices, DlSTCO(i) also maximizes its profit
while satisfying its load constraints, that is,
The arguments of this sub-problem are functions of the prices and are denoted
by and The maximum profit of any self-committing DISTCO
(with controllable 0/1 variable is also non-negative, that is,
4.3.3 The TRANSCO Merchandising Surplus Maximization Sub-

Problem
Assuming that a single Transco operates the network, this sub-problem con-
sists of maximizing the merchandising surplus over the planning horizon, sub-
ject to the various network security constraints, Thus,
The argument of this sub-problem is a function of the prices and is denoted by

a solution that is de facto security feasible, as the TRANSCO en-
forces security constraints on all network variables.
We emphasize that in (24) the TRANSCO does not physically adjust the
voltage phase angles or the power injections, it merely computes the injec-
tions that are then submitted to the Price Caller. What is perhaps surprising is
that giving the TRANSCO the freedom to maximize its merchandising sur-
plus is consistent with the primal problem objective of maximum social wel-
fare under the set of network constraints,
Even so, the abuse of monopoly power by a TRANSCO in the process of
maximizing its merchandising surplus is a possibility that must be addressed.
For example, manipulation of the line flow limits in could artificially cre-
ate transmission congestion and nodal price differences, thereby increasing
the merchandising surplus. We believe that this practice can be virtually ruled
out by making the TRANSCO data and response to prices public and open to
scrutiny by all market participants. This is consistent with the rulings of
FERC [24,25]. Furthermore, because its revenues and/or profits are regulated,
there is no incentive for a TRANSCO to create congestion artificially, unless
it were in an anti-competitive collusion with a GENCO or a DISTCO.
4.4 Profit Optimality
Let and be the main arguments of the pri-

mal problem (1) at its optimum. The corresponding profit of GENCO(i) is
Analogously, the profit of DISTCO(i) is
The TRANSCO surplus under the primal optimum conditions is given by
With these results, we can now define the notion of profit optimality.
DEFINITION 1 (Profit Optimality): A primal solution is said to be profit opti-

mum if,
Profit optimality states that the primal profit of GENCO(i), is

the maximum achievable profit, A similar statement applies to the
profit of DISTCO(i) and to the merchandising surplus of the TRANSCO.
4.5 Equivalence between Primal and Dual

PROPOSITION: Assume that PUC and DUC are each unique. DUC and
PUC yield identical solutions if and only if the solution of is
profit optimum.
PROOF:
Necessity: Let PUC and DUC have identical solutions, that is,
Since maximizes the individual profits in DUC
for the prices then also maximizes the same profits for Thus, the
solution of PUC is profit optimum.
Sufficiency: Let the solution to PUC, be profit optimum. We

show that PUC and DUC yield identical solutions. First calculate the dual
function at the primal solution prices, From (18)–(24),
Next, applying the profit optimality conditions in (28), and then refining with
equations (25)–(27),
Since the power balance at all nodes and times is guaranteed by the primal
solution, the second right-hand side term above disappears. Thus,
which is the value of the primal. This implies that at the primal prices, the
dual function is maximized and the duality gap is nil. Making use of the gen-
erally accepted assumption that the solutions of DUC and PUC are unique, it
then follows that these solutions are equal, that is, Q.E.D.
5. BASIS OF THE DECENTRALIZED NODAL-PRICE

SELF-SCHEDULING AUCTION (DNSA)
5.1 Profit Optimum Unit Commitment

We know (see Case C in Section 6) that the standard formulation of the
unit commitment problem, (1)-(5), can yield solutions that are not profit opti-
mum. Although mathematically feasible, such solutions may require some
participants to operate at a profit below the maximum possible for the current
prices, a result that we contend is unfair in the context of competition. In fact,
the current practice of “uplift” charges in Pool operation is a mechanism for
correcting such inequities, albeit in a heuristic and sub-optimal manner. To
preclude the potential of unequal treatment of participants’ profits and to do
so systematically, we assert that the next generation unit commitment formu-
lation should include the condition of profit optimality. This is a subject of our
ongoing research for both the centralized PUC and for DNSA.
5.2 Basic Steps of DNSA
DNSA differs from standard dual methods (SDM) in several respects:

On the centralized nature of the method: all scheduling and price updating
is centralized in SDM, whereas in DNSA, only price updating is centralized.
GENCOs, DISTCOs, and TRANSCO self-schedule.
On the requirements of price updating and commercial information: In
SDM, as price updating is based on maximizing the central scheduler
needs full knowledge of private cost and benefit functions. In DNSA, as price
updating is based on solving the Price Caller needs only a point-
wise price-response by GENCOs and DISTCOs.10
On the requirements of network information: In SDM, all TRANSCO
data is required by a central scheduler to assure network security. In DNSA,
the TRANSCO assures security. The Price Caller needs only a reduced set of
network data and sensitivities.
On the completeness of network modeling: In SDM, network constraints,
particularly those describing nonlinear behavior, have been often simplified or
ignored. In DNSA, most network constraints can be or have been included.
On the evolution of bids through the auction process: In SDM, bid func-
tions are fixed throughout the solution process. In DNSA, response to prices
may evolve during the price-updating trials subject to the “convexifying” rule,
which is defined next.
8
Here
9
The state-of-the-art methods for updating dual variables are the sub-gradient method [26], the
bundle method and the cutting plane method. In the current implementation of DNSA, a
Newton approach is used to update prices.
10
The Price Caller requires no explicit bid functions. To prove that the self-scheduling scheme
yields the same solution as PUC, however, we assume the existence of such functions. This
implies that GENCOs and DISTCOs behave according to some deliberate strategy.
5.3 Convergence of DNSA

We now address the question of equilibrium attainment, that is, if a com-
petitive equilibrium is feasible, in the sense that PUC is profit optimum, how
does the market attain it? The difficulty of this question is old and well known
to economists. Indeed, it was first contemplated by Leon Walras, one of the
progenitors of modern equilibrium theory [12]. In his conjecture, known as
Walrasian tatônnement, Walras states that attaining equilibrium would in-
volve a process by which a market groped toward equilibrium with the help of
a fictitious auctioneer who announced prices, then collected demands from
consumers as to how much of each good they would wish to purchase at the
announced prices. If demand exceeded supply, the price was adjusted upward.
If supply exceeded demand, the price was adjusted downward. Walras conjec-
tured that this procedure would cause the market to eventually settle into equi-
librium. This simple and intuitively plausible idea was widely accepted until
Herbert Scarf demonstrated, by his famous examples, the existence of an open
set of economies having a unique equilibrium that was unstable, therefore un-
attainable, under basic tatônnement [27]. In this study, we came to the conclu-
sion that it was possible to construct an improved tatônnement procedure in
decentralized electricity market. We conjecture that this procedure will con-
verge, of course assuming the existence of a competitive equilibrium, by im-
posing a restriction on the participants’ bidding here called the convexifying
rule. This constraint is not a theoretical limitation of DNSA but a practical
necessity for attaining a competitive equilibrium in both decentralized as well
as centralized auctions. A consequence of the convexifying rule is that during
the price-updating trials, participants are prevented from derailing the auction
process or rendering it chaotic.
DEFINITION 2 (Convexifying Market Rule): An electricity market is said to

operate under the convexifying market rule if, for each time period, the col-
lection of trial price-response pairs of GENCO(i),
to the k-th auction iteration – can be extended to a monotone increasing
incremental cost [monotone decreasing incremental benefit] function. 11
11
A real-valued function of a real variable x is said to be monotone increasing if the value of
decrease as increases; that is,
CONJECTURE (Convergence of DNSA): Extensive simulations and related

experience [28] indicates that DNSA converges well when PUC is profit opti-
mum and the competing participants behave rationally according to the con-
vexifying market rule.
5.4 Price Updating

In this subsection, a vector variable with the superscript k represents its
value at the k-th price-updating trial.
One important innovation of DNSA is the use by the Price Caller of a
price-updating algorithm that uses the first and second derivatives of the dual
function. This Newton algorithm [29] solves the nodal power balance equa-
tions, that is,
via the iterative procedure,
The second derivative matrix or Hessian is
The first two terms in the right-hand side of (35) are based on confidential
competing participant information and can only be estimated by the Price
Caller. Any estimation method may be used, among others, first differences,
and regression using all the past price-response pairs. In the simulations of
this chapter, each generator responds myopically so that the first two terms
are diagonal matrices whose elements can be estimated from the participant’s
behavior during the price-updating trials. The convergence of Newton’s
method does not however hinge on knowing the exact Hessian, a good ap-
proximation being sufficient. The third term in (35) is computed from public
network data available to the Price Caller.
This Hessian term corresponds to the case with no active line flow constraints.
Under congestion, the TRANSCO is required to disclose the Lagrange multi-
pliers associated with the active line flow limits. These values may then be
used by the Price Caller to refine the Hessian for faster convergence [30].
6. CASE STUDIES
To illustrate how DNSA works, we use a 5-bus system with network and gen-
erator data shown in Table 1 and 2. The resistance, R, reactance, X, and total
line charging susceptance, Bcap, are per unit on a base of 100 MVA.
We named the generators and loads according to their location in the network.
We chose identical generator costs because this case has been challenging for
traditional centralized UC algorithms that do not adequately represent the
transmission system. In this chapter, we report three simulations. In Cases A
and C, line flow limits are large enough to rule out congestion, while in Case
B some line capacities are reduced as indicated later to create congestion.
The planning horizon consists of three equal length periods. The load pat-
terns over time of DISTCO(2) and DISTCO(5) appear in Table 3 and are as-
sumed unaffected by nodal prices.
12
Following the notation in (6), the variable cost of GENCO(i) is given by
CASE A (Primal solution is profit optimal): Figure 4 and Figure 5 show

that convergence to within is obtained in seven trials (or
Newton price-updating iterations), at which point the nodal prices have also
steadied. Plainly, the auction is stopped successfully when the maximum
nodal power imbalance (defined in footnote 8, page 12) is less than 1 MW.
This is common practice in actual power system operation. The generator
output and profit profiles are reported in Table 4 and Table 5, respectively.
All GENCOs make some profit over the planning horizon even though
GENCO(l) and GENCO(3) must supply power at a loss during Period 3 ow-
ing to their minimum up-time constraint, In contrast,
GENCO(4), with a lower minimum up time, elects to operate
during Period 1 and Period 2 only, de-committing during Period 3 when there
is no financial incentive because of the low demand and low nodal price.
Table 6 shows the merchandising surplus of the TRANSCO that, without
congestion, is due entirely to transmission losses.
CASE B (Effects of transmission congestion): Here, the data of Case A is

modified by setting the flow limit of line 1–5 to 35 (MW), producing a light
congestion during Period 1 and a heavier one during Period 2.
As evidenced by Figure 6 and Figure 7, DNSA convergence is obtained
within 11 trials. Price updating exploits the Lagrange multipliers of the active
flow limits made available by the TRANSCO. Without this information, the
DNSA still converges but at a slower rate (47 iterations).
The final generator output and profit profiles are shown in Tables 7 and 8,
respectively. Comparing these results to those of Case A, we note a redistribu-
tion of the generation profits and a substantial increase in the Transco mer-
chandising surplus (Table 6 versus Table 9).
CASE C (Primal solution is profit sub-optimum): The same data as in case

A are used except that the minimum up-times of GENCO(l) and GENCO(3)
are decreased to two hours
As the generating units can now switch off during Period 3, under the
profit maximizing trial iterations of DNSA the participants do not come to a
power balancing agreement at this time period. The resulting cycling behavior
in prices and outputs during period 3 is shown in Figure 8 and Figure 9. This
is a signal to the Price Caller that profit optimality is infeasible and that an
alternative price-updating scheme must be initiated, as discussed in the Con-
clusions.
The primal unit commitment solution without profit optimality does how-
ever exist as shown in Table 10,
Table 11, and Table 12.13 Comparing the last two columns of Table 12,
we see that each generator primal profit, is positive but less than the
maximum possible, The unfairness of primal solutions that are
profit sub-optimal can be observed from the relative differences between pri-
mal and maximum profits.
13
During those intervals where the DNSA and the primal solutions are theoretically equal, the
numerical results may differ slightly due to convergence tolerance.
7. CONCLUSIONS
As an alternative to centralized pool auctions and operation, we propose a

Decentralized Nodal-Price Self-Scheduling Auction (DNSA) for self-
commitment and dispatch of electricity market participants in a power pool.
The scheme exploits the decomposability of the dual unit commitment prob-
lem. Each competing GENCO (DISTCO) self-computes its level of produc-
tion (consumption) and the 0/1 commitment by maximizing profit for a given
set of trial nodal prices. Similarly, the Transco independently maximizes its
merchandising surplus while satisfying all network security constraints.
The Price Caller, a centralized entity without direct access to proprietary

cost information, updates prices through an effective Newton algorithm until
the power balance at each bus is satisfied. DNSA does not assume a perfect
market and accounts for the AC load flow model including transmission losses
and line congestion, in addition to integer variables, ramping rates, start-up
costs, and minimum up and down times. The convergence of DNSA hinges
on the notions of profit optimality and the convexifying market rule. The latter
rule is imposed to ensure that the response to the Price Caller by market par-
ticipants will not derail the price updating process. Interestingly, the convexi-
fying rule, while preventing irrational and chaotic bidding, also offers the
competing participants some latitude to refine their bids during the auction.
In study cases A and B, we illustrate the convergence characteristics of
DNSA with and without transmission congestion. In addition, we solve the
traditionally difficult case of generators with identical cost functions. In Case
C, we illustrate two important points as regards profit sub-optimal problems.
First, the primal profits of individual participants are not the maximum possi-
ble for the primal nodal prices. We contend that under free competition this
result is unfair to some participants. Second, price cycling is a clear signal that
profit optimality is infeasible.
To ensure fairness of treatment for all competing participants, the next
generation of unit commitment models should allow for self-scheduling, and
should therefore include profit optimality constraints. Ongoing research is
considering various mathematical reformulations and computational solutions
of the profit optimal primal problem, as well as DNSA derivations that solve
this new problem in a decentralized manner [30]. We are looking at DNSA
price updating schemes that will enforce profit optimality through the trial
iterations for those cases that are not naturally profit optimum, that is for
which the primal unit commitment formulation is not already profit optimum.
The crucial point here is to enforce profit optimality whenever possible while
achieving maximum social welfare.
REFERENCES
1. K.J. Arrow. Social Choice and Individual Values, Second Edition. New Haven: Yale Uni-
versity Press, 1963 ed. 1951).
2. R. Green. “The Political Economy of the Pool.” In Power Systems Restructuring: Engi-
neering and Economics. Kluwer Academic Press, 1998.
3. M. Ilic, and F.D. Galiana. “Power Systems Operation: Old vs. New.” In Power Systems
Restructuring: Engineering and Economics. Kluwer Academic Press, 1998.
4. M. Huneault, F.D. Galiana, and G. Gross. “A Review of Restructuring in the Electricity
Business.” In Proc. 13th PSCC Conf., Tronheim, Norway, 1999.
5. F. Wu, and P. Varaiya. Coordinated multilateral trades for electric power networks: The-
ory and implementation. Elec. Power Energy Syst., 21(2): 75-102, 1999.
6. C. Imparato. “Market-Making Attributes of Alternative ISO Structures.” Comments in a

Panel Session, In Proc. IEEE PES Winter Meeting, 1999.
7. Federal Energy Regulatory Commission (U.S.). “Regional Transmission Organizations:
Rulemaking.” Docket No. RM99-2-000, Order No. 2000, December 1999.
8. R. Wilson. “Activity Rules for the Power Exchange.” Report to the California Trust for
Power Industry Restructuring, March 3, 1997.
9. G. Gross, E. Bompard, P. Marannino, and G. Demartini. “The Uses and Misuses of Opti-
mal Power Flow in Unbundled Electricity Markets.” In Proc. of the IEEE PES, Summer
Meeting, Edmonton, Alberta, Canada, July 1999.
10. R. Johnson, S.S. Oren, and A. Svoboda. “Equity and Efficiency of Unit Commitment in
Competitive Electricity Markets.” In Proc. POWER Conf. and paper PWP-039, Univ. of
California Energy Institute, March 15, 1996.
11. C.W. Smith. Auctions: The Social Construction of Value. London: The Free Press, 1989.
12. M.-E.L. Walras. Éléments d'Économie Politique Pure; ou la Théorie de la Richesse So-
ciale. First edition, 1874. [English translation: Elements of Pure Economics, or; The The-
ory of Social Wealth. Homewood, I11.: Published for the American Economic Association
and the Royal Economic Society, by R.D. Irwin, 1954.
13. P.H. Griffes. “Iterative Bidding in the PX Market.” A report to the California Power Ex-
change, http://www.calpx.com/news/publications/index.htm., January 2000.
14. H. Everett. Generalized Lagrange multiplier method for solving problems of optimal allo-
cations of resources.” Op. Res. 11: 399-417, 1963.
15. F. Zhuang, and F.D. Galiana. Towards a more rigorous and practical unit commitment by
16. A. Merlin, and P. Sandrin. A new method for unit commitment at Electricité de France.
IEEE Trans. PAS -102, 5: 1218-1225, 1983.
17. P.A. Samuelson. Foundations of Economic Analysis. Harvard University Press, 1947.
18. L.A. Wolsey. Integer Programming. John Wiley & Sons, Inc., 1998.
19. T.S. Dillon, K.W. Edwin, H.D. Kochs, and R.J. Tand. Integer programming approach to
the problem of optimal unit commitment with probabilistic reserve determination.” IEEE
Trans. PAS-97, 6:2154-2166, 1978.
20. J.M. Arroyo and A.J. Conejo. Optimal response of a thermal unit to an electricity spot
market. To appear in the IEEE Trans. Power Syst. Paper PE040PRS (10-99).
21. A. Brooke, D. Kendrick, A. Meeraus, and R. Raman. GAMS: A User’s Guide. GAMS
Development Corp., 1998.
22. J.A. Muckstadt, and S.A. Koenig. An application of Lagrangian relaxation to scheduling
in power generation systems. Op. Res., 25(3): 387-403, 1977.
23. D.P. Bertsekas, G.S. Lauer, N.R. Sandel, and T.A. Posberg. Optimal short-term schedul-
ing of large-scale power systems. IEEE Trans. Autom. Cont., 28(1): 1-11, 1983.
24. N. Nadira, and A.S. Cook. “On the Availability of Data Required by Optimal Power
Flows under Increased Competition,” IEEE PES, Tutorial Course – Optimal Power Flow:
Solution Techniques, Requirements, and Challenges, 96 TP 111-0, 1996.
25. Federal Energy Regulatory Commission (FERC–US). “Form 715: Annual Transmission
Planning and Evaluation Report.” http://www.ferc.fed.us/electric/F715/, January 2000.
26. B.T. Polyak. Introduction to Optimization. Optimization Software, 1987.
27. H.E. Scarf. Some examples of global instability. Int. Econ. Rev., 1: 157-172, 1960.
28. M. Huneault. “An Investigation of the Solution to the Optimal Power Flow Problem In-
corporating Continuation Methods.” Doctoral thesis, McGill University, August 1988.
29. J.M. Ortega and W.C. Rheinboldt. Iterative Solution of Nonlinear Equations in Several
Variables. Academic Press, 1970.
30. A.L. Motto. “Equilibrium of Electricity Market Under Competition.” Ph.D. thesis in pro-
gress, McGill University.
Chapter 16
DECENTRALIZED UNIT COMMITMENT IN

COMPETITIVE ENERGY MARKETS
Jinghai Xu and Richard D. Christie

University of Washington
Abstract: In a competitive energy market, instead of, or in addition to, a centralized unit
commitment, individual generation owners will make independent unit com-
mitment decisions. They will seek to maximize their profits against the predicted
market clearing price. Their unit commitment strategy will be expressed in their
bids, so that they shut-down or start-up when the market price indicates such ac-
tivity. In this chapter, we develop a unit commitment based price-taking (UCPT)
bidding strategy with a simple price prediction mechanism and explore it using a
market simulator. Simulation results show that an individual generator has
higher profits with UCPT bidding than with simple price-taking bidding, and
that the cost of supplying price-inelastic loads achieved by the market is lower
when all generators use UCPT bidding. It appears that UCPT bidding gives re-
sults similar to those from a Lagrangian relaxation unit commitment (LRUC),
without a fix-up step, and it has problems with convergence and feasibility simi-
lar to LRUC. We observe cyclic behavior in market prices with UCPT bidding,
and we show that it depends on the price prediction mechanism. Alternative
price prediction mechanisms can reduce cyclic behavior. Finally, we conceptu-
ally explore potential strategic behavior and market power arising from unit
commitment constraints.
1. INTRODUCTION
The unit commitment problem – scheduling generator start-ups and shut-
downs over a period of time to minimize the cost of serving expected loads -
has been applied by the power industry and studied by researchers for dec-
ades. Since unit commitment was typically performed for a set of generators
all owned by one company—to meet load exclusively served by that com-
pany—it was natural for the algorithm to assume that one central authority
controlled the status of every generator. This case is called centralized unit
commitment.
Deregulation has invalidated the assumption of centralized control. A
number of different companies now own generators. Each company must
make its own individual start-up and shut-down decisions, and cannot control
the decisions made by other companies. This case is called decentralized unit
commitment, because the commitment decision making is carried out in a de-
centralized control structure.
Recent publications that consider deregulation and unit commitment deal
mostly with centralized unit commitment. In some cases (such as the England
and Wales Power Pool), the market structure requires generators to submit to
centralized unit commitment [1]. In other cases, researchers have assumed the
existence of a centralized unit commitment in decentralized markets [2, 3, 4].
Oren et al. [5], however, identify the problems inherent in the use of cen-
tralized unit commitment. Specifically, they point out that due to the near-
optimal nature of the solutions obtained by practical unit commitment algo-
rithms, small changes in total cost can have large consequences for individual
generators. When all generators are owned by one company, these differences
are not important. When generators are owned by different companies, these
differences are highly problematical.
The problems with centralized unit commitment have been recognized in
practice by various deregulated markets. California and the Nord Pool market
in Norway have no centralized unit commitment in the market process. The
PJM Interconnection has a voluntary centralized unit commitment, but allows
market participants to self-commit. Based on economic simulations the Cali-
fornia tariff proposed an iterative energy market bidding scheme [13] to ac-
count for start-up costs, but has so far decided not to implement it due to the
cost of implementation and time constraints of the required communications.
Even if centralized unit commitment is required by, for example, connec-
tion agreements, it seems unlikely that it can practically be enforced. If a gen-
erator is required to run, and thinks that it should be shut down, it may suffer
an operating problem of some kind that forces it off-line. Generators are well
known to be the least reliable components of the power system and separating
intentional shut-downs from truly inadvertent ones is likely to be impractical.
If a generator wants to run, but is required to shut down, it could, perhaps,
claim restraint of trade, especially if its bid for the time period in question
shows that it is willing to run at minimum power for any price.
Note that we base this discussion of generator non-compliance on eco-
nomic motivations over a relatively long period of time, and it does not ad-
dress compliance during emergency conditions.
In [6], Li et al. introduce a market model that uses generator self-
commitment to determine generator bids over a fixed time period, and then
the usual market resolution process resolves the bids to determine market
prices. Then generator bids are redetermined for the same time period using
the new prices. The concept of individual self-commitment to maximize prof-
its given future market prices and the decentralized nature of the commitment
problem are new. But the idea that bidders would have several opportunities
to bid for the same time period is unworkable, as mentioned above.
Decentralized Unit Commitment 295
In the same journal, Huse et al. [7] introduce a method of computing gen-
erator bids that includes self-commitment and generates one bid at a time.
This is better-suited for most existing market resolution processes and for the
general desire in most markets to move bidding as close to operation as possi-
ble. This technique, however, is introduced in the context of using a market
simulation to solve a centralized unit commitment problem. Once market
clearing prices have been obtained for the time periods of interest, the bidding
process is repeated for the same time periods, so an iteration loop is present
which is not present in real markets.
In this chapter, we apply the bidding strategy developed in [7] to a some-
what more realistic market simulation. We then use the simulation to explore
the following questions: Is there an incentive for individual generators to pay
attention to unit commitment when bidding? What is the impact of unit self-
commitment on the market, in particular, in comparison to a centralized unit
commitment?
Section 2 gives the mathematical notation of this model. Section 3 dis-
cusses how generators can control their commitment through the form of their
bids. Section 4 describes several different bidding strategies that pay more or
less attention to unit commitment. Section 5 illustrates the market simulation
used, including the generation and loads. Section 6 gives results from the use
of the different strategies, identifies an interesting cycling behavior in the
market, corrects it, and addresses the first question. Section 7 addresses the
second question, and Section 8 discusses strategic bidding issues.
2. NOTATION
We used the following notation in our research:
thermal time constant for the generator, in hours
cold start cost, in $
fixed start cost, in $
C(P) : cost function, in $/hour
modified cost function, in $/hour
IC(P) :incremental cost function, in $/MWh
M : unit self-commitment period
profit, in $/hour
P : power output, in MW
high power output limit, in MW
low power output limit, in MW
p : market price, in $/MWh
predicted market price, in $/MWh
STC : start-up cost, in $
SDC: shut-down cost, in $/hour

t: time that the generator was cooled, in hours
T: cyclic period of price
u : unit status, u = 1 if running, u = 0 if not running
w : price prediction weight
3. CONTROLLING COMMITMENT WITH BIDDING

In a Marshallian market where all bidders are paid the market clearing
price (MCP), the price-taking bidding strategy sets bids equal to incremental
cost for generators operating between minimum and maximum output power
(segment CD in Figure 1a). What happens to the bid when these limits are
approached? For the upper power limit, the answer is easy. The generator
cannot produce more power no matter how much it is paid, so the bid curve is
simply vertical (segment DE in Figure 1a). For the lower limit the answer is
both more difficult and more interesting.
If the generator is willing to stay on line at minimum power once MCP

goes below its incremental cost at its low output power limit at
then the curve simply extends downward, as shown by the segment BC in
Figure la. When MCP reaches a level at which the generator is no longer
willing to operate, (the minimum acceptable price, or MAP), it will shut
down. The generator transitions from to zero output, as shown by the dot-
ted line AB. Thus, by submitting the bid curve ABCDE, the generator informs
the market about its willingness to operate, and the market tells the generator
whether to operate or shut down via the MCP.
It is certainly possible that the MAP for operation could be greater than
In this case, the MAP supercedes the incremental cost. The bid curve for
this case is shown in Figure 1b.
The transition from operating to shutdown poses feasibility problems. The
generator cannot actually operate below its minimum power but it is
quite possible for a market solution to return an MCP equal to MAP, with a
scheduled output between zero and (on section AB of Figure 1a). Resolu-
tion of this infeasibility depends on the market type.
In an ex ante market, the generator has at least three possible ways to deal
with the problem. There is some time before the bid period starts in which the
generator may be able to adjust its schedule by making bilateral contracts. The
generator also has an implicit contract with the ISO, since there must be a
generation adjustment mechanism of some kind to account for real time varia-
tions in load and inadvertent generator outages. The generator could simply
shut down at the start of the hour and accept any costs charged for deviating
from its schedule. Finally, what is actually scheduled is the energy to be de-
livered over the bid period. If metering is done on an hourly basis, the genera-
tor could simply operate at minimum power for the amount of time required
to deliver the scheduled energy, and then shut down. The time needed to re-
duce generator power to zero may cause the generator to supply some energy
to the power system while it shuts down even if it is scheduled for zero out-
put, aiding this strategy. The generator will choose the lowest cost strategy,
and it can factor any associated costs into its bids, as described later.
In an ex post market, a generator should simply shut down when assigned
an output less than The real time nature of the ex post market will change
bids in the next resolution period to account for changes in generator status.
Thus, if a minimum acceptable price (MAP) can be computed, generators
can use this to modify their ideal price-taking bids, and the interaction of the
modified bid with the market will signal the generator whether to stay on-line
or shut down in a given bid period.
4. BIDDING BEHAVIOR
Two bidding strategies with different approaches to finding the MAP are
discussed in this section. In developing these strategies, we make four as-
sumptions: (1) generators are price-takers; (2) generators try to maximize its
profit in the market; (3) each generator bids independently in the market; and
(4) the generator’s incremental cost function is monotonically increasing.
4.1 Finding Minimum Acceptable Price

If a generator has no shut-down and start-up costs, the minimum accept-
able price at which it will choose to operate is, therefore, the MCP at which its
profit is zero. It is well known from microeconomic theory that this occurs at
the point where incremental cost equals average cost, which is also the mini-
mum of average cost, i.e. [8]
The bid curve of the generator with this MAP is shown in Figure 2 (dark
line). P* is the power output where average cost (AC) and incremental cost
(IC) intersect.
4.2 Applying Power Output Limits

The power output limits of a generator, and modify the MAP.
When P* is within the power output limits, i.e., the MAP is
unaffected, as shown in Figure 3a.
With at the profit of the generator is
and is positive (see Figure 3b). Since the profit is zero at
which is the minimum value of average cost in the range and

the bid curve is shown in Figure 3b. Similarly, when the profit is
zero at as shown in Figure 3c. In general,
4.3 Including Shut-down and Start-up Costs

The MAP is also modified by non-zero shut-down cost or start-up cost.
Consider a non-zero shut-down cost. The MAP is the price that sets the gen-
erator’s profit equal to the negative of shut-down cost (SDC), i.e.,
or
This is similar to equation (2) with a modified cost function:
Then the MAP is:
The MAP given by equation (8) may even be negative if SDC is high
enough, giving the bid curve in Figure 4.
If a generator is shut-down in hour n-1, then the MAP for hour n may
have to include the start-up cost STC. STC here can be treated as an extra
fixed cost of operation. The modified cost function can be further revised to
and
If the generator is operating in hour n-1, then value of STC in hour n is zero.
4.4 Shut-Down Price-Taking Strategy

The MAP given by equation (10) controls commitment decisions based
on cost and income in hour n only. This strategy is referred to as shut-down
price-taking (SDPT).
The SDPT strategy is naive, since it considers only those costs incurred in
a single period. A better approach to finding the MAP would include the ef-
fect of the commitment decision on future profit.
4.5 Unit Commitment-Based Price-Taking Strategy

A bidding strategy that incorporates the future profitability of the genera-
tor was first suggested in the unit commitment literature by Huse et al. [7].
The future profitability here is evaluated by a unit self-commitment approach,
so this strategy is referred to as unit commitment-based price-taking (UCPT).
In the UCPT strategy, when preparing the bid curve for hour n, a genera-
tor first computes its future profits for the two cases of running and shutting
down in hour n. Then the difference of the two projected future profits, or the
lost profit due to shutting down, is treated as a extra cost and incorporated into
the MAP computation.
Suppose is the future profit if the generator is running in hour n, and
is the future profit if the generator is shut down in hour n. The lost future
profit (which is referred to as in [7]) is given by:
Then the MAP for UCPT is computed by extending equation (10) as follows:
4.6 Unit Self-Commitment for UCPT

We find the future profit by solving a unit self commitment problem,
formulated as:
Given a known operating state in hour n and predicted prices for hours
n+1 to n+M, find the commitment schedule for a generator that maximizes its
profit in the interval n+1 to n+M, subject to generator output limits and mini-
mum shut-down time constraints. Profits are income (which is market price
times power output) minus cost (fuel cost, and start-up cost) if operating, or
the negative of shut-down cost if shut down. The commitment does not con-
sider minimum run time, ramp rate, or spinning reserve constraints, in order to
simplify the computation. Time varying start-up costs are considered.
In mathematical form,
Given
Find and to maximize
s.t.
and minimum down-time constraints
Where
This problem is simple enough to be solved by forward dynamic program-

ming. It resembles the single generator subproblem in Lagrangian relaxation
unit commitment.
4.7 Price Prediction for UCPT
As indicated in equation (13), a future price profile is required when

computing future profits. This price profile has to be predicted by each indi-
vidual generator. In this paper, we use two simple price prediction methods
based on historical prices. In real life, various sophisticated price prediction
algorithms will be employed.
The first method is called the single period prediction (SPP). It assumes
that the market price is cyclic with period T, because the load profile used in
the example is also cyclic with period T. Then the estimated future market
price is just the historical market price for the preceding cycle. Suppose that
the historical prices before hour n are known. Then predicted future prices
starting from hour n are as follows (and are illustrated in Figure 5):
Note that is not computed because is unknown. The unit self-

commitment is therefore actually performed for m = l,...,T-1. Omitting the
last hour of a 168-hour week should have little effect on the result of future
profit differential
If the historical price is not available (in the initial period), the predicted
price is set as:
As discussed later, the use of SPP leads to a cycling market price pattern
with period 2T. To correct this, we use a weighted average price (WAP).
Similar to SPP, a generator using WAP also assumes that the market prices
are cyclic with period T, but the predicted price here is a weighted combina-
tion of the last two historical prices, i.e.,
Setting of the price prediction weight is discussed in the next section.
5. BID STRATEGIES IN MARKET SIMULATION

In this section, we test SDPT and UCPT in a simulated power market en-
vironment. In addition, we address questions regarding the incentive for using
the UCPT and the impact on the market efficiency.
5.1 Simulated Power Market

We constructed a 15-generator test case was constructed on a market
simulator platform. There are no transmission constraints or costs. The gen-
erator parameters, listed in Appendix 1, were chosen randomly from the 110-
generator system published in [9]. In the market simulation, we include start-
up and shut-down costs in the cost computation. The minimum shut down
time constraint of the generator is also enforced. We do not enforce other con-
straints, such as ramp rate limits and minimum run time. We took a weekly
load (energy demand) shape from published market loads [12] and scaled it to
the generation in the example (Appendix 2). The same weekly load profile
was applied in every week of the simulation. The load bid for each hour is
price inelastic.
Market resolution proceeds hour by hour for ten weeks. In each hour,
each generator submits a bid. The market price is decided by crossing the ag-
gregated generator and load bids.
5.2 SDPT Results

When all generators use SDPT, the weekly average price is
$13.220/MWh. The hourly price trajectory is shown in Figure 6. As expected,
the prices for each week are identical.
5.3 UCPT Results

When all generators use UCPT with SPP price prediction with a one-
week time horizon, we observe a weekly price oscillation with a period of two
weeks, presented in Figure 7. After a high average price week, price estimates
for the coming week are also high. With high predicted future prices in the
UCPT computation, lower MAPs result. With lower MAPs in supply bids,
more generators stay on-line, which decreases the actual market prices and
vice versa. We show weekly average prices in Figure 8.
When the price prediction horizon is extended to two weeks, we observe a

four-week period of price oscillation (Figures 9, 10). This periodicity may be
because near term predicted prices dominate the profit differential calculation
and thus the value of MAP. The near term prices come from the initial hours
of the preceding period and are unaffected by recent market results.
To damp the weekly price oscillation, we replaced the SPP by WAP. Dif-
ferent price weights, i.e. w = 0, 0.1... 1.0, in the WAP were tested. We find
the least price oscillation when w is 0.7, shown in Figure 11. The weekly av-
erage prices in this case exhibit only small changes after five weeks as shown
in Figure 12. In general, oscillations are well-damped for
Figure 13 shows the last four weeks of hourly prices from UCPT overlaid
on each other. It is clear that the prices vary from week to week, indicating
that there is not a clean convergence of the process.
Results show that WAP is better suited to predict future market prices. In
later sections we used only WAP (with w = 0.7) in UCPT.
5.4 Addressing Questions

1. If all generators use SDPT, is one generator motivated to use UCPT?
To answer this question, we constructed a test case with Generator 27 us-
ing UCTP and all other generators using SDPT. We chose Generator 27 be-
cause it has the most status changes in the all SDPT case.
In the all SDPT case, Generator 27’s profit in the last four weeks is
$5,367. When the generator changes its bidding strategy to UCPT, its profit
increases to $32,443. When Generator 27 uses UCPT in its bid, because of the
long-term profitability, lower MAPs are used in its bids. Lower MAPs allow
it to run continuously, saving the start-up and shut-down costs.
Thus, Generator 27 is motivated to use UCPT by increased profits.
2. If all generators use UCPT, is a generator motivated to use SDPT?
When all generators use UCPT, Generator 27’s profit for the last four
weeks is $15,807. If Generator 27 changes to SDPT while the rest of the gen-
erators stay with UCPT, Generator 27’s profit falls to $5,614. Generator 27 is
motivated to use UCPT.
Profits of Generator 27 for the four scenarios are listed in Table 1.
3. Is the market better off with UCPT?

With a price inelastic load, the total cost of generation is related to social
welfare. The best market, with the highest social welfare, has the lowest cost
of generation.
We compare generation costs of the four scenarios in previous subsec-
tions. To avoid the initial unstable results, we use the data from the last four
weeks. A minimum total generation cost is observed in the scenario where all
generators use UCPT. The maximum total cost is found when all generators
bid with SDPT. Total generation costs in the last four weeks of different sce-
narios are listed in Table 2.
When all generators use UCPT, the lowest average weekly price also oc-
curs. This indicates that the price inelastic load also benefits from the sup-
plier’s UCPT bidding.
The result with lower costs and lower average prices in the UCPT case
indicates that UCPT improves the efficiency of the specific market.
To broaden this observation, the 110-generator system and its associated
daily load curve from [9] are simulated. In this case, the market process is
shortened to 10 days but the daily load curves are identical. The price predic-
tion period is 23 hours. Other assumptions made in Section 5.1 still apply.
We test two cases. First, all generators use SDPT, then UCPT. In each
day of the SDPT case – except the first day when no start-up cost applies –
the daily average price is $15.955/MWh, and the daily average cost and profit
are $3,793,107 and $1,313,937 respectively. In the UCPT case, the result is
stabilized after four days when using WAP with w = 0.7. Table 3 shows the
daily average prices, total cost and total profit for the UCPT case.
In the 110-generator system, after the initial day, the market efficiency of
UCPT is also higher than that of SDPT.
6. COMPARING UCPT WITH A LAGRANGIAN

RELAXATION SOLUTION
With the same 110-generator system and the same daily load curve used
in Section 5, we solve the unit commitment problem with a Lagrange Relaxa-
tion Unit Commitment (LRUC) algorithm. The initial unit state for LRUC is
taken from the UCPT solution for the last hour before the final day of the
simulation, allowing a reasonable comparison of LRUC results with the last
day of UCPT. We iterate LRUC 20 times and present the results in Table 4.
The minimum generation cost was found in iteration 11. The total cost of the
LRUC solution for the 11th iteration was only 0.02% less than UCPT. The
profit was 0.47% lower in LRUC.
Note that the total daily cost of LRUC is lower than those reported in [7]
and [9], because constraints such as minimum up-time and spinning reserve
are not observed and the initial state differs.
LRUC and UCPT had very similar commitment patterns as shown in Ap-
pendix 3. The LRUC commitment pattern is shown, with UCPT differences
highlighted. Only three of 110 generators had different commitment patterns,
two with different run time durations and one with a three-hour shut-down.
The differences in operating durations of generators 13 and 15 may be attrib-

uted to the way LRUC deals with infeasible solutions, such as generator
power outputs between zero and These are allowed by UCPT, as dis-
cussed in Section 3, but not by LRUC.
Overall, comparison of LRUC and UCPT results gives some assurance
that centralized unit commitment solutions can be used as reasonable predic-
tors of energy market prices, assuming no exercise of market power, and thus
as benchmarks for market efficiency measures.
7. STRATEGIC BIDDING
We assumed in previous sections that generators are price takers. This as-
sumption may not hold in reality. In a real world power pool, a generator will
have non-zero market power – the ability to affect the price of electricity. In-
stead of bidding at its incremental cost, generators are likely to explore other
bidding strategies that maximize long term profit. This approach to bidding is
often called strategic bidding.
A common form of strategic bidding is to anticipate the response of other
bidders to one’s own bid. By examining the expected commitment behavior
of other generators, a strategic bidder can arrive at a computed profit differen-
tial far different fromt that in a price-taking calculation, changing its MAP.
The technical constraints of unit commitment offer another opportunity
for strategic bidding. If competitors have generators with long time constants,
such as long minimum shut-down times or long minimum run times, a strate-
gic bidder with market power can attempt to drive market price down far
enough to force the slow generators off-line at a strategic time, for example,
just as load starts to increase, and then recoup the cost of this by exercising
market power while the slow units are off-line.
This problem is reminiscent of entry problems from economics. In gen-
eral, slow units are cheap ones, and expensive units have short response times.
The potential for entry (start-up) of an expensive unit should control the ef-
fects of this sort of exercise of market power.
This discussion has only scratched the surface of unit commitment-aware
strategic bidding strategies. A lot of work can be done developing and evalu-
ating such strategies, and developing methods to detect their employment,
measure their effects on market efficiency and limit their negative effects.
8. CONCLUSIONS AND FUTURE WORK

We have explored two methods of including start-up costs and shut-down
constraints in market-based generation bidding. We found that prices depend
on the price prediction mechanism, and we showed that a simple weighted

price prediction methodology resulted in more stable market prices. Examples
show that the Unit Commitment Price Taking (UCPT) strategy results in
higher profits for generators using it and in improved market efficiency.
A comparison between market based UCPT solutions and centralized La-
grange relaxation unit commitment shows that the UCPT solution is very
close to the LRUC solution, indicating that centralized unit commitment can
be used for price prediction in decentralized markets.
The area of strategic bidding, taking unit commitment considerations into
account, appears fruitful for future research. The possibility of a theoretical
link between the decentralized market-based unit commitment process with
UCPT bidding, and centralized Lagrangian relaxation unit commitment, is an
intriguing theoretical topic.
ACKNOWLEDGEMENTS
This work was funded by the Advanced Power Technologies Center at the
University of Washington, Chen-Ching Liu, Director. The authors are grateful
to Karl Seeley for discussions on economic issues.
REFERENCES
1. D.P. Mendes and D.S. Kirschen, “Modelling of a competitive electricity power pool.” 32nd
Universities Power Engineering Conference, UPEC ’97, UK, 1: 387-390, 1997.
2. G. Gross and D.J. Finlay. “Optimal Bidding Strategies in Competitive Electricity Markets.”
In Proc. Twelfth Power Systems Computation Conference, PSCC, Dresden, 2: 815-823,
1996.
3. J.J. Ancona. A bid solicitation and selection method for developing a competitive spot
priced electric market. IEEE Trans. Power Syst., 12 (2): 743-748, 1997.
4. G. Huang and Q. Zhao. “A Multi-objective Formulation for Competitive Power Market.”
In Proc. 1998 Large Engineering Systems Conference on Power Engineering, LESCOPE
98, Canada: 317-22, 1998.
5. S.S. Oren, A.J. Svoboda, and R.B. Johnson. “Volatility of Unit Commitment in Competi-
tive Electricity Markets.” In Proc. Thirtieth Hawaii International Conference on System
Sciences, 5: 594-601, 1997.
6. C. Li, A.J. Svoboda, X. Guan, and H. Singh. Revenue adequate bidding strategies in com-
petitive electricity markets. IEEE Trans. Power Syst., 14(2): 492-497, 1999.
7. E.S. Huse, I. Wangensteen, and H.H. Faanes. Thermal power generation scheduling by
simulated competition. IEEE Trans. Power Syst., 14(2): 472-477, 1999.
8. R.S. Pindyck and D.L. Rubinfeld. Microeconomics, Fourth Edition. Prentice Hall, 1998.
9. S.O. Orero and M.R. Irving. Large scale unit commitment using a hybrid genetic algorithm.
Elec. Power Energy Syst, 19(1): 45-55, 1997.
10. A.J. Wood and B.F. Wollenberg. Power Generation, Operation, and Control, Second edi-
tion. John Wiley & Sons, Inc., 1996.
11. S. Takriti, J.R. Birge, and E. Long. A stochastic model for the unit commitment problem.
IEEE Trans. Power Syst., 11(3): 1497-1508, 1996.
12. California ISO Web Site, http://www.caiso.com.
13. P.H. Griffes. “A Report to the California Power Exchange: Iterative Bidding in the PX
Market.” ANALYSIS GROUP/Economics, 1999, http://www.calpx.com/news/publications
/index.htm.
APPENDIX 1. PARAMETERS OF 15 GENERATORS
MDT: minimum down time

PUC: A, B, and C: coefficients of polynomial cost function
Pmax / Pmin: maximum/minimum power output
SCC: cold start-up cost
SCF: fixed start-up cost
ST: time constant of cold start-up cost
SDC: shut-down cost
APPENDIX 2. WEEKLY LOAD PROFILE
APPENDIX 3. UNIT COMMITMENT PATTERN FROM

LRUC FOR 110 GENERATOR CASE
INDEX
ABACUS 154, 161 cutting plane 154, 155, 167, 168,

AC transmission 9, 75, 78, 271, 278 171
AMPL 72, 161 DC power flow approximation 10,
ancillary services 6, 9, 16, 18, 20, 75, 78, 79
22, 26, 29, 33, 41, 76, 93, 95, decision trees 252
96, 103, 107, 120, 122, 169, demand-side 23, 26, 30, 44, 48, 56,
194, 251, 261, 266 242
arbitrage 33 demand-side bidding 53, 67, 276
ARIMA 148 dispatch 114, 123, 193, 195, 231
artificial neural network 255, 265 DISTCO 188, 271, 274, 276, 280
auctions 35, 185, 228 distributed energy 10
Australia 190, 253 DSI-OPF 51
autonomous agent models 190, 203, duality gap 72, 76, 89, 134, 167,
253 175, 176, 181,278
bc-opt 154, 161 dynamic programming 2, 71, 75, 77,
bid optimization 46, 231 81, 93, 125, 129, 162, 171,
bidding 9, 16, 20, 21, 32, 34, 43, 55, 193, 212, 219, 221, 230, 260,
93, 185, 191, 207, 233, 256, 303
283 dynamic programming, stochastic
bidding strategies 251, 252, 255, 99,109,117,122,141,144,
257, 259, 264, 267, 295, 296, 260, 263
304, 311 economic efficiency 32
bids, multipart 33 Edgeworth approximation 147, 150
bilateral trading 17, 19, 26, 28, 29, emissions 10
185, 188, 228, 274 EMOS 40, 50
BPMPD 85 EMOSL 160, 161, 162, 163
branch-and-cut 154, 156, 157 EPRI ANN-STLF 51
bundle method 77, 264 EPRI-DYNAMICS 51
California 17, 20, 22, 24, 32, 39, 46, EPRI-OTS 51
97, 135, 188, 249, 259, 294 equivalencing method 211, 214
Chicago Board of Trade 191 ESCO 40, 44, 49, 188, 190
competitive market 15, 41, 53, 55, experimental economics 91, 188,
94, 185, 227, 249, 274, 275, 189
293 fairness 90, 294
complementarity problem 230 FERC 16, 96
cost recovery 175, 176, 181 fuel price 123, 135
Cournot model 230 fuel price forecasts 97
CPLEX 4, 6, 153, 235, 241, 243 game theory 252, 268
GAMS 161, 241 market simulation 9, 39, 187, 189,

GENCO 40, 44, 185, 188, 190, 193, 190, 203, 228, 250, 268, 293,
271, 274, 276, 279 304
genetic programming 2, 185, 196, Markov process 94, 95, 96, 97, 108,
204, 212, 253 141, 146, 260, 262, 263, 265
genetic programming automata 205 Mathematical Programs with Equi-
hydropower 9, 41, 230, 233, 235, librium Constraints 9
237, 239, 244, 260, 262, 263 MATLAB 85, 88
Iberdrola 248 MATPOWER 85
India 217 MINOS 72, 85, 90
integer programming 2, 4, 125, 153, MINTO 154, 161
155, 159, 171, 212, 235, 238, MIPO 154, 161
262 modeling environments 160
interior point method 167, 168, 171, modeling languages 161
174, 182 Monte Carlo 96, 103, 104, 105, 106,
interruptible loads 9 109, 141, 147, 150
ISO 3, 9, 17, 18, 19, 23, 24, 25, 26, MPL 161
31, 39, 43, 44, 46, 48, 140, MultiMATLAB 88, 90
188, 250, 274, 276 New England 16, 20, 22, 23, 24, 33,
Ito processes 135 97, 188, 250, 252, 260, 266
Lagrange multipliers 2, 9, 60, 70, New York 20, 21, 23, 24, 25, 42, 97
77, 81, 143, 170, 176, 263, New Zealand 42, 252
279, 285 Newton algorithm 285
Lagrangian relaxation 2, 4, 54, 57, nominal group 7
70, 75, 78, 79, 84, 88, 89, 90, Nord Pool 294
94, 118, 127, 133, 162, 167, normal approximation 147, 150
170, 176, 193, 212, 219, 252, Northern Ireland Electricity 163
256, 258, 260, 262, 278, 293, numerical application 60, 85, 100,
303, 310 106, 127, 130, 147, 162, 178,
limits constraints fixing technique 179, 200, 217, 241, 259, 266,
223 286, 304
LINDO 153 NYMEX 135
linear programming 6, 153, 158 OPF 42, 84, 85, 90
load recovery 54, 57, 62 OPL 161
locational pricing 23, 25, 42, 46, 48, optimal self-commitment 93
271, 287, 290 optimization 2
LSE 50 options 103, 121, 129, 139
maintenance 203 Order 2000 18, 27
market design 15, 18, 28, 31, 35 Order 888 17, 19
market power 9, 18, 20, 28, 35, 227, ordinal optimization 253, 258, 259
231,293,311 OSL 153, 235
Index 319
parallel implementation 88 risk management 47

penalty bundle method 168 RTO 18, 27, 30, 31
PJM 16, 18, 20, 21, 23, 24, 25, 26, scenario analysis 262
40, 42, 46, 97, 188 scenario tree 117, 127, 129
POOLCO 16, 168, 274, 275, 282 schedule coordinators 3, 16, 40, 47
price 54, 67, 120, 143, 193, 244, settlement systems 21, 23
250, 254, 261, 296, 304, 306, simulated annealing 212
307 simulation 43, 54
price elasticity 9 Spain 248
price forecasting 9, 45, 96, 103, 107, start-up 21, 55, 93, 95, 100, 107,
187, 192, 197, 231, 260, 268, 120, 142, 193, 195, 233, 238,
293, 303, 306, 311 239, 255, 263, 273, 299, 302
price inversion 22 statistical data-mining 96, 190
price uncertainty 93, 98, 135, 143, statistical methods 137
145 stochastic dynamic programming:
price volatility 104, 111, 140, 251 See dynamic programming,
priority lists 2, 72, 212 stochastic
probabilistic production-costing stochastic programming 118
model 140, 145 strategic behavior 65
profit optimality 271, 281 sub-gradient technique 71, 77, 127,
PROFITMAX 106 167, 172, 264
programming 233 Tractebel 162
pumped storage 48, 233, 237, 245, TRANSCO 188, 272, 280
260, 262, 263, 265, 266 transmission 18, 20, 23, 27, 30, 32,
PX 17, 19, 20, 28, 39, 47, 49, 250, 42, 202, 272, 277, 280, 283,
259 285, 286, 288
ramp rate constraints 76, 90, 93, 95, transmission rights 25, 28
170, 195, 214, 216, 239, 262, UK 53, 54, 252, 294
277 uncertainty 44, 251, 266
Rational Buyer Protocol 22 unit decommitment method 117,
reactive power 78, 79 125
regression model 45, 258 variable duplication technique 79
regulation 15, 17, 186 voltage limits 76, 78, 80
reliability 9, 10, 33, 188, 202 Walrasian equilibrium 284
reserves: See ancillary services XPRESS-MP 153, 158, 163
revenue adequacy 252 zonal pricing 24

The Next Generation of Electric Power Unit Commitment Models PDF

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

The Next Generation of Electric Power Unit Commitment Models PDF

Hochgeladen von

Copyright:

Verfügbare Formate

THE NEXT GENERATION OF

Saigal, R. / LINEAR PROGRAMMING: A Modern Integrated Analysis

KLUWER ACADEMIC PUBLISHERS

©2002 Kluwer Academic Publishers

Print ©2001 Kluwer Academic Publishers

All rights reserved

Created in the United States of America

Visit Kluwer Online at: http://kluweronline.com

I. The Evolving Context for Unit Commitment Decisions

2. Regulatory Evolution, Market Design and Unit Commitment

3. Development of an Electric Energy Market Simulator

II. New Features in Unit Commitment Models

4. Auctions with Explicit Demand-Side Bidding in Competitive

5. Thermal Unit Commitment with a Nonlinear AC Power Flow

6. Optimal Self-Commitment under Uncertain Energy and Reserve

7. A Stochastic Model for a Price-Based Unit Commitment Problem

8. Probabilistic Unit Commitment under a Deregulated Market

III. Algorithmic Advances

9. Solving Hard Mixed-Integer Programs for Electricity Generation

10. An Interior-Point/Cutting-Plane Algorithm to Solve the Dual Unit

11. Building and Evaluating Genco Bidding Strategies and Unit

12. An Equivalencing Technique for Solving the Large-Scale Ther-

IV. Decentralized Decision Making

13. Strategic Unit Commitment for Generation in Deregulated Elec-

14. Optimization-Based Bidding Strategies for Deregulated Electric

15. Decentralized Nodal-Price Self-Dispatch and Unit Commitment

16. Decentralized Unit Commitment in Competitive Energy Markets

WHY THIS BOOK?

William R. Stewart Jr.

1. PURPOSE OF THE BOOK

2. EXAMPLE OF NEW CAPABILITIES: SOLVING

seven-day (168-hour) period. The generators have different operating char-

Test UC Mixed Integer Linear Program:

and are the maximum ramp rates (in MW/period) for

The seven-day MIP problem is non-convex, with one integer variable

The marked improvement in times seen in Version 6.5 (Table 1) can be

3. NEW NEEDS: RESEARCH PRIORITIES

Identify features, characteristics, or requirements of system operation and

Approximately 40 of the workshop attendees addressed the first two tasks

not a high priority with this group, as inclusion of AC representations fell in

generation; emissions modeling; and particular stochastic methods for

REGULATORY EVOLUTION, MARKET DESIGN

Richard P. O’Neill, Udi Helman, and Paul M. Sotkiewicz

William R. Stewart Jr.

2. REGULATORY EVOLUTION AND THE

frequent curtailments of transactions, justified on the basis of reliability but

els. Firms can exercise market power in electricity markets because of

2.1 Experience with ISO and Bilateral Markets

electricity markets can be complex, with multiple interdependent products

in a sequence, even though several of the products may be alternative (substi-

The settlement systems can be characterized as either single settlement or

2.2 Order 2000 and RTOs

framework. Each PX would act as a single scheduler, submitting schedules

3. DESIGN PRINCIPLES FOR DAY-AHEAD RTO

The following is a list of recommended design principles derived from the

Principle 4: Bidding protocols should promote flexibility of participation.

actual marginal (incremental or going forward) costs, a good approximation

sion; administrative measures should be at best transitions to market designs

Principle 9: Minimize the incentives for market participants to engage in

The other primary argument in the chapter is that a well-designed RTO

Atif Debs and Charles Hansen

Abstract: The paper outlines the development of an electricity market operation

2. ELEMENTS OF EMERGING ELECTRICITY