Beruflich Dokumente
Kultur Dokumente
Herausgegeben von
C. Bierwirth, Halle, Deutschland
B. Fleischmann, Augsburg, Deutschland
M. Fleischmann, Mannheim, Deutschland
M. Grunow, München, Deutschland
H.-O. Günther, Bremen, Deutschland
S. Helber, Hannover, Deutschland
K. Inderfurth, Magdeburg, Deutschland
H. Kopfer, Bremen, Deutschland
H. Meyr, Stuttgart, Deutschland
K. Schimmelpfeng, Stuttgart, Deutschland
Th. S. Spengler, Braunschweig, Deutschland
H. Stadtler, Hamburg, Deutschland
H. Tempelmeier, Köln, Deutschland
G. Wäscher, Magdeburg, Deutschland
Diese Reihe dient der Veröffentlichung neuer Forschungsergebnisse auf den Gebie-
ten der Produktion und Logistik. Aufgenommen werden vor allem herausragende
quantitativ orientierte Dissertationen und Habilitationsschriften. Die Publikatio-
nen vermitteln innovative Beiträge zur Lösung praktischer Anwendungsprobleme
der Produktion und Logistik unter Einsatz quantitativer Methoden und moderner
Informationstechnologie.
Herausgegeben von
Professor Dr. Christian Bierwirth Professor Dr. Herbert Kopfer
Universität Halle Universität Bremen
Kontakt
Professor Dr. Thomas S. Spengler
Technische Universität Braunschweig
Institut für Automobilwirtschaft
und Industrielle Produktion
Katharinenstraße 3
38106 Braunschweig
Thomas Kirschstein
Integrated Supply
Chain Planning in
Chemical Industry
Potentials of Simulation
in Network Planning
Foreword by Prof. Dr. Claudia Becker
und Prof. Dr. Christian Bierwirth
Thomas Kirschstein
Halle (Saale), Germany
Springer Gabler
© Springer Fachmedien Wiesbaden 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or
part of the material is concerned, speci¿cally the rights of translation, reprinting, reuse of illus-
trations, recitation, broadcasting, reproduction on micro¿lms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by
similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a speci¿c statement, that such names are
exempt from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.
Foreword
In literature on operations management, chemical industry is primarily dealt with focused
case studies. The limited flexibility of production processes as well as the volatility of
customer demand were identified as the most prominent challenges for chemical supply
chain management. Most chemical production assets are complex, inflexible, and incur
high set-up costs. Hence, such assets can only be used in an economically reasonable way
if operated continuously. Therefore, logistical processes offer substantial contributions to a
chemical company’s value added. E.g. logistical processes ensure the supply of downstream
assets in case of asset break downs. More important, however, logistical processes allow
balancing demand variations within chemical production networks. Therefore, transport
and distribution planning is essential in both, intra-site logistics, mostly relying on pipeline
transportation, and inter-site logistics, mostly relying on rail and ship transportation.
With this book, Thomas Kirschstein covers this topic in an encompassing, profound,
and general way. Starting with basic chemical production processes, which are modelled
with time series methods, established and new logistical planning problems for distributed
chemical production networks are presented. Finally, these components are integrated in a
simulation-based planning framework which is implemented in a decision support system.
The developed decision support system is validated by means of case studies relying on
historical records of a real world chemical production network.
The present work is an important contribution to scientific literature from a methodolo-
gical and application-oriented point of view. It develops systematically the basic elements
for modelling complex chemical production networks and illustrates the benefits of advan-
ced decision support systems in chemical supply chain management. We wish the book
continued success and wide acceptance.
Preface
This book is the product of a process which started in 2006 as a cooperation seminar
between the Dow Olefinverbund GmbH and the chairs of Production & Logistics, Stati-
stics, and Operations Research. At this time I couldn’t imagine a result like this. During
this time not only the project has developed and changed, also my experiences, skills, and
expectations have grown. These developments would have been impossible without the
support of a bunch of persons. The space provided here does not suffice to duly thank all
these persons. So, the following thanks are rather exemplary than encompassing.
Special thanks are due to my supervisors Prof. Dr. Claudia Becker, Prof. Dr. Christian
Bierwirth, and Prof. Dr. Taïeb Mellouli which offered invaluable support by their advices
and constant commitment in all those years.
No less important was the support of all employees of Dow Olefinverbund GmbH which
allowed me to study not only Dow’s production and logistics processes but also their day-
to-day business. In particular, I thank Wolfgang Schnabel, Andreas Kroupa, and Wubbe
Prins for organizing and coordinating the practical part of this project. Without their
support, this book would not exist.
Besides my direct supporters, I’d like to thank all colleagues at the department for
Economics and Business Administration. Thanks to them research and teaching at the
university was always a pleasant and (almost) always a productive challenge. Represen-
tative for all colleagues I thank Heidrun Rudolph, Lisiane Schnegelsberg, and Dr. Steffen
Liebscher from the chair of Statistics as well as Ute Lorenz, Dorota Mańkowska, Jens
Kuhpfahl, and Prof. Dr. Frank Meisel from the chair of Production & Logistics.
An important contribution to the successful completion of this book offered my family
and friends. Not only by numerous discussions about the contents of the projects but also
by the constant interest in its progress, I never lost sight of the goal of this project.
But, it is unthinkable that this book would exist without the support of my wife Susanne
and my daughters Antonia and Emilia. All three of them are constant sources of joy,
inspiration, and recuperation in particular in times of frustration and self-doubts. For
their support, I’m deeply grateful.
Thomas Kirschstein
IX
Contents
List of Tables XI
List of Figures XV
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Bibliography 207
Appendix 229
XI
List of Tables
A.1 Information criteria for VARX models of order 1 to 6 for the de-alkylation
plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
A.2 Coefficients of the V ARX(3) model with outlier correction . . . . . . . . . 230
A.3 ANOVA table for the initial V ARX(3) model (without outlier correction) 230
A.4 ANOVA table for the initial V ARX(3) model with outlier correction and
variable selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
A.5 Stationary flow rates (in t/h) for exemplary chemical production network . 234
A.6 Coefficients of the time series models for cracker at site 1 (V ARX(2)+AR(3))234
A.7 Coefficients of the time series models for cracker at site 2 (V ARX(2)+AR(3))234
A.8 Coefficients of the time series models for hydrogenation plant at site 1
(V ARX(3) + AR(2)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
A.9 Coefficients of the time series models for hydrogenation plant at site 2
(V ARX(3) + AR(3)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
A.10 Coefficients of the time series models for Butex plant at site 2 (V ARX(1)+
AR(4)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
A.11 Coefficients of input and output time series models of SISO and MISO plants235
A.12 Un-/loading capacities, initial stock of empty RTCs, target and initial stock
levels per transported chemical at both production sites . . . . . . . . . . . 236
A.13 Cost rates for the MC-RTP instances . . . . . . . . . . . . . . . . . . . . . . . 236
A.14 Distributions of inter-arrival time and deliver quantities for external cus-
tomers/suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
XIII
List of Figures
A.1 Diagnostic plots for the ARX(1) model of the Naphtha time series . . . . . 229
A.2 Residual diagnostic plots for V ARX(3) model with outlier compensation. 233
A.3 QQ-plots of residuals for models (4.19)-(4.21) . . . . . . . . . . . . . . . . . . 240
A.4 Density function of the Weibull distribution with k = 1.5 and λ = Γ( 365
5
)
. . . 241
3
A.5 Loss functions for both sites during pipeline inspection . . . . . . . . . . . . 242
A.6 Diagnostic plots for (4.22) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
XVII
List of Abbreviations
AB agent-based
ACF auto-correlation function
AIC Akaike information criterion
AR(X) auto-regressive [model] (with exogenous variables)
auto-regressive conditional heteroscedasticity [model]
ARCH(X)
(with exogenous variables)
auto-regressive integrated moving average [model]
ARIMA(X)
(with exogenous variables)
auto-regressive moving average [model] (with
ARMA(X)
exogenous variables)
ARD average relative deviation
CV corporate value
DE discrete event
DSS decision support system
EACF extended auto-correlation function
(s)ELSP(-BP- (sequence-dependent) economic lot sequencing problem
IF) (with global stocks and interface handling)
FIR finite impulse-response
generalized auto-regressive conditional
GARCH(X)
heteroscedasticity [model] (with exogenous variables)
GA genetic algorithm
GDP gross domestic product
h hour
HQIC Hannan-Quinn information criterion
IRP inventory routing problem
IPPSP integrated production planning and scheduling problem
(s)LP (stochastic) linear program
MA(X) moving average [model] (with exogenous variables)
MAD median absolute deviation (from median)
MC-RTP multi-chemical rail transportation problem
(s)MILP (stochastic) mixed-integer linear program
MIMO multiple input, multiple output
MINLP mixed-integer non-linear program
MIP mixed-integer program
MIRP maritime inventory routing problem
MIRSP maritime inventory routing and scheduling problem
XVIII
List of Notation
General notation
a, ..., z, α, ..., ω vectors
A, ..., Z, Γ, ..., Ω matrices
A, ..., Z sets
i, j, k, b indices
t, T time index, time horizon
E(⋅)/P (⋅)/V ar(⋅)/Cov(⋅, ⋅) expectation/probability/variance/covariance of ⋅
Γ(⋅) Gamma function
ˆ
x̂/x̂ estimate of x / estimate of estimate of x
Bin (n, p) Binomial distribution with n trials and probability p
N (μ, σ 2 ) Normal distribution with mean μ and variance σ 2
WB (k, λ) Weibull distribution with shape k and scale λ
Notation of Chapter 2
N index and number of lags for exogenous variables
L/M number of output/exogenous variables
K number of parameters
p/q/P /Q order of AR/MA/ARCH/GARCH process
α, β coefficients of (G)ARCH models
γ/ρ, Corr(⋅, ⋅) partial correlation/ correlation
φ, θ AR and MA coefficients
Φ, Θ VAR and VMA coefficient matrices
coefficient scalars/vectors/matrices for exogenous
υ/υ/Υ
regressors
δ(B), ω(B), θ(B), φ(B) univariate lag polynomials
Φ(B), Θ(B), Υ(B) multivariate lag polynomials
σ2 variance
Σ covariance matrix
μ/μ mean scalar/vector
, ε, ξ, η/, ξ/Ξ error scalars/vectors/matrices
B back-shift operator
C(t) concentration function
X(t) turnover function
XX
Cs starting concentration
A reactant
EA activation energy
R universal gas constant
tR Damköhler number
m reaction order
r reaction rate
k reaction rate constant
W (t) Wiener process
Notation of Chapter 3
K, H set of state combinations
I set of lot positions
N set of access points
O set of states
S set of modes
ρCap pumping rate capacity
ρ pumping rate
Q transition matrix
qst transition probability from state s to state t
π steady state vector
ωs /ωs flow level of state s/state combination s
Xj flow rate at access point j
yh total material balance of state combination h
ph probability of state combination h
μC expected net deficit
α, β service level
r stock level
V (⋅) expected loss function
L/L̄ stock level / maximum stock level
T stock /T f ill stock-up time/pipeline fill time
τ time for pipeline transport
s /cs
cset hold set-up / holding costs for product s
ctrans
st transition costs for a change from product s to t
binary, 1 if at position i = 0, ..., I chemical s is
bsi
scheduled
Ts cycle time of chemical s ∈ S
tbi starting time of position i
binary, 1 if from position i − 1 to position i a transition
xsti
from chemical s to t occurs
XXI
1 Introduction
1.1 Motivation
The chemical industry is a key industry in developed countries. Not only its size but also
its role as a raw material supplier for almost all other industries establishes the chemical
industry as one of the most important industries in a nation’s economy. The chemical
industry is defined by the type of production process. Chemical industry focuses on the
modification of a substance’s chemical and/or physical properties.1
In this thesis the chemical industry and, particularly, the basic chemical industry is
the subject of interest. The basic chemical industry provides industrial goods (e.g. basic
plastics, coatings, fertilizer) as well as consumer products (e.g. detergents, tires, insulat-
ing materials). This industry can be seen as an indicator for a nation’s economy since
shocks in economic development (or the expectation of it) typically first affect the basic
industries. A general change in (expected) economic development induces a revision of
raw material stock levels to match the expected production plans in a broad range of com-
panies in various industries. Since almost all industries use various chemical products as
raw materials, the chemical industry is confronted with the accumulated "revision effect"
manifesting itself either in an increase or decrease of total customer demand. This is typi-
cally one of the first measurable and tangible effects of a change in economic development.
Hence, companies in chemical industry are faced with a volatile and hardly influenceable
customer demand. Moreover, most chemical products are commodities, i.e. there are only
few possibilities for product diversification such that the market for chemical products is
highly competitive with a strong focus on prices.
Another crucial characteristic is that many chemical production processes are only cost-
efficient if operated continuously. Continuously operated chemical production plants are
expensive and complex technical systems. Typically, they are optimized for operation
at a specific conversion rate such that operational costs per output unit are minimized.
Deviations from the optimal production rates typically cause additional operational costs
stressing the small contribution margins.
Chemical transformation processes are typically split into multiple stages that are or-
ganizationally separated but technologically interdependent. Chemical production plants
are interconnected via direct energy and product flows but managed locally. Due to diverg-
1
See Schönsleben (2011).
T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,
DOI 10.1007/978-3-658-08433-2_1, © Springer Fachmedien Wiesbaden 2015
2
ing production processes in basic chemicals’ production as well as the processes’ continuity
and interconnectedness, chemical production plants are often clustered at large-scaled in-
tegrated chemical production sites. This form of organization is economically preferable
since co-products can be processed further with little logistical effort. However, managing
such industrial complexes is a challenging task since all product flows have to be kept in
balance simultaneously to maintain all production processes. This is particularly compli-
cated if unforeseen events occur. Therefore, local production and consumption as well as
material imports and exports have to be planned accordingly. This perspective is called
the intra-site perspective.
Basic chemical companies often operate a certain number of such integrated production
sites such that a geographically disperse network of production sites exists. To keep all
materials balanced at all sites, material transports between sites have to be considered.
This perspective is called the inter-site perspective.
In summary, basic chemical companies face a challenging combination of difficult market
conditions, inflexible and complex production processes, and a wide-spread, interdepen-
dent production network. Hence, there is an inevitable strain for minimizing the oper-
ational costs of production from a company-wide perspective. To identify and exploit
such a network’s cost saving potentials, elaborate decision support tools are necessary
taking into account not only the complex production network of each site but also the
interrelations between sites. Such a company-wide perspective typically addresses tac-
tical and strategical decisions since, on the operational level, the local management of
the production sites is responsible. However, operational processes have to be modelled
accurately to be able to anticipate the network’s reactions on an aggregated level. On the
tactical and strategical level, environmental influences have to be incorporated to reflect
the uncertain nature of future developments.
At this level of aggregation, multiple objectives are typically pursued. Beside the op-
erational costs of a certain network configuration, e.g. investment costs to realize such a
configuration as well as robustness aspects and service measures come into scope.2 Thus,
tactical/strategical decision support tools have to provide
• a reliable model of the production sites (including the production processes and the
local logistical processes),
• the capacities of the logistical system (for transport, turn-over, and stock holding),
This thesis aims at providing a general framework for building a tactical/strategical deci-
sion support model to optimize (large-scaled) chemical production networks. This frame-
work consists of three core parts:
1. To model local chemical production processes at production sites, time series mod-
els are used which are able to accurately represent the dynamic, time-dependent
structure of chemical production processes.
Throughout this thesis, all methods and techniques described are illustrated by small-
scaled examples or case studies based on real-world data. In Chapter 4, a complex case
study inspired by a cooperation project with a chemical company is presented, integrating
the very most of the aforementioned methods. Figure 1.1 shows a graphical overview on
the structure of the thesis and the associated examples.
Chapter 1: Introduction
section example
Motivation —
Outline —
Chapter 5: Conclusion
intra-site inter-site
S1 C1
intra-site
intra-site
S2 C2
From Figure 2.1, it becomes obvious that modelling the production plants provides the
basic data to describe the material flows within the whole network. Before the chemical
production processes are described in detail, some characteristics and key figures about
chemical products are provided.
Basically, two major groups of chemicals can be distinguished: inorganic and organic
chemicals. The latter subsumes all chemicals containing at least one bond of a hydrogen
and carbon atom. The former group encompasses all chemicals without such a bond.1
This work primarily focuses on organic chemicals which are produced by the (basic)
chemical industry. For the industrial production of organic chemicals, three basic natural
resources are available: natural gas, coal, and crude oil. Of these three materials crude oil
1
In particular, pure carbon, e.g. in the form of diamonds is handled as an inorganic chemical. See
e.g. Seager and Slabaugh (2007) for more details.
T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,
DOI 10.1007/978-3-658-08433-2_2, © Springer Fachmedien Wiesbaden 2015
6
is the exceedingly most important raw material for the production of organic chemicals.
The (basic) chemical industry is an intermediary who transforms these raw materials
into basic and intermediate (organic) chemicals. Three main production phases can be
distinguished:
1. raw substance splitting: raw materials are split into (short-chained) basic chemicals
The category of basic chemicals comprises about 20-30 chemical substances building the
basis for all subsequent substances. This class summarizes basic organic substances as
well as basic gases and inorganic basic substances (such as Chlorine or Ammonia).2 These
substances are purely intermediate and not sold to consumer markets.
Intermediate chemicals are usually simply structured chemicals most often composed
by basic chemicals that are only exceptionally made for (private) consumer markets.
This class comprises, e.g., alcohols and many kinds of acids.3 Examples for marketable
intermediate chemicals are ammonia compounds used as basic fertilizers in agriculture.
The classes of basic and intermediate chemicals consist of a fairly small number of sub-
stances which are mainly fluids or gases. But these substances are used in vast quantities
to produce final chemical products. To give an impression, Figure 2.2 shows the produced
quantities of basic and intermediate chemicals in 2008-2010 for Germany (grouped as
stated above).4
production quantities in 1,000 tons
100000
organic chemicals
inorganic chemicals
basic gases
80000
60000
40000
20000
0
2
See e.g. Behr et al. (2010, p. 10 ff.) or Baerns et al. (2006, ch. 16).
3
See e.g. Behr et al. (2010, p. 10 ff.) or Baerns et al. (2006, ch. 17).
4
These production statistics are calculated based on the European goods classification scheme NACE.
The three classes "basic gases", "inorganic chemicals", and "organic chemicals" correspond to the
NACE classes 20.11, 20.13, and 20.14, see Eurostat (2011).
7
The group of organic chemicals represents about 50% of the total quantity of produced
chemicals with a total quantity of about 40 million tons per year in Germany. These prod-
ucts are typically crude oil derivatives. Thereof, approximately 20 million tons account
for organic basic chemicals such as alkenes or aromatic compounds where the remain-
ing 20 million tons comprise organic intermediate chemicals such as alcohols or chlorine
derivatives.5
Final products in chemical industry are produced by chemical reactions of intermediate
and basic chemicals. These final products are widely used in almost all other industries.6
Final chemicals can be categorized into
1. polymers/plastics
5. pharmaceuticals.
Among these categories, pharmaceuticals and body care products set up own sub-
industries due to the special characters of their products and production processes.
These companies directly serve (private) consumer markets. The remaining three cat-
egories comprise "classic" chemical companies mainly providing intermediate products
for other industries such as mechanical engineering industry, building industry, textile
industry, and plastics industry.
In the next section chemical production processes are characterized and categorized.
The theoretical modelling of the underlying chemical reactions is described subsequently.
Chemical production processes realise chemical reactions in industrial scale in chemical
plants. Based on models of chemical reactions, methods are provided to describe the
behaviour of chemical production processes.
side chemical reactions, also physical transformation processes (so-called basic operations)
are used to alter specific properties of the reactants.
Usually, chemical reactions require an initiation, i.e. they only take place under specific
circumstances. The list of reaction parameters is vast. Basic physical parameters are
pressure, temperature, electricity, or light. Specific constellations of these parameters
influence the reaction rate, i.e. how fast or slow a reaction takes place.8 If a reaction
can only take place by means of auxiliary chemicals, it is called catalytic and such an
auxiliary reactant is called catalyst.9 Another important measure of chemical reactions
is the conversion rate of a reaction, i.e. how many percent of the input reactants’ mass
is transformed into the substances of interest.10 This measure is important to decide
whether a reaction can be realized in an economically profitable way.
The molecular structure of reactants can be changed in numerous ways:11 First the
reactants’ molecules can be combined which is called synthesis. A prominent example is
the hydrogenation of carbon dioxide to produce methanol.12
Second, a molecule or molecular fragment can also be split which is called decomposition.
To recycle bottles made of polyethylene terephthalate (PET) catalytic depolymerization
is used to split the PET in valuable components.13
In a substitution, a molecular fragment of a reactant is replaced by a fragment of another
reactant. A prominent example is the alkylation where an alkyl group is transferred from
one molecule to another. E.g., the production of Ethylbenzene from Benzene and Ethylene
by the so-called Friedel-Crafts alkylation is a standard process in chemical industry.14
If more than one molecular fragment is substituted, this is categorized as a metathesis.
A recent industrial application is the olefin metathesis to produce e.g. Propene from
Ethene and 2-Butene.15
The types of reactions presented above can be combined with the ordinary classifi-
cation of production processes in convergent, divergent, and transformation processes.
Depending on the number of input and output products this process classification can be
enhanced as Table 2.1 shows.
A SISO process is a single input-single output process and corresponds to transforma-
tion processes or substitution reactions. MISO processes (multiple inputs-single output)
comprise convergent processes and SIMO processes (single inputs-multiple output) com-
prise divergent processes similar to decomposition and synthesis, respectively. MIMO
8
E.g. see Baerns et al. (2006, p. 32).
9
The catalyst is not part of the molecular reorganization, i.e. no part of the catalyst is part of the
resulting chemicals. See Behr et al. (2010, ch. 12) for more information.
10
E.g. see Baerns et al. (2006, ch. 11.2) for more details.
11
A similar classification scheme particularly addressing organic reactions can be found in Jones and
Bunnett (1989).
12
See Bill (1997).
13
The resulting components depend on the specific depolymerization process applied, see e.g. Mishra
et al. (2002) or Paszun and Spychaj (1997).
14
E.g. see Degnan et al. (2001).
15
See Mol (2004).
9
```
``` # outputs
``` single multiple
# inputs ```
```
single SISO SIMO
multiple MISO MIMO
16
However, the production coefficients may vary.
17
This depends primarily on the further use of the output products and/or their market prices. For
example steam crackers are usually optimized for Ethylene production because Ethylene is used in a
wide variety of final products.
10
Depending on the product portfolio as well as the structure of product demand, multi-
purpose plants are operated either in batch mode or continuously. There are plenty of
definitions about both terms.18 Here, a technological point of view is used, i.e. batch
processes are characterized by a fixed production capacity which is defined as the quan-
tity of produced chemicals after which a process interruption is required. In contrast,
continuous processes are characterized by a production rate, which is defined as the quan-
tity of goods produced in a given time. The time between process interruptions is not
technically limited. The former especially occurs for specialty chemicals and pharmaceu-
ticals, where the latter is typical for polymers/plastics and some agrochemicals.19 For
body care products the production technology is mixed depending on the product variety.
Low-volume products with many product variations and often changing recipes, such as
cosmetics, are usually produced in batch mode, whereas high-volume products with few
product variations, such as detergents, are usually produced continuously. In both cases,
the production processes are interruptible. Table 2.2 summarizes the above-mentioned
characteristics.20
final chemicals
basic chemicals intermed. chemicals
commodity speciality
mode continuous continuous continuous batch
vergence divergent di-/convergent convergent/transform.
purpose single single multiple
18
See e.g. Loos (1997, p. 48 ff.) and references therein.
19
Low-volume products such as specialized pesticides are often produced in batch mode, see e.g. Loos
(1997, p. 70 ff.).
20
Far more typical characteristics of production processes could be included, see e.g. Loos (1997, sec. 3.1).
However, the chosen characteristics are the most important with respect to logistical implications.
21
The head sites of BASF in Ludwigshafen and Dow Chemical in Midland are popular, large-scaled
examples of such integrated production sites.
11
e.g. Ethylbenzene which in turn can be de-hydrogenated to Styrene. Styrene can then be
used as raw material for Styrene-Butadiene rubber which is an important raw material
e.g. for the production of tires. Other branches of Benzene application are the production
of Aniline (which is widely used in polyurethane production) and Cumene which is a
composition of Benzene and Propylene and is mainly processed to resins.22
Naphta
cracking
Pygas
dehydrogenation hydrogenation
co-polymerization polymerization
Note that the provided final products depicted in Figure 2.3 rather represent a set of sub-
products than a uniform substance. These sub-products have the same basic molecular
structure but can differ e.g. in certain physical characteristics or colour.23
The vessels containing chemical reactions are called (chemical) reactors. Their design
and size depends on the intended reaction. The performance of a reactor is measured in
terms of the conversion rate, the purity of the substances of interest, and economic aspects
such as resistance, energy consumption etc. Basic types of reactors can be categorized
by different characteristics.24 The vessel design is a first category. Roughly, one can
distinguish between (stirred-)tank reactors and pipe reactors.
To effectively execute chemical reactions, the involved reactants have to be provided in
sufficient purity. Moreover, many chemical reactions result not only in one pure output
product but in a mixture of output products. Thus, preparation and post-production
22
See e.g. Baerns et al. (2006, ch. 16) for details.
23
This can be obtained e.g. by various additives or process settings, see Behr et al. (2010).
24
E.g. see Baerns et al. (2006) or Trambouze and Euzen (2004) for more detailed classification schemes.
12
processes have to be carried out to provide valuable reactants for subsequent chemical
reactions. These operations are called basic operations and can be categorized in thermic
and mechanic operations.25
The most prominent thermic basic operation is the rectification or distillation to sepa-
rate individual components from a mixture.26 For separation, the mixture is evaporated
completely and successively cooled down. The components can be effectively separated if
they have different boiling points or dew points, respectively. In the case of close boiling
points for the components in the feed mixture, distillation is still applicable if an auxiliary
chemical (so-called solvent) is available changing the boiling or dew points of at least one
component (so-called fractional distillation).27
A distillation column is a metal tube which is separated in compartments by so-called
trays or plates. The mixture to be separated is fed into the middle of the column. In-
side the column temperature and pressure are variable depending on the height. At the
column’s top temperature is maximal and pressure minimal. Conversely, at the bottom
temperature is minimal and pressure maximal. Depending on the boiling points of the
mixture’s components, the composition of liquids and gases differs in each compartment
and at each tray, respectively. Ideally, at each tray a single fraction/component of the
feed mixture can be obtained.28 To guarantee constant conditions regarding pressure and
temperature, a surplus of residue liquids at the bottom of the column is (re-)boiled and
fed back into the column (so-called reflux). Similarly, a surplus of gases at the top of col-
umn is condensed and fed back. Figure 2.4 shows a schematic overview of a prototypical
column.29
Separation processes are designed and optimized for a specific mixture to be separated
which determines e.g. the number and position of trays as well as the atmospheric con-
ditions. Such processes are single-purpose plants for SIMO or SISO30 processes. The
production rate often can be varied in certain ranges without causing serious variations
of the separation accuracy.
The distinction in reactions and basic operations to prepare and post-process chemical
reactions is valid in most cases. However, there is a principal advantage to integrate both
chemical process steps. Despite organizational and technical drawbacks, such an integra-
tion is physically and chemically advantageous due to more favourable energy balances.
One example is the reactive distillation where chemical reactions take place inside the
25
E.g. see Baerns et al. (2006) for details.
26
For a more general overview on (thermal) separation processes see e.g. Seader et al. (2011) or Baerns
et al. (2006, ch. 9).
27
E.g. see Behr et al. (2010, p. 88-89),Hoffman (1977) or Seader et al. (2011) for details.
28
Regarding the determination of the maximal or optimal number of trays a large body of literature has
been expanded and is still growing, e.g. see Seader et al. (2011) for an overview and Yeomans and
Grossmann (2000) or Viswanathan and Grossmann (1993) for more detailed insights.
29
In practice many adaptations and subtypes of this basic form have been derived depending on the
specific processes and circumstances the distillation column is intended for. See e.g. Smith (2005,
ch. 11 and 12) for an overview.
30
If only one component is considered as valuable.
13
condenser
gas
Prod. D
reflux
Prod. C
mixture/feed
trays
Prod. B
vapor
liquid
Prod. A
reboiler
distillation column.31 With the exception of such special techniques, chemical operations
can be categorized as displayed in Figure 2.5.
operations
Chemical production processes can be divided in chemical reactions and basic opera-
tions (i.e. physical transformations). In chemical production plants, multiple processes
from both classes are combined and take place in sub-plants which are closely intercon-
nected. The planning and configuration of such plants is very complex and expensive.
Hence, a detailed modelling of the underlying chemical and physical processes is necessary
to avoid misinvestments. The next section outlines an overview on the steps necessary to
31
For an overview on reactive distillation see e.g. Baerns et al. (2006, p. 322 ff.) and for details see
e.g. Taylor and Krishna (2000).
14
model and control a chemical production plant beginning with a brief introduction into
the mathematical modelling of chemical reactions.
• What are the requirements and outcomes of the intended reaction(s) regarding en-
ergy and reactants?
Thermodynamics Kinetics
In the next subsection, kinetic definitions and concepts are introduced. Subsequently,
relevant properties of chemical operations w.r.t their modelling are outlined. Special
15
chemical kinetics
In reaction kinetics, the object of interest is the progress of a chemical reaction measured
in terms of the change in concentration of the corresponding reactants and products.33 The
change of a component’s amount in time is defined as the reaction rate of this component.34
Assuming constant volume this equals the concentration of this component. The reaction
rate rj of component j can be expressed as the change in concentration cj over time:
dcj
rj = . (2.1)
dt
The change in concentration typically depends on the reaction time, the concentration
of the other components, the temperature, and reaction specific properties. Temperature
32
In the strict sense the term kinetics only refers to reaction kinetics, i.e. the temporal development of
chemical reactions. Mass and energy transport processes are usually categorized as a pure physical
phenomenon. However, here a classification scheme based on the topics’ time domain is chosen. For
an extensive introduction in chemical kinetics, see e.g. Levine (2005).
33
In the following, the term component subsumes reactants and products as well.
34
There exist also some other definitions of reaction rates based on reaction-specific characteristics,
e.g. volume, mass or activated catalyst surface, see e.g. Baerns et al. (2006, p. 59-60) or Behr et al.
(2010, p. 41-42).
16
where c is the vector of concentrations and T is the temperature. The temperature effect
is usually assumed to be constant (also called reaction rate constant) and obtained by the
Arrhenius equality:
EA
f (T ) = k = k0 ⋅ e R⋅T (2.3)
where k0 is a reaction specific factor, EA is the energy necessary to start the reaction
and R is the universal gas constant.35 In the simplest case, the reaction rate depends
multiplicatively on the powers of the reactant’s concentrations:
rj = k ⋅ f (c) = k ⋅ cm
1 (t) ⋅ ... ⋅ cN (t)
1 mN
(2.4)
where mi is the reaction order of reactant i and ∑N i=1 mi is the order of the whole reaction.
36
The type of reaction orders depends on the structure of the reaction. Assume a simple
decomposition of one reactant, e.g. in cracking reactions of hydrocarbons in gasoline
production:
ν1 ⋅ A1 → ν2 ⋅ A2 + ... + νn ⋅ An (2.5)
where A1 is the (only) reactant and A2 , .., An are the products. The stoichiometric co-
efficients νj refer to the number of molecules of chemical j consumed/produced during
the reaction. For all components their concentration depends on the concentration of the
reactant, i.e. for reactions of order m1 = 1 follows for the reaction rate of chemical j
rj = νj ⋅ k ⋅ c1 (t). (2.6)
If one molecule of A1 is split into the products, it follows ν1 = −1. For the change in
concentration of the reactant follows:
r1 = dc1
dt = −k ⋅ c1 (t) (2.7)
which is the simplest form of a homogeneous linear differential equation of order 1 with
the solution:
where cs,1 is the concentration of A1 at the beginning of the reaction. Here, (2.8) implies
an exponentially decaying concentration of the reactant over time.
35
E.g. see Behr et al. (2010, p. 42).
36
There are also other forms of reaction rate equations, such as hyperbolic relations. See Baerns et al.
(2006, p. 62 ff).
17
dc1 (t)
r1 = = −k1 ⋅ c1 (t) (2.9)
dt
dc2 (t)
r3 = = k1 ⋅ c1 (t) − k2 ⋅ c2 (t) (2.10)
dt
dc3 (t)
r3 = = k2 ⋅ c2 (t). (2.11)
dt
This simple system of differential equations can be solved iteratively. The solution of
(2.9) is given in (2.8) and, hence, can be substituted in (2.10):
dc2 (t)
r2 = = k1 (cs,1 ⋅ e−k1 ⋅t ) − k2 ⋅ c2 (t). (2.12)
dt
Assuming k1 ≠ k2 and cs,2 = 0, (2.12) can be solved using standard results for homogeneous,
linear differential equations of first order:
cs,1
c3 (t) = cs,1 + (k2 ⋅ e−k1 ⋅t − k1 ⋅ e−k2 ⋅t ) . (2.14)
k2 − k1
These examples of some simple types of chemical reactions show the general solution
methodology for determining the behaviour of chemical reactions over time. Especially
the modelling of catalytic and reversible reactions requires more sophisticated mathemat-
ical methods to determine the concentration rates.38 However, in the end more or less
complicated systems of differential equations have to be solved.
Another prominent thermodynamical relation often relevant in describing chemical pro-
cesses is the heat conduction formalized in Fourier’s law.39 This partial differential equa-
tion describes the diffusion of heat (energy) in time depending on (contact) area and
temperature gradient. This partial differential equation leads to some other prominent
37
For more detailed information about the deduction of reaction rates for other elementary reactions and
reaction networks e.g. see Levine (2005) or Baerns et al. (2006, ch. 4).
38
See e.g. Baerns et al. (2006, p. 65-72) or Levine (2005) for details.
39
See e.g. Behr et al. (2010), p. 49 ff.
18
thermodynamical equations such as the heat and diffusion equation.40 The diffusion and
behaviour of energy is the main controller of most reactions. E.g. an energy-consuming
chemical reaction system is arranged such that the energy loss due to the chemical trans-
formation process is compensated by external energy supply (e.g. heating). Distillation
processes are typical examples for such processes.41
Another class of equations subsumes kinetic and phase transformations of all involved
reactants. Such equations describe how the reactants’ molecules are transformed and
distributed in the reactor depending on time and other environmental parameters. De-
pending on the kind of chemical processes under consideration, both classes of equations
are of varying importance for modelling. E.g. for catalytic packed-bed reactors the chem-
icals’ reaction rates heavily depend on local physical conditions at the (solid) catalyst
material. The precise modelling of the local physical conditions and the mixture of chem-
icals flowing is important and complex in this case. In contrast, for classic stirred-tank
reactors kinetic and phase transformations are comparatively easy to model.
Most mass balances are deduced from the chemical transformation process directly,
whereas most energy equations are mainly deduced from thermodynamic relations. The
kinetics of a chemical reaction subsume relationships determining how the reaction rate
depends on conditions such as pressure, temperature and catalysts.
A1 → A2 + ... + An
40
There are a lot of standard textbooks regarding this topic see e.g. Sandler (2006), Koretsky (2004), or
Letcher (2004).
41
See e.g. Seader et al. (2011).
42
See Behr et al. (2010, p. 59 ff.) or Baerns et al. (2006, p. 145 ff.).
43
See e.g. Behr et al. (2010, p. 64-66).
19
Assuming a constant volume of the reactants and processing the reaction in a stirred-
tank reactor one important question is to determine the time needed to achieve a desired
turnover rate. To answer this question, (2.15) is solved exactly to t = − k1 ln(1 − X) for a
given turnover ratio X. Normalizing this result by the reaction constant k leads to the
so-called Damköhler number for this specific class of reactions:
dc1 (t)
= − k1 ⋅ cn1 (t) (2.17)
dt
d(1 − X)
⇔ = − k ⋅ (1 − X)n ⋅ cns,1 (2.18)
dt
dX
⇔ − = − k ⋅ (1 − X)n ⋅ cns,1 (2.19)
dt
1
⇔ dt = dX (2.20)
k ⋅ (1 − X)n ⋅ cns,1
1 XR 1
⇒ tR = n
cs,1 ⋅ k ∫0
dY (2.21)
(1 − Y )n
1 1 1
⇔ tR = ⋅( ( − 1)) (2.22)
cns,1 ⋅ k n − 1 (1 − X)n−1
For more complex reactions the derivation of Damköhler numbers is much more compli-
cated.44
Under adiabatic conditions heat flows have to be modelled accordingly with respect to
the mass balances. For exothermic reactions usually a stimulation of the chemical reaction
is required due to the higher energy level at the beginning of the reaction. In general,
the interactions between heat and mass balances are more difficult to model.45 More
complicated cases occur when both conditions are mixed, i.e. under polytropic conditions.
Here, solutions for heat and mass balances are hardly available analytically. Instead,
44
See e.g. Baerns et al. (2006, ch. 5.2) for more details.
45
E.g. see Behr et al. (2010, p. 67).
20
where dW denotes a stochastic process (the so-called Wiener process, which is described in
section 2.3) and σ is some variability parameter indicating the influence of the stochastic
component. Equation (2.23) constitutes a specific Ornstein-Uhlenbeck process.47 For this
specific class of SDEs an analytic solution can be derived as48
t
c1 (t) = cs,1 ⋅ e−k⋅t + σ ⋅ ∫ ek⋅(s−t) dW (s) (2.24)
0
which extends (2.8) by the addition of a stochastic term. Note that (2.23) can be solved
since the stochastic component is constant in time and independent from the concentration
process c1 (t). For most other types of SDEs or systems of SDEs analytic solutions are
not available.49
The design of chemical plants usually combines a bundle of chemical reactions and
basic operations, where the chemical reaction of interest is rather the core of a network
surrounded by preparation and follow-up operations. Predicting the behaviour of such
a network is difficult and usually not manageable by analytic models even if theoreti-
cal models are available for all components. Therefore, simulation approaches are often
applied.50
To give an example, Figure 2.8 shows the schematic sketch of a typical steam cracker
splitting Naphtha (or similar hydrocarbons) into short-chained components.
The focal product is usually the C2 -fraction, known as Ethylene.52 The chemical process
46
See e.g. Behr et al. (2010, p. 68).
47
Here, the parameter μ = 0, see Uhlenbeck and Ornstein (1930).
48
See e.g. Hassler (2007).
49
See e.g. Hassler (2007).
50
See e.g. Behr et al. (2010).
51
See Behr et al. (2010, p. 179).
52
Ethene or Ethylen is the simplest alkene consisting of only two carbon atoms.
21
"#$
!
!
!
of splitting the feed’s long-chained molecules into smaller ones is processed in a pipe
reactor (located in a furnace to provide the necessary temperature). To separate the
produced mixture of short-chained chemicals a bundle of separation steps is processed in
a system of distillation columns.
Modelling the complete production process as a SIMO process requires the modelling of
the splitting process first. This process can be characterized as a continuously processed
decomposition reaction in a pipe reactor.53 A model to describe the complete chemical
splitting process contains 680 single chemical reactions describing the dependencies of 37
substances.54 This decomposition describes the kinetic behaviour of the reaction depen-
dent on energy (heat) and pressure.55 Additionally, the environmental parameters have
to be modelled in terms of energy and mass equations.56 This model simulates the output
of the splitting process depending on certain control variables. In this case, mainly the
residence time of the feed in the pipe (between 0.1 and 2 seconds), the composition of the
feed, and the cracking temperature (between 800 and 900 °C) determine the composition
of the output mixture. Depending on the costs associated to energy and feed as well
as the value of the output mixtures, it is still an ongoing effort to derive optimization
models maximising the profit by finding the optimal control setting.57 However, not only
the composition of the cracking mixture but also the quality of the separation processes
is necessary to predict the outcome of the whole plant. For these components additional
models were developed.58
53
Classification according to Figure 2.5.
54
See Edwin and Balchen (2001).
55
Details can be found in Sundaram and Froment (1978),Willems and Froment (1988a), and Willems
and Froment (1988b). The latter references gives information about the necessary activation energies
of the specific reactions.
56
Here specific models were developed, see e.g. Ghashghaee and Karimzadeh (2007) for a comprehensive
one.
57
E.g. see Edwin and Balchen (2001).
58
See e.g. Qi et al. (2002) for a reactive distillation model for C4 separation process or specialized
22
Hence, to model a chemical plant precisely requires a lot of effort. This effort is typically
expended when a new plant is planned. However, such a precise process model is not
necessary or useful for other purposes. In short-term control of chemical production plants,
so-called prediction control models are used which focus on the measurement of system
responses to certain changes of control variables on an empirical basis. These models use
basic relations of the theoretical models introduced above. The next subsection outlines
a brief overview of such models.
disturbances
estimator
estimates/model
The chemical process can be controlled by various variables which affect physical con-
ditions as well as the product and energy inflows of the chemical plant. Output variables
comprise the physical properties of the plant’s outflows. In case of continuously oper-
ated plants, outputs are measured as flow rates. Beside the output quantity, the output
quality e.g. in terms of the composition of the output mixture is an important measure.
To link inputs/controls with output measures, an estimator quantifies the relationship
between both components. This estimator relies on a predefined model representing the
theoretical background of the estimated process. This model is completed by estimated
parameters which are based on historical records of the process’ performance. Note that
the estimator has to be able to distinguish between undesired, uncontrollable disturbances
and desired previous control effects. The resulting parametrized model can then be used
to predict the future behaviour of the process depending on the control settings. This
model is reported to the controller unit. The controller unit varies the control variables
to meet the desired target configuration of the production process. The optimal sequence
of control variables may depend on economic parameters (e.g. input prices) resulting in
a simultaneous optimization of controls and outputs. This implies a dynamic process
control. Usually, however, predefined static optimal target values for output measures are
to be met.60
Since this work deals with the aggregated simulation and planning of chemical produc-
tion processes, the focus is laid upon methods to determine estimations of the process
models. For process control this task is the crucial one as the estimations’ accuracy de-
termines the accuracy of the whole control process. The task to find an accurate process
model is often called process identification.61 To describe the input-output behaviour of
(continuously operated) chemical production plants finite impulse response (FIR) models
are widely used. These models can be seen as regression models where the historical
records of input/control measures determine the output measure. The term "finite" indi-
cates that a finite number of historical records is used to predict the process’ outputs.62
Often, chemical processes show a significant time-dynamic behaviour which is typically
reflected in auto-correlated and cross-correlated process measures. However, classic re-
gression models do not incorporate auto-correlation explicitly which in turn leads to a loss
in estimation efficiency or, even worse, biased estimates.63 Therefore, time series meth-
ods can be applied to incorporate auto-correlation effects. According to the classification
shown in Table 2.1 four basic types of FIR models can be distinguished.
SISO model
For the simplest case of univariate input variables and univariate output measures the
following model describes a linear dependency between the input (or control) variable
x = (x1 , ..., xT )′ and the output variable y = (y1 , ..., yT )′
N
yt = μ + ∑ υb ⋅ B b xt + ξt (2.25)
b=0
where T denotes the time horizon and B denotes the so-called back-shift operator
B b xt = xt−b . The parameters υb reflect the (lagged) influence of the control variable
on the output variable. μ is the average output level. The noise vector ξ = (ξ1 , ..., ξT )′
contains the disturbances that cannot be explained by the model. Note that ξ can be
derived from some coloured noise model as well as from a classic white noise model.
White noise refers to a stochastic process with zero mean and no auto-correlation, while
coloured noise accounts for auto-correlation. If the process under study fulfils the white
noise assumptions, (2.25) can be interpreted as a simple regression model and can be
60
About the conception of model predictive control schemes see e.g. the brief overview in Darby et al.
(2009) or textbooks such as Camacho and Bordons (2004).
61
E.g. see the overview on chemical process modelling provided by Lewin et al. (2002) or textbooks about
the topic such as Ljung (1999).
62
This finite number of regressors can be chosen quite large which is indicated by the term "non-
parsimonious" FIR models, see Dayal and MacGregor (1996).
63
See the seminal papers Yule (1921) and Yule (1926) or some standard textbook on time series analysis,
e.g. Cryer and Chan (2008, pp. 260-265).
24
SIMO model
For multiple outputs the univariate control variable approach (2.25) has to be re-
formulated using matrix notation
N
yt = μ + ∑ υ b ⋅ B b xt + ξ t (2.26)
b=0
where yt = (y1,t , ..., yL,t )′ is the vector of L output variables, υ b = (υ1,b , ..., υL,b )′ the vector
of regression parameters, and ξ t = (ξ1,t , ..., ξL,t )′ the vector of output errors. Similarly,
Ξ = (ξ1,1 , ..., ξL,1 , ..., ξ1,T , ..., ξL,T ) is an L×T matrix which is derived from an L-dimensional
stochastic process either white or coloured. μ = (μ1 , ..., μL )′ denotes the vector of output
levels.
MISO model
The MISO model is used to describe converging production processes. Here, the structure
is similar to a SISO model, where the input is some (say) M -dimensional vector such that
N
yt = μ + ∑ υ ′b ⋅ B b xt + ξt (2.27)
b=0
with xt = (x1,t , ..., xM,t )′ the vector of control variables and υ b = (υ1,b , ..., υM,b )′ the vector
of regression parameters for lag b. For the noise term the same remarks hold as for the
SISO model.
MIMO model
The last and most general class of processes combines both previous models. Here, inputs
and outputs are multivariate leading to the following multivariate regression model
N
yt = μ + ∑ Υb ⋅ B b xt + ξ t (2.28)
b=0
with x = (x1,t , ..., xM,t )′ the vector of control variables, yt = (y1,t , ..., yL,t )′ the vector of
output variables, Υb = (υ 1,b , ..., υ L,b ) = (υ1,1,b , ..., υ1,L,b , ..., υM,L,b ) the matrix of regression
parameters for lag b, and ξt = (ξ1,t , ..., ξL,t )′ the vector of noise. Again Ξ is some L × T
matrix sampled from a coloured or white L-dimensional stochastic process.
64
E.g. see Dayal and MacGregor (1996) for more information.
65
Note that there is still the option to ignore the time-correlation in favour of increasing the input lag
N . It can be shown that this model yields the correct estimates asymptotically, see Ljung (1999).
However, the estimates based on finite samples may be poor, see Dayal and MacGregor (1996).
25
with
t ∼ wn(0, σ 2 )67 for all t ∈ {p+1, ..., T }. This denotes an autoregressive process of order
p (in short AR(p)). Typically, the stochastic nature of the analysed process is assumed
to be stable in time. I.e. the stochastic characteristics (such as variance, autocorrelations
and mean) of the recorded process do not change in time. This is called the stationarity
premise which implies some restrictions for the parameters φi . E.g. suppose an AR(1)
process with φ = 1.68 It follows
t
yt = yt−1 +
t = yt−2 +
t−1 +
t = ... = ∑
i . (2.30)
i=0
66
Alternatively, the noise type can be ignored applying so-called non-parsimonious finite impulse-response
(FIR) models by increasing N , the number of lagged observations of the control variable(s). This
method has some serious drawbacks e.g. biased and instable estimates/forecasts in case of finite
samples. See e.g. Dayal and MacGregor (1996) and references therein.
67
The term wn(μ, σ 2 ) refers to a white noise process with zero mean and (constant) variance σ 2 .
68
This defines a so-called random walk.
26
t t
E(yt ) = E (∑
i ) = ∑ E(
i ) = 0 (2.31)
i=0 i=0
t t
V ar(yt ) = V ar (∑
i ) = ∑ V ar(
i ) = t ⋅ σ 2 (2.32)
i=0 i=0
t s t s
Cov(yt , ys ) = Cov (∑
i , ∑
i ) = ∑ ∑ Cov(
i ,
j ) = t ⋅ σ 2 with s ≥ t. (2.33)
i=1 i=1 i=1 j=1
Hence, this process is not stationary due to the time-varying, unbounded variance and
covariance.69 Only AR(1) processes with ∣φ∣ < 1 are stationary. This result can be gener-
alized for any AR(p) model for which the following conditions must hold:70
p
∑ φi < 1 (2.34)
i=1
∣φi ∣ < 1 i = 1, ..., p. (2.35)
In the moving average (MA) model yt depends on the previous errors
t−i instead of
previous observations. An MA model of order q (MA(q)) is represented by
q
yt = ∑ θj ⋅
t−j +
t (2.36)
j=1
with
t ∼ wn(0, σ 2 ) for all t ∈ {1, ..., T }. For MA processes no stationarity conditions have
to be considered, even an infinite MA(∞) process remains stationary.71 MA processes
have another special characteristic: their non-uniqueness with respect to the (first lag)
auto-correlation ρ. Suppose a time series generated by an MA(1) process with parameter
θ: yt = θ ⋅
t−1 +
t . The autocorrelation ρ = ρ (yt , yt−1 ) can be expressed as
V ar(yt ) = (1 + θ2 ) ⋅ σ 2 , (2.38)
Cov (yt , yt−1 ) = θ ⋅ σ 2 , (2.39)
θ
ρ = . (2.40)
1 + θ2
69
The covariance expression can be understood intuitively as the sum of variances due to the error terms
commonly shared by both variables yt and ys . For more detailed deductions see Cryer and Chan
(2008, pp. 12-13) or Brockwell and Davis (2002, pp. 16-17).
70
These conditions are necessary but not sufficient. Precisely, for the characteristic equation φ(x) = 0 all
roots must have absolute value greater 1 (with the characteristic polynomial φ(x) = 1 − φ1 ⋅ x − φ2 ⋅ x2 −
... − φp ⋅ xp ), see Schlittgen and Streitberg (2004, pp. 121-132) or Cryer and Chan (2008, pp. 72-77).
71
See Cryer and Chan (2008, pp. 55-56).
72
See e.g. Cryer and Chan (2008).
27
However, setting θ′ = 1θ yields exactly the same auto-correlation ρ. Hence, there are at least
two possible values for θ producing exactly the same time series w.r.t. the auto-correlation
structure.73 This problem is related to the stationarity condition of AR processes. To solve
this problem the invertibility condition is introduced. An MA process must be invertible
into an infinite AR process. This holds if and only if the characteristic equation for the
characteristic polynomial θ(x) = 1 + θ1 ⋅ x + θ2 ⋅ x2 + ... + θq ⋅ xq has roots with absolute value
larger than 1.74 Given a special MA process with known order and unknown parameter
(set), there exists only one parameter (set) such that this MA process is invertible.75
Both types of auto-correlation models are rarely found in real world problems in genuine
form, but in combination they build a huge class of time series patterns summarized as
so-called ARMA models. An ARMA(p,q) model can be formalized as:
p q
yt = ∑ φi ⋅ yt−i + ∑ θj ⋅
t−j +
t (2.41)
i=1 j=1
with
t ∼ wn(0, σ 2 ) for all t ∈ {1, ..., T }. To assure stationarity and invertibility the process
has to satisfy stationarity conditions from its AR part and invertibility conditions from
its MA part.76 Basic ARMA models only include ex ante stationary models but, by little
extensions, this class can be extended to comprehend also difference stationary time series.
Consider again a random walk model (which is clearly non-stationary):
yt = yt−1 + t . (2.42)
Taking the first differences results in a stationary time series since the white noise process
remains
Δyt = yt − yt−1 =
t . (2.43)
The time series Δyt is stationary. Hence, the original series yt is called difference station-
ary.77 Differencing (i.e. the concept of taking differences) extends the ARMA class to the
most popular ARIMA class where ARIMA stands for autoregressive integrated moving
average. The concept of differencing can be extended to higher orders leading to complex
mathematical formulation in traditional notation.78 Therefore, the general ARIMA model
is formulated compactly as
φ(B)(1 − B)d yt = θ(B)
t (2.44)
where φ(B) and θ(B) denote the characteristic polynomials for the AR and MA part as
73
For details see Schlittgen and Streitberg (2004, p. 117).
74
See Schlittgen and Streitberg (2004, pp. 124-132).
75
See e.g. Cryer and Chan (2008, p. 80).
76
For details see Shumway and Stoffer (2006, pp. 93-97).
77
Alternatively, also the term integrated of order 1 is used. This term suggests that this concept can be
applied to differenced time series again, see e.g. Pfaff (2006, pp. 23-24).
78
For more details about higher-order differencing see e.g. Shumway and Stoffer (2006, pp. 98-103) or
Cryer and Chan (2008, pp. 87-98).
28
described above.
A useful extension to classic ARIMA models are models incorporating exogenous ex-
planatory variables, so called ARIMAX models. As the name suggests, additional param-
eters are added to the ARIMA model. The use of explanatory variables usually results
from knowledge about the process’ external dependencies, but sometimes it is also reason-
able to model outliers or structural changes of the time series using auxiliary exogenous
variables.79
Assume the SISO model (2.25) where ξ is coloured noise and follows an ARMA process
with lag polynomials φξ (B) and θξ (B). Without loss of generality μ is assumed to be
zero and (2.25) changes to
N
φξ (B)yt = ∑ υb B b xt + θξ (B)
t (2.45)
b=0
N
θx (B)
φξ (B)yt = ∑ υb B b εt + θξ (B)
t (2.46)
b=0 φx (B)
where εt is a white noise series.
The first approach to solve such a problem is to perform a regression analysis be-
tween the time series y and the regression variable x ignoring the fact of auto-correlated,
stochastic variables. Afterwards, the residuals from this first step could be obtained and
a time series could be fitted for these residuals, e.g. using ARIMA models. Unfortunately,
this approach can lead to biased and inefficient estimates even if the sample is large.81
Due to the time series characteristics of the dependent and independent variables, the
cross-correlations between x and y might be overlaid by the individual temporal depen-
dencies of x and y. To obtain (reasonable) estimates for the impulse response parameters
′
υ = (υ0 , ..., υN ) , the time series y has to be "cleaned" from the time series effects of the re-
gressor variables x. This procedure is called prewhitening and performed in the following
steps:
In the case of the MISO model with multiple, say M , input variables the prewhitening
approach can be easily adapted by iteratively performing steps 1 and 2 of the prewhitening
procedure for all input variables to obtain estimates for υ i , i = 1, ..., M . The residual series
is finally calculated by ξˆt = yt − ∑N
b=0 υ̂1,b B x1,t − ... − ∑b=0 υ̂M,b B xM,t .
1 b NM b 83
where W (t) is the Wiener process which is characterized by the following properties:84
• P (W (0) = 0) = 1,
• W (t) has normally distributed increments W (t) − W (s) ∼ N (0, t − s) with 0 ≤ s < t,
It can be shown that for a linear SDE with coefficient functions ci (t) (which are con-
tinuous in t) an explicit, unique solution always exists and the SDE has finite first and
second-order moments.85
Time series models and SDEs deal with the same sort of stochastic process. Both differ
only in the domain of variables, which are either discrete or continuous. For instance,
chemical processes are continuous by nature. In practice, however, the condition of a
82
Details about prewhitening can be found e.g. in Box et al. (2008, p. 417 ff.) or Cryer and Chan (2008,
p. 265 ff.).
83
See e.g. Wei (1990, p. 328 ff.).
84
See Hassler (2007, p. 117) and Iacus (2008, p. 18 ff.)
85
See Hassler (2007) and Øksendal (2003).
30
chemical reaction is measured by various sensors in discrete time intervals. Hence, the
question arises whether discrete time series models can be seen as discretized SDEs.
It can be shown that ARCH and GARCH models are able to approximate stochastic
differential processes if the latter fulfil certain properties.86 Albeit the goodness of fit
is limited,87 both types of methods are related and can be converted into each other.88
Moreover, simple stochastic processes show quite simple auto-correlation structures simi-
lar to basic ARMA models. For instance, the Ornstein-Uhlenbeck process can be seen as
the continuous equivalent of the AR(1) process. In other words, an Ornstein-Uhlenbeck
process measured in discrete intervals can be interpreted/modelled as an AR(1) process
(see also (2.23), (2.60), and (2.61)).89
This subsection briefly introduces GARCH models as discrete counterpart of continuous
stochastic processes. In contrast to ARMA models, the basic idea is that the variance/
volatility in time is no longer deterministic and constant but depends on previous errors
and volatility, i.e.
y t = μ +
t (2.48)
σt2 = f (σt−1
2 2
, ..., σt−p ,
t−1 , ...,
t−q ) (2.49)
t ∼ wn(0, σt2 ). (2.50)
The model above is called a GARCH(p,q) model if f (⋅) is a linear function such that
(2.49) changes to
If the volatility depends only on the previous errors
i (i.e. q = 0), the corresponding
models are called ARCH(p). (G)ARCH processes are known to be weakly stationary if
q p
∑ αi + ∑ βi < 1.90
i=1 i=1
(G)ARCH models have become very popular in empirical economic research. In the
past decades a vast amount of extensions and generalizations have been published.91
These numerous adaptations of the basic model consisting of (2.48), (2.50) and (2.51)
are able to incorporate very specific time series characteristics and are usually driven by
empirical observations and experiences.92 A relevant extension in the context of chemical
86
See Nelson (1990) and Duan (1997).
87
See Wang (2002).
88
E.g. see Fornari and Mele (2001) for an application of GARCH models as diffusion approximations
fitted to financial data sets.
89
See e.g. Hassler (2007, p. 236) about the similarity between both processes and Uhlenbeck and Ornstein
(1930) or Vasicek (1977) for details about the Ornstein-Uhlenbeck process.
90
E.g. see Hassler (2007, p. 96). To assure stationarity for higher moments (>2) more restrictive condi-
tions have to be met, see Ling and McAleer (2002).
91
See Engle (2002) for a review.
92
E.g. see Bauwens et al. (2006) for a review of recent multivariate GARCH models and Bollerslev (2008),
Hansen and Lunde (2005) or Degiannakis and Xekalaki (2004) for a review of univariate (G)ARCH
31
production processes are so-called GARCH-M and ARMA-GARCH models. For this class
of models the mean process in (2.48) is not a constant, but a stochastic process itself.
The GARCH-M(p,q) mean process is modelled as function of σ:93
yt = μ ⋅ σt + t . (2.52)
P Q
yt = ∑ φi ⋅ yt−i + ∑ θj ⋅
t−j +
t (2.53)
i=1 j=1
This class of models allows formulating a great variety of dependency patterns and is used
in many empirical applications.94 Its most appealing advantage is the possibility to model
both a time-dependency structure for the observed series and for their volatility combined
in one model. For modelling chemical production processes, the latter property is useful
for describing the variations in a chemical process’ equilibrium. In normal operation,
chemical reactions take place under controlled physical and chemical conditions. The
parameters of these reactions such as conversion rates and reaction times depend on
these conditions and determine the stochastic processes of output and input measures.
Variations in process control variables affect these conditions and have an immediate
impact on the parameters of the underlying reactions. This may change the structures
of the input and output processes. As a consequence, a destabilization of the chemical
process might occur resulting in an increase in the measures’ volatility.
The incorporation of exogenous regressors (leading to ARMAX-GARCH models) is less
emphasized in the literature since it is similar to ARMAX models. In this case (2.53) is
substituted by (2.45). For estimation of the corresponding parameters, prewhitening can
be applied to eliminate the influence of the exogenous time series variables. Afterwards,
the ARMA-GARCH model can be estimated on the residuals.95
models.
93
See Hassler (2007).
94
See Li et al. (2002) for an extensive review and references therein for applications.
95
It has to be noted that the estimation of ARMA-GARCH models requires some more sophisticated
methods compared to simple ARMA models, see Francq and Zakoian (2004).
32
concepts can be transferred more or less easily. This section briefly introduces the key
models for multivariate time series models, so-called vector autoregressive moving average
processes (VARMA) and their extensions with exogenous regressors (VARMAX). Multi-
variate GARCH processes (MGARCH) are not discussed in detail here.96
Suppose a multivariate time series of dimension L collected at T points of time, say
Y ∈ RL×T . The basic types of dependency, auto-regression and moving average, are used to
formulate a model explaining the behaviour of this multivariate time series. Suppose that
the observations of Y at a given time t, say yt ∈ RL , depend on the previous observations
yt−i with i = 1, ..., p and previous error terms, say t−j with j = 1, ..., q. This relation can
be formulated as a VARMA(p,q) model
p q
yt = ∑ Φi yt−i + ∑ Θj t−j + t (2.56)
i=1 j=1
p q
⇔ yt = ∑ Φi B i yt + ∑ Θj B j t + t (2.57)
i=1 j=1
96
For an introduction see e.g. Bauwens et al. (2006).
97
About the stationarity and invertibility conditions see e.g. Wei (1990, pp. 335 ff.).
98
For an application of an MIMO model in chemical process modelling see Barceló et al. (2011).
99
For more details about the relationship between multivariate finite impulse response models and (par-
simonious) VARMAX models in chemical process analysis, see e.g. Seppala et al. (2002).
100
See e.g. Hall and Nicholls (1980) for a direct maximum likelihood estimation or Metaxoglou and Smith
(2007) for estimation based on a transformation into state-space models.
101
I.e. these variables can be included in Y.
33
number of parameters to be estimated and, hence, may hamper the model’s feasibility.102
If exogeneity can be assumed (e.g. in case of control variables), a direct modelling of the
exogenous variables’ stochastic models is the preferred option. Similar to the prewhiten-
ing approach, a time series model for the exogenous variables has to be fitted which can
be used to improve the estimates of the original VARMA model.103
Let Δt denote the recording interval and let h = m ⋅ Δt be a multiple of Δt. Then
yt+m = y(t + m ⋅ Δt) is the discrete equivalent of y(t + h). For the autocorrelation of two
measures holds
m
ρ(y(t), y(t + m ⋅ Δt)) = ρ(yt , yt+m ) = e−k⋅m⋅Δt = (e−k⋅Δt ) = φm . (2.61)
Note that (2.61) represents the autocorrelation function of an AR(1) process with pa-
102
See e.g. Horváth (2003).
103
In the state-space reformulation of the VARMAX model, the time series model of the exogenous
variables is used to predict initial values of the regressor matrix, see Metaxoglou and Smith (2007).
104
Intervals about or shorter than one second are frequently found, see Darby et al. (2009).
105
I.e. the parameters p and q in the ARIMA models introduced before, see e.g. (2.41).
106
For the underlying expressions for asymptotic variance and autocovariance, see Hassler (2007, p. 237).
34
107
However, there is typically no possibility to check these assumptions explicitly.
108
In case of over-specification (synonymously over-fitting) additional variables or parameters are incor-
porated in the model which show a spurious correlation with the error terms. This may lead to less
efficiently estimated parameters and poor prediction accuracy, see Cryer and Chan (2008, pp. 185-188)
or Schlittgen and Streitberg (2004, pp. 347-349).
109
See e.g. Schlittgen and Streitberg (2004, pp. 332-345).
35
be expressed as
K
IC = −2 ⋅ ln (LT (K)) +
⋅ κ(T ) (2.62)
T
where K denotes the number of parameters fitted in the model, κ(T ) is a function of the
number of observations T (this function distinguishes the various criteria), and LT (K)
denotes the model’s likelihood. The number K of parameters to be estimated depends
on both, the order of the model and the number of exogenous variables. For example,
assuming a VARX(p) of dimension d with non-zero mean and k = 1 exogenous variable
(which has a contemporary effect only), the number of parameters to be estimated is
K = p ⋅ d2 + d + d.
From (2.62) it can be taken that the model’s goodness of fit and the number of parame-
ters used are counterbalanced. Among a set of model specifications, the specification with
minimal IC value is recommended. In other words, information criteria aim at minimiz-
ing the residuals’ variance with as few parameters as possible. Often used information
criteria for time series models are the Akaike information criterion (AIC),110 the Schwarz
information criterion (SIC)111 or the Hannan-Quinn information criterion (HQIC)112 with
the following κ functions:
⎧
⎪
⎪
⎪ 2 for AIC
⎪
⎪
κ(T ) = ⎨ ln (T ) for SIC . (2.63)
⎪
⎪
⎪
⎪
⎩ ln (ln (T ))
⎪ for HQIC
Among these IC, the AIC has a non-zero asymptotic probability for over-fitting.113 To
avoid an over-fitting of the model order, a corrected AIC (AICc ) can be calculated by
AICc = AIC + 2⋅(K+1)⋅(K+2)
T −K−2 .114
The SIC is deduced from Bayesian arguments. It consistently estimates the true order
of ARMA(p, q) processes and is probably the most widely used information criterion in
univariate time series analysis.115 The HQIC is the most recent IC and especially designed
for multivariate time series models.116 In practice, multiple ICs are simultaneously cal-
culated which allows the analyst to cross-check the recommendations of the various ICs.
Strongly deviating recommendations may indicate an inappropriate model structure.
One source of mis-specification is the type of the error process. Homogeneity of vari-
ances is typically assumed by standard time series models. However, in the context of
chemical production processes e.g. an adjustment of a plant’s production rate may lead
to an imbalance of the underlying chemical reaction(s). This instability may materialize
in fluctuations of the output flow rate(s) which may also affect flow rates in subsequent
110
See Akaike (1974).
111
This criterion is often called Bayesian information criterion (BIC), see Schwarz (1978).
112
See Hannan and Quinn (1979).
113
In other words it is biased, see Cryer and Chan (2008, p. 131).
114
See Hurvich and Tsai (1989).
115
See Cryer and Chan (2008, p. 131).
116
See Quinn (1980).
36
of an empirical time series by the standard sample correlation formula.117 These values
are usually displayed in a correlogram. If the auto-correlation for a specific lag exceeds
some critical value,118 a significant auto-correlation is to be assumed. Hence, the ACF
allows detecting the order of pure MA processes which can be easily seen from the formal
definition of MA processes in (2.36). I.e. the ACF of an MA(q) process theoretically shows
significant values for the first q auto-correlations.119
To detect AR-like correlations, the PACF corrects the ACF values by the effect that is
caused by preliminary auto-correlations120
the MA order q. To determine a time series’ ARMA order, the coordinates of the upper
left vertex of a triangle of os correspond to the order (p, q) of the underlying ARMA
process.123
ACF, PACF and EACF are applied to time series data to initially estimate the type and
order of the underlying time series model. Once an initial model is fitted, ACF, PACF and
EACF are applied to the corresponding residuals to check the white noise assumptions. If
significant auto-correlations are found, the model type or order is adjusted accordingly and
the procedure is repeated until the residuals do not show any serious auto-correlations.
A normality test such as Jarque-Bera’s test or graphical tools such as QQ-plots are also
part of the standard procedure of residual checks. Although normality is often assumed to
derive arguments for theoretical deductions, this is not a necessary condition for consistent
estimates.124 Nonetheless, if normality can be assumed, an identical distribution of the
residuals and variance homogeneity follows immediately. However, if normality cannot be
confirmed, the weaker concept of identical variances can be checked by calculating ACF,
PACF and EACF for the squared and absolute residuals of a model.125 If noticeable
correlations occur, heterogeneous variances must be presumed and the time series model
should be extended by a sub-model for the variances (i.e. an ARCH or GARCH model).
The following examples illustrate how the above mentioned methods can be used to
find appropriate time series models for empirical samples. The first example shows the
application of the methods by means of a simulated univariate time series. This example
emphasizes the effect of the length of time series and gives an impression how to interpret
the described graphical tools. The second example describes the analysis of a sample of
the input flow rate of a cracker plant. This example focuses on practical problems (such as
outliers) in analysing even simple univariate time series of chemical production processes.
The third example shows how the transformation process of a chemical production plant
can be modelled using multivariate time series methodology based on a real-world sample
from a de-alkylation plant. Again practical problems such as the determination of the
model order and multivariate outliers are discussed.
123
For details about the deduction of the EACF test statistic, see Tsay and Tiao (1984) and for a practical
overview see e.g. Cryer and Chan (2008, pp. 116-118) or Tsay (2002, pp. 51-53).
124
Consistent estimates can be obtained for independent and identically distributed residuals with quasi-
maximum likelihood estimators and Gaussian MLE, see Yao and Brockwell (2006).
125
See Cryer and Chan (2008, pp. 277-280).
38
1.0
1.0
0.5
0.5
0.5
ACF
ACF
ACF
0.0
0.0
0.0
−0.5
−0.5
−0.5
5 10 15 20 5 10 15 20 5 10 15 20
Lag Lag Lag
(a) ACF for T = 100 (b) ACF for T = 1000 (c) Theoretical ACF
1.0
1.0
1.0
0.5
0.5
0.5
Partial ACF
Partial ACF
partial ACF
0.0
0.0
0.0
−0.5
−0.5
−0.5
5 10 15 20 0 5 10 15 20 25 30 5 10 15 20
Lag Lag Lag
(d) PACF for T = 100 (e) PACF for T = 1000 (f) Theoretical PACF
Figure 2.10: Theoretical and empirical ACF and PACF for an ARMA(2,1) process
The number of observations available for estimation considerably influences the em-
pirical ACF and PACF. For the small sample with T = 100 the ACF correlogram (Fig-
ure 2.10a) does not show any similarity to the corresponding theoretical ACF correlogram
(Figure 2.10c). None of the ACF correlograms shows an easily interpretable pattern. In
contrast, for all PACFs (Figures 2.10d - 2.10f) the first two lags show the highest cor-
relation pointing to a second order AR component. Tables 2.3 - 2.5 show the results of
the sample EACFs and the theoretical EACF for this example ("x" denotes a significant
correlation and "o" denotes a non-significant one).
For the small sample, the EACF values (Table 2.3) show an ambiguous pattern. Proba-
bly, an ARMA(1,3) or ARMA(3,1) model would be chosen as initial choices. In contrast,
the large sample (Table 2.4) shows a good fit to the theoretical EACF (Table 2.5) and an
ARMA(2,1) model seems an appropriate initial choice. Since the analysis of ACF, PACF
and EACF cannot reveal an unambiguous favourite model for both samples, all possible
order specifications with p ∈ {0, ..., 4} and q ∈ {0, ..., 4} are estimated for both samples.127
The corresponding AICc , SIC, and HQIC values are displayed in Table 2.6 where the min-
imum for each criterion and time series is set in bold font and the true model order is
coloured in grey.128
126
The samples are simulated by means of the arima.sim function with R 2.15.1.
127
For all models no intercept is estimated.
128
Only a subset of all tested orders is displayed in the table for the sake of brevity (q ≤ 1).
39
AR lag
AR lag
2 o x o x o o o 2 x o o o o o o 2 x o o o o o o
3 x o o o o o o 3 x x o o o o o 3 x x o o o o o
4 x x o o x o o 4 x x x o o o o 4 x x x o o o o
5 x x x o o o o 5 x o o x o o o 5 x x x x o o o
6 x x o x o o o 6 x o o x o o o 6 x x x x x o o
Table 2.3: EACF for T = Table 2.4: EACF for T = Table 2.5: theoretical
100 1000 EACF
AR MA T = 100 T = 1000
order order AICc SIC HQIC AICc SIC HQIC
1 0 4.273 4.175 4.160 4.486 4.479 4.476
2 0 4.141 3.943 3.912 4.272 4.258 4.252
3 0 4.330 3.987 3.941 4.272 4.246 4.237
4 0 4.556 4.022 3.960 4.291 4.250 4.238
0 1 4.258 4.160 4.145 4.463 4.456 4.453
1 1 4.197 3.999 3.968 4.277 4.262 4.256
2 1 4.330 3.987 3.940 4.269 4.244 4.234
3 1 4.566 4.032 3.970 4.291 4.250 4.238
4 1 4.835 4.062 3.984 4.317 4.257 4.242
Table 2.6: Information criteria for both sampled time series and various model specifica-
tions
For both samples all three criteria receive their minimum for a unique model order. In
the small sample case, all criteria suggest an ARMA(2,0) whereas in the large sample
case an ARMA(2,1) is suggested (which is the correct order). Fitting these models to both
time series and analysing the corresponding residuals results in diagnostic plots without
remarkable features as Figures 2.11a - 2.11f show.
For the residuals’ ACFs only spurious significant auto-correlations occur at lag 9 for
the small sample and lag 30 for the large sample. Hence, it can be concluded that no auto-
correlation remains and the model successfully captures the time-dependency pattern.129
For the squared residuals’ ACFs, slightly significant auto-correlations occur at lag 11 for
the small sample and at lags 6, 17, and 26 for the large sample. Therefore, also identical
variances of the residuals can be assumed.130 In summary, identically and independently
distributed residuals can be assumed for both models. This is also confirmed by the QQ-
plots (Figure 2.11c and Figure 2.11f) which show a good fit to the normal distribution for
both models. This confirms the assumption of independently and identically distributed
residuals indicating adequate models. Note that for the small sample the true model order
129
Since the ACFs show no attractions, the PACF need not to be analysed.
130
Also the ACFs of absolute residuals show no serious attractions.
40
is not identified despite the fact that the residual analysis reveals no hints for an inaccurate
model specification.
0.4
0.4
4
2
0.2
0.2
0
0.0
0.0
ACF
ACF
−2
Sample Quantiles
−0.2
−0.2
−4
−0.4
−0.4
5 10 15 20 5 10 15 20 −2 −1 0 1 2
Lag Lag Theoretical Quantiles
(a) ACF of residuals for T = 100 (b) ACF of squared residuals for T = 100 (c) QQ-plot of residuals for T = 100
0.4
0.4
5
0.2
0.2
0
0.0
0.0
ACF
ACF
Sample Quantiles
−0.2
−0.2
−5
−0.4
−0.4
5 10 15 20 25 30 5 10 15 20 25 30 −3 −2 −1 0 1 2 3
Lag Lag Theoretical Quantiles
(d) ACF of residuals for T = 1000 (e) ACF of squared residuals for T = 1000 (f) QQ-plot of residuals for T = 1000
Figure 2.11: ACF of residuals, ACF of squared residuals and QQ-plot of residuals for ARMA(2,0) model (small sample, T = 100) and for
ARMA(2,1) model (large sample, T = 1000)
41
42
Example 2 (Univariate time series analysis for chemical production data). An application
of univariate time series analysis is the modelling of univariate input flows of chemical
production processes (as observed for SISO or SIMO processes). To apply prewhitening
procedures, the auto-correlation structure of the (exogenous) time series of the input flow
has to be modelled. In this example, the input flow rate of a Naphtha cracker is analysed.131
Figure 2.12 shows the average hourly input flow rate for one week (T = 168).
41.62
input flow rate in t/hour
41.60
41.58
41.56
0 50 100 150
Time
The time series shows a stationary pattern with a homogeneously oscillating flow rate
around a constant level. However, an additive outlier occurs at time index 128. To
analyse the time dependency pattern of the Naphtha time series, the effect of the outlier
has to be removed. Therefore, the outlier’s effect (i.e. the raise beyond the mean level)
is estimated by modelling a simple linear regression model and the corresponding value of
the time series is corrected by this estimate. For the corrected Naphtha time series, ACF
and PACF are calculated (Figure 2.13a and Figure 2.13b).
The ACF shows a decaying cyclic pattern whereas the PACF clearly indicates a signifi-
cant partial auto-correlation at lag 1. Therefore, an AR(1) model is chosen as the initial
model. To incorporate the outlier at index 128, an ARX(1) model is fitted to the original
Naphtha time series where a dummy variable xt accounts for the effect of the outlier132
yt = μ + φ ⋅ yt−1 + β ⋅ xt + t . (2.66)
The estimated parameters of the ARX(1) model are displayed in Table 2.7.
The corresponding residuals’ standard deviation is estimated as σ̂ = 0.002 and the
model’s AICc is calculated as -8.73. An analysis of the residuals reveals no hints for a vio-
lation of the underlying assumption of independent and identically distributed residuals.133
131
Data is provided by Dow Chemical. The scale of the time series has been transformed due to reasons
of privacy protection.
132
The dummy variable is binary, i.e. xt = 1 for t = 128 and zero otherwise.
133
Diagnostic plots can be found in the Appendix in Figure A.1.
43
1.0
0.8
0.6
ACF
0.4
0.2
0.0
0 5 10 15 20
Lag
(a) ACF of corrected Naphtha time series
0.8
0.6
Partial ACF
0.4
0.2
0.0
5 10 15 20
Lag
(b) PACF of corrected Naphtha time series
Jarque-Bera’s test for normality confirms that the hypothesis of normally distributed resid-
uals cannot be rejected (at a 5% level), i.e. it is assumed that
ˆ ∼ N (0, 0.0022 ). Due to
the already convincing results of the ARX(1) model (which has minimal order) no other
ARMA models of higher order are estimated and (2.66) is taken as the best model for the
Naphtha time series.
parameter μ̂ φ̂ β̂
estimate 41.6 0.79 0.0564
standard error 0.001 0.046 0.002
Toluene + Xylene
Methane
Benzene
H2
inflow mixture
All chemicals entering or leaving this plant are recorded. Figure 2.15a shows the average
hourly inflow and outflow rates for two weeks (T = 336).137 Although hydrogen is a
necessary component to undergo the relevant chemical reactions, the external inflow of
hydrogen has shown no individual effect on the outflow rate. Thus, external hydrogen
supply has been dropped from the further analysis. The outputs are pure Benzene and the
134
See e.g. Ma and Trimm (1996).
135
See Ozokwelu (2006).
136
See e.g. Ozokwelu (2006).
137
This data is provided by Dow Chemical. The scales of all variables are transformed due to privacy
protection reasons.
45
● ●
● ●
●●●
● ●
●
● ●●
●●
●
●●
●
●
●● ●●
120
●●● ●
●
●●●
● ●●
●
●
● ●
●●
130
● ● ●●
●●
●
●● ●
●
●
●
●
● ●●●
● ●●
●
●●
●
●● ● ●
● ● ● ●● ●
●
●●
●
●
● ● ●
●
● ●
input rate
● ● ●●
●
●●
●●● ●
●● ●●● ●
●●
●
● ●●
● ●● ●
● ● ●
●●●
●●●● ●
●●● ●●
●●
● ●
●● ● ●
Benzene rate
●●●● ● ● ● ●● ●● ●
120
● ●● ●● ● ●●
●● ●
●
●
●●
● ●●●
110
●
●●
● ● ●
● ● ● ●
●
●
●● ●●
● ●
●
● ● ● ●
● ●
●
●●
● ●
●●
●
●
●
●●
● ●
●
●●●
110
● ●
● ●●
●
● ● ● ● ●
● ●
● ●
● ● ●
● ● ●
100
●
●
● ●
120
●
● ●
Benzene rate
● ●●
● ●
●●
90
●●
●
●●●
●
110
● ● ●
3.0 ●
● ●● ●
100
● ●
● ●●
● ● ●●
●●
● ●
●
●
● ●●
● ●
●●
●
●●
● ●
● ● ●●●
● ●
●●
●●● ●
●●
●
● ● ●
●●
●
●
● ●●●
● ● ● ●
●
90
●●
●●
● ●
●● ● ● ● ●
●● ●
●●
●●● ●●●
●●●●●● ●●
● ●
●
●
●
●
●
● ●●
●●
● ●
● ●● ●●
C9 rate
● ● ●●
●
●●●
●● ● ●● ●
●
2.5
●●
●● ●
●● ● ●
●
● ● ● ●
●●
3.0
●●● ● ● ●● ● ● ●
● ● ● ●●●●
● ● ● ● ●
●
●
●● ● ●●
●
●
●●● ●● ● ●● ●
●●
●●
●
● ● ● ● ●
●
●● ● ●
● ● ● ●●
C9 rate
●● ●
●
● ●● ●● ●
●● ●
●
●● ● ● ●●
●
●
● ● ● ●
2.5
●●
●
● ● ●
● ● ● ●
●●● ●
●
● ●
●
2.0
● ●
●
● ●
● ● ● ● ●
●
● ●
●● ●
●
2.0
●● ● ●
●
●
● ●
●●
●
●
0 50 100
200 150
250 300 110 120 130
Time input rate
(a) Inflow and outflow rates for de-alkylation plant (b) Scatter plots for Benzene and C9 outflow rates
dependent on inflow rate
From Figure 2.15b it can be observed that the output rate of Benzene depends linearly
on the input stream, while the C9 outflow rate more likely shows a slightly quadratic
dependency pattern. Furthermore, from the general nature of chemical production plants,
an auto-correlation structure of both output components can be presumed.138 Let Y be
the matrix of both output flow rates with dimension 336 × 2. yt ∈ R2 is the vector of both
output rates at time index t. The inflow rate x is a vector of length 336 and xt denotes
the inflow rate at time index t. To model the outflow rates yt , a vector auto-regression
(VARX) model is supposed with inflow rate x and squared inflow rate x2 as exogenous
variables. Figure 2.15a clearly indicates a non-stationary time series of inflow rates which
is primarily characterized by multiple changes of its mean. This can be interpreted as
actions to control the plant’s performance. An analysis of the stationary parts of the
inflow time series shows no indication for auto-correlation structures. It is assumed that x
follows a white noise process with varying means and is handled as an exogenous regressor.
138
A raw initial analysis by checking the residuals of a simple regression of the input rate on both output
rates confirms this presumption. The results are not shown here for the sake of brevity.
46
To select an appropriate order for the supposed VARX model, models of various orders are
fitted and compared based on standard information criteria.139 The information criteria
suggest a model order of p = 3 or p = 4.140
For reasons of parsimony a V ARX(3) model is chosen, i.e.
Table 2.8: Coefficients of the initial V ARX(3) model for de-alkylation plant
Afterwards, the residuals are analysed for deviations from the underlying assumptions
by checking for remaining correlations and normality. Checking for remaining correla-
tions among the residuals is performed by analysing the auto-correlation for both residual
components and the cross-correlation between both components. These are shown in Fig-
ure 2.16a. Checking for normality is done by analysing the quantile-quantile plots (QQ-
plots) for both residual components (shown in Figure 2.16b) and a Jarque-Bera-Test for
normality.
From Figure 2.16a no clearly significant auto- and cross-correlations are obvious. At
lag 16 of the ACF of the C9-residuals and at lags 0 and -6 of the residuals’ CCF slightly
significant correlations can be found (at a 5% level). It is to conjecture that the significant
findings are spurious since it is hardly reasonable to assume a delay of 6 or 16 hours
before the outflow rates are affected.
The QQ-plots displayed in Figure 2.16b suggest a good fit to the normal distribution.
However, the multivariate Jarque-Bera-test rejects the normality hypothesis (at a 5% α-
level). This result is probably caused by some fluctuations in the outflow rates occurring
at changes of the production level (e.g. see time intervals 40-70, 100-170 and 270-310).
These outliers might be caused either by process instabilities or measurement errors. It
is to conjecture that a better fit to the normal distribution can be achieved by explicitly
139
Computations are performed using the function VARselect from the vars package with R 2.15.1. See
Pfaff (2008).
140
The detailed results are kept for the appendix in Table A.1.
141
Computations are performed using the VAR function from the vars package with R 2.15.1, see Pfaff
(2008). The model’s ANOVA can be found in Table A.3 in the appendix.
0.3
Benzene residuals
6
Benzene residuals
0.2
4
0.1
2
−0.1
0
autocorrelation
−2
0.3−0.3
C9 residuals
Sample Quantiles
−4
0.2
0.1
−6
C9 residuals
0.2
−0.1
autocorrelation
0.1
−0.3
0 5 10 15 20 25
Lag
0.3
Benzene & C9 residuals
0.0
0.1
Sample Quantiles
−0.1
−0.1
crosscorrelation
−0.2
−3 −2 −1 0 1 2 3
−0.3
−20 −10 0 10 20 Normal Quantiles
Lag
(a) Autocorrelation and cross-correlation for residuals of both outflow (b) QQ-plots of residuals’ marginal distributions for both outflow rates
rates
Figure 2.16: Residual diagnostic plots for V ARX(3) model of the de-alkylation plant.
47
48
199
●
● ordinary residual
0.2
outlying residual
64
● 61
158 ● ●
●
● ●
●
●
●
● ●
● ● ● ● ●● 302
249 ●
0.1
●
● ●● ●
● ● ●
● ● ● ● ●● ● ●
residuals (C9)
● ● ● ●
● ●
●● ●
● ●
● ● ●
● ● ●
●● ●● ●●● ●
●● ● ● ● ●
● ●
51 ● ●● ●● ●
● ●● ●
● ● ● ● ● ●
● ● ● ● ● ●● ●●●
●●
● ●●
● ●
● ●
●● ●● ● ●● ●● ● ● ● ● ●
● ● ● ●●
●
● ●●●●● ● ● ● ●
●
● ● ● ● ●● ● ● ● ● ●● ●
0.0
●
●●● ● ● ●● ● ● ●●● ● ●
● ● ●●
● ● ● ● ● ●● ● ●● ● ● ●
●
● ●
● ●●●● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ●● ●
● ● ●● ● ●●
●● ● ● ● ● ● ●
● ● ● ● ● ●● ● ● ●●● ● ● 155
●●
● ● ●
●● ● ●● ●● ●
●●
●
● ●
●● ● ● ●
● ● ● ● ● ● ●
207
● ● ●● ● ● ● ●● ●● ● ●
● ● ●
● ● ●●● ● ● ●● ●
● ● ●●●
● ● ●
−0.1
297 ●
●
●●
● ● ●● ● ●
●● ●
●
● ● ●
●
●
● ●
20 ● ●
●
65 ●●
●
●
188 202 156 109
● ●
−0.2
● ●
−6 −4 −2 0 2 4 6
residuals (Benzene)
Figure 2.17: Scatterplot of residuals for outlier corrected V ARX(3) model of the de-
alkylation plant
Based on the estimated robust covariance matrix, the Mahalanobis distances for all residu-
als are calculated. All residuals whose Mahalanobis distances exceed a critical value based
on the 97.5% quantile of the χ22 -distribution are handled as outliers.143 For each identi-
fied outlier, a dummy variable is constructed which has entry 1 at the time index of the
identified outlier (and is zero otherwise).144 Figure 2.17 shows the pairs of residuals at
each time. Grey-coloured ◾-marked points refer to identified outliers whose time indices
are superimposed.
Subsequently, (2.67) is enhanced by the dummy variables for the outliers. To provide the
142
The robust estimator is the minimum covariance determinant estimator (MCD), carried out by covMcd
function with default settings from the robustbase package, see Rousseeuw and van Driessen (1999).
143
See e.g. Maronna et al. (2006) or Liebscher et al. (2012).
144
I.e. so-called innovative outliers are modelled. The procedure can be seen as a multivariate adaptation
of the univariate standard procedure for outlier detection as described e.g. in Cryer and Chan (2008,
pp. 257-260).
49
most parsimonious model, only parameters with significant effects are finally selected.145
The corresponding re-fitted coefficients of the selected model can be found in Table 2.9.146
Table 2.9: Coefficients of the V ARX(3) model with outlier correction and variable selec-
tion
While the estimates of the autocorrelation coefficients for the C9 time series (lower rows
in Φ̂1 to Φ̂3 ) only change slightly, the estimates the autocorrelation coefficients for the
Benzene time series (upper rows in Φ̂1 to Φ̂3 ) are clearly affected since three parameters
are dropped from the model. The remaining coefficients are affected, too. In particular,
the lagged cross-correlations to the C9 time series change from 1.67 to 2.51 and from
-2.91 to -2.67 (right upper entries in Φ̂1 and Φ̂3 ). This confirms the serious effect of even
unobtrusive outliers in multivariate times series analysis. By incorporating the outliers’
effects, the model’s AIC decreases from -4.22 to -4.72. Similarly, SIC decreases from -4.05
to -4.17. The analyses of residuals show a similar pattern as for the initial model and
reveal no serious hints for cross- or auto-correlation.147 Now, the multivariate Jarque-
Bera test does not reject the hypothesis of multivariate normally distributed variables (at
a 5% level). The residuals’ empirical covariance matrix is finally estimated as
Using this model, the outflow rates can be fitted quite well, as Figure 2.18 shows.
145
The estimated coefficients for the V ARX(3) model with outlier compensation can be found in the
appendix, Table A.2.
146
The model’s ANOVA can be found in Table A.4 in the appendix.
147
The corresponding correlograms can be found in the appendix in Figure A.2.
50
125
115
105
Benzene rate
real
fitted
outlier
90 95
3.0
2.5
C9 rate
real
2.0
fitted
outlier
Figure 2.18: Real and fitted time series of both outflow rates (outlier indices superimposed)
51
intra-site inter-site
S1 C1
intra-site
intra-site
S2 C2
This chapter deals with methods to plan the distribution of chemicals in chemical
production networks from a focal point of view. Specific planning problems occurring for
the transport modes pipeline, rail and ship are described.
T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,
DOI 10.1007/978-3-658-08433-2_3, © Springer Fachmedien Wiesbaden 2015
52
road
[43.6%]
road
[70.2%]
pipeline
[14.6%]
open sea open sea
[7.7%] [8.8%]
short sea
[5.8%] short sea
rail [9.6%]
pipeline [8.4%]
rail
[23.4%]
[8.0%]
(a) Modal split - all companies (total: 313 (b) Modal split - chemical companies (total: 107
mill. tons) mill. tons)
Figure 3.2: Modal split for chemical products in Germany in 2009 (based on total trans-
ported quantity in mill. tons)2
Figure 3.2a shows the modal split for all transports of chemical products in Germany
in 2009. This includes all transports along the complete supply chain, in particular final
chemical transports to customers as well as raw and intermediate chemical transports
1
See Shah (2005).
2
See VCI (2011).
53
between chemical companies. The modal split is fairly similar to the total modal split
over all products.3 Road transports are dominating with a share of about 70% of the
total transported quantity. This picture changes if only transports performed by chemical
companies are considered (Figure 3.2b). Here, the share of road transports is considerably
smaller (~43%). Most transports are handled via high-capacity transporters. E.g. 70%
of the total amount of transported alkenes such as Ethylene and Propylene are shipped
via pipeline.4 This shows the special relevance of these transport modes for the chemical
industry. It is to presume that this picture is even more biased in favour of high-capacity
transports in basic chemical industry.
For all considered means of transport the problem is to plan transport flows and local
stocks such that total costs for transports and stock holding in the production network are
minimized. Depending on the mean of transport, various technical restrictions have to be
considered. In the case of rail cars or ships, transport flows have to be accompanied by a
corresponding flow of transport carriers such that carrier flows are planned simultaneously.
Depending on the modelling of the carriers, the problem structure is similar either to
inventory-routing problems (IRP)5 or to capacitated network design problems (CNDP).
In the former case, routes for each carrier have to be determined whereas in the latter case,
carrier flows are modelled as dynamic arc capacities.6 The literature on IRPs considering
ship transports constitutes the sub-class of maritime IRPs which is briefly reviewed in
subsection 3.4.2. No approach is known addressing IRPs for rail transports.7 This is
probably because in the rail context transports are not performed by a single transport
carrier (such as a ship). Rail transports require a joint motion of locomotives and rail
cars. The routing of both types of equipment would complicate the problem considerably.
Therefore, in subsection 3.3.2 a hybrid CNDP-IRP model for rail transports in chemical
industry is formulated as a multi-layer multi-commodity network design model. Here, the
transport traction, offered by locomotives, is modelled as a binary decision incurring a
fixed charge whereby the transport capacity, offered by rail cars, is modelled by rail car
flows routed through the network.
In contrast to all other means of transport, pipeline transports do not require mobile
transport devices. Therefore, the planning of pipeline transports is not concerned with
the matching of transport flows and transport carrier flows. Nonetheless, other problems
come into scope such as the planning of pumping sequences when the pipeline is used by
multiple products.
3
See Statistisches Bundesamt (2012).
4
Further 20% are transported by ships, see Association of Petrochemicals Producers in Europe (2004).
5
For a recent and encompassing overview on IRPs see Andersson et al. (2010).
6
These capacities are typically associated with fixed charges, see e.g. Gendron et al. (1999) or SteadieSeifi
et al. (2014).
7
See Andersson et al. (2010).
54
The last fact induces that investment and management of pipeline systems are com-
pletely organized by the participating companies and (almost) all costs are internalized.
Investments in pipeline infrastructure make sense under specific circumstances only:
These requirements are given e.g. for the transport of raw and intermediate chemicals
within a chemical production site. For trans-regional transports, most prominent exam-
ples for pipeline transports are crude oil and natural gas pipelines. However, also some
basic chemicals such as Ethylene or Ammonia are transported via pipelines over long
distances to connect partners in a chemical supply chain.9
and corrosion. Using the pipeline for other chemicals is usually not possible or associated
with high preparation and/or repair costs.
To categorize pipeline systems from an organizational point of view, two basic criteria
can be used: The number of products to be transported and the number of access points.
Commonly, pipelines are designed for one product only, e.g. in the case of crude oil trans-
ports. This allows optimizing the pipeline’s technological configuration with respect to
the energy consumption. However, in chemical industry also multi-product pipelines are
met when the intended chemicals to be transported are chemically similar.11
The structure of the pipeline network is another aspect to be considered. Since in
pipelines the product flow is unidirectional in very most cases,12 the network structure
can be categorized according to the number of sources and sinks. In the simplest case,
there is a single source and a single sink (one-to-one network). Similarly, if there are
multiple sinks to be supplied this is labelled as a one-to-many network. Conversely, there
exist many-to-one networks and, finally, for multiple sources and sinks also many-to-many
networks. Table 3.1 displays this categorization scheme.
system structure
many-to-
one-to-one one-to-many many-to-one
many
single
flow split junction flow
batch batch
multiple batch flow batch split
junction network flow
Of special relevance among these types are the one-to-many and many-to-many net-
works as these show the most challenging planning problems. In the single product case,
the main problem is to keep the flow in balance along the pipeline or pipeline segments
such that inflow and outflow are equal at any time.
A more complicated problem is met in the multi-product case where batches of products
have to be routed through the pipeline (network) whereby a batch may also be split into
sub-batches at pipeline forks. In the cases of multiple sources, additionally a coordinated
pumping planning is required under consideration of the pumping capacity, the due dates
of the batches, and the segment-wise flow balances.
Pipelines are typically operated with a unidirectional material flow. In typical set-
tings the associated partners along the pipeline can be categorized distinctly in (pure)
11
A necessary condition is that chemicals to be transported do not react with each other. See e.g. Asso-
ciation of Petrochemicals Producers in Europe (2004).
12
However, there exist exceptions as described in Magatão et al. (2004).
56
providers/sources and (pure) consumers/sinks. Here, the providers are located at the
head of the pipeline and the consumers are located at the bottom. This is called a seri-
ally operated pipeline.
In contrast, if the partners along the pipeline can act as providers and consumers, each
partner has to be able to receive from and feed in the pipeline. These situations are met
in local pipeline systems, e.g. at large-scaled integrated chemical production sites, where
multiple plants and storage facilities are connected to a single-product pipeline. Here,
the storage facilities buffer the total pipeline flow and serve as receivers and suppliers of
products at the same time. This requirement is met by building orbital or cyclic pipelines
connecting all partners. Still the material flow is unidirectional, but all partners can take
or feed material from/to the pipeline as long as the (segment-wise) total flow is in balance.
There might be additional chemical or technical restrictions that have to be taken into
account for pipeline transport planning such as the interruptibility of pipeline operation.
E.g. if a product remains too long in a pipeline without movement, chemical reactions
may take place which may cause corrosion or polymerization processes. This would lead
to a failure and cause expensive maintenance and/or cleaning operations. However, the
most obvious technical restriction is probably the pump rate of a pipeline. The pump rate
measures the transport speed on a volume-per-time basis. Typically, there is an upper
bound for the tolerable pressure inside a pipeline determining the maximum pump rate
(say ρcap ). The realized pump rate might be smaller and depends on the power of pumps
at the pumping stations.13
13
Typically the pumping power can be controlled. In a reasonably planned pipeline system typically the
pumps’ maximal power suffices to exploit the maximum pump rate ρcap .
14
A time indexing of flow rates is dropped for the sake of simplicity. In principle, flow rates are time-
continuous variables, i.e. ρi (t) is the precise notation of flow rate functions.
57
i=2
i=1 ρ2
ρ1
j=3 j=4 ρ4
ρ3
15
See e.g. Matko et al. (2000) for an overview or Herrán-González et al. (2009) for the special case of gas
pipelines.
16
See e.g. Blaẑiĉ et al. (2004).
17
E.g. see Cameron et al. (2001).
18
See e.g. Association of Petrochemicals Producers in Europe (2004) for empirical figures.
58
assured additionally that sufficient storage capacity is available to maintain a minimal flow
rate at any time. As the producing plants along the pipeline are not working constantly
at the same level, buffers have to be integrated in the pipeline system to assure material
availability. A crucial question arising is the distribution of stocks along the pipeline. A
typical setting consists of tanks/inventories at each access point.
In contrast to common assumptions in inventory management, the demands of chem-
icals at chemical production sites are composed by the input and output flows of the
related plants instead of a total number of orders within a period. This implies that
the uncertainty covered by safety stocks arises rather from technological sources than
from external markets. Moreover, make-to-stock production is predominating in chemical
industry because almost all products are commodities.
In the case of continuously operated single-purpose plants the set of input and output
chemicals is fixed (i.e. there is only one recipe used). However, the throughput of the plant
can vary in most cases. I.e. there is a finite set of production modes each one associated
to specific flow levels of all chemicals. As such plants are typically designed for maximal
efficiency, the production mode associated with the maximum throughput is the intended
state of production. A change to another mode is often forced by disturbances such as
technical failures.19
To sum up, chemical production plants are characterized by a finite number of produc-
tion modes (say ∣S∣ = k) corresponding to a finite number of flow levels ω(s) with s ∈ S
for the associated chemicals.20 Each production mode s lasts for some time (reflecting
e.g. repair times).
The unintended change of a plant’s production mode can be interpreted as a stochastic
process. Two components are necessary to model such a process: If a plant enters a
certain production mode, the sojourn time (the time until the mode changes again) must
be expressed. When the production mode changes, a transition model determines which
mode is entered next. Since a finite number of production modes exists, the transition
from one mode to another is a discrete process with finite state space. The sojourn time
of a production mode is a continuous variable in principle. However, the analysis and
modelling of sojourn times as continuous variables requires a lot of historical records to fit
an accurately parametrized sojourn time distribution for each mode.21 In most cases, this
empirical problem can be circumvent by considering a discrete time scale. A stochastic
model in discrete time assumes that a plant is in exactly one mode/state per period of
time. The sojourn time of a mode can then be reflected as the length of a sequence of
periods without a mode change. Thus, a joint sojourn time and transition model can be
19
The shut-down of a plant can also be seen as a specific production mode with zero throughput. Beside
unintended changes of the optimal production mode, sometimes planned shut-downs occur e.g. for
regular technical inspections or due to a severe drop in demand.
20
Note that the number of flow levels of a certain chemical is smaller than the number of production
modes if multiple modes have the same flow level of the specific chemical.
21
This is because some modes last only for a few days and rarely occur, e.g. plant break downs.
59
expressed by transition probabilities ruling the transition from one production mode to
another at the beginning of each time period. This constitutes a discrete Markov process
or a Markov chain.22
Assume that there are probabilities that a plant’s production mode changes in a period
from mode s to mode t denoted by qst ≥ 0. Transitions from mode s to s are allowed:
qss ≥ 0 reflecting the probability that the plant remains in mode s in the next period
(sojourn time). Assume that transition matrices Q exist for all plants which are squared
matrices composed by the mode transition probabilities qst . Such a matrix is called
a stochastic matrix if the sum of each row or each column equals 1, respectively. By
convention, it is assumed that the rows sum up to 1, i.e. ∑kt=1 qst = 1, s = 1, ..., k. A
stochastic matrix constitutes a discrete time Markov process/chain.23 Let u0 be a binary
(row) vector of length k which has entry 1 at some starting position s and zero entries
otherwise, i.e. u0 represents the starting state of the Markov process in period 0.24 Then
the probability vector for period 1 is calculated by u1 = u0 ⋅ Q whose entries represent the
probabilities that the process enters the corresponding state. If Q constitutes a regular
Markov chain, a steady state vector π = {π1 , ..., πs , ..., πk } exists where πs represents the
long-term probability that the modelled process is in mode s.25 The steady state vector
π can be obtained by Π = lim Qn where each row of Π equals the steady state vector
n→∞
π.26
Another property of the steady state vector is its temporal invariance: If the Markov
process starts with π as the initial distribution (u0 = π) the starting distribution remains
unchanged for all following periods: u1 = πQ = πQn = un = π for n > 0.27
Let X jm denote the flow rate of the considered chemical of plant m ∈ {1, ..., M j } at
location j. Each plant has a finite number of production states/modes S jm with ∣S jm ∣ =
k jm . Each state is associated to a flow rate of the considered material that is consumed
or produced per period denoted by ω(s) with s ∈ S jm or, more compact, ωsjm ∈ Ojm . The
corresponding transition matrix Qjm is of dimension k jm ×k jm . The transition probability
jm
from state s to state t is denoted by qst for plant m at location j.
j
Let X be the total flow rate at location j in a period of time (i.e., X j = ∑M
j
m=1 X
jm ).
The set of states of X , denoted by O , depends on the combination of the states of all
j j
j
location j. Its cardinality is given by ∣S ∣ = k = ∏M
j j
m=1 k
jm . Hence, an element sj ∈ S j is
a vector of states for all plants at location j, sj = (sj1 , ..., sjM ) with sjm ∈ S jm . Under
j
22
Since the term "chain" refers to finite state space, the term "Markov chain" is probably more precise,
see e.g. Ibe (2008, ch. 2).
23
See Grinstead and Snell (1997, pp. 405-407).
24
To ease calculus all vectors are row vectors in the following.
25
The term regular Markov chain, simply spoken, refers to the property that all possible states can be
reached in a finite number of subsequent periods independent from the starting distribution. For
more information about definition, conditions, and properties of regular and ergodic Markov chains
see e.g. Grinstead and Snell (1997, pp. 433 ff.).
26
See e.g. Grinstead and Snell (1997, pp. 435 ff.)
27
More formally, π is the eigenvector of Q with eigenvalue 1.
60
j
independence of the production plants and their states,28 the transition probability qst
from state combination s to state combination t is given by
Mj
j
qst = ∏ qsjm
m tm . (3.1)
m=1
Mj
ωsjj = ω(sj ) = ∑ ω(sjm )
m=1
with ω j ∈ Oj . Note that there might exist state combinations with identical total
material balance, i.e. ∣Oj ∣ ≤ ∣S j ∣.30 Without loss of generality define that if X j > 0 there is
a (net) demand, whereas if X j < 0 the considered material is surplus at location j. Often,
for a given location j, X j is either exclusively positive (Oj ⊂ R+ ) or exclusively negative
(Oj ⊂ R− ).
The total material balance over all locations associated to the considered pipeline is the
sum of all X j : X = ∑j∈N X j . Again, X depends on the combination of the locations’ total
production states. Let S denote the set of the locations’ state combinations and O the
set of associated total production balances of X. To calculate the transition probabilities
qst from a total state combination s ∈ S to a total state combination t ∈ S the same
procedure as mentioned before is applied assuming independence across locations. As
different locations typically do not share a direct technical relation (in contrast to the
plants at a location), it can be assumed that independence is not a crucial assumption here.
Therefore, s, t ∈ S are vectors of ∑N j
j=1 M elements assigning each plant at each location
a state, i.e. s = (s , ..., s ) with s ∈ S and N = ∣N ∣. Consequently, Q is constructed
1 N j j
28
I.e. that a mode change at a plant does not depend on the states of all other plants. In other words,
changes of a plant’s production mode are not influenced by the other plants’ modes.
29
This follows immediately for regular Markov chains because the corresponding transition matrices
must contain positive entries only. For more information about the summation of Markov chains see
Rozhkov (2010).
30
Equality holds if all state combinations result in unique total material balances.
61
b
ph = πs1 ⋅ ∏ qsl−1 sl . (3.2)
l=2
b
yh = ∑ ω(sl ) (3.3)
l=1
where the total sum over all production quantities at all locations and plants is ω(sl ) =
31
Otherwise, the supply system is dis-balanced in expectation which would be an indication for a badly
managed system or not properly modelled suppliers/consumers.
32
See e.g. Tempelmeier (2005, pp. 397 ff.).
33
This again implies independence. This can be seen as a special case of a hidden Markov model with
certain observations, see e.g. Ephraim and Merhav (2002). Although the number of sequences grows
exponentially, calculating the corresponding probabilities can be reduced considerably (O(k 2 b)) by
iteratively calculating the probabilities of sub-sequences (dynamic programming). In the context of
hidden Markov models this is called forward-backward algorithm, see e.g. Yu and Kobayashi (2003).
62
N j
∑j=1 ωsj .
l
Let yh(g) denote the total material balances in ascending order (yh(1) = min yh ) and ph(g)
h∈H
the associated probability vector. For a given total safety stock level along the pipeline r,
−1
the α-service level of the pipeline storage system is defined by ∑fg=1 ph(g) with f determined
such that r > yh(f ) and r ≤ yh(f +1) . To define a safety stock level satisfying a desired
34
α-service level (say rα ), the quantile y α with 0 < α ≤ 1 needs to be defined by y α = yh(f )
−1
with f determined such that ∑fg=1 ph(g) ≥ α and ∑fg=1 ph(g) < α. It follows that setting the
safety stock to
leads to stock-out situations at most with probability 1 − α (within the critical time span
b).
Moreover, the safety stock level can be determined based on the β-service level. I.e. at
least β ⋅100 percent of the average net deficit occurring along the pipeline within b periods
b
can be provided from stock. Here, the average net deficit is defined by μC = ∑kg=f ph(g) ⋅ yh(g)
(with f determined such that yh(f ) > 0 and yh(f −1) ≤ 0). For the expected loss in dependence
b
of a certain stock level r follows V (r) = ∑kg=f (yh(g) − r) ⋅ ph(g) (with f determined such that
yh(f ) > r and yh(f −1) ≤ r). Then, 1 − Vμ(r)
C defines the β-service level of a stock level r for the
pipeline storage system.35 The total safety stock level satisfying a predefined β-service
level can be determined by
rβ ≥ V −1 ((1 − β) ⋅ μC ) . (3.5)
The setting of a global safety stock level based on the lower bounds introduced above
implies that
transport capacity. The former risk is neglected in the deductions since pipelines are
typically highly reliable transporters. In the latter case, the described procedure can be
used to determine the total safety stock to be held at the consumer locations.
A further assumption is that the maximum flow rate ρcap is not exceeded at a single
location, i.e. X j ≤ ρcap
j ≤ ρcap at all locations j. Otherwise, a local safety stock level has
to be incorporated covering situations when the local demand X j exceeds the maximum
flow rate ρcapj . To calculate the local safety stock level, the procedure described above can
be adapted. Consider the transition matrix Qj and recall the set of state combinations
S j as well as the corresponding set of local material balances Oj ∋ ωsj = ω(sj ) with
sj ∈ S j . To incorporate the local flow restriction, define a critical material balance by
j ). Subsequently, define the local critical demand surplus yh
ω C (sj ) = max (0, ω(sj ) − ρcap jC
over b periods for a state combination h = (s1 , ..., sb ) by yh = ∑i=1 ω (si ). The probability
j j jC b C j
of a state sequence remains as pjh = πsjj ⋅ ∏bl=2 qsjj sj . The required local safety stock rj can
1 l−1 l
∣Hj ∣
j = ∑g=f πh(g) ⋅ yh(g)
j jC
be determined e.g. by assuming that the expected demand surplus is μC
(with f determined such that yhjC(f ) > 0 and yhjC(f −1) ≤ 0).36 The local α- or β-service level
constraints can be deduced by means of (3.4) and (3.5) accordingly. Note that since
the calculation for the total safety stock along the pipeline does not differentiate states
(w.r.t. the local pipeline capacity), local safety stock levels serve as a part of the global
safety stock as well and are implicitly included in the total safety stock level calculation.
Conversely, the pipeline capacity at location j also restricts the feed rate. Hence, the
distribution of safety stock shares among the provider locations has to take into account
these constraints as well. In general, the number of covered periods should be as equal as
possible for all local safety stocks. I.e. the share of the total safety stock r for a particular
provider location j can be assigned by requiring an identical number of periods covered,
ρcap
i.e. rj = r ⋅ ∑ min(ρj cap ,ρcap ) for all j ∈ N in .
j j
The following example exercises the aforementioned procedure to derive safety stock
levels for a serially operated pipeline system with two provider locations and two consumer
locations
36
Here, yhjC(g) denotes the surplus demand in ascending order and πhj (g) its corresponding probability.
64
Example 4 (Serial pipeline supply network). Suppose that the pipeline system depicted
in Figure 3.3 connects four locations consuming or providing Ethylene. The pipeline is
serially operated with two providers of Ethylene at its head and two consumer plants at its
bottom. Provider P 1 is an Ethylene producing plant (e.g. a steam cracker) with a capacity
of 3,000 tons per day. Provider P 2 is a seaport where Ethylene can be unloaded from
tanker ships into tanks feeding the pipeline. Feeding the pipeline is operated in batch mode.
I.e. when the Ethylene stock level is sufficiently high, Ethylene is fed into the pipeline with
a constant rate of 1,000 tons per day. Otherwise, there is no feeding. The consumer plants
are supposed to be continuously operated (e.g. to produce Ethylbenzene) with consumption
rates of 2,500 tons and 1,500 tons per day, respectively. All production plants are restricted
to two production modes, either full capacity working (s = 2) or a breakdown (s = 1) with
no production/consumption. Tables 3.2a- 3.2d show the transition matrices of providers
(QP 1 and QP 2 ) and consumers (QC1 and QC2 ) as well as the corresponding production
modes and associated Ethylene quantities given in 100 tons.
s 1 2 s 1 2 s 1 2 s 1 2
ωsP 1 0 -30 ωsP 2 0 -10 ωsC1 0 25 ωsC2 0 15
0 0.60 0.40 0 0.85 0.15 0 0.60 0.40 0 0.90 0.10
-30 0.20 0.80 -10 0.05 0.95 25 0.10 0.90 15 0.10 0.90
(a) QP 1 (b) QP 2 (c) QC1 (d) QC2
Table 3.2: Transition matrices and production modes for providers and consumers
1 2 1 3 1 4 1 1
πP 1 = ( , ) , πP 2 = ( , ) , π C1 = ( , ) , π C2 = ( , ) . (3.6)
3 3 4 4 5 5 2 2
Altogether, there are 24 = 16 state combinations. The state space of the total Ethylene
balance X is O =(-40, -30, -25, -15, -10, -5, 0, 5, 10, 15, 25, 30, 40) with ∣O∣ = 13. The
system is in balance (ωs = 0) if all locations are working (s = (2, 2, 2, 2)) or if all locations
are off (s = (1, 1, 1, 1)). Tables 3.3a and 3.3b show the state combinations constituting
all unbalanced total states.
For the total material balance X the corresponding transition matrix Q is calculated
using (3.1). For instance, q99 (i.e. that there is no transition from state s9 with ωs9 = 0)
is calculated as q99 = ∏4j=1 qljj lj = q22 q22 q22 q22 = 0.8 ⋅ 0.95 ⋅ 0.9 ⋅ 0.9 = 0.62. Analogously, the
P 1 P 2 C1 C2
t 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
ωst -40 -30 -25 -15 -15 -10 -5 0 0 5 10 15 15 25 30 40
-40 0.41 0.02 0.05 0.27 0.00 0.10 0.01 0.01 0.03 0.01 0.00 0.07 0.00 0.00 0.01 0.00
-30 0.06 0.37 0.01 0.04 0.04 0.02 0.24 0.09 0.00 0.00 0.03 0.01 0.01 0.06 0.00 0.01
-25 0.05 0.00 0.41 0.03 0.02 0.01 0.00 0.00 0.27 0.10 0.01 0.01 0.01 0.00 0.07 0.00
-15 0.07 0.00 0.01 0.62 0.00 0.02 0.03 0.00 0.07 0.00 0.00 0.15 0.00 0.01 0.02 0.00
-15 0.01 0.04 0.06 0.00 0.37 0.00 0.03 0.01 0.04 0.02 0.24 0.00 0.09 0.01 0.01 0.06
-10 0.21 0.01 0.02 0.14 0.00 0.31 0.01 0.02 0.02 0.03 0.00 0.21 0.00 0.01 0.02 0.00
-5 0.01 0.06 0.00 0.10 0.01 0.00 0.55 0.02 0.01 0.00 0.06 0.02 0.00 0.14 0.00 0.02
0 0.03 0.18 0.00 0.02 0.02 0.05 0.12 0.28 0.00 0.01 0.01 0.03 0.03 0.18 0.00 0.02
0 0.01 0.00 0.07 0.07 0.00 0.00 0.00 0.00 0.62 0.02 0.03 0.02 0.00 0.00 0.15 0.01
5 0.02 0.00 0.21 0.02 0.01 0.03 0.00 0.00 0.14 0.31 0.01 0.02 0.02 0.00 0.21 0.01
10 0.00 0.01 0.01 0.01 0.06 0.00 0.06 0.00 0.10 0.00 0.55 0.00 0.02 0.02 0.02 0.14
15 0.03 0.00 0.00 0.31 0.00 0.05 0.02 0.00 0.03 0.01 0.00 0.46 0.00 0.02 0.05 0.00
15 0.00 0.02 0.03 0.00 0.18 0.01 0.01 0.03 0.02 0.05 0.12 0.00 0.28 0.02 0.03 0.18
25 0.01 0.03 0.00 0.05 0.00 0.01 0.28 0.05 0.01 0.00 0.03 0.07 0.01 0.41 0.01 0.05
30 0.00 0.00 0.03 0.03 0.00 0.01 0.00 0.00 0.31 0.05 0.02 0.05 0.00 0.00 0.46 0.02
40 0.00 0.00 0.01 0.01 0.03 0.00 0.03 0.01 0.05 0.01 0.28 0.01 0.05 0.05 0.07 0.41
tonomous is set to b = 4 days. Hence, all ∣S∣ = 65, 536 sequences of state combinations are
evaluated. The corresponding probabilities and total material balances are calculated using
(3.2) and (3.3). Figure 3.4a shows the states of the total material balances Y (4) and the
associated cumulated probabilities. The corresponding loss function V (r) is depicted as a
function of the safety stock level r in Figure 3.4b.
To assure material availability with a probability of at least α = 0.95, a safety stock level
of at least rα = 95 ⋅ 100 tons is required. This safety stock level also ensures a β-service
level of at least β = 0.95 where μC = 22.45 (as defined above). Both values are depicted in
figures Figure 3.4a and Figure 3.4b, respectively.
66
1.0
●
95% ● ●
● ● ● ●
● ● ● ● ● ● ●
● ● ●
●
● ●
0.8
cumulated probabilities
● ●
●
●
●
●
●
●
●
0.6
●
●
●
●
●
0.4
●
●
●
●
● ●
●
0.2
●
●
●
● ●
●
●
●
● ● ●
● ●
0.0
● ● ●
● ● ● ● ● ●
●
15
expected loss V(r)
●
10
●
5
●
●
●
μC ⋅ 0.05 = 1.122
●
●
●
●
●
● ● ● ● ● ● ●
0
● ● ● ● ●
0 50 100 150
V−1 ⋅ (μC ⋅ 0.05) = 95 safety stock level r
(b) Loss function V (r)
Figure 3.4: Cumulative distribution function and loss function for Y (4)
67
• separation PIGs or
• interfaces.
For the first option, special PIGs are used. Separation PIGs are technical devices acting
as an impenetrable membrane. This separates subsequent chemicals perfectly at a point
of destination. A disadvantage is the effort to plug-in the PIGs when a transition occurs.
Beside the investment costs for PIGs, they have to be re-distributed to the feed points
and must be constantly maintained. This is economically reasonable only under spe-
cific conditions, e.g. when the transported materials are of high value and comparatively
heterogeneous such that a mixing of materials induces high re-processing costs.
The second option is simply to allow a mixture of successively injected materials. This
transition mixture is often labelled as interface.38 This option is preferable as in many
cases the set of materials to be transported is chemically homogeneous (e.g. crude oil
derivatives). However, the interface is still a problem as it typically does not meet the
chemical specifications of one of the parental materials. There are two ways to deal with
the interface: Either the interface is added to one of both parental materials where it
does not harm in the further production processes39 or the interface has to be extracted
(e.g. at the end of a serial pipeline) for a special treatment.40 No matter how interfaces are
treated, transition efforts have to be faced. Hence, minimizing the number of interfaces
is one objective to aim at when planning multi-product pipeline schedules.
If transition efforts can be expressed in terms of costs, the total costs for pipeline
transport can be used as the planning objective. In this context, total costs comprise
transition costs, transport costs, and inventory costs. Transport costs account for oper-
ational pipeline costs, such as the energy costs for pumping. These costs are affected by
the pipeline’s pump rate which needs to be explicitly controlled and planned when the
transported materials differ in viscosity or other physical properties.41 Inventory costs are
37
E.g. see Rejowski and Pinto (2008) and references therein.
38
See e.g. Hull (2005).
39
Especially as its concentration declines to 0 when getting mixed with large amounts of the parental
material.
40
E.g. see Hull (2005).
41
See Rejowski and Pinto (2008) for an example. However, often exists a regular pump rate for which
the operational costs per flow unit are minimized. Deviations from this optimized pump rate usually
incur additional operational costs due to increased (relative) energy consumption. See Hane and
Ratliff (1995) or Hull (2005).
68
distinguished in stock holding costs and shortage costs. Note that in literature primarily
the first category is taken into account.42 According to the classification presented in
Table 3.1 the next subsection deals with the batch flow case, i.e. multi-product pipelines
with only two access points.
In the batch flow system, the pipeline connects two locations where a set of products
is produced and consumed. The subset of products that are used at both locations is
denoted by S. These products have to be balanced between both locations by pipeline
transports.
In the literature on batch flow planning, primarily planning problems from the oil in-
dustry are prevalent. Here, pipeline systems are more extensively used than in all other
industries where pipelines mainly distribute chemicals from a refinery to customer mar-
kets. The problem is to meet the customer demands in time and replenish local inventories.
These planning problems are typical operational planning problems where the production
quantities and customer demands are given for a certain time horizon. At both locations
tanks are used to buffer demand and supply over time. The connecting pipeline is used
to transport the materials to the consuming location and/or to the producing location.
The aim is to find a pumping schedule that meets the demand at the consuming loca-
tion by minimizing the operating costs. Technical restrictions to be considered encompass
e.g. settling periods, tank pumping restrictions or interface expenditure approximations.43
In the literature, basically two directions can be distinguished. A concise comparison of
the problem features tackled in both directions is contained in Table 3.5.
Table 3.5: Comparison of problem features for Magatão et al. (2004) and Relvas et al.
(2006)
42
This is because in most studied planning problems stock outs are not allowed at all and, hence, shortages
are explicitly prohibited. See MirHassani (2008) or Cafaro and Cerdá (2008) for examples where both
costs are considered. Note that in Cafaro and Cerdá (2008) shortages are defined as tardily delivered
batches.
43
E.g. see Relvas et al. (2006) or Rejowski and Pinto (2008).
69
Both directions commonly assume that the pipeline remains completely filled over the
total time horizon. The pipeline is fed by only one tank at a time and can only feed one
tank at a time. Conversely, a tank can either be filled or emptied at a time. It is assumed
that for each product a set of tanks is available and it has to be determined when to use
which tank for pumping.
Magatão et al. (2004) assume a one-to-one pipeline system with reversal flow option.44
The objective is to minimize interfaces and tank change-overs. The problem is modelled
as a large-scaled, time-discrete mixed-integer program (MILP). Magatão et al. (2005)
provide a solution approach relying on the decomposition of the monolithic MILP into
two sub-MILPs and a main MILP.
In contrast, Relvas et al. (2006) assume a unidirectional pipeline flow system. The
corresponding problem is formulated as a time-continuous MILP. Time-discrete customer
demands are dynamically assigned to certain batches depending on the batches’ arrival
times. The objective is to maximize the utilization of the pipeline, i.e. to maximize the
product flow from the provider to the customer depot. Originally, the pipeline’s flow
rate is fixed and the pipeline has to be operated continuously. In Relvas et al. (2007)
the original model is extended by relaxing both restrictions. To deal with the increased
complexity, Relvas et al. (2009) propose an efficient heuristic for determining desirable
product sequences.
In principle both types of operational pipeline planning problems can be applied to
chemical pipeline systems, too, as the technical specifications in both industries are sim-
ilar. However, in chemical industry the distribution of final chemicals is rarely based on
multi-product pipelines. More likely, supply processes of raw and intermediate materials
are performed by pipeline transports. This implies that chemical production sites exist
at both locations. As most chemical production sites are continuously operated, it can be
presumed that in normal operation a constant demand or surplus per period of time is
observed. I.e. there is a fixed (net) consumption and a fixed (net) production rate for each
product and site. The flow rate between both sites is then the minimum of consumption
and production rate. The remaining surplus or deficit has to be handled differently, not
affecting the pipeline operations planning at all. If the production system is in balance,
both rates are equal for each product. In any case, it can be concluded that local con-
sumption and production rates for each product are assumed to be equal. Hence, in the
long run a pumping sequence has to be determined minimizing the total costs consisting
of stock holding and transition costs.
Let ω denote the production and consumption rate for a certain product and ρ denote
the pump capacity at which the pipeline is operated. Furthermore, let T denote the cycle
time between two successive batches of the considered product and τ denote the time
necessary to transport a batch of a product from provider to consumer location.45 The
44
A many-to-many pipeline network with reversal flows is provided in Moura et al. (2008).
45
I.e. it is assumed that the pipeline is continuously operated at a constant pump rate for all products.
70
cset chold ⋅ ωs ⋅ Ts ωs
T C ELSP = ∑ ( s
+ s ⋅ (1 − )) (3.7)
s∈S Ts 2 ρs
12
L
8
Δ
Δω − Δ(ρ − ω)
Δ
4
stock level
● ●
0
00 ⎛ ω⎞ 6
1T ⎛ ω⎞ 12
2T Time
T 1− T 2−
⎝ ρ⎠ ⎝ ρ⎠
Pipeline
12
F
10
L
8
6
Δρ − Δρ
4
Δ Δ
stock level
2
0
● ●
0 6 12 Time
0 τ ⎛ ω⎞ 1T τ+T ⎛ ω⎞ 2T τ + 2T
T 1− T 2−
⎝ ρ⎠ ⎝ ρ⎠
Consumer
12
10
L
8
6
Δ
Δ(ρ − ω) − Δω
Δ
stock level
2
0
● ●
0 ω 6 12 Time
0 τ 1T τ + T⎛1 − ω⎞ τ + T 2T τ + T⎛2 − ω⎞ τ + 2T
τ−T
ρ ⎝ ρ⎠ ⎝ ρ⎠
The proposed problem with equal consumption and production rates can be seen as
a special case of the classic ELSP by assuming constant pumping rates for all products,
i.e. ρs = ρ and no set-up times. Additionally, in this case the cost function has to be
adjusted according to the total stock level in the entire supply system (provider, pipeline,
and consumer stock). Hence, (3.7) changes to
cset ωs
T C ELSP −BP = ∑ ( s
+ chold
s ⋅ ωs ⋅ (τ + Ts ⋅ (1 − ))) (3.8)
s∈S Ts ρ
where τ denotes the pumping time to transport the material through the pipeline and
L̄ = ωs ⋅ (τ + Ts ⋅ (1 − ωρs )) is the total stock level along the pipeline. If ∑s∈S ωs ≤ ρ, a
feasible schedule always exists.51 If sequence feasibility is neglected, the optimal cycles
Ts∗ can be derived from (3.8) with respect to Ts which results in
1
2 cset
Ts∗ =2
2
3
s
. (3.9)
chold
s ⋅ ωs ⋅ (1 − ωρs )
If the set of independent solutions results in a schedule feasible w.r.t. sequence and ca-
pacity, this is the optimal solution for (3.8) at the same time. Otherwise, heuristics for
the classic ELSP can be applied.52
If subsequent batches in the pipeline are separated by interfaces, the transition costs
may depend on the product sequence as e.g. reprocessing costs may differ for different
mixtures. Therefore, a sequence-dependent economic lot sequencing problem (sELSP)
has to be formulated to represent the case of interface separation. To reformulate (3.8),
let ctrans
st denote the transition costs for subsequent batches of products s and t.
Then, the set-up cost part per period of time is replaced by the total transition costs
occurring in a fundamental cycle divided by the fundamental cycle’s total length
1
min → T C sELSP −BP = ( ∑ ∑ ∑ (ctrans ⋅ xsti )) (3.10)
T cycle s∈S t∈S i∈I st
ωs
+ ∑ (chold
s ⋅ ωs ⋅ (τ + Ts ⋅ (1 − ))) .
s∈S ρ
hold
Note that replacing the stock holding costs part in (3.10) by ∑s∈S cs 2ωs Ts (1 − ωρs ) leads
to the classic sELSP.53
To assure sequence feasibility, a mixed-integer non-linear program (MINLP) is formu-
lated for the ELSP-BP and sELSP-BP. Table 3.6 contains the required set of parameters
and variables.
Sequence feasibility requires that the lot sizes are determined such that the reach of
51
This can be seen as an adaptation of the so-called capacity feasibility constraint of the classic ELSP,
see e.g. Chatfield (2007).
52
For example Dobson’s approach (Dobson, 1987) or Haessler’s approach (Haessler, 1979).
53
See Wagner and Davis (2002) or Dobson (1992).
73
Sets
S = 1, ..., S set of chemicals
I = 1, ..., I set of lot positions
Parameters
M large number
τ pumping time
ρ pump rate
ωs demand rate of chemical s ∈ S
ctrans
st transition costs for a transition from chemical s ∈ S to t ∈ S (e.g. due to
PIG insertion or interface handling)
chold
s holding costs for chemical s ∈ S
Decision variables
bsi binary, 1 if at position i = 0, ..., I a chemical s ∈ S is scheduled
Ts cycle time of chemical s ∈ S
tbi starting time of position i ∈ I
Variables
xsti binary, 1 if from position i−1 to position i a transition from chemical s ∈ S
to t ∈ S occurs
osij binary, 1 if chemical s ∈ S is lastly scheduled j positions before i ∈ I
rsi binary, 1 if position i ∈ I is the last position where chemical s ∈ S is
scheduled
rrs binary, 1 if chemical s is the finally scheduled chemical in the fundamental
cycle
T cycle fundamental cycle time
Table 3.6: Set of parameters and decision variables for the sELSP
a certain chemical’s lot covers the time span until the production of the next lot of this
chemical starts. Therefore, assume that for a set of chemicals S a set of lot positions I is
available, with ∣S∣ ≤ ∣I∣. If the number of lot positions equals the number of chemicals, the
inequality Ts ≥ ∑t∈S Tt ⋅ ωρt = T cycle holds for each chemical’s cycle time Ts . In this case, each
chemical is scheduled at exactly one position in the fundamental cycle. The sequence with
minimum transition costs can be determined by solving a Travelling Salesman problem
where the transition cost rates ctrans
st serve as distances between cities. However, such a
solution implies similar production rates ωs and similar cost rates for each product. If
these parameters vary, an optimal fundamental cycle can be composed by heterogeneous
individual cycle lengths Ts . As a consequence, chemicals with short cycles have to be
scheduled more than once in the fundamental cycle. Hence, either ∣I∣ > ∣S∣ or ∣I∣ ≫ ∣S∣
must hold for the number of positions. Sequence feasibility in such a schedule requires that
for each chemical the time span between two successive lots of this chemical is equal to
its cycle time Ts . Formally, the set of constraints (3.11)-(3.27) assures sequence feasibility
74
ωs
tbi+1 ≥ tbi + Ts ⋅ − M ⋅ (1 − bsi ) ∀i ∈ I, s ∈ S (3.11)
ρ
ω s
T cycle ≥ tbi + Ts ⋅ − M ⋅ (1 − bsi ) ∀i ∈ I, s ∈ S (3.12)
ρ
Ts ≥ tbi + T cycle − tbj − M ⋅ (3 − osii − bsi − rsj ) ∀i, j ∈ I, j ≥ i, s ∈ S (3.13)
Ts ≥ tbi − tbi−j − M ⋅ (3 − osij − bsi−j − bsi ) ∀i, j ∈ I, j < i, s ∈ S (3.14)
Ts ≤ tbi + T cycle − tbj + M ⋅ (3 − osii − bsi − rsj ) ∀i, j ∈ I, j ≥ i, s ∈ S (3.15)
Ts ≤ tbi − tbi−j + M ⋅ (3 − osij − bsi−j − bsi ) ∀i, j ∈ I, j < i, s ∈ S (3.16)
I
rsi ≥ bsi − ∑ bsl ∀s ∈ S, i ∈ I (3.17)
l=i+1
∑ rsi = 1 ∀s ∈ S (3.18)
i∈I
∑ bsi ≥ 1 ∀s ∈ S (3.19)
i∈I
∑ bsi ≤ 1 ∀i ∈ I (3.20)
s∈S
bs0 = 1 ∀s ∈ S (3.21)
∑ bsi ≤ ∑ bsi−1 ∀i ∈ I (3.22)
s∈S s∈S
Inequalities (3.11) assure that the starting time of a lot at position tbi+1 is not smaller
than the previous lot’s starting time tbi plus the time required to pump the quantity of the
scheduled chemical at position i. Constraints (3.12) restrict the fundamental cycle time
T cycle of the fundamental cycle to the maximum of all starting times tbi (plus correspond-
ing pumping time). The time span between two successive lots of a certain chemical s is
exactly Ts . There are two cases to be distinguished: First, when a new fundamental cycle
starts, the time span between the last position, say i, at which this chemical is scheduled
in the old fundamental cycle (rsi = 1) and the first position in the new fundamental cycle,
75
say i, at which the chemical is scheduled has to be restricted. Then, osii = 1 and bsi = 1
indicate that chemical s is scheduled at position i and has been lastly scheduled i periods
before, which is in period 0 or in the previous fundamental cycle. This case is modelled
by constraints (3.13) and (3.15).
Second, when a chemical is scheduled at least at two positions in a fundamental cycle,
say i and k with i > k, and nowhere in between, then this chemical is previously scheduled
j = i − k positions before position i. Hence, osi(i−k) = osij = 1 and bsi = bsk = bs(i−j) = 1.
Thus, constraints (3.14) and (3.16) ensure that tbk − tbi = Ts .
Constraints (3.17) and (3.18) define the last position at which each chemical s is sched-
uled. Constraints (3.19)-(3.20) assure that each chemical is at least scheduled once in the
fundamental cycle. No more than one chemical is scheduled at each position. Constraints
(3.21) indicate that all chemicals are scheduled in the (previous) fundamental cycle.
Constraints (3.22) avoid intermediate idle positions within the fundamental cycle. Con-
straints (3.23)-(3.26) set the transition variables xsti = 1 iff bsi−1 = 1 and bti = 1. Here,
constraints (3.25)-(3.26) define xst1 = 1 iff chemical s is the last chemical and chemical t
is the first chemical scheduled in a fundamental cycle.
Constraints (3.27)-(3.28) set osij = 1 iff chemical s is previously scheduled at position
i − j (i.e. bsi−j = 1). Finally, constraints (3.29) set rrs = 1 iff chemical s is the chemical
scheduled at last in a fundamental cycle.
The sELSP-BP constituted by (3.10)-(3.28) refers to the situation when transition costs
are caused by efforts for PIG injection to separate the batches. However, if no PIG is
used for separation, an interface is built. This interface has to be treated like one of the
two parental chemicals or is specially treated.
In the former case, some adaptations are necessary. First, let dst represent the quantity
of material which is subtracted from (dst < 0) or added to (dst > 0) a batch due to an
interface of the chemical sequence s → t. Note that this quantity typically does not depend
on the quantity of the parental batches but on the pressure and diameter of the pipeline.54
As both parameters are assumed to be constant, dst is a parameter. Moreover, it holds
that dst = −dts as the loss of one parental chemical is the surplus of the other one.55
The size of a batch received at the consumer location depends now on its predecessor
and successor. In turn, the injected batch size has to compensate the correction of the
received batch size due to the interfaces. Let qsi denote the total batch size correction at
the consumer location for the batch at position i and material s. Then, qsi = ∑t∈S dts ⋅xtsi +
∑u∈S dsu ⋅ xsu(i+1) holds. Note that due to the specific structure of dst also ∑s∈S ∑i∈I qsi = 0
holds, i.e. no quantity is lost. Figure 3.6 illustrates a situation for three subsequent batches
of chemicals t, s and u at positions i − 1 to i + 1, respectively.
54
This holds as long as the batch is larger than the critical mixing quantity. See e.g. Hall and Nicholls
(1980) or Cafaro and Cerdá (2004).
55
Note that dst = −dts is not a necessary condition as it implies equal parental shares constituting the
interface. It suffices to force sign(dst ) = −sign(dts ) where sign(⋅) denotes the sign function. However,
this generalization requires some more notation and, therefore, is not discussed here.
76
As Figure 3.6 shows, the interface between chemicals t and s is added to the batch of
chemical s (i.e. dts < 0 and dst > 0). In contrast, the interface between chemicals s and u is
added to the batch of chemical u (i.e. dus < 0 and dsu > 0). It holds bt(i−1) = bsi = bu(i+1) = 1
and it follows from (3.23) and (3.24) that xtsi = xsu(i+1) = 1. Consequently, qsi = dts −dsu < 0
holds since ∣dts ∣ < ∣dsu ∣ and the batch of chemical s at position i is reduced by qsi units.
The batch size of a chemical s is given by Ts ⋅ ωs , i.e. due to the interfaces the realized
batch size is Ts ⋅ ωs + qsi . Due to the correction of the batch sizes, the pumping times have
to adjusted accordingly. I.e. for qsi > 0 pump times can be reduced such that the pumping
time expression changes to Ts ⋅ωρs −qsi . Thus, (3.11) and (3.12) change to
Ts ⋅ ωs − qsi
tbi+1 ≥ tbi + − M ⋅ (1 − bsi ) ∀i ∈ I, s ∈ S (3.30)
ρ
Ts ⋅ ωs − qsi
T cycle ≥ tbi + − M ⋅ (1 − bsi ) ∀i ∈ I, s ∈ S (3.31)
ρ
which implies that the pumping times depend on the pumping sequence, too. Additionally,
the objective function is affected since the batch size corrections depend on the sequence.
I.e. (3.10) changes to
1
min → T C sELSP −BP −IF = ( ∑ ∑ ∑ (ctrans ⋅ xsti )) (3.32)
T cycle s∈S t∈S i∈I st
ωs
+ ∑ (chold
s ⋅ (∑ qsi + ωs ⋅ (τ + Ts ⋅ (1 − )))) .
s∈S i∈I ρ
The remaining constraints (3.13)-(3.28) remain unchanged. This constitutes the sELSP
model with global stocks and interface handling (sELSP-BP-IF).
A third option to handle interfaces is extraction and external post-processing. This
implies that all batches reduce to a certain amount due to the associated interfaces. To
reflect this specific interface handling procedure in the sELSP-BP-IF, the values of dst
have to be strictly negative such that qsi are negative, too. Note that this implies a total
loss of material due to the pumping process.
Note further that the MINLPs introduced above are hard to solve to optimality even
for very small problem sizes. Presumably, heuristics developed for the classic ELSP
77
can be adopted to the sELSP, sELSP-BP, and sELSP-BP-IF with a little effort, e.g. by
solving a TSP-like sub-problem to account for the sequence dependency.56 However, the
development of efficient heuristics is not in the scope of this work. For the relevant
problem sizes with a few chemicals discussed here, optimization is still possible within
reasonable time (i.e. within some minutes at most). To show the applicability of the
sELSP-BP and sELSP-BP-IF, the following example 5 describes a case study from basic
chemical industry.
56
Another approach which promises a good prospect for adoption is the genetic optimization approach
of Chatfield (2007).
78
Example 5 (Batch pipeline sequencing). Suppose a one-to-one serial pipeline that con-
nects two chemical sites. The provider site produces the raw and intermediate materials
for the consumer site. Suppose that the consumer site produces Styrene-based chemicals
such as Polystyrene and Styrene-Butadiene rubber. Raw materials for Styrene production
are Benzene and Ethylene. Among others, a steam cracker provides Ethylene and Pygas,
which is subsequently refined to Benzene. It is assumed that the capacities of these plants
are unbalanced such that the deficit of Pygas and Benzene has to be imported. Addition-
ally, the cracker feed, Naphtha, has to be imported.
These three chemicals are mixable, chemically related liquids and can be transported
via the same pipeline. Pygas and Naphtha are mixtures of hydrocarbons, whereby Pygas
can be interpreted as a sub-mixture primarily consisting of aromatic chemicals with main
component Benzene. At the provider site these three chemicals are surplus and, thus, are
transported via a pipeline. Ordering these chemicals to increasing purity57 leads to the
sequence Naphtha, Pygas, Benzene. Assume that two possibilities for pipeline operation
are discussed: Either the batches are separated by PIGs such that no interfaces occur or
the interfaces are "downgraded" to the less pure parental chemical.
The transition costs in case of PIG separation depend on the monetary effort for PIG
handling, e.g. redistribution of PIGs as well as depreciation of the PIG due to capital
commitment. Since Benzene is a highly reactive chemical it is assumed that two kinds of
PIGs are used. For separation of Naphtha and Pygas a PIG with lower investment costs
can be used whereas for the separation of Benzene a more robust device has to be used
which induces higher investments and, thus, higher capital commitment costs. Table 3.7
shows the transition costs ctrans
st given for this scenario.
For the downgrading option the interfaces are re-processed. E.g. an interface consisting
of Pygas and Naphtha is handled as Naphtha and stored in the Naphtha tanks where
the concentration of aromatics increases due to the Pygas inflow. The cracking process
separates these components again. This implies in turn, that downgraded ingredients
are re-processed leading to additional processing costs. I.e. an interface of Benzene and
Naphtha has to be downgraded to Naphtha and the Benzene part is re-processed twice.
Hence, sequence-dependent transition costs ctrans
st are calculated by the quantity to be re-
57
Purity refers to the fact that Naphtha and Pygas are mixtures of multiple hydrocarbons whereby
Pygas is a sub-fraction of Naphtha (obtained after cracking) and Benzene is a pure component of
Pygas (obtained after distillation).
79
processed multiplied with the processing cost rate and the number of re-processing steps.
The processing cost rate is assumed to be 100 e per ton for cracking of Naphtha and
separation of Benzene from Pygas as well. Table 3.8 shows the interface quantities dst
and associated transition costs ctrans
st for this scenario.
Typically, it can be assumed that the market value of (basic) chemicals increases with
passing more processing steps. Hence, stock holding costs increase, too. The net demand
rates provided in tons per day for Naphtha, Pygas, and Benzene under normal processing
conditions can be found in Table 3.9 accompanied by stock holding cost rates in e per ton
and day.
Table 3.9: Net demand rate ωs and holding cost rates chold
s for both scenarios
• interface downgrading.
To determine the optimal pumping schedule, for the first scenario an sELSP-BP and
for the second scenario an sELSP-BP-IF has to be solved. To compare these scenarios,
both sELSP-BP and sELSP-BP-IF are solved to optimality under the parameter setting
provided in Table 3.9 as well as Table 3.7 or Table 3.8, respectively. Up to 15 lot positions
are allowed in each model (i.e. ∣I∣ = 15).58 Both solutions differ substantially with total
58
Both MINLPs are solved on an Intel Core 2 Q6700 at 2.66 GHz with BONMIN in about 140 seconds, see
Bonami and Lee (2007).
80
costs per period of 3,270 e and 3,737 e for the sELSP-BP and sELSP-BP-IF, respectively.
Table 3.10 shows the cycle times as well as starting and pump times for both models.
Table 3.10: Resulting optimal pumping cycles for the sELSP-BP and sELSP-BP-IF
It can be taken from Table 3.10 that for the sELSP-BP a unique cycle time for all
products is the optimal solution such that each chemical is pumped every 2.066 days.
For the sELSP-BP-IF the fundamental cycle time is 2.769 days which is probably caused
by the generally higher level of transition costs forcing larger batches. The solution of
the sELSP-BP-IF yields more diverse pumping times. Here, Pygas is scheduled twice in a
fundamental cycle. This is caused by the exceedingly high transition costs between Naphtha
and Benzene which induces an intermediate batch of Pygas between Naphtha and Benzene
batches. Table 3.11 shows the pumping times of the chemicals and the idle times for each
batch position.
Table 3.11: Pumping times and idle times of the optimal schedules for the sELSP-BP and
sELSP-BP-IF
It can be observed in Table 3.11 that for the sELSP-BP-IF solution pumping times of
Naphtha and Benzene are prolonged in favour of a decrease of the Pygas batch size since
Pygas is scheduled twice. Note that the pumping times for the sELSP-BP-IF solution also
reflect the correction due to the interfaces. Figure 3.7 illustrates the fundamental cycle
and the associated interface quantities.
It can be observed that q31 = −20 and q13 = 20 which results in an increase of the
pumping time of Benzene and a decrease of the pumping time of Naphtha by approximately
700 ⋅ 24 ⋅ 60 ≈ 4 minutes which is almost negligible.
2
Comparing both scenarios merely based on the period cost, the PIG separation shows a
cost advantage. However, not all costs are included in this model since the redistribution
of the PIGs is not considered in the sELSP-BP. Moreover, investment and maintenance
of the PIGs causes additional costs. In result, the moderate advantage in period costs
59
Behind the idle time the index of the chemical scheduled at position i is provided separated by a slash.
Indices of chemicals are: Naphtha...1; Pygas...2; Benzene...3.
81
compared to the interface downgrading option may vanish if an encompassing cost analysis
is conducted.
As the literature on pipeline scheduling has its roots in the petrochemical industry, a
straight-forward extension from batch flow pipeline systems to batch split systems is
natural. Here, typically a single source has to serve a set of distinct sinks with a set
of (petro-)chemicals. Typically, the source represents a refinery and the sinks represent
local fuel depots from which the local customer demand is served. Conversely, a harbour
may also serve as a supplier of different types of raw materials for a set of refineries.60
These sinks are connected to the source via a serial pipeline system. At the source and
the sinks a set of storage tanks is available for each product. Common assumptions for
the proposed models in this branch of literature are summarized as follows:
• General characteristics
– finite time horizon
– interface re-processing
– known and deterministic demands for each product at each depot
– known and deterministic production rates for each product at the source
• Pipeline system
– unidirectional flow
– pump rate range
– initial pipeline filling
– constant pressure
• Tank system
– tank level ranges
– initial tank levels.
60
See e.g. Más and Pinto (2003).
82
In the following, the literature focusing on batch split pipeline scheduling is reviewed.
Integrative approaches containing pipeline transports in a supply chain context are out
of scope in this paragraph.61 Such approaches provide crude pipeline models to reflect
the transports in production networks over time but simplify technical details of pipeline
operation (such as the precise material tracking in time or interface handling) for the sake
of solvability of the entire model.62
Regarding the objectives, all approaches try to minimize the total costs of pipeline
operations over the planning horizon. Four cost categories are typically considered:
• Pumping costs (in terms of energy consumption multiplied with energy cost rate)
• shortage costs
The proposed approaches differ in the considered cost categories and in the type of
formulation which is either a linear or a non-linear model. Inventory costs depend on the
duration and level of stock holding. Hence, if stock levels and the time periods are decision
variables, inventory cost calculation is non-linear as both variables have to be multiplied.
The pumping costs primarily depend on the energy consumption for pumping. Typically,
pipelines are assumed to be operated at constant pressure denoted by the constant pump
rate. In some cases, however, an adjustment of the pump rate is technologically possible
and economically reasonable, e.g. to speed up the transport of the chemicals. If this is the
case, the pump rate determines the duration of pumping. As the size of an injected batch
is in any case a decision variable, the time for pumping is represented by the ratio of batch
size and pump rate which determines the total energy consumption and, consequently,
the pumping costs. Hence, pumping costs are non-linear if the pump rate is a decision
variable. Given these criteria the literature classification scheme is set up as follows:
The indicator κ ⊆ (p, h, b, i) is a tuple representing the considered cost categories where p
indicates pumping costs, h indicates inventory holding costs, b indicates backorder costs,
and i indicates interface costs. Note that not all cost categories are considered in all
61
Examples for such integrative approaches can be found in Pitty et al. (2008); Neiro and Pinto (2004);
Pinto et al. (2000).
62
See Rejowski and Pinto (2008) and the remarks about the literature review therein.
83
references, i.e. ∣κ∣ ≤ 4. The tuple λ ∈ {n, l}∣κ∣ indicates whether the corresponding cost
category is linearly or non-linearly modelled.
The second classification category distinguishes between approaches with continuous
or discrete formulation for time and pipeline. In a continuous formulation, continuous
coordinates are used to describe e.g. the position of a batch or a pipeline branch. The in-
dicator ν ∈ {c, d}2 indicates whether a continuous or discrete formulation is chosen, where
the first component refers to the time aspect and the second refers to the pipeline. In
general, time-discrete formulations are more intuitive, but to the costs of larger model di-
mensions in terms of the number of equations and variables. In contrast, time-continuous
formulations are more compact but less intuitive.63
For the pipeline a discrete formulation implies that the pipeline is represented as a
sequence of discrete segments. This eases the tracking of product batches along the
pipeline and the formulation of product removals at the depots. Table 3.12 summarizes
the relevant literature according to the proposed classification scheme.
Reference κ λ ν
Rejowski and Pinto (2003) (p, h, i) (l, l, l) (d, d)
Cafaro and Cerdá (2004) (p, h, i) (l, l, l) (c, c)
MirHassani (2008) (i) (l) (d, d)
Cafaro and Cerdá (2008) (p, h, b, i) (l, l, l, l) (c, c)
Rejowski and Pinto (2008) (p, h, i) (n, n, l) (c, d)
MirHassani and Fani Jahromi (2011) (p, h, i) (l, l, l) (c, d)
The literature referred to in Table 3.12 aims at minimizing the total operational costs
for a planning period. Most works take stock holding costs, pumping costs, and costs
for reprocessing the interface into account. Two exceptions are noteworthy: MirHassani
(2008) and Cafaro and Cerdá (2008).
MirHassani (2008) proposes a comparatively simple pipeline scheduling model consid-
ering only interface reprocessing costs to determine optimal schedules. In this model the
pipeline is subdivided into equally-sized segments. Similarly, the time horizon is subdi-
vided into equal periods of time. However, in contrast to most other models, a branching
pipeline is modelled. I.e. the depots are not located at a serial main pipeline, but con-
nected via a sub-pipeline that branches from the main pipeline. Figure 3.8 depicts both
types of pipeline systems with one source (S1) and three depots (D1, D2, D3).
Albeit the problem structure for branching pipelines is generally similar to a pipeline
structure with multiple depots located along one main pipeline, it complicates the
mathematical formulation by forcing to track batches along the pipeline branches
which equals the modelling of additional pipelines. In the follow-up paper (MirHas-
63
I.e. they contain less decision variables. See e.g. Rejowski and Pinto (2008) or Maravelias and Gross-
mann (2003) for a brief or more extensive discussion, respectively.
84
D1
S1 D1 D2 D3 S1 D3
D2
sani and Fani Jahromi (2011)), the mathematical formulation is improved towards a
time-continuous model showing a better performance which is able to handle larger
problem instances.
Most of the literature on pipeline transportation considers customer demands for the
products at the depots to be served at the end of the planning horizon. An exception is
proposed in Cafaro and Cerdá (2008) where customer demands at the depots are associ-
ated with due-dates such that backorder costs are incorporated in the objective function.
Moreover, the work proposes a rolling horizon model for updating and re-scheduling pre-
viously determined schedules according to updated demand characteristics.
Beside the incorporation of backorder costs in multi-period models, the handling of
pipeline operation costs differs. In general, pipeline operation costs are driven by the
energy consumption which, in turn, depends on the quantity and speed of pumping. As
stated before, most pipeline models are adapted for a specific pump rate, i.e. models are
based on a fixed pump rate or pump rate range (see Rejowski and Pinto (2003); Cafaro
and Cerdá (2004, 2008)). However, these works incorporate time-varying energy cost
rates depending on the time of day. I.e. in specified periods (typically when total energy
consumption in a region is high) higher energy cost rates have to be faced. Hence, pipeline
operation is shifted in tendency to less expensive periods.
In contrast, in Rejowski and Pinto (2008) the speed of pumping is a decision variable
which leads to a non-linear term in the objective function since both pump rate and
batch length have to be determined. Furthermore, this work is noteworthy as it explicitly
accounts for the inventory holding costs, but at the expense of another non-linear term
in the objective function. Note that the time-continuous formulation requires that both
the period’s length and the tank level in a period are decision variables. The product of
these terms determines the holding costs. Cafaro and Cerdá (2004, 2008) approximate
the holding costs by averaging the stock levels at the depots.
Despite the complexity of the scheduling decisions, all proposed models can be solved
with standard solvers in a reasonable amount of time.64 This is because only small problem
instances are solved w.r.t. the number of products to be scheduled and the number of
64
Either e.g. CPLEX in the MILP case (see e.g. Rejowski and Pinto (2003); Cafaro and Cerdá (2008))
or e.g. CONOPT in the MINLP case (see Rejowski and Pinto (2008)).
85
To tackle more realistic planning problems of real-world pipeline systems, models for
multi-source pipeline systems have been developed recently. This extension of one-to-
many systems considerably increases complexity. Hence, first approaches provide a heuris-
tical decomposition that decomposes a problem into three blocks:66
1. Allocation decision: determine which locations are potential candidates for injecting
and receiving batches.
3. Batch scheduling: determine the exact starting and receiving dates for each batch.
this increase in complexity, the resulting extended MILP can still solve realistic problem
instances to optimality in a reasonable amount of time. Surprisingly, for some examples
computation times are decreasing compared to the one-at-a-time approach.71 More im-
portant, the resulting schedules are more compact (increasing the pipeline utilization)
and reduce total operation costs.
In general, many-to-many systems include all other pipeline system configurations as
special cases. Since computational complexity of the models proposed in Cafaro and
Cerdá (2009, 2010) is limited and the general characteristics of pipeline system are the
same as for the previously described systems, the model proposed by Cafaro and Cerdá
(2010) can be seen as a generic basic model encompassing the previously published works
on one-to-many and many-to-many pipeline systems.
With respect to the technical characteristics of pipeline transports in basic chemical
industry the differences to petrochemical industry are limited. However, multi-product
pipelines in basic chemical industry are typically less widespread. Here, single-product
pipelines are dominating. For local material transports at chemical production sites,
multi-product pipelines do not make sense because a constant flow of materials among
the set of local production plants has to be realized. For trans-regional transports, multi-
product pipelines might be used although most intermediate and basic chemicals are
chemically not very similar. This induces that finding a common technical standard
configuration of a multi-product pipeline is a challenging task. Incompatible chemicals
can not be transported without physical separation. Costs for separation or interface
reprocessing are often very high. Consequently, if multi-product pipelines are used in the
basic chemical industry, the number of different chemicals transported is usually smaller
than in petrochemical industry. Moreover, the number of partners along the pipeline can
be assumed to be smaller, since no depots for serving customer markets are required.
Instead, the pipelines are used for balancing supply and demand of raw and intermediate
chemicals among a set of (interdependent) chemical production sites. As a consequence,
the resulting scheduling problems are expected to be less complex than in petrochemical
industry. Available operational scheduling models, which have been originally designed
for petrochemical industry, can be adapted here. Since pipelines in chemical industry are
used to interconnect chemical production sites, the quantities to be distributed via the
pipeline system can be assumed to be rather constant. Hence, not an operational reactive
pumping schedule but a long-term pumping schedule (as provided by the sELSP-BP) is
more likely to be useful for this specific industry.
71
See Cafaro and Cerdá (2010).
87
72
At least in most developed countries (such as most countries in Europe and the USA) a denser railway
network can be assumed.
73
For transcontinental transports ships and rail transports are not competing. However, in Europe
open sea and rail transports may compete. E.g. transport relations to and from middle and western
European countries passing the Mediterranean or Black Sea can also be performed by trans-balkan
rail transports.
74
In the sense that the railway network is denser and, hence, more locations can be reached by rail
transports.
75
For instance, chemical companies operating in the US often own huge rail car fleets for domestic rail
transports, see Closs et al. (2003).
88
76
See e.g. Boysen et al. (2012) for an overview on shunting yard optimization models.
77
See Schenker (2007).
78
To get an impression about RTC types, see the list of RTC types offered by lessors such as GATX
(GATX (2009)) or VTG (VTG (2011)).
79
The unloaded rest may account for approximately 10% of the total payload depending on the specific
unloading technique, see Compressed Gas Association (1999, pp. 104-115) for details.
89
transferring RTCs from one fleet to another.80 In this case, no rinsing costs occur for
rearranging the RTC fleets but re-processing costs for the contaminated chemical have to
be faced. It has to be noted that this option is only possible among fleets of the same
RTC type (i.e. gas or fluid RTC).
For transferring the relevant chemicals to and from RTCs, special technical devices
are required. For loading and unloading activities, the RTCs have to be connected to
the local pipeline system at so-called transfer arms. This is done by shunting the RTCs
to the transfer stations. Typically, an RTC has one access valve, i.e. either unloading or
loading can take place at a time. The transfer time depends on the technical specifications
of the local pipeline system and the RTC. Once the transfer procedure is finished, the
RTC is decoupled from the local pipeline system, shunted, and transported to its next
destination. The total process time for unloading/loading of RTCs consists of
1. preparation time for shunting and connecting the RTC to the transfer station,
Steps 1) and 3) depend on the availability of manpower and the layout of the transfer
stations. Since transfer stations are often serially located along dead-end rail tracks, RTCs
can interfere with one another. Hence, the planning of turnover activities may have severe
influence on the total turnover time and on the RTCs’ total cycle time.81 After shunting
and transfer, RTCs have to be transported from some point of origin to some point of
destination.
Often, chemical companies act as shippers of chemicals and manage the RTCs while
the transport is organized by a rail operator.82 Basically, the shipper has two transport
options:
In the first case, the rail operator has to offer the locomotive(s) for hauling and has to
determine route as well as schedule of the block train on a rail network. In the second case,
the RTCs handed over have to be consolidated at central shunting yards with other rail
cars in order to form block trains. Afterwards, block trains have to be routed, scheduled,
and re-composed. Obviously, the latter option is far more complex to plan and organize.
Hence, higher shipping fees can be expected on an RTC basis in contrast to the block train
80
This is similar to interface "downgrading" in pipeline batch planning, see subsection 3.2.3.
81
On the impact of cycle times on supply chain performance see Closs et al. (2003).
82
This differs from the typical organization in road freight transport where the transport operator man-
ages the transport and owns/manages the transporters.
90
option. To save transport costs, chemical companies often prefer to organize complete
block train transports. This option is the more useful the higher and more frequent
transport volumes are.
To sum up, a lot of technical and organizational restrictions have to be considered
for planning rail transports in the chemical industry. Beside technical constraints, the
planning of transport processes is integrated in the production and replenishment planning
procedures since the consolidation of transport volumes is often organized already at the
chemical production sites. Moreover, large chemical companies often manage a wide-
spread network of production sites. Hence, RTCs can be used for a variety of different
transport tasks. To support decisions on when and how many chemicals to be shipped
from which origin to which destination, the following subsection proposes a short-term
rail transportation model.
To distribute chemicals among the sites, RTCs are required. Therefore, each chemical
s ∈ S has a fleet of RTCs distributed among the sites. Each fleet is assumed to be
composed by a homogeneous RTC type with maximum payload capacity rsCap , length rsLe ,
and tare weight rsWe . At each site, these RTCs can be consolidated to trains, can be
unloaded, and/or can be loaded.
For transports between sites, trains can be composed and dispatched.84 The set of rail
links available for transport is denoted by L ⊆ I × I. Note that each site has access to at
least one rail link. Each site can dispatch a maximum number of trains yitCap in period t
which acts as a proxy for a site’s shunting capacity. The maximum number of trains to
be chartered for a rail link (i, j) ∈ L, ȳijt
Cap
, reflects varying contingents negotiated between
the shipper and the rail operator(s). 85
Le We
Similarly, individual maximum train lengths trij and train weights trij are assumed
for each rail link (i, j) ∈ L reflecting technical train specifications such as the power and
number of locomotive(s) committed by the rail operator as well as legal restrictions e.g. for
maximum train length. Typically, either the length or the weight constraint is restrictive.
The former restriction is often restrictive if only empty RTCs are hauled whereby the
latter is restrictive if only loaded RTCs are hauled.
83
See Newman and Yano (2000) for a similar intermodal transportation planning problem.
84
Loaded and empty RTCs are shunted and composed to a train load. The train load is handed over to
the rail operator who offers the traction (i.e. the locomotive) and performs the transport.
85
This allows incorporating periods without available train dispatches, e.g. on weekends or holidays.
92
The discretized travelling time of a train operating on rail link (i, j) ∈ L is denoted by
tTrv
ij which is an integer value and reflects the number of basic periods elapsing during
shipment. Obviously, this is a notable simplification of real-world conditions and a draw-
back of all time-discrete models. In this case, however, the simplification does not severely
interfere the modelling since travelling times in trans-regional rail freight transports are
typically relatively large.86 Transport times include the pure travelling times as well as
waiting and inspection times, e.g. when borders are crossed or when cargo trains have to
wait for passenger trains to pass. Hence, even for comparatively short distances to bridge,
travelling time is counted in hours. As a rule of thumb, the basic period can be set to
the maximum over the greatest common divisor of all time parameters including train
travel times as well as e.g. turnover times. The total planning horizon is restricted since
demand and production rates have to be updated frequently. Planning horizons of 7 to
14 days seem appropriate for most problem instances in practice.
An additional advantage of a time-discrete model is that e.g. work shifts for the turnover
staff are discrete in time. Under regular conditions turnover and transfer activities do not
take place around the clock but e.g. in two shifts with 8 hours each (or one shift with 12
hours). Hence, a time-discrete formulation allows to easily incorporate such regulations.
Therefore, the unloading and loading capacities, eCap Cap
its and lits , are expressed in tons of
chemical s to be handled at site i in period t. Both parameters reflect the work force
availability and technical restrictions such as the number and capacity of transfer arms. It
is assumed that exclusive turnover capacities are provided for each chemical at a site. This
implies single-product transfer stations for each product. This is reasonable if handled
chemicals are dissimilar and a joint unloading/loading is too dangerous or expensive.87
However, in case of similar chemicals, commonly used transfer stations may exist at the
sites.88 In this case, a slightly adjusted turnover capacity formulation is required. In basic
chemical industry, however, individual transfer stations are predominant.
To properly model the RTC handling processes at the sites, the time necessary to un-
load/load the RTCs has to be considered. The total processing time for RTC handling
is summarized in parameter tTrn
is which is the number of basic periods required to make
an RTC available for transport (after loading or unloading). I.e. if RTCs are unload-
ed/loaded, tTrn
is is the minimum time span an RTC has to spend at site i. In most cases,
a turnover time span of one work shift, i.e. one basic period, is assumed before an RTC
86
A reasonable choice for the basic period is e.g. 24, 12 or 8 hours. In practice, more granular planning
accuracy is seldom required since travelling times in trans-regional rail transports are not much shorter
than 12 hours. See Newman and Yano (2000) for a more detailed discussion.
87
Dissimilar chemicals are likely to react with each other which is typically unintended as it bears the
risk of debris, hazardous reactions or even explosions.
88
Such stations are often used in crude oil industry when different types of fuel or other oil derivatives
are handled.
93
is available for transport. Note that RTCs which are not unloaded/loaded are available
for transport immediately.89
If the number of RTCs to be unloaded/loaded in a period at a site exceeds the respective
capacities eCap Cap
its and lits , the remaining number of RTCs have to wait at the site’s shunting
L
yard. Loaded RTCs, denoted by srits , may wait on rail tracks ready for forwarding.
This equals an additional inventory denoted by sLits which induces stock holding costs
for committed capital and costs for supervising the RTCs. Since most chemicals are
hazardous, storing requires permanent supervision and control. In regular tanks this is
typically assured by sensor systems. Although RTCs are designed for recurring transports
of hazardous chemicals, they typically do not meet e.g. legal requirements for a permanent
technical supervision system.
To evaluate the operational effort for realizing transports and keeping stocks, three types
of costs have to be considered:
• transport costs
• turnover costs.
Stock holding costs are calculated as follows: At each site safety stocks of each handled
chemical are hold. They protect against extraordinary demand peaks e.g. caused by plant
breakdowns. With respect to the inventory holding costs, the stock level sits should meet
the desired safety stock level sTar
is . An overshooting of the safety stock levels causes un-
necessary stock holding costs (e.g. due to capital commitment) whereas an undershooting
increases expected shortfall costs since in emergency cases plant shutdowns menace due
to a shortfall of supply. Because the costs induced by a missing or surplus ton of a certain
chemical depend on the target stock level sTar is , over- and undershooting (oits and uits )
are measured as relative deviations from sTar Os
is . The corresponding cost rates cis and cis
Us
are measured in monetary units per relative deviation of the stock level of chemical s
from its target stock level at site i. Often it holds cUsis ≫ cis because a plant shutdown
Os
90
causes much higher costs compared to stock holding. Additionally, for chemicals stored
intermediately in (loaded) RTCs, the cost rate cAdd reflects the additional supervision and
maintenance effort measured in monetary units per ton and period.
Transport costs reflect the monetary effort to be faced when trains are dispatched, i.e. a
rail operator performs a transport service. Therefore, cost rate cTrv
ij accounts for the charge
89
E.g. if empty RTCs arrive in a period at a site and are not to be loaded, they can be forwarded to
another site in the same period.
90
In general, expected shortfall costs increase non-linearly with increasing undershooting of target stock.
However, for the sake of solvability this non-linearity is dropped in this model.
94
paid when a train is dispatched on rail link (i, j) ∈ L. These charges are typically based
on general agreements with the responsible rail operator and depend on the distances to
be travelled, borders crossed, train specifications and other details.91
The turnover cost rate cTrn
is comprises direct costs for unloading and loading RTCs. This
cost rate reflects labour costs for shunting, unloading/loading, and supervision activities.
To sum up, a rail transportation model for multiple products shipped by trains between
chemical production sites can be formulated as a multi-layer, multi-commodity time-
space expanded network flow model. In other words, an operational multi-chemical rail
transportation problem is provided (in short MC-RTP).
In the network model each chemical corresponds to a main layer representing the infras-
tructure of this chemical. The main layer consists of four sub-layers which are dedicated
to
Flows between two nodes of the same sub-layer correspond to storing or shipping flows
of RTCs or chemicals, respectively. Flows between nodes of different sub-layers indicate
loading or unloading flows. At each node inflows and outflows are in balance. Each flow is
restricted by stock, RTC or train capacities. Transport flows of chemicals are restricted by
the accompanying flow of loaded RTCs, i.e. flows in the second and third sub-layer match
each other. Hence, transport capacities are dynamically modelled as in a classic IRP. In
difference to classic IRPs, no routes for individual carriers are determined. Instead, the
RTC flow balances ensure that the correct transport capacity is modelled. The available
transport capacity, offered by RTCs, can only be exploited by chartering trains. This
component assigns arc-based flow capacities incurring fixed costs which is a characteristic
of fixed-charge network design models. Hence, the proposed model can be seen as a hybrid
model bearing characteristics of both prominent model classes. Figure 3.9 illustrates an
exemplary flow network for five periods and three sites.
Grey arrows indicate (passive) flows of chemicals and (empty) RTCs. Marked by black
arrows is the following sequence of activities: In period 1 empty RTCs are shipped from
site 2 to site 1 where they are needed to load a certain quantity of the chemical in period
2. I.e. a flow between forth and third sub-layer indicates a change in the RTCs’ states.
91
Note that additional flexibility can be incorporated by time varying charges e.g. for weekend transports.
95
layer
r)
b- s
su TC
ye
1,2 2,2 3,2 4,2 5,2
la
R
ship
(4 pty
1,1 ping 2,1 3,1 4,1 5,1
em
th
loading
unloading
1,3 2,3 3,3 4,3 5,3
b- Cs
r)
ye
1,2 2,2 3,2 4,2 5,2
su RT
la
(3 ded
g
1,1 2,1 3,1 shippin 4,1 5,1
a
rd
lo
1,3 2,3 3,3 4,3 5,3
ye s
la C
b- T
r)
su in R
1,2 2,2 3,2 4,2 5,2
nd ls
g
shippin 4,1
(2 ica
1,1 2,1 3,1 5,1
em
ch
un
lo
ad
in
1,3 2,3 3,3 4,3 5,3
g
g
ye s
din
la nk
r)
b- ta
loa
su in
e
sit
st s
storing
(1 ical
1,1 2,1 3,1 4,1 5,1
em
time
Correspondingly, a quantity of the chemical is transferred from the tanks (first sub-layer)
to the mobile stock (second sub-layer). In period 3 the loaded RTCs depart by train from
site 1 carrying the chemical to site 2 where they arrive in period 4 (second and third
sub-layer). In this period the RTCs are unloaded such that the local chemical stocks are
replenished and, simultaneously, the RTCs are emptied. In this sequence two shipments
are performed, i.e. (at least) two trains are chartered.
To formalize the sketched network flow model, the remaining notation is introduced
as follows: A shipment flow of chemical s from site i to site j in period t is denoted by
xijts (second sub-layer). Shipment flows of loaded and empty RTCs are denoted by rxLijts
and rxEijts , respectively (third and forth sub-layer). Unloading flows and loading flows of
E L
chemical s in period t at site i are described by zits and zits (flows between first and second
sub-layer). These flows are accompanied by the corresponding RTC flows (between third
E L
and forth sub-layer) rzits and rzits , respectively. To enable a shipment of loaded and/or
empty RTCs from node i to node j a number of trains (yijt ) has to be dispatched.
The local chemical stock in tanks is denoted by sits (first sub-layer) whereas the mo-
bile stock in RTCs is denoted by sLits (second sub-layer). The number of loaded and
L E
empty RTCs at hand are described by srits and srits . Table 3.13 summarizes all notation
introduced above (and some more).
96
Sets
Cost rates
cOs
is over-shooting cost rate for chemical s ∈ S
cUs
is under-shooting cost rate for chemical s ∈ S
cTrn
s turnover costs for chemical s ∈ S
cTrv
ij costs for chartering a train on rail link (i, j) ∈ L
(artificial) cost rate for carrying an empty RTC of chemical s ∈ S
c̃Tr−E
ijs
on rail link (i, j) ∈ L
(artificial) cost rate for carrying a loaded RTC of chemical s ∈ S
c̃Tr−L
ijs
on rail link (i, j) ∈ L
cAdd cost rate for storing chemical in RTCs
Utilization parameters
Capacities
sCap
is inventory capacity for chemical s ∈ S at site i ∈ I
Initialization parameters
sIni
is initial stock level for chemical s ∈ S at site i ∈ I
sL−Ini
is initial stock level stored in RTCs for chemical s ∈ S at site i ∈ I
L−Ini
ris initial loaded RTCs for chemical s ∈ S available at site i ∈ I
E−Ini
ris initial empty RTCs for chemical s ∈ S available at site i ∈ I
quantity of chemical s ∈ S dispatched in period τij ∈ {1 − tTrv
ij , ..., 0}
xIni
ijτij s
on rail link (i, j) ∈ L
number of loaded RTCs dispatched in period τij ∈ {1 − tTrvij , ..., 0}
rxL−Ini
ijτij s
on rail link (i, j) ∈ L
number of empty RTCs dispatched in period τij ∈ {1 − tTrvij , ..., 0}
rxE−Ini
ijτij s
on rail link (i, j) ∈ L
L−Ini
rziτis s
number of RTCs loaded in period τis ∈ {1 − tTrn
is , ..., 0} at site i ∈ I
E−Ini
number of RTCs unloaded in period τis ∈ {1 − tTrn
is , ..., 0} at site
rziτis s
i∈I
Time parameters
tTrv
ij integer, travelling time for trains on rail link (i, j) ∈ L
tTrn
is integer, turnover time for chemical s ∈ S at site i ∈ I
Decision variables
E
number of empty RTCs for chemical s ∈ S in period t ∈ T at site
srits
i∈I
relative over-shooting in % of the target stock of chemical s ∈ S in
oits
period t ∈ T at site i ∈ I
relative under-shooting in % of target stock of chemical s ∈ S in
uits
period t ∈ T at site i ∈ I
Table 3.13: Sets, parameters, variables, and decision variables for the MC-RTP
Using the aforementioned notation, the total operational costs for balancing the chem-
ical production network by rail transports is expressed in (3.33) which constitutes the
MC-RTP’s objective function to be minimized.
T C = ∑ ∑ ∑ (oits ⋅ cOs
is + uits ⋅ cis ) +
Us
i∈I t∈T s∈S
∑ ∑ yijt ⋅ cTrv
ij +
(i,j)∈L t∈T
∑ ∑ ∑ (rzits
E
+ rzits
L
) ⋅ cTrn
s +
i∈I t∈T s∈S
The first part of (3.33) accounts for the stock holding costs arising from deviations of
stock levels from the desired target stock levels. The second part sums up the total train
charges over all rail links by multiplying the number of dispatched trains yijt with the
corresponding train cost rate cTrv
ij . Turnover costs for turnover activities taking place at
the sites are summed up in the third part of (3.33). The last part accounts for costs for
supervision and control of loaded RTCs waiting at the sites which can be interpreted as
costs for dynamic storage capacity extensions.
The considered restrictions can be categorized in balancing and capacity constraints.
Constraints (3.34)-(3.37) represent the balancing constraints.
− ∑ xijts + ∑ xji(t−tTrv
ji )s
∀i ∈ I, t ∈ T , s ∈ S (3.35)
j∈I j∈I
L
srits = sri(t−1)s
L
− rzits
E
+ rzi(t−t
L
Trn )s
is
E
srits = sri(t−1)s
E
− rzits
L
+ rzi(t−t
E
Trn )s
is
99
Here, (3.34) and (3.35) track the quantities of chemical s available at site i in period t
E L
in tanks (3.34) and in loaded RTCs (3.35). Unloading and loading flows (zits and zits )
connect both inventories inversely, e.g. unloading RTCs (i.e. zits E
> 0) reduces sLits and
increases sits . Furthermore, tank inventories buffer the net consumption or surplus ωits .
In contrast, the stock loaded in RTCs sLits absorbs incoming chemical flows from other
sites (xji(t−tTrv
ji )s
) and provides outgoing chemical flows to other sites (xijts ).
Equations (3.36) and (3.37) represent stock balances for loaded and unloaded RTCs.
E L
Both stocks are interconnected by RTC transfer flows rzits and rzits . E.g. unloading of
E
rzits RTCs in period t at site i immediately reduces the number of loaded RTCs. However,
tTrn
is periods are required for unloading and shunting before these RTCs can be handled
as empty RTCs again. Both stocks are replenished by incoming RTCs from other sites
(rxLji(t−tTrv )s and rxEji(t−tTrv )s ) and provide outgoing RTC flows to other sites (rxLijts and
ji ji
rxEijts ).
Capacity constraints can be separated into two subgroups: Static capacities which refer
to infrastructural conditions (such as tank, train, or turnover capacities) and dynamic
capacities referring to restrictions that are planned simultaneously (such as RTC stocks
and flows). Constraints (3.38)-(3.42) express static capacity constraints.
E
zits ≤ eCap
its ∀i ∈ I, t ∈ T , s ∈ S (3.38)
L
zits ≤ lits
Cap
∀i ∈ I, t ∈ T , s ∈ S (3.39)
sits ≤ sCap
is ∀i ∈ I, t ∈ T , s ∈ S (3.40)
yijt ≤ ȳijt
Cap
∀(i, j) ∈ L, t ∈ T (3.41)
∑ yijt ≤ yit ∀i ∈ I, t ∈ T
Cap
(3.42)
j∈I
Constraints (3.38) and (3.39) restrict the unloading and loading flows to the unloading and
loading capacities of the corresponding periods and sites. Similarly, (3.40) ensures that in
all periods stock levels are less or equal to the tank capacities. Constraints (3.41) require
that the maximum number of trains dispatchable on a rail link per period is not exceeded.
Similarly, constraints (3.42) ensure that the sum of all trains composed on site i in period
t is smaller than the corresponding maximum shunting capacity. Constraints (3.43) and
100
(3.44) refer to technical train specifications assuring that maximum weight and length
are not exceeded. On the right-hand side, the total available capacities are calculated as
the product of maximum train weight and length multiplied with the number of trains
dispatched on a link (i, j) ∈ L. On the left-hand side, the trains’ total hauling weight is
calculated as the sum of all chemicals loaded in RTCs (∑s∈S xijts ) plus tare weights of all
RTCs attached to the train(s). The total length of the trains is the sum of the lengths of
all RTCs attached.
If yijt > 1 for a specific trip (i, j, t) and only one of both constraints is restrictive, an
integer feasible assignment of RTCs to individual trains is assured by constraints (3.43)
and (3.44). However, if both constraints are (almost) restrictive92 and yijt > 1, an integer
feasible assignment of RTCs to trains only exists if the number of RTCs is a multiple of
the number of trains yijt . To illustrate this situation, consider a specific trip with y = 2.
For simplicity assume that a homogeneous fleet of RTCs exist with equal tare weight
rWe = 2, equal length rLe = 1, and equal capacity rCap = 8. Let trLe = 20 and trWe = 100.
Le
Then, the maximum number of RTCs due to (3.44) is trrLe ⋅y = 20 1 ⋅2 = 40 with an associated
tare weight of 40 ⋅ rWe = 80. According to (3.43) the total remaining (weight) capacity for
payload is then y⋅trWe −80 = 120 which implies a total number of r120 Cap = 15 of loaded RTCs.
In consequence, 25 empty RTCs and 15 loaded RTCs with a maximum payload of 120
could be carried by two trains according to (3.43) and (3.44). However, the numbers of
loaded and empty RTCs are not multiples of two. The most balanced integer assignment
of loaded and empty RTCs to two trains is displayed in Table 3.14.
From Table 3.14 it can be taken that for train # 1 a feasible RTC assignment is made
while for train # 2 the maximum total weight of 100 is exceeded. In case when multiple
trains can be dispatched on a trip, a reformulation of (3.43) and (3.44) would overcome
this problem. However, to reformulate both constraints additional variables are necessary
indicating whether both constraints are (almost) restrictive for a specific trip. This would
lead to a more complicated model. Since the problem only occurs if the total payload is
close to its maximum, it suffices to introduce a safety buffer for the total train weight.
We
I.e. a train’s allowed total weight trij We
is calculated as trij ˜ We
= tr ij − max rs
Cap ˜ We
where tr ij
s∈S
92 Le
To be precise, it has to hold trij ⋅ yijt − ∑s∈S rsLe ⋅ (rxE L We
ijts + rxijts ) = 0 and trij ⋅ yijt − ∑s∈S xijts +
We E L Cap
∑s∈S rs ⋅ (rxijts + rxijts ) ≤ max rs .
s∈S
101
rsCap
is the technical maximum of the train’s total weight. Note that in practice the ratio ˜ We
tr ij
is quite small (approximately 1-5 %) such that no serious loss in train utilization is to be
expected.
Dynamic capacity constraints are formulated in (3.45)-(3.49).
E
zits ≤ rzits
E
⋅ rsCap ∀i ∈ I, t ∈ T , s ∈ S (3.47)
L
zits ≤ rzits
L
⋅ rsCap ∀i ∈ I, t ∈ T , s ∈ S (3.48)
sLits ≤ srits
L
⋅ rsCap ∀i ∈ I, t ∈ T , s ∈ S (3.49)
Here, (3.45) and (3.46) ensure that the shipped quantity of a chemical s matches the
total payload capacity of all loaded RTCs to be shipped.93 Constraints (3.47) and (3.48)
restrict the loading and unloading flows to the payload capacities of the simultaneously
loaded and unloaded RTCs. Constraints (3.49) and (3.50) enforce that the quantity of
chemicals designated to be stored in RTCs matches the total payload capacity of loaded
RTCs. Constraints (3.45) and (3.46) in combination with (3.49) and (3.50) also restrict
loading and unloading flows to be matched with the numbers of loaded and unloaded
RTCs in a period.
To assess relative under- and over-shooting of target stock levels, constraints (3.51)
assign the deviations of the ratio ssTar
its
from 1 to the variables oits and uits . If sits > sTar
is ,
is
it follows that sTar − 1 > 0. Because oits and uits are non-negative variables and both are
sits
is
associated to positive cost rates, the cost-minimal values for oits and uits satisfying (3.51)
are obtained by oits = ssTar
its
− 1 and uits = 0. Otherwise, uits = ssTar
its
− 1 and oits = 0. Finally,
is is
(3.52) and (3.53) define the variables’ domains.
sits
oits − uits = −1 ∀i ∈ I, t ∈ T , s ∈ S (3.51)
sTar
is
L
xijts , zits E
, zits , sits , sLits , oits , uits ∈ R+ ∀i ∈ I, j ∈ I, t ∈ T , s ∈ S (3.52)
Constraints (3.33)-(3.53) constitute the MC-RTP. The optimal solution of the MC-
93
Note that dropping (3.46) would circumvent the distinction in empty and loaded RTCs. This would
allow sending empty RTCs declared as loaded ones to another site where these could be loaded
immediately. Hence, the turnover times would not be considered.
102
RTP minimizes the total costs for transports, turnover activities, and stock holding. The
solution provides flows of RTCs and chemicals within the considered time horizon. At
the end of the planning horizon, no transports are planned because the arrival periods
are beyond the time horizon.94 In practice, therefore, the MC-RTP has to be applied in
a rolling horizon environment with overlapping time horizons to update schedules with
new information e.g. about consumption and production rates.95 I.e. starting with an
initial MC-RTP instance and for fixed a length of the planning horizon, trains and RTCs
are scheduled in accordance to the corresponding (optimal) solution for the first peri-
ods (the re-planning interval).96 After the re-planning interval has elapsed, an updated
MC-RTP instance is set up and solved. Therefore, updated initial values for stocks and
available RTCs etc. and additional production/consumption rates for the new periods to
be planned are incorporated. Additionally, trains that have departed but not arrived yet
have to be considered. To use the MC-RTP in a rolling horizon environment, additional
initialization equations have to be added.
At first, initial stock levels in tanks sIni L−Ini
is , quantities available in RTCs sis , the corre-
L−Ini E−Ini
sponding numbers of loaded RTCs ris , and the available numbers of empty RTCs ris
for each chemical s and site i at the beginning of the planning horizon are incorporated
(see (3.54)-(3.57)).
si0s = sIni
is ∀i ∈ I, s ∈ S (3.54)
E
sri0s = ris
E−Ini
∀i ∈ I, s ∈ S (3.55)
L
sri0s = ris
L−Ini
∀i ∈ I, s ∈ S (3.56)
sLi0s = sL−Ini
is ∀i ∈ I, s ∈ S (3.57)
If RTCs and chemicals are dispatched in previous periods (and arrive during the planning
horizon), these have to be considered. This affects variables xijτij s , rxLijτij s , and rxEijτij s
with 1 − tTrv ij ≤ τij ≤ 0 which have to be set to the corresponding initial values denoted by
xIni
ijτij s , rx L−Ini E−Ini
ijτij s , and rxijτij s as shown in (3.58)-(3.60).
xijτij s = xIni
ijτij s ∀(i, j) ∈ L, s ∈ S, τij ∈ {1 − tTrv
ij , ..., 0} (3.58)
rxEijτij s = rxE−Ini
ijτij s ∀(i, j) ∈ L, s ∈ S, τij ∈ {1 − tTrv
ij , ..., 0} (3.59)
rxLijτij s = rxL−Ini
ijτij s ∀(i, j) ∈ L, s ∈ S, τij ∈ {1 − tTrv
ij , ..., 0} (3.60)
94
A similar problem occurs in pipeline operations planning when batches planned at the end of the
planning horizon are solely injected to keep the pipeline working. Such batches do not arrive at
depots during the planned time horizon. Hence, they cannot satisfy demands at depots.
95
For a similar rolling horizon approach for pipeline operations planning see Cafaro and Cerdá (2008).
96
Note that the re-planning interval should be considerably smaller than the planning horizon, see
e.g. Cafaro and Cerdá (2008).
103
Similarly, RTCs already shunting at the yards at the beginning of the planning horizon
have to be considered by setting the variables rziτ L
is s
E
and rziτis s
for 1 − tTrn
is ≤ τis ≤ 0 to
L−Ini E−Ini
their corresponding initial values rziτis s
and rz iτis s
E
rziτis s
= rziτ
E−Ini
is s
∀i ∈ I, s ∈ S, τij ∈ {1 − tTrn
is , ..., 0} (3.61)
L
rziτis s
= rziτ
L−Ini
is s
∀i ∈ I, s ∈ S, τij ∈ {1 − tTrn
is , ..., 0} (3.62)
site i
2 2 - 1 2 20 20 20 rsWe 5 5 10
3 4 1 - 3 25 20 25 rsLe 20 20 25
(a) tTrv
ij
E−Ini
(b) ris (c) rsCap , rsWe , rsLe
Table 3.15: Technical parameters for the rail operations planning example
For each chemical a fleet of RTCs is available. No previous transports are assumed
to be on track such that the initialization variables in constraints (3.58)-(3.62) are set to
zero. Similarly, all available RTCs at the sites at the beginning of the planning horizon
are empty and ready for loading. Hence, initialization variables in constraints (3.56) and
(3.57) are zero, too. Tables 3.15b and 3.15c show the distribution of the RTC fleets among
the sites and the technical configurations for the RTCs in the chemicals’ fleets, respectively.
All turnover times tTrn
is are set to one period. Furthermore, train specifications are given
We
by trij = 800 and trij
Le
= 1, 000 for all (i, j) ∈ L. The unloading and loading capacities eCap its
Cap
and lits are set to 700 tons such that a complete train’s load can be (un-)loaded within a
period.
For simplicity it is assumed that the chemicals’ consumption/production rates ωits are
constant over the planning horizon. Table 3.16 displays the assumed consumption/pro-
duction rates ωis as well as stock capacities sCap Ini
is , initial stocks sis , and target stock levels
104
sTar
is for each chemical s and site i.
site i 1 2 3
chemical s 1 2 3 1 2 3 1 2 3
ωis -50 25 100 -50 25 -200 100 -50 100
sCap
is 3000 2500 2500 1500 1500 5000 5000 5000 5000
sIni
is 400 500 700 500 300 1000 700 500 1000
sTar
is 500 250 500 500 250 1000 750 500 1000
cAdd = 10 for all i, j ∈ I and s ∈ S. Note that overshooting the target stock levels induces
only a thousandth of the cost for undershooting. The train cost rate cTrvij is comparatively
small forcing extensive rail activities.
Solving this instance of the MC-RTP leads to the optimal solution after 8 seconds with
an objective value of 20,845.97 Note that network design problems are N P − hard prob-
lems.98 However, the computational complexity depends on the relation of the parts of the
objective function and the size of the network (particularly the number of links). Here,
a comparatively small problem instance with a favourable ratio of costs99 is given which
eases solving to optimality.
The optimal transport flows and stock levels are displayed in Figure 3.10 for the first
eight periods by means of a time-space expanded network. Each cell shows the stock levels
at a site in a particular period. Dispatched trains are indicated by arrows between two
nodes where the associated transport quantities (in tons of chemicals) and/or the number
of empty RTCs are assigned to each arrow.100
In t = 1, from sites i = 1 and i = 3 repositioning transports are dispatched enabling
chemical transports in later periods. For example, empty RTCs of chemicals s = 1 and
s = 3 are dispatched in period t = 1 from site i = 3 to site i = 2 arriving there in t = 2.
In period t = 3, empty RTCs of chemical s = 1 are partially forwarded101 (4 out of 20)
to site i = 1 together with some loaded RTCs of chemical s = 3.102 Similar sequences of
re-positioning and supply transports can be found for the remaining chemicals and sites.
These sequences are repeated if the production/consumption rates do not change in time
which results in recurring re-positioning cycles.
97
Calculations are performed on a 2.6 GHz machine with IBM CPLEX 12.5, see IBM (2010).
98
See e.g. Crainic (2000).
99
To be precise, the inventory costs dominate transport costs such that network flow primarily determines
the total costs.
100
For clarity, the indices of variables annotated at the arcs are reduced and indicate the number of the
chemical only.
101
The rest is used for transports of chemical s = 1 back to site i = 3 in periods t = 5 and t = 7.
102
To be precise, 26 loaded RTCs of chemical s = 3 are dispatched in period t = 3 carrying 520 tons from
site i = 2 to site i = 1. The train’s maximum weight is totally exploited.
t =1 t =2 t =3 t =4 t =5 t =6 t =7 t =8
450 500 430 480 410 460 510 560
600 500 400 300 720 620 520 420
i =1
475 450 425 400 375 350 325 300
E E E
rx s=3 = 13 rx s=3 = 12 rx s=3 = 26
x s=1 = 120 x s=1 = 120
E
x s=3 = 520 rx s=2 = 12
E
rx s=1 =4
550 600 650 700 510 560 445 495
1200 1400 1080 1280 1160 1360 1280 1480
i =2
275 250 225 320 295 270 245 310
x s=3 = 320 x s=3 = 280
x s=1 = 240 x s=2 = 90 x s=1 = 285
E
rx s=3 = 20 x s=2 = 120 x s=2 = 90
E E E
rx s=1 = 25 rx s=3 =5 rx s=1,3 = 16
600 500 400 300 200 340 360 545
900 800 700 600 500 720 620 800
i =3
550 600 530 580 540 590 550 600
For settings with realistically sized finite planning horizons, only the beginnings of such
cycles are visible. At the end of the planning horizon no transports are reasonable as
the savings in stock holdings become effective in periods beyond the planning horizon
only. Therefore, stock levels are the less balanced the closer the planning horizon ap-
proaches. This end-of-horizon effect becomes strikingly apparent for chemical s = 1 (dark-
grey curves) at sites i = 1 and i = 3 as it is illustrated by Figure 3.11.103 The figure shows
the stock levels for all chemicals, periods, and sites including the corresponding target
stocks (as dashed lines).
It can be observed that the stock levels meet the corresponding target stocks well in most
cases. For chemical s = 2 (medium grey curves) the stock levels are close above the target
stocks at all sites and in (almost) all periods.104 Chemicals s = 3 (light grey curves) and
s = 1 are oscillating around their target stock levels at sites i = 1 and i = 2 whereas at
site i = 3 both chemicals are clearly below their target stocks. Transports of chemical s = 3
from site i = 2 to i = 3 could reduce holding costs at both sites.105 However, this option
is not realized either due to a lack of RTCs or due to economic reasons (e.g. the related
transport costs).
This example provides some first insights into the interplays of re-positioning and stock-
piling transports crucially depending on both the availability of RTCs and the heights
of target stock levels. Moreover, it becomes apparent that in practical applications the
MC-RTP has to be applied in a rolling horizon environment to overcome end-of-horizon
problems like stock decreases and undone repositioning transports.106
In reality, the production networks typically consist of more than three sites107 whereby
a time horizon of one or two weeks is reasonable in most cases. The time horizon depends
on the forecasting stability of production/consumption estimates as well as the transport
times. Hence, the model’s complexity increases with increasing numbers of nodes and
periods. In cases of very large instances, a heuristical procedure can be set up like this:
1. Build a relaxed MC-RTP by replacing the train cost part ∑(i,j)∈L ∑t∈T yijt ⋅ cTr
ij in the
objective function (3.33) with
∑ ∑ ∑ (rxLijts ⋅ c̃Tr−L
ijs + rxijts ⋅ c̃ijs )
E Tr−E
750 1000
500
stock level
250
0
location i = 2
1250
750
stock level
250
0
location i = 3
750
stock level
250
0
0 2 4 6 8 10 12 14
period
Figure 3.11: Optimal stock levels for the MC-RTP instance of example 6
107
108
A considerable part of the model’s complexity is caused by the integer decisions to deter-
mine the number of dispatched trains. Since these decision variables are associated with
fixed charges per train, a variabilization of these fixed costs simplifies the problem consider-
ably. Variable transport costs depend linearly on the number of RTCs dispatched on a rail
link. To define variable transport cost rates c̃Tr−L
ijs and c̃Tr−E
ijs , calculate the maximal num-
trWe trLe
ber of empty and loaded RTCs to be carried, i.e. nCap−L
ijs = ⌊ rWe +r
ij
Cap ⌋ and nijs
Cap−E
= ⌊ rLe
ij
⌋,
s s s
respectively. Then, the transport cost rates can be calculated as
cTr
ij
c̃Tr−E
ijs = Cap−E
(3.63)
nijs
cTr
ij
c̃Tr−L
ijs = Cap−L
. (3.64)
nijs
The proposed heuristic reduces the computation time by approximately 50% for the in-
stance described in this example.108 The heuristical solution yields total costs of 21,117.75
which is about 1.3 % above the optimal value.
108
Note that the relaxed MC-RTP was solved to optimality although this is not necessary for a heuristic.
The procedure can be fastened by terminating optimization at a threshold for computation time or
optimality gap.
109
Chemical tankers for inland waterway transports have typically a deadweight of at most 10,000 tons.
109
scheduling of tankers is not integrated with inventory and replenishment planning per
se. Transports are organized by ship operators which own fleets with different types of
tankers. Similar to rail transports, the contracts are either quantity-based or tanker-
based, i.e. either the transport fee depends on the quantity to be shipped or a complete
tanker is chartered.116 The former option is typically used for small- to medium-sized
transport orders (say up to 10,000 tons) and organized on large-scaled tankers with mul-
tiple tanks. Typically, such large-scaled tankers operate on predefined routes acquiring
numerous transport orders of various materials at various ports.117 Due to scale effects,
such tankers can operate at comparatively low cost rates per m3 ⋅ km, but they are in-
flexible due to the fixed routes. If a single transport order has sufficient size, chartering
a complete (smaller-scaled) tanker is often economically preferable. Chartering smaller
tankers generates comparatively high cost rates but offers more flexibility.
Both types of transport contracts imply a distinction between the shipper (a chemical
company) placing the transport orders and the ship operator realizing these transport
orders. Despite the fact that the chemical companies do not own or rent tankers for long
periods, an integration of transportation and inventory planning is reasonable and can
be organized by a more or less close cooperation between chemical companies and ship
operators. From the perspective of chemical companies, a close cooperation with ship
operators requires a collaborative planning of tanker routes visiting a set of production
sites under consideration of the stock levels at the sites. Such problems are categorized
as maritime inventory routing problems. In contrast, if the cooperation is less close,
only transport orders are placed e.g. on a spot market for chemical shipments. Here, the
problem is to determine start and end node, timing, and composition of transport orders
under consideration of the stock levels at the nodes. Such problems are categorized as
maritime inventory shipping problems. The next two subsections briefly review available
literature in both classes.
subtypes of the classic inventory routing which are characterized by the following facts:119
• There is a set of ports, where at least one product is consumed or produced (at a
constant rate).
• Ships are used to transport products between the ports in a finite time horizon.
If multiple periods are considered, the routes have to be coordinated in space and
time. Such problems are also categorized as maritime routing and scheduling problems
(MIRSP). Typically, the objective is to minimize the total operational costs over a finite
time horizon consisting of
• transport costs for operating the ships (mainly fuel and manpower),
Note that inventory costs are omitted by many approaches121 with the argument that all
inventories are owned by the same company, i.e. the total inventory among all nodes of the
network is assumed as independent from the routing and scheduling decisions. However,
this argument is not entirely correct in the long run since the frequency of transports
affects the total stock among the network.122 In other words, the higher the transport
frequency (i.e. the transport costs) the lower the total stock in the network (and, hence,
total inventory holding costs). In the short run, however, inventory holding costs (in the
sense of capital commitment costs) are probably sufficiently small to be disregarded as an
optimization criterion.
Contributions for MIRSPs mostly differ in two categories: the number of products
considered and the solution method(s). Basically, the underlying problem structure is
a vehicle routing problem (VRP) and/or a pick-up and delivery problem (PDP) under
maritime-specific constraints such as port and turnover capacities as well as inventory
constraints. Since VRP and PDP are hard to solve to optimality even in their basic
form, solution methods for MIRSP mainly rely on sophisticated and tailored heuristical
procedures. First contributions (Christiansen and Nygreen (1998); Christiansen (1999))
119
For a comparison with "classic" IRPs for other modes of transport see e.g. Christiansen et al. (2004)
or Andersson et al. (2010).
120
This is not a necessary condition, however most work assumes heterogeneous fleets as this is typical
for real world problems.
121
See Christiansen (1999); Al-Khayyal and Hwang (2007); Siswanto et al. (2011).
122
Also compare the discussion about pipeline scheduling, e.g. see Figure 3.5.
112
focus on one-product problems and model the underlying real-world problem as an Inven-
tory PDP with Time Windows (IPDPTW). It can be shown that inventory constraints
can be reformulated to time windows. To solve the IPDPTW, a column generation ap-
proach is proposed. Therefore, the IPDPTW is reformulated allowing a decomposition
into ship routing and harbour visiting sub-problems.123 By solving the sub-problems,
feasible ship routes and visiting sequences for all ships and harbours are generated.124
In the master problem, a combination of the identified routes and sequences has to be
determined. Therefore, a tailored branch-and-bound procedure is proposed to find an
optimal combination of routes and sequences.
MIRSPs for multiple products require a sophisticated modelling of the tanker capacity
which has to be modelled as a set of tanks dedicated to a specific product on a specific trip.
In Persson and Gothe-Lundgren (2005) it is shown that in such a formulation inventory
constraints cannot be converted into time windows. In Al-Khayyal and Hwang (2007)
the problem is simplified by fixing the assignment of tanks to products. Siswanto et al.
(2011) show that for MIRSPs considering non-dedicated tanks and multiple products even
small instances are much harder to solve to optimality than the single product problems
with dedicated tanks. Contributions also differ in the technological restrictions at ports,
e.g. how many ships or products can be served simultaneously. Christiansen and Nygreen
(1998); Christiansen (1999) and Siswanto et al. (2011) allow at most one ship to be handled
at a time in a harbour, whereas Al-Khayyal and Hwang (2007) allow multiple ships to be
docked at the same time.
Almost all contributions to MIRSP use a time-continuous formulation and consider con-
stant consumption and/or production rates for the products at the sites. One exception
is Persson and Gothe-Lundgren (2005) using a time-discrete formulation. Moreover, pro-
duction planning decisions are incorporated besides routing and inventory management.
The second exception propose Christiansen et al. (2011) where a solution procedure for a
MIRSP with time-varying consumption rates is described (but no mathematical model).
The literature on MIRSPs is large and still growing, particularly addressing specific
aspects of industrial applications. E.g. in Christiansen et al. (2011) and Stålhane et al.
(2012) applications of MIRSPs in cement industry and for LNG transports are proposed.
A related branch of literature focuses on the problem of scheduling shipments when the
ships are not owned/managed by the shipper. These maritime inventory shipping prob-
lems are discussed in the next subsection.
123
For each ship a routing model is formulated which is independent from all other ships. Similarly, for
each harbour a visiting model is proposed which is independent from all other harbours.
124
Promising routes and sequences are identified by solving the LP-relaxation of the master problem
iteratively to determine columns with least reduced costs.
113
These decisions are made based on the inventory and production constraints as well as
the cost rates offered for different types of ships by the ship operators.
Ronen (2002) proposes a MISP for a single ship type where the ships’ capacities are
not modelled in detail, e.g. no tanks are modelled and the transport capacity is only
restricted by a port-based upper bound. Only one ship can leave a departure port per
period but multiple ships may arrive at a destination port in a period. The objective
function minimizes the total shipping and inventory costs where the inventory costs are
measured as the weighted sum of absolute deviations from a desired safety stock level,
i.e. no adjustments according to the target stock level are made. The shipping costs
consist of a fixed part per dispatched ship and a variable part depending on the quantity
shipped. To overcome some lacks of detailedness of the approach presented by Ronen
(2002), the following extensions are made:
• Different types of tanks on a ship are modelled with varying sets of loadable prod-
ucts.
• Transport costs consist of a daily charter rate (depending on ship type) and fixed
port fees (depending on the departure and arrival port and the ship type).
• Inventory costs are measured as the weighted sum of relative deviations from a
target stock level.
Due to the great variety of ship types available for chemical transports (in particular
among gas tankers) it seems reasonable to reflect this option by modelling a heterogeneous
114
set of tanker types. The main features characterizing tankers from an economic point of
view are the offered transport capacity and the associated costs. The transport capacity
depends on the size and type of tanks on board. The type of tank also determines the
set of products that can be loaded. In the case of chemical tankers, this set is mainly
determined by the safety class of a tank. For gas tankers the liquefaction technology
plays an additional role.125 Hence, the number and the types of tanks installed on a
tanker determines its flexibility and capacity.
Altogether, a MISP is proposed that minimizes the total costs for shipping and stock
holding. It has to be decided on the transport volumes to be shipped, the ship types
used for the shipments and, subsequently, the assignment of chemicals to tanks on a trip.
The following maritime inventory shipment problem with ship type and tank assignment
(MISP-STA) uses the following notation given in Table 3.17.
Sets
cOs
is over-shooting cost rate
cUs
is under-shooting cost rate
cTrn
is turnover costs for chemical s ∈ S at port i
cTrv
v cost rate for chartering a tanker of type v for one period
cPort
iv port fee for tanker type v at port i
local net balance of chemical s ∈ S at port i ∈ I in period t ∈ T ; if
ωits
ωits > 0 there is a net deficit, otherwise a surplus
Cap
qsk capacity of tank type k for chemical s ∈ S
nCap
vk number of tanks of type k installed on tanker type v
sCap
is inventory capacity for chemical s ∈ S at port i ∈ I
sTar
is target stock level for chemical s ∈ S at port i ∈ I
sIni
is initial stock level for chemical s ∈ S at port i ∈ I
initial flows of chemical s ∈ S from i to j in period t ∈ T in period
xIni
ijτ s
τ ∈ {1 − tTrv
ij , ..., 0}
tTrv
ij integer valued travelling time on link (i, j) ∈ L
Decision variables
125
E.g. Ethylene can be transported efficiently under super-cooled conditions only. For most other basic
chemicals also liquefaction by pressure is possible. See Stopford (2009).
115
Table 3.17: Sets, parameters, variables, and decision variables for the MISP-STA
Since it is assumed that tankers are chartered for a specific transport order, it is rea-
sonable to assume that a time-dependent charter rate is charged when a tanker is booked.
Hence, the charter costs on link (i, j) ∈ L are calculated by a charter rate per period cTrv
v
multiplied by the trip’s duration tTrv
ij .
126 Furthermore, port fees are to be expected de-
pending on the handling time (for unloading or loading) as well as the size of the ships.127
These together constitute the transport costs depending on the travel time, ship type,
and ports (see first two rows of (3.65)).
As a counterpart, inventory costs have to be faced. As discussed for the MC-RTP model,
it is assumed that target stock levels (in the sense of safety stocks) are predefined for each
port/production site and product. Deviations from these target levels are to be minimized
and are included in the objective function multiplied by specific cost rate equivalents.
These cost rate equivalents are higher for negative deviations (i.e. undershooting) and
lower (or zero) for positive deviations (i.e. overshooting). As before, undershooting cost
rate equivalents reflect the expected costs for a plant shutdown due to raw material short-
age whereas overshooting cost rate equivalents reflect unnecessary stock holding costs. As
in the MC-RTP, stock deviations are measured relative to the target stock levels taking
into account varying target stock levels.128 The total costs of transports and inventory
holding are minimized and formulated as
T C = ∑ ∑ ∑ (yijtv ⋅ (tTrv
ij ⋅ cv + civ
Trv Port
jv )) +
+ cPort
(i,j)∈L t∈T v∈V
∑ ∑ ∑ ∑ (zijtsk ⋅ (cTrn
is + cjs )) +
Trn
126
To reflect distance-dependent discounts the cost rates can be assumed to be negotiated for each trip
individually, i.e. cTrv
ijv .
127
See Stopford (2009) for more details.
128
See the discussion in subsection 3.3.2.
116
∑ ∑ ∑ (cOs
is ⋅ oits + cis ⋅ uits ) .
Us
(3.65)
i∈I s∈S t∈T
− ∑ xijts ∀i ∈ I, s ∈ S, t ∈ T (3.66)
j∈I
sits
− 1 = oits − uits ∀i ∈ I, s ∈ S, t ∈ T (3.69)
sTar
is
si,0,s = sIni
is ∀i ∈ I, s ∈ S (3.70)
sits ≤ sCap
is ∀i ∈ I, t ∈ T , s ∈ S (3.71)
xijτ s = xIni
ijτ s ∀(i, j) ∈ L, τ ∈ {−tTrv
ij ...0}, s ∈ S (3.72)
xijts ∈ R+ ∀i ∈ I, j ∈ I, t ∈ T , s ∈ S (3.73)
Constraints (3.66) constitute the inventory balance equations consisting of the previous
period’s stock, local consumption/production (ωits ), and in-/outflows from/to other ports.
Constraints (3.67) assure that a sufficient number of tanks is assigned to chemical s in
order to store the shipping quantity xijts on board. Similarly, constraints (3.68) restrict
the total number of dedicated tanks on a trip to the number of tanks installed on board
of the chartered ships. Relative stock deviations from the desired target stock levels
are calculated by constraints (3.69). Constraints (3.70) initialize the stock levels at the
beginning of the planning horizon whereas (3.71) assure that stock capacities are not
exceeded. Finally, constraints (3.72) incorporate flows from previous periods which allows
applying the model in a rolling horizon environment.
The following example shows the applicability of the MISP-STA by means of the same
artificial setting as described in example 6.
Example 7 (Maritime inventory shipping). Assume three ports with associated production
sites where three chemicals are produced/consumed. Shipments are planned for a 14-
117
1 1 2 4 0 500 0
2 1 4 8 450 100 600
cTrv
v 200 300 400 – – –
Table 3.18: Technical specification of tankers and tanks (nCap Cap Trv
vk , qsk , cv )
optimality in about 5.7 hours, the corresponding minimal total cost is 13,514.130 Again, the
MISP-STA can be categorized as a specific network design problem where arc capacities
are associated with fixed costs for chartering ships. Therefore, the same remarks w.r.t. the
model’s complexity hold as for the MC-RTP. In this problem instance, however, the ratio
between transport and stock holding cost is less favourable as the transport cost rates are
considerably higher than in the MC-RTP instance. Moreover, the transport capacity is
not restricted by a predefined set of RTCs but purely depends on the number of chartered
ships.
Figure 3.12 shows the resulting optimal transport quantities dispatched in the first eight
periods.131
Mainly tanker type v = 1 is chartered since handled quantities are sufficiently small.
Exceptions can be observed in periods t = 1, t = 4 and t = 7 where chemicals s = 1 and s = 3
are shipped jointly which requires ship type v = 2. Primarily, chemical s = 3 is transported
from port/site i = 2 to i = 1 and i = 3. Chemical s = 1 is shipped from i = 1 and i = 2 to
i = 3. The transports from i = 1 to i = 3 are routed via i = 2. Note that direct transports
from i = 1 to i = 3 last four periods whereas indirect transports via port i = 2 last only
129
Due to the longer transport and handling times, a rougher time scale can be expected in the mar-
itime/waterway context compared to rail context. I.e. a period’s length is expected to be one or two
day(s).
130
The calculation was performed on a 2.6 GHz machine with IBM CPLEX 12.5, see IBM (2010).
131
Associated indices refer to the transported chemical and the used ship type on a trip (i, j, t).
118
t =8 t =2 t =3 t =4 t =5 t =6 t =7 t =9
450 500 550 600 300 350 400 450
600 500 400 600 500 400 700 600
i =8
475 450 425 400 375 350 325 300
x s=3 = 400
x s=3 = 300 y v=8 = 8
y v=8 = 8 x s=8 = 350
y v=8 = 8
400 450 500 400 450 500 400 450
100 900 8000 8000 900 8000 900 8000
i =2
275 250 225 350 325 300 275 250
three periods. Hence, indirect transports induce lower transport costs. Chemical s = 2 is
transported from the only producer site i = 3 to site i = 2.132
Comparing Figure 3.12 and Figure 3.10, it can be observed that the total quantity shipped
in the first periods is higher for the maritime setting than for the corresponding rail setting
(example 6). This is caused by the restricted number of RTCs in the rail setting. The
RTCs have to be re-positioned to perform material transports. This leads to comparatively
more balanced stock levels in the maritime setting. To illustrate this fact, Figure 3.13
shows the stock levels and target stocks of all chemicals at all ports over the complete
planning horizon.
In the maritime setting, all stock levels are closer to their corresponding target stock
levels compared to Figure 3.11. This is particularly apparent for chemicals s = 1 and s = 3
at site i = 3. In the rail setting the stocks of both chemicals were below their target stocks
(almost) over the complete planning horizon. In the maritime setting, both stocks are
considerably closer to their target stocks. This corresponds to the stock patterns of both
chemicals at site i = 2 where the corresponding stocks’ overshooting is reduced. I.e. in
the maritime setting transports of chemicals s = 1 and s = 3 from site i = 2 to i = 3 are
economically preferable (in contrast to the rail problem).133
In general, the maritime setting offers more transport flexibility in the sense that ships
can be dispatched independent of the availability of RTCs. On the other hand, the freight
consolidation on ships is restricted by the technical constraints of the tanks available on
board. In the given examples, the first fact prevails resulting in smoother stock patterns.
In contrast to the MC-RTP, the MISP-STA is more complex due to the larger number
of integer variables to be determined. To solve larger instances, a heuristical procedure
can be set up as follows:
1. Build a relaxed MISP-STA by dropping the ship flow variables yijtv and (3.68). Re-
place the first part of (3.65) by
∑ ∑ ∑ ∑ zijtsk ⋅ tTrv
ij ⋅ c̃k
Trv
(3.75)
(i,j)∈L t∈T s∈S k∈K
Similar to the MC-RTP, a considerable part of the model’s complexity is inherited from
the integer decisions determining the number of tankers. For variabilization, the cost rates
132
Transports to site i = 1 take place in later periods.
133
Compare also the transport flows displayed in Figure 3.12 and Figure 3.10.
120
chemical s = 1
location i = 1 stock level
chemical s = 2 target stock
chemical s = 3
750 1000
500
stock level
250
0
location i = 2
1250
750
stock level
250
0
location i = 3
750
stock level
250
0
0 2 4 6 8 10 12 14
period
Figure 3.13: Optimal stock levels for the MISP-STA instance of example 7
121
c̃Trv
k approximate the shipment charges on a tank basis and are calculated as follows
The heuristic delivers a solution for the described problem instance within 8 seconds. The
solution’s total cost is 14,379 which is approximately 6% above the optimum.
123
• multiple units: Several business units (companies) are responsible for the transfor-
mation processes.
Hence, a chemical’s SC can be defined as all (business) units and processes involved in
the production and distribution of this chemical. From this point of view, chemical pro-
duction networks of large-scaled chemical companies are major parts of the SCs of many
chemicals. A particular property of chemical SCs is that a comparatively small number of
independent companies is involved in the SC. This is because chemical companies are of-
ten deeply integrated, i.e. many subsequent production steps are processed by a chemical
company. On the supply side, raw material suppliers such as oil-producing companies and
refineries are to be taken into account. Moreover, logistical service providers managing
transport and distribution processes play a vital role in chemical SCs.
To manage SCs, a broad body of (scientific) literature has emerged since the 1980s.1
Supply chain management (SCM) aims at the planning, execution, and control of all
transformation processes in a defined SC such that the SC’s performance is optimized.2
The main focus of SCM is on the coordinated planning of the SC processes. For planning,
1
See Croom et al. (2000) for an overview.
2
This is often operationalized as minimizing the SC’s total costs or maximizing the SC’s total profit.
See Stadtler (2005) for a more detailed discussion and definition of SCM.
T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,
DOI 10.1007/978-3-658-08433-2_4, © Springer Fachmedien Wiesbaden 2015
124
SCM draws many approaches from related disciplines (such as logistics and operations
management), but focuses on their integration and interactions.
In the management of chemical SCs, product flows within chemical production networks
are to be planned from a focal point of view. I.e. local and network-wide processes have to
be managed such that the production network’s total performance is optimized. Figure 4.1
shows the schematic overview of the exemplary chemical SC where relevant elements for
integrated planning are highlighted.
intra-site inter-site
S1 C1
intra-site
intra-site
S2 C2
Figure 4.1: Chemical SC scheme for integrated planning (relevant elements highlighted)
For integrated planning approaches, all elements in the considered chemical SC are
relevant including transport relations to (external) customers and suppliers. Only external
suppliers or customers themselves are not modelled in detail.
Core elements of chemical SCs considered in integrated planning approaches at the
tactical/strategical level are capacities and parameters of the production systems and
the logistical systems (e.g. turnover or inventory capacities). Since sites and plants are
interconnected, local adjustments more or less immediately affect the remaining sites/-
plants. Integrated approaches capture these spill-over effects by a combined modelling of
the interdependent components.
Integration not only focuses on the spatial or temporal dimension of an SC but also
on separated planning (sub-)problems. These (sub-)problems and their interdependencies
are modelled in a common framework. Basically, two options for integration can be
distinguished:3 The first option is to merge the sub-problems into a monolithic model
(also called deep integration). The second option is to stay with the decomposition into
3
See Dolk and Kottemann (1993) or Geoffrion (1999).
125
1. The decomposition of the monolithic model into suitable sub-problems and setting
up a functional integration scheme.4
The following subsection reviews the literature on integrated planning problems with
special focus on (basic) chemical industry. The papers reviewed are categorized according
to methodical features and problem characteristics.
4
Note that the suitable sub-problems do not need to be the same as the original sub-problems.
5
See Meyr et al. (2008) and Stadtler (2005). Note that there is a lot of work from numerous scientific
perspectives labelled with SCM, see Croom et al. (2000) for a detailed analysis and categorization of
SCM approaches. Based on this classification scheme this work reviews SCM approaches at a network
level modelling material flows.
126
the demand planning module which comprises models to compute demand forecasts for
different regions and time horizons by assuming specific types of demand processes.
At the operational level, problems are considered on a detailed perspective dealing with
the short-term planning and control of basic SC operations. The aggregated material flows
determined at the tactical level are broken down to the local level on a short-period basis
like a daily or hourly time frame. For the pre-determined production quantities assigned
to a specific production site on the tactical level, plans and schedules for all plants at this
site are derived in the production planning & scheduling modules. Similarly, replenishment
and transport decisions are determined in the inventory management and distribution &
transport planning module, respectively.
The general SCM matrix can be adapted to better reflect the specific characteristics of
basic chemical industry.6 In the traditional SC planning matrix, the material requirement
planning modules are concerned with determining order quantities of raw and intermediate
materials necessary to realize a pre-defined production plan. These tasks are particularly
challenging if the variety of handled materials is large.7 In basic chemical industry, how-
ever, the variety of raw and intermediate chemicals is comparatively small while their
demand is high and mostly determined by the plants’ technical specifications.
Moreover, the means of transport differ for procurement and distribution transports.
For raw and intermediate transports often pipeline, ship, and rail are used. In contrast,
for the distribution of final chemicals rail and road transports are prevalent. As pipeline,
ship, and rail are comparatively inflexible and differently organized compared to road
transports, an independent planning of procurement and customer transports is justified.
Therefore, the procurement planning module encompasses the planning of procurement
transports of raw and intermediate chemicals and the planning of their local stocks. In
contrast, the distribution transports module comprises the tasks of the classic distribution
& transport planning module for final chemicals.
The transport capacities provided along a chemical SC have to be planned on the tac-
tical level. In the case of ships or RTCs, the corresponding fleet sizes are determined. To
enlarge fleets, equipment can be bought or rented. To reduce fleets, equipment can be sold
or renting contracts may not be prolonged. Such fleets are sometimes used commonly by
multiple chemical companies e.g. by short-term renting contracts or by exchanging RTCs
or ships (so-called swaps). Therefore, the original master production planning module is
complemented by the master transportation planning module capturing all tactical de-
cisions about the capacities of the logistical system in the SC.8 Pipeline capacities are
determined by technological details and, thus, are subjects to strategical issues as they
cannot be adapted easily in the short run.
6
See Zoryk-Schalla et al. (2004) for a report from an Aluminium-producing company about implementing
an advanced planning system based on an adapted SCM matrix.
7
See Meyr et al. (2008).
8
Note that this module also contains planning problems determining turnover capacities at the sites if
they limit the transport quantities.
127
The distinction into a strategical, tactical, and operational level is retained but re-
named into design, configuration, and operations to better reflect the terms most used in
literature.9 Figure 4.2 shows the adapted SCM matrix.
The basic planning problems corresponding to the specific modules are briefly described
as follows:
• configuration:
– Master production planning: This module encompasses decisions about the
usage of production capacities of the considered SC, e.g. the determination of
production quantities and their assignment to production sites.11
– Master transportation planning: Determination of aggregated transport flows
among the participants of the SC and provided transport capacities.12
– Demand planning: Determination of demand forecasts for a finite planning
horizon. They are based on historical records of former periods, external pre-
9
See e.g. Melo et al. (2009).
10
See e.g. Vidal and Goetschalckx (1997) or Tsiakis et al. (2001) for an overview.
11
In chemical industry this is often encompassed by campaign planning problems. See Kallrath (2005) or
Grunow et al. (2002).
12
Examples are e.g. rail car fleet sizing models (Cheon et al., 2012), container fleet sizing (Dong and
Song, 2009), or ship fleet sizing (Ronen, 1993).
128
dictors (e.g. economic growth indicators), and assumptions about the demand
process.13
• management
– Production planning: Determination of local production quantities and assign-
ment to local plants.14 For multi-product batch processes, the aim is to define
batch sizes for different chemicals and to assign them to available plants.15 In
case of multi-product continuous processes, the aim is to determine the produc-
tion mode, i.e. the physical conditions and raw material composition specifying
the product mixture.16
– Production scheduling: Determination of mode sequences/schedules of batches
assigned to the plants. The changeovers between production modes typically
induce costs for cleaning, maintenance, and/or lost material.17
– Procurement planning: Determination of transport modes and transport ca-
pacities for procurement transports as well as planning of inventories of raw
and intermediate chemicals at a site.18
– Distribution planning: Planning of distribution processes, e.g. management of
distribution stocks, routing, and scheduling final chemical transports to cus-
tomers.19
– Demand fulfilment: Management of customer orders, e.g. tracking of orders
along the production process, acceptance of new orders, setting of initial order
due dates.20
The modules displayed in Figure 4.1 address different planning problems and associated
planning models. There are plenty of interdependencies between the blocks as the out-
come of a particular module is typically the basis for subsequent lower-level modules. In
hierarchical planning, long-term/strategic decisions form the basis for tactical planning
whose outcome typically guides operational plans. In a hierarchical planning approach,
top-level decisions anticipate the low-level decisions to a certain extent.21 However, the
13
See Stadtler (2005).
14
The distinction between production planning and scheduling problems is not always consistent, see
Kallrath (2002) or Berning et al. (2004).
15
For an overview on integrated planning and scheduling, a detailed description in the context of multi-
product plants can be found in Maravelias and Sung (2009).
16
This problem is extensively studied for refinery operations planning leading to non-linear process mod-
els, see Zhang et al. (2001), Li et al. (2005), or Alhajri et al. (2008).
17
See e.g. Méndez et al. (2006) for an extensive overview and classification for scheduling of chemical
batch processes.
18
This comprises e.g. inventory routing models for procurement via ship or pipeline (see Ronen (2002)
or Moura et al. (2008)) and local tank management models (see e.g. Saharidis et al. (2009)).
19
This block comprises e.g. the branch of MIRSPs (see e.g. Christiansen et al. (2004)) and most pipeline
scheduling models (see e.g. Cafaro and Cerdá (2010) or Neiro and Pinto (2004)).
20
See Kilger and Meyr (2008).
21
See e.g. Fleischmann and Meyr (2003).
129
anticipation is never perfect such that the final solution of top- and base-level problems is
typically not (globally) optimal. To overcome this deficiency, integrated planning models
are often build by aggregating planning problems from adjacent blocks either horizontally
or vertically.
To categorize the literature on integrated planning, the modelling technique is an im-
portant criterion. Different types of models are suitable depending on the planning prob-
lem’s requirements. In general, when addressing long-term problems, assumptions about
the occurrence of future events have to be made. However, the future is known for its
uncertainty. Hence, techniques capable to handle uncertainty are prevalent. Such ap-
proaches can be further distinguished according to the stochastic elements considered
(which is typically the demand). For short-term problems, uncertainty is (most often)
not an obstacle. For operational problems, mathematical optimization is widely used,
albeit simulation is still a reasonable option.22 Mathematical optimization models are
categorized w.r.t. linearity (linear vs. non-linear), the domain(s) of decision variables (in-
teger/binary vs. continuous), and stochasticity (stochastic vs. deterministic). Simulation
techniques come to the fore when at least one of the following criteria prevails:23
In general, these criteria often hold for planning problems with a rather long planning
horizon. Therefore, simulation-based models are prevalent especially for SC configuration
problems.24 Special classes of simulation techniques are distinguished.
In this work the term simulation refers to stochastic and dynamic models.25 If all
components are modelled continuously, this is typically called a system dynamics model.26
Such a model is reasonable either at a high aggregation level27 or when all modelled
processes are indeed continuous.28 The system studied is tracked continuously and its
(aggregated) behaviour can be described at any point in time.
22
See Papageorgiou (2009) for details and Adhitya and Srinivasan (2010) for an application of simulation
for a detailed process modelling.
23
See Law (2007) or Carson (2004) for a more detailed discussion about these prerequisites. The sub-class
of deterministic simulation is typically referred to as computer experiments, see Santner et al. (2003)
for more details. In this work the focus is on stochastic simulation models.
24
See Kleijnen and Smits (2003) for an overview.
25
I.e. there exist stochastic process elements to be modelled and the system is studied over time, see Law
(2007) for a more detailed classification.
26
See Ogata (2003) for an introduction.
27
E.g. when (discrete) objects handled in a network can be aggregated to continuous flows.
28
For more details about system dynamics models in SCM see e.g. Kleijnen (2005). For applications see
e.g. Rabelo et al. (2007) or Venkateswaran et al. (2004).
130
In contrast, if the processes of the modelled system are discrete, the corresponding
model is called a discrete-event simulation. Here, the system’s states are calculated at a
finite number of points in time. This type of simulation is often used for atomic simula-
tions, i.e. when a system is modelled in detail.29
A third category are so-called agent-based simulation models. This sub-class of sim-
ulation models is characterized by a network of interacting agents. Each agent reacts
to stimuli from other agents and environmental variables depending on mathematically
formulated decision rules.30 This type of simulation is particularly useful to model inter-
actions between interrelated, but independent entities such as business units, companies,
or market actors.31
To categorize the literature on integrated SC planning in chemical industry the following
criteria are used:
• SCM matrix refers to the modules of the adapted SCM matrix (Figure 4.2) addressed
by the reviewed reference. Capital letters indicate the planning level (design, con-
figuration, and operation) with at most two superscripts. The first superscript is
p, t or d referring to transportation, production, or demand. The potential sec-
ond superscript specifies the sub-problem on the operational level with p, s or d
referring to procurement/production, scheduling or distribution/demand fulfilment.
E.g. the combination Otp refers to the procurement planning module whereas the
combination Opp refers to the production planning module.
• Modelling technique refers to the technique used to formalize the planning problem
and comprises mathematical optimization approaches such as (mixed-integer) linear
programs [(MI)LP], (mixed-integer) non-linear programs [(MI)NLP], or stochastic
programs [s(MI)LP] as well as simulation techniques such as agent-based models
[AB], discrete-event models [DE], or system dynamics models [SD]. If mathematical
optimization models are embedded in a simulation framework, this is indicated by
a connector. E.g. DE-MILP indicates that a MILP is embedded in a discrete-event
simulation framework.
• Objective refers to the objective pursued (such as costs, net present value [NPV],
corporate value [CV], or profit). If multiple objectives are considered, this is stated
by [mu].
• Time scale describes the way the time dimension is represented. Continuous [co] or
discrete [di] models are distinguished if time is modelled explicitly. Otherwise, the
model refers to a single period [si].
• Transport mode refers to the transport mode(s) considered (ship [sh], pipeline [pi],
road [ro], or rail [ra]). For aggregated models only flows might be relevant. Then,
the mode is unspecified [us].
• Solution method refers to the method applied to solve the proposed model e.g. (com-
mercial) standard solvers [standard] (such as CPLEX or CONOPT), a specific op-
timization method [specific], a hierarchical decomposition approach [decomp.], or a
heuristical procedure [heur.]. For simulation models, often a finite set of scenarios
is evaluated and compared [scen.]. But also genetic algorithms are sometimes used
for simulation optimzation [GA].32
If a criterion is not included in the proposed model, this is indicated by a horizontal line.
In general, simulation approaches are advantageous if a static framework of entity-
processing units exists, i.e. the system’s general structure is static. But for strategic
problems the structure of a SC is typically the subject to be altered. As this work aims at
a framework relying on simulation, literature primarily focusing on strategical problems
is out of focus.
At the tactical level, however, more simulation-based approaches can be found. Ta-
ble 4.1 shows a classification of literature for integrated SC configuration and operations
planning where 7 out of 20 reviewed articles propose simulation-based approaches.
Among the simulation approaches Garcia-Flores and Wang (2002), Mele et al. (2006),
and Puigjaner and Guillén-Gosálbez (2008) propose agent-based simulation models where
the agents (partially) utilize MILPs to determine their decisions (hybrid model).33 These
approaches primarily investigate the organizational structure of an SC, i.e. the interac-
tion of involved decision making units. In Garcia-Flores and Wang (2002) a chemical SC
producing paints and coatings is modelled by using six classes of agents: retailers, lo-
gistic service providers, warehouses, purchasing departments, plants, and suppliers. The
production processes at the plants are organized as multi-purpose, multi-product batch
processes. The precise operational planning and scheduling of batches is carried out by a
MILP implemented in a standard software package. Decision fields are the parameters of
the inventory policy and the parameters of the negotiation scheme between the warehouse
in cases of stock-outs. I.e. in case of a stock-out at a warehouse, the requested order is
32
This classification encompasses the solution methods applied in the reviewed literature. A more detailed
classification (e.g. addressing meta-heuristics, math-heuristics, etc.) is omitted for the sake of brevity.
33
E.g. site managing agents decide about the assignment of production orders to plants whereas financial
agents decide about loans to be taken.
132
type of
time transport uncer- solution
Reference SCM matrix modelling objective produc-
scale mode tainty method
tion
McDonald and Karimi (1997) (Opp , Ops , C p , C t ) M ILP cost ba di us de standard
Vidal and Goetschalckx (2001) (C p , C t ) N LP profit co si us — heur.
Gjerdrum et al. (2002) (C p , C t ) M IN LP multiple us di us — specific
pp
Gupta and Maranas (2003) (O , Ops , C p , C t ) sM ILP cost ba di us de decomp.
Jackson and Grossmann (2003) (C p , C t ) N LP profit co di us — decomp.
Gupta and Maranas (2004) (C p , C t ) sM ILP NPV/ROV us di us de standard
Ryu et al. (2004) (C p , C t ) sLP profit us si us de specific
Chen and Lee (2004) (C p , C t ) M IN LP multiple ba di us de+pr standard
Oh and Karimi (2006) (C p , C t ) LP profit co di us — standard
Yi and Reklaitis (2007) (Opp , Ops , C p , C t ) M IN LP cost ba co us — specific
Amaro and Barbosa-Póvoa
(Otd , Opp , Ops , C t , C p ) M ILP profit ba di mu — standard
(2008)
Al-Othmann et al. (2008) (C t , C p ) sLP profit co di us de+pr standard
pp
Kim et al. (2008) (O , C p , C t , D) M IN LP profit co si mu — standard
Garcia-Flores and Wang (2002) (Opp , C t , C p ) AB + M ILP cost ba di us mu scen.
Jung et al. (2004) (Opp , Ops , C p ) DE + M ILP cost ba di us de specific
Mele et al. (2006) (Ops , Opp , C t , C p ) AB + M ILP profit ba di us mu GA
Puigjaner and Guillén-Gosálbez
(C t , C p ) AB + M ILP multiple us co us mu GA
(2008)
ps pp t p
Jung et al. (2008) (O , O , C , C ) DE + M ILP inventory ba di us de scen.
Pitty et al. (2008) (Otp , Opp , C t , C p ) DE profit co di pi/sh mu scen.
Adhitya and Srinivasan (2010) (Opp , Ops , C t , C p ) DE profit ba co us mu scen.
abbr.: ba...batch; co...continuous; de...demand; di...discrete; mu...multiple; pi...pipeline; pr...price; ra...rail; ro...road; sh...ship;
si...single; us...unspecified
Table 4.1: Classification of literature on integrated SC configuration and management planning in chemical industry
133
delivered from another warehouse and the charge of additional transport costs is subject
of the negotiation process between the agents. However, logistical details determining the
additional transport cost are out of scope.34 The optimization is based on an evaluation
of a pre-defined set of scenarios. Mele et al. (2006) propose a similar simulation approach
with the same assumptions for the logistical processes and production processes as well
as the same optimization model for operational planning and scheduling of batches. This
work proposes an optimization procedure based on a genetic algorithm which seeks for an
inventory parameter constellation maximizing the SC’s expected total profit. Puigjaner
and Guillén-Gosálbez (2008) extend this work by dealing with multiple objectives. The
applied multi-objective genetic algorithm is NSGA-II.35
Another hybrid simulation model is proposed by Jung et al. (2004, 2008). Here, a
discrete-event simulation is chosen to model the stochastic environment. The focus is on
the organization of production processes, i.e. optimization of operations is in focus. In
Jung et al. (2004) a single-stage production-distribution problem is considered where the
production planning and scheduling of multiple batch plants is embedded in a stochastic
rolling horizon environment.36 The objective is to maximize the expected profit by decid-
ing on the safety stocks at the plants without falling below a minimum customer service
level. The customer demand is the only stochastic component. For optimization a stochas-
tic gradient approach is applied which iteratively increases the safety stocks for each com-
bination of product and plant until the expected profit cannot be increased any further.
Jung et al. (2008) extend this work by considering a multi-stage production-distribution
system with pure inventory holding facilities and combined facilities for production and
inventory holding. For both types of facilities the influence of inventory parameters on
performance measures is analysed by simulation experiments. Thee approaches incorpo-
rate operational decisions by MILPs. Hence, optimal reactions on varying system states
are modelled on the operational level. However, both models omit logistical details on
the operational level (only flows between facilities are considered).
Pure simulation approaches are proposed by Pitty et al. (2008) and Adhitya and Srini-
vasan (2010). Pitty et al. (2008) propose a discrete-event simulation model for a refinery
supply chain. Operational decisions such as unloading schedules and production planning
are made based on simple priority rules. Various configurations of the modelled SC are
studied and compared to reveal optimization potentials. This approach explicitly consid-
ers some details of ship and pipeline transports. Adhitya and Srinivasan (2010) describe a
discrete-event simulation model for an SC producing and distributing lubricant additives.
Here, batch production is modelled. Again, operational production decisions are made by
priority rules and a scenario analysis is conducted to evaluate the effects of other priority
34
A classic transportation problem is solved to determine the total transport effort.
35
See Deb et al. (2007).
36
For production planning the MILP proposed by McDonald and Karimi (1997) is used. Production
schedules are determined by heuristics to ease solvability.
134
planning depends on the realized demand, production planning is based on the estimated
demand. A two-stage stochastic linear program is created seeking for a configuration
with minimum expected total costs. In Gupta and Maranas (2004) the objective of the
model is modified by considering future payments. For discounting future payments, two
approaches are discussed: The net present value (NPV) and the real-options-based value
(ROV).42 In Ryu et al. (2004) the hierarchy of production and distribution planning is
inverted, i.e. the distribution problem rules the production planning problem. Further-
more, both components are affected by the uncertainty of demand. To solve the problem,
a parametric programming approach is described. Al-Othmann et al. (2008) present a
scenario-based sLP for the optimization of a petroleum SC consisting of companies from
the crude oil sector as well as the refinery, petroleum, and basic chemical sector. The
objective is to find a configuration maximizing the expected profit. A two-stage stochas-
tic linear program is proposed whereby decisions about the production quantities in the
crude oil sector constitute the first stage problem43 whereas the decisions for the other
sectors are made based on a discrete number of demand scenarios.
A special part of the chemical industry is the oil-producing industry. Core production
units in this industry are refineries. When refinery operations are modelled, typically
non-linear models are used due to the non-linear production yield curves of distillation
units.44 Approaches for refinery production-distribution planning are proposed by Kim
et al. (2008), Amaro and Barbosa-Póvoa (2008), Jackson and Grossmann (2003), Oh
and Karimi (2006), and Yi and Reklaitis (2007). Among these, only Kim et al. (2008)
and Amaro and Barbosa-Póvoa (2008) incorporate some logistical aspects by considering
multiple transport modes with varying transport cost rates. In Kim et al. (2008) the
production network of a Korean petroleum company is modelled. The model considers
multiple, non-specified modes of transport (with varying cost rates). Amaro and Barbosa-
Póvoa (2008) also deal with unspecific transport modes but the proposed model considers
batch production processes. Jackson and Grossmann (2003) use non-linear expressions
for modelling continuous production processes (similar to Kim et al. (2008)). To solve the
production-distribution problem, a decomposition approach is proposed by splitting the
model into the spatial and temporal dimension. A simplified production process model
is used by Oh and Karimi (2006) where the product fractions are handled as constants
and no reflux streams are considered. This work focuses on financial impacts of duty
drawbacks for planning production and distribution in multi-national companies. An
LP is presented which seeks for the production-distribution plan maximizing the total
after-tax profit. The influence of tax differences and exchange rates between different
currencies is considered by Yi and Reklaitis (2007). Here, batch processes are assumed for
production and a periodic continuous-time model is formulated. An optimal configuration
42
Both differ in the interest rate used for discounting and the definition of the expected return. In case
of the NPV the expected rate of return is used whereby the ROV utilizes the risk-free rate of return.
43
I.e. the decisions are independent from demand fluctuations by utilizing expectations.
44
See Li et al. (2005).
136
of the studied network is exhibited by deriving analytical expressions for production and
order lot-sizes.
For tactical production planning roughly assigning production quantities to production
sites suffices for deriving reliable mid-term plans.45 On the operational level, however,
a more detailed view on the involved production processes is needed. In particular, the
planning of batch processes is a challenging task as it involves
Planning of continuous processes faces a similar problem structure except that the batch
size is not limited by technological restrictions. Beside production planning, the planning
of logistical activities needs to be involved. Since (operational) production planning typi-
cally focuses on one production site, intra-site transports (which are typically based on the
pipeline mode) need to be considered but also inter-site transports or customer deliveries
are of importance in daily business in some cases.
Table 4.2 shows a literature classification for integrated approaches primarily addressing
operational planning problems in chemical SC management.
All approaches summarized in this table contain a production planning and scheduling
component and are deterministic. The integrated production planning and scheduling
problem (IPPSP)49 is the core component of most approaches. The approaches differ in
details regarding the production system and the incorporation of logistical aspects. The
reviewed approaches can be subdivided into approaches modelling batch and continuous
production.
Batch production is considered by the first seven approaches shown in Table 4.2. Timpe
and Kallrath (2000) propose a time-discrete MILP for the IPPSP with multiple production
sites/plants, multiple production modes, intermediate storages, sales points, and trans-
ports between sites, storages, and sales points. The model uses different time scales for
production and sales planning which allows a more precise planning of production. In a
production period, a site can produce in different modes whereby changeover times and
costs are considered. To store intermediate and raw materials, one-product and multi-
product tanks are available and an assignment of products to tanks is implemented. The
45
Rough assignment refers to an assignment to aggregated production facility groups such as similar
plants or even complete sites.
46
This is due to the fact that typically batch processes have limited capacity, e.g. the volume of the
reactor is limited.
47
This is necessary in case of multi-plant networks which are typical for producing e.g. speciality chemi-
cals.
48
This might be necessary due to sequence-dependent set-up efforts and due dates.
49
See Maravelias and Sung (2009), Li et al. (2010a) or Shao et al. (2009) for an introduction and literature
review. For an application of the ELSP in the pure batch scheduling context in chemical industry see
Cooke and Rohleder (2006).
type of
model- time transport uncer- solution
Reference SCM matrix objective produc-
ling scale mode tainty method
tion
Timpe and Kallrath (2000) (Otd , Opp , Ops ) M ILP sales ba di us — standard
Grunow et al. (2002) (Otd , Opp , Ops ) M ILP cost ba di us — decomp.
Neumann et al. (2002) (Opp , Ops ) M ILP other ba co — — heur.
Romero et al. (2003) (Opp , Ops ) M ILP revenue ba di — — standard
Berning et al. (2004) (Opp , Ops ) other profit ba di — — GA
Guillén et al. (2007) (Otd , Opp , Ops ) M ILP equity ba di us — standard
Sung and Maravelias (2009) (Opp , Ops ) M IP cost ba di — — decomp.
tp
Bok et al. (2000) (O , Otd , Opp , Ops ) M ILP profit co di us — decomp.
Pinto et al. (2000) (Otd , Opp ) M ILP profit co si pi — standard
Neiro and Pinto (2004) (Otp , Opp , Ops ) M IN LP cost co di pi — standard
Erdirik-Dogan and Grossmann
(Opp , Ops ) M ILP profit co co — — decomp.
(2006, 2008)
tp pp ps
Shah and Ierapetritou (2011) (O , O , O ) M ILP profit co co pi — standard
DE +
own approach (Otp , Opp , C t , C p ) mu co di pi/ship/rail mu specific/GA
M ILP
abbr.: ba...batch; co...continuous; de...demand; di...discrete; mu...multiple; pi...pipeline; pr...price; ra...rail; ro...road; sh...ship;
si...single; us...unspecified
model is solved with standard solvers for multiple objectives. Maximization of sales is
finally recommended which best fits to the intended application in practice.
Grunow et al. (2002) propose a sequential planning approach consisting of the pro-
duction planning step (called campaign planning) as well as the assignment and schedul-
ing step (called assignment model). The focus is on one production site with multiple
multi-purpose processing units. Transports between plants are mentioned but not ex-
plicitly modelled.50 The individual models are solved with standard solvers seeking for a
minimum-cost solution. A similar decomposition scheme is proposed by Neumann et al.
(2002). Here, the batch planning problem is formulated as a MILP whereas the schedul-
ing problem is solved by a heuristic considering limited plant capacities. This way, a
re-scheduling step is avoided and intermediate storages are introduced. A quite simi-
lar but integrated model for batch sizing and scheduling is proposed by Berning et al.
(2004). Due to the model’s complexity, a genetic algorithm is used to derive near-optimal
solutions. Sung and Maravelias (2009) propose an algorithm to determine the set of fea-
sible solutions for a production planning problem by evaluating infeasible polytopes of
the underlying scheduling problem. The resulting linear inequalities can be added to the
production planning model to find an optimal and feasible production plan (w.r.t. to the
scheduling constraints).
In Romero et al. (2003) a simplified IPPSP for chemical batch production is described
which is enhanced by budget constraints. A one-stage batch production on parallel,
identical plants is considered. The goal is to maximize the total revenue over the planning
horizon without violating budget and resource constraints. Due to the simple production
model, the proposed MILP can be solved with standard solvers. This approach is extended
by Guillén et al. (2007) where multiple plants as well as transports between plants are
incorporated next to cash flow and budget constraints. The objective is to maximize the
accumulated equity during the planning horizon, i.e. the difference between accumulated
cash and outstanding liabilities.
For continuous production planning and scheduling, there is no technological production
capacity limit. The number of production modes is typically smaller for continuous pro-
duction plants which eases the scheduling problem. Erdirik-Dogan and Grossmann (2006,
2008) consider a single production site with continuous single-stage production. Erdirik-
Dogan and Grossmann (2006) present the "basic" production planning and scheduling
problem for a single production unit. The production unit can be set up to multiple
products but only one product can be produced at a time. Deterministic demand fig-
ures serve as lower bounds for the portfolio of chemicals to be produced. The model
is formulated in continuous time51 whereby a rough division of the planning horizon in
time periods is used. This allows the introduction of demand due dates and eases the
50
To overcome resource conflicts in the sequential procedure, a post-optimization step performs a re-
scheduling of the schedule provided by the assignment model.
51
I.e. production times are continuous variables.
139
calculation of inventory levels. The objective is to maximize total profit over the planning
horizon by considering approximated inventory costs as well as changeover and process-
ing costs. To solve the proposed MILP, an iterated decomposition approach is described
by splitting the planning model into a master problem52 and a sub-problem.53 Here,
the master problem provides an upper bound for the original problem whereas the sub-
problem provides a lower bound. The iteration is stopped if the difference between both
bounds is sufficiently small. Erdirik-Dogan and Grossmann (2008) extend this model by
considering multiple parallel production units. This model additionally requires to assign
production lots to the production units which complicates the problem considerably. A
similar iterated decomposition approach as used by Erdirik-Dogan and Grossmann (2006)
is presented to solve larger instances of this problem.
In practice, large-scaled chemical companies operate multiple different plants at mul-
tiple production sites. Such a situation requires to model materials transports between
plants/sites and is considered by Bok et al. (2000), Neiro and Pinto (2004), Pinto et al.
(2000), and Shah and Ierapetritou (2011). Bok et al. (2000) present a discrete-time pro-
duction planning and scheduling model with multiple plants. Each plant can be set-up
to a specific production mode once at the beginning of each period. Transports between
sites affect the provider’s and the receiver’s stock levels.54 Raw material transports are
associated with fixed ordering costs. However, further logistical details such as turnover
capacities are not considered. The objective is to maximize the total profit over the plan-
ning horizon. To solve larger problem instances, a two-stage decomposition procedure
is described which partitions the monolithic model into a master problem and a sub-
problem. The master problem determines a production and transport plan by ignoring
changeover calculations. In the sub-problem the production plan is determined given the
transport flows from the solution of the master problem. Based on the solution of the
sub-problem, cuts are added to the master problem in the next iteration.
A related problem for the operational production and distribution planning of refineries
is presented by Pinto et al. (2000). Here, a single-period planning model is formulated
which determines the blending and distribution of refinery products via a network of
pipelines and tanks. Multiple distillation units are assumed to provide a continuous flow
of multiple (raw) materials being transported by multi-product pipelines to tanks. Each
tank is designated to one product type only. Transition processes for switching from one
product to another are modelled.55 The model aims at determining a pumping schedule
maximizing the profit. The resulting MILP is solved with standard solvers. Neiro and
Pinto (2004) extend this work by modelling petroleum supply and production planning
52
The master model determines production quantities and inventories per period.
53
The sub-model determines the precise schedule using the restricted set of products planned by the
master problem.
54
Customer transports are not considered because products are sold ex factory.
55
Each customer market is supplied via a multi-product pipeline from these tanks. The distribution
pipelines can be fed simultaneously from multiple tanks to allow a blending of the (raw) products to
final products.
140
for multiple periods. Transition processes of multi-product pipelines are neglected. The
production is formulated as a generic transformation model that accounts for multiple
production modes. For each unit exists an intended production mode. Deviations from
the intended mode generate additional processing costs. Albeit the model is formulated as
a profit maximization model, the demand is deterministic and no stock-outs are allowed
such that, in fact, the model is a cost minimization problem considering production,
transport, and storage costs.
Shah and Ierapetritou (2011) present an integrated planning model for refinery pro-
duction based on distillation and blending units. Production planning is simplified by
assuming semi-continuous production modes.56 Changeovers of production modes are
penalized in the objective function.57 The proposed MILP is formulated in continuous
time and can be solved with standard solvers for realistically sized problem instances.
The objective is to maximize a profit-like measure considering cost-inducing performance
measures such as the number of setups or the plants’ utilization rates. The weights for
the performance measures have to be interpreted rather as preference measures than as
cost rates. The setting of these weights significantly affects the model’s computational
complexity.
To sum up, the review of existing literature reveals that a lot of effort has been spent in
developing integrated approaches to tackle deficiencies of sequential planning processes.
Prevalent in scientific literature are analytical optimization approaches allowing a high
degree of both portability and conciseness. On the tactical level, however, stochastic
components have to be considered to enable an accurate modelling of real systems. Op-
timization approaches, typically, handle only one source of uncertainty.58 But for an
encompassing model focusing on multiple system components, usually multiple sources
of uncertainty have to be considered. Simulation approaches offer this capability and,
therefore, are prevalent in literature on integrated tactical problems (see Table 4.1).
The literature review also reveals that most simulation-based approaches lack a system-
atic procedure for deriving recommendations of improvements. Instead, most approaches
still bear the stigma of rather descriptive models where a (pre-)selected number of scenar-
ios is evaluated and compared. Potentials for improvement are not revealed directly this
way. However, a systematic procedure for post-analysis of simulation models can lead to
an exploration of the space of possible system configurations which allows deriving rec-
ommendations for improvements on a (more) reliable basis. This topic will be discussed
in section 4.3.4.
A further fact is that most integrated approaches do not consider transport processes
accurately, in particular at the operational level. Most approaches focus on production
planning decisions and handle logistical processes as auxiliary components. Exceptions
56
I.e. for the set of production tasks the product recipes can be varied in certain ranges only
57
Changeovers are also considered for multi-purpose tanks whereby the changeover implies a downgrade
of the interface of mixed products.
58
This is most often the demand, see Peidro et al. (2009) for more detailed information.
141
such as Pinto et al. (2000), Neiro and Pinto (2004), and Shah and Ierapetritou (2011)
consider refinery operations where the logistical system differs from basic chemical indus-
try in the sense that primarily the pipeline management and tank management constitute
critical logistical components. In basic chemical industry, however, also rail transports
constitute a relevant logistical aspect, in particular for inter-site transports. Also other
logistical details such as turnover capacity are often not modelled in detail.
These drawbacks are overcome by the simulation framework proposed in this thesis by
integrating MILPs for operational distribution planning considering transport-mode spe-
cific restrictions in a simulation environment. It aims at integrating both the production
and logistics system in basic chemical industry at the same level of aggregation. As the
proposed framework is designed to support tactical decisions, multiple sources of uncer-
tainty are handled. The characteristics of the proposed framework are briefly summarized
in the last row of Table 4.2. The next section categorizes and discusses typical sources of
uncertainty relevant for integrated supply chain planning in chemical industry.
60
See Park et al. (2006) for a complex inventory management model for a single site.
143
64
See e.g. Vidal and Goetschalckx (2001) or Oh and Karimi (2004).
65
See e.g. Van der Vorst and Beulens (2002).
66
For a similar categorization see Klibi et al. (2010).
67
For related examples see also Applequist et al. (2000).
68
The so-called sample space.
69
The probability of an outcome may depend on other variables or decisions made in advance, see e.g. Law
145
internal external
technological
produc-
finan- person- materi- infra-
tion demand supply 3PLs
cial nel ally structure
system
work- demand transport
quan- credit stock pumping yield avail-
force varia- capaci-
tity line levels capacity rates ability
size tion ties
qua- quali- reli- off-spec product product reli-
interests decay
lity fication ability rate specs specs ability
evapo- pro-
due pumping due delivery transport
time speed ration cessing
date rates dates dates times
rate times
of outcomes and the process’ basic structure are known, information about the underlying
probabilities can be deduced.
For example, the output rate of a simple SISO reactor depends on various conditions.
To model the transformation of input to output, knowledge about the chemical reaction
and the chemical reactor can be used. E.g., a linear model might be used to describe this
relationship properly on an aggregated level (e.g. the hourly production rate). Neglecting
minor influences70 leads to a simplification of the process model. Additionally, measure-
ment errors may hinder a perfect description of the process and lead to uncertainty in
the observed process measures.71 This uncertainty is expressed e.g. by a (normal) error
process. The resulting linear regression model can be verified using historical records of
the process. Often historical records allow analysts to deduce a proper stochastic model
of such a process. For more complex production processes more sophisticated stochastic
models (as described in section 2.3) can be necessary.
However, not all relevant stochastic processes are typically known at such a level of
detail. For external sources of uncertainty, typically, the underlying processes are not
known (entirely) or not describable in detail.72 Even the set of outcomes might not be
perfectly known. Modelling such processes is often based on historical records and rough
assumptions about the modelled process. E.g. the total customer demand faced by a
company or entire SC builds a conglomerate of a manifold of individual orders placed in
a given interval of time. However, neither models for each individual ordering process can
be derived nor an agglomerative model describing the ordering processes of all potential
customers could be handled. Hence, to model and estimate demands, one has to rely
on historical records about the total customer orders in history. A model to predict the
total demand to be faced in a period relies on indicators (such as GDP growth rate)
(2007).
70
E.g. the composition of input flows or precise temperature and pressure within the reactor.
71
For a more detailed discussion and definition about uncertainty in the context of operations manage-
ment see Zimmermann (2000).
72
According to Zimmermann (2000) this refers to a lack of information causing the uncertainty.
146
and on common assumptions about the customers’ ordering processes.73 However, there
is no reliable information about the uncertainty inherent in the estimation procedure
since the individual ordering processes cannot be modelled adequately. Hence, no reliable
information about the degree of uncertainty is available.74 This phenomenon often occurs
in long-term planning problems, e.g. when an SC configuration is to be found which fits
best to a fuzzy set of scenarios.75
For the purpose of modelling SC operations on a detailed level, precise knowledge about
the stochastic processes to be modelled is required. For an existing SC the status quo
of operations is known and documented in most cases. Hence, internal processes can
be described and information about the processes’ stochasticity is (at least) partly avail-
able. This allows the proper modelling of internal uncertainty. E.g. chemical production
processes can be modelled using the methods described in section 2.3 and illustrated in
example 3. However, such a process model is insufficient to model the production pro-
cess as a whole since only the output streams can be described with respect to the input
stream. To set up a proper model of the entire production process the stochasticity of
the input stream(s) has to be modelled as well. Typically, there exists a finite set of
operation states for a plant which are overlaid by stochastic noise. To model the states
of a plant, the Markov methodology introduced in subsection 3.2.2 can be used, enriched
by a stochastic model for the noise process. The following example extends the analysis
of the de-alkylation plant introduced in example 3 by analysing and modelling the inflow
process.
140
100
120
median inflow rate per day
80
100
frequency
80
60
60
40
40
20
20
0
0
Figure 4.3: Daily median inflow rates of aromatic hydrocarbons for a de-alkylation plant
sons of these perturbations reveals that they are caused by technical re-adjustments due
to preceding repair and maintenance activities. Hence, two general states of operation
are considered as relevant, namely: full capacity production or a complete shutdown. Let
O = {0, 141} denote the set of these two states.
To model the state of the plant, a discrete Markov process is used. To calculate the
transition matrix Q of a discrete Markov process, the transition probabilities between both
states have to be estimated. All transitions of the recorded inflow data is used. The time
series of plant states ωt are calculated by
⎧
⎪
⎪ 141 ft > 0
ωt = ⎨ (4.1)
⎪
⎪ ft = 0
⎩ 0
where ft is the median inflow rate at day t depicted in Figure 4.3. The estimated transition
matrix Q̂ is then given by
0 141
0 ⎛ 0.875 0.125 ⎞
Q̂ = .
141 ⎝ 0.002 0.998 ⎠
The corresponding steady state vector is π = (0.087, 0.913). In other words, in the long
run, on 8.7% of days, the de-alkylation plant suffers from an (unintended) break-down.
The probability to face a break-down at a day which started in normal operation is 0.2%.
When the plant operates at full capacity, there are still variations in the realized inflow
rates. To model these fluctuations, the time series of average hourly inflow rates is anal-
ysed in an appropriate interval where the plant operated at full capacity. As in Example
148
144
average inflow rate per hour
143
142
141
140
Figure 4.4: Hourly average inflow rate of aromatic hydrocarbons (including confidence
intervals at a confidence level of α = 0.1%)
In general, the inflow rates oscillate slightly around an average of 141 units. However,
some outliers occur which are probably caused by measurement errors rather than real vari-
ations of the flow rate. To account for these observations, mean and standard deviation
are robustly estimated by calculating median and median absolute deviation from median
(MAD). This yields the following robust estimates: μ̂median = 141 and σ̂M AD = 0.18. As-
suming a normal distribution of the inflow rate, both estimates are used to calculate a
confidence interval with a confidence level of α = 0.1%. The corresponding quantiles are
displayed in Figure 4.4 by the dashed lines. Observations exceeding these intervals are
handled as outliers.
Having removed the outlying observations, the implied normal distribution is checked by
Shapiro-Wilk’s test. Figure 4.5a shows the QQ-plot and Figure 4.5b the autocorrelation
values for the trimmed time series of inflow rates.
The trimmed time series shows a good fit to the normal model and no hints for auto-
correlation.
149
141.4
141.2
Sample quantiles
141.0
140.8
140.6
140.4
3 2 1 0 1 2 3
Theoretical quantiles
(a) QQ-Plot of trimmed inflow rates of aromatic hydrocarbons
1.0
0.8
0.6
ACF
0.4
0.2
0.0
0 5 10 15 20 25
Lag
(b) Autocorrelations of trimmed inflow rates of aromatic hydrocar-
bons
Figure 4.5: Diagnostic plots of trimmed inflow rates of aromatic hydrocarbons for a de-
alkylation plant
150
simulation model
perform
experiments
meta-model
decision support
of the simulation outcomes. Keeping run times in acceptable ranges and deriving concise
information about the system performance at the same time, is a challenging task. Var-
ious approaches are provided in the literature to tackle this problem (often labelled as
simulation optimization).78
In the case of single-objective problems, the aim is to find the optimal configuration of
the system under study. Once found, this configuration is the sole recommendation for
the management. The objective typically represents an expected performance measure.
However, since stochastic events may influence the performance measure, it is advisable
to incorporate measures of variability (like the variance of a performance measure) as a
further decision criterion. This immediately leads to a multi-objective problem.79
In the multi-objective case, a set of Pareto-optimal configurations has to be found (so-
called Pareto set). In many cases not all possible Pareto-optimal configurations can be
determined. Instead, a subset is returned by a multivariate simulation optimization ap-
proach. Based on this subset, analytical models can be fitted to this data describing the
relationship between the control parameters and performance measures. These analytical
models are often lower-polynomial models and are called meta-models, response surface
models, or surrogate models.80 In other words, meta-models are mathematical functions
representing interdependencies of variables in a (simulated) system. Meta-models can be
used to derive Pareto-optimal configurations without the need to simulate these configu-
rations. Meta-models can help to find the most suitable system configuration in an easy
way (e.g. by weighting input and performance measures with costs and/or income rates).
In the remainder of this section the outlined planning process is described step-by-step.
Along these theoretical deductions, an exemplary simulation model of a chemical SC is
successively described, developed, and analysed in the example blocks.
78
See e.g. Law (2007), Banks et al. (2005) or Kleijnen (2007).
79
In this context Pareto-optimal configurations are often called stochastically dominating, see Klibi et al.
(2010).
80
See e.g. Box and Draper (2007) or Kleijnen (2007).
81
Following Law (2007) and Banks et al. (2005).
152
modelling
programming
conceptual simulation
model verification
model
the responsibility of SC managers. The perceived deficiencies define the aims and scope of
the project as well as the system components to be improved. Modellers have to propose
a concept for identifying the reasons why the perceived deficiencies occur. In a first
phase this concept description is called a conceptual model or assumptions document.82
In this phase of the project all stakeholders commonly discuss and argue about the scope
and capabilities of the model to be developed. From the modeller’s point of view, the
conceptual model is a simplification of the real-world system.83 Simplification can be
achieved by reducing complexity or scope of the model:84
• complexity reduction
– subsume system components
– drop randomness
– approximate dependencies
• scope reduction
– drop system components
– drop variables
– restrict domains of variables.
Together, managers and modellers have to agree upon simplifications and implicit assump-
tions made, finally leading to the definition of a suitable conceptual model. As Figure 4.7
suggests, the development of simulation models is an iterative process, where the appro-
priateness of model specifications is checked in a structured way. Errors in implementing
the conceptual model in a computer program are checked by verification.
82
See Law (2007) or Banks et al. (2005) for the latter term.
83
See Robinson (2006).
84
See Robinson (2006).
153
In the validation phase the behaviour of the simulation model is compared with the real
system’s behaviour. If errors are revealed, the modelling and programming steps have to
be re-started to ensure the required level of accuracy.
All sub-processes of the development process, beginning with the perception of deficien-
cies, require the analysis of process data. The data characterizes the system’s performance
and behaviour. This is necessary to parametrize the simulation model and build a credible
basis for comparing the model’s behaviour and reality. A common database maintained
by both modellers and managers allows linking necessary and available information. A
lack of data may lead to uncertainty, more complexity, and a loss in accuracy. Uncertainty
is reflected e.g. by stochasticity at variables which in turn may lead to a larger number
of simulation experiments and more volatile performance measures.85 On the other hand,
the information collected in the project database needs to be relevant for the project.
Otherwise, a flood of data to be evaluated and analysed hinders an efficient development
process as it is time and resource consuming.
The analysis of available and relevant data typically requires a set of statistical methods.
They are applied to extract deeper knowledge about the processes to be modelled. The
set of methods is vast and the choice of methods depends on the needs of the specific
project. Time series methodology is one prominent branch of methods.
The following example describes the conceptual model of an artificial chemical SC
analysed in this section.
85
See Zimmermann (2000).
86
This is partly due to privacy protection reasons and partly because the proposed example does not
match the real SC entirely.
87
Additionally, natural gas is admixed if the local fuel production is insufficient. In general, cracking is
154
site 1 PE C1
H2
C2
Benz. Styr. PS C2
C5+
C9+
H2
C3
C1 Cum. C3
C4
Naphtha
C4
S1 rail transports
Naphtha site 2 PE C5
H2
C2
Benz. Styr. C6
C5+
C9+ H2
C3
C1 PP C7
C4
Buta. SBR C8
Raf.
supplier storage customer production plant
Imbalances are compensated either by heating with H2 or by external supply from producers
of industrial gases nearby both sites.88
Steam crackers produce a fraction of C4 hydrocarbons which is a mixture of differently
structured chemicals all containing four carbon atoms.89 In this mixture, 1,3-Butadiene
is the largest sub-fraction with the highest applicability in downstream processes. There-
fore, this sub-fraction is extracted by rectification.90 The remaining sub-fractions of C4
hydrocarbons (so-called raffinate) is sold to the market ex factory.
At both sites imbalances of the processed intermediate chemicals occur when all plants
produce at full capacity. Unbalanced intermediate chemicals are indicated by a coloured
storage symbol when inter-site transports are possible. Correspondingly coloured arrows
between both sites indicate the direction of the product flow. Table 4.4 quantifies these
imbalances and provides an overview on the chemicals’ associated storage capacities.
A net demand (indicated by negative values in the balance columns in Table 4.4) of
intermediate chemicals has to be satisfied by product flows between both sites or from
a highly endothermic reaction which in turn leads to a net demand of natural gas.
88
Note that this is a typical situation since most chemical production sites have demand for industrial
gases not only for chemical reactions but also e.g. for cooling (N2 ).
89
See Sun and Wristers (2006).
90
Since all C4 -sub-fractions possess similar boiling points, extractive distillation has to be applied, i.e. a
solvent chemical is used to drive the boiling points apart. See Sun and Wristers (2006).
155
Table 4.4: Material balances (in t/hour) and storage capacities (in t) per site and chemical
91
In the case of of plants with multiple inputs, for each input stream its variation is calculated based on
its current state.
92
This is because reciprocal imbalances exist at both sites for all these chemicals. E.g. Ethylene is short
at site 1 but surplus at site 2. Hence, transports of Ethylene from site 2 to site 1 are reasonable.
156
time. After an expected travelling time of approximately 24 hours the train arrives at the
arrival site’s shunting yard where the attached RTCs are decoupled and further processed
(i.e. loaded or unloaded). The train’s travelling time is a normally distributed random
variable with N(24, 1). Loaded RTCs become available for transport the day after un-
loading. During this time, turnover processes as well as security checks and shunting
operations take place.
Unloading and loading times (in hours) are normally distributed random variables with
N(2, 0.25). The total turnover capacity depends on the number of transfer arms available
and the number of operating hours. Turnover operations take place 10 hours per day from
Monday to Saturday. Due to strikes, technical failures, etc. operating hours may reduce
to 5 or 0 hours. A discrete Markov model is used to model these disturbances constituted
by the following transition matrix:
10 5 0
10 ⎛ 0.99 0.005 0.005 ⎞
⎜ ⎟
Qturn−over = 5 ⎜⎜ 0.15 0.80 0.05 ⎟⎟
.
0 ⎝ 0.20 0.05 0.75 ⎠
Since both sites are geographically separated and do not share a common workforce, the
Markov model is applied for both sites independently.
The number of transfer arms and corresponding turnover capacities can be found in
Table A.12 in the appendix.93
To dynamically generate plans for train departures and RTC flows on the operational
level, the MC-RTP is used. The interval of time for discretization is a 24-hour rhythm.
The target stocks for the corresponding chemicals are calculated such that the maximal
demand of a chemical can be supplied for approximately 10 days from stock. The target
and initial stock levels as well as the number of initially available empty RTCs can be
found in Table A.12. Cost rates for the MC-RTP instances are shown in Table A.13 in
the appendix. Note that the costs for additional storage in RTCs are given as cAdd = 1.94
For Styrene, C4 hydrocarbons, Naphtha, and the final chemicals the network is imbal-
anced which requires exports and imports to and from external partners. Naphtha is sup-
plied via pipeline which connects the sites with the exclusive supplier (a refinery nearby).
The supply is organized in batches injected by the supplier when an order is placed by a
production site.
Batch injection and transport is associated with costs for pipeline operation (denoted by
cbatch ). Stock holding at the sites is associated with a stock holding cost rate chold . The
fixed pipeline transport capacity ρ is 7,500 tons per day (which exceeds both consumption
rates ωi ), optimal batch sizes q opt and order intervals topt can be determined by solving the
93
The turnover capacities are symmetric, i.e. loading capacity at the sending site equals unloading ca-
pacity at the receiving site. In an asymmetric setting both capacities do not match.
94
This allows a virtually cost-neutral storage extension.
157
To calculate the optimal order quantities qiopt and corresponding order intervals topt
i requires
information about holding and batch injection costs, chold and cbatch .
Holding costs chold typically consist of the costs for maintaining storage facilities and
committed capital. Maintenance costs of storage facilities mainly encompass labour costs
for the maintenance staff and the facilities’ depreciations. Both are independent from
the stock levels and, hence, do not affect the stock level decision. In contrast, costs for
committed capital are typically assumed to grow linearly with the stock level as they reflect
the opportunity costs for investments which cannot be realized. With a market price for
Naphtha of about 700 e per ton and a return on assets of about 5% in basic chemical
industry, the storage cost rate can be approximated by chold = 35 e per ton and year.96
Batch injection costs are comparatively hard to assess as they depend on the techni-
cal specification of the pipeline. Pipeline operations primarily induce costs for energy
consumption. A pipeline transport generates energy costs which are proportional to the
transport volume. For modern basic chemical pipelines an energy consumption of about
42 kWh per ton and kilometre is estimated.97 Depending on the energy supply contract,
the costs of a batch transport can be calculated. However, with a contract charging energy
costs proportional to the energy consumption, batch injection and transport costs grow
linear with increasing batch size. Hence, the fixed cost part for pipeline operations cbatch
only consists of operational costs for preparing pipeline operation which are comparatively
small. Hence, it is concluded that qiopt → ωi which implies a continuous operation of the
pipeline at the level of the consumption rate. This result is reasonable as long as no con-
siderable fixed costs for providing a Naphtha batch occur.98 In this situation, both sites are
supplied continuously, i.e. the pipelines’ feed rate is 5,000 tons a day with both crackers
operating at full capacity. The flow rate is reduced accordingly when one or both of the
95
Note that this implies that both pipelines can be operated independently, see e.g. Günther and Tem-
pelmeier (2004).
96
Prices for Naphtha are partially transparent on the spot markets and can be found e.g. in Kellermann
(2012). However, most quantities are traded based on long-term contracts. Hence, the spot mar-
ket prices are potentially slightly over-estimated. The return on assets is a key figure reported in
companies’ economic reports, see e.g. BASF (2012) for one specific example or Fortune (2006) for an
industry-wide figure.
97
See van Essen et al. (2003).
98
Note that typically refineries are operating continuously where Naphtha results as a direct by-product.
Hence, production-specific set-up costs due to a variation of the production process are not to be
expected.
158
crackers are shut-down. The safety stocks to be held at both sites are defined such that for
5 days the crackers can be supplied from stock. I.e. safety stock levels are set to 15, 000
and 10, 000 tons for site 1 and 2, respectively. Details on the supply system will be studied
in example 12.
Final chemicals (such as Polystyrene, Polyethylene, etc.) are sold to customers ex
factory (i.e. customer transports are not considered). Styrene and C4 hydrocarbons are
transported via RTCs. To plan these transports the MC-RTP instance is extended by an
artificial sink and an artificial source representing customers and suppliers of Styrene and
C4 hydrocarbons, respectively. These fictional nodes are always able to absorb or provide
a required quantity of chemicals. They represent the spot market for both products where
orders are placed and a certain company accepts (at a certain price). Transports from/to
this source/sink represent transports to ex ante unknown partners and, hence, transport
times are stochastic. Here, a normally distributed transport time is assumed with N(5, 1)
(given in days).
Logistical disturbances are incorporated reflecting RTC defects and variations in the
customers’ pick-up quantities and frequencies (see Table A.14 in the appendix). RTC
defects occur with probability p = 0.025 per day and RTC. I.e. the number of defect RTCs
per day is binomial distributed with sritsf ail
∼ Bin (srits , p) where srits is the number of
RTCs of chemical s checked at site i in period t.99 The time for repair is given in days
and normally distributed with N (7, 1).
Auxiliary processes may fail e.g. steam or electric power supply breaks down which
causes plant breakdowns or a site-wide breakdown. For simplicity only two events are
modelled here: First, a breakdown of the steam network which forces all plants with energy-
consuming reactions to shut-down. This set includes the cracker, Butadiene extraction,
Benzene production (hydrogenation plant) as well as Cumene and Styrene production.100
To model such an event, a Markov approach is used with the following transition matrix
where state ω = 1 indicates normal operations and ω = 0 indicates a breakdown:
1 0
1 ⎛ 0.995 0.005 ⎞
Qsteam = .
0 ⎝ 0.35 0.65 ⎠
The second auxiliary process affecting the production network is a breakdown of the hy-
drogen network which immediately affects all plants consuming or producing hydrogen.
This includes the cracker, Benzene production, and Styrene production. The following
99
Note that any RTC is inspected before it leaves a site or is loaded. Hence, this encompasses all RTCs
E
waiting for further processing (srits ).
100
The alkylation reaction taking place in Cumene and Styrene production is exothermic (see Vora et al.
(2006)) which induces a production of steam. The same holds for the cracker where the quenching
process produces steam. Hence, these plants are connected to the site’s steam network and affected
when steam distribution cannot be managed.
159
1 0
1 ⎛ 0.99 0.01 ⎞
Qhydro = .
0 ⎝ 0.50 0.50 ⎠
Note that both auxiliary processes are assumed to be highly reliable and failures can be
bypassed or repaired in comparatively short time.
In the next step, all components and (statistical) models described in this example are
implemented in an R program. The implementation is described in the next example.
101
See Kleijnen (2005).
160
• production facilities,
Table 4.5: Examples for processors according to processor type and processing attributes
Processors may merge, split, or route ingoing flows. For production processors this
corresponds to the general classification of chemical production processes (see Table 2.1).
161
For logistical processors these attributes correspond to the technical details of the mod-
elled equipment. Splitting storages can be used to supply multiple consuming processors
at the same time. Similarly, splitting turnover processors can handle multiple carriers at
the same time or may feed multiple storages at the same time. Transport processors that
are able to split flows simultaneously are e.g. one-to-many pipeline. Analogous examples
are given for merging and routing processors.
Depending on the real system’s specifications, a crucial attribute of processors is their
handling of time deciding whether ingoing flows are delayed or not. An example of
delaying production processors are batch processes where typically some time elapses until
the chemical transformation process is finished. Storage and turnover processors are also
delaying whereby in the former case the delay is an intended feature and flexible. In the
latter case, the delay occurs due to the processing times for turnover preparation (e.g. for
RTC shunting and connection to the transfer system). The time delay for transport
processors accounts for the transport time which is typical for e.g. rail and ship transports.
In contrast, pipeline transport can be non-delayed if the pipeline is completely filled once
and continuously operated under constant conditions.
Another characteristic is whether a processor conserves the inflowing quantity or not.
If non-usable or invaluable materials are produced during a transformation process, these
can be dropped from the simulation model for the sake of brevity such that a total loss
of material is modelled. Examples of such losses are off-spec materials102 produced after
a mode changeover in a multi-purpose plant and interfaces in multi-product pipeline
operation.
The kind of modelling of auxiliary processes depends on their structure and importance
for the entire system. Often auxiliary processes are modelled on an aggregated level as
(global) environmental variables affecting the attributes (such as capacities) of processors.
Such processes are often not modelled explicitly by means of processors or flows but rather
as (environmental) events which are ruled by stochastic processes.
Similar to SC planning problems, the elements of simulation models (i.e. flows, pro-
cessors, and environmental variables) can be structured according to the level of detail
involved. Based on the basic structure of a chemical SC as depicted in Figure 4.10, three
aggregation levels are distinguished:
• plant level
• intra-site level
• inter-site level.
Table 4.6 shows a brief overview on the related model components and aggregation levels.
At the plant level one or more production processors form a material transformation
process. Chemical production plants can be modelled as processors with technologically
102
See page 141 for the definition.
162
logistic
production auxiliary
storage turnover transport
plant ✓ — — — (✓)
intra-site — ✓ ✓ ✓ ✓
inter-site — (✓) — (✓) —
fixed rules of operation (e.g. a fixed material flow inside the plant). Here, static models
are used to represent the transformation process. Considered stochastic elements are
e.g. yield rates, general production mode, and processing times. Operating rules at the
plant level are e.g. priority rules for material processing (such as first-in-first-serve).
The intra-site level is characterized by a set of multiple plants which are interconnected
by logistical processors (such as pipelines and tanks). Logistical processors enable inter-
mediate storage and a connection to external sources of materials. Examples for local
processor rules are inventory policies for each storage processor. At this level, a coordina-
tion of production processors addresses e.g. production planning and scheduling problems
for multi-product batch processes. In this case, the central control of production proces-
sors tackles an operational planning problem. A similar intra-site coordination problem
occurs in the management of local stocks within a multi-product pipeline system.103 Such
coordination problems can either be solved by applying heuristics locally or by an em-
bedded analytical optimization model.104 Considered stochastic influences in modelling
logistical processors reflect e.g. variations in their capacity due to technical failures or the
processing times of handled objects.
Coordination problems also occur on the inter-site level where multiple integrated pro-
duction sites are interconnected by (material) flows. At the inter-site level, only logistical
processors are added to the model if any at all.105 Here, the main focus is on the man-
agement of inter-site transports by utilizing the logistical facilities located at the sites.
To utilize these facilities, transport carriers have to be modelled, i.e. trucks, rail cars, or
ships. The flow of these entities forms the capacity for inter-site material flows. Hence,
at the inter-site level primarily flows (e.g. of materials, transport carriers, or information)
have to be planned for which operational distribution planning and routing models can
be used. Stochastic influences often affect the transport times but also the quality of
materials which may decay during transport.
If the modelled real system relies on discrete objects, a discrete-event simulation is
the natural choice for simulation. However, in chemical industry production processes
are continuously operated. Hence, material flows are usually continuous with variable
103
See Neiro and Pinto (2004).
104
See Garcia-Flores and Wang (2002), Jung et al. (2004) or Jung et al. (2008) for examples of the latter
case.
105
An example are inter-site pipelines.
163
flow rates. In contrast, the associated logistical sub-systems (e.g. for rail and ship trans-
ports) rely on discrete entities which requires discrete-event simulation. This implies that
continuous processes have to be simulated for discrete time intervals. Hence, the ideal
simulation model is a continuous time discrete-event model which integrates continuous
chemical process models. If operational planning models are included in the simulation
model, this forces a continuous time formulation of these optimization models. However,
typical operational planning models are complex and time continuous formulations are
often not adequate to solve such problems efficiently. Often a time-discrete formulation
is preferred to keep the problems solvable for instances of practical size.106 Therefore,
time-discrete, discrete-event formulations are often appropriate for hybrid simulation ap-
proaches.107 Fortunately, continuous chemical production processes can be approximated
by discrete time series models quite accurately as long as the time increment is chosen to
be not too large, see chapter 2.
The following example describes the implementation of the conceptual model described
in example 9.
106
See e.g. Pinto et al. (2000).
107
See Garcia-Flores and Wang (2002), Jung et al. (2004) or Jung et al. (2008) for examples.
108
Site 2 and information flows are omitted for the sake of brevity.
109
Details are described in example 9 and Table A.6-Table A.11. Flows from sources and sinks (repre-
senting suppliers and customers) are generated as described in example 9 and Table A.14.
110
See IBM (2010).
111
For the configurations used in the simulation study, all MC-RTP instances are solved to optimality in
less than two seconds.
112
To determine the re-planning interval, simulation experiments with varying re-planning intervals are
conducted. In tendency, it holds that the shorter the replanning interval is, the better the system
164
C2 PE
Benz. Styr. PS
C5+
C3
Cum.
C4
storage processor
production processor turnover processor
storage processor for empty RTCs
fictional sink/source transport processor
storage processor for loaded RTCs
Figure 4.9: Flow chart of a part of the exemplary chemical SC using the notation of
Table 4.5
ture of the implemented simulation model is displayed in Figure 4.10. Numbers annotated
to the blocks refer to the lines in the pseudo-code provided in Table 4.7.
In each iteration (i.e. for each day), it is initially checked whether a new transport
plan is to be determined. If this is the case, initial data is gathered reflecting the current
state of the SC. This comprises the stocks of chemicals and RTCs at the sites (loaded or
empty) as well as the currently flowing chemicals and RTCs. Additionally, the expected
production rates are estimated by linearly extrapolating the last week’s (hourly) production
rates. The capacities of the storage facilities as well as expected values for turnover and
transport times are constants and, hence, fixed parameters for each MC-RTP instance.
The turnover and train capacities follow regular train and workforce schedules which are
assumed to be deterministic and, hence, known for all instances. The cost rates ruling the
MC-RTP are fixed for all instances (see Table A.13).
After solving an MC-RTP instance, the resulting optimal flows of materials and RTCs
are passed to the simulation model. Whether these flows can be realized depends on the
availability of resources which are affected by stochastic elements. Hence, scheduled flows
can be realized only if sufficient
• RTCs,
• turnover capacities,
performance and the longer the computation time is. An interval length of three days yields a
reasonable compromise between run time and solution quality.
165
set-up matrices
2-10
initialize inidicators
t=1
12
is t re-plan- yes generate
13
ning period? MC-RTP-instance
no
solve 14
instance
no
end of
horizon?
yes
evaluate
performance
are available. Breakdowns of turnover capacities are ruled by a Markov approach and
immediately affect all flows. A loss of empty RTCs due to unexpected maintenance activ-
ities can affect scheduled flows if the remaining number of empty RTCs is insufficient to
transport the scheduled flow. Moreover, dispatched trains may arrive later than estimated
such that scheduled turnover processes are delayed. A breakdown of production plants may
induce e.g. stock-outs such that planned chemical flows cannot be realized.
Anyway, when a scheduled flow cannot be realized as intended, it is postponed to the
next period. Hence, stochastic influences typically lead to a postponement of transport
activities. In the next re-planning period all flow schedules are re-set to the new optimal
solution such that formerly postponed flows are integrated in the new schedule. Table 4.7
shows the detailed pseudo-code of the implemented simulation model.
Configuring the SC can be done by a set of parameters of the simulation model reflecting
the adaptation of capacities (e.g. RTC fleet sizes, inventory, or turnover capacities) and
of operational rules (e.g. target stocks and cost rates for the MC-RTP). Note that the cost
rates ruling the MC-RTP instances reflect the operational costs for transports, turnover,
166
and stock holding. They can be seen as policy parameters as they control the level of
preference for stock holding or transportation. In particular, the setting of inventory cost
rates for over- and undershooting determines the smoothness of stocks and can be seen as
an alternative for the variation of target stock levels.
To evaluate the performance of a simulation run, the flow and stock matrices are re-
turned by the model. These matrices can be used to calculate a desired number of perfor-
mance measures (e.g. the number of dispatched trains, average stock levels, etc.). These
depend on the purpose of the study. It has to be noted that each simulation run is a stochas-
tic experiment. Hence, multiple replications are necessary to evaluate the performance of
a certain SC configuration properly.
113
See Law (2007).
114
See Law (2007).
115
See Sargent (2005).
116
See Sargent (2005).
117
See Balci (1997).
168
model with a reference system (e.g. the real system or the conceptual model). They can
be categorized according to the type of test and the type of criterion used. Basically,
the model structure can be evaluated by testing the internal logic of the model. I.e. it
is tested whether the relations ruling the model match the assumed relations deduced
from the reference system. On the other hand, the output of a model can be checked for
similarity with the reference system.
The criteria used for V&V can be qualitative or quantitative. Qualitative criteria
involve a subjective assessment on the logic or similarity of the tested model. In contrast,
quantitative criteria measure the similarity or consistency by a numerical comparison
of a reference value from the reference system with a (calculated) test value obtained
from the tested model.118 Table 4.8 shows examples of V&V techniques according to the
categorization criteria discussed.119
type of test
logical similarity
These techniques are more or less applicable for the different V&V activities. For
verifying the implemented simulation model, software engineering techniques are most
widely used which refer to the techniques displayed in the upper left cell in Table 4.8.
Cause-effect-diagrams visualize relations between events and system outcomes. They
can be compared with commonly accepted assumptions about relations between events
and system outcomes to detect unintended model behaviour. Similarly, animation and
trace analysis often help modellers and managers to check the plausibility of the model’s
behaviour over time. However, not all programming mistakes uncover themselves in an
animation or in the analysed traces. Moreover, rare events are hard to detect since
animation is applied for rather short time frames and trace analysis is applied for a
118
Similarly, both extremes can be labelled with subjective and objective. See Rabe et al. (2008) for a
continuous-scale classification on the degree of subjectivity.
119
For an extensive list of V&V techniques see Balci (1998).
169
120
See Sargent (2005) and Rabe et al. (2008) for more information.
121
I.e. conditions which are barely realistic (e.g. when parameters are set to values which are judged as
just tolerable in reality), see Sargent (2000).
122
See Kleijnen and Sargent (2000) and Kleijnen (1995).
123
See e.g. Sargent (2005); Kleijnen (1995); Rabe et al. (2008).
124
See e.g. Rabe et al. (2008).
125
Therefore, real historical records and simulated records are presented and the expert/manager is asked
to identify the simulated record.
126
See Schruben (1980).
127
See Sargent (2000).
170
process than by a specific technique.128 Once the model development has passed all V&V
steps successfully, the simulation model is assumed to accurately reflect the real system’s
behaviour (w.r.t. the scope of the project). In the next steps, simulation experiments are
planned.
128
See Rabe et al. (2008).
129
Based on Hinkelmann and Kempthorne (2005a).
130
See e.g. Law (2007).
171
statistical model
Experimental design
meta-modelling
decision support
131
Note that in structural equation models measurable indicator variables are grouped into factors. These
factors are unobservable/latent and influence latent/unobservable response variables. Such models
are typically designed and estimated to confirm a theory about the construction and relations of the
latent factors, see Pearl (2000).
132
I.e. the existence of such a variable is assumed, but it cannot be measured precisely e.g. intelligence or
trustworthiness.
133
See e.g. Borsboom et al. (2003) for an example and definition.
134
In principle, also latent variables might be considered in simulation studies. However, in a simulation
study the subject studied is explicitly modelled in detail and not as a black box. Hence, the intro-
duction of latent variables in a simulation model yields typically no additional insights in the general
relation of its elements.
172
discrete)
response
ordinal
latent (dependent) uncontrollable
nominal nuisance
simulating a business system, response variables are usually measures of the system’s per-
formance variables. These variables are assumed to contribute to the overall success of
the business system. The contributions of performance variables to the overall success
are often operationalized as costs or revenue rates.
A prominent measure of overall success is the profit obtained with a given set of resources
in a given time span. Basically, the profit depends on costs, revenues and assets employed,
whereby those depend on system-internal and -external influences. In SC simulation the
aim is to investigate the behaviour of the system and explain the relations between its
components. Here, the focus is on studying internal relations under external disturbances.
Let x ∈ Rm denote the vector of controllable independent variables and y ∈ Rn denote the
vector of dependent response variables defined in the model building phase. Furthermore,
e ∈ Ro denotes the vector of uncontrollable variables which are not considered in detail
in the model but are captured as environmental effects. The function H(⋅) describes the
real relation between dependent and independent as well as environmental variables
y = H(x∣e). (4.4)
Typically, the function H(⋅) is unknown and reflects the behaviour of the real system.
Since the controllable variables x are often strategic decisions, experiments in reality are
often too expensive. Hence, simulation models are developed mimicking the real system’s
behaviour. Formally, the simulation model can be seen as a function Ĥ(⋅) such that
ŷ = Ĥ(x∣e). (4.5)
All these components influence the profit u of the real system, i.e.
where U (⋅) is a function describing the relation between influencing variables and obtained
profit u. Typically, U (⋅) is unknown in all details, but at least the direction of influences
of the variables are known.135 As the vector of controllable variables x reflects the set of
options to take, it is sought for an optimal configuration xopt that either maximizes U (⋅)
or, if the focus is purely on internal processes, minimizes the associated total costs where
C(⋅) denotes the cost function of the system under study. I.e.
xopt = arg max U (x, H(x)∣e) or xopt = arg min C (x, H(x)∣e) . (4.7)
x x
x̂opt = arg max U (x, Ĥ(x)∣e) or x̂opt = arg min C (x, Ĥ(x)∣e) . (4.8)
x x
135
I.e. the signs of the partial derivatives of w.r.t. x and y are known at least in some ranges.
136
In case of a linear relation, only one parameter (the cost rate) is required. But also more complex
relations are imaginable.
137
See Sürie and Wagner (2008).
138
See Kaplan et al. (1992) for the seminal work on the balanced scorecard and Bhagwat and Sharma
(2007) for performance measurement systems in SCM based on the balanced scorecard approach.
139
See Sürie and Wagner (2008) or Gunasekaran et al. (2004, 2001) and Agami et al. (2012) for an overview
of performance measurement systems.
174
SCOR model as well as their attributes.140 Table 4.10 shows a classification of performance
measures according to aggregation level and type of measure.
amortization
revenue/
cash flow total profit corporate value
value-based
cost-to-income return on
relational cost-to-asset ratio
ratio investment
cash flow cycle product
time order lead times
non-financial
Depending on the scope of the simulation study, performance measures should be chosen
w.r.t. to the aggregation level as well as managerial impact. Ideally, an SC evaluation
system containing an already existing system of performance measures can be adapted
to the simulation environment.141 However, this procedure bears the risk that too many
criteria have to be evaluated which increases the computational effort for evaluating SC
configurations. Moreover, traditional performance measurement systems aim at reflecting
the current situation of an SC whereas simulation aims at finding a future configuration.142
Hence, the choice of performance measures to be evaluated in the simulation model should
account for cost and profit drivers instead of a direct evaluation of financial measures
since they typically imply assumptions about environmental financial drivers such as raw
material prices or interest rates. This information, however, is uncertain and complicates
the system’s/model’s analysis. Furthermore, such measures distract from the underlying
processes to be optimized. In general, a configuration of the SC is pursued that works well
under almost every environmental state. This ability is often labelled as flexibility in SC
performance measurement systems.143 To operationalize this dimension in SC planning,
a robust configuration is sought.144
Robustness is often an important feature, in particular when an objective’s variation
depends on the configuration.145 Robustness can be measured in various ways, e.g. by con-
140
For the latter case see e.g. Beamon (1998).
141
See Kleijnen and Smits (2003).
142
See Kleijnen and Smits (2003).
143
For a discussion on flexibility see Gunasekaran et al. (2001).
144
See Kleijnen and Smits (2003).
145
I.e. a configuration x not only influences the mean level of responses H(x) but also their variance.
175
sidering the inverse coefficient of variation instead of the pure expected performance146 or
by considering both expectation and variance of a measure as independent objectives.147
Another possibility to invoke robustness is to optimize a certain quantile instead of the
expected performance. This has the advantage that also measures with a skewed distri-
bution can be handled.148
From the deductions above it can be concluded that simulation studies on SC con-
figuration put focus on multiple objectives/performance measures which depend on the
configuration x. The configuration and its performance affect the total cost and/or total
profit. The combination of a configuration and its performance z = (x, y) is called a
constellation. In simulation studies on SC configuration typically a set of efficient con-
stellations is sought.149 Per definition an efficient constellation is not dominated by any
other constellation. A constellation z = (x, y) is (strictly) dominated by a constellation
z̃ (formally z̃ ≻ z) if z̃ is (at least) indifferent for all variables (z̃j ≿ zj for each j with
j = 1, ..., m + n) and strictly preferred for at least one variable (∃j ′ such that z̃j ′ ≻ zj ′ ).150
The set of efficient constellations is denoted by Z ef f .
To sum up, there are two cases to be distinguished in simulation optimization: If a
unique objective can be defined and calculated, an optimal configuration is searched for.
In case of multiple objectives, the set of efficient constellations has to be determined.
Both tasks are hindered by the following properties of simulation models:
Also, the selection of xopt or Z ef f is a challenging task and much research has been devoted
to this topic.
From a historical point of view, experimental designs were developed to provide methods
for planning real-world experiments.152 Here, typically the set of controllable variables
is large but their domains are restricted. Since most variables are considered as discrete
variables, they are often simplified to dichotomous variables or their domain is reduced
146
The so-called Taguchi approach see e.g. Shang et al. (2004).
147
See e.g. Kleijnen (2007, sec. 4.6) for an example.
148
See e.g. Law (2007, sec. 9.4).
149
Sometimes also called Pareto front.
150
The notation a ≻ b denotes that a is better than b, whereas a ≿ b declares that a is at least as good
as b. This notation is chosen to allow flexibility in ordering relation. E.g. if for all components of z
holds "the smaller the better", ≻ could be replaced by <. However, using e.g. service level measures or
output measures as response variables, requires a converse ordering relation. Due to this, the common
order relations are generalized by preference relations.
151
The sample space depends on the number and domains of the control variables.
152
See e.g. the seminal work of Fisher (1971).
176
to a small number of values. Assume that for all control variables xi with i = 1, ..., m
such that x = (x1 , ..., xm ), a finite number of states mi exist. Then a configuration x is
constituted by a combination of m active states, where each control variable has exactly
one level. Let bi denote a binary vector of length mi reflecting the level of variable
xi by a 1-entry at the corresponding position. Hence, the set of vectors bi is a binary
representation of configuration x. Furthermore, let y denote the vector of responses of
length n. This allows modelling the outcome of configuration x by a simple linear model
(multivariate analysis of variance):
m
ˆ
ˆ x = E (ŷx ) = μ + ∑ Γi ⋅ bi = E (Ĥ
ŷ (x∣μ, Γ1 , ..., Γm )) (4.9)
i=1
where Γi are matrices of dimension n × mi such that the entry in the kth row and lth
column (γkli ) is the effect of the lth level of control variable xi on response variable yk .
Furthermore, μ ∈ Rn denotes the mean vector.
The aim is to (consistently) estimate the coefficients μ and Γi with as little compu-
tational effort as possible. I.e. only the minimal number of configurations153 necessary
to estimate these coefficients should be evaluated. A set of configurations constitutes an
experimental design. A simple experimental design is the full-factorial design generated
by evaluating all possible configurations in the sample space. In this case, ∏m i=1 mi config-
urations have to be simulated. Obviously, this approach can only be realized in situations
when the sample space is small. For most realistic simulation models this approach is
computationally intractable since the number of configurations grows exponentially and
computational power is still limited.
From (4.9) it can be seen that for a particular response variable yj the number of
coefficients to be estimated is 1 + ∑mi=1 mi .
154 Hence, there are only
m
1 + ∑ mi − m (4.10)
i=1
can be estimated simultaneously by a single design’s results such that multiple objectives
do not induce more configurations to be evaluated.159
Resolution III designs are applicable when all interactions between control variables
are assumed to be 0. If only mean and main effects are to be estimated despite the
presence of non-zero two-variable interactions, resolution III designs have to be enhanced
to consistently estimate mean and main effects.160 For enhancing the resolution III design
in the case of dichotomous variables the so-called foldover theorem can be applied.161 The
main idea is that any configuration carries information of both the variables’ main effects
and interaction effects. To separate both effects the mirrored configuration is evaluated
which carries the same interaction effects but the inverse main effects. Hence, the number
of configurations to be evaluated doubles if two-variable interactions are assumed and a
so-called resolution IV design is created.162
If two-variable interactions are assumed to be non-negative and should be estimated,
the MANOVA model (4.9) changes to
m
⎛ b′i ⋅ Γ1(ij) ⋅ bj ⎞
ˆ ⎜ ⎟
ŷx = E (ŷx ) = μ + ∑ Γi ⋅ bi + ∑ ⎜ ⋮ ⎟ (4.11)
i=1 i,j∣j>i
⎜ ⎟
⎝ b′i ⋅ Γn(ij) ⋅ bj ⎠
where Γk(ij) is the matrix of interaction effects between control variables xi and xj of
response variable yk . Matrix Γk(ij) has dimension mi × mj , but the first row and the first
column is always zero.163 Then, the number of coefficients to be estimated for a particular
response variable is given by
m
1 + ∑ ((mi − 1) + ∑ (mi − 1)(mj − 1)). (4.12)
i=1 j>i
ˆ x = E (ŷx ) = μ + Γ ⋅ x
ŷ (4.13)
where Γ is the n × m matrix of regression coefficients. In such a case the control variables’
domains can be reduced to a binary set containing only the corresponding intervals’ ex-
tremes.168 This is because γki (the regression coefficient of control variable xi on response
variable yk ) equals γki ⋅ (xk1 − xk0 ) = γk1i − γk0i where xkl are the chosen interval points of
the control variable and γkli are the coefficients at both levels according to (4.9). Due to
the assumed linearity, efficient constellations ẑˆ = (x, ŷ) ˆ can only exist on the boundary
of the simplex spanned by the control variables’ intervals. Hence, after estimating all
ˆ
coefficients, i.e. μ̂ and Γ̂i , the meta-model Ĥ (x∣μ̂, Γ̂1 , ..., Γ̂m ) can be used to derive the
efficient configurations by checking for dominance of the vectors ẑˆ = (x, ŷ ˆ x ) of all ∏m mi
i=1
possible configurations.
166
For these numbers see e.g. Hinkelmann and Kempthorne (2005b, p. 535).
167
See Kleijnen (2007, ch. 2) for some generators for dichotomous variables.
168
Or any other distinct pair of values from the interval.
179
effect of making Sunday available for train dispatching and RTC processing should be
tested to investigate whether an improvement in the network’s robustness and inventory
balancing can be achieved. Furthermore, the impacts of doubling the turnover capacity for
all chemicals is investigated which might reduce RTC cycle times and, hence, reduce the
number of required RTCs. In brief, Table 4.11 shows the values for each variable.
Table 4.11: Values of the control variables for the experimental design
Since for all variables two states are considered, in total 26 = 64 possible configurations
constitute the corresponding full factorial design. It is assumed that a variable’s effects
on the responses depend on the values of the other variables. E.g. it can be assumed that
the individual effect of extending the turnover capacity of a certain chemical potentially
reduces turnover times. In this case, RTCs might be repositioned faster and, hence, in
a given amount of time transport quantities could be increased. This might manifest
in more smoothed inventory levels and a reduced risk for inventory shortfalls. However,
these potentials might only be exploited if the transport capacities are extended, too. I.e. to
realize these effects, train dispatches on Sunday might be necessary.
In other words, interaction effects between the variables are considered to reflect the
systems’ dynamics. To find a balance between modelling accuracy and computational
effort, only two-variable interactions are considered. Note that, since mi = 2 for all
variables, main effects and two-variable interaction effects correspond to 22 coefficients to
be estimated according to (4.12). Hence, a resolution V design is constructed with 25 = 32
configurations. Table A.17 shows the 32 evaluated configurations. For each configuration
a time span of 180 days is simulated. To assess the responses’ variability, 100 replications
are conducted per configuration.
The system’s reactions under specific constellations of variables manifest themselves in
various measures. Three classes are distinguished here: First, inventory-related measures
reflect costs incurred by holding stocks. On the other hand, stock-out costs occur when
supplied plants have to be shut down due to a shortage of material. These costs typically
reflect the loss in revenues when customer orders cannot be fulfilled due to a shortage
of final products. It is assumed that the target stock levels reflect the points where stock
holding costs and expected stock-out costs are in balance.169 Hence, any deviation from the
desired target stock level implies an increase in total inventory-related costs. Therefore,
169
Note that a re-configuration of the logistical system may lead to an improved logistical performance
(e.g. faster cycle time) and may influence the determination of target stock levels. However, this effect
is not investigated here.
180
tr = ∑ ∑ ∑ trijt , (4.15)
i∈I j∈I t∈T
where trijt denotes the number of trains dispatched in period t on link (i, j).
As a third category production-related performance measures reflect the monetary cash-
flow resulting from production operations. This includes costs for raw and auxiliary mate-
rial consumption, labour costs of operators, opportunity costs for tied capital, and revenues
due to sales of final and intermediate chemicals. A configuration of the network’s logisti-
cal system affects the production system by providing prerequisites for normal production
operations. I.e. a configuration assures that raw materials are available and that final and
intermediate chemicals can be stored and delivered to customers. In this simulation study,
these effects result in the availability or storability of intermediate chemicals. If material is
unavailable or not storable (e.g. due to fully utilized stock capacities), a disturbance of the
production processes results such that plants have to reduce their production outcome or
shut down. An appropriate measure to account for production disturbances is to measure
the gap between planned and realized production for all plants associated to the considered
170
Note that rip can be seen as a relative root mean squared error (RMSE) for chemical p at site i.
171
Since there are two sites and four products in focus.
181
chemicals. This measure equals the β-service level of each chemical considered:
∑t∈T ωipt
real
βip = plan
. (4.16)
∑t∈T ωipt
real plan
In (4.16), ωipt and ωipt denote the realized and planned balance of chemical p in period t
at site i.172 Note that the planned balance ωipt plan
of a chemical also incorporates variations
of the plants’ production states due to technical or other reasons such that the difference
real plan
between ωipt and ωipt only accounts for deviations due to logistical reasons.
To sum up, for the chemicals considered in inter-site transport in total 16 responses
are defined by (4.14) - (4.16).173 To simplify the following analyses it is assumed that all
chemicals at both sites have equal priority. Service level and inventory-related responses
can be merged into average measures by
and
∑i∈I ∑p∈P βip
β= . (4.18)
∣I∣ ⋅ ∣P∣
Then, a scenario comprises six input variables whose effects on three response variables
are evaluated. To calculate the effects, linear models with two-way interaction effects
are estimated for each response variable. The average β-service level is restricted to the
interval [0, 1]. An ordinary linear model is not an appropriate model for probabilistic
measures since its prediction space is not restricted. Therefore, a logit model is used to
describe the input variables’ effects on the average β-service level. I.e.174
6 6
β
log ( ) = μβ + ∑ (γiβ + ∑(γiβ γjβ )ij ) (4.19)
1−β i=1 j>i
where γiβ is the (main) effect of setting input variable xi to its higher value and (γiβ γjβ )ij
is the two-way interaction effect of setting variables xi and xj to their higher values. Note
that μβ corresponds to the basic configuration where all input variables are set to their
lower/initial values.
For both other responses ordinary linear models are estimated, i.e.
6 6
tr = μtr + ∑ (γitr + ∑(γitr γjtr )ij ) (4.20)
i=1 j>i
6 6
r = μr + ∑ (γir + ∑(γir γjr )ij ) (4.21)
i=1 j>i
172 plan
Note that in the considered scenarios ∑t∈T ωipt ≠ 0 for all sites, periods, and products.
173
For C4 at site 1 no consuming plant exist. Hence, no local β-service level exists such that only seven
service levels are calculated.
174
For detailed information about logit models see e.g. Faraway (2005).
182
where the coefficients (γ) are defined as described above. Note that the relative average
sCap
inventory deviation r is restricted to the interval [0, max (1, ip
sTar
)]. This violates the as-
i∈I,p∈P ip
sumption of classic linear models that variables have continuous and unbounded support.
However, an explicit modelling of truncated variables does not reveal a significant advan-
tage of this model as the simulated values are sufficiently far away from their extremes.
Hence, a linear model is used since the violation of model assumptions is expected to be
ignorable. Beside the input variables, a main influence on the responses is the number
and duration of plant break-downs due to technological reasons. To take this influence
into account, the total quantity of the corresponding chemicals produced at both sites is
incorporated as an additional explanatory variable.175 Figures 4.12a - 4.12c show the box-
plots for the three responses. The estimated responses based on the models (4.19)-(4.21)
are superimposed by the symbol ▲.
The corresponding estimated coefficients γ can be found in Table 4.12.
The corresponding QQ-plots for all models’ residuals are depicted in Figures A.3a-A.3c.
The QQ-plots displayed in Figure A.3b and Figure A.3c show a good fit to the normal
175
Note that this variable is independent from the experimental design such that no collinearity exists.
183
1.00
0.95
●
●
● ●
● ●
● ● ● ● ●
●
● ● ● ● ●
● ●
● ● ● ●
●
● ● ● ● ● ● ●
● ● ● ● ●
● ●
● ● ● ● ● ● ● ●
●
● ● ● ● ● ● ● ● ● ● ●
● ●
● ● ● ● ● ●
●
● ● ● ●
● ●
● ● ● ● ● ●
β−service level
● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ●
● ● ● ● ● ● ●
● ● ● ● ● ●
● ●
● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ●
0.90
● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ●
● ● ●
● ● ● ● ● ●
● ●
● ● ●
● ●
● ● ● ●
● ●
● ● ● ●
● ● ●
● ●
● ● ● ● ● ●
● ● ●
●
● ● ● ●
● ● ● ●
● ●
0.85
● ● ●
● ● ●
● ●
●
● ●
●
●
●
● ● ●
● ●
0.80
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
scenario
(a) Boxplots of simulated and fitted β-service levels per scenario
320
● ●
●
●
●
●
300
●
280
# trains
260
● ●
● ● ●
● ● ●
● ●
240
● ● ● ●
●
●
● ● ●
● ●
● ● ●
220
●
●
●
●
fitted value per scenario
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
scenario
(b) Boxplots of simulated and fitted number of trains per scenario
●
●
0.7
● ●
● ●
●
inventory deviation
0.6
0.5
0.4
●
●
●
●
● ●
● ●
●
0.3
● ● ●
● ● ● ●
●
● ●
●
●
fitted value per scenario
●
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
scenario
(c) Boxplots of simulated and fitted inventory deviation per scenario
distribution.176 For the logit model (4.16) the QQ-plot has only limited relevance, since the
residuals are not assumed to be normally distributed.177 Thus, Figure A.3a shows a typical
pattern for truncated variables with increasing concentration towards the truncation. From
the QQ-plots no serious violations of the model assumptions can be detected for all three
fitted models. All models significantly reduce the corresponding residuals’ variances178
whereby the explained variance is about 50% for (4.16) and (4.20) but only 5% for (4.21).
The low level of explained variance for (4.21) is an indicator that either not all relevant
influences are included in the regression or the investigated variables only slightly affect
this measure.
From Figure 4.12 it can be deduced that the variances of the response variables are
similar in the experimental configurations. This presumption is supported by Bartlett’s test
for variance homogeneity for all three responses (at a significance level of 1%).179 Hence,
an effect of the input variables on the variation of the responses cannot be verified such
that a discussion on the robustness of an experimental configuration is obsolete w.r.t. to
the selected three responses. Table 4.12 contains the estimates and t-test probabilities
of the coefficients of models (4.19)-(4.21). It can be observed that the number of trains
dispatchable within a week (shipment capacity) is the only variable that has a significant
effect on all responses. As expected, the effect on the β-service level and the total number
of trains dispatched is positive, whereas the inventory deviation can be significantly reduced
by additionally available train transports. An extension of the turnover capacity for C2
has significant negative effects on the number of trains and the stock deviation.
Only few (slightly) significant interaction effects can be found. A negative impact on
the β-service level occurs for a simultaneous extension of turnover capacities for C2 and
C3 as well as the simultaneous extension of Styrene’s turnover capacity in combination
with extended RTC processing times. Negative two-way interaction effects are not to be
expected whereas positive ones indicate an effective exploitation of extended capacities.
Due to the fact that only 100 replications were performed it is presumed that the (slightly)
significant two-way interaction effects are spurious and reflect a noise effect.
A positive interaction effect on the total number of dispatched trains is observed for the
simultaneous extension of Styrene’s turnover capacity in combination with an extended
RTC processing time. I.e. more trains are dispatched as RTCs can be handled faster.
This might be an indication that the fleet of Styrene RTCs is too small. However, this
176
However, Jarque-Bera tests do not support this founding for both models. A closer inspection of the
QQ-plots reveals indications for slightly fat-tailed distributions of the residuals. Hence, the models
were re-fitted for Box-Cox-transformed dependent variables (with λtrain = 1.4747 and λstock = 1.6363).
For these models Jarque-Bera tests confirm the normality of residuals. For the sake of brevity the
detailed model specifications are not shown here as the general results are the same as for the original
model.
177
See Faraway (2005).
178
This founding is verified by F-tests for models (4.20) and (4.21) as well as Wald’s test for model (4.16)
at a significance level of 1%.
179
Note that the left-skewed distribution of the simulated β-service levels as depicted in Figure 4.12a is
compensated by the logit transformation.
185
180
Configuration # 64 is e.g. dominated by configuration # 60.
181
For details see e.g. configuration 3 in Table A.18.
182
See e.g. configuration # 17 in Table A.18.
183
On average by three trains, see e.g. configuration # 18 in Table A.18.
184
Re-configuration costs are e.g. investments for turnover capacity extensions.
185
Financial impacts may realize e.g. as cost savings due to more smooth stock levels.
186
See Barton and Meckesheimer (2006) for details on meta-modelling using RSM.
187
See Kleijnen (2007).
188
Classic RSM applies a simple steepest descent approach, see Law (2007) or Kleijnen (2007). Typically,
RSM assumes an unconstrained experimental space (except for upper and lower bounds). However,
an extension to incorporate (additional) constraints on the experimental space is provided by Kleijnen
(2008).
186
quantified by various measures.189 If the lack of fit exceeds a critical value, a local poly-
nomial of higher-order (i.e. second-order) is used to approximate the local responses.190
If the local meta-model fits sufficiently accurately, this model is used to determine the
optimal configuration.191
Unfortunately, in classic RSM the local RSA relies on local polynomial relations and
requires that all general assumptions of linear models have to be fulfilled. This includes
homogeneous variances among the configurations and normally distributed errors. In
many simulation experiments these assumptions are violated which hinders a consistent
fit of linear models and interferes the optimization process. Therefore, in advanced RSM,
local linear meta-models are replaced by other functions known as surrogate functions.
In the literature, five main classes of surrogate functions are distinguished: Kriging mod-
els,192 neural networks,193 regression splines,194 support vector regression,195 and radial
basis functions.196 The general procedure of the RSM remains unchanged, but the gra-
dient defining the local sub-space to be explored in the next iteration is calculated based
on the aforementioned local meta-model functions.197
As an alternative to RSM, simulation responses can be used directly to explore the sam-
ple space of control variables. To do so, a lot of combinatorial optimization approaches
were adapted for simulation optimization. In general, there are four main classes of
methods that have shown a particular applicability in (multi-objective) simulation opti-
mization: Meta-heuristics, gradient-based procedures,198 random search,199 and sample
path optimization.200 Of particular interest are meta-heuristics as they have shown a
good performance for a wide range of combinatorial optimization approaches. Therefore,
commercial simulation software primarily uses these techniques to incorporate simulation
optimization routines.201 Among meta-heuristics, tabu search, scatter search, and genetic
algorithms are most widely used.202 Table 4.13 provides an overview on all aforementioned
techniques.
In non-commercial applications often genetic algorithms are proposed.203 In general,
189
Such as the adjusted R2 or the Karesh-Kuhn-Tucker conditions, see Kleijnen (2007).
190
The most prominent design to fit a second-order polynomial is the so-called central-composite design,
see e.g. Kleijnen (2007). Here, a resolution V design is enhanced by 1 + 2 ⋅ m configurations to be able
to estimate quadratic effects, too.
191
See Wilson et al. (2001) for details on applying RSM to find efficiency frontiers.
192
See e.g. Kleijnen (2009) for details.
193
See e.g. Sabuncuoglu and Touhami (2002) or Fonseca et al. (2003).
194
See Li et al. (2010b) or Barton and Meckesheimer (2006).
195
See e.g. Clarke et al. (2005).
196
See e.g. Barton and Meckesheimer (2006) or Barton (2009) for an overview and Shin et al. (2002) for
an application of radial basis functions.
197
See Li et al. (2010b) for a comparison of these meta-models.
198
See e.g. Fu (2006).
199
See Andradóttir (2006).
200
See Fu et al. (2005) for an overview.
201
See e.g. Law (2007) or Banks et al. (2005).
202
See e.g. the commercial build-in optimizer OptQuest Laguna and Martí (2003) and Laguna (2011), or
WITNESS optimizer (Lanner Group, 2002).
203
Albeit, there is also commercial simulation software using this type of meta-heuristics, see e.g. Barton
187
Response surface
direct optimization
approximation
• meta-heuristics
• linear & quadratic regres-
sion (classic RSA) – tabu search
– scatter search
• regression splines
– genetic algorithms
• support vector regression
• gradient-based methods
• radial basis functions
• random search
• neural networks
• sample path optimization
5. repeat step 2 and 4 until the set of non-dominated configurations is sufficiently large.
While the first goal comes along with single-objective optimization, the second goal is
characteristic for multi-objective optimization problems. Hence, multi-objective optimiza-
tion procedures try to spread the pool of non-dominated configurations along the Pareto
front. To do so, the diversity of non-dominated configurations has to be incorporated in
the solution procedure.
206
See Deb et al. (2002) and Srinivas and Deb (1994).
207
See Srinivas and Deb (1994).
208
See Deb et al. (2002).
209
See e.g. Deb et al. (2007) or Bin et al. (2010) for engeneering applications and Puigjaner and Guillén-
Gosálbez (2008) or Mele et al. (2006) for applications in the SCM context.
210
Otherwise, the maximal number of efficient configurations to be found equals the population size in
each iteration. This loss of non-dominated configurations is called Pareto drift and the improved
algorithm is called NSGA-IIa, see Goel et al. (2007).
211
However, typically contracts with only few suppliers will be made to cover the regular demand of
Naphtha. Nonetheless, the sea ship terminal provides the opportunity to order additional supplies
e.g. in case of a shortfall of the regular supplier(s).
189
Unloading is possible every day. The tank capacities at both sites are shown in Table 4.4
whereas the tank capacity at the sea port is assumed to be 75, 000 tons.
The inventory management follows local (si , Si ) policies at both production sites (i =
{1, 2}) and the harbour (i = h). Six parameters have to be determined to fully specify the
SC’s raw material inventory management. If the available stock at location i ∈ {1, 2, h} in
period t (say lit′ ) falls below the corresponding order point si , an order is placed with the
order quantity qit = Si −lit′ . The available stock comprises the local inventory level (lit ) and
all non-delivered orders (∑Tτ=t qiτ ). If orders are placed at the production sites, they are
immediately fulfilled via pipeline as long as a sufficient quantity of Naphtha is available
at the harbour tanks.
The pipeline is assumed to be highly reliable. In certain intervals, however, the pipeline
control system reports failures which are partly caused by misestimations of pressure but
sometimes also due to micro-leakages. In any case, an inspection of the identified pipeline
segment is necessary which forces an interruption of pipeline operations. It is assumed
that the time span between two alarms talarm (given in days) is Weibull distributed with
talarm ∼ WB (k = 1.5, λ = Γ( 5 ) such that E(t
365
)
alarm ) = 365. I.e. one year is expected to
3
elapse before a new alarm occurs.
The inspection time before the pipeline can be used again (tinsp ) is also assumed to be
Weibull distributed with (tinsp − 1) ∼ WB (k = 1.5, λ = Γ(35 ) ) with a minimum inspection
3
time of 1 day in case of a false alarm. If the pipeline has to be repaired, the time span
expands considerably. The formula of the density function and a plot of the density for
talarm are given in the appendix in (A.1) and Figure A.4, respectively.
At the harbour, Naphtha orders are placed according to the same inventory policy. Here,
however, the ordered quantity is delivered after some transport time when the ship arrives
at the harbour. This time span depends on the supplier and the shipper organizing the
transport. The order lead time (tship ) is considered as a Weibull-distributed random vari-
able with (tship − 2) ∼ WB (k = 2, λ = Γ(1.5)
5
). It is assumed that under optimal conditions
(i.e. closest supplier and immediately available shipment capacities) at least two days
elapse before an order can be received.
The objective is to minimize the inventory holding costs which comprises a) ordering/-
shipment costs, b) stock holding costs and c) shortfall costs. The drivers of these costs
are a) the number δ of shipments, b) the average total inventory ¯l and c) the plant uti-
lization β.212 As discussed above, the precise cost rates of these three drivers are assumed
to be unknown and time-varying. The aim is to determine efficient configurations of the
inventory parameters and to quantify the trade-off between the three cost drivers.
From example 9 it is known that the pipeline transport is assumed to induce negligible
fixed costs such that a continuous supply of both sites minimizes the cycle stocks. Hence,
it is concluded that Si = si + ωimax for i = 1, 2 such that at both sites only safety stocks
212 ∑t ∑i lit
β is calculated analogously to (4.18) and (4.16). ¯l is calculated as ¯l = T
. δ simply counts the
number of non-zero orders at the harbour (qht > 0).
190
are at hand to cover shortfalls of the pipeline supply but no cycle stocks due to batch
building.213 If a pipeline is under repair or inspection, the demand to be covered from
stock depends on the inspection time tinsp (which is Weibull distributed) and production
modes of the crackers during this time (which follow a discrete Markov process). Hence,
a (closed-form) mathematical expression of the distribution of Naphtha consumption at a
site during a pipeline inspection is hard to derive. However, Monte-Carlo simulation can
be used to approximate the distributions of this measure at both sites.
Based on the density estimates, the corresponding loss functions are calculated straight-
forwardly. Since the Markov processes of both crackers are very similar, the loss func-
tions show a very similar pattern, differing primarily in the scale (which corresponds to
the crackers’ capacities). Figures 4.13a-4.13c show the estimated density functions and a
QQ-plot depicting the re-order levels to be held at both sites to achieve a specific β-service
level.214 I.e. each point in Figure 4.13c corresponds to a specific β-service level that is
achieved by setting the associated re-order levels at the sites.215 The loss functions for
both sites are shown in Figure A.5 in the appendix.
Based on the estimated density functions, the mean and the 95%-quantile are superim-
posed by the dashed lines in Figure 4.13a and Figure 4.13b whereby the latter coincides
with the re-order level ensuring a 95% α-service level. For both sites the density functions
of total consumption are bimodal and skewed to the right. The first mode is at zero con-
sumption and is inherited from the Markov chain of the production models. It corresponds
to situations when a pipeline inspection coincides with a cracker shut-down. The second
mode is inherited from the Weibull distribution determining the pipeline inspection time
which also causes the skewness.
From Figure 4.13c it can be taken that a linear relation between the re-order levels at
both sites exists. This relation can be used to reduce the set of parameters to be explored by
defining s2 = s1 ⋅ 23 whereby the slope of the fitted regression line in Figure 4.13c corresponds
almost exactly to the ratio of cracker capacities.
Accordingly, the relation between the local re-order level si and the expected loss during a
pipeline inspection period can be quantified.216 Local stocks protect against the shortfall risk
due to a pipeline break-down. Additionally, the variability of supply has to be examined.
Since it is generally advantageous to consolidate inventories (risk pooling), stocks at the
harbour should be held to buffer supply and production risk. Therefore, the parameters
(sh , Sh ) have to be determined controlling the Naphtha availability of both crackers and
the number of shipments for replenishment. The density of the total Naphtha consumption
during the order lead time is a convolution of Markov processes and the Weibull distributed
213
Here, ωimax denotes the maximum consumption rate of cracker i.
214
The density functions are derived by Gaussian kernel estimation with the default settings of the density
function from the stats package based on a 10,000-replicates sample, see R Core Team (2012). The
loss functions are calculated straightforwardly from the density functions’ estimates.
215
The leftmost point implies a β-service level of 5% whereas the rightmost point constitutes a β-service
level of 99.5%.
216
This corresponds to the expected cracker utilization and, hence, expected shortfall costs.
191
mean 95%−quantile
0.00006
0.00004
density
0.00002
0.00000
mean 95%−quantile
0.00008
density
0.00004
0.00000
●
15000
●
safety stock site 2
●
●
●
●
●
10000
●
●
●
●
●●
●●
●●
●●
●●
●●
●●
●●
●●●●●
●●●
5000
●●●
●●●
●●●
●●●
●●●●
●●●●
●●
●●●●●
●●●
●
●
●●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●
0
Figure 4.13: Loss functions for Naphtha consumption during pipeline inspection and their
relation
lead time. Figure 4.14a and Figure 4.14b show the simulated density and loss function of
192
this measure.217
mean 95%−quantile
0.000030
0.000020
density
0.000010
0.000000
V−1⎛⎝E(X) ⋅ (0.05)⎞⎠
25000
expected loss V(r) (in t)
15000
5000
E(X) ⋅ (0.05)
0
Figure 4.14: Density and loss function for total Naphtha consumption during order lead
time
Figure 4.14a shows no (clear) bimodality whereby the density for a zero-consumption is
positive. This is because the independent Markov chains of both crackers form a collapsed
Markov process whose state space and steady state vector is smoother. In both figures,
95%-service levels are superimposed by dashed lines (the α- service level in Figure 4.14a
and the β-service level in Figure 4.14b). Using this information about the central re-order
level at the harbour in combination with the information about the local re-order level, the
system’s total expected backlog could be calculated using the density and loss functions of
each stage. However, the interaction of local safety stocks and the central safety stock at
the harbour needs to be formulated explicitly which is quite cumbersome. Since the local
217
Again a Monte-Carlo simulation with a sample size of 10,000 is used. The estimation procedure is the
same as described before.
193
safety stocks also buffer lacks of supply due to delayed shipments, an independent setting
of the parameters si based on the individual loss functions underestimates the real capacity
utilization (and the β-service level). Furthermore, the deduction of the average stock level
at all three locations cannot be expressed in a closed form.218 Therefore, the complete
inventory system is simulated.
The system is modelled on a daily basis and implemented as an R program. Due to the
simplicity of the model, the program’s pseudocode is kept for the appendix (Table A.19).
The NSGA-II algorithm is used for optimization with three parameters specifying a con-
figuration: s1 , sh , and Sh . The remaining parameters are set according to the relations
discussed above. I.e. the re-order point of location 2 is given by s2 = s1 ⋅ 2/3 whereas the
order levels are expressed as S1 = s1 + 3, 000 and S2 = s2 + 2, 000. The shipment quantity
Sh − sh is restricted to be at least 15,000 tons which avoids too frequent deliveries and
ensures a sufficient utilization of the tanker fleets. Furthermore, lower bounds for the
re-order points are set to s1 , sh ≥ 5, 000 to ensure a minimal safety stock at both crackers.
The parameters’ upper bounds are given by the local tank capacities. The implementation
of the NSGA-II algorithm in the nsga2 function contained in the mco package of the R
environment is used.219 In this optimization procedure, the population size is set to 1,024
which is a sufficiently large sample for further statistical analysis. The number of gener-
ations is set to 50 which results in a fairly good convergence to the Pareto front.220 The
remaining parameters are set to the default values.221
Each configuration is simulated for a 10-years period to account for the rather long
intervals between pipeline inspections. No replications of a configuration are performed
since the variability of the performance measures primarily depends on the length of the
simulated time span. Averaging over a number of replications would in turn overweight
the warm-up phase of the simulation. In the warm-up phase the performance measures
are quite stable due to fixed initial stocks.222 The optimization using NSGA-II generates a
Pareto front consisting of 1,024 observations. Figure 4.15 visualizes the found solutions.
In Figure 4.15 the average total stock in tons per period is plotted against the number
of shipments ordered per period. Each Pareto-optimal solution is represented by a grey-
scaled point whereby the shade indicates the realized β-service level of this solution. Bright
solutions show a β-service level of about 40% whereas dark-coloured refer to a β-service
level close to 100%.
This sample can be used to describe the Pareto front by an appropriate surrogate function
218
See Schneider et al. (1995) for an approximation.
219
See Trautmann et al. (2010).
220
Experiments with 100 and 200 generations and smaller population size reveal no significant shifts in
the Pareto front.
221
The defaults are e.g. a mutation probability of 0.2 and a crossover probability of 0.7, see Trautmann
et al. (2010).
222
Initial stocks are 15,000, 10,000, and 50,000 tons for locations 1,2, and the harbour stocks, respectively.
In general, for such a simulation study there is a trade-off between multiple replications and one long
simulation run, see Law (2007, sec. 9.5).
194
● ●● ● ● ●
●● ●●● ● ●
● ● ●● ●● ●
β−service
● ● ● ● ●●
●●
●● ●● ● ●●
●●●
●
● ● ●
●● ● ● ●●● ● ● ●● ●
●
●●
● ●● ● ● ● ●● ● ●●
●
● ● ● ● level●
●
number of shipments (per period)
● ● ● ●●● ● ● ●
● ●● ●●● ● ●● ● ●
● ● ● ● 100 %
0.20
● ● ● ●● ● ●
●●●
● ● ● ● ●● ●
●
● ●● ● ●
●●●●●
● ● ● ● ● ●●●●● ●
●● ● ●
●● ●● ●●● ●● ● ●
●● ●●●●● ● ●●● ● ● ●●
●
●● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ●●● ● ● ● ●● ● ●● 84 %
●●● ● ●● ● ● ● ● ●
●
●● ● ● ● ● ● ● ●
● ● ●● ● ● ● ● ● ●
● ●● ● ●● ● ● ●
●● ●●● ●●
● ●● ●● ● ● ● ●
●● ● ●● ●
●● ● ● ● ● ● ●
●●●● ●● ● ●● ● ●● ● ●● ●● ●
● ●
● ●● ● ● ● ● ● ● ● 46 %
0.15
●
● ● ● ●● ● ● ●
● ● ● ●● ● ● ● ● ● ●
● ● ●
●
● ●
● ● ● ● ● ● ● ●● ●
●● ●● ●●● ● ● ●● ● ● ● ●
●● ● ●●● ● ●
● ● ● ● ●
●
●● ●●●
● ● ● ●
●● ● ● ● ●
●
●●●●●● ●●● ●● ● ●
● ● ●●
● ●● ● ● ● ● ● ●
● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●●
● ●●●●●
●●
●●● ●● ● ●●
● ●●●
● ●●
● ●● ● ● ● ●● ●● ●● ●
● ● ● ●● ●
●
●
●
●
●●●
●●●
●●●●●●
● ● ●●●● ● ● ● ●● ● ●
●●●
●●
● ● ●
● ● ● ● ●● ● ●
●
●●●
●
●
●
●
●● ●
● ●● ●
● ●● ● ● ●●
●
● ● ●
● ● ●
●
●●
●
●●
●
●●●
● ●
●●●● ●
●
●●● ● ●●● ● ●
● ● ●●●● ● ●
●●
●● ● ● ● ● ● ● ●
●
●
●●
●
●● ●● ● ● ● ● ● ●
●● ● ●● ●● ● ●●●
0.10
●
●●●
●
● ● ● ●● ●
●
●
●●
●●●●●
● ●● ●●● ● ● ●●
●
●●●
●●
●
●●●●●● ● ● ● ● ● ● ●
●
●
● ●
●●
●●●● ●● ● ● ●
● ●●
● ●● ● ●● ●
● ●●● ●● ●
● ●
●●
●●●
● ●●●● ●
● ● ● ● ● ●● ●
● ● ●
●
● ●●● ● ● ● ●●● ● ●
●● ● ●● ●
●●
●●●●●
●●
●● ● ●● ● ● ● ● ● ●
●●●
●● ●
●●
●●●
●● ● ●
●●
●
●● ●● ● ● ● ● ● ●
● ●●● ● ● ● ● ● ● ●
●●●●● ● ●●● ● ● ● ●
● ● ● ● ●● ● ●
●●●● ●● ● ● ● ● ● ●● ● ●
● ● ●● ●● ●
●
●●●●● ●●
● ●● ● ● ● ●●●●●
●
●●● ● ● ●
●● ●● ●● ●● ●●
●●
●
●● ● ● ● ● ●● ●● ● ●●●●●
●●
●●● ●●●● ● ●● ● ●● ● ●
●
0.05
●●●●●●
●● ●
●
●●●
● ●
● ●● ●●
●
●●●●
● ●●● ●●
which eases the determination of optimal configuration in the following decision support
steps.
Among the three performance measures, the realized β-service level depends on both
other performance measures. Let i be the index of Pareto optimal solutions and define by
¯li , δi and βi the corresponding performance measures. Then, the following model can be
used to approximate the observed configurations
βi
log ( ) = μ + γ11 ⋅ ¯li + γ12 ⋅ δi + γ21 ⋅ ¯li2 + γ22 ⋅ δi2 + (4.22)
1 − βi
γ31 ⋅ log(¯li ) + γ32 ⋅ log(δi ) + γ1 ⋅ δi ⋅ ¯li +
i .
The model summary of (4.22), including parameter estimates, can be found in Ta-
ble A.20. The model shows a good overall fit which is indicated by unsuspicious residual
plots (see Figure A.6) and McFadden’s pseudo-R2 of about 0.99. Using this model, the
Pareto front can be described continuously which is depicted in Figure 4.16.
In Figure 4.16 the Pareto front is evaluated over the area spanned by the sample depicted
in Figure 4.15. The grey scale, representing the β-service level, is the same as before.
Additionally, contour lines are superimposed indicating levels of constant β-service level.
Figure 4.16 quantifies the trade-off between stock holding and the frequency of shipping
w.r.t. the β-service level. The shape of the substitution lines remains stable for all β-
service levels but the substitution area (i.e. the length of the contour line) decreases for
high β-service levels (β > 99.5%). This implies that a decrease in average stock cannot
be compensated arbitrarily by an increase of the shipping frequency. The minimum stock
level to achieve a specific β-service level forms the upper end of each contour line. For
195
β−service
level
1.0
number of shipments (per period)
0.20
0.9
99
0.9
0.9
0
97
0.9
5
950
0.8
0.9
900
0.15
0.9
750
0.7
0.9
0.5000
500
0.9
000
0.6
0.60
0.10
0.8
00
00
0.7
0
00
0.5
0
instance, the average stock level to achieve a β-service level of 95% varies between 20,000
and 38,000 tons approximately.
Based on the individual loss functions depicted in Figure A.5a and Figure A.5b local
re-order levels s1 and s2 of 18,000 and 13,000 tons are indicated, respectively. Setting
the re-order level sh at the harbour to about 42,000 tons (as Figure 4.14b suggests) and
the minimal order-up-to level Sh to 57,000 tons, results in an average stock level of ap-
proximately 50,000 tons accompanied with a β-service level close to 100% and a shipment
frequency of about 0.28 ships per day. As 0.28 exceeds the maximum of Pareto-optimal
solutions, this configuration is clearly inefficient. Obviously, a component-wise setting of
inventory parameters does not lead to a sufficiently accurately modelled system perfor-
mance. Based on the estimated relation between the performance measures depicted in
Figure 4.16, however, the optimal configuration for given weights of all three measures
can be easily determined.
196
223
See Power (2002, ch. 1).
224
For an encompassing classification see e.g. Power (2002, ch. 1).
225
See Power (2002, ch. 1).
226
The fast access to (statistical) analyses of huge databases is often referred to as on-line analytical
processing (OLAP), see e.g. Shim et al. (2002).
227
See Power and Sharda (2007).
228
Such as an analytic hierarchy process, see Power and Sharda (2007).
229
So-called decision analyses, see Power and Sharda (2007).
230
See Power and Sharda (2007).
231
See Shim et al. (2002).
197
ˆ
then be used to estimate the configuration x̂ that corresponds to ŷ by evaluating Ĥ −1 for
ˆ −1
ˆ = Ĥ
ŷ, i.e. x̂ (ŷ). The procedure is briefly described as follows:
ˆ
3. estimate a local (linear) model based on (xi , yi ) = zi ∈ Z̄ŷef f , i.e. xi = Ĥŷ−1 (yi ) +
i
ˆ
4. estimate a desired configuration x̂ by x̂ = Ĥŷ−1 (ŷ).
In the described procedure, two components have to be specified: First, which type of model
ˆ
to use for the local model Ĥŷ−1 and, second, how to determine Z̄ŷef f . In this example for
ˆ
the first component Ĥŷ−1 a simple first order linear regression model with interaction terms
is estimated for each performance measure. I.e. for all (xi , yi ) ∈ Z̄ŷef f holds
234
If this restriction would be violated, either a simpler model (e.g. by dropping interaction effects) had
to be chosen or the neighbourhood Z̄ ef f had to be enlarged.
235
See Delaunay (1934) for the original paper.
236
As long as the data set is in general position, i.e. no co-circular or degenerated subsets exist, see De Berg
et al. (2008, ch. 9) for a precise definition.
199
ˆ
of Ĥŷ−1 .237 In the two-dimensional case this fact is supported by the property that the
Delaunay triangulation maximizes the minimum angle of all simplices.238 Alternatively,
the procedure could be briefly described as finding all adjacent points of ŷ in the Delaunay
graph.239
To illustrate the determination of Z̄ŷef f , Figure 4.17 shows a Delaunay triangulation for
an artificial data set drawn from a bivariate uniform distribution.
● ●
●
●
●
● ●●
●
●
●
● ●
●
●
● ●
●
8
●
●
●
●
●
●
● ● ●
●
●
● ●
●
●
6
●
●
●
●
● ●
● ●
●
● ● ●●
●
●
● ●
●
4
●
●
●
●
● ●
●
●
●
●
●
●
● ●
●
2
●
●
●●
●
●
●
●
● ●
●
●
●
●
● ●
●●
●
0
0 5 10 15 20
Here, the ∎-marked point represents the performance vector for which the corresponding
configuration is to be determined. All other points represent the set of Pareto-optimal
solutions Z̄ ef f . ▲-marked points are adjacent to ŷ. The corresponding simplices incident
to ŷ are highlighted in blue.
To assess the performance of the proposed procedure for the Pareto-set of the example
(see Figure 4.15), the goodness of approximation is tested by comparing the desired perfor-
mance vector ŷ with the performance of the configuration determined with the described
procedure (Ĥ(x̂)). Ideally, both performance vectors are equal. To measure the similarity
of both performance vectors the relative deviation of ŷ w.r.t. its corresponding simulated
∣Ĥ(x̂j )−ŷj ∣
performance Ĥ(x̂) is calculated by ej = Ĥ(x̂ . I.e. ej contains the relative deviation of
j)
each performance component. As a one-dimensional similarity measure, let ej denote the
average of ej . The average relative deviation (ARD) is calculated for a set of 1,000 ran-
domly distributed points Ŷ = {ŷj } with ∣Ŷ∣ = 1, 000. Note that implicitly also the accuracy
of fit of (4.22) is measured. Hence, this test serves as a test for the overall meta-modelling
237
The simplices incident to ŷ build a convex and compact sub-space containing ŷ. Since this sub-space
is constituted by simple geometric elements, the constituting points have to be distributed around ŷ.
238
See De Berg et al. (2008, ch. 9).
239
See De Berg et al. (2008, ch. 9) for the precise definition of the Delaunay graph.
200
accuracy. Figure 4.18 shows a grey-scaled scatterplot of ARDs for the sample of 1,000
performance vectors tested.
As Figure 4.18b displays, Ŷ is uniformly distributed over the area of the Pareto-front.
The grey scale is logarithmized to account for the skewed distribution of ARDs as Fig-
ure 4.18a shows. From Figure 4.18a it can be observed that the performances of more
than 80% of the estimated configurations x̂j deviate at most 10% from the intended per-
formance ŷj . This is seen as a very good fit given the fact that the variation of the
simulation model, the variation of the Pareto-front model (4.22), and the variation of the
ˆ
local linear model Ĥŷ−1 is all reflected in ej . Nonetheless, it is apparent that comparatively
bad fits are more frequent in the uppr-right region of the sample space. In this region
the density of evaluated configurations is sparse (compare Figure 4.15) which hampers
accurately fitted local regression models. This region of the Pareto-front corresponds to
configurations with extreme β-service levels close to 100%. I.e. there are numerical and
statistical problems to decide on the efficiency of configurations.240 Moreover, it is at least
questionable whether this region is of economic relevance, as configurations with such a
high shipment frequency and/or average stock level are likely to be disregarded due to the
small benefit in β-service level compared to the exploding logistic costs.
To illustrate a more realistic decision problem, assume a stock holding cost rate for
Naphtha chold of 35 e per ton and year as calculated in example 9 and average ordering
costs per shipment corder of 15,000 e. The total cost for stock holding and shipment is
hold
C = c365 ⋅ ¯l+corder ⋅δ. Furthermore, a minimum β-service level of 99% is required. Then, the
optimal performance vector based on (4.22) and the total cost function can be calculated
as ŷopt = (0.99; 42, 850; 0.0945) with total cost of 5,526 e per day. Figure 4.19 displays
the found solution graphically.
The corresponding configuration x̂opt is calculated as x̂opt = (15, 591; 29, 667; 76, 538).
I.e. based on the given cost parameters and the results of the simulation study, the optimal
inventory parameters can be derived as shown in Table 4.14. They differ from the baseline
values derived at the end of example 12 by component-wise deductions.
Clearly, the optimal values are smaller than the baseline parameters (except for Sh )
indicating a smaller average stock level. Moreover, for the optimal configuration, the
order frequency at the harbour is smaller since Shopt −sopt
h > Sh
base
−sbase
h . This implies fewer
240
This is caused by rounding errors during simulation as well as the natural stochasticity of the responses.
I.e. it cannot be decided whether a (marginal) difference in a response is caused by noise or an
advantageous constellation of parameters.
● ●● ● ● ● ● ● av. relative
●●●● ●● ● ● ● ● ●●●●● ● ● ●●●●
● ●
● ● ●●● ● ● ●● ● ● deviation
● ● ●● ● ● ●●●
0.5
● ● ● ● ● ● ● ● ● ●
● ● 54 %
● ●●● ● ● ● ●● ● ● ● ● ●●● ● ●
● ●● ● ● ● ●● ● ●●●●●● ● ● ● ●
●● ● ●● ● ● ● ●
● ●● ●
● ● ● ●
● ● ●
● ●
●● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ●●
● ●
● ● ● ● ● ● ● ● ●● ● 5%
● ● ●
●●●
● ● ●●●● ● ● ● ●
0.20
● ● ● ● ● ● ●●●●
● ●●
● ● ●● ●● ● ● ● ● ● ●●● ●
● ● ● ● ●●●●●●
●● ● ● ● ● ● ● ● ● ●●
0.4
● ●●● ● ● ●●● ● ●● ●● ● ● ● ● ● ● ● ●
●● ● ● ●● ● ●● ● 0%
● ● ●
● ●●●● ●● ● ●● ● ● ● ●
● ●●●● ●● ● ●
● ● ● ●
● ●
● ● ●●● ● ● ●● ● ● ● ●●
●● ●
● ● ●● ●
● ●● ●● ●
● ●
●●●●● ● ●● ●
● ● ● ● ●● ● ● ● ● ● ● ● ●●
● ● ●● ● ● ●
● ● ● ● ●● ● ●
● ● ● ●
● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ●
● ●
●● ●● ● ● ● ● ● ● ● ● ● ● ●
●●
●●●● ● ●●
●●
●● ● ● ●
● ● ●
● ● ●
● ● ● ● ● ● ●●
0.3
● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●●
● ● ●● ●● ●
● ● ●
● ●●● ● ● ● ● ● ● ●●● ● ●
0.15
● ● ● ● ● ● ● ● ● ● ●● ●●
● ●●
● ●●● ● ●
● ● ● ●● ● ● ● ●● ● ● ● ●● ●
● ●● ●
●● ● ● ● ●● ●●● ● ● ●●
● ● ● ●
●●● ●
●● ● ●● ● ● ●● ● ● ● ●
● ●●
●● ● ● ● ● ●
● ● ● ● ● ●●●
● ●
●● ●
●● ● ● ● ● ● ● ● ● ●●● ● ●●●● ● ● ●
●
●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ●
● ●● ●
● ● ●
● ● ● ● ● ● ●● ●● ● ●
●
●●● ●
● ● ● ●● ●
● ● ● ●●
● ●●● ● ● ● ●●●●
0.2
● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ●
● ● ● ● ● ●●● ● ● ● ●
●
0.10
● ● ● ● ● ● ● ● ●
● ●● ● ● ● ● ● ●●
● ● ● ● ● ● ● ●
● ● ● ●● ●● ● ● ●● ●● ● ●
● ●● ● ● ● ●
● ● ●●● ●
● ● ● ●●●●●● ● ●● ● ●● ●
0.1
● ●
●● ● ●
● ● ●● ● ● ●●
●● ●●● ● ●● ●
● ● ●
● ●●● ● ● ●● ●
● ● ● ● ● ● ● ● ●
●● ●● ● ● ● ● ●
● ● ●● ● ●●● ●
● ● ●●●
● ● ● ● ●●● ● ● ● ●
● ●● ●●● ● ● ● ● ●
●●● ●●●● ● ●●●
●
●
● ● ●
0.0
0.05
10000 20000 30000 40000 50000 60000
av. total stock level
β−service
level
1.0
number of shipments (per period)
0.20
0.9
99
0.9
0 .9
0
97
0.9
5
950
0.8
0.9
900
0.15
0.9
750
0.7
0.9
0.5000
500
optimal
0.9
performance
000
vector 0.6
0.60
0.10
0.8
00
00
0.7
0
00
0.5
0
shipments and, hence, smaller shipping costs. To verify the performance of the estimated
optimal configuration x̂opt , a simulation run over 10 years is performed revealing an ARD
to the estimated performance of 3.2%. Table 4.15 shows the detailed performance measures
for the optimal and the baseline configuration.
Table 4.15: Performance measures for optimal and baseline values of inventory parameters
Table 4.15 reveals that the baseline configuration is not dominated by the optimal con-
figuration (due to the maximum β-service level) but total cost are 8,995 e. In contrast,
the simulated total costs for the optimal configuration account for only 5,266 e per day
ˆ
which is approximately 5% below the estimate based on Ĥ(x̂opt ). However, this comes to
the cost of a reduced reliability of the inventory system as the simulated β-service level
(98%) is smaller than the desired 99%. In practice, it is probably advisable to round up
the optimal parameter (as shown in Table 4.14) to a desired accuracy (e.g. whole hundreds
or thousands).
203
2. transportation models
T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,
DOI 10.1007/978-3-658-08433-2_5, © Springer Fachmedien Wiesbaden 2015
204
To model chemical production processes, time series methods are used. Theoretical and
empirical indications are given that this class of models is able to capture the characteristic
short-term time-dependency patterns of physical and chemical transformation processes.
Simultaneously, the models’ complexity is kept at an acceptable level. To model the
change of production modes of chemical production plants in the long run, Markov chains
are used. The application of these models is illustrated by real-world case studies.
Transports in basic chemical industry are performed mainly by means of pipeline, ship,
and rail. This work provides an overview on typical operational distribution planning
problems in chemical industry for each of these three modes of transport. For pipeline
distribution planning a classification of the literature is provided by means of the topology
of the pipeline network and the number of products transported. For single-product
pipelines with multiple access points a model for determining safety stocks along the
pipeline is proposed. Thereby, the safety stock calculation relies on a convolution of
multiple Markov chains which represent the chemical production processes at the access
points. In case of one-to-one pipelines for multiple products with product-specific but
constant production/consumption rates, a typical planning problem is to find batch sizes
and a pumping sequence that minimize stock holding and pumping costs. It is shown
that this planning problem can be modelled as an adapted and enhanced economic lot-
scheduling problem. A small-scaled but realistic case study illustrates the application of
the developed model.
Rail transports are a key component in basic chemical logistics in particular for pro-
duction sites which have no waterway access. This mode of transport is comparatively
flexible but comes along with only slightly higher transport costs. A typical operational
planning problem in chemical production networks is the determination of transport flows
between the sites under consideration of the sites’ logistical capacities (such as turnover
and storage), the available number of rail cars, and restrictions of the rail operator. In the
literature no approach has yet been published addressing this type of planning problem.
Therefore, the MC-RTP, a time-discrete, multi-layer network design model, is proposed
that is able to determine optimal distribution plans. For larger problem instances a
heuristical procedure based on a relaxed MC-RTP is described.
Ship transports are used for most raw material transports in basic chemical industry.
Typical planning problems address the determination of transport flows among a set
of ports where production sites are located at which chemicals are surplus or required.
Two basic types of planning sub-problems can be distinguished: inventory routing and
inventory shipping problems. In the former case, the ships/tankers are owned or managed
by the shipper. Hence, routes and schedules for the available ships/tankers have to be
planned in detail depending on the available stocks at the ports and technical restrictions.
The corresponding literature on maritime inventory routing is briefly reviewed. In the
latter case, transport capacities are booked on the spot market. Hence, the problem is
to schedule and quantify transport flows between ports under consideration of technical
205
and logistical restrictions. A MILP addressing this type of problem is proposed by Ronen
(2002). This model is extended by considering different tanker types which differ in
size, type and number of tanks installed on-board. Hence, beside the quantification and
scheduling of transport flows, decisions about the tankers booked and the assignment of
tanks are modelled. Again a heuristic to solve larger instances is described.
The above mentioned models can be used to support operational distribution decisions
under consideration of the available capacities. On the tactical and strategical level such
capacities are subjects of interest. To be able to find improved SC configurations, the
effects of alterations of the system’s capacities have to be measured. These effects are
hard to anticipate since stochastic processes often complicate theoretical deductions which
also affect the planning at the operational level. Therefore, this work advocates simula-
tion models as an appropriate approach to model complex chemical SCs with multiple
stochastic processes and multiple objectives. This modelling approach integrates oper-
ational planning models in a stochastic environment to accurately reflect the dynamic
effects of alternative SC configurations.
Typical sources of uncertainty in chemical SCs are categorized and their modelling is
discussed. The simulation technique advocated is discrete-event simulation whose generic
components are divided into production, logistical, and auxiliary processors. To support
strategical and tactical decisions often multiple objectives are pursued. Therefore, the two
basic classes of simulation optimization approaches are briefly described and illustrated
by means of a large-scaled example: experimental designs and simulation optimization.
The former class is particularly attractive in case of discrete decision variables and keeps
the computational effort at a minimum. The latter class primarily addresses situations
when decision variables are continuous or a mixed set of variables occurs. From the various
approaches proposed in the literature, the genetic algorithm NSGA-II1 is briefly described
and applied to an exemplary simulation model. This algorithm returns a sample of the
Pareto-front of the multi-objective optimization problem. Based on this sample, a method
to quickly construct efficient SC configurations without evaluating additional simulation
runs is presented which relies on the Delaunay triangulation of the Pareto-front sample.
For the presented example, the procedure shows a convincing performance.
Simulation models can help to understand and investigate very complex systems, in
particular in case of a highly stochastic environment. The interactions between the differ-
ent logistical sub-systems (rail, ship, and pipeline) in chemical SCs provide potential for
further research. Besides, also hybrid simulation approaches incorporating more complex
production planning and scheduling decisions (as apparent to produce speciality chemi-
cals) may reveal potentials for further improvements. On the level of isolated planning
problems further research in rail distribution planning may address continuous time mod-
els or the development of heuristical solution procedures for e.g. an integrated fleet sizing
and distribution planning model.
1
See Deb et al. (2007).
207
Bibliography
Adhitya, A. and Srinivasan, R. Dynamic Simulation and Decision Support for Multisite
Specialty Chemicals Supply Chain. Industrial & Engineering Chemistry Research, 49:
9917–9931, 2010.
Agami, N., Saleh, M., and Rasmy, M. Supply chain performance measurement approaches:
Review and classification. Journal of Organizational Management Studies, pages 1–20,
2012. doi: 10.5171/2012.872753.
Al-Othmann, W., Labadibi, H., Alatiqi, I., and Al-Shayji, K. Supply chain optimization
of petroleum organization under uncertainty in market demands and prices. European
Journal of Operational Research, 189:822–840, 2008.
Alhajri, I., Elkamel, A., Albahri, T., and Douglas, P. A nonlinear programming model for
refinery planning and optimisation with rigorous process models and product quality
specifications. International Journal of Oil, Gas and Coal Technology, 1(3):283–307,
2008.
Andersson, H., Hoff, A., Christiansen, M., Hasle, G., and Løkketangen, A. Industrial as-
pects and literature survey: Combined inventory management and routing. Computers
& Operations Research, 37(9):1515–1536, 2010.
Applequist, G., Pekny, J., and Reklaitis, G. Risk and uncertainty in managing chemical
manufacturing supply chains. Computers & Chemical Engineering, 24(9-10):2211–2222,
2000.
T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,
DOI 10.1007/978-3-658-08433-2, © Springer Fachmedien Wiesbaden 2015
208
Baerns, M., Behr, A., Brehm, A., Gmehling, J., Hofmann, H., Onken, U., and Renken,
A. Technische Chemie. J. Wiley & Sons, Hoboken, NJ, 2006.
Banks, J., Carson, J., Nelson, B., and Nicol, D. Discrete-Event System Simulation.
Prentice Hall, Upper Saddle River, NJ, USA, 4th edition, 2005.
Barceló, S., Vidal-Puig, S., and Ferrer, A. Comparison of multivariate statistical methods
for dynamic systems modeling. Quality and Reliability Engineering International, 27
(1):107–124, 2011.
Bauwens, L., Laurent, S., and Rombouts, J. Multivariate GARCH models: a survey.
Journal of Applied Econometrics, 21(1):79–109, 2006.
Beamon, B. M. Supply chain design and analysis: Models and methods. International
Journal of Production Economics, 55(3):281–294, 1998.
Behr, A., Agar, D., and Jörissen, J. Einführung in die technische Chemie. Springer,
Berlin, 2010.
Berning, G., Brandenburg, M., Gürsoy, K., Kussi, J. S., Mehta, V., and Tölle, F.-J. Inte-
grating collaborative planning and supply chain optimization for the chemical process
industry (i)—methodology. Computers & Chemical Engineering, 28(6):913–927, 2004.
209
Bill, A. Carbon dioxide hydrogenation to methanol at low pressure and temperature. PhD
thesis, École Polytechnique Fédérale de Lausanne, Lausanne, 1997.
Bin, X., Nan, C., and Huajun, C. An integrated method of multi-objective optimization
for complex mechanical structure. Advances in Engineering Software, 41(2):277–285,
2010.
Blaẑiĉ, S., Matko, D., and Geiger, G. Simple model of a multi-batch driven pipeline.
Mathematics and Computers in Simulation, 64:617–630, 2004.
Bok, J.-K., Grossmann, I., and Park, S. Supply chain optimization in continuous flexible
process networks. Industrial & Engineering Chemistry Research, 39(5):1279–1290, 2000.
Borsboom, D., Mellenbergh, G., and Van Heerden, J. The theoretical status of latent
variables. Psychological Review, 110(2):203, 2003.
Boschetto, S., Felizari, L., Yamamoto, L., Magatão, L., Stebel, S., Neves-Jr, F., Arruda,
L., Lüders, R., Ribas, P., and Bernardo, L. An integrated framework for operational
scheduling of a real-world pipeline network. Computer Aided Chemical Engineering, 25:
259–264, 2008.
Box, G. and Draper, N. Response surfaces, mixtures, and ridge analyses. Wiley-
Interscience, Hoboken, NJ, 2nd edition, 2007.
Box, G., Jenkins, G., and Reinsel, G. Time series analysis: Forecasting and control. John
Wiley & Sons, Hoboken, NJ, 4th edition, 2008.
Boysen, N., Jaehn, F., and Pesch, E. New bounds and algorithms for the transshipment
yard scheduling problem. Journal of Scheduling, 15:499–511, 2012.
210
Brockwell, P. and Davis, R. Introduction to Time Series and Forecasting. Springer, New
York, 2nd edition, 2002.
Cafaro, D. and Cerdá, J. Optimal scheduling of refined products pipelines with multiple
sources. Industrial & Engineering Chemistry Research, 48(14):6675–6689, 2009.
Camacho, E. and Bordons, C. Model predictive control. Springer, New York, 2004.
Cameron, D. B., Ødegaard, R. J., and Glende, E. On-line modelling in the petroleum
industry: Successful applications and future perspectives. In Gani, R. and Jørgensen,
S. B., editors, European Symposium on Computer Aided Process Engineering, volume 9
of Computer Aided Chemical Engineering, pages 111–116. 2001.
Chatfield, D. The economic lot scheduling problem: A pure genetic search approach.
Computers & Operations Research, 34(10):2865–2881, 2007.
Chen, C.-L. and Lee, W.-C. Multi-objective optimization of multi-echelon supply chain
networks with uncertain product demands and prices. Computers & Chemical Engi-
neering, 28(6):1131–1144, 2004.
Chen, C.-L., Wang, B.-W., and Lee, W.-C. Multiobjective optimization for a multien-
terprise supply chain network. Industrial & Engineering Chemistry Research, 42(9):
1879–1889, 2003.
Cheon, M.-S., Furman, K. C., and Shaffer, T. D. A modeling framework for railcar fleet
sizing in the chemical industry. Industrial & Engineering Chemistry Research, 51(29):
9825–9834, 2012.
Christiansen, M. and Nygreen, B. Modelling path flows for a combined ship routingand
inventory management problem. Annals of Operations Research, 82:391–413, 1998.
Christiansen, M., Fagerholt, K., and Ronen, D. Ship routing and scheduling: Status and
perspectives. Transportation Science, 38(1):1–18, 2004.
Christiansen, M., Fagerholt, K., Flatberg, T., Haugen, Ř., Kloster, O., and Lund, E.
Maritime inventory routing with multiple products: A case study from the cement
industry. European Journal of Operational Research, 208(1):86–94, 2011.
Clarke, S., Griebsch, J., and Simpson, T. Analysis of support vector regression for ap-
proximation of complex engineering analyses. Journal of Mechanical Design, 127(6):
1077–1087, 2005.
Closs, D., Keller, S., and Mollenkopf, D. Chemical rail transport: The benefits of relia-
bility. Transportation Journal, 42(3), 2003.
Coello, C., Lamont, G., and Van Veldhuizen, D. Evolutionary algorithms for solving
multi-objective problems. Springer, New York, 2nd edition, 2007.
Compressed Gas Association. Handbook of Compressed Gases. Springer, New York, 4th
edition, 1999.
Cooke, D. and Rohleder, T. Inventory evaluation and product slate management in large-
scale continuous process industries. Journal of Operations Management, 24(3):235–249,
2006.
Croom, S., Romano, P., and Giannakis, M. Supply chain management: an analytical
framework for critical literature review. European Journal of Purchasing & Supply
Management, 6(1):67–83, 2000.
Cryer, J. and Chan, K.-S. Time Series Analysis - With Applications in R. Springer, New
York, 2nd edition, 2008.
Darby, M., Harmse, M., and Nikolaou, M. Mpc: current practice and challenges. In
Symposium Preprints of IFAC Symposium on Advanced Control of Chemical Processes
ADCHEM 2009, Part 1, Istanbul, Turkey, pages 88–100, 2009.
212
De Berg, M., Cheong, O., and Van Kreveld, M. Computational geometry: algorithms and
applications. Springer, New York, 2008.
Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. A fast and elitist multiobjective
genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2):
182–197, 2002.
Deb, K., Rao N, U., and Karthik, S. Dynamic multi-objective optimization and decision-
making using modified nsga-ii: a case study on hydro-thermal power scheduling. In
Evolutionary Multi-Criterion Optimization, pages 803–817. Springer, New York, 2007.
Degnan, T. et al. Alkylation of aromatics with ethylene and propylene: recent develop-
ments in commercial processes. Applied catalysis A: general, 221(1-2):283–294, 2001.
Delaunay, B. Sur la sphere vide. Izvestija Akademii Nauk SSSR, Otdelenie Matematich-
eskii i Estestvennyka Nauk, 7(793-800):1–2, 1934.
Dobson, G. The cyclic lot scheduling problem with sequence-dependent setups. Operations
Research, 40(4):736–749, 1992.
Dong, J.-X. and Song, D.-P. Container fleet sizing and empty repositioning in liner ship-
ping systems. Transportation Research Part E: Logistics and Transportation Review,
45(6):860–877, 2009.
Duan, J. Augmented GARCH(p, q) process and its diffusion limit. Journal of Economet-
rics, 79(1):97–127, 1997.
Engle, R. New frontiers for ARCH models. Journal of Applied Econometrics, 17(5):
425–446, 2002.
213
Faraway, J. Extending the linear model with R: generalized linear, mixed effects and
nonparametric regression models. Chapman & Hall/CRC, London, 2005.
Fonseca, D., Navaresse, D., and Moynihan, G. Simulation metamodeling through artifi-
cial neural networks. Engineering Applications of Artificial Intelligence, 16(3):177–183,
2003.
Fornari, F. and Mele, A. Recovering the probability density function of asset prices using
GARCH as diffusion approximations. Journal of Empirical Finance, 8(1):83–110, 2001.
Francq, C. and Zakoian, J. Maximum likelihood estimation of pure garch and arma-garch
processes. Bernoulli, 10(4):605–637, 2004.
Fu, M., Glover, F., and April, J. Simulation optimization: a review, new developments,
and applications. In Proceedings of the Winter Simulation Conference, pages 83–95,
2005.
Gjerdrum, J., Shah, N., and Papageorgiou, L. G. Fair transfer price and inventory holding
policies in two-enterprise supply chains. European Journal of Operational Research, 143
(3):582–599, 2002.
Goel, T., Vaidyanathan, R., Haftka, R., Shyy, W., Queipo, N., and Tucker, K. Re-
sponse surface approximation of pareto optimal front in multi-objective optimization.
Computer Methods in applied mechanics and engineering, 196(4):879–893, 2007.
Grunow, M., Günther, H.-O., and Lehmann, M. Campaign planning for multi-stage batch
processes in the chemical industry. OR Spectrum, 24(3):281–314, 2002.
Guillén, G., Badell, M., and Puigjaner, L. A holistic framework for short-term supply
chain management integrating production and corporate financial planning. Interna-
tional Journal of Production Economics, 106(1):288–306, 2007.
Gunasekaran, A., Patel, C., and Tirtiroglu, E. Performance measures and metrics in a
supply chain environment. International Journal of Operations & Production Manage-
ment, 21(1/2):71–87, 2001.
Gunasekaran, A., Patel, C., and McGaughey, R. E. A framework for supply chain per-
formance measurement. International journal of production economics, 87(3):333–347,
2004.
Haessler, R. An improved extended basic period procedure for solving the economic lot
scheduling problem. AIIE transactions, 11(4):336–340, 1979.
Hall, A. and Nicholls, D. The evaluation of exact maximum likelihood estimates for varma
models. Journal of Statistical Computation and Simulation, 10(3-4):251–262, 1980.
Hansen, P. and Lunde, A. A forecast comparison of volatility models: does anything beat
a GARCH (1, 1)? Journal of Applied Econometrics, 20(7):873–889, 2005.
Hayler, W. B., Keever, J. M., and Seiler, P. M. American Merchant Seaman’s Manual.
Schiffer Publishing, Atglen, PA, 7th edition, 2003.
Hsu, W. On the general feasibility test of scheduling lot sizes for several products on one
machine. Management Science, 29(1):93–105, 1983.
Hughes, R. and Powell, W. Mitigating end effects in the dynamic vehicle allocation model.
Management Science, pages 859–879, 1988.
216
Hull, B. Oil pipeline markets and operations. In Journal of the Transportation Research
Forum, volume 44, page 111. Transportation research forum, 2005.
Hurvich, C. and Tsai, C. Regression and time series model selection in small samples.
Biometrika, 76(2):297–307, 1989.
Iacus, S. Simulation and inference for stochastic differential equations: with R examples.
Springer, Berlin, 2008.
Ibe, O. Markov Processes for Stochastic Modeling. Academic Press, Waltham, MA, 2008.
Jones, R. and Bunnett, J. Nomenclature for organic chemical transformations. Pure and
Applied Chemistry, 61:725–768, 1989.
Jung, J., Blau, G., Pekny, J., Reklaitis, G., and Eversdyk, D. A simulation based opti-
mization approach to supply chain management under demand uncertainty. Computers
& Chemical Engineering, 28(10):2087–2106, 2004.
Jung, J. Y., Blau, G., Reklaitis, G., and Eversdyke, D. Integrated safety stock manage-
ment for multi-stage supply chains under producttion capacity constraints. Computers
& Chemical Engineering, 32:2570–2581, 2008.
Kallrath, J. Solving planning and design problems in the process industry using mixed
integer and global optimization. Annals of Operations Research, 140(1):339–373, 2005.
Kaplan, R. S., Norton, D. P., et al. The balanced scorecard – measures that drive perfor-
mance. Harvard Business Review, 70(1):71–79, 1992.
Kellermann, C. Rohstoffpreise ziehen seit jahresbeginn wieder an. In chemie report. VCI,
2012. URL https://www.vci.de/Downloads/Publikation/chemie-report/cr2012_
03.pdf. access: 28.10.2012.
Kilger, C. and Meyr, H. Demand fulfilment and atp. In Stadtler, H. and Kilger, C.,
editors, Supply Chain Management and Advanced Planning, pages 181–198. Springer,
Berlin, 2008.
217
Kim, Y., Yun, C., Park, S. B., Park, S., and Fan, L. An integrated model of supply
network and production planning for multiple fuel products of multi-site refineries.
Computers & Chemical Engineering, 32(11):2529 – 2535, 2008.
Kleijnen, J. Supply chain simulation tools and techniques: a survey. International Journal
of Simulation and Process Modelling, 1(1):82–89, 2005.
Klibi, W., Martel, A., and Guitouni, A. The design of robust value-creating supply chain
networks: a critical review. European Journal of Operational Research, 203(2):283–293,
2010.
Koretsky, M. Engineering and chemical thermodynamics. J. Wiley & Sons, Hoboken, NJ,
2004.
Laguna, M. and Martí, R. The optquest callable library. In Voß, S., Woodruff, D. L., and
Sharda, R., editors, Optimization Software Class Libraries, volume 18 of Operations Re-
search/Computer Science Interfaces Series, pages 193–218. Springer, New York, 2003.
Law, A. Simulation Modeling and Analysis. Mcgraw-Hill, Tucson, AZ, 4th edition, 2007.
Lewin, D., Seider, W., and Seader, J. Integrated process design instruction. Computers
& chemical engineering, 26(2):295–306, 2002.
Li, W., Ling, S., and McAleer, M. Recent theoretical results for time series models with
GARCH errors. Journal of Economic Surveys, 16(3):245–269, 2002.
Li, W., Hui, C.-W., and Li, A. Integrating cdu, fcc and product blending models into
refinery planning. Computers & Chemical Engineering, 29(9):2010–2028, 2005.
Li, X., Zhang, C., Gao, L., Li, W., and Shao, X. An agent-based approach for integrated
process planning and scheduling. Expert Systems with Applications, 37(2):1256–1264,
2010a.
Li, Y., Ng, S., Xie, M., and Goh, T. A systematic comparison of metamodeling techniques
for simulation optimization in decision support systems. Applied Soft Computing, 10
(4):1257–1273, 2010b.
Ma, L. and Trimm, D. Alternative catalyst bed configurations for the autothermic con-
version of methane to hydrogen. Applied Catalysis A: General, 138(2):265–273, 1996.
Magatão, L., Arruda, L., and Neves, F. A mixed integer programming approach for
scheduling commodities in a pipeline. Computers & Chemical Engineering, 28(1–2):171
– 185, 2004.
Magatão, L., Arruda, L., and Neves, F. Using CLP and MILP for scheduling commodities
in a pipeline. Computer Aided Chemical Engineering, 20:1027–1032, 2005.
Maronna, R., Martin, D., and Yohai, V. Robust statistics: Theory and methods. J. Wiley
& Sons, Hoboken, NJ, 2006.
Más, R. and Pinto, J. A mixed-integer optimization strategy for oil supply in distribution
complexes. Optimization and Engineering, 4(1):23–64, 2003.
Matko, D., Geiger, G., and Gregoritza, W. Pipeline simulation techniques. Mathematics
and computers in simulation, 52(3):211–230, 2000.
Mele, F. D., Guillén, G., Espuña, A., and Puigjaner, L. A simulation-based optimization
framework for parameter optimization of supply-chain networks. Industrial & Engi-
neering Chemistry Research, 45(9):3133–3148, 2006.
Melo, M., Nickel, S., and da Gama, F. S. Facility location and supply chain management
- a review. European Journal of Operational Research, 196(2):401–412, 2009.
Méndez, C., Cérda, J., Grossmann, I., Harjunkoski, I., and Fahl, M. State-of-the-art
review of optimization methods for short-term scheduling of batch processes. Computers
& Chemical Engineering, 30:913–946, 2006.
Meyr, H., Wagner, M., and Rohde, J. Structure of advanced planning systems. In Stadtler,
H. and Kilger, C., editors, Supply Chain Management and Advanced Planning, pages
109–115. Springer, Berlin, 2008.
Mishra, S., Zope, V., and Goje, A. Kinetic and thermodynamic studies of depolymerisa-
tion of poly (ethylene terephthalate) by saponification reaction. Polymer international,
51(12):1310–1315, 2002.
Moura, A., de Souza, C., Cire, A., and Lopes, T. Planning and scheduling the operation of
a very large oil pipeline network. In Principles and Practice of Constraint Programming,
pages 36–51. Springer, New York, 2008.
220
Mula, J., Peidro, D., Díaz-Madroñero, M., and Vicens, E. Mathematical programming
models for supply chain production and transport planning. European Journal of Op-
erational Research, 204(3):377 – 390, 2010.
Narro Lopez, M. and Kingsman, B. The economic lot scheduling problem: theory and
practice. International Journal of Production Economics, 23(1-3):147–164, 1991.
Neiro, S. and Pinto, J. A general modeling framework for the operational planning of
petroleum supply chains. Computers & Chemical Engineering, 28:871–896, 2004.
Neumann, K., Schwindt, C., and Trautmann, N. Advanced production scheduling for
batch plants in process industries. OR Spectrum, 24(3):251–279, 2002.
Newman, A. and Yano, C. Scheduling direct and indirect trains and containers in an
intermodal setting. Transportation Science, 34(3):256–270, 2000.
Ogata, K. System dynamics. Prentice Hall, Upper Saddle River, NJ, 4th edition, 2003.
Oh, H.-C. and Karimi, I. Regulatory factors and capacity-expansion planning in global
chemical supply chains. Industrial & Engineering Chemistry Research, 43(13):3364–
3380, 2004.
Papageorgiou, L. Supply chain optimisation for the process industries: Advances and
opportunities. Computers & Chemical Engineering, 33(12):1931–1938, 2009.
Park, M., Park, S., Mele, F. D., and Grossmann, I. E. Modeling of purchase and sales
contracts in supply chain optimization. Industrial & Engineering Chemistry Research,
45(14):5013–5026, 2006.
Pearl, J. Causality: models, reasoning and inference. Cambridge University Press, Cam-
bridge, 2000.
221
Peidro, D., Mula, J., Poler, R., and Lario, F.-C. Quantitative models for supply chain
planning under uncertainty: a review. The International Journal of Advanced Manu-
facturing Technology, 43:400–420, 2009.
Pfaff, B. Analysis of Integrated and Cointegrated Time Series with R. Springer, Berlin,
2006.
Pfaff, B. Var, svar and svec models: Implementation within r package vars. Journal of
Statistical Software, 27(4):1–32, 2008.
Pinto, J., Joly, M., and Moro, L. Planning and scheduling models for refinery operations.
Computers & Chemical Engineering, 24:2259–2276, 2000.
Pitty, S., Li, W., Adhitya, A., Srinivasan, R., and Karimi, I. Decision support for in-
tegrated refinery supply chains:: Part 1. dynamic simulation. Computers & Chemical
Engineering, 32(11):2767–2786, 2008.
Power, D. J. Decision Support Systems: Concepts and Resources for Managers. Green-
wood Publishing Group, Westport, CT, 2002.
Qi, Z., Sundmacher, K., Stein, E., Kienle, A., and Kolah, A. Reactive separation of
isobutene from c4 crack fractions by catalytic distillation processes. Separation and
Purification Technology, 26(2-3):147–163, 2002.
Rabe, M., Spieckermann, S., and Wenzel, S. Verifikation und Validierung für die Simu-
lation in Produktion und Logistik: Vorgehensmodelle und Techniken. Springer, Berlin,
2008.
222
Rabelo, L., Eskandari, H., Shaalan, T., and Helal, M. Value chain analysis using hybrid
simulation and ahp. International Journal of Production Economics, 105(2):536–547,
2007.
Rejowski, R. and Pinto, J. A novel continuous time representation for the scheduling of
pipeline systems with pumping yield rate constraints. Computers & Chemical Engi-
neering, 32(4):1042–1066, 2008.
Relvas, S., Matos, H., Barbosa-Póvoa, A., Fialho, J., and Pinheiro, A. Pipeline scheduling
and inventory management of a multiproduct distribution oil system. Industrial &
engineering chemistry research, 45(23):7841–7855, 2006.
Relvas, S., Matos, H., Barbosa-Póvoa, A., and Fialho, J. Reactive scheduling frame-
work for a multiproduct pipeline with inventory management. Industrial & engineering
chemistry research, 46(17):5659–5672, 2007.
Relvas, S., Barbosa-Póvoa, A., and Matos, H. Heuristic batch sequencing on a multi-
product oil distribution system. Computers & Chemical Engineering, 33(3):712–730,
2009.
Romero, J., Badell, M., Bagajewicz, M., and Puigjaner, L. Integrating budgeting models
into scheduling and planning models for the chemical batch industry. Industrial &
Engineering Chemistry Research, 42(24):6125–6134, 2003.
Ronen, D. Ship scheduling: The last decade. European Journal of Operational Research,
71(3):325–333, 1993.
Rousseeuw, P. and van Driessen, K. A fast algorithm for the minimum covariance deter-
minant estimator. Technometrics, 41:212–223, 1999.
Ryu, J.-H., Dua, V., and Pistikopoulos, E. N. A bilevel programming framework for
enterprise-wide process networks under uncertainty. Computers & Chemical Engineer-
ing, 28(6):1121–1129, 2004.
223
Saharidis, G. K., Minoux, M., and Dallery, Y. Scheduling of loading and unloading
of crude oil in a refinery using event-based discrete time formulation. Computers &
Chemical Engineering, 33(8):1413 – 1426, 2009.
Sandler, S. Chemical, biochemical, and engineering thermodynamics. John Wiley & Sons,
Hoboken, NJ, 2006.
Santner, T., Williams, B., and Notz, W. The design and analysis of computer experiments.
Springer, New York, 2003.
Schneider, H., Rinks, D. B., and Kelle, P. Power approximations for a two-echelon invetory
system using service levels. Production and Operations Management, 4(4):381–400,
1995.
Seader, J., Henley, E., and Roper, D. Separation Process Principles. J. Wiley & Sons,
Hoboken, NJ, 3rd edition, 2011.
224
Seager, S. and Slabaugh, M. Chemistry for today: General, organic, and biochemistry.
Brooks/Cole Publishing Company, Stamfort, CT, 2007.
Seppala, C., Harris, T., and Bacon, D. Time series methods for dynamic analysis of
multiple controlled variables. Journal of Process Control, 12(2):257–276, 2002.
Shah, N. Process industry supply chains: Advances and challenges. Computers & Chem-
ical Engineering, 29:1225–1235, 2005.
Shang, J. S., Li, S., and Tadikamalla, P. Operational design of a supply chain system
using the taguchi method, response surface methodology, simulation, and optimization.
International Journal of Production Research, 42(18):3823–3849, 2004.
Shao, X., Li, X., Gao, L., and Zhang, C. Integration of process planning and scheduling
- a modified genetic algorithm-based approach. Computers & Operations Research, 36:
2082–2096, 2009.
Shim, J., Warkentin, M., Courtney, J. F., Power, D. J., Sharda, R., and Carlsson, C.
Past, present, and future of decision support technology. Decision support systems, 33
(2):111–126, 2002.
Shin, M., Sargent, R., and Goel, A. Gaussian radial basis functions for simulation meta-
modeling. In Proceedings of the Winter Simulation Conference, 2002, volume 1, pages
483–488, 2002.
Shumway, R. and Stoffer, D. Time Series Analysis and Its Applications - With R Exam-
ples. Springer, New York, 2nd edition, 2006.
Siswanto, N., Essam, D., and Sarker, R. Solving the ship inventory routing and scheduling
problem with undedicated compartments. Computers & Industrial Engineering, 61(2):
289–299, 2011.
Smith, R. Chemical process design and integration. J. Wiley & Sons, Hoboken, NJ, 2005.
Sürie, C. and Wagner, M. Supply chain analysis. In Stadtler, H. and Kilger, C., editors,
Supply Chain Management and Advanced Planning, pages 37–63. Springer, Berlin, 2008.
Stålhane, M., Rakke, J., Moe, C., Andersson, H., Christiansen, M., and Fagerholt, K. A
construction and improvement heuristic for a liquefied natural gas inventory routing
problem. Computers & Industrial Engineering, 62(1):245–255, 2012.
SteadieSeifi, M., Dellaert, N., Nuijten, W., Woensel, T. V., and Raoufi, R. Multimodal
freight transportation planning: A literature review. European Journal of Operational
Research, 233(1):1 – 15, 2014.
Stopford, M. Maritime Economics. Taylor & Francis, New York, 3rd edition, 2009.
Sun, H. N. and Wristers, J. P. Butadiene. John Wiley & Sons, Hoboken, NJ, 2006.
Taleb, N. Black swans and the domains of statistics. The American Statistician, 61(3):
198–200, 2007.
Trautmann, H., Steuer, D., and Mersmann, O. mco: Multi criteria optimization algo-
rithms and related functions, 2010. URL http://CRAN.R-project.org/package=mco.
R package version 1.0.9.
Tsay, R. Analysis of Financial Time Series. J. Wiley & Sons, Hoboken, NJ, 2002.
226
Tsiakis, P., Shah, N., and Pantelides, C. C. Design of multi-echelon supply chain networks
under demand uncertainty. Industrial & Engineering Chemistry Research, 40(16):3585–
3604, 2001.
Uhlenbeck, G. and Ornstein, L. On the theory of the Brownian motion. Physical Review,
36(5):823, 1930.
Van der Vorst, J. and Beulens, A. Identifying sources of uncertainty to generate supply
chain redesign strategies. International Journal of Physical Distribution & Logistics
Management, 32(6):409–430, 2002.
van Essen, H., Croezen, H., and Nielsen, J. Emissions of pipeline transport compared
with those of competing modes, 2003. Delft, Association of Petrochemicals Producers
in Europe.
Venkateswaran, J., Son, Y., and Jones, A. Hierarchical production planning using a
hybrid system dynamic-discrete event simulation architecture. In Proceedings of the
2004 Winter Simulation Conference, volume 2, pages 1094–1102, 2004.
Vidal, C. J. and Goetschalckx, M. A global supply chain model with transfer pricing
and transportation cost allocation. European Journal of Operational Research, 129(1):
134–158, 2001.
Viswanathan, J. and Grossmann, I. Optimal feed locations and number of trays for
distillation columns with multiple feeds. Industrial & engineering chemistry research,
32(11):2942–2949, 1993.
Vora, B. V., Kocal, J. A., Barger, P. T., Schmidt, R. J., and Johnson, J. A. Alkylation.
John Wiley & Sons, Hoboken, NJ, 2006.
227
Wagner, B. and Davis, D. A search heuristic for the sequence-dependent economic lot
scheduling problem. European Journal of Operational Research, 141(1):133–146, 2002.
Wilson, B., Cappelleri, D., Simpson, T., and Frecker, M. Efficient pareto frontier ex-
ploration using surrogate approximations. Optimization and Engineering, 2(1):31–50,
2001.
Yao, Q. and Brockwell, P. Gaussian maximum likelihood estimation for arma models. i.
time series. Journal of time series analysis, 27(6):857–875, 2006.
Yeomans, H. and Grossmann, I. Optimal design of complex distillation columns using rig-
orous tray-by-tray disjunctive programming models. Industrial & engineering chemistry
research, 39(11):4326–4335, 2000.
Zhang, J., Zhu, X., and Towler, G. A simultaneous optimization strategy for overall
integration in refinery planning. Industrial & Engineering Chemistry Research, 40(12):
2640–2653, 2001.
228
Zoryk-Schalla, A., Fransoo, J., and de Kok, T. Modeling the planning process in advanced
planning systems. Informations and Management, 42:75–87, 2004.
229
Appendix
0.15
0.05
0.05
ACF
ACF
−0.05
−0.05
−0.15
−0.15
5 10 15 20 5 10 15 20
Lag Lag
(a) ACF of residuals (b) ACF of squared residual
0.006
41.60
−0.002
41.58
41.56
−0.006
−2 −1 0 1 2 0 50 100 150
Theoretical Quantiles Time
(c) QQ-plot of residuals (d) Fitted and original Naphtha time series
Figure A.1: Diagnostic plots for the ARX(1) model of the Naphtha time series
VARX order
p 1 2 3 4 5 6
AIC(p) -3.81 -3.95 -4.22 -4.24 -4.23 -4.20
SIC(p) -3.72 -3.81 -4.04 -4.01 -3.95 -3.88
HQIC(p) -3.77 -3.89 -4.15 -4.15 -4.12 -4.08
Table A.1: Information criteria for VARX models of order 1 to 6 for the de-alkylation
plant
T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,
DOI 10.1007/978-3-658-08433-2, © Springer Fachmedien Wiesbaden 2015
230
Table A.3: ANOVA table for the initial V ARX(3) model (without outlier correction)
231
Φ̂11
1 – – –
Φ̂12
1 2.51 0.628 0.00∗∗∗
Φ̂21
1 0.04 0.002 0.00∗∗∗
Φ̂22
1 0.85 0.043 0.00∗∗∗
Φ̂11
2 – – –
Φ̂12
2 – – –
Φ̂21
2 0.01 0.002 0.00∗∗∗
Φ̂22
2 -0.25 0.000 0.00∗∗∗
Φ̂11
3 -0.15 0.045 0.00∗∗∗
Φ̂12
3 -2.67 0.487 0.00∗∗∗
Φ̂21
3 -0.02 0.002 0.00∗∗∗
Φ̂22
3 0.18 0.033 0.00∗∗∗
υ̂11 1.05 0.041 0.00∗∗∗
υ̂12 -0.03 0.003 0.00∗∗∗
υ̂21 -0.00028 0.0002 0.01∗
υ̂22 -0.00003 0.000 0.00∗∗∗
dummy.out.20.benz -4.14 1.414 0.00∗∗∗
dummy.out.20.c9 -0.15 0.063 0.02∗
dummy.out.51.benz -5.98 1.413 0.00∗∗∗
dummy.out.51.c9 – – –
dummy.out.61.benz 4.02 1.411 0.00∗∗∗
dummy.out.61.c9 0.16 0.063 0.01∗∗
dummy.out.64.benz -2.74 1.418 0.05∗
dummy.out.64.c9 -0.17 0.063 0.01∗∗
dummy.out.65.benz -5.28 1.420 0.00∗∗∗
dummy.out.65.c9 -0.17 0.064 0.01∗∗
dummy.out.109.benz 2.96 1.426 0.04∗
dummy.out.109.c9 -0.19 0.064 0.00∗∗∗
dummy.out.155.benz 4.63 1.422 0.00∗∗∗
dummy.out.155.c9 – – –
dummy.out.156.benz 2.74 1.416 0.05∗
dummy.out.156.c9 -0.20 0.064 0.00∗∗∗
dummy.out.158.benz -6.43 1.433 0.00∗∗∗
dummy.out.158.c9 0.16 0.064 0.01∗∗
dummy.out.188.benz -3.81 1.413 0.01∗∗
dummy.out.188.c9 -0.19 0.063 0.00∗∗∗
232
Table A.4: ANOVA table for the initial V ARX(3) model with outlier correction and
variable selection
233
0.3
Benzene residuals
0.2
0.1
autocorrelation
0.0
−0.1
−0.2
0.3 −0.3
C9 residuals
0.2
0.1
autocorrelation
Benzene residuals
0.0
Sample Quantiles
−0.1
2
−0.2
0
−0.3
−2
0 5 10 15 20 25
Lag
0.3
C9 residuals
0.2
Sample Quantiles
0.05
crosscorrelation
0.1
0.0
−0.05
−0.1
−0.15
−0.2
−0.3
−3 −2 −1 0 1 2 3
−20 −10 0 10 20 Normal Quantiles
Lag
(a) Autocorrelation and cross-correlation for residu- (b) QQ-plots of residuals’ marginal distributions for
als of both outflow rates both outflow rates
Figure A.2: Residual diagnostic plots for V ARX(3) model with outlier compensation.
234
Simulation model
Data for conceptual model
site 1 site 2
Cr. Hy. EBS Al. PE PS Cr. Hy. EBS B.Ex. PP SBR PE
Naph. -3,000 -2,000
C1 600 400
C2 1050 -50 -1,200 700 -120 -380
C3 450 -195 300 -600
C4 300 200 -600
C5+ 600 -600 400 -400
C9+ 60 40
Buta. 240 -240
Raff. 360
Benz. 540 -150 -390 360 -360
Styr. 200 -600 480 -80
Cum. 585
PE 1,200 380
PS 600
PP 600
SBR 320
Table A.5: Stationary flow rates (in t/h) for exemplary chemical production network
Φ1 Φ2 β
⎛ 0.50 0.20 0.10 0.00 0.00 ⎞ ⎛ 0.30 0.10 0.00 0.00 0.00 ⎞ ⎛ 0.20 ⎞
⎜ 0.20 0.40 0.30 0.00 0.00 ⎟ ⎜ 0.10 −0.10 0.10 0.00 0.00 ⎟ ⎜ 0.35 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ 0.00 −0.30 0.50 0.30 0.00 ⎟ ⎜ 0.00 0.10 −0.20 −0.20 0.00 ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜ 0.15 ⎟
⎜ 0.20 −0.60 0.30 ⎟ ⎜ ⎟ ⎜ ⎟
⎜ 0.00 0.00 ⎟ ⎜ 0.00 0.00 0.10 0.10 0.10 ⎟ ⎜ 0.10 ⎟
⎝ 0.00 0.00 0.00 −0.10 0.40 ⎠ ⎝ 0.00 0.00 0.00 0.20 −0.20 ⎠ ⎝ 0.20 ⎠
coefficients for input time series: φ = (−0.7, −0.3, −0.15)
Table A.6: Coefficients of the time series models for cracker at site 1 (V ARX(2)+AR(3))
Φ1 Φ2 β
⎛ 0.50 0.20 −0.20 0.00 0.00 ⎞ ⎛ −0.20 0.00 0.00 0.00 0.00 ⎞ ⎛ 0.20 ⎞
⎜ 0.30 0.20 0.40 0.00 0.00 ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ −0.10 0.10 0.10 0.00 0.00 ⎟ ⎜ 0.35 ⎟
⎜ 0.00 −0.20 0.40 0.20 0.00 ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ 0.00 −0.10 −0.10 0.10 0.00 ⎟ ⎜ 0.15 ⎟
⎜ 0.00 0.00 0.20 −0.40 0.20 ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ 0.00 0.00 0.10 0.10 0.10 ⎟ ⎜ 0.10 ⎟
⎝ 0.00 0.00 0.00 −0.20 0.20 ⎠ ⎝ 0.00 0.00 0.00 0.10 −0.10 ⎠ ⎝ 0.20 ⎠
coefficients for input time series: φ = (0.5, −0.2, 0.3)
Table A.7: Coefficients of the time series models for cracker at site 2 (V ARX(2)+AR(3))
235
Φ1 Φ2 Φ3 β
−0.40 0.10 −0.10 −0.20 0.20 0.00 0.90
( ) ( ) ( ) ( )
0.20 0.30 −0.10 0.15 0.10 0.05 0.10
coefficients for input time series: φ = (−0.4, 0.1)
Table A.8: Coefficients of the time series models for hydrogenation plant at site 1
(V ARX(3) + AR(2))
Φ1 Φ2 Φ3 β
0.40 0.30 0.10 −0.20 −0.20 0.00 0.90
( ) ( ) ( ) ( )
0.20 0.30 0.10 −0.15 0.00 0.00 0.10
coefficients for input time series: φ = (−0.6, 0.2, 0.1)
Table A.9: Coefficients of the time series models for hydrogenation plant at site 2
(V ARX(3) + AR(3))
Φ1 β
0.70 −0.30 0.40
( ) ( )
0.20 0.50 0.60
coefficients for input time series:
φ = (0.5, 0.3, −0.25, −0.4)
Table A.10: Coefficients of the time series models for Butex plant at site 2 (V ARX(1) +
AR(4))
Table A.11: Coefficients of input and output time series models of SISO and MISO plants
236
C2 C3 C4 Styrene
# un-/loading arms 2 3 4 3
# un-/loaded rail cars 10 15 20 15
un-/loading capacity (in t) 500 750 1,000 750
E−Ini
site 1 20 20 40 40
ris
site 2 30 30 30 30
site 1 10,000 4,000 1,000 6,000
sTar
is
Table A.12: Un-/loading capacities, initial stock of empty RTCs, target and initial stock
levels per transported chemical at both production sites
site 1 site 2
PE PS Cum. C4 PE Raff. PP SBR
inter-arrival time E(2.5) E(1.25) E(0.488) E(1/12) E(0.8) E(3/10) E(1.25) E(2/3)
order quantity N(20, 5) N(20, 5) N(50, 5) N(50, 2) N(20, 5) N(50, 2) N(20, 5) N(20, 5)
Table A.14: Distributions of inter-arrival time and deliver quantities for external cus-
tomers/suppliers
600 0 200 0
600 0.98 0.02 200 0.99 0.01
Q1Hyd. = ( ) Q1EBSM = ( )
0 0.25 0.75 0 0.15 0.85
585 0 600 0
585 0.97 0.03 600 0.94 0.06
Q1Alkyl. = ( ) Q1P S = ( )
0 0.20 0.80 0 0.30 0.70
2, 000 1, 000 0
400 0 480 0
2, 000 ⎛ 0.98 0.01 0.01 ⎞
400 0.995 0.005 480 0.98 0.02
Q2Cracker = 1, 000 ⎜ 0.20 0.75 0.05 ⎟ Q2Hyd. = ( ) Q2EBSM = ( )
⎝ 0.08 0 0.30 0.70 0 0.22 0.88
0 0.02 0.90 ⎠
600 0 600 0
600 0.99 0.01 600 0.97 0.03
Q2But.Ex. = ( ) Q2P P = ( )
0 0.25 0.75 0 0.15 0.85
320 0 380 0
320 0.975 0.025 380 0.95 0.05
Q2SBR = ( ) Q2P E = ( )
0 0.2 0.8 0 0.3 0.7
Planning of experiments
Experimental designs
Table A.17: Resolution V design for 6 dichotomous variables ("-1" encodes the variable’s
lower level and "1" the upper level)
239
Table A.18: Estimated responses for all possible configurations (dominated configuration
gray coloured)
240
●
●
●
●●●
3
●
●●
●
●●●●
●●
●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●●
●●
●
2
●
●
●
●
●
sample quantiles
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
1
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
0
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
−1
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●●
●
●●
●●
●●
●
●●●●
●●
●●●
●●●●
● ●
−2
−2 0 2
quantiles of normal distribution
(a) QQ-plot of residuals for (4.19)
40
●
●
●● ●
●●●
●
●●
●●●●●
●
●●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
20
●●
●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
sample quantiles
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
0
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
−20
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●●
●●●●
●
●●
−40
●●
●
●
−2 0 2
quantiles of normal distribution
(b) QQ-plot of residuals for (4.20)
●
●
0.2
●
●●
●●
●●
●●
●●●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●●
0.1
●
●
●●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
sample quantiles
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●●
0.0
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
−0.1
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
−0.2
●●
●
●
●
●●
●●
●
●●
●●
●
●●●
●
●●●●●
●
●
●
−2 0 2
quantiles of normal distribution
(c) QQ-plot of residuals for (4.21)
Simulation optimization
k x (k−1) −(x/λ)k
f (x) = ⋅( ) ⋅e . (A.1)
λ λ
0.0015
0.0010
probability
0.0005
0.0000
Figure A.4: Density function of the Weibull distribution with k = 1.5 and λ = 365
Γ( 53 )
242
V−1⎛⎝E(X) ⋅ (0.05)⎞⎠
expected loss V(r) (in t)
2000 4000 6000 8000
E(X) ⋅ (0.05)
0
V−1⎛⎝E(X) ⋅ (0.05)⎞⎠
7000
expected loss V(r) (in t)
5000
3000
1000
E(X) ⋅ (0.05)
0
Figure A.5: Loss functions for both sites during pipeline inspection
243
models)
2 generate production, flow, order & inventory matrices
3 initialize inventory matrices
4 for i ∈ {1, 2} ∧ t = 1 → T do
5 generate ωitplan ⊳ generate Markov chain samples
6 end
7 talarm ← W B(1.5, 365/Γ(5/3)) ⊳ sample pipeline inspection
8 tinsp ← 0 ⊳ initial inspection time
9 for t = 1 → T do
10 for all locations i ∈ {1, 2, h} do
11 if lit + ∑Tτ=t qiτ ≤ si then
12 if i == h then
13 θ ← 2 + W B(2, 5/Γ(1.5)) ⊳ sample order lead time
14 qi(t+θ) ← qi(t+θ) + Si − lit − ∑Tτ=t qiτ ⊳ assign order
15 else
16 qit ← qit + Si − lit − ∑Tτ=t qiτ ⊳ assign order
17 end
18 end
19 end
20 if t == talarm then
21 tinsp ← t + 1 + W B(1.5, 3/Γ(5/3)) ⊳ sample inspection time
22 talarm ← t + W B(1.5, 365/Γ(5/3)) ⊳ sample next inspection
23 end
24 for i ∈ {h, 1, 2} do
25 if i == h then
26 fit ← min (qit , licap − li(t−1) , ρcap
out ) ⊳ calc. unloading flow
27 else
28 if t > tinsp then
29 in ) ⋅ ∑j∈{1,2}q )
fit ← min (qit , licap − li(t−1) , min (lh(t−1) + fht , ρcap qit
jt
30 else
31 fit ← 0 ⊳ omit supply from harbour
32 end
33 ωitreal ← min (ωitplan , li(t−1)+fit ) ⊳ assign realized consumption
34 lit ← li(t−1) − ωitreal + fit ⊳ update inventory
35 end
36 qi(t+1) ← qi(t+1) + qit − fit ⊳ update outstanding orders
37 end
38 lht ← lh(t−1 + fht − ∑j∈{1,2} fjt ⊳ update harbour inventory
39 end
40 end
Table A.20: Summary for the logistic regression model of responses of efficient settings
245
●
● ●
●
●● ●
●●●
●●●●●
●●●●
●●●
●●●●
●●
●●
4
●●
●●
●●
●●
●
●●
●●
●●
●
●●
●●
●
●●
●●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●●
●
●●
●●
quantiles of residuals
●
●
●●
●●
●
●
●●
●
●
●●
●
●
2
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●●
0
●
●●
●
●
●●
●
●●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
−2
●
●
●●
●
●●
●●
●●
●
●
●●
●
●●
●●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
−4
●●●
●●
●
●●●
●●●●
●●●●
●●●●
● ●●●
●
●
−6
● ●
●
●
−3 −2 −1 0 1 2 3
quantiles of normal distribution
(a) QQ-plot of residuals for (4.22)
6
●
● ●
●
● ● ● ●
● ● ●
● ● ● ●
●
●●
●
● ● ●● ●
● ● ●● ●
4
● ●● ● ● ●
● ● ● ● ●
● ● ● ● ●
● ● ● ●●
● ● ●● ● ● ● ●
● ●
● ● ● ● ● ●●
● ● ●●
● ● ●● ● ● ●
●●● ● ●●
●● ● ● ● ●
● ●● ● ●
● ●● ●● ● ● ● ●
● ● ●● ● ● ●
● ● ●
● ●● ● ●●● ● ●
● ● ●● ● ● ● ● ●
●
● ●● ● ● ●
● ●● ● ●
2
●● ●●●●
●● ● ●● ● ●● ● ● ● ● ●●
●●●● ● ●
● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ●
● ●
●●● ● ● ●● ● ● ● ● ● ●
● ● ● ● ● ● ●● ●
● ● ● ● ●● ●● ● ● ● ●● ●● ● ● ●
●● ● ●● ●●● ●● ● ●●● ●● ●● ●● ● ●● ●
residuals (εi)
●● ● ● ● ●
● ● ●● ●●● ● ● ● ●● ● ● ●●● ●
● ●● ● ● ● ●
●
●●●●●● ● ● ● ●
●●
● ● ● ● ● ● ●
● ● ●● ● ●●●
● ● ● ● ●● ● ● ● ● ● ● ● ● ●
●
●● ● ●●● ●
●
●● ● ● ●● ● ● ● ●● ●●● ●● ●●
●●●●
●● ●
● ●● ●
● ● ● ● ●
● ● ● ● ●●
●●● ●●● ●●● ● ●
●●●● ● ● ● ● ● ● ●
●●●● ●●● ● ● ● ● ●●● ● ●● ● ● ●● ● ● ●● ● ● ● ●
● ●● ●●
● ●●●●●● ● ● ● ●
●● ●
●● ●● ●● ● ● ● ●
● ●●● ● ●
● ●
● ●
●
● ● ● ● ●●●●●●●●●● ● ● ●● ● ●
●●● ● ●
●
●● ● ●● ● ● ●● ●●
● ●●
0
● ●● ●● ●●●● ●● ●
● ●● ● ● ●
●
● ●
●● ●●●● ● ● ● ●
● ● ●
●
● ●●●●●●●
● ● ● ● ●●● ●
● ●● ●
●● ●
●● ● ● ●●● ●● ●● ● ● ● ●
●● ● ●●●● ● ● ●
● ●
●● ● ●●● ●● ● ●● ● ●
● ●● ● ● ● ●
● ● ● ● ●
●● ● ●● ● ● ●● ● ●
●●
● ●●●●● ● ●● ●●● ● ●● ●●
●● ● ● ●
●● ●● ● ● ● ●● ● ● ● ●
● ●● ● ●
●● ● ● ● ● ● ● ●● ●●●●● ● ● ●●● ●●●
● ●
●●
●
●●●●●●● ●● ● ● ● ●●●
● ● ●● ● ● ●●●●● ● ● ● ● ●
●●● ● ●
●●●● ●● ●● ● ●● ● ● ● ● ● ●
● ●●● ● ●●● ● ● ● ●●● ● ● ● ● ● ● ●● ● ●
●● ●● ● ● ● ● ●
●● ● ●●
● ●● ● ● ●
● ● ●● ●●● ●
● ● ● ●●
● ● ●● ● ● ●● ● ● ● ● ● ●
● ●● ●●●●
●●● ●● ● ● ● ● ● ● ● ● ●
●● ●●● ● ● ●
−2
● ● ● ● ● ● ●
●●● ● ●
● ● ●
● ●● ● ● ● ●● ●● ●
●● ● ●●● ●● ● ●
● ●● ● ● ● ● ●
●●● ●● ● ● ● ● ●
● ● ● ● ● ●
● ●
● ● ● ●
● ●
● ● ●● ●● ● ●●●
●● ● ● ● ●● ● ●●
● ● ● ● ●
● ●
● ●● ● ●
●
● ● ●
−4
●●● ● ● ●
●● ●
●● ● ● ●
●
● ● ●
● ●
● ●
●
●
●
●
−6
● ●
●
●
0 2 4 6 8
fitted values (logits)
(b) Scatterplot of residuals vs. fitted values of (4.22)