2008 Phd-Thesis: Modeling and Optimizing The Offshore Production of Oil and Gas Under Uncertainty (Steinar M. Elgsæter)

Steinar M.
Elgster
Modeling and Optimizing the
Oshore Production of Oil and
Gas under Uncertainty
Thesisfor the degree of philosophiae doctor
Trondheim, October 2008
Norwegian University of Science and Technology
Faculty of Information Technology, Mathematics,
and Electrical Engineering
Department of Engineering Cybernetics
NO-7491 Trondheim
2
NTNU
Norwegian University of Science and Technology
Doctoral thesis
for the degree of philosophiae doctor
Faculty of Information Technology, Mathematics,
and Electrical Engineering
Department of Engineering Cybernetics
NO-7491 Trondheim
c 2008 Steinar M. Elgster.
ISBN ISBN 978-82-471-1230-4 (printed version)
ISBN ISBN 978-82-471-1231-1 (electronic version)
ISSN 1503-8181 ITK Report 2008-8-W
Doctoral theses at NTNU, 2008:267
Printed by Tapir Uttrykk
Summary
The topic of this thesis is optimizing the oshore production of oil and gas
on a day-to-day basis. Optimization is based on a model of the production.
Designing such production models is very complex and challenging, and as
a result models will always be subject to some uncertainty in practice.
In this thesis a data-driven approach to production modeling and model
updating is suggested. Simple model structures are inferred from observa-
tions of measured production, motivated by the concepts of system identi-
cation and a desire for models which are suciently accurate while being
easily solved numerically. A production model needs to be updated peri-
odically to reect changes in reservoirs and wells. Updating models against
observations of recently measured data describing normal operations is sug-
gested, motivated by a desire for an updating scheme which requires little
human intervention and does not require frequent additional experimenta-
tion, for instance in the form of well-tests.
The suggested approach to model-updating is based on parameter es-
timation against production data stemming from normal operations. As
data describing normal operations may have low information content, pa-
rameters estimated in this manner may be subject to signicant uncertainty.
Bootstrapping is considered as a means of estimating uncertainty in tted
parameters. Methods for the explicit treatment of such uncertainty based
on Monte-Carlo analysis are investigated. Methods for estimating lost po-
tential due to uncertainty, result analysis, excitation planning and choosing
active decision variables are suggested.
Since uncertainty will never be completely eliminated in practice, some
method is needed to adapt to uncertainty when implementing changes in pro-
duction suggested by optimization. Motivated by how operators implement
such changes in practice, an operational strategy approach is suggested to
link optimization with implementation and monitoring by iteratively imple-
menting smaller changes while monitoring prots and constraints.
The models considered are low in complexity, which may simplify design-
i
ii Summary
ing and updating models and may reduce numerical issues. Since the sug-
gested methods rely on measurements, the success of the suggested methods
will depend on measurement quality. Also, although the methods suggested
do not require exotic measurements, the suggested methods could benet
from additional instrumentation on some elds. The methods suggested to-
gether constitute a novel, structured approach to optimization of oshore
oil and gas production in the presence of uncertainty.
Modeling methods are applied to real-world production data from two
oshore oil elds. Implementation of the suggested methods are studied by
simulations against an analog of one of the elds considered.
Preface
I would especially like to thank my two supervisors Tor Arne Johansen and
Olav Slupphaug. I am grateful for Tor Arne for giving my a ying start and
guiding me along my path. This thesis has greatly beneted from Olavs
insight into the practical side of oshore oil and gas production and his
innovative ideas.
In addition to my supervisors, this thesis would not have been possible
in its present form without the valuable discussions with and feedback from
many working practitioners in the oil and gas industry. I would like to thank
Anastasios Siamos, Espen Halvorsen and their colleagues at StatoilHydros
oces in Sandsli for their assistance in obtaining and interpreting data. I
would like to thank Are Mjaavatten, Kristin Hestetun and Marta Dueas
Dez at StatoilHydros research center in Porsgrunn for stimulating discus-
sions, and I would especially like to thank Vidar Alstad for his support.
I would like to thank John-Morten Godhavn and Petter Tndel and their
colleagues at StatoilHydros research center in Trondheim for their assis-
tance in obtaining and interpreting data. I would like to thank Hans-Petter
Bieker, Bjrn Bringedal and their colleagues at ABB Enhanced Oil Produc-
tion for their valued feedback. I also would like to thank my colleagues at
the institute of Engineering Cybernetics.
I would like to acknowledge the Research Council of Norway, StatoilHy-
dro, and ABB for funding this research.
I am indebted to my girlfriend Ines for the love, support and encourage-
ment shown to me during the course of this work, and I am grateful for the
support shown to us by our families.
Trondheim, October 2008 Steinar M. Elgster
iii
iv Preface
Contents
Summary i
Preface iii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Summary and thesis overview . . . . . . . . . . . . . . . . . . 8
1.3.1 Publications . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Main contributions . . . . . . . . . . . . . . . . . . . . 12
1.4 Notation and denitions . . . . . . . . . . . . . . . . . . . . . 12
2 Preliminaries 19
2.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Mathematical preliminaries . . . . . . . . . . . . . . . . . . . 21
3 Challenges in parameter estimation of models for oshore
oil and gas production optimization 25
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1.1 Prior work . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 The relationship between production optimization, models
and data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Production data: A case-study . . . . . . . . . . . . . . . . . 29
3.3.1 Field description . . . . . . . . . . . . . . . . . . . . . 30
3.3.2 Ratio between well tests and total measured rates . . . 30
3.3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.4 Observations . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.6 Challenges in production optimization . . . . . . . . . 39
v
vi Contents
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4 Production optimization; system identication and uncer-
tainty estimation 41
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.1.1 Problem formulation . . . . . . . . . . . . . . . . . . . 42
4.2 Modeling for production optimization under uncertainty . . . 43
4.2.1 Local production models for optimization . . . . . . . 43
4.2.2 Estimating parameter uncertainty . . . . . . . . . . . 44
4.2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 Synthetic examples and a case study . . . . . . . . . . . . . . 45
4.3.1 Production model . . . . . . . . . . . . . . . . . . . . 46
4.3.2 Numerical solution of the parameter estimation problem 47
4.3.3 Physical knowledge in parameter estimation . . . . . 47
4.3.4 A synthetic example . . . . . . . . . . . . . . . . . . . 49
4.3.5 Case study: Production data from a North Sea oil eld 53
4.3.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5 Oil and gas production optimization; lost potential due to
uncertainty 61
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.1.1 Prior work . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Estimating the loss due to uncertainty . . . . . . . . . . . . . 65
5.2.1 Denitions . . . . . . . . . . . . . . . . . . . . . . . . 66
5.2.2 Estimating the potential of production optimization . 67
5.2.3 Estimating lost potential due to uncertainty . . . . . . 67
5.2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3.2 Synthetic example . . . . . . . . . . . . . . . . . . . . 72
5.3.3 Case study: North Sea eld . . . . . . . . . . . . . . . 72
5.3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.5 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Contents vii
6 A structured approach to optimizing oshore oil and gas
production with uncertain models 79
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.1.1 Prior work . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.2 Production optimization under uncertainty . . . . . . . . . . 83
6.2.1 A structured approach to optimization with uncertain
models . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.2.2 Operational strategy . . . . . . . . . . . . . . . . . . . 84
6.2.3 Result analysis . . . . . . . . . . . . . . . . . . . . . . 90
6.2.4 A cost/benet approach to excitation planning . . . . 91
6.3 Case study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.3.1 Field description . . . . . . . . . . . . . . . . . . . . . 94
6.3.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7 Optimization of oshore oil and gas production under un-
certainty; active decision variables 111
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.1.1 Previous Work . . . . . . . . . . . . . . . . . . . . . . 112
7.1.2 Problem Formulation . . . . . . . . . . . . . . . . . . . 113
7.2 Choosing active decision variables . . . . . . . . . . . . . . . . 113
7.3 Case study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.3.1 Field Description . . . . . . . . . . . . . . . . . . . . . 116
7.3.2 Measurements used and well types . . . . . . . . . . . 118
7.3.3 Production optimization . . . . . . . . . . . . . . . . . 119
7.3.4 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.3.5 Implementation . . . . . . . . . . . . . . . . . . . . . . 128
7.3.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.3.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
8 Conclusions and suggestions for further work 143
8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.1.1 A data-driven approach to production modeling and
model updating . . . . . . . . . . . . . . . . . . . . . . 143
8.1.2 An explicit treatment of uncertainty in production op-
timization . . . . . . . . . . . . . . . . . . . . . . . . . 145
viii Contents
8.1.3 An operational strategy approach to link optimization
with implementation and monitoring . . . . . . . . . 147
8.2 Directions for further work . . . . . . . . . . . . . . . . . . . . 147
8.2.1 Integration of production models with physics-based
simulators . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.2.2 Active decision variables . . . . . . . . . . . . . . . . . 148
8.2.3 Subsea wells . . . . . . . . . . . . . . . . . . . . . . . . 148
8.2.4 Comparing alternative methods of estimating uncer-
tainty . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.2.5 The benet of single-rate well tests . . . . . . . . . . . 148
8.2.6 The value of reducing measurement uncertainty . . . . 149
8.2.7 Constraint uncertainty . . . . . . . . . . . . . . . . . . 149
8.2.8 Reservoir modeling . . . . . . . . . . . . . . . . . . . . 149
8.2.9 Studying the link to reservoir management . . . . . . . 150
8.2.10 Managing transients in operational strategies . . . . . 150
8.2.11 Exploiting shutdowns for model updating . . . . . . . 150
A Data preprocessing 151
Bibliography 153
Chapter 1
Introduction
1.1 Motivation
In the oshore production of oil and gas, oil, gas, water and possibly other
components are produced from a number of production wells, before pro-
duced uids are routed through pipes as multiphase ow to a processing
facility for separation. Production in the context of oshore oil and gas
elds can be considered the total output of production wells.
In this thesis production optimization is dened as the mathematical
problem of daily maximization of prots through choosing a change in deci-
sion variables subject to a production model and production constraints. To
solve the production optimization problem generally requires using numeri-
cal optimization techniques. Production optimization is sometimes referred
to as production planning (Bieker, Slupphaug and Johansen, 2007a). The
mathematical solution of the production optimization problem requires a
production model, some equation which expresses the relationship between
change in decision variables and change in production, both with respect
to prots and constraints. A production model typically predicts change in
rates of oil, gas, water, and possibly other components of production as for
instance choke openings or gas-lift rates change.
The most common practice in industry is an engineering approach to
production modeling, integrating detailed and complex rst-principles mod-
els and empirical equations in commercial simulators
1
, often through nodal
analysis (Beggs, 2003). Optimization algorithms either query simulators
directly (Wang, 2003), or simulator predictions are aggregated in proxy-
models which are then queried by optimization algorithms (Zangl, Graf and
1
One such simulator is PROSPER by Petroleum Experts, www.petex.com
1
2 Introduction
Al-Kinani, 2006).
The complexity of production modeling is a key challenge in production
optimization for a number of reasons:
Production is dicult to model due to the shear complexity of the
systems considered, since production can span a large number of wells
and ow lines and a large amount of surface and subsurface process
equipment.
The oshore production of oil and gas deals for a large part with many
uids, often in dierent phase, owing through the same pipe, which
is a complex and dicult physical process to model.
Measurements of oshore oil and gas production are often inaccurate,
either due to the diculty of accurately measuring multiphase ow, or
due to a lack of calibration and maintenance of measurement equip-
ment.
Production changes due to changes in wells and reservoirs, for instance
due to depletion.
Another key challenge in production optimization is model updating, up-
dating models to account for un-modeled changes in production. If un-
modeled changes require production optimization to update decision vari-
ables at intervals of hours or days, then some method for model maintenance
at these intervals is also necessary. Challenges in model updating may be
linked to challenges in modeling, as timely model updating may be more
challenging with increasing model complexity.
Yet another challenge in production optimization is numerical and op-
timization issues. As production models may need to be re-solved perhaps
many times on the timescale of hours or days as part of model updating and
production optimization, eciency of the numerical solver is important. It is
also important that models can be solved in a manner which is numerically
robust. The structure of the production model may make the production op-
timization problem non-convex, which may make numerical optimization of
modeled production challenging. The challenge of numerical issues is linked
to model complexity, as more complex models may be more challenging to
solve and more prone to issues of numerical stability and convexity.
A control approach to production optimization is also possible, which
can consider production optimization as a controller which based on ob-
servations made from measurements attempts to alter decision variables so
that prots increase. Such a controller would still need to be based on
1.2. Related work 3
a production model. The models used in control are often inferred from
data, and the study of such methods is the topic of system identication
(Ljung, 1999). System identication has been found to often achieve high
performance with very simple models and updating of such models is of-
ten simple (Gevers, 2006). The properties of models found with system
identication motivate this thesis, as they have the potential to address
the challenges in modeling, model updating and numerical issues faced in
production optimization.
1.2 Related work
System identication in the broadest sense deals with inferring dynamical
models from data (Ljung, 1999). Sjberg, Zhang, Ljung, Benveniste, Delyon,
Glorennec, Hjalmarsson and Juditsky (1995) divides models into white-box,
black-box or grey-box depending on whether their structure is based on
physical rst principles, empiricism or a combination, respectively. System
identication is sometimes dened more narrowly as purely empirical black-
box modeling, while in this thesis the term is used according to the broader
denition. System identication takes a pragmatic view of the choice of
model structure, seeking model structures in a trial-and-error fashion which
can be relatively easily tted to data yet describe observations with su-
cient accuracy. Emphasis is usually on keeping the number of parameters
to be tted low while introducing some physical knowledge to achieve re-
quired performance (Sjberg et al., 1995). In this thesis static models with
parameters that vary with time to reect changes in reservoir and well are
considered, but the methodology used is inuenced by system identication.
In practice any model will need some adjustment to match observations
due to structural model errors or un-modeled disturbances. Parameter esti-
mation is a procedure belonging to the system identication domain which
compares the measured and modeled output of a system for a set of tuning
data and adjusts parameters, usually with a least-squares numerical solver,
until measured and modeled output match as closely as possible. For it to
be possible to determine parameters to be tted uniquely, there needs to
be sucient information content in the tuning data set. If the information
content in production data is low, more than one set of tted parameter
values may describe the tuning set equally well and parameter uncertainty
may result. Information content in data can be increased by introducing
variation in decision or disturbance variables, or excitation, but excitation
often has an associated cost which may make eliminating parameter uncer-
tainty impractical. Due to uncertainty, it is widely recognized that a model
4 Introduction
is of little use if no quality tag is attached to it which describes uncertainty
and is taken into account when the model is used in practice (Campi and
Weyer, 2005).
Production optimization is related to the topic of real-time optimization
in process control, controllers which attempt to maximize economic value by
choosing setpoints subject to a model and constraints. Real-time optimiza-
tion usually consider nonlinear steady-state process models and is usually
implemented with sampling times in the range of hours, and model parame-
ters are usually updated using parameter estimation (Engell, 2006). Deter-
mining the level of accuracy required in model and parameter estimates and
determining the benet of improving accuracy are considered open questions
for research in real-time optimization (Forbes, Marlin and Yip, 2005).
Friedman (1995) states that a challenge to real-time optimization in the
process industries is that the use of steady-state process models prevents
the frequent re-setting of targets in smaller steps and that detailed accurate
models and complex simulations are required to perform real-time optimiza-
tion. Friedman (1997) points out that real-time optimization is based on
models which can be invalidated by several circumstances, for instance vio-
lated modeling assumptions or measurement errors. It is stressed that con-
trol schemes are required to function perfectly at constraints or risk being
turned o, as operators are uncomfortable with operating under such con-
ditions, yet if the model is inaccurate for any reason real-time optimization
will not function perfectly at constraints and the optimal setpoint calculated
by real-time optimization is no longer optimal in practice.
As parameter estimation is considered part of real-time optimization,
changes in setpoints made by real-time optimization will inuence accu-
racy of parameter estimation. In chemical engineering, experimental design
has considered planned excitation to reduce parameter uncertainty. Classi-
cal optimum experimental design considers the information content of data
sets without considering the cost of performing experiments, suggesting ex-
periments to reduce correlation between tted parameters (Pritchard and
Bacon, 1978). Recently, Yip and Marlin (2003) have proposed an expected
prot approach which weighs costs of experiments against the expected gain
over a future horizon in the objective function.
It is common to consider layered controllers in control in the process
industries. Real-time optimization is implemented above a layer of model-
predictive control, which manages transients, and model-predictive control is
again above a layer of regulatory control, which stabilizes and ensures that
the process is operated close to the settings specied by the layer above.
Online model-based optimization has found widespread use in processes in-
1.2. Related work 5
dustries, especially the methods of model-predictive control. It is interesting
to note that the vast majority of implementations of model-predictive con-
trol use empirical models (Qin and Badgwell, 2003), and this may be related
to the lower costs of model design and updating of such models.
In the context of oil and gas production real-time optimization has been
dened as a process of measure-calculate control at a frequency which main-
tains the systems optimal operating conditions within the time-constant
constraints of the system (Saputelli, Mochizuki, Hutchins, Cramer, An-
derson, Mueller, Escorcia, Harms, Sisk, Pennebaker, Han, Brown, Kabir,
Reese, Nuez, Landgren, McKie and Airlie, 2003). A review of real-time
optimization case histories in the production of oil and gas can be found in
Mochizuki, Saputelli, Kabir, Cramer, Lochmann, Reese, Harms, Sisk, Hite
and Escorcia (2006). The production optimization problem is related to
the reservoir management problem, which considers the optimization of in-
jection and drainage on the time scales of months and years. Production
optimization and reservoir management consider real-time optimization at
dierent time-scales, and it is common to treat reservoir management and
production optimization as separate subproblems linked through constraints
to limit complexity (Saputelli et al., 2003). There is a possibility that short-
term gains in prots gained by production optimization come at the ex-
pense of ultimate recovery as some researchers have observed in simulations
(Naus, Dolle and Jansen, 2006). In such cases short-term gains may need to
be weighed against long-term gains. Mochizuki et al. (2006) state that the
most important measure of the success of a real-time optimization system
in oil and gas production is the value added to the organization, however
the measurement of value is a very complex issue. A critical factor for
the successful adoption of new production optimization technology is that
there is tremendous pressure to keep assets producing at agreed-upon rates,
therefore operators are generally risk-averse (Jalilova, Tautiyev, Forcadell,
Rodriguez and Sama, 2008). A common practice is to generate a production
and injection plan periodically at typical intervals of between a week and
a month, which species for instance target choke openings or gas-lift rates
with the aim of maximizing daily production (Bieker et al., 2007a).
The concepts of black-box or grey-box modeling have been applied to
reservoir management previously (Rowan and Clegg, 1963), (Chierici, 1995)
(Renard, Dembele, Lessi and Mari, 1998), and (Saputelli, Nikolaou and
Economides, 2005). Saputelli et al. (2005) suggest that a persistently excit-
ing process can be implemented in cases where data captured from the eld
is not informative enough. Foss (1987) suggested a strategy for the design
of well tests such that the measured data is more sensitive to the properties
6 Introduction
of near-well reservoir models which are of interest.
The concepts of parameter estimation have also been applied to t con-
ventional reservoir models to data, in what is referred to as automated his-
tory matching. It is generally accepted that any physical reservoir model
history matched to data will be subject to signicant non-uniqueness (Little,
Fincham and Jutila, 2006), and many authors have attempted to describe
this non-uniqueness (Almeida Netto, Schiozer, Ligero and Maschio, 2003),
(Schulze-Riegert, Haase and Nekrassov, 2003), and (Yang, Nghiem, Card
and Bremeier, 2007). In many cases history matching of reservoir models is
performed manually, and manually history-matching a reservoir model to a
eld has been reported to take as much as a year, which is impractical for
short-term decision making (Awasthi, Sankaran and Nikolau, 2008).
The nodal approach to production optimization has numerical issues re-
lated to infeasible solutions in the case of shut-in wells, change in model
dimensions as wells are shut-in, and disappointing computational eciency
according to Wang, Litvak and Aziz (2002). Barragn-Hernndez, Vzquez-
Romn, Rosales-Marines and Garca-Snchez (2005) attempted an alterna-
tive modeling approach still based on rigorous physical modeling and re-
ported that solving the resulting set of dierential-algebraic equations was
impractical numerically. Some authors have suggested that the main bot-
tleneck in production optimization is the computational eort required to
solve rigorous physical production models (Zangl et al., 2006).
A vast number of strategies for real-time optimization of oshore oil and
gas production have been proposed in the literature, for a recent review see
Bieker et al. (2007a). Bieker et al. (2007a) noted that very few of these
strategies were designed for on-line use and that few real-world implementa-
tions exist, and suggested that this may be linked to diculties with model
maintenance due to the complexity of the models in many of the proposed
strategies. The authors also indicate that developing strategies for handling
model uncertainty is a key challenge in real-time optimization.
The concepts of system identication have been applied to production
modeling previously in Poulisse, van Overschee, Briers, Moncur and Goh
(2006) and Poulisse (2008). The authors suggest re-routing each well to a
test separator and toggling decision variables to determine a dynamic nger-
print for each well. Fitted parameters are determined so that a production
model consisting of tted parameters and dynamic ngerprints combined
match measured rates as closely as possible. Mjaavatten, Aasheim, Saelid
and Gronning (2006) have considered production optimization of gas-coned
wells using relatively simple models based on a partial dierential equation
that describes the near-well area and which is tted to data, what could be
1.2. Related work 7
described as a grey-box approach to production modeling.
In practice rigorous physical models of oshore oil and gas production
often require some form of empiricism to achieve desired accuracy. For
instance, rigorous physical models used in multiphase ow meters to esti-
mate oil, gas and water rates from measured pressures and temperatures
are often only considered accurate for a range of operating points for which
models have been calibrated (van Dijk, Goh and van Lienden, 2008). Mod-
els used in production optimization are also usually subjected to some form
of model-tting. A practice commonly referred to in the literature is to
t proxy-models found from commercial simulators to recently measured
single-rate well tests, see for instance Jalilova et al. (2008). A disadvantage
of this approach is that it can only compensate for biases in the produc-
tion models, and modeled change in production as decision variables change
is not validated against measurements. Nadar, Schneider, Jackson, McKie
and Hamid (2008) have observed that the change in production predicted by
an undisclosed commercial simulator did in fact not match changes in pro-
duction observed on an oshore oil and gas eld as decision variables were
changed. To circumvent this shortcoming, the authors incrementally im-
plemented small changes in decision variables, observed resulting change in
production and updated models to compensate for these model inaccuracies.
Model-tting and strategies to incrementally implement small changes can
be considered implicit strategies for handling uncertainty in optimization of
oshore oil and gas production.
A signicant hurdle to model-tting is measurement uncertainty, which
can be signicant in the oshore oil and gas production. Measurements of oil
production are generally most accurate at point-of-sale, where scal mea-
surements are usually available, and become increasingly uncertain toward
the well level (Little et al., 2006). Some consider scal meters which mea-
sure exported oil the only reliable measurement of oil production (Jalilova
et al., 2008).
Although there seems to be much agreement that uncertainty is an im-
portant factor in production optimization (Bieker et al., 2007a), and much
attention has been paid to uncertainty in reservoir management, explicit
treatment of uncertainty in production optimization has received little at-
tention. One of the few explicit treatments of uncertainty in production
optimization can be found in Bieker, Slupphaug and Johansen (2006).
Parameter estimation is a form of regression problem. A measure of
uncertainty in estimated parameters is the standard error, the standard de-
viation of the sampling distribution of the sample mean. The standard error
of a regression problem can be analyzed analytically when the regression
8 Introduction
function is linear and tting method is least squares, while for general re-
gression models that have no analytical solution this error can be evaluated
computationally using bootstrapping (Efron and Tibshirani, 1993). Boot-
strapping is based on the concept of drawing multiple random samples from
an empirical distribution, or resampling. When bootstrapping residuals in
a regression problem, the residuals, the dierence between measured and
modeled values for a data set, are resampled to construct new data sets and
parameter estimation repeated for each such set. An alternative method
of bootstrapping in parameter estimation is bootstrapping pairs (Efron and
Tibshirani, 1993). Bootstrapping provides multi-variable sampling distri-
butions that capture the dependence between between marginal and condi-
tional parameter distributions (Frey and Burmaster, 1999). Other avors of
bootstrapping also exist, for instance parametric bootstrap which samples a
parametric rather than empirical distribution. The jackknife is a technique
which bears close resemblance to the bootstrap. The number of bootstrap
replications required depends on the information desired, but 200 resam-
ples or less are often sucient to determine a standard error, while 1000
resamples is considered sucient to determine condence intervals (Efron
and Tibshirani, 1993). There also exist variations of bootstrapping methods
intended to improve computational eciency (Gigli, 1996).
Some recent research on assessing uncertainty in dynamic models may
have some relevance to this thesis, although dynamic models will not be
considered. In system identication the asymptotic error of Prediction Error
Methods is traditionally used to assess model error, which allows condence
regions for parameters to be found from data analytically (Ljung, 1999).
Garatti, Campi and Bittanti (2004) note that this asymptotic theory requires
the absence of un-modeled dynamics, and is only rigorously correct when
when the number of data tends to innity in such a way that the total
amount of information on the system parameters grows unbounded, and
only when the cost function used in parameter estimation has a unique
minimizer. Garatti et al. (2004) further note that when data sets have
low information content or data sets are nite and small, asymptotic theory
providing misleading results, i.e. this theory can fail when uncertainty levels
are high. One alternative to asymptotic theory is suggested in Campi and
Weyer (2005).
1.3 Summary and thesis overview
Chapter 2 contains a problem description and mathematical preliminaries,
which forms the basis for subsequent chapters.
1.3. Summary and thesis overview 9
Chapter 3 contains an analysis of a set of production data from a North
Sea oil and gas eld. In the set of production data considered, decision
variables are seen to vary little except during times when all wells are shut
down, which might not reveal much useful information. Decision variables
are also observed to vary within a narrow range of values outside times of
shutdowns, which will make it dicult to determine parameters from data
which are valid globally for all setpoints. Observations made in Chapter 3
motivate the study of the link between the information content in produc-
tion data, uncertainty in the tted parameters of production models and
uncertainty in setpoints suggested by production optimization.
Chapter 4 presents an approach to modeling production which is mo-
tivated by the concepts of system identication and which is suitable for
production optimization. The model structure suggested is intended to be
locally valid around the rates and settings measured at the most recent well
test. The methods are applied to a case study of the eld studied in Chap-
ter 3, considering 20 gas-lifted wells. It is suggested how constraints can
aid parameter estimation by expressing physical knowledge, for instance of
rate-independent gas-oil or water-oil ratios. Uncertainty estimation in t-
ted parameters by the method of bootstrapping is applied to the suggested
model. As expected, the uncertainty in parameters when tted to the set
of production data considered is shown to be considerable even though the
tted model matches the tuning data set well. Although parameter uncer-
tainty will depend on model structure, production data and the estimation
strategy, it is reasonable to expect any model structure to have signicant
uncertainty due to the low information content in data. An advantage of the
bootstrapping method is that it does not make any requirements on model
structure.
Chapter 5 considers a Monte-Carlo approach to assessing the uncer-
tainty in production optimization that results from estimated uncertainty
in production models, and an approach to relate model uncertainty to sim-
ulated loss in prot is suggested, the concept of lost potential of production
optimization. The potential for production optimization that is lost due
to uncertainty will depend on how setpoint changes are implemented, and
a rudimentary operational strategy, a strategy for gradual setpoint change
while monitoring prots and constraints, is simulated when estimating lost
potential due to uncertainty. The suggested method is applied to the same
eld and data set as that of Chapter 4. Both total gas and total water ca-
pacity constraints are implemented, as well as some well constraints due to
ow assurance issues.
Chapter 6 studies production optimization based on uncertain produc-
10 Introduction
tion models, utilizing an elaborated operational strategy combined with it-
erative re-tting of model parameters and re-optimization. The operational
strategy is motivated by the way operators currently implement changes in
setpoints in small, iterative steps and may be applied regardless of model
structure used or of type of uncertainty, as long as capacities and prots are
measured and the sign of the response in prots and capacity utilization to
change in each decision variable is known. When production is optimized
based on uncertain production models, the resulting change in realized prot
is uncertain, it may even be negative. In Chapter 6 a Monte-Carlo algorithm
is suggested which gives probabilistic predictions of possible changes in re-
alized prots resulting from a setpoint change, exploiting an estimate of
parameter uncertainty.
When production is optimized based on uncertain production models,
convergence to optimum cannot be guaranteed in general. It may be that
at a given iteration of the iterative optimization strategy, implementing ex-
citation of a particular well will reduce parameter uncertainty in a manner
which allows a much larger prot increase in subsequent optimization and
implementation. Chapter 6 investigates estimating this additional prot in-
crease caused by excitation, or benet of excitation, using a Monte-Carlo
method which exploits an estimate of parameter uncertainty.
Chapter 7 presents a case study of a second North Sea oil and gas eld,
consisting of 16 wells where the production choke setting is considered the de-
cision variable, four gas-lifted wells and ve subsea wells. A model structure
found with an approach similar to that of Chapter 4 is suggested for plat-
form wells where choke setting is considered the decision variable. Subsea
wells are not considered part of production optimization, but a rudimentary
subsea well model is suggested to estimate and subtract the contribution of
these wells from total measured production. When bootstrapping is applied
to parameter estimation, parameter uncertainty is found to be signicant
and of similar magnitude as that of the previous case study even though
the model matches the tuning set well. Production optimization consid-
ers constraints in total gas and total water capacity, as well as some well
constraints.
Altering all decision variables simultaneously at each implementation
of the operational strategy may not be practical, therefore a Monte-Carlo
method for evaluating subsets of decision variables to be altered, sets of
active decision variables, is suggested in Chapter 7. It is proposed how the
Monte-Carlo method can be used to nd active decision variables so that a
setpoint change can be suggested which can be expected to increase prots
with condence.
1.3. Summary and thesis overview 11
Chapter 8 contains conclusions drawn from this work and proposes some
directions for further work.
1.3.1 Publications
This thesis has resulted in a number of publications, as outlined below.
Chapter 3:
Elgsaeter, S.M., Slupphaug, O. and Johansen, T.A. (2007), Challenges
in Parameter Estimation of Models for Oshore Oil and Gas Produc-
tion Optimization , International Petroleum Technology Conference,
Dubai, December 2007, IPTC 11728
Chapter 4:
Elgsaeter, S.M., Slupphaug, O. and Johansen, T.A. (2008), Produc-
tion Optimization; System Identication and Uncertainty Estimation,
Intelligent Energy 08, Amsterdam, February 2008, SPE 112186
Chapter 5:
Elgsaeter, S.M., Slupphaug, O. and Johansen, T.A. (2008), Oil and
Gas Production Optimization; Lost Potential due to Uncertainty, IFAC
World Congress, Korea, July 2008
Chapter 6:
Elgsaeter, S.M., Slupphaug, O. and Johansen, T.A. (2008), Optimiz-
ing oshore oil and gas production with uncertain production models,
Computers & Chemical Engineering (submitted)
Chapter 7:
Elgsaeter, S.M., Slupphaug, O. and Johansen, T.A. (2008), Optimiza-
tion of Oshore Oil and Gas Production under Uncertainty; Choosing
Active Decision Variables and Case-Study, SPE Production & Opera-
tions (submitted)
As main author, my contribution to these papers are in the modeling,
development of methodology and implementation of the case studies pre-
sented. My co-authors have assisted in determining the scope and direction
of the work, as well as providing feedback and review.
12 Introduction
1.3.2 Main contributions
The main contributions of this thesis, all with regard to the oshore pro-
duction of oil and gas, are
a data-driven approach to production modeling and model updating
(Chapters 4 and 7),
an explicit treatment of uncertainty in production optimization (Chap-
ters 4, 5, 6 and 7), and
an operational strategy approach to link optimization with implemen-
tation and monitoring (Chapters 5 and 6).
Chapter 3 contains a study of a set of real-world production data which
motivated the approach described in later chapters.
1.4 Notation and denitions
Symbols introduced in Chapter 2:
u decision variables
x internal variables
d modeled and measured disturbances indepen-
dent of u
parameters to be determined through parameter
estimation
M prot measure
c production constraints
y measured variables exploited in parameter esti-
mation
t time
N the number of time steps in the tuning data set
Z
N
a tuning data set spanning N time steps
residuals in parameter estimation
D
y
normalization matrix for y
w weighting of residuals in parameter estimation
V
s
soft constraints on parameters
c
hard constraints on parameters

R routing matrix
1.4. Notation and denitions 13
z relative production choke opening
q
gl
gas lift rate
q
o
well oil production rate
q
g
well gas production rate
q
w
well water production rate
q
tot
o
total rate of oil produced from all wells consid-
ered
q
tot
g
total rate of gas produced from all wells consid-
ered
q
tot
w
total rate of water produced from all wells con-
sidered
q
test
o
oil rate from a well measured at the most recent
well test
q
test
g
gas rate from a well measured at the most recent
well test
q
test
w
water rate from a well measured at from the most
recent well test
F
o
ratio between well tests and total measured rates
for oil
F
g
for gas
F
w
for water
q
l
the oil, gas and water rates measured at the most
recent well tests for each well
d
l
the value d at the time of the most recent well
tests for each well
t
l
the times of the most recent well tests for each
well
f
d
a kernel function describing the changes in pro-
duction of a well to changes in d
f
u
duction of a well to changes in u
f
t
duction of a well to changes in t
P
the covariance matrix of the asymptotic distri-

bution of prediction error estimates
0
the variance of residuals
14 Introduction
e a stochastic process component of y
Z
N
r
a resampled dataset
N
s
the number of resamples in a Monte-Carlo re-
sampling
q
gl
the normalized change in gas-lift rates since rel-
ative to value tested at most recent well test
r
parameters found from a resampled dataset
z
l
the choke openings measured at the most recent
well test
q
l
gl
the gas lift rate measured at the most recent well
test
o
tted parameter in oil performance curve
g
tted parameter in gas performance curve
w
tted parameter in water performance curve
o
tted parameter in oil performance curve
g
tted parameter in gas performance curve
w
tted parameter in water performance curve
y
tted parameters describing biases in measure-
ments
r
m
n
the ratio between production rates of phases m
and n
R
L
lower constraints on ratios r
m
n
R
U
upper constraints on ratios r
m
n
m mass ow rate
o
density of oil
g
density of gas
w
density of water
reg
soft constraint on parameters
reg
soft constraint on parameters
w
reg
weight of soft constraints on
w
reg
weight of soft constraints on
n
w
number of wells considered in production opti-
mization
k a design parameter in kernel functions f
z
K
GOR
gas-oil ratio
K
WC
water cut
parameter values which are true in some sense

P
o
the potential for production optimization
P
o,r
realizable potential for production optimization
L
u
lost potential of production optimization
Q the quotient of potential that is lost due to un-
certainty
f
L
u
probability function for loss due to uncertainty
f
P
o
probability function for the potential for produc-
tion optimization
f
probability function for parameters

H
0
the event that production is currently optimal
H
1
the event that production is currently sub-
optimal
W
c
the set of indices of wells whose gas-lift rate may
not be reduced
U
prc
a limit on the size of change in decision variables
q
max
g
maximum gas processing capacity
q
max
w
maximum water processing capacity
b
o
bias added to modeled oil production in produc-
tion optimization
b
g
bias added to modeled gas production in pro-
duction optimization
b
w
bias added to modeled water production in pro-
duction optimization
L the length of the interval at the end of the tun-
ing set used over which biases b
o
, b
g
and b
w
are
calculated
s the portion of a decision variable change that is
implemented before a constraint is reached
N
t
number of resamples used in Algorithm 5.2
N
f
number of resamples used in Algorithm 5.2
t
samples of used in Algorithm 5.2
f
samples of used in Algorithm 5.2
f
P
o
|H
1
probability function for loss due to uncertainty,
given that production is sub-optimal
f
L
u
|H
1
probability function for the potential for produc-
tion optimization, given that production is sub-
optimal
f
Q|H
1
probability function for Q, given that production
is sub-optimal
16 Introduction
u
min
the lower end of the range of possible values for
u
u
max
the higher end of the range of possible values for
u
u
T
a target value for an operational strategy
U
the set of indices of u which are to be decreased

as an operational strategy approaches u
T
from
u
0
U
+
the set of indices of u which are to be increased
as an operational strategy approaches u
T
from
u
0
J
M
a cost function against which measured prot
change during implementation of an operational
strategy is to be compared to determine whether
to terminate the operational strategy before the
target is reached
J
T
M
a threshold value for J
M
t
o
s
the time at which the operational strategy is ini-
tiated
t
o
e
the time at which the operational strategy is ter-
minated
S the user-specied size of setpoint change at each
iteration of an operational strategy
J
r
a cost function for the expected prot attainable
from moving setpoints toward a new target
J
T
r
M
B
E
the benet of excitation
C
E
the cost of excitation
n
the nominal parmater estimate, in the context
of benet of excitation
J
E
the cost-benet tradeo of an excitation
J
E,T
E
h a general function which maps (u, d) to y
I
a
set of indices of active decision variables
I
C
a
set of indices of inactive decision variables
I
A
, I
B
, I
C
I
D
, I
E
indices of wells of dierent categories A-E
I
MC
the set of indices of wells which have maximum
constraints on choke openings due to sand pro-
duction
I
MBHP
the set of wells which have minimum constraints
on bottomhole pressure
I
MGS
the set of indices of wells which have constraints
on gas-lift rates due to slugging
I
MR
the set of indices of wells which have maximum
constraints on rates due to sand production
z
tting parameters describing production chokes
s
tting parameters describing subsea chokes
L
lower constraints on
U
upper constraints on
p
bhp
bottomhole pressure
18 Introduction
Chapter 2
Preliminaries
Abstract
This chapter introduces a preliminary problem description and mathematical
formulation which will apply in all subsequent chapters.
2.1 Problem description
Production in the context of oshore oil and gas elds, can be considered
the total output of production wells, a mass ow with components including
hydrocarbons, in addition to water, CO
2
, H
2
S, sand and possibly other
components. Hydrocarbon production is for simplicity often lumped into oil
and gas. Production travels as multiphase ow from wells through ow lines
to a processing facility for separation. Wells can be divided into platform
wells, where production from each well travels in separate ow lines to a
processing facility, and subsea wells, where production from several wells
are routed to a subsea manifold before production is routed through one
or more shared ow lines to a processing facility. Water and gas injection
are used for optimizing hydrocarbon recovery of reservoirs. Gas lift can
increase production to a certain extent by increasing the pressure dierence
between reservoir and well inlet. The schematic model of oshore oil and
gas production is shown in Figure 2.1.
Production is constrained by several factors, including the capacity of
the facilities to separate components of production and the capacity of fa-
cilities to compress gas at the eld level. The production of groups of wells
may travel through shared ow lines or inlet separators which have a lim-
ited liquid handling capacity. The production of individual wells may be
19
20 Preliminaries
Reservoir(s)
.
.
.
.
.
.
.
.
Routing
.
.
.
.
Water
injection
.
.
.
.
Gas
injection
.
.
.
.
. . . .
Separation
.
.
.
.
Separation
g
.
.
.
.
Gaslift
Export
Oil
Gas
Water
Platform
wells
Reinjection/To sea
Lift gas/
gas injection
Gaslift
Subsea
wells
Gaslift
Figure 2.1: A schematic model of oshore oil and gas production.
constrained due to slugging, other ow assurance issues or due to reservoir
management constraints.
Multiphase ows are hard to measure and are usually not available for
individual ow lines in real-time, however measurements of total single-phase
produced oil rates are usually available, and estimates of total gas and water
rates can often be found by aggregating dierent measured gas and water
rates after separation. To determine the rates of oil, gas and water produced
from individual wells, the production of a single well is usually routed to a
dedicated test separator at intervals where the rate of each separated single-
phase component is measured, a well test. Well tests allow biases in models
of individual wells to be updated, and well tests which measure rates for
dierent settings, multi-rate well tests, also allow responses to changes in
for example choke settings predicted by models to be validated.
The aim of production optimization is to determine setpoints for a set
of chosen decision variables which are optimal by some criterion. These set-
points are implemented by altering the settings of production equipment.
Decision variables may be any measured or calculated variables associated
with production which are inuenced by changes in settings, but the num-
ber of decision variables is limited by the number of settings. We may for
instance determine the setting of a gas lift choke by deciding a target lift gas
rate, a target annulus pressure or a target gas lift choke opening. On short
timescales the ow from individual wells can be manipulated by production
choke settings, by gas lift choke settings and possibly by routing settings.
Production optimization requires production models, equations which ex-
press the relationship between change in decision variables and resulting
2.2. Mathematical preliminaries 21
change in production. Production models are usually nonlinear to describe
signicant nonlinear phenomena in production. Parameter estimation is ad-
justment of tted parameters so that predictions of the production model
match a set of recent historical production data as closely as possible. Pa-
rameters of production models should be tted to production data through
parameter estimation to compensate for un-modeled eects or disturbances
and to set reasonable values for physical parameters which cannot be mea-
sured directly or determined in the laboratory. Erosion of production chokes
is an example of a typical such un-modeled disturbance. If there is little
variation of decision variables in the production data against which the pro-
duction model is tted we term the data as having low information content.
Fitted parameters may be uncertain if tted to data with low information
content.
Planned excitation is some planned variation in one or more decision
variables designed to reveal information on production through measure-
ments, for instance a multi-rate well test. Normally production is at set-
points where some capacities are at their perceived constraints, therefore a
multi-rate well test cannot be performed without simultaneously reducing
production at some other well, which may cause lost production and a cost.
There is also a risk that changes in setpoints during testing may cause some
part of the facilities to exceed the limits of safe operation, which may force
an expensive shutdown and re-start of production.
2.2 Mathematical preliminaries
Throughout this thesis variables with a hat () will denote estimates and
variables with a bar () will denote measurements, and variables with a star
( ) are true values in some sense.
In this thesis we will consider production a system, an object in which
variables of dierent kinds interact and produce observable signals. External
signals are often grouped into those which are manipulated and and those
which are not, and non-manipulated signals are referred to as disturbances
(Ljung, 1999).
Let x be a vector of internal variables and let u be a vector of decision
variables. In this thesis we will consider a production optimization problem
22 Preliminaries
on the form
_
u() x()
= arg max
u,x
M(x, u, d) (2.1)
s.t 0 = f(x, u, d, ) (2.2)
0 c(x, u, d), (2.3)
where is a vector of parameters to be determined. u() and x() are
the optimal solution of (2.1)-(2.3) for a given . d is a vector of modeled
and measured disturbances independent of u. M(x, u, d) is a prot measure
which we seek to maximize subject to a process model (2.2) and process
constraints (2.3) for a given parameter value . For oshore oil and gas
production, x may be the production rates of each uid from each well, u
may be relative choke openings or gas-lift rates,c may describe constraints
in total water and gas processing capacity and M may express total oil
production. Which components to model is a design question that depends
on the choice of c(x, u, d) and M(x, u, d), and typically at least oil, gas and
water production is modeled.
We only consider instantaneous optimization, determining ( u(), x())
at the current time. Normally x is not measured. Let y be a vector of
measured variables, which may include separator rate measurements, scal
or other total rate measurements or multiphase rate measurements where
available. Let y(u, d, , t) be an estimate of measured variables based on the
model (2.2), and y(t) is an estimate of y, y may for instance be the measured
total rates of oil, gas and water. Let the tuning data set be a set of historical
process data spanning N time steps spanning the time interval t [1, N]
Z
N
=
_
y(1)]

d(1) u(1) y(2)

d(2) u(2) . . .
y(N)

d(N) u(N)
(2.4)
with residuals
(t, ) = y(t) y(u(t), d(t), , t) t {1, . . . , N} . (2.5)
Parameter estimation determines

by minimizing the sum of the squared
residuals for the tuning data set:
=arg min
t=1
w(t)D
y
(t, )
2
2
+V
s
(), s.t. c
() 0 (2.6)
where w(t) is a user-specied weighting function, V
s
() is an optional soft-
constraint and c
() an optional constraint on . The components of y may

2.2. Mathematical preliminaries 23
have dierent ranges, yet parameter estimation should give similar weight
to minimizing residuals of all measurements, which motivates normalizing
residuals with a diagonal matrix D
y
. The process model (2.2) should only
be tted to historical process data that are recent, in the sense production
during the time interval spanned by the tuning set should be consistent
with current production and the production model. It may be impossible to
determine an accurate estimate

uniquely with (2.6) when the information
content in Z
N
is low. An improved model structure may enable the use
of a longer tuning set, which may in turn reduce parameter uncertainty.
Therefore parameter uncertainty is linked to other forms of uncertainty such
as structural uncertainty.
In general the time interval spanned by the tuning set Z
N
is limited by
un-modeled disturbances, as it is assumed that all variations observed in
Z
N
can be described by the chosen model structure. To update the model
as changing states or disturbances cause production to change, the window
spanned by Z
N
can be moved as new data becomes available, a moving-
horizon approach (Nikolaou, Cullick and Saputelli, 2006).
To describe the relationship between rates x and measurements y, and
to capture the typically time-varying nature of routing, we dene a routing
matrix R of ones and zeros such that when measurement uncertainties can
be neglected,
y(t)

=

R(t)x(t). (2.7)
Routing R(t) may change with time, as wells are routed to dierent pro-
cessing trains or to test separators. y may consist of test separator rate
measurements, total rate measurements and in some cases multiphase rate
measurements at some or all wells.
From (2.1)(2.3) it may be seen that

will inuence u(
), while (2.4)
(2.6) illustrate that

is inuenced by the information content in Z
N
, which
again can be inuenced by planned excitation. These relationships are illus-
trated in Figure 2.2.
In the following chapters we will consider production models which are
tted to data from normal operations using (2.6) and exploited in (2.1)(2.3)
and consider the uncertainty that arises.
24 Preliminaries
Production
optimization
Production
Planned
excitation
Production
constraints
State and model
parameters
Measurements
Decision
variables
Disturbances
Parameter and
state estimation
Figure 2.2: Relationship between subproblems in production optimization.
Chapter 3
Challenges in parameter
estimation of models for
oshore oil and gas production
optimization
A version of this chapter was presented at the International Petroleum
Technology Conference, Dubai, December 2007
Authors: Steinar M. Elgsaeter, Olav Slupphaug, and Tor Arne Johansen
Abstract
In this chapter model-independent assertions about uncertainty in produc-
tion optimization are made through a case study of actual production data
from a North Sea oil and gas eld. As the information content in produc-
tion data describing normal day-to-day operations was observed to be low,
the explicit treatment of uncertainty may be as relevant for production op-
timization as it is for reservoir management. Based on the observations
made, three challenges for research that have received little attention in the
literature until now are highlighted. Firstly, the lost potential incurred due to
uncertainty should be estimated to assess its signicance. Secondly, costs and
values of actions to reduce uncertainty should be estimated to allow proposed
actions to be evaluated through structured business cases. Thirdly, strategies
for making decisions in day-to-day operations under uncertainty should be
investigated.
25
26
Challenges in parameter estimation of models for oshore oil and
gas production optimization
3.1 Introduction
3.1.1 Prior work
Quantifying uncertainty in reservoir models has attracted much interest
in recent years. Authors have described tting multiple models to data
(Griess, Diab and Schulze-Riegert, 2006), describing measurement uncer-
tainties (Little et al., 2006) and sensitivity analysis using prior knowledge
of parameter uncertainty (Costa, Schiozer and Poletto, 2006). Very few
references have been found which attempt to quantify uncertainty in pro-
duction models. An estimate of uncertainty derived from well tests for
wells with rate-independent gas-oil ratio and water-oil ratios is derived in
(Bieker et al., 2006), and a strategy for production optimization of such
wells that treats uncertainty explicitly is suggested in (Bieker, Slupphaug
and Johansen, 2007b). The lack of attention to the subject of uncertainty
in production optimization seems paradoxical as production models are t-
ted to much the same production data as reservoir models. As quantied
uncertainty depends on the model used, it is hard to discern from studies of
uncertainty in reservoir models whether observed uncertainties are a result
of the low information content in production data or of the models. No
references have been found which address the characteristics of typical pro-
duction data itself and what implications these characteristics will have in
terms of modeling.
The concept of estimating the value of information in uncertain systems
has roots in stochastic programming, which uses the concept of value of
perfect information, and the idea of assigning a value to information have
found applications in reservoir modeling (Branco, Pinto, Tinoco, Vieira,
Sayd, Santos and Prais, 2005). No references have been found which address
the value of information in the context of production optimization.
3.1.2 Overview
This chapter rests on the assumption that any production model will re-
quire tting its parameters to production data to account for un-modeled
disturbances, and therefore the characteristics of production data will have
implications for production optimization. Model-independent assertions on
the signicance of uncertainty in production optimization are sought through
a case study of production data. Initially some preliminary considerations
on the relationship between production data and production optimization
are discussed. A case study of a North Sea oil and gas eld is used to illus-
trate the typical state of real-life production data. Finally, key challenges
3.2. The relationship between production optimization, models
and data 27
motivated by the observations made are outlined.
3.2 The relationship between production optimiza-
tion, models and data
As the number of logged measurements from oil and gas production on
one eld may run into the thousands, it is neccessary to focus on selected
measurements in the case study. This section describes the requirements
for production data in the context of production optimization, which will
motivate the focus of the case-study in the following section.
This chapter is a study of production data and what implications the
state of production data will have for the design of the model. An attempt is
made to keep the discussion as independent of the choice of production model
as possible. In practice, production models are either commercial simulators
of well and pipeline, or a lookup-table or other proxy-model derived from
such simulators (Zangl et al., 2006).
This chapter aims to reach conclusions about the implications production
data have on production modeling which are as independent of the choice
of model structure as possible. Selection of variables, both of measured
variables y and decision variables u, is a design choice (Halvorsen, Skogestad,
Morud and Alstad, 2003). The emphasis of this chapter is not to discuss
the selection of variables, but for our discussion a choice of variables y and
u will still have to be made.
As our emphasis is on the optimization of wells and ow lines, rather than
the reservoir or topside facilities, the main settings through which production
can be inuenced are related to routing, gas lift and production chokes.
Routing is accounted for through the matrix R in (2.7). For simplicity
relative production choke opening z and gas lift rates q
gl
as chosen as our
decision variable u =
_
z q
gl
. The only modeled disturbance considered is

routing d = R. In this chapter the information content in production data
Z
N
refers to sensitivities
dy
du
= R
dx
du
. (3.1)
This idea is illustrated in Example 3.1.
Example 3.1 Consider that the objective of production optimization (2.1)
(2.3) is to nd the gas lift rate u = q
1
gl
of a single gas-lifted well which maxi-
mizes oil production, M(x, u, d) = q
1
o
, and assume that the rate of produced
oil is measured y = q
1
o
. Assume that the relationship between gas lift rate and
28
q
gl
1
q
o
1
Dq
o
1
Dq
gl
1
q
o
l
q
gl
l
Figure 3.1: Example 3.1: An illustration of the hypothetical true perfor-
mance curve (solid), along with sensitivity
dq
1
o
dq
1
gl
(dotted).
produced oil is as shown in Figure 3.1, and that production is currently at
(q
l
gl
, q
l
o
). To solve the production optimization problem, the production model
should express how changing gas lift rates will inuence change in produced
oil. The model may be found through physical modeling, through experimen-
tation or a combination thereof. Ideally the production model should predict
the response in oil production q
1
o
to all conceivable changes q
1
ql
in gas lift
correctly, but as a minimum, Z
N
should allow at least, a form of (3.1), to
be determined.
Information content in production data is related to changes in u, which
will be referred to as excitation. If one or more elements in u do not change
or change only very little in Z
N
, then the data set is weak in some way,
sensitivities cannot be determined from production data, (2.6) will be ill-
conditioned and determining certain elements uniquely may become dicult.
It is possible in principle to introduce planned excitation to increase the
information content in Z
N
, for instance by performing multi-rate well tests
at intervals for all wells. As planned excitation may have both a cost and
3.3. Production data: A case-study 29
a risk, it would be preferable to t production models to production data
stemming from normal day-to-day operations if possible. The objective of
decomposing production optimization into reservoir management and pro-
duction optimization is that each subproblem may consider simpler models
considering dierent time scales (Saputelli et al., 2003). Reservoir man-
agement may consider a coarse production model and a detailed reservoir
model, while production optimization may consider a coarse reservoir model
and a detailed production model. The length of the tuning set Z
N
is limited
by the interval for which the chosen model can describe observed variations
in Z
N
, and for production models with a coarse description of reservoir dy-
namics this may mean that Z
N
must span a relatively short time interval.
The information content in Z
N
will decrease with decreasing length in prac-
tice, which introduces the possibility that (2.6) may be subject to signicant
uncertainties.
It is conceivable that the ability of the model to describe the data set (2.4)
can be improved by exploiting auxiliary measurements, measured variables
which may vary with production and are dependent on u. Measurements
which are dependent on the choice of u may for instance be the measured
pressure or temperature upstream of a production choke when production
choke opening is an element in u.
The disadvantage of including auxiliary measurements in the produc-
tion model is that a second model that expresses auxiliary measurements in
terms of x, u and d. is needed to close the production model. A key dier-
ence between models for production monitoring and models for production
optimization is that models for production monitoring may use auxiliary
measurements without the need for this second model. The disadvantage
of introducing a second model for closure is that it increases the chance of
introducing structural model errors, and measured pressures and tempera-
tures may themselves be subject to signicant measurement uncertainties.
For these reasons, measured pressures or temperatures will not be considered
in the case study, although it cannot be ruled out that model performance
could improve by including such measurements.
3.3 Production data: A case-study
This section contains a study of a set of real-world production data. The
emphasis of the study is to highlight characteristics of production data that
may have implications for production optimization
30
3.3.1 Field description
One eld is considered in this case study, a mature North Sea eld in decline
with 20 platform production wells producing oil, gas and water, all with the
aid of gas-lift. A data set spanning 240 days is considered. Production
data from another eld was also studied, but as similar observations were
made a discussion of this eld is omitted for brevity. No multiphase ow
meters are available while one test separator with rate measurements of oil,
gas and water is available. The total rate of produced oil is measured by
scal measurements, while total rates of gas and water can be estimated by
aggregating several measured rates in the processing facilities. No allocated
rates are analyzed in this chapter; conclusions are drawn entirely based on
measured variables. The owner of the production data requested it be kept
anonymous; therefore all gures display normalized variables.
3.3.2 Ratio between well tests and total measured rates
Total production is measured continuously by rate measurements in the
processing facilities for the eld considered. In addition, an estimate of total
production can be found by summing the rates of each measured component
at the most recent well test for each well. When the eld is produced close
to the settings of the most recent well tests and disturbances since the last
well tests are small, these two sets of rates should have comparable values.
The ratio between these two sets of rates (F
o
(t), F
g
(t), F
w
(t)) are dened as
F
o
(t)

=
q
tot
o
(t)
n
w
i=1
z
i
(t)q
test,i
o
(t)
(3.2)
F
g
(t)

=
q
tot
g
(t)
n
w
i=1
z
i
(t)q
test,i
g
(t)
(3.3)
F
w
(t)

=
q
tot
w
(t)
n
w
i=1
z
i
(t)q
test,i
w
(t)
(3.4)
where q
test,i
o
(t), q
test,i
g
(t) and q
test,i
w
(t) are the rates of oil, gas and water,
respectively, that have been measured for well i at the most recent well test
at time t. z
i
is the production choke opening for well i, normalized so that
z
i
= 1 implies a fully open choke, and n
w
is the number of wells.
As long as gas-lift rates q
gl
and choke-openings z are close to the values
at which wells were tested, values of (F
o
(t), F
g
(t), F
w
(t)) close to one indi-
cate that measurements of total production match the sum of the last well
tests, i.e. that there is little measurement uncertainty, and that wells are
tested frequently relative to the rate at which production changes due to dis-
turbances. Values (F
o
(t), F
g
(t), F
w
(t)) dierent from 1 may be an indication
of the opposite.
3.3.3 Data
The average number of single-rate and multi-rate well test per year was 4.87
and 0.29 respectively. Figure 3.2 shows how produced rates of oil and liquid
vary with changing gas lift rates measured during single-rate well tests for
one chosen well. Figure 3.3 shows gas-oil ratios and water cut measured
at the test separator as a function of time. Figure 3.4 shows an example
of an observed response in measured total rates as the gas lift rate of a
well is changed. Figure 3.5 shows a comparison of summed test separator
rates and measured total rates and the estimated ratios F
o
, F
g
, F
w
. Figure
3.6 illustrate the information content of gas lift rates for a time interval,
compared with the number of well tests in the same interval.
3.3.4 Observations
Oftentimes, only single-rate well tests are performed: Single-rate well tests
do not directly provide information on sensitivities, yet such tests far out-
number multi-rate well tests, as 4.87 single-rate well tests were performed
a year relative to only 0.29 multi-rate well tests. This observation can be
attributed to the cost of performing a multi-rate well test. The low fre-
quency of multi-rate well tests will make it dicult to verify sensitivities in
production models and parameters describing sensitivities may be uncertain.
It may be dicult to determine sensitivities by aggregating single-rate
well tests: In principle it is feasible to combine pairs of single-rate well tests
to determine sensitivities. Figure 3.2 illustrates that single-rate well tests
are often performed over a narrow interval of decision variable values. When
determining sensitivities by aggregating single-rate well tests performed in
this manner, estimates may become very sensitive to disturbances that oc-
curred in the time interval between those well-tests. Observed variation in
gas-oil ratios may be due to uncertainty in measured gas-lift rates in this
case.
Some well characteristics remain similar for days, weeks and even months,
while other characteristics vary in ways that may be dicult to model and
may require limiting the length of the tuning set: Figure 3.3 illustrates vari-
ations in gas-oil ratios and watercuts, as measured by well tests for each
individual well.
32
q
o
q
gl
q
o
+q
w
q
gl
t = 2544
t = 3552
t = 3720
t = 4656
t = 4872
t = 5400
0.69 0.7 0.71 0.72 0.73 0.74
0.69 0.7 0.71 0.72 0.73 0.74
0.8
1
1.2
0.4
0.6
0.8
1
Figure 3.2: Case-study: Well tests for a representative well. Top graph:
normalized measured oil rate q
o
against normalized measured gas-lift rate
q
gl
. Bottom graph: sum of normalized oil q
o
and water rates q
w
against
normalized gas-lift rate.
[%]
[%]
[%]
[%]
[%]
[%]
t[hours]
t[hours]
[%]
t[hours]
1 5833 1 5833
1 5833
-40
20
-20
20
-5
5
-10
10
-20
10
-10
10
-10
10
-10
10
-5
10
-20
20
-20
60
-5
5
-50
100
-5
15
-20
10
-50
50
-40
40
0
1
-20
20
-10
10
Figure 3.3: Case-study: Change in gas-oil ratios (diamond) and watercut
(square) relative to the average as measured at the test separator for all
wells plotted in ascending order according to well index.
34
q
tot
q
gl
z
t[hours]
2200 2205 2210 2215 2220 2225 2230 2235
2200 2205 2210 2215 2220 2225 2230 2235
2200 2205 2210 2215 2220 2225 2230 2235
0
0.5
0.5
1
0.9
1
Figure 3.4: Case-study: The observed response in total measured rate of oil
(solid), gas (dashed) and water (dashdot) to the change in a single gas lift
rate q
gl
while all production valve openings z were kept constant.
q
tot
o
q
tot
g
q
tot
w
F
t[hours]
1000 2000 3000 4000 5000
1000 2000 3000 4000 5000
1000 2000 3000 4000 5000
1000 2000 3000 4000 5000
0.8
1
1.2
1.4
1.6
1.8
0
0.2
0.4
0.6
0.8
1
0
0.5
1
1.5
0
0.5
1
1.5
Figure 3.5: Case-study: Top three graphs: a comparison of summed rates
measured at the test separator (solid) to measured total rates of oil, gas and
water (dotted), down-sampled by a factor of 10 for legibility. Bottom graph:
ratios F
o
, F
g
, F
w
.
36
q
tot
z, q
gl
t [hours]
1000 2000 3000 4000 5000
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
0.5
1
1.5
Figure 3.6: Case-study: Top graph: measured total oil production q
tot
o
(solid), measured total gas production q
tot
g
(dashed) and measured total
water production q
tot
w
(dotted), downsampled by a factor 10 for legibility.
Bottom graph: gas lift rates q
gl
(dashed) and relative choke opening z (solid)
and well tests (circles) for all wells plotted in ascending order according to
well index (well index i along y-axis).
The absence of variation in ratios may aid production modeling: For the
eld in question it may be reasonable to make the simplifying assumption
that watercut is time-invariant for the timescales considered.
When variation in ratios are observed, this may be attributable to time-
varying well characteristics, measurement uncertainty or a combination of
the two: For the eld in question and on the timescales considered, variations
of gas-oil ratios on the order of 20% relative to the average are observed. If
this variation is attributed to changes in reservoir or well, either a model of
the dynamics in the production model must be included or these variation
can be treated as an un-modeled disturbance and the permissible length of
the tuning set limited to intervals short enough for this disturbance to be
negligible. In general, the shorter the time interval spanned by the tuning
set, the lower the information content and the more signicant uncertainty
will be in (2.6). As the operator reports that there is no free gas on this
eld, and as the gas-oil ratio depends on gas rate measurements that may be
prone to a higher measurement uncertainty than liquid rate measurements,
one may speculate that observed variation in gas-oil ratios is a result of
measurement uncertainty, especially in measured gas-lift rates. If this is
the case, it may be dicult to uniquely determine the gas-oil ratio from
historical well tests, resulting in uncertainty in the production model.
Measurements of total rates contain information on production of indi-
vidual wells: Figure 3.4 illustrates the observed response in measured total
rates to a change in the gas lift rate of a single well. This response reveals
information on sensitivities which may be exploited in production modeling.
The response considered is large and isolated and could easily be determined
through manual inspection, but oftentimes several wells are excited simul-
taneously or responses are far less obvious, and in such instances it may be
advantageous to extract information using a parameter estimation procedure
of the form (2.6).
The information content of total production rate measurements is a sig-
nicant addition to the information in well test measurements: Figure 3.6
allows the frequency at which decision variables are changed to be compared
with the well test frequency. In principle each variation in z or q
gl
produces
a response in measured total rates, although not all such responses may be
discernible. Figure 3.6 illustrates that the frequency at which z or q
gl
are
changed during normal operations is high comparable to the frequency at
which well tests are performed. Furthermore, most well tests are single-rate,
giving no direct information on sensitivities, but changes to z or q
gl
do yield
information on sensitivities. The overall information content of data, both
of total rates and of test separator measurements, is still quite low. Figure
38
3.6 illustrates that gas lift rates q
gl
and production choke openings z on some
wells remain constant for months. When considering tuning sets spanning
short time intervals, this may mean that some uncertainty should still be
expected in (2.6) even if information in both total rate measurements and
test separator measurements is used to update the production models.
Production varies within a narrow band around what is believed to be
optimum production: Figure 3.6 illustrates that, if shutdowns are excluded,
gas lift rates vary within a relatively small range. This may mean that a
production model can only be veried for a narrow range of decision vari-
able values. Uncertainty in a tted model should be expected to be larger
outside the range of decision variable values observed in Z
N
, and production
optimization may need to take this into account when suggesting changes in
decision variables.
Ratios F
o
, F
g
, F
w
indicate the presence of measurement uncertainties:
Figure 3.5 illustrate that the ratios F
o
, F
w
of oil and water are in the re-
gion of 1.3 and in F
g
is in the region of 1.1 on average. Values of ratios
F
o
, F
g
, F
w
dierent from 1 can indicate measurement uncertainty, but will
also be aected by decline, well dynamics and change in decision variables
since the time of the last well test. As single-rate well tests are performed
relatively frequently, and relatively few changes are made in decision vari-
ables on this eld, as illustrated by Figure 3.5, the observed F
o
, F
g
, F
w
can
be interpreted as indication of the presence of measurement uncertainty on
the eld considered.
Seemingly predictable reservoir dynamics on short timescales: The mea-
sured total oil rate illustrated in Figure 3.6 show a slow and steady decline
in total oil rates. For the eld considered, this may indicate that a relatively
simple description of depletion may be sucient in a production model valid
for short time intervals.
3.3.5 Discussion
The intent of this chapter is to document experiences gained from assessing
production data with regards to modeling for production optimization, and
to illustrate these using data for a particular eld. Not all observations made
in this chapter are expected to be eld-independent, but in my experience
the diculties posed by the state of production data with regards to infor-
mation content and measurement uncertainty are not unique to this eld.
A measurement uncertainty which has not been considered in this chapter,
but which may be signicant, is uncertainty in measurements of the deci-
sion variable. The sensitivities of relative choke openings on production may
vary with time as a production chokes are eroded. This chapter has consid-
ered measured gas lift rates q
gl
, but these rates may be improved through
model-based reconciliation with auxiliary measurements, which has not been
considered. A visible seemingly random or noisy component is visible in
measurements y. This component may be due to measurement uncertainties
or un-modeled disturbances in the topside facilities. The signal-to-noise ra-
tio may become low as the number of wells increases, as the relative change
in total measured rates caused by altering production of an individual well
will diminish, while disturbances stemming from processing facilities and
ow dynamics will not diminish with increasing number of wells.
3.3.6 Challenges in production optimization
Based on the observations made in the previous section, it seems reasonable
to assume that production models will be subject to a degree of uncertainty,
due both to measurement uncertainties and the low information content in
production data. What challenges do these observations raise in the context
of production optimization, and what requirements do they pose for new
production optimization technology? The challenges outlined here will be
addressed in the following chapters.
Determine the signicance of uncertainty: The signicance of produc-
tion model uncertainty depends on how model uncertainty inuences the
settings suggested by production optimization. If model uncertainty causes
production optimization to suggest settings which are suboptimal, a loss in
production prot will result. The signicance of model uncertainty should be
expressed in terms of this loss. An explicit treatment of model uncertainty
will only be warranted on elds where this loss is estimated to be signi-
cant. Research could focus on devising methods to assess the signicance of
uncertainty on a case-by-case basis.
Business cases for uncertainty reduction: Reducing uncertainty can take
on many forms, but will usually incur a cost. As it may not be cost-ecient
to completely eliminate uncertainty, the costs and value of a proposed action
to reduce uncertainty should be weighed. A truly structured approach to
uncertainty reduction would be to formulate a business case, a structured
proposal for business change that is justied in terms of costs and benets,
for each candidate action. Uncertainty due to low information content in
production data can be mitigated through systematic excitation planning.
The cost of excitation is linked to the possible temporary reduction in pro-
duction prot that may result from exciting production, the risks of trigger-
ing a shutdown and possibly costs associated with using the test separator.
The value of excitation is the increase in production prots that results from
re-optimizing production with a production model that has been updated
40
with information gained from a particular excitation. Measurement uncer-
tainty can be mitigated through the careful calibration of measurements
by installing improved measurement equipment or by installing redundant
measurement equipment to allow data reconciliation (Crowe, 1996), and such
actions will also have costs and values which should be weighed. Estimating
costs and values of uncertainty reduction to allow proposed actions to be
evaluated through structured business cases should be a topic of research.
Decision making under uncertainty: Through analysis of structured busi-
ness cases for uncertainty reduction, it is expected that it will not be cost-
ecient to eliminate uncertainty completely. In that case day-to-day oper-
ational decisions will have to be made under uncertainty. If a framework
which describes uncertainty in production models has been developed and
uncertainties are quantied, this framework may be exploited to reduce the
consequences of uncertainty in day-to-day decision making. One example of
explicit uncertainty handling is Bieker et al. (2007b), which deals with pri-
oritizing wells under uncertainty. Decision making under uncertainty should
be a topic of research.
3.4 Conclusion
Production data was observed to have a low information content and be sub-
ject to signicant measurement uncertainties in the case study considered.
Based on these observations, it seems reasonable to assume that production
optimization may be subject to signicant uncertainty. Three challenges for
research are highlighted, which until now have all received little attention in
the literature. Firstly, the lost potential incurred due to uncertainty should
be estimated to assess its signicance. Secondly, costs and values of actions
to reduce uncertainty should be estimated to allow proposed actions to be
evaluated through structured business cases. Thirdly, strategies for making
decisions in day-to-day operations under uncertainty should be investigated.
This thesis will attempt to address these challenges in the following chap-
ters.
Chapter 4
Production optimization;
system identication and
uncertainty estimation
A version of this chapter was presented at Intelligent Energy 08,
Amsterdam, February 2008
Abstract
The contribution of this chapter is twofold. Firstly, this chapter discusses es-
timation of uncertainty in production optimization resulting from tting mod-
els to production data with low information content, a concept that has pre-
viously mainly been applied in reservoir management. Secondly, this chapter
illustrates how a system identication approach can be used to nd produc-
tion models which can be solved with little computational eort and which
are designed to be easily tted to production data.
The method is demonstrated on a synthetic example before being applied to a
case study of a North Sea oil and gas eld. In oshore oil and gas production,
the suggested method is expected to have applications in the development of
structured approaches to uncertainty handling, for instance excitation plan-
ning and production optimization under uncertainty.
41
42
Production optimization; system identication and uncertainty
estimation
4.1 Introduction
There are many reasons why a production model may not describe produc-
tion accurately. One reason may be structural uncertainty, the model may
have a structure which makes it impossible to describe production truly
regardless of the choice of parameters. An important cause of structural
uncertainty may be un-modeled disturbances, inuences which are not ac-
counted for in the model but which cause production to change with time. A
second reason may be measurement uncertainty, measured production may
dier from the actual production for some reason, for instance due to in-
correct calibration of measurement equipment. A third reason may be lack
of informative data, the data may have insucient excitation to uniquely
determine the parameters of the model. In practice all of these factors are
usually present to some extent.
Some physical phenomena in production are only modeled using em-
pirical relationships, due to limited understanding of the physics involved.
Multiphase ow in the reservoir and through ow lines, including gas-lift
performance curves, ow through restrictions, and inow from reservoir into
well, require empirical closure relations (for recent discussion of such em-
pirical relations for ow line, see for instance Ullmann and Brauer (2006),
Tribbe and Mller-Steinhagen (2000), for ow through restrictions see Ras-
toin, Schmidt and Doty (1997), for inow relations see Beggs (2003), and
references therein). Empirical relationships can be tted against laboratory
experiments, but experiments can be costly and small deviations between
laboratory model and eld can produce large dierences in observed ow
(Utvik, Rinde and Valle, 2001). Even the most carefully constructed pro-
duction model will require some tting against production data to reect
the inuence of un-modeled disturbances and structural uncertainty, for in-
stance skin eects near the well, erosion of chokes or the build up of wax or
hydrates in ow lines.
4.1.1 Problem formulation
Cost-eective methods for the designing and updating production models
is a signicant hurdle for the proliferation of production optimization in oil
and gas production. This chapter makes two contributions toward reducing
the cost of designing and implementing production optimization. Firstly,
modeling production with the general methods of system identication is
investigated, motivated by a desire for models which can be easily tted to
production data. Secondly, a study of tting production models to recent
historical production data describing normal operations is made, as such
4.2. Modeling for production optimization under uncertainty 43
data are available at little or no cost. Suggestions are made on how to
quantify uncertainty that may result when such data has low information
content.
4.2 Modeling for production optimization under un-
certainty
In this section outlines a system identication approach to modeling for
production optimization and an approach to quantify parameter uncertainty
when models are tted to data with low information content.
4.2.1 Local production models for optimization
Assume that the oil, gas and water rates measured at the most recent well
test are q
l
, measured at time t
l
and at setpoint (u
l
, d
l
). Consider a model
which is locally valid in the sense that it attempts to predict the rates q
i
for values of (u, d) close to (u
l
, d
l
). The choice to consider locally valid
production models is motivated by two observations. Firstly, a locally valid
model may be sucient as long as production optimization only attempts to
suggest new setpoints close to u
l
, and as long as d
l
has not changed signi-
cantly since the last well test. Secondly, decision variables often vary within
a narrow range in production data from normal operations, as illustrated in
Chapter 3.
To simplify modeling it is assumed that the eects of changes in (u, d, t)
from (u
l
, d
l
, t
l
) on q
i
, the vector of modeled rates for well i, can be described
by separate kernel functions f
u
, f
d
, f
t
:
q
i
= q
l,i
f
i
d
(d
i
, d
l,i
, ) f
i
u
(u
i
, u
l,i
, ) f
i
t
(t, t
l
, ) (4.1)
Each kernel function may be further separated as necessary, f
t
may for in-
stance be divided into kernel functions describing depletion and transients,
and f
u
and f
d
may be divided into kernel functions describing dierent com-
ponents of u and d as necessary. In addition, terms describing measurement
uncertainty may be added as appropriate for joint data reconciliation and
parameter estimation (Crowe, 1996). Kernel functions can be found by dif-
ferent means, either simulators, physical knowledge, well tests or tted to a
tuning set in a black-box manner. The model structure (4.1) may need to
be tailored to describe the characteristics of each particular eld.
A balance is required between models which are too rigid to describe the
observed tuning set and models which are too exible. Models which are too
exible can suer from a phenomenon known as over-tting, where the tted
44
estimation
model describes the tuning data set well while the model describes other data
poorly. Fitted models which are almost equally capable of describing a set
of independent data, a validation data set and the tuning set should be
preferred, as such models do not suer from over-tting (Ljung, 1999).
4.2.2 Estimating parameter uncertainty
It is desirable to exploit production data from normal operations as much
as possible as such data can be obtained at low costs, yet low information
content can result in signicant parameter uncertainty if models are t-
ted to such data. If

is erroneously assumed to describe production while
parameter uncertainty is signicant, production optimization may suggest
setpoint changes u(
) that are infeasible, sub-optimal or may even reduce

prot. Rather than abandon the use of production data from normal oper-
ations, it is proposed to quantify parameter uncertainty. Further work may
focus on exploiting this quantication of uncertainty to devise strategies for
production optimization under uncertainty.
A standard result of system identication is that the matrix
P
= lim
N
1
N
N
t=1
_
y(t,

)
__
y(t,

)
_
T
1
(4.2)
is the covariance matrix of the asymptotic distribution of prediction-error
estimates (Ljung, 1999). Estimates

P
are an expression of parameter un-

certainty which can be found from the tuning set and model, the diagonal
elements of

P
are the variance of each component of .

The derivation of (4.2) requires (t,

) to be a sequence of zero mean
independent variables with variance
0
, as well as some technical conditions
and invokes the central limit theorem. Approximations of P
can be obtained
from a nite set of data of length N, but this approximation can introduce
errors, especially when N is small. In nonlinear identication, numerical
solvers may return estimates

which are one of several local optima rather
than a global optimum of (2.6). In such cases it is not clear that (4.2) will
be a valid description of parameter uncertainty. The matrix product in (4.2)
may be an ill-conditioned matrix when Z
N
has low information content, in
which case matrix inversion may be numerically inaccurate.
Parameter uncertainty can also be estimated numerically using boot-
strapping (Efron and Tibshirani, 1993). Bootstrapping decomposes y(t)
into a systematic, modeled component y(u, d,

) and a stochastic process
e: y(t) = y(u, d, t) + e(t), and assumes that the observed residuals are
4.3. Synthetic examples and a case study 45
a representative distribution of e(t). Bootstrapping uses this assumption
to construct a large number of synthetic measurements with a set of re-
sampled stochastic component, and re-estimates

for each of the synthetic
measurements. Residuals for time-instances where measurement errors or
large un-modeled disturbances cause gross errors should be detected in pre-
analysis and excluded.
The advantage of bootstrapping over asymptotic analysis is that it makes
assumptions about the model and data set which are possibly less stringent,
but nding estimates of uncertainty requires signicant computational ef-
fort as the parameter estimation problem is solved a large number of times
numerically.
4.2.3 Summary
The suggested approach to modeling and parameter uncertainty estimation
is summarized in Algorithm 4.1.
Algorithm 4.1 Given a dataset Z
N
(2.4) of historical data describing pro-
duction, and let N
s
be the desired number of re-samples.
Choose a model structure (2.2) of the form (4.1) as applicable to the
particular eld, and
estimate a nominal

from (2.6). Apply constraints to assist parameter
estimation as applicable.
For the number of re-samples N
s
generate a data set Z
N
r
, by sampling observed residuals N times
and adding them to the output estimated using the process model
and the nominal

, and
determine a re-sampled parameter estimate

r
for Z
N
r
from (2.6).
The distribution of

r,i
i = 1, . . . , N
s
is an estimate of the parameter
uncertainty.
4.3 Synthetic examples and a case study
In this section the suggested approach is validated on a synthetic example
and a case study of actual eld data is performed. All simulations in this
46
estimation
chapter are implemented and solved in MATLAB
1
using the TOMLAB
2
toolbox.
4.3.1 Production model
A well is considered decoupled from other wells when changes in its produc-
tion do not inuence the production of other wells. In this chapter the case of
a eld with n
w
decoupled, gas-lifted wells producing predominantly oil, gas
and water is considered. Change in gaslift rates q
i
gl
=
q
i
gl
q
l,i
gl
1, i 1, 2, . . . n
w
is considered the decision variable u = q
gl
. The relative production choke
opening z [0, 1] is considered a modeled disturbance. Let the most re-
cently measured rate of a given component of production be q
l,i
, measured
for gas lift rate q
l,i
gl
and relative choke opening z
l,i
at time t
l
. As prot
depends on total oil production and constraints are linked to total produc-
tion of gas and water, it is chosen to model the production of oil, gas and
water for each well, ( q
i
o
, q
i
g
, q
i
w
) i = 1, 2, . . . , n
w
. To simplify estimation
of

it is assumed that kernel functions f
i
z
(z
i
, z
l,i
) based solely on physical
knowledge will describe production suciently well. The error introduced
by this assumption should be small, as the production chokes for most wells
are either fully opened or fully closed most of the time. As gas lift rates
are the decision variable, it is chosen to t the parameters of gas-lift kernels
f
i
gl
(q
i
gl
, q
l,i
gl
, ) to production data. It is feasible to model time-variant eects
by interpolating between single-rate well tests, but it was chosen to not in-
clude such eects in the model as test-separator measurements of gas are
unreliable and often exhibit large non-physical variations. As few transients
are visible on the time-scales and sampling rate considered, it was chosen to
not model transients. Discrepancies between test separator and total rate
measurements are assumed to be described suciently by identifying bias
terms
y
for each total rate measurement, and no attempt to compensate
further for measurement uncertainty is made. The form of (4.1) chosen in
this chapter is
q
i
o
= max{0, q
l,i
o
f
z
(z
i
, z
l,i
) (1 +
i
o
q
i
gl
+
i
o
(q
i
gl
)
2
)} (4.3)
q
i
g
= max{0, q
l,i
g
f
z
(z
i
, z
l,i
) (1 +
i
g
q
i
gl
+
i
g
(q
i
gl
)
2
)} (4.4)
q
i
w
= max{0, q
l,i
w
f
z
(z
i
, z
l,i
) (1 +
i
w
q
i
gl
+
i
w
(q
i
gl
)
2
)} (4.5)
Kernel functions f
gl
(q
i
gl
, q
l,i
gl
, ) are chosen as second order polynomials, where
i
o
,
i
g
,
i
w
expresses gradient information and
i
o
,
i
g
,
i
w
expresses curvature
1
The Mathworks,Inc., version 7.0.4.365
2
TOMLAB Optimization Inc., version 5.5
in oil, gas and water gas-lift curves for well i, respectively. f
z
(z
i
, z
l,i
) is a non-
linear kernel which should express the nonlinear relationship between choke
opening and production, which obeys f
z
(0, z
l,i
) = 0 and f
z
(z
l,i
, z
l,i
) = 1. In
this chapter
f
z
(z
i
, z
l,i
) =
1 (1 z
i
)
k
1 (1 z
l,i
)
k
, (4.6)
where k is a design parameter determined oine. The total rates of oil,
gas and water q
tot
o
n
w
i=1
q
i
o
, q
tot
g
n
w
i=1
q
i
g
, q
tot
w
n
w
i=1
q
i
w
are considered
elements in the measurement vector y

=
_
q
tot
o
q
tot
g
q
tot
w
T
. Let the modeled
output for a given parameter

be
y(u(t), d(t),

, t) = R(t) x(u(t), d(t),

) +

y
(4.7)
where

y
is the vector of measurement biases due to calibration inaccuracies
to be determined, R(t) is a matrix of zeros and ones describing routing,
and x(u(t), d(t),

) is an estimate of x(t) found with the production model.
The components of in this chapter are
i
o
,
i
g
,
i
w
and
i
o
,
i
g
,
i
w
for all
wells i = 1, 2, . . . , n
w
and
y
. (4.3)-(4.5) does not assume rate-independent
ratios, and may therefore be able to describe wells with oil-, gas- or water
coning in steady state.
4.3.2 Numerical solution of the parameter estimation prob-
lem
Estimating the parameters of (4.1) is a nonlinear programming problem in
general, while (4.3)(4.5) has a linear-in-variables structure, and

is esti-
mated with the linear-least squares solver lssol in TOMLAB.
The parameter estimation problem can be divided into three separate
sub-problems, one for oil, one for gas and one for water, but it is chosen
to solve these problems as a single parameter estimation problem and use
constraints to link these problems, as will be discussed in the section below.
4.3.3 Physical knowledge in parameter estimation
It is impossible to give an exhaustive list of all conceivable physical con-
straints on , but a short review is given.
The ratios between phases, such as gas-oil ratio or water-oil ratio, are
rate-independent for some wells, and this qualitative knowledge can be used
48
estimation
to simplify the parameter estimation problem. The ratio r between phases
m and n for well i is r
i
m,n
=
q
i
m
q
i
n
. A rate-independent ratio r
i
m,n
obeys
r
i
m,n
=
q
l,i
m
(1 +
i
m
q
i
gl
+
i
m
(q
i
gl
)
2
)
q
l,i
n
(1 +
i
n
q
i
gl
+
i
n
(q
i
gl
)
2
)
q
i
gl
. (4.8)
(4.8) could be enforced by hard constraints
i
m
=
i
n
and
i
m
=
i
n
, or
alternatively by a soft constraint
V
s
() = w
r
_
i
m

i
n
_
2
+w
r
_
i
m

i
n
_
2
. (4.9)
On wells where the gas-oil or water-oil ratios are known to be rate-dependent
it may still be known that these ratios do not vary by more than a given per-
centage and this may be included as dierence constraints. If it is expected
that
q
l,i
m
q
l,i
n
R
L
<
q
i
m
(q
i
gl
)
q
i
n
(q
i
gl
)
<
q
l,i
m
q
l,i
n
R
U
, q
i
gl
, (4.10)
where 0 < R
L
< 1 < R
U
, constraints
i
o
R
L

i
g
,
i
o
R
L

i
g
,
i
o
R
U

i
g
and
i
o
R
U
i
g
could be enforced.
For gas-lifted wells it is reasonable to expect the increasing friction at
increased gas-lift rates to result in performance curves with negative curva-
ture

2
m
i
tot
q
2
gl
< 0, and this qualitative knowledge can be expressed in terms
of . Let m
i
(q
i
gl
) be the mass rate of production from well i for a given
gas lift rate q
i
gl
, and let m
i

= m
i
(q
i
gl
) m
i
(0). If the rate of produced
mass from well i can be assumed to consist predominantly of oil, gas and
water with densities
o
,
g
and
w
, respectively:
m
i
=
o
q
l,i
o
_
i
o
q
i
gl
+
i
o
(q
i
gl
)
2
_
+
g
q
l,i
g
_
i
g
q
i
gl
+
i
g
(q
i
gl
)
2
_
+
w
q
l,i
w
_
i
w
q
i
gl
+
i
w
(q
i
gl
)
2
_
+
g
(q
i
gl
q
l,i
gl
).
(4.11)
with second derivative
2
m
i
tot
(q
i
gl
)
2
= 2 (
o
q
l,i
o

i
o
+
g
q
l,i
g

i
g
+
w
q
l,i
w

i
w
) < 0. (4.12)
(4.12) could be added as a soft-constraint. One technique that can help
reduce over-tting is regularization (Johansen, 1997). Let be a vector
of
i
o
,
i
g
,
i
w
, i = 1, . . . , n
w
and let be a vector of
i
o
,
i
g
,
i
w
, i =
1, . . . , n
w
. Regularization terms
V
s
() = w
reg
(
reg
)
T
(
reg
)+
w
reg
(
reg
)
T
(
reg
)
(4.13)
can be added as soft constrains, penalizing deviation from =
reg
and
=
reg
. w
reg
and w
reg
are weighting parameters.
reg
and
reg
could be
estimates of gradients and curvature determined from multi-rate well tests.
Upper and lower bounds on

, based on knowledge of the expected shape of
the gas lift performance curve, may improve the performance of some solvers.
Such bounds should be loose, i.e. so wide that the solver is expected to nd
solutions within rather than on these bounds when there is any excitation
of the decision variables associated with a given parameter. Loose bounds
also ensure that unrealistic parameter values are not chosen in those cases
when there is very little excitation of certain modes.
4.3.4 A synthetic example
In this subsection the properties and capabilities of the proposed modeling
approach are illustrated on a synthetic example. Consider a eld with 8
gas-lifted wells, each with production of oil q
i
o
, gas q
i
g
, and water q
i
w
, at
rates which vary with the gas lift rate q
i
gl
. Wells are grouped into pairs with
performance given by similar equations. Wells 1 and 2 have rate-independent
gas-oil ratio and water-oil ratio:
q
i
o
= 10 + 0.025q
gl
+
_
q
i
gl
0.000013(q
i
gl
)
2
(4.14)
q
i
g
= 30q
i
o
(4.15)
q
i
w
= 0.5q
i
o
(4.16)
Wells 3 and 4 have gas coning, and a rate-dependent gas-oil ratio is chosen,
while the water-oil ratio is rate-independent:
q
i
o
= 10 + 0.033q
gl
+
_
q
i
gl
0.000017(q
i
gl
)
2
(4.17)
q
i
g
= 30(
q
i
gl
1000
1)q
i
o
(4.18)
q
i
w
=
0.5
1 0.5
q
i
o
(4.19)
50
estimation
Wells 5 and 6 have water coning, and have a rate-dependent water-oil ratio,
while the gas-oil ratio is rate-independent:
q
i
o
= 15 + 0.028q
gl
+
_
q
i
gl
0.000022(q
i
gl
)
2
(4.20)
q
i
g
= 40q
i
o
(4.21)
q
i
w
=
0.4(1 + 0.3(q
gl
/1000 1))
1 0.4(1 + 0.3(q
gl
/1000 1))
q
i
o
(4.22)
Wells 7 and 8 have oil coning, and have a rate-dependent water-oil ratio
while the gas-oil rate is rate-independent:
q
i
o
= 14 + 0.15q
i
gl
0.00003(q
i
gl
)
2
(4.23)
q
i
g
= 35q
i
o
(4.24)
q
i
w
=
0.6(1 0.20(q
i
gl
/1000 1))
1 0.6(1 0.20(q
i
gl
/1000 1))
q
i
o
(4.25)
Suppose that a set of historical measured total production rates and accom-
panying gas lift rates of each well are available, against which the model
can be tted, shown in Figure 4.2. The model is tted with Algorithm 4.1,
rstly with only loose bounds on , secondly knowledge of rate-independent
gas-oil and water-oil ratios were applied as soft constraints where appropri-
ate. The resulting models and model uncertainties are shown in Figure 4.1.
This example illustrates that the parameter estimation is able to gain infor-
mation from measurements of total rates for a synthetic eld with a small
to medium number of wells with varying characteristics. As the synthetic
example has pairs of similar wells, the eect of the level of excitation on the
estimates of uncertainty is visible in Figure 4.1. The matrix P
was close
to singular, with cond(P
) 10
19
, which would make it dicult to obtain
meaningful estimates of parameter uncertainty using asymptotic analysis in
this case.
Although the suggested methodology nds uncertainty estimates which
describe the changing production as gas-lift changes well for most wells,
the method can break down if the information on the gas lift performance
of some wells is virtually non-existent, as is the case for well 2 as shown
in Figure 4.1. Such cases should be detected in post-analysis, and either
further excitation of such wells should be performed, or these wells should
be left out of production optimization.
q
1
o
q
1
g
q
1
w
q
2
o
q
2
g
q
2
w
q
3
o
q
3
g
q
3
w
q
4
o
q
4
g
q
4
w
q
5
o
q
5
g
q
5
w
q
6
o
q
6
g
q
6
w
q
7
o
q
7
g
q
7
w
q
8
o
q
i
gl
q
8
g
q
i
gl
q
8
w
q
i
gl
Figure 4.1: Synthetic example: True performance curve (solid), bootstrap
performance curves without physical constraints (dashed), bootstrap per-
formance curves with physical constraints (dotted), local operating point
(circle) and the span of gas lift rates observed in tuning set (vertical solid).
52
estimation
q
tot
o
q
tot
g
q
tot
w
t
q
gl
0 50 100 150 200 250 300
8
7
6
5
4
3
2
1
660
680
700
720
1.8
2
550
600
Figure 4.2: Synthetic example: Top three graphs: Measured total rates of
oil, gas and water (dotted) compared with estimates of the tted model
without soft constraints (dashed) and with soft constraints (solid). Bottom
graph: normalized gas lift rates of wells 1 through 8 plotted in ascending
order.
4.3.5 Case study: Production data from a North Sea oil eld
In this section, the proposed method is applied to the same set of data con-
sidered in Chapter 3, production data from an oshore North Sea oil eld
producing mainly oil, water, and gas from 20 gas-lifted wells. The opera-
tor of the eld requested that data be kept anonymous and all results are
therefore presented in terms of normalized variables. A tuning set spanning
ve months with sampling time of one hour was considered. To compare the
signicance on estimated uncertainty of including physical assumptions as
soft constraints, the model is tted to production data in two runs. In the
rst run, only loose bounds 0 < < 1 and 1 < < 0 are implemented.
In the second run soft constraints on rate-independent ratios were enforced,
dierence constraints were applied to limit the rate-dependency of gas-oil ra-
tios, chosen as R
L
= 0.7,R
U
= 1.3 for all wells, and the curvature constraint
was enforced. Past well tests indicate that the watercut is rate-independent,
and soft-constraints are included in the second simulation run to enforce
this rate-independence. k is chosen k = 5, so that changes in z
i
have a large
inuence on production when z
i
is small and a small inuence on production
when z
i
is large. To reduce the impact of reservoir depletion and un-modeled
disturbances on estimates, measured oil rates were de-trended as is shown in
Figure 4.3 and in addition residuals in (2.6) were weighted with a forgetting
factor (Ljung, 1999)
w(t) =
Nt
, (4.26)
chosen as = 0.5
1
N1
. As the production model is intended to be a local
description around (q
l
, u
l
, d
l
), total shutdowns were omitted from the tuning
set.
4.3.6 Results
The t between the nominal production model and the tuning data set is
shown in Figure 4.4. A set of N
s
= 100 bootstrap replications were designed
and the parameters were re-estimated, as outlined in Algorithm 4.1. The
tuning sets used when bootstrapping are shown in Figure 4.5. Estimates of
the bootstrap models are compared with the validation set in Figure 4.6. The
resulting models are compared in Figure 4.7. Nominal parameter estimates
result in models and estimates similar to the those shown in and Figure 4.7
and Figure 4.6.
54
estimation
q
tot
o
t[hours]
0 500 1000 1500 2000 2500
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
Figure 4.3: Case-study: Measured oil rates (dotted) and de-trended oil rates
used in tuning set (solid).
q
tot
o
q
tot
g
q
tot
w
t[hours]
2000 2200 2400 2600 2800
0.8
0.9
1
0.6
0.8
1
0.8
1
Figure 4.4: Case study: Last portion of tuning set (solid) compared with
estimates of the model including physical constraints (dashed). Sampling
time reduced to ve hours for clarity.
56
estimation
q
tot
o
q
tot
g
q
tot
w
t[hours]
2000 2200 2400 2600 2800
0.8
0.9
1
0.6
0.8
1
0.8
1
Figure 4.5: Case-study: Tuning set (solid) and one modeled production for a
selected paramter estimate found through bootstrapping (dashed), for model
including physical constraints.
q
tot
o
q
tot
g
q
tot
w
t[hours]
3000 3200 3400 3600 3800
0.8
0.9
1
0.9
1
1.1
0.8
0.9
1
Figure 4.6: Case-study: Estimates found with model for dierent parameter
estimates found through bootstrapping (dotted) compared with the valida-
tion data set (solid), for model including physical constraints.
58
estimation
q
i
o
i = 1
q
i
g
q
i
w
i = 2
i = 3
i = 4
i = 5
i = 6
i = 7
i = 8
i = 9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
q
i
gl
i = 20
q
i
gl
q
i
gl
Figure 4.7: Case-study: Modeled rates of oil, gas and water for dierent
tted parameters found through bootstrapping, without physical constraints
(dashed), and with physical constraints (dotted). Local operating point
(circle) and the span of gas lift rates observed in tuning set(vertical solid).
Lower left bound of all plots is (0,0). Well indices along left margin.
4.3.7 Discussion
An advantage of the chosen modeling approach is that estimated models
can be interpreted by the shape of performance curves, as in Figure 4.7,
with which industry practitioners may already be familiar. In Figures 4.4
and 4.6 several large changes in measured rates are observed which cannot
be explained with changes in z or q
gl
. These deviations could be the result
of measurement error or disturbances stemming from the reservoir or top-
side process facilities. As the model is intended for determining day-to-day
operating setpoints rather than prediction for long time intervals, the bias
between model estimates and measurements observed in Figure 4.6 is less
signicant than the ability of the model to predict the change in production
resulting from changes in setpoints shown in Figure 4.6. Figure 4.7 illus-
trates that the uncertainty varies greatly between dierent phases and for
dierent wells. Comparing the relative degree of uncertainty for dierent
wells may be useful for excitation planning. In this chapter it has been as-
sumed that the only measurement uncertainty is a static bias in total rate
measurements
y
. Measurement uncertainty in rate measurements at the
test separator may cause errors in the operating point, and the signicance
of such errors on estimates or how to mitigate or estimate these uncertainties
has not been studied. The assumption of time-invariant production in the
tuning set is clearly an approximation, as a falling trend in oil production
due to depletion is visible and some wells have varying phase ratios during
the time period spanned by the tuning set. A relatively long tuning set
errs on the side of caution with regards to estimated uncertainties, but risks
introducing biases in parameter estimates. Reservoir dynamics and changed
processing conditions are treated as un-modeled disturbances in this chapter
and only the most recent well test was used. A visible decline in measured oil
production was treated by de-trending and introducing a forgetting factor.
It may be possible to improve on this approach through modeling distur-
bances and by exploiting older well tests. Although focus has been on gas-
lifted wells in this chapter, the system identication approach to modeling
is general and has extensions to other types of elds and wells, for instance
wells where coupling is signicant or wells where production choke settings
are decision variables. The approach to estimating uncertainty resulting
from a low information content is general and has extensions to estimating
the uncertainty in proxy-models derived from commercial simulators as well.
The model and uncertainty estimates described may be applied to estimate
the signicance of uncertainty on production prots, for formulating struc-
tured business cases for uncertainty reduction or for designing structured
approaches to decision making under uncertainty, as proposed in Chapter 3.
60
estimation
4.4 Conclusion
The contribution of this chapter is twofold. Firstly, it is discussed how pa-
rameter uncertainty can be quantied when models are tted to data with
little information content, a concept that has been little explored in the con-
text of modeling for production optimization. Secondly, it is suggested how
a system identication approach can be used to nd models for production
optimization with the aim of reducing costs of design and maintenance of
production optimization. The method was demonstrated on a synthetic ex-
ample before being applied to a case study of real eld data from a North
Sea oil and gas eld. The method suggested in this chapter has applications
in day-to-day production optimization and in the development of structured
approaches to handling uncertainty, such as excitation planning/well test
planning or production optimization under uncertainty.
Chapter 5
Oil and gas production
optimization; lost potential due
to uncertainty
A version of this chapter was presented at the IFAC World Congress,
Korea, July 2008
Abstract
The information content in measurements of oshore oil and gas production
is often low, and when a production model is tted to such data, uncer-
tainty may result. If production is optimized with an uncertain model, some
potential production prot may be lost. The costs and risks of introducing
additional excitation are typically large and cannot be justied unless the po-
tential increase in prots are quantied. In Chapter 4 it is discussed how
bootstrapping can be used to estimate uncertainty resulting from tting pro-
duction models to data with low information content. In this chapter it is
proposed how lost potential resulting from estimated uncertainty can be esti-
mated using Monte-Carlo analysis. Based on a conservative formulation of
the production optimization problem, a potential for production optimization
in excess of 2% and lost potential due to the form of uncertainty considered
in excess of 0.5% was identied using data from a North Sea eld.
61
62
Oil and gas production optimization; lost potential due to
uncertainty
5.1 Introduction
Lost prot can result if production models are tted to production data with
a low information content, as illustrated in Example 5.1.
Example 5.1 (Uncertainty in production optimization) This exam-
ple illustrates a synthetic eld which is producing at its processing con-
straints, where only single-rate well tests are available and where a produc-
tion model matches historical production data with a low information content
closely. The example is motivated by observations of production data from
actual elds. The eld produces oil q
o
, water q
w
and gas q
g
from two wells,
each with a gas lift rate q
gl
according to equations
q
i
o
= c
i
1
+c
i
2
(q
i
gl
q
l,i
gl
) +c
i
3
(q
i
gl
q
l,i
gl
)
2
i {1, 2} (5.1)
q
i
g
= K
i
GOR
q
i
O
(5.2)
q
i
w
=
K
i
WC
1 K
i
WC
q
o
+w
1
. (5.3)
q
l,i
gl
are local operating points and q
l,1
gl
= q
l,2
gl
= 1000. Let
i

=
_
c
i
1
c
i
2
c
i
3
K
i
WC
K
i
GOR
w
1
(5.4)
and
,1
,2
39.62 0.025 4 10
6
0.173 30 0
76.62 0.030 4 10
6
0.173 30 0
39.62 0.033 6 10
6
0.185 30 0.85
76.62 0.032 12 10
6
0.160 30 1.5
(5.5)
The ratio of water to produced liquid (watercut) and the gas-oil ratio are
assumed rate-independent. True parameters (
,1
,
,2
) and tted parameters
(
1
,

2
) produce predictions of measured total oil q
tot
o
= q
1
o
+ q
2
O
, total gas
q
tot
g
= q
1
g
+ q
2
g
and total water q
tot
w
= q
1
w
+ q
2
w
which match closely for a
typical set of gas lift rates with little excitation, shown in Figure 5.1, and
which match the single-rate well test available for (q
1
gl
, q
2
gl
) = (1000, 1000).
Assume that at the current setpoint (q
1
gl
, q
2
gl
) = (1000, 1000) with oil pro-
duction q
tot
o
= 116.1, all gas processing capacity (associated gas + gas lift)
and water processing capacity is utilized. The current setpoint is optimal
according to numerical optimization with parameters (
1
,

2
), yet with the
parameters (
,1
,
,2
) (q
1
gl
, q
2
gl
) (750, 1250) is optimal with oil production
q
tot
o
= 116.4. The two models and the setpoints are illustrated in Figure 5.1.
5.1. Introduction 63
q
1
gl
q
1
o
q
1
gl
q
1
g
q
1
gl
q
1
w
q
2
gl
q
2
o
q
2
gl
q
2
g
q
2
gl
q
2
w
Figure 5.1: Example 5.1: True (solid) and modeled (dashed) production of
oil, gas and water, well 1 (upper graphs) and well 2 (lower graphs). Optimal
operating point of the true production (square) and optimal operating point
of modeled production (circle).
64
uncertainty
q
tot
O
q
tot
G
q
tot
W
q
i
gl
t[hours]
0 50 100 150 200 250 300
0 50 100 150 200 250 300
0 50 100 150 200 250 300
0 50 100 150 200 250 300
500
1000
1500
24
26
28
3200
3400
3600
100
110
120
Figure 5.2: Example 5.1: True (solid) and modeled (dashed) total produc-
tion rates of oil, gas and water resulting from gas lift rates (lower graphs)
q
1
gl
(dotted) and q
2
gl
(dashdot).
5.2. Estimating the loss due to uncertainty 65
This example illustrates that lost prot can result even if production op-
timization is based on a model that ts production data well if production
data has low information content.
5.1.1 Prior work
The concept of a loss resulting from uncertainty is well established in con-
trol theory, it is for instance used to select controlled variables in Halvorsen
et al. (2003). The concept of the value of information has been applied
previously in reservoir management. In Branco et al. (2005) an analysis of
what the authors called the value of the new information was used to jus-
tify investments to determine static and dynamic reservoir behavior. Monte
Carlo simulations have been applied for analysis of uncertainty in produc-
tion optimization previously in (Bieker et al., 2006), where it was applied
to well test planning, and in (Bieker et al., 2007b), where it was used for
optimal well management. In both cases, the focus was on uncertainty in
gas-oil and water-oil ratios for wells where these ratios do not change when
production change and where no gas-lift was applied and only test separator
measurements were considered.
In Bieker et al. (2007b) it has been proposed that uncertainty in produc-
tion optimization should be related to production prot to allow structured
business cases for uncertainty mitigation to be created. In Chapter 4 it was
suggested how uncertainty due to low information content in production
data can be quantied using bootstrapping. No references have been found
which discuss the value of information in the context of daily oil and gas
production optimization.
The objective of this chapter is to develop an analytical framework to quan-
tify the lost potential that results from the low information content in pro-
duction data used for production optimization, suitable to oshore oil and
gas production.
5.2 Estimating the loss due to uncertainty
Bootstrapping is a computational approach which generates an estimate of
uncertainty by re-solving (2.6) a number of times to t the model to number
of synthetically generated, or re-sampled, data sets similar to (2.4). In
Monte-Carlo simulations in subsequent sections the entire parameter vectors
are sampled rather than sampling parameter values for each component of
66
uncertainty
separately, ensuring that co-variance between terms in

are implicitly
accounted for.
5.2.1 Denitions
Let the potential for production optimization P
o
be the dierence between
the production prot at the optimal operating point (x
, u
, d) and the cur-

rent operating point (x, u, d):
P
= M(x
, u
, d) M(x, u, d) 0. (5.6)
If there were no uncertainty, (2.1)(2.3) would yield u() = u
and x() = x
,
P
o
could be found from (5.6) and the entire potential could be realized by
implementing u().
When

is uncertain, the solution to (2.1)(2.3) will be uncertain. Un-
certainty may cause estimates u() and

P
o
(
) to dier from u
and P
o
, and
when implementing u() it must be expected that some of the potential P
o
to remain unrealized. This discussion motivates
expected potential of production optimization(

P
o
)

=
expected realizable potential (

P
o,r
) +
expected lost potential due to uncertainty (
L
u
).
(5.7)
Let
Q

=

L
u
P
o
, (5.8)
be the quotient of potential that is lost due to uncertainty.
In this chapter, probability density functions f
Po
(

P
o
), f
Lu
(
L
u
) will be
determined. In addition to depending on parameter uncertainty, these prob-
ability density functions will depend on production constraints and on the
operational strategy, dened below.
P
o
= 0 and

L
u
= 0 can be interpreted as production being optimal. The
probability density functions f
L
u
(
L
u
) and f
P
o
(

P
o
) may consist of spikes near
P
o
= 0 and

L
u
= 0 and a distribution in interval away from zero. To allow
these two parts of f
L
u
(
L
u
) and f
P
o
(

P
o
) to be studied separately, the two
events
H
0
: production is optimal (5.9)
H
1
: production is suboptimal (5.10)
are dened. Estimates p(H
0
) and p(H
1
) can be found from estimated prob-
ability density functions f
L
u
(
L
u
) and f
P
o
(

P
o
).
5.2. Estimating the loss due to uncertainty 67
5.2.2 Estimating the potential of production optimization
An estimate

P
o
(
) of (5.6) for a given parameter estimate

when (x, u, d)
is the current operating point is
P
o
(
) = M( x(
), u(
), d) M(x, u, d). (5.11)

When

is uncertain, an estimate of the probability density function f
)
may be exploited to estimate a probability density function for (5.11) using
a Monte Carlo method, generating suitable random numbers and observing
the fraction of the numbers obeying some property or properties (Metropolis
and Ulam, 1949). The approach is outlined in Algorithm 5.1 in Section 5.5.
5.2.3 Estimating lost potential due to uncertainty
A setpoint u(
) calculated with an uncertain

may be found to violate
production constraints (2.3) when implemented. Therefore, a strategy for
constraint handling is required. Lost potential due to uncertainty will de-
pend on the strategy employed to deal with constraint handling, and to
estimate lost potential due to uncertainty this strategy must be expressed
as an algorithm.
Let u
0
be the setpoint implemented on the eld prior to implement-
ing u(
). When altering setpoints from u

0
along some path toward u(
),
constraints may be encountered at an operating point (x
c
, u
c
) due to un-
certainties. Lost potential due to uncertainty can be estimated in a Monte
Carlo fashion by determining the prot that is lost when a setpoint u(
f
) is
implemented and the true parameter value is actually

t
, drawing estimates
(
t
,

f
) from the estimated probability density function f
):
L
u
(
t
,

f
, d) = M( x(
t
), u(
t
), d)
M(x
c
(
f
,

t
), u
c
(
f
,

t
), d).
(5.12)
The rst term in (5.12) is the optimal prot with the true parameter value
t
, while the second term is an estimate of the prot obtained when

f
is erroneously assumed to be the true parameter. (x
c
(
f
,

t
), u
c
(
f
,

t
)) is
determined by an operational strategy function. The proposed algorithm
for determining f
L
u
(
L
u
) is summarized in Algorithm 5.2 in Appendix 5.5.
5.2.4 Discussion
In Algorithms 5.1 and 5.2 samples are drawn at random, using a Monte-Carlo
approach. It may be possible to improve accuracy and reduce computa-
68
uncertainty
tional burden through the use of sampling techniques based on Hammersley-
sequences, see Diwekar and Kalagnanam (1997). Algorithms 5.1 and 5.2 can
also be combined with more ecient bootstrap methods, see for instance
Gigli (1996). To limit the scope and complexity of this chapter, such exten-
sions are left for further work.
5.3 Applications
This section studies applications of the suggested approach, rst on a set of
synthetic examples, then on data from a real eld.
5.3.1 Introduction
Oil, gas and water rates for each well i are considered elements in x, that is
q
i
o
, q
i
g
and q
i
w
, respectively. The production of individual wells are assumed
independent of each other, or decoupled, as will typically be the case for so-
called platform wells. Two alternative production models, (5.13) and (5.14),
are considered
q
i
p
= max{0, q
l,i
p
f
z
(z
i
, z
i,l
)
_
1 +
i
p
q
i
gl
_
} (5.13)
q
i
p
= max{0, q
l,i
p
f
z
(z
i
, z
i,l
)
_
1 +
i
p
q
i
gl
+
i
p
(q
i
gl
)
2
_
}, (5.14)
for all wells i and for oil, gas and water p {o, g, w}, based on Chapter 4.
q
i
gl
is gas lift rate and z
i
[0, 1] is the relative production choke opening of
well i. q
i
gl
=
q
i
gl
q
l,i
gl
1, where q
l,i
gl
is the gas lift rate measured at the time
of optimization for well i, and q
l,i
p
is the measured rate of component p of
well i at the most recent well test, respectively. f
z
(z
i
, z
l,i
) is a nonlinear
kernel which expresses the nonlinear relationship between choke opening
and production, which obeys f
z
(0, z
l,i
) = 0 and f
z
(z
i
, z
l,i
) = 1. z
l,i
is the
relative choke opening at the most recent well test. In this chapter f
z
(z
i
, z
l,i
)
is chosen
f
z
(z
i
, z
l,i
) =
1 (1 z
i
)
k
1 (1 z
l,i
)
k
, (5.15)
5.3. Applications 69
with the design parameter k = 5. Let
q
tot
o
=
n
w
i=1
q
i
o
(5.16)
q
tot
g
=
n
w
i=1
q
i
g
(5.17)
q
tot
w
=
n
w
i=1
q
i
w
(5.18)
where n
w
is the number of wells. Let q
tot
o
(
), q
tot
g
(
) and q
tot
w
(
) be estimates
derived by combining (5.16)(5.18) with (5.13) or (5.14) for a given

.
This chapter considers the following measurement vector when solving
(2.6)
y =
_
q
tot
o
q
tot
g
q
tot
w
T
. (5.19)
Let the modeled output for a given parameter

be
y(u(t), d(t),

, t) = R(t) x(u(t), d(t),

) +

y
(5.20)
where

y
is the vector of measurement biases due to calibration inaccu-
racies to be determined and x(u(t), d(t),

) is an estimate of x(t) found
with the production model. u, d and are chosen u = q
gl
, d = z,
=
_
o

g

w

y
for (5.13), and =

_
o

g

w

o

g

w

y
for (5.14). The objective function is chosen as (5.21) and production con-
straints as (5.22)-(5.25):
u(
) = arg max
u
q
tot
o
(
) +b
o
(
) (5.21)
s.t. q
tot
g
(
) +
n
w
i=1
q
i
gl
+b
g
(
) q
max
g
(5.22)
q
tot
w
(
) +b
w
(
) q
max
w
(5.23)
U
prc
u U
prc
(5.24)
u
i
> u
l,i
i W
c
. (5.25)
(5.22) is a constraint on gas processing capacity and q
max
g
is the process-
ing capacity for gas. (5.23) is a constraint on water processing capacity and
q
max
w
the processing capacity for water. (5.25) constrains the gas lift rate of
wells with indices in the set W
c
which may not have their gas-lift rates de-
creased due to ow assurance issues. b
o
(
), b
g
(
) and b
w
(
) are biases which

70
uncertainty
may be calculated separately for each

. All constraints are considered hard
in this chapter.
What model structure describes the relation between u and y is itself
uncertain. It is reasonable to expect a tted model to be a valid description
locally around setpoints (u, d) similar to those observed in the tuning set.
The mismatch between modeled outputs y and measurements y will grow
when setpoints outside the range of setpoints observed in (2.4) are consid-
ered. If production optimization considers large changes in u, structural
uncertainties may dominate and this observation motivates (5.24). U
prc
lim-
its u to within U
prc
100% of its current value and is a design paramter that
must be chosen suciently small.
q
max
g
and q
max
w
are themselves uncertain in practice, but identifying or
mitigating this uncertainty is beyond the scope of this chapter. To avoid
overstating q
max
g
and q
max
w
, and thereby overestimating

L
u
and

P
o
, it is
assumed conservatively that the eld is producing at its constraints in water
and gas capacity at the time of optimization. The aim is to solve (5.21)
(5.25) at time t = N, the end of the tuning set (2.4). Measurements y have
a high-frequency, low-amplitude, seemingly random component which is not
modeled. If q
max
g
= q
tot
g
(N) and q
max
w
= q
tot
w
(N), solutions of (5.21)(5.25)
will depend to some degree on these high-frequency disturbances. For this
reason, tuning sets are considered where no signicant changes are visible in
u(t) and no large disturbances visible in y(t),

d(t) for the interval t [N
L, N] where L << N. Capacities are estimated as average measured rates
during this interval, to reduce dependency on high-frequency disturbances:
q
max
g
= avg( q
tot
g
(t) +
n
w
i=1
q
i
gl
(t)) t [N L, N] (5.26)
q
max
w
= avg( q
tot
w
(t)) t [N L, N]. (5.27)
Similarly, q
tot
o
(
), q
tot
g
(
) and q
tot
w
(
) should equal averaged observations over

t [N L, N] which motivates
b
o
(
)
b
g
(
)
b
w
(
= avg(
q
tot
o
(t) q
tot
o
(
, t)
q
tot
g
(t) q
tot
g
(
, t)
q
tot
w
(t) q
tot
w
(
, t)
) t [N L, N]. (5.28)
b
o
(
) is included in the objective function (5.21) so that (5.11) yields

P
o
= 0
if (5.21)(5.25) returns u() equal to the implemented u at the time of
optimization. (b
o
(
), b
g
(
), b
w
(
)) may dier from elements of

y
, as

y
is
an estimate of the bias over in the tuning set, while b
o
(
), b
g
(
) and b
w
(
)
are biases calculated for a short time interval at the end of the tuning set.
The operational strategy function
u
c
(
f
,

t
) = u
0
+s( u(
f
) u
0
) (5.29)
0 = f(x
c
(
f
,

t
), u
c
(
f
,

t
), d,

t
) (5.30)
where s = max s (5.31)
s.t 0 s 1 (5.32)
u = u
0
+ s( u(
f
) u
0
) (5.33)
0 c(x, u, d). (5.34)
0 = f(x, u, d,

t
) (5.35)
is considered. (5.29)-(5.35) simulates moving the setpoint from u
0
toward
u(
f
) in a straight line until a constraint is met, when the system is de-
scribed by

t
. If implementing (x
c
(
f
,

t
), u
c
(
f
,

t
)) results in a decline in
production prots, it is assumed that the operating point is returned to its
value prior to optimization when solving (5.12).
The set of operating points which are feasible is non-convex in general,
and the conservative statement of gas and water capacities q
max
g
and q
max
w
means that (5.29)(5.35) may meet a constraint immediately if starting at
u
0
, which would result in large estimates of loss due to uncertainty

L
u
.
Instead, it is chosen to start simulations of (5.29)(5.35) in the lowest value
of u which obeys constraints, which results in far more conservative estimates
of

L
u
in simulations. The real-world analogy to this choice is a eld where
production is moved toward a calculated setpoint u(
) as production is re-
started after a shutdown.
Two dierent formulations of the problem (5.21)-(5.25) are considered:
Formulation A: production model (5.13), a linear programming problem
Formulation B: production model (5.14), a nonlinear programming prob-
lem
The applications in the following section were implemented in MATLAB
1
and all optimization problems were solved with the TOMLAB
2
toolbox. All
nonlinear programming problems were solved sequentially with global solvers
(Jones, Perttunen and Stuckman, 1993) and local solvers based on sequen-
tial quadratic programming, for parameter estimation based on (Huschens,
1994), and for production optimization based on (Schittkowski, 1983). In
all cases analytical Jacobians and Hessians were supplied.
1
2
72
uncertainty
5.3.2 Synthetic example
Example 5.1 is revisited in the following, to illustrate the suggested method-
ology on a set of data where the true model is known.
Example 5.2 (Example 5.1 revisited: f
P
o
(

P
o
) and f
L
u
(
L
u
)) The results
of applying Algorithms 5.1 and 5.2 to the synthetic production data described
in Example 5.1 with N
t
= N
f
= 100 for dierent U
prc
are shown in Figure
5.4. For this example, all estimates produced by Algorithm 5.1 gave

P
o
> 0
using both formulations A and B, or p(H
1
) = 100%.
The use of a nonlinear production model in formulation B resulted in
estimates

P
o
and

L
u
that were less dependent on U
prc
than those found with
formulation A. While estimates of

P
o
remained in the region of P
o
= 0.29%
with formulation B as U
prc
grows, estimates

P
o
found with formulation A
grow without bounds as U
prc
grows. A closer analysis reveals that formula-
tion A tends to nd estimates u() at the constraint (5.24), while formula-
tion B nds estimates u() within this constraint, and this may explain the
dierence observed.
As seen in Figure 5.3, the linear models used in formulation A predict oil,
gas and water rates within a narrow span as q
i
gl
change compared with the
nonlinear models used in formulation B. The narrow span in predicted rates
with formulation A causes comparatively small estimates (
L
u
,

Q) compared
with formulation B. Estimates of

Q found with formulation A decreases as
U
prc
increases, which seems counter-intuitive.
This example indicates that nonlinear models on the form (5.14) may
be more suitable than linear models (5.13) in conjunction with the methods
suggested in this chapter.
5.3.3 Case study: North Sea eld
The set of data studied is from a North Sea eld with 20 gas lifted platform
wells, studied previously in Chapters 3 and 4, producing predominantly
oil, gas and water. Sporadic single-rate well tests are available, and no
routing options are considered in u. The operator of the eld requested that
all data be kept anonymous, therefore all variables will be represented in
normalized form. A tuning set spanning 180 days is chosen. The sampling
time was one hour, and production was not signicantly changed for the
last L = 10 hours prior to t = N. f
P
o
(

P
o
) was estimated with Algorithm
5.1 and f
L
u
(
L
u
) was estimated with Algorithm 5.2, both with N
f
= N
t
=
100. When estimating parameters, bounds 0 < < 1 and 1 < < 0
based on physical knowledge were applied. Based on the knowledge that
q
1
gl
q
1
o
q
1
gl
q
1
g
q
1
gl
q
1
w
q
2
gl
q
2
o
q
2
gl
q
2
g
q
2
gl
q
2
w
Figure 5.3: Example 5.2: Production models (5.13) (dotted) and (5.14)
(dashed) based on parameters found with bootstrapping, operating point
(x
l
, u
l
) (circle). Vertical line illustrates span of decision variable values in
tuning set. The solid line illustrates the true model. To reduce clutter, only
every fth model based on re-sampled parameters are plotted.
74
uncertainty
f
P
o
|H
1
(%)
f
L
u
|H
1
(%)
f
Q|H
1
U
prc
0.1 0.2 0.3 0.4 0.5
0.1 0.2 0.3 0.4 0.5
0.1 0.2 0.3 0.4 0.5
0.2
0.4
0.6
0.8
0.05
0.1
0.15
0.2
0.2
0.4
0.6
0.8
1
Figure 5.4: Example 5.2: estimated potential P
o
, lost potential L
u
and
quotient of potential lost to uncertainty Q for formulation A (plus) and
B(circles). Center graphs are expected values, upper and lower graphs de-
note 95% condence intervals.
5.4. Conclusions 75
the watercut is rate-independent, soft constraints which penalize deviation
from
o
=
w
and
o
=
w
were added to the objective function. A decline
in q
tot
o
was visible in the tuning set, and it was chosen to de-trend q
tot
o
and chose w(t) such that older measurements were weighed less than newer
measurements when solving (2.6). For further details on the modeling, refer
to Chapter 4. The set of indices of wells with ow assurance issues was
W
c
= {1, 4, 10, 13, 14, 16}.
5.3.4 Results
The distribution of estimates

P
o
,

L
u
and

Q for varying U
prc
are shown in
Figure 5.5. The models used are shown in Figure 5.6. For formulation A
p(H
1
) = 100% while p(H
1
) = 95% for formulation B.
5.3.5 Discussion
U
prc
limits change in decision variables to a certain range for which struc-
tural uncertainties can be neglected, and

P
o
and

L
u
depend on this design
parameter. As it is not known with certainty when structural uncertainties
become signicant, a single value for U
prc
cannot be determined and so when
estimating

P
o
and

L
u
a range of U
prc
is considered. A conservative assump-
tion of U
prc
= 0.1 indicates P
o
2%. The quotient Q is relatively insensitive
to U
prc
, so from the analysis it seems fair to conclude that on average, a loss
of about 25% of

P
o
due to uncertainty should be expected, or

L
u
0.5%
when U
prc
= 0.1. Unfortunately, estimates

P
o
and

L
u
made in this chap-
ter cannot be veried without further experiments, and such experiments
are associated with great cost and risk. However, the methods presented
in this chapter may be used motivate such experiments in further work, as
the presented methods allow the associated benets to be estimated. In this
chapter constraint uncertainty has not been considered. It is possible that at
the time of the analysis the eld was producing below its actual processing
capacities. In this case, the potential for production optimization would be
larger than our estimates.
5.4 Conclusions
Based on a conservative formulation of the production optimization problem,
a potential for production optimization in excess of 2% and lost potential due
to the form of uncertainty considered in excess of 0.5% could be identied.
As sales revenues from an oil and gas eld may be very large and the costs
of mitigating uncertainty are independent of revenue, it is believed that
76
uncertainty
f
P
o
|H
1
(%)
f
L
u
|H
1
(%)
f
Q|H
1
U
prc
0.1 0.2 0.3 0.4 0.5
0.1 0.2 0.3 0.4 0.5
0.1 0.2 0.3 0.4 0.5
0.1
0.2
0.3
0.4
0.5
0
1
2
3
2
4
6
8
10
Figure 5.5: Case study, North Sea eld data: estimated potential P
o
, lost
potential L
u
in percent of current production, and quotient of potential lost
to uncertainty Q for formulation A (plus) and B (circles). Center graphs are
expected values, upper and lower graphs denote 95% condence intervals.
5.4. Conclusions 77
q
i
o
i = 1
q
i
g
q
i
w
i = 2
i = 3
i = 4
i = 5
i = 6
i = 7
i = 8
i = 9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
q
i
gl
i = 20
q
i
gl
q
i
gl
Figure 5.6: Case study, North Sea eld data: Production models (5.13)
(dotted) and (5.14) (dashed) based on parameters derived by bootstrapping,
operating point (x
l
, u
l
) (circle). Vertical lines illustrate the range of decision
variable values in the tuning set. Lower left bound of all plots is (0,0). Well
indices along left margin.
78
uncertainty
further work may prove that these potentials are sucient to justify the
costs of reducing uncertainty and optimizing production.
The methodology presented may have extensions to other processes op-
timized using models tted to process measurements with low information
content.
5.5 Algorithms
Algorithm 5.1 (Estimating f
P
o
(

P
o
)) Given a dataset Z
N
(2.4)
determine

using (2.6),
determine f
), for instance by bootstrapping as described in Chapter

4,
sample f
) N
t
times,
determine ( x(), u()) using (2.1)-(2.3) for each of the N
t
samples,
and
insert all determined ( x(), u()) in (5.11) to obtain the probability
density function f
P
o
(

P
o
).
L
u
(
L
u
)) Given a dataset Z
N
of the form
(2.4), an operational strategy function, the current operating point (x, u, d)
and a probability density function f
).
do N
t
times
draw a sample

t
from f
).
draw N
f
samples

f
from f
).
for each of the N
f
samples of

f
and the sample

t
obtain a point
estimate of L
u
by solving (5.12)
the distribution of N
t
N
f
point estimates of

L
u
is an estimate of
f
L
u
(
L
u
)
Chapter 6
A structured approach to
optimizing oshore oil and gas
production with uncertain
models
A version of this chapter has been submitted to Computers & Chemical
Engineering
Abstract
Production model uncertainty may cause a setpoint suggested by optimization
to violate constraints when implemented. The change in prots that result
from implementing a suggested setpoint change may also be uncertain, it
may even be negative, and therefore some result analysis of the optimization
may be required. Implementing excitation in specic decision variables prior
to optimization may benet optimization enough to justify the costs, and a
structured approach to excitation planning to estimate and weigh these costs
and benets is desirable.
Two approaches toward structured optimization with uncertain models are
suggested which are suitable for optimization of oshore oil and gas pro-
duction. Firstly, an operational strategy scheme to ensure feasibility by
exploiting measurements postoptimization is investigated. Secondly, the ex-
ploitation of estimates of parameter uncertainty for structured optimization
result analysis and excitation planning is investigated.
79
80
A structured approach to optimizing oshore oil and gas
production with uncertain models
The suggested approaches are evaluated in simulations on a model designed
to mimic a real-world oil eld including its typical low-information content
data as closely as possible.
6.1 Introduction
As a consequence of the complexity of the process considered, of measure-
ment diculties, and of the low information content in data, production
models may be subject to signicant uncertainty that can be reduced against
a cost or not at all. As a result more than one setpoint can be the plausible
optimum, which raises several practical challenges. One challenge is that the
result of a setpoint change is uncertain, it may even be negative. Another
challenge is that ensuring that an implemented setpoint is feasible, i.e. that
it obeys production constraints, is non-trivial in the presence of uncertainty.
A third challenge is that when uncertainty can be reduced at a cost, some
method of determining how and when uncertainty reduction is cost-eective
may be required.
The current common practice in oshore production of oil and gas is to
either neglect uncertainty or to assess uncertainty subjectively based on the
experience and process insight of the practitioner. A structured approach to
handling uncertainty would provide support for practitioners and improve
transparency of decision-making.
6.1.1 Prior work
Production optimization can be seen as a form of real-time optimization
system, a closed-loop controller which attempts to locate the optimal set-
point through a series of steps, each consisting of a smaller setpoint change.
Real-time optimization systems, originally developed for chemical processes,
consider an economic objective function maximized subject to a rigorous
steady-state nonlinear process model and process constraints. Real-time
optimization is a form of model-based control, and obtaining the process
model is considered the single most dicult and time-consuming task in
the application and maintenance of model-based control and optimization
(Andersen, Rasmussen and Jorgensen, 1991) (Terwiesch, Agarwal and Rip-
pin, 1994) (Zhu, 2006).
Some parameters in process models are typically tted to data in real-
time optimization. Ogunnaike (1995) has suggested that modeling for con-
trol should be based on criteria related to the actual end use, and that tting
equations to data may be inadequate in the context of controller design.
6.1. Introduction 81
A constraint on the maximum setpoint change is often applied in real-time
optimization, and smaller changes in setpoints implemented iteratively with
re-identication and re-optimization between each step, a two-step approach.
Uncertainty in real-time optimization falls into four main categories
(Zhang, Monder and Fraser Forbes, 2002): market uncertainty, the imprecise
knowledge of process economics, process uncertainty, the imprecise knowl-
edge of operation due to process disturbances or uncertain inputs, measure-
ment uncertainty, the imprecise knowledge of measured process variables
due to sensor or transmission errors, and model uncertainty, plant/model
structural and parametric mismatch.
Approaches to handling uncertainty in optimization can be divided into
stochastic optimization, sensitivity analysis, back-o and robust optimiza-
tion methods. Stochastic optimization attempts to directly solve a problem
given uncertainty by formulating the stochastic optimization problem in de-
terministic form using probabilities (Kall and Wallace, 1994). The sensitivity
analysis approach consists of augmenting the objective function of the opti-
mization problem with a penalty term intended to minimize parametric sen-
sitivity (Becker, Hall and Rustem, 1994). The back-o method adds a vector
constraint on decision variables to ensure that operating points are feasible,
and the back-o vector is computed at intervals (Loeblein, Perkins, Srini-
vasan and Bonvin, 1999). Robust optimization optimizes the expected value
for a chosen performance index for a given level of risk, formulated in terms of
worst-case, mean-variance and so forth (Darlington, Pantelides, Rustem and
Tanyi, 1999) (Mulvey and Vanderbei, 1995). All these approaches introduce
conservativeness to ensure feasibility rather than exploiting measurements
postoptimization.
Optimization methods which attempt exploit measurements fall into the
two main categories: modied two-step approaches and approaches which do
not update process models. Roberts and Williams (1981) suggests a modied
two-step approach which add a gradient modication term to the cost func-
tion of the optimization problem to ensure that iterates converge to a point
at which the necessary conditions for optimality are satised. Methods which
do not require model updating can be classied into model-free and model-
xed methods. Some model-free methods mimic various iterative numerical
optimization algorithms, such the Nelder-Mead simplex algorithm (Box and
Draper, 1969). Other model-free methods recast the nonlinear program-
ming problem into a problem of choosing decision variables whose optimal
values are approximately invariant to uncertainty. Self-optimizing control
considers determining which variables to keep at constant setpoints to keep
the process acceptably close to optimum (Skogestad, 2000). Extremum-
82
seeking control are adaptive control methods related to self-optimizing con-
trol which attempt to move process setpoints toward values which result in
an extreme value (maximum or minimum) of a measured output without a
process model. Extremum-seeking control usually introduces ample excita-
tion in setpoints to achieve this goal, for instance in form of a low-frequency
sinusoidal reference (Krstic, 2000). NCO-tracking is a model-free method
which uses o-line analysis of the process model to determine functions for
setpoint values at which the necessary conditions for optimum are enforced,
parameterized in terms of measured variables (Franois, Srinivasan and Bon-
vin, 2005). Fixed-model methods utilize both the available measurements
and a process model for guiding the iterative scheme toward an optimal op-
erating point, but update constraint and cost function rather than model
at each iteration (Forbes and Marlin, 1994). Methods such as extremum-
seeking or those which mimic iterative numerical optimization algorithms
introduce many variations to setpoints. These methods may be unsuitable
to the oshore production of oil and gas where the costs and risk of each
change in setpoint may be signicant. So-called model-free methods still
require models in the design of controllers, and obtaining accurate models
of the oshore production of oil and gas oine may be challenging.
There is some precedence for analyzing the risk of implementing a set-
point change. Miletic and Marlin (1998) have proposed result analysis using
multi-variable statistical hypothesis testing to determine whether the pre-
dicted increase in prot from implementing changes in setpoints is statisti-
cally signicant or a result of process noise.
A classic theoretic approach to uncertainty reduction is dual control the-
ory, which deals with the controller design for processes which are initially
unknown (Feldbaum, 1961a) (Feldbaum, 1961b). The theory is called dual
as the objectives of such a controller are twofold, rstly to control the system
as well as possible based on current system knowledge, secondly to experi-
ment with the system so as to better learn how to control it in the future.
The problem of determining uncertainty reduction is also related to the eld
of reinforcement learning, an area of machine learning concerned with how
an agent ought to take actions in an environment so as to maximize some
notion of long-term reward (Sutton and Barto, 1998), focused on on-line
implementation while making a tradeo between exploration of uncharted
territory and exploitation of current knowledge. Reinforcement learning has
seen some application in modeling of batch and semi-batch chemical reac-
tors (Martinez, 2000). Optimal experiment design for control has focused
on deriving an input signal that minimizes some control-oriented measure
of plant/model mismatch under a constraint on total input power. Optimal
6.2. Production optimization under uncertainty 83
experiment design is usually performed with the aim of achieving control
that is robust to disturbances (Gevers and Ljung, 1986).
The aim of this chapter is to suggest and study a structured approach to
optimizing the oshore production of oil and gas with uncertain production
models. The suggested approach is outlined in Section 6.2, a simulation
case-study is described in Section 6.3 before conclusions are drawn.
6.2 Production optimization under uncertainty
6.2.1 A structured approach to optimization with uncertain
models
As production data are often local in nature, i.e. setpoints are only varied
within a narrow range of values, a model tted to data may only be locally
valid, only accurately able to describe production for a narrow range of
setpoints. Motivated by this observation, an iterative two-step approach
to production optimization is considered. An iterative two-step approach
ensures that the model is only used to make predictions for setpoint values
for which the model is the most accurate, which could be achieved in practice
by enforcing
max{u
min
, u
0
U
prc
U} u min{u
max
, u
0
+U
prc
U} (6.1)
in optimization, where U denes the scale elements of u, u
0
is the initial
value of the decision variable, U
prc
< 1 is a design parameter which limits
the magnitude of setpoint change at each step, and u
min
, u
max
are minimum
and maximum values of u, respectively.
In practice setpoint changes suggested by optimization are not imple-
mented instantaneously in oshore oil and gas production, instead the set-
point is moved toward a target in a series of smaller steps to limit transients
and ensure feasibility. Operators need to approach the target setpoint while
observing actual responses, and if constraints are met the approach may
need to be halted before the target is reached. Conversely, constraints ex-
pected to be active when the target is reached may not be so in practice, in
which case operators may move setpoints beyond the target until a constraint
is reached. In this manner, measurements are exploited postoptimization
to ensure feasibility in a manner which adds little conservativeness due to
model uncertainty. This gradual implementation of setpoint change will be
84
referred to as the operational strategy, and the operational strategy will be
considered a component in the structured approach to optimization under
uncertainty.
The change in prots that results from implementing suggested setpoint
change may be uncertain, it may even be negative. If an estimate of uncer-
tainty is available, some form of result analysis could gauge the uncertain
change in prots that will result from an operational strategy altering set-
points toward a target. The decision to implement a setpoint change could
then be based on this result analysis.
Optimizing production under uncertainty requires weighing implement-
ing excitation intended to reduce uncertainty against maximizing production
based on current knowledge. One possible structured approach to weigh
these two considerations is to exploit an estimate of uncertainty to predict
the costs and benets of implementing excitation, a form of excitation plan-
ning.
Some components of the decision variable u may have changed so little
in the tuning set that little or no information on the response to changes
in those decision variables is available. It may also be laborious for a hu-
man operator to simultaneously change all decision variables when dim(u)
is large. For these reasons an optional component of the proposed approach
is to determine a set of active decision variables, variables which are to be
varied by optimization while other decision variables are kept constant. To
limit the scope of this chapter, choosing active decision variables will not be
considered further.
The suggested strategy, motivated by the above discussion, is outlined in
Figure 6.1. Operational strategy design is discussed in Section 6.2.2, result
analysis in Section 6.2.3, and excitation planning in Section 6.2.4.
6.2.2 Operational strategy
Example 6.1 illustrates the intended behavior of the operational strategy
Example 6.1 Consider a hypothetical oshore oil and gas eld, produc-
ing from two wells with u =
_
u
1
u
2
T
, one decision variable aecting
the production of each well. Each well produces oil and water at rates
x =
_
q
1
o
q
1
w
q
2
o
q
2
w
T
which depend nonlinearly on u, and the objective
is to maximize total oil production q
tot
o
= q
1
o
+q
2
o
while the total rate of pro-
duced water q
tot
w
= q
1
w
+ q
2
w
must not exceed a constraint. Suppose that a
target setpoint has been determined by solving the production optimization
problem. Suppose that the current setpoint, feasible and infeasible regions
Perform excitation
planning
Perform production
optimization
Optionally: select
active decision
variables
Implement setpoint
change suggested
by production
optimization
according to
operational strategy
Is the cost/benefit
tradeoff of any
planned excitation
favorable?
Implement
planned
excitation
Yes
Update model:
Estimate parameters
and parameter
uncertainty
Is result analysis
favorable?
No
Yes
Wait until new data
becomes avialable
No
Perform result
analysis
Figure 6.1: Flowchart of the proposed structured approach to optimization
of oshore oil and gas production with uncertain models.
86
are as illustrated in Figure 6.2. The boundary between the feasible and in-
feasible regions is the set of setpoints at which all water processing capacity is
utilized, and parameter uncertainty may cause the position of this boundary
and the optimal setpoint to be uncertain. To ensure that the implemented
setpoint remains in the feasible region, an operational strategy could alter-
nately decrease u
1
and increase u
2
in smaller steps as illustrated in Figure
6.2.
An attempt is made to design an operational strategy which generalizes
the concepts illustrated in Example 6.1. For the purpose of analysis the
operational strategy is stated formally, while a human operator may follow
the suggested algorithm more informally.
The operational strategy attempts to implement changes in u toward a
target u
T
by sequentially increasing and decreasing elements of u and mon-
itoring responses in prots and constraints. u
T
may be any desired value of
u, for instance that found by solving the production optimization problem
numerically. The operational strategy is intended as a postoptimization
strategy for the implementation of the new desired setpoint, and no produc-
tion optimization or parameter estimation is performed during the course
of the operational strategy. It is assumed that the sign of the response in
prots and constraint utilization to a change is setpoints is known. This will
usually be the case in oshore oil and gas production, where decision vari-
ables may for instance be choke openings or gas-lift rates and the response to
an increase in u is usually positive for the range of setpoints considered. Let
U
be the set of indices corresponding the components of u

T
u
0
which are
negative, and let U
+
be the indices of the components of corresponding the
components of u
T
u
0
which are positive. u
T
does not have to be feasible
in practice, the operational strategy will ensure feasibility of implemented
setpoints.
The operational strategy should have a dened criterion for when to ter-
minate. If the termination criterion considers only the sign of prot change,
the operational strategy may not be able to proceed if the initial setpoint
is close to a local optimum. However, if the termination criterion consid-
ers the dierence between monitored and predicted prot, the operational
strategy can proceed in the face of prot decrease, as long as this decrease
is in reasonable agreement with the model used in optimization. If prots
decrease below the predicted values then the decision to terminate should
weigh the magnitude of this decrease against the predicted prot increase at
the target. The rst scenario illustrated in Figure 6.3. shows how a prot
decrease is predicted by the model and it is reasonable for the operational
strategy to continue, while in the second scenario the prot decrease is not
u
2
u
1
Infeasible
Feasible
t
s
o
t
e
o
Figure 6.2: Example 6.1: Illustration of an operational strategy initated at
t
o
s
and terminating at t
o
e
. Solid line illustrates the unknown border between
feasible and infeasible regions, the dotted lines illustrate the uncertainty of
this border, while the dashed line illustrates how an operational strategy
might move setpoints in closed-loop. The circle illustrates the setpoint u
0
,
and crosses illustrate uncertain calculated optimal setpoints.
88
predicted by the model and it is reasonable for the operational strategy to
terminate and return to the initial setpoint. These considerations can be
formulated in terms of estimated and measured prot change and evaluated
by comparing a cost function J
M
against a threshold value J
T
M
.
An operating strategy which is initiated at time t
o
s
and terminated at
time t
o
e
, increases the realized potential by:
P
o,r
= P
o,r
(t
o
s
) P
o,r
(t
o
e
), (6.2)
and in the section below it is suggested how a probability density function
f
Pr
(

P
o,r
) can be found prior to implementing the operational strategy.
The operational strategy can continue altering setpoints beyond u
T
as long
as each step of the operational strategy causes increased prot. As the
operational strategy may overshoot the most protable setpoint, the nal
step should be to assess all steps and return to the most protable one. The
suggested operational strategy is outlined in Algorithm 6.1.
Algorithm 6.1 (An operational strategy) Given a process that is ini-
tially at a feasible setpoint (u
0
, d
0
) and given measured process constraints c,
measured prot

M and decision variables u, a target value u
T
and probability
distributions f
Pr
(

P
o,r
), let 0 < S 1 denote step size.
1. Let the measured prot prior to implementing the operational strategy
be

M
0
. Let k be the index of the current step, and set k = 0 initially.
2. Determine sets (U
, U
+
) to correspond with u
T
u
0
.
3. Repeat (3a),(3b),(3c)
(a) k = k+1
(b) decrease u
i
, i U
by an amount proportional to u
i
T
u
i
0
, i U
,
so that u
i
k
= u
i
k1
S (u
i
T
u
i
0
), i U
,
(c) increase u
i
, i U
+
by an amount proportional to u
i
T
u
i
0
, i U
+
while observing c and

M, as long as all elements of c obey c < 0
and

M
u
> 0, let the resulting measured prot be

M
k
, and the
resulting setpoint u
k
,
while (u
k
u
0
> u
T
u
0
and

M
k

M
k1
) or ( u
k
u
0
<
u
T
u
0
and J
M
(

M
k
, f
P
(

P
o,r
)) > J
T
M
).
4. Implement the setpoint which resulted in the highest measured prot in
steps 1-3.
M(t)
t
t
M(t)
M
0
M
0
M +
0 o,r
P
M +
0 o,r
P
^
^
Scenario 1
Scenario 2
t
a
t
a
t
s
o
t
e
o
t
e
o
t
s
o
increasing process model uncertainty
negative profit bias
Figure 6.3: The two plots illustrate how measured prot (solid lines) may
evolve compared to a calculated prediction interval (dotted lines) in two dif-
ferent cases. M
0
is the prot as the operational strategy is implemented,
and M
0
+

P
o,r
(dashed line) is the predicted increase in prot from imple-
menting the operational strategy.
90
One of the operators tasks is to ensure that the operational constraints
are obeyed, therefore it is reasonable to assume that (u
0
, d
0
) is feasible in
Algorithm 6.1. It is assumed that in practice cases of infeasible (u
0
, d
0
) are
handled operationally, i.e. an operator will apply process knowledge instead
of mathematical optimization to move the process to a feasible setpoint.
The magnitude of S is a user preference and preferably S << 1. The
magnitude of S will also depend on how often production optimization is
implemented. Smaller S may result in an operational strategy which is more
labor-intensive to implement, but may in some cases cause the operational
strategy to terminate closer to optimum.
6.2.3 Result analysis
This section aims to investigate whether an estimate of the parameter uncer-
tainty can be exploited in result analysis. The approach chosen is simulating
the prot change P
o,r
that results from implementing a setpoint change
suggested by production optimization on the production model with dier-
ent parameter estimates drawn from the estimated parameter uncertainty
f
) in a Monte-Carlo manner. The approach is outlined in Algorithm 6.2

Pr
(

P
o,r
)) Given an operational strategy,
the current operating point (x
0
, u
0
, d
0
), U
prc
and probability distributions
f
) and f
u
( u),
repeat N
t
times
draw a sample

t
from f
),
obtain a point estimate of P
o,r
by simulating the operational
strategy running on a process described by the process model with
negligible structural model uncertainty and with parameters equal
to

t
, while enforcing the constraint (6.1).
The distribution of N
t
point estimates

P
o,r
is an estimate of
f
Pr
(

P
o,r
).
As a copy of the production model cannot account for structural uncertainty
which will grow with the magnitude of uu
0
, a constraint of the form (6.1)
should be enforced in simulations. The magnitude of estimates of

P
o,r
will
depend on the choice of U
prc
, but a conservative choice of U
prc
should yield
conservative estimates

P
o,r
on which result analysis could be based.
The aim of the result analysis is to determine whether to implement the
setpoint change suggested by production optimization. This decision could
be based on the evaluation of a cost function which depends on f
Pr
(

P
o,r
):
J
r
(f
Pr
(

P
o,r
)) J
T
r
. (6.3)
J
r
(f
Pr
(

P
o,r
)) could for instance be the expected or worst-case value. The
threshold J
T
r
should ensure that setpoint changes are only performed when
the resulting increase in prots is expected to be signicant and to justify the
risks and operational costs which are associated with any setpoint change,
such as the risk of triggering an unplanned shutdown.
6.2.4 A cost/benet approach to excitation planning
The value of implementing a planned excitation prior to implementing pro-
duction optimization is that the excitation may reduce model uncertainty,
allowing production optimization and the operational strategy to nd a more
protable setpoint than otherwise possible. This chapter investigates ex-
ploiting estimates of parameter uncertainty f
) found from recent histori-

cal production data to form the basis of a structured approach to rank the
benet of exciting dierent decision variables. For simplicity this discussion
is restricted to wells which are decoupled, i.e. where implementing change in
the production of one well will not inuence the production of other wells,
which is often a reasonable assumption for so-called platform wells. The
discussion is restricted to excitation of single decision variables, although
simultaneously exciting several variables is also conceivable.
The benet of excitation B
E
is proposed dened as the dierence be-
tween the prot which would be realized if no excitation is performed, P
o,r
,
and the prot which would be realized after implementing an excitation,
P
E
o,r
:
B
E

= P
E
o,r
P
o,r
. (6.4)
The idea of the benet of excitation is illustrated in Example 6.2.
Example 6.2 Consider a hypothetical eld producing oil and water from
two decoupled wells. Well 1 produces (q
1
o
, q
1
w
) at rates which depend on u
1
,
well 2 produces (q
2
o
, q
2
w
) at rates which depend on u
2
. The relationships be-
tween (q
1
o
, q
1
w
) and u
1
and between (q
2
o
, q
2
w
) and u
2
are uncertain, but a model
which expresses q
1
o
(u
1
,
1
), q
2
o
(u
2
,
2
), q
1
w
(u
1
,
1
), q
2
w
(u
2
,
2
) is available. Mod-
els for well 1 depend on parameters
1
and models for well 2 on
2
. Two
parameter estimates

a
=
_
1
a

2
a
and

b
=
_
1
b
2
b
have been determined

92
which result in estimates of q
tot
o
= q
1
o
+ q
2
o
and q
tot
w
= q
1
w
+ q
2
o
which match
measurements equally well for a set of recent historical data. Optimization
of the objective M = q
1
o
(u
1
,
1
) + q
2
o
(u
2
,
2
) subject to the model and the
constraint q
1
w
(u
1
,
1
) + q
2
w
(u
2
,
2
) < q
max
w
is desired. If no excitation is im-
plemented the target u
T
= u(
a
) is planned implemented using an operational
strategy. If an excitation could be performed which would reveal which of

a
and

b
better describe production, what would be the benet of that excita-
tion?
If

a
describes production better B
E
= 0 as P
E
o,r
= P
o,r
. If

b
describes
production better B
E
will depend on how much prot increase the opera-
tional strategy would be able to implement if given the target u
T
= u(
b
)
as opposed to u
T
= u(
a
). In the case that

b
describes production better
while u
T
= u(
a
), B
E
could be found through simulations on the production
models, provided that structural model errors are small.
Example 6.2 illustrates that B
E
will depend on the interaction between un-
certainty in

, uncertainty in u(
) and the ability of the operational strategy

to compensate for uncertainty in u(
). The benet of excitation will also

depend on what target the operational strategy will pursue if no excitation
is performed. In the oshore production of oil and gas, exciting all compo-
nents of u is usually not feasible, so some method for ranking components
of u to be excited is desirable, as is illustrated in Example 6.3.
Example 6.3 Consider again the eld described in Example 6.2. Consider
that our aim is to choose whether to introduce excitation in u
1
, in u
2
or not at
all. Assume that an excitation of u
1
would allow us to distinguish between

1
a
and

1
b
, while

2
would still be uncertain, and that an excitation of u
2
would
allow us to distinguish between

2
a
and

2
b
while

1
would still be uncertain.
With the assumptions made it can be expected that an excitation of well 1
will either result in implementing u
T
= u([
1
a
,

2
a
]) if production is described
by

a
, or implementing u
T
= u([
1
b
,

2
a
]) if production is described by

b
, as
it was planned to implement u
T
= u([
1
a
,

2
a
]) if no excitation was performed,
and

2
would still be uncertain after excitation of u
1
. The outcome on the
production models can be simulated in both cases provided that structural
model errors are small, which would give two estimates of

B
E
. A structured
approach to planning excitation could then be to excite the well associated
with the highest positive average

B
E
.
Examples 6.2 and 6.3 illustrate the principles of a Monte-Carlo simulation
approach to estimating B
E
for the simplest possible case of two wells, two
parameter estimates and one constraint. In our case study the same prin-
ciples are applied to estimate

B
E
for a larger number of wells, constraints,
and parameter estimates

.
Let

n
be the nominal parameter estimate, the parameter estimate found
by solving (2.6) directly. The estimation of B
E
will consider the case when
the operational strategy attempts to implement the target u
T
= u(
n
) if
no excitation is performed. Note that although production of wells are de-
coupled, the components of u() as calculated by (2.1)(2.3) are coupled,
so updating models describing one well may inuence suggested settings for
other wells, and this will need to be reected when estimating

B
E
. Simu-
lations on the production model assume that structural model uncertainty
is negligible, and, as locally valid models are considered, this assumption
is reasonable as long as change in decision variables is small, which moti-
vates enforcing the constraint (6.1) in simulations, and which should result
in conservative estimates

B
E
.
The approach considered is outlined in Algorithm 6.3.
Algorithm 6.3 (Benet of excitation, u
i
) Given that it is planned to
implement u(
n
) if no excitation is performed. Given N
t
, the distribution
f
), an operational strategy, a process model which depends on parameters

, U
prc
, and that the aim is to estimate the benet of exciting u
i
, a single
decision variable in the vector u.
repeat N
t
times
draw a sample

t
from f
).
determine

P
o,r
(
t
) by simulating the operational strategy using
u
T
= u(
n
) on the process model with parameters

t
while setpoint
change is limited by (6.1).
let

i
t
be parameters of

t
excited by u
i
. Let
E
(
n
,

i
t
) be a vector
of parameters that is equal to

i
t
for parameters excited by u
i
and
equal to

n
for all other parameters.
determine

P
E,i
o,r
(
t
) by simulating implementing the operational
strategy using u
T
= u(
E
(
n
,

i
t
)) on the process model with pa-
rameters

t
while setpoint change is limited by (6.1).
let

B
E,i
(
t
) =

P
E,i
o,r
(
t
)

P
o,r
(
t
).
the distribution of simulated values

B
E,i
is an estimate of the proba-
bility density function for the benet of exciting u
i
, f
E
B
(

B
E,i
)
A structured approach to excitation planning could be to excite those deci-
sion variables u
i
which have highest

B
E,i
on average.
94
As components of u() as calculated by (2.1)(2.3) are coupled, excita-
tion of u
i
can cause u(
E
(
n
,

i
t
)) to dier from u(
n
) in components other
than u
i
. If uncertainty for non-excited wells is signicant, implementing
u(
E
(
n
,
i
t
)) may inadvertently cause lower prots than u(
n
), and in such
cases estimates

B
E,i
may be negative. Negative

B
E,i
is an indication that
the prot increase resulting from excitation of u
i
is variant to uncertainty
in models describing other wells.
It is reasonable to assume that production is initially utilizing at least
one processing capacity fully, so that any excitation will require production
to back o from full capacity for a period of time, incurring a cost. It is
conceivable to simulate the cost of excitation C
E
in a Monte-Carlo fashion
for dierent samples of f
). To limit the scope of this chapter, estimation

of C
E
will not be considered. It is proposed that candidate excitations can
be ranked by the value of a function
J
E
(f
E
B
(

B
E
), f
E
C
(

C
E
) J
E,T
, (6.5)
where J
E,T
is a user-specied threshold and J
E
(f
E
B
(

B
E
), f
E
C
(

C
E
) is a user-
specied metric such as for instance the expected value of the dierence
B
E

C
E
.
6.3 Case study
This section describes the application of the suggested approach to a case-
study modeled on the North Sea oil and gas eld and the set of real-world
production data considered in Chapters 3, 4 and 5.
6.3.1 Field description
The case study considered in this chapter is motivated by a North Sea oil
and gas eld with 20 gas-lifted platform wells producing predominantly oil,
gas and water. The eld has a layout as depicted in Figure 2.1, with one
production separation train and one test separator. Measurements of the
total rates of produced oil, gas and water are available. The operator of the
eld requested that all data be kept anonymous, therefore all variables will
be presented in normalized form. The production data are characterized by
little variation in decision variables. The aim of production optimization
on the eld is to distribute available lift gas so as to maximize total oil
production while keeping total produced water and gas rates below capacity
constraints.
6.3. Case study 95
6.3.2 Method
The cost and risks of implementing a trial of the suggested approaches on an
actual eld are signicant, and it may be dicult to compare strategies by
implementation on an actual eld as production will vary with time due to
disturbances. This motivates the choice of studying the suggested approach
in simulations.
Due to uncertainty there may be many plausible descriptions of pro-
duction, and an approach to optimization under uncertainty should ideally
perform well for all such plausible descriptions. This chapter will consider
simulating optimization on a model that is plausible in the sense that it con-
forms with production data and refer to this model as the production analog.
When optimizing production, the production analog is considered unknown,
production optimization can only infer knowledge of the production analog
through measurements.
To ensure that the production analog is a plausible description of the
eld considered, parameters in (6.7) are estimated through the use of boot-
strapping. (6.7) combined with one of the plausible parameters determined
in this manner is chosen as the the production analog for setpoint values
similar to those observed in the tuning set.
Let q
max,i
gl
be the maximum values observed in the tuning set. To sim-
ulate our lack of knowledge about process behavior for values of u outside
those observed in the tuning set, the kernel function f
gl
is replaced in (6.7)
with
f
i
u
(q
i
gl
) =
1
2
c
i
(q
i
gl
q
i,max
gl
)
2
+b
i
(q
i
gl
q
i,max
gl
) +a
i
(6.6)
when q
i
gl
> q
max,i
gl
. a
i
and b
i
are chosen so that q
i
(u) is smooth and continu-
ous for u around u
max
, and the curvature coecient c
i
is chosen as a random
negative value. Operating points q
l,i,a
of the production analog are chosen
so that the tuning set can be described by the production analog with bias
terms.
Production modeling of the eld considered based on the concepts of
system identication was considered in Chapter 4, and similar methods are
applied here. The oil, gas and water rates q
i
o
, q
i
g
, q
i
w
of each well i are modeled
as the product of two kernel functions, one describing the eects of changes
in production valve opening z
i
, and one describing the eects of changes in
gas lift rates q
i
gl
. Models are local around the most recent well test. When
modeling the response to change in gas-lift rate, a rst-order linear kernel
with a single tted parameter for each uid and phase with dim() = 63
(formulation A) will be compared with a second-order nonlinear kernel with
96
two tted parameters for each uid and phase and dim() = 123 (formula-
tion B). By comparing these formulations the signicance of model structure
and parameterization on the suggested methodology can be assessed. Con-
straints and regularization terms on parameter values are added to include
physical knowledge.
Production optimization considered choosing gas-lift rates for all wells
with the objective of maximizing total produced rate of oil without increasing
utilization of water and gas processing capacity, a conservative constraint. In
addition an upper and lower constraint on u of the form (6.1) was enforced.
Some wells are in danger of slugging if gas-lift is decreased, and on these wells
a constraint which prohibits gas-lift from being decreased was implemented.
The production model
q
i
p
= max{0, q
l,i
p
f
i
z
(z
i
, z
l,i
)f
i
gl
(q
i
gl
, q
l,i
gl
)}, (6.7)
is considered, which is intended to be valid locally around the most recent
well test rates q
l,i
p
i = 1, . . . , n
w
p {o, g, w}, similarly to Chapter 4.
z
i
[0, 1] is the relative valve opening of well i, q
i
gl
is the gas-lift rate of
well i and q
i
gl
=
q
i
gl
q
l,i
gl
1 is the normalized, relative gas lift rate for well i.
Kernels are chosen as
f
i
z
(z
i
, z
l,i
) =
1 (1 z
i
)
k
1 (1 z
l,i
)
k
, k = 5 (6.8)
f
i
gl
(z
i
, z
l,i
) =
_
1 +
i
p
q
i
gl
+
i
p
(q
i
gl
)
2
_
(6.9)
for all wells i.
The measurement vector y(t) =
_
q
tot
o
(t) q
tot
g
(t) q
tot
w
(t)
T
is considered
and an attempt is made to nd so that estimates
y(, t)

= 1 x(, t) +

y
, (6.10)
t measurements as close as possible for the tuning set, where

y
is the vector
of measurement biases due to calibration inaccuracies to be determined and
1 is a matrix of ones.
Based on the knowledge that the watercut is rate-independent, soft con-
straints which penalize deviation from
o
=
w
and
o
=
w
were added to
the objective function. A decline in q
tot
o
was visible in the tuning set, and
it is chosen to de-trend q
tot
o
and weigh older measurements less than newer
ones using the weighting-term w(t) in (2.6). In formulation B, and for
all wells and for oil, gas and water, as well as
y
are considerd part of the
6.3. Case study 97
vector of tted parameters. Only the rst-order term in (6.9) is considered
in formulation A.
The production optimization problem considered was of the form
u() = arg max
u
iI
a
q
i
o
(u, d, ) +b
o
() (6.11)
s.t. (6.12)
q
l,i
gl
q
i
gl
i I
MGS
(6.13)
max{u
min
, u
0
U
prc
U} u min{u
max
, u
0
+U
prc
U} (6.14)
iI
a
q
i
g
(u, d, )

iI
B
q
i
gl
+b
g
() q
tot,c
g
(6.15)
iI
a
q
i
w
(u, d, ) +b
w
() q
tot,c
w
. (6.16)
I
a
is the indices of all wells considered in production optimization. Biases
b
o
(), b
g
(), b
w
() were determined for each so that modeled and mea-
sured production match at the time of optimization. For wells with indices
I
MGS
= {1, 4, 10, 13, 14, 16} gas-lift could not be decreased without the risk
of triggering slugging. q
i
o
, q
i
g
, q
i
w
are modeled rates.
The case study was implemented in MATLAB
1
. Linear least-squares
parameter estimation problems and linear real-time optimization problems
were solved using the TOMLAB
2
solver lssol. Nonlinear production opti-
mization problems were solved using the sequential quadratic programming
solvers based on Schittkowski (1983).
The implementation of the suggested strategy was simulated for four
iterations, both with formulation A and formulation B. Algorithms 6.2 and
6.3 were run for each iteration with N
t
= 200, but no excitations were
implemented and setpoint change was implemented regardless of risk-reward
estimates. Unless otherwise stated U
prc
= 0.2, a second simulation with
U
prc
= 0.5 was also performed for comparison.
The simulated benet of excitation of u
i
was found by replacing tted
parameters excited by variation of u
i
with those that best describe the pro-
duction analog, and simulating production optimization and implementation
with the operational strategy.
1
2
98
6.3.3 Results
A comparison of the predictions of the production analog against eld data
is shown in Figure 6.4. The production analog is shown along with models of
formulation A and B at iteration 1 in Figures 6.5 and 6.6. The distribution
of u found through Monte-Carlo simulation prior to iteration 1 are compared
for the two models in Figure 6.7. Figure 6.8 illustrates the changes in set-
point implemented by the operational strategy for formulation B. Figure 6.9
compares the realized prots P
o,r
of formulations A and B at each iteration,
for U
prc
= 0.2 and U
prc
= 0.5, indicating that between 30% and 80% of the
prot potential was realized. Figure 6.10 compares f
Pr
(

P
o,r
) found with
Algorithm 6.2 with realized prot for each iteration of formulation A and
B. Figures 6.11 and 6.12 compare the simulated benet of excitation with
estimates found with Algorithm 6.3 for iterations 1 and 4. Figure 6.13 com-
pares target setpoints before and after excitation for a particular strategy,
well and iteration.
6.3.4 Discussion
Figure 6.4 illustrates that the production analog is a plausible description of
production, as it matches production data well. Figures 6.5 and 6.6 illustrate
that there is signicant parameter uncertainty for both model formulations
as a result of low information content in the tuning data set, while Figure
6.7 illustrates that uncertainty in u() is signicant due to parameter uncer-
tainty. Figure 6.8 illustrates how the operational strategy is able to increase
prots while ensuring that capacity constraints are not exceeded.
Figure 6.9 illustrates that despite production model uncertainty, a sig-
nicant prot increase is realized with either model formulation and with
either choice of U
prc
. None of the simulations converge to the optimum,
which should be expected as no guarantees of convergence can be given in
general. A larger prot increase was realized in simulations with formulation
B, and this is because model formulation B better describes the production
analog. In simulations with model formulation B, U
prc
= 0.2 actually re-
sulted in higher realized prots than U
prc
= 0.5. This is due to parameter
estimation better exploiting information introduced by setpoint changes in
previous iterations when U
prc
is small.
Figure 6.10 illustrates that estimates

P
o,r
found with Algorithm 6.2
matched simulated P
o,r
signicantly better for formulation B than model
formulation A. This observation is caused by the structural inability of model
formulation A to accurately model the production analog for a range of
decision variables, and this illustrates that Algorithm 6.2 requires models
6.3. Case study 99
q
tot
o
q
tot
g
q
tot
w
t[hours]
0 500 1000 1500 2000 2500
0 500 1000 1500 2000 2500
0 500 1000 1500 2000 2500
0.8
1
0.8
1
1.2
1.4
0.6
0.8
1
1.2
Figure 6.4: Case study: Production data (dotted), the production analog
(dashed), estimates with nominal nonlinear production model (solid).
100
q
i
o
i = 1
q
i
g
q
i
w
i = 2
i = 3
i = 4
i = 5
i = 6
i = 7
i = 8
i = 9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
q
i
gl
i = 20
q
i
gl
q
i
gl
Figure 6.5: Case-study: Iteration 1, formulation A. Each row shows modeled
rates of oil, gas and water for dierent parameter estimates found through
bootstrapping (dashed) compared with modeled rates for the nominal pa-
rameter estimate (dotted) and the production analog (solid), global optimum
(crosses) and currently implemented setpoint (circles).
6.3. Case study 101
q
i
o
i = 1
q
i
g
q
i
w
i = 2
i = 3
i = 4
i = 5
i = 6
i = 7
i = 8
i = 9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
q
i
gl
i = 20
q
i
gl
q
i
gl
Figure 6.6: Case-study: Iteration 1, formulation B. Each row shows modeled
rates of oil, gas and water for dierent parameter estimates found through
bootstrapping(dashed) compared with modeled rates for the nominal param-
eter estimate (dotted) and the production analog (solid), global optimum
(crosses) and currently implemented setpoint (circles).
102
i = 1
Formulation A
i = 2
i = 3
i = 4
i = 5
i = 6
i = 7
i = 8
i = 9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
q
i
gl
i = 20
Formulation B
q
i
gl
Figure 6.7: Case-study: Distribution f
u
( u) for formulations A and B shown
for each well after iteration 1 (bars). Constraints on u (dotted), u
0
(circle)
and u
T
(stem), and optimal setpoint (crosses).
6.3. Case study 103
q
tot
o
q
tot
g
q
tot
w
q
gl
q
l
gl
t [iterations]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
0 1 2 3 4
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.985
0.99
0.995
0.98
0.985
0.99
0.995
1
1.01
1.02
1.03
Figure 6.8: Case study: An example of how the operational strategy im-
plements setpoint changes while obeying process constraints for formulation
B. Top graphs show normalized prots (q
tot
o
), gas capacity utilization (q
tot
o
)
and water capacity utilization (q
tot
w
), lower graphs show normalized relative
changes in gas lift rates q
gl
for all wells.
104
P
o,r
P
o,r
100%
t[iterations]
0
20
40
60
80
100
Figure 6.9: Case-study: The percentage of the initial potential P
o,r
that
was realized in simulations of formulation A (cirles) and B (squares) for
U
prc
= 0.2 (line) and U
prc
= 0.5 (dashed).
6.3. Case study 105
it:1
Formulation A Formulation B
it:2
it:3
it:4
P
o,r
(%) P
o,r
(%)
0 1 2 0 2
0 1 2 0 2
0 1 2 0 2
0 1 2 0 2
Figure 6.10: Case-study: Estimated f(P
o,r
) for formulations A and B,
one row for each iteration. The increase in prots that was implemented in
practice is shown as stems.
106
i = 1
Formulation A
0.69
i = 2
1.18
i = 3
0.10
i = 4
-0.01
i = 5
-0.16
i = 6
0.38
i = 7
0.97
i = 8
0.03
i = 9
1.17
i = 10
0.15
i = 11
0.28
i = 12
2.78
i = 13
0.03
i = 14
2.14
i = 15
0.69
i = 16
0.00
i = 17
0.24
i = 18
1.23
i = 19
-0.11
i = 20
B
E
(%)
0.86
Formulation B
0.29
1.16
0.49
-0.01
0.08
0.56
1.13
0.42
0.24
0.27
1.64
1.44
-0.10
0.97
-0.01
-0.00
-0.24
0.70
0.17
B
E
(%)
-0.36
-10 0 10 -10 0 10 20
Figure 6.11: Case-study: Distribution of

B
E
of exciting well i in isolation
at iteration 1 of the simulation, found with Algorithm 6.3 (bars), B
E
found
by simulating excitation on the production analog (stems). Numbers show
average

B
E
as predicted by Algorithm 6.3. All values are expressed in
percent of total unrealized potential.
6.3. Case study 107
i = 1
Formulation A
0.88
i = 2
-0.22
i = 3
0.25
i = 4
0.03
i = 5
-0.03
i = 6
0.85
i = 7
0.83
i = 8
0.62
i = 9
1.41
i = 10
0.28
i = 11
0.81
i = 12
5.16
i = 13
-0.12
i = 14
2.50
i = 15
0.59
i = 16
0.00
i = 17
0.02
i = 18
0.79
i = 19
-0.07
i = 20
B
E
(%)
1.19
Formulation B
1.80
6.40
0.38
0.18
-0.14
3.46
9.80
11.98
11.64
2.37
15.25
6.51
1.66
2.71
-0.01
0.12
0.82
18.32
2.88
B
E
(%)
1.35
-50 0 50 100 -20 0 20
Figure 6.12: Case-study: Distribution of

B
E
of exciting well i in isolation
at iteration 4 of the simulation, found with Algorithm 6.3 (bars), B
E
found
by simulating excitation on the production analog (stems). Numbers show
average

B
E
as predicted by Algorithm 6.3. All values are expressed in
percent of total unrealized potential.
108
i = 1
i = 2
i = 3
i = 4
i = 5
i = 6
i = 7
i = 8
i = 9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
q
i
gl
i = 20
Figure 6.13: Case-study: Formulation B, iteration 4, candidate excitation
u
9
. Upper and lower constraints on u (dashed), currently implemented set-
point (circles), the target setpoint that would be implemented if no exci-
tation is performed of u
9
(stems with circles) and target setpoint after an
excitation of u
9
is simulated (stems with squares).
6.3. Case study 109
with little structural uncertainty to achieve good estimates. In iteration
1 estimates

P
o,r
found with Algorithm 6.2 underestimate the simulated
P
o,r
, and this is due to the conservativeness introduced by the constraint
(6.1).
The models used in this simulation case study are local around an op-
erating point which could be the most recent well test or allocated rate.
When the currently implemented operating point has changed signicantly
from the point around the operating point used in modeling, this operating
point may need to be updated through estimation or single-rate well tests.
This explains the mismatch between observed and estimated prot change
in later iterations in Figure 6.10, and the reduction in implemented total oil
production seen in iteration 4 in Figure 6.8.
A correlation between high actual benet of excitation and a high average
estimated benet of excitation can be seen in Figures 6.11 and 6.12, and this
could be exploited in excitation planning.
Figures 6.11 and 6.12 illustrate that the benet of excitation is small in
magnitude during iteration 1 compared with iteration 4. As a large increase
in prots can be obtained without excitation in iteration 1, while no increase
in prots is obtained without excitation in iteration 4, this result is to be
expected.
Note that both estimated and simulated benet of excitation sometimes
have negative values in Figures 6.11 and 6.12. This is due to the depen-
dency between elements of u(). An excitation which reduces parameter
uncertainty for some parameters may cause a new u() which diers from
the previous estimate in ways which depends on the still present parameter
uncertainty. Because overall parameter uncertainty is lower in iteration 4
due to the excitation introduced by setpoint changes in previous iteration,
negative estimated and simulated benet of excitation is less prevalent.
Figure 6.13 illustrate how excitation of a particular well can have im-
plications for the target value of other wells, as the components of u() are
coupled.
Since analyzing the benet of exciting a single well is subject to an ele-
ment of chance, it may be easier to gain acceptance among practitioners for
iteratively implementing changes in multiple decision variables in a manner
which simultaneously increases prots and introduces excitation. Design
and application of such methods is an interesting topic of further work.
110
6.4 Conclusion
In the simulation case study the method suggested realized between 30%
and 80% of the available prot potential, despite that models were tted
to data with low information content similar to that found on real-world
oil elds. Feasibility was ensured despite model uncertainty by exploiting
measurements postoptimization in a manner which introduced little con-
servativeness. The method requires measurements of prots and constraints,
some knowledge of the sign of responses to setpoint change, and in our sim-
ulations transients have not been considered explicitly.
Simulations indicate that estimates of parameter uncertainty derived
from recent historical data can be used both for result analysis and exci-
tation planning as long as structural errors are small.
Chapter 7
Optimization of oshore oil
and gas production under
uncertainty; active decision
variables and case-study
A version of this chapter has been submitted to SPE Production &
Operations
Abstract
The aim of production optimization is to suggest changes in decision vari-
ables which increase prots, based on a model and constraints. In many
practical cases model uncertainty in production optimization will be signif-
icant and cannot be trivially reduced or eliminated, due to the costs and
risk of uncertainty reduction. Previous chapters have considered exploiting
information in production data to build and update models for production
optimization and to explicitly analyze resulting uncertainty.
The contributions of this chapter are twofold. First, the methods developed in
previous chapters are applied to a eld and set of real-world production data
not previously considered. Some extensions of the methods of earlier chap-
ters are suggested and insight into the practical applicability of the method is
gained. It was found that due to uncertainty it may be dicult to determine
a change in all decision variables which will increase prots with a high de-
gree of certainty. This observation motivates the second contribution of this
111
112
Optimization of oshore oil and gas production under
uncertainty; active decision variables
chapter, a study of choosing active decision variables to adapt to estimated
uncertainty.
Simulations indicate that it may be possible to suggest setpoint changes which
increase prots with high condence through the analysis of real-world pro-
duction data despite signicant model uncertainty.
7.1 Introduction
Chapter 4 considered a method for modeling production based on analysis
of production data and on updating the parameters of such models with pa-
rameter estimation. An explicit analysis of resulting parameter uncertainty
using the method of bootstrapping was conducted and the method applied
to a set of real-world eld data. As the ndings of such an analysis may be
case-dependent, applying the method to a eld and set of production data
that has not previously been considered may reveal insight into the general
applicability of the method.
Parameter uncertainty can be reduced by the introduction of additional
variation of decision variables, however the benet of completely eliminating
uncertainty may not justify the cost in all cases. In such cases a strategy for
adapting to uncertainty may be desirable, so that some prot increase may
still be realized.
The gradual change of setpoints while adapting to uncertainty using what
was referred to as an operational strategy is studied in Chapter 6. Decision
variables are not changed from an original setpoint to a target setpoint in-
stantaneously in oshore oil and gas production, rather decision variables are
changed gradually toward a target setpoint to monitor constraints and limit
transients. The target setpoint can be modied during this process to adapt
to information gained from observed responses. The operational strategy
ensures that the setpoint reached is feasible, and can ensure that prots are
increased until at least one measured processing capacity constraint is met.
While an operational strategy adapts to nd a setpoint near the target
which is feasible and as protable as possible, the resulting setpoint found
depends to a large extent on the target setpoint specied, and the target
setpoint may itself be highly uncertain. For this reason, the manner in which
target values themselves are chosen may need to adapt to uncertainty.
7.1.1 Previous Work
Approaches to handling uncertainty in optimization can be divided into
stochastic optimization, sensitivity analysis, back-o and robust optimiza-
7.2. Choosing active decision variables 113
tion methods. It is not obvious how to devise a penalty term in the case of
sensitivity analysis or a performance index in the case of robust optimization
which takes into account the adaptation of the operational strategy and re-
sulting robustness. The back-o approach is mostly concerned with ensuring
feasibility, which also the operational strategy can ensure.
Of these methods, only stochastic optimization can be formulated to
explicitly account for the operational strategy, through a recourse term in
the objective function. Most solvers of stochastic optimization problems
consider linear models and constraints, which would require linearization
methods if these were to be applied to oshore oil and gas production.
7.1.2 Problem Formulation
The aim of this chapter is to apply and extend methods for production
optimization developed in prior chapters to a eld and a set of real-world
production data that has not previously been considered to reveal insight
into the general applicability of the method. The observation that an op-
erational strategy may have diculty in adapting to the decision variable
uncertainty for the case considered motivates the study of active decision
variables as a means of adapting to uncertainty.
Section 7.2 outlines the suggested method and Section 7.3 contains the
case study.
7.2 Choosing active decision variables
If uncertainty is signicant it may not be possible to suggest a setpoint
change that will increase prots with a high degree of condence. On the
other hand, the uncertainty associated with altering dierent decision vari-
ables will often be of dierent magnitude. Therefore, it may be possible
to suggest a setpoint change involving some decision variables which will
increase prots with condence, even if no such setpoint change involving
all decision variables can be found.
Another advantage of altering fewer decision variables is that it may
reduce the workload on human operators, which in oshore oil and gas pro-
duction often implement setpoint changes manually. The attainable prot
increase from altering fewer decision variables may be less than than the
maximum prot increase possible, i.e. there is not guarantee of optimality.
In practice suggesting setpoint changes which with condence increase prof-
its without requiring uncertainty reduction may be of interest regardless of
optimality.
114
Let the production of well i be a function of certain decision variables
u
i
, modeled disturbances d
i
, and the response be characterized by certain
tted parameters
i
. For many elds these variables are distinct for each
well, so that
=
_
2
. . .
n
(7.1)
d =
_
d
1
d
2
. . . d
n
(7.2)
u =
_
u
1
u
2
. . . u
n
. (7.3)
and measured production y can be decomposed into contributions from in-
dividual wells as
y = h(f
1
(u
1
, d
1
,
1
), f
2
(u
2
, d
2
,
2
), . . . , f
n
(u
n
, d
n
,
n
)), (7.4)
where f
1
(u
1
, d
1
,
1
), f
2
(u
2
, d
2
,
2
), . . . , f
n
(u
n
, d
n
,
n
) are possibly nonlinear
functions describing the contributions from individual wells. Through the
knowledge of which parameters
i
are associated with which decision vari-
ables u
i
, it may be possible to select which decision variables to alter when
some parameters
i
are signicantly more uncertain that others.
(2.1)(2.3) is usually a nonlinear programming problem that will require
numerical solution, which may benet from providing analytical Jacobians
and Hessians
M(x, u, d)
u
(7.5)
2
M(x, u, d)
u
2
(7.6)
f(x, u, d, )
u
(7.7)
2
f(u, d, )
u
2
(7.8)
c(u, d, )
u
(7.9)
2
c(u, d, )
u
2
(7.10)
to the numerical solver to obtain a solution with sucient computational
speed and accuracy. The expressions (7.5)(7.10) will be dierent for each
choice of u, and supplying correct Jacobians and Hessians for all conceivable
u will add complexity to an algorithm for evaluating dierent choices of
u. The same Jacobians and Hessians (7.5)(7.10) can be used in all cases
if candidates for active decision variable sets are evaluated by enforcing
7.2. Choosing active decision variables 115
dierent constraints on a single u. This observation motivates including the
constraint
u
i
= u
i
0
i I
C
a
. (7.11)
to (2.1)(2.3). Variables u
i
, i I
a
are termed active decision variables
and I
a
is the set of indices of the active decision variables, all other indices
are the inactive decision variables u
i
, i I
C
a
.
It is feasible in principle to determine I
a
by solving an optimization
problem. As determining an optimal I
a
may be complex, such a problem
may be both stochastic, mixed-integer and nonlinear, this will not be con-
sidered further. For simplicity this chapter will instead consider determining
I
a
which according to simulations can be varied protably with condence
and which obey the following conditions:
Condition 1: the sign of the active decision variable change u
i
()u
i
o
, i
I
a
suggested by production optimization is invariant to uncertainty,
and
Condition 2: the prot increase resulting from altering u
i
, i I
a
is ex-
pected to be positive regardless of parameter uncertainty.
Condition 1 is motivated by the properties of operational strategies, and
Condition 2 is motivated by risk aversion.
This chapter considers a Monte-Carlo approach summarized in Algo-
rithm 7.1 for evaluating a given candidate I
a
, according to the criteria stated
above.
Algorithm 7.1 (Evaluating a candidate I
a
) Given I
a
, production ini-
tially at setpoint (u
0
, d
0
), N
s
, a production optimization problem (2.1)(2.3),
a parameter estimate

and a probability density for parameter uncertainty
f
)
1. add the constraint (7.11) to (2.1)(2.3) and solve in a Monte-Carlo
fashion for N
s
samples of

drawn from f
)
2. evaluate Conditions 1 and 2 by observing the distribution of u
i
(
), i
I
a
and M( x(
), u(), d
0
) for the N
s
samples.
Probability densities for parameter uncertainty f
) can be found through

bootstrapping. The number of samples N
s
should be a suciently large
number, a rule of thumb often used in the bootstrapping literature is that
at least 200 samples should be considered (Efron and Tibshirani, 1993).
116
Candidates I
a
can be evaluated in a trial-and-error fashion using Algo-
rithm 7.1. If several candidates are found which obey Conditions 1 and 2,
then these can be ranked by the prot increase expected by altering the
decision variables.
7.3 Case study
In this section the method suggested in Section 7.2 is applied to a case study
of a real world oil and gas eld. The approach to production modeling and
optimization is based on Chapter 4, but has not previously been applied to
this eld and some extensions are suggested.
7.3.1 Field Description
The eld considered is a North Sea oil and gas eld. The layout of the eld is
illustrated in Figure 7.1. A total of 31 wells produce predominantly oil, gas
and water from ve reservoirs. Reservoir 1 is produced by eight gas-lifted
wells which are connected to a subsea manifold which routes production
to separation facilities through two owlines, and the production of these
wells are measured with multiphase meters. During startup only gas-lift is
applied to the owlines. Reservoirs 2-5 are produced by 26 platform pro-
duction wells. Several of these wells can be gas-lifted and the production of
some wells are measured by multiphase measurements, and some wells have
downhole pressure measurements. The production of these wells are either
routed to a production separation train or to one of two test-separators.
Produced oil is measured before being stored and later exported by ship.
Produced water is re-injected after treatment. Produced gas can be exported
through one of two pipelines, injected in a gas-injection well for storage, used
as fuel gas or ared. There is also a gas-lift system. Gas export pipelines
are also used to import gas from other elds for injection and storage in an
gas-reservoir associated with the eld.
Many wells produce from thin layers, as a result variation in production
from individual wells is relatively signicant for the eld in question. Some
wells have production which is transient in nature, with gas-oil ratio that
increases steadily after startup and where so-called ushed production, a
transient peak in oil production with low water production, is visible after
startup. Well tests are performed relatively frequently for wells with sig-
nicant transients, at frequencies ranging from a week to several months
depending on the rate of change in well production. Multi-rate well tests
are performed less frequently for other wells ranging from once every four
7.3. Case study 117
Reservoir 1
Reservoir 2
Reservoir 3
Reservoir 4
Reservoir 5
Water
injection
Water
injection
Routing
2 subsea
manifolds
2 flowlines
Gaslift
Gaslift
Production
separation
train
2 test
separators
Oil
Gas
Water
Oil
Gas
Water
Total produced Oil
Export
Total produced Gas
2 Gas Injection
wells
Water-in-oil
Flare
Fuel
Import/
Export 1
Import/
Export 2
FT
FT
FT
FT
FT
FT
FT
FT
FT
FT
FT
FT
FT
MFT
5 wells
Gaslift
Gaslift
28 platform wells,
varying instrumentation
not shown
Total produced Oil
Total produced Gas
Total produced Water
Figure 7.1: Top: an illustration of the layout of the real-world oil and gas
eld considered. Bottom: an illustration of how total produced oil and gas
can be aggregated from other measurements.
118
months to once every 18 months.
One well is considered the swing-producer and production of this well
is frequently varied by a closed-loop controller to ensure that inlet separa-
tor pressure remains close to its constraint, in order to improve processing
capacity utilization.
Currently wells are optimized with the objective of maximizing total oil
production, and the active constraint is total gas processing capacity. Some
wells are constrained to a maximum liquid rate and others to a maximum
choke opening to avoid sand production. The wellhead pressure of some
wells are constrained to avoid damage in the near-well area of the reservoir.
Some wells exhibit slugging, which production optimization must take into
account. Currently, monthly targets for production choke openings and gas-
lift rates are determined. Expected wellhead pressures are also specied as
targets for operators, although these are expected to vary signicantly for
some wells.
7.3.2 Measurements used and well types
As the wells for the eld considered have characteristics which vary greatly,
dierent model structures may be required for dierent wells. Based on
observations of recent historical production data it is proposed that, for the
purpose of modeling, the wells for the eld considered can be divided into
the following categories:
Type A: platform wells where production optimization should consider the
opening of production chokes as decision variables. No rapid transients
are observed in measurements (16 wells).
Type B: platform wells where production optimization should consider the
gas-lift rates as decision variables. No rapid transients are observed in
measurements (5 wells).
Type C: wells attached to subsea manifolds (5 wells).
Type D: swing producer varied by closed-loop controller (1 well).
Type E: wells with rapidly transient production. Gas-oil ratio increases
continuously after startup on these wells, and these wells are shut
down periodically. All such wells have multiphase meters (4 wells).
Throughout this chapter I
A
, I
B
, I
C
, I
D
, I
E
are used to describe the sets of
indices of all wells of category A, B, C, D and E, respectively.
7.3. Case study 119
Let q
f,i
o
, q
f,i
g
, q
f,i
w
be oil, gas and water rates for well i measured with
multiphase meters. In this chapter the process model will be tted to the
measurement vector
y(t)

=
q
tot
o
(t)
q
tot
g
(t)
q
tot
w
(t)
(7.12)
where, with reference to Figure 7.1,
q
tot
o
= Oil export rate (100 Water in oil%)/100
iI
D
,I
E
q
f,i
o
, (7.13)
q
tot
g
= Flare gas + Fuel gas
+ Gas injected
+ Gas export/import line 1
+ Gas export/import line 2
iI
D
,I
E
q
f,i
g
(7.14)
q
tot
w
= Produced water
iI
D
,I
E
q
f,i
w
(7.15)
In (7.13)(7.15) it is exploited that the production of wells of type D and E
are measured with multiphase meters to subtract their contribution to total
measured rates.
7.3.3 Production optimization
Let z denote relative choke openings of platform production wells, z
f
denote
the relative choke openings of ow line chokes, let z
s
denote the relative
choke openings of subsea production wells, and let q
gl
denote gas-lift rates.
The decision variable which will be considered is
u =
_
z
i
i I
A
q
i
gl
i I
B
_
. (7.16)
The modeled disturbance d considered is
d =
z
i
i I
B
z
i
s
i I
C
z
i
f
i I
C
. (7.17)
120
The production optimization problem for the case considered can be
formulated as
u() = arg max
u
iI
a
q
i
o
(u, d, ) +b
o
() (7.18)
s.t. (7.19)
z
i
z
u,i
i I
MC
(7.20)
q
l,i
gl
q
i
gl
i I
MGS
(7.21)
z
i
z
u,i
i I
MBHP
(7.22)
u
0
(1 U
prc
) u u
0
(1 +U
prc
) (7.23)
q
i
o
(u, d, ) + q
i
w
(u, d, ) q
u,i
l
i I
MR
(7.24)
iI
a
q
i
g
(u, d, ) +
iI
B
q
i
gl
+b
g
() q
tot,c
g
(7.25)
iI
a
q
i
w
(u, d, ) +b
w
() q
tot,c
w
(7.26)
u
i
= u
i
0
i I
C
a
. (7.27)
I
MC
is the set of indices of wells which have maximum constraints on choke
openings due to sand production. I
MGS
is the set of indices of wells which
have minimum constraints q
l
gl
on gas-lift rates due to slugging. I
MR
is the
set of indices which have maximum constraints q
u
l
on the total rate of liquid
produced due to sand production. I
MBHP
is the set of indices of wells
that have minimum constraints on bottomhole pressure, which was chosen
mapped to maximum constraints z
u
on choke opening, as is discussed below.
U
prc
1 is a design parameter which limits the amount that decision
variables are allowed to be altered, intended to limit the inuence of model
uncertainty on production optimization. Larger setpoint changes can be im-
plemented by re-tting parameters and re-optimizing production iteratively
as suggested in Chapter 6.
Biases
b
o
() =
1
L
N
t=NL

iI
a
q
i
o
(u(t), d(t), ) q
tot
o
(t)
(7.28)
7.3. Case study 121
b
g
() =
1
L
N
t=NL

iI
a
q
i
g
(u(t), d(t), ) +

iI
B
q
i
gl
q
tot
g
(t)
(7.29)
b
w
() =
1
L
N
t=NL

iI
a
q
i
w
(u(t), d(t), ) q
tot
w
(t)
(7.30)
are chosen as the dierence between the production of wells with active
decision variables and measured production of all wells, averaged over a
time interval [N L, N]. Adding the bias b
o
() to the objective function
does not inuence the resulting u(), but allows the comparison of prots for
dierent values of tted parameters and dierent choices of active decision
variable sets I
a
.
Figure 7.2 shows the relationship between measured bottomhole pressure
and relative choke opening for a typical well on the eld considered. There is
signicant pressure dynamics due to changes in the reservoir and near-well
areas which may be challenging to model, yet the gradient between these two
variables seems to vary little. This observation motivates using the simple
model
p
i
bhp
= p
i
bhp,0
_
p
bhp
z
_
i
(z
i
z
i
0
) (7.31)
to determine a maximum choke opening, where z
i
0
is the choke opening of
well i and p
i
bhp,0
is the measured bottomhole pressure at the time production
optimization is performed, and
_
p
bhp
z
_
i
is a parameter determined from
recent historical data.
7.3.4 Modeling
Solving (7.18)(7.27) requires models for the production of oil, gas and water
for wells of type A, B and C. Chapter 4 considered the use of models which
are locally valid around the operating point of the most recent well test.
These models were of the form
q
i
= q
l,i
f
i
d
(d
i
, d
l,i
, ) f
i
u
(u
i
, u
l,i
, ) f
i
t
(t, t
l,i
, ) (7.32)
122
z
6
p
6
bhp
0 0.1 0.2 0.3 0.4 0.5
0.75
0.8
0.85
0.9
0.95
1
Figure 7.2: Measured bottomhole pressures for dierent relative valve open-
ings over time interval spanned by the tuning set (dotted lines). The lower
pressure constraint (solid horizontal), the pressure at the time of optimiza-
tion (circle), the predicted relationship between bottomhole pressure and
relative choke opening (dashed) and the predicted choke opening at which
the lower pressure constraint is reached (cross). All pressures are normalized.
7.3. Case study 123
where f
d
, f
u
, f
t
are kernels found through the methods of system identica-
tion, is a vector of parameters to be tted, and q
l,i
is the rate measured at
the most recent well test at setpoint (u
l,i
, d
l,i
) and time t
l,i
. In prior chapters
a eld where all wells were gas-lifted was considered, and this approach is
extended in this chapter.
Production choke-throttled platform wells
Example 7.1 motivates the choice of model structure for choke-throttled
wells.
Example 7.1 Consider a single well producing oil at rate q
1
o
, which varies
nonlinearly with relative choke opening z
1
. Consider a local linear model
around the operating point (q
l,1
o
, z
l,1
) of the form
q
1
o
= (z
1
z
l,1
) +q
l,1
o
. (7.33)
Consider that (7.33) is tted to the dataset shown in Figure 7.3 while the true
nonlinear relationship between q
1
o
and z
1
is as illustrated. As a result of the
shutdown, the parameter of the model (7.33) which described the dataset
best is a poor estimate of the local gradient around (q
l,1
o
, z
l,1
), and therefore
the tted model would not be well suited for production optimization.
Although data of the shutdown could have been removed from the tuning
set in this trivial example, on a eld with 20 or more wells, one or more
modeled wells may be shut down at any time. To exploit more of the available
data, a model structure which is able to describe the local gradient in spite
of shutdowns is desirable.
A model of the form q = q
l
f
z
(z, z
l
, ) is desired, where f
z
(z, z
l
, ) should
have the following properties
f
z
(0, z
l
, ) = 0 (7.34)
f
z
(z
l
, z
l
, ) = 1 (7.35)
2
f
z
(z, z
l
, )
z
2
0 z [0, 1] (7.36)
Eq. (7.34) implies that modeled production is zero when the production
choke is closed. Eq. (7.35) implies that modeled production matches most
recent single-rate well test q = q
l
when the opening matches the choke
opening tested z = z
l
. Eq. (7.36) implies that opening the choke an equal
amount has more inuence when choke opening is small than when choke
opening is large. Let x(u, ) be a vector of the rates of oil, gas and water
from each well as predicted by the production model.
124
z
1
q
o
1
t
q
o
1
z
1
Figure 7.3: Example 7.1: To the left: an imagined dataset which includes a
shutdown of well 1 (solid), and the predictions of a tted model (dotted) and
a model with correct local gradient (dotted). To the right: an illustration
of the true nonlinear model (solid), and the tted linear model (dotted) and
the model with correct local gradient (dashed) around the operating point
(q
l,1
o
, z
l,1
) (circle).
A production choke model on the linear-in parameters form
x(u, d, ) = (u, d) (7.37)
with only a single tted parameter assigned to describe the production of
each modeled uid of each well would ease parameter estimation. To make
(7.18)(7.27) easier to solve numerically, it would also be desirable for the
model to be dierentiable for all values of z, or smooth.
Choke-throttled wells are modeled as
q
i
= q
l,i
_
1 (1 z
i
)
k
i
1 (1 z
l,i
)
k
i
_
_
1 +
i
z

_
z
i
z
l,i
1
__
. (7.38)
This model was found by curve tting through trial and error and is mo-
tivated by the requirements stated above. k
i
are design parameters deter-
mined oine, while
i
z
are elements of .
Physical knowledge about the shape of the performance curves of pro-
duction choke-throttled wells can be included in the choice of k
i
and in
constraints on
i
z
. Although it is feasible that increasing the choke opening
z
i
may under special circumstances reduce the production of some compo-
nent of production, in most cases it is reasonable to expect increasing z to
7.3. Case study 125
cause production to either increase or remain the same. Based on this as-
sumption k
i
is chosen depending on z
l,i
so that
z
= 0 results in a
f
z
(z)
z
= 0.
Plots of the resulting shapes of (7.38) for 0 a
z
1 are shown in Figure
7.4 for dierent z
l,i
. As (7.34)(7.36) was found to be obeyed when k = 20
if z
l
= 0.1, and k = 6 if z
l
= 1, as shown in Figure 7.4, k
i
i chosen by the
heuristic k
i
= 6 + 14 (1 z
l,i
), interpolating between these values.
Gas-lifted platform wells
Gas-lifted well are modeled as in Chapter 4, where change in gas-lift rates
are considered the decision variable u = q
gl
and production choke openings
are considered a modeled disturbance d = z. The model structure
q
i
o
= max{0, q
l,i
o
f
d
(z
i
, z
l,i
) (1 +
i
o
q
i
gl
+
i
o
(q
i
gl
)
2
)} (7.39)
q
i
g
= max{0, q
l,i
g
f
d
(z
i
, z
l,i
) (1 +
i
g
q
i
gl
+
i
g
(q
i
gl
)
2
)} (7.40)
q
i
w
= max{0, q
l,i
w
f
d
(z
i
, z
l,i
) (1 +
i
w
q
i
gl
+
i
w
(q
i
gl
)
2
)} (7.41)
is used, where
f
d
(z
i
, z
l,i
) =
1 (1 z
i
)
k
i
1 (1 z
l,i
)
k
i
, (7.42)
and k
i
are design parameters determined oine.
Subsea wells
The aim of this case study was originally to optimize platform wells, i.e. wells
of types A and B, and to use multiphase meters to subtract the production
rates of subsea wells in (7.13)(7.15). Unfortunately, the multiphase meters
installed at the subsea wells were non-functional at the time of this analysis.
Modeling subsea wells that share a common owline is challenging as
changes in the production of one subsea well will inuence production of
other wells which share a common owline. If the production of subsea wells
were not modeled, parameter estimation could attempt to explain changes in
y that were due to change in the production of subsea wells by altering tted
parameters that describe platform wells, biasing parameter estimates. Based
on this reasoning and the diculties of modeling subsea wells, a very simple
model of subsea wells is considered in this chapter. No attempt to optimize
production of subsea wells is made, rather production from subsea wells
are treated as modeled disturbances, modeled to reduce bias in parameters
describing other wells.
126
f
z
f
z
f
z
z
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0
0.5
1
0
1
2
0
5
10
Figure 7.4: Graphs show f
z
(z) dened by (7.38) for dierent values
z
for
dierent z
l
(circle) and dierent k. Top graphs: z
l
= 0.1 and k = 20, center
graphs: z
l
= 0.5 and k = 13 and bottom graphs: z
l
= 1 and k = 6. In all
cases 0 <
z
< 1.
7.3. Case study 127
The ow from each subsea well inuenced by a owline choke opening
z
i
f
and a subsea production choke opening z
i
s
is modeled as
q
i
= q
l,i
f
z,f
(z
i
f
, z
l
f
) f
z
(z
i
s
, z
l,i
s
)
_
1 +
i
s

_
z
i
s
z
l,i
s
1
__
(7.43)
where
i
is a tted parameter, and f
z,f
(z
i
f
, z
l
f
) and f
z
(z
i
s
, z
l,i
s
) are of the
form (7.42).
Parameter estimation
Parameter estimation considers residuals which are related to the vector y by
(t) = y(u(t), d(t), ) y(t). The estimated measurement vector considered
is
y(u(t), d(t), ) = R x(u(t), d(t), ) +

y
(7.44)
x is a vector of oil, gas and water rates of the modeled wells and R is a
routing matrix of ones and zeros which maps x to y, and

y
are estimated
biases in each component of the measurement vector, in this case
y
=
_
o

g

w
T
. The parameter vector considered is
=
o,i
z
,
g,i
z
,
w,i
z
i I
A
o,i
gl
,
g,i
gl
,
w,i
gl
,
o,i
gl
,
g,i
gl
,
w,i
gl
i I
B
o,i
s
,
g,i
s
,
w,i
s
i I
C
(7.45)
Note that
o
,
g
,
w
are calculated over the time interval [1, N], while b
o
(),
b
g
(),b
w
() are calculated over a much shorter interval, therefore these two
sets of biases have dierent values in general.
The resulting vector of tted parameters has dim() = 96. Inequality
constraints of the form
L

U
(7.46)
are applied, and it is chosen to constrain 0 <
gl
< 1, 1 <
gl
< 0,
min
z

z

max
z
and 0
s
1 for all uids and for all wells.
The ratios between uids produced are rate-independent for some wells,
such as gas-oil ratio or water-oil ratio, and this qualitative knowledge can
be used to simplify the parameter estimation problem. The ratio between
128
phases m and n for well i is r
i
m,n
=
q
i
m
q
i
n
. Based on (7.38), a production-choke
throttled well with a rate-independent ratio r
i
m,n
obeys
r
i
m,n
=
q
l,i
m
(1 +
i
m
(z
i
z
l,i
))
q
l,i
n
(1 +
i
n
(z
i
z
l,i
))
z
i
. (7.47)
This relationship is enforced by soft constraints
V
s
() = w
r
_
i
m

i
n
_
2
(7.48)
for production choke-throttled platform wells and similarly by
V
s
() = w
r
_
i
m

i
n
_
2
+w
r
_
i
m

i
n
_
2
. (7.49)
for gas-lifted platform wells.
7.3.5 Implementation
The approach was implemented in MATLAB
1
and all optimization problems
were solved with the TOMLAB
2
toolbox. Solvers based on Huschens (1994)
were used for parameter estimation and based on Schittkowski (1983) for
production optimization, with analytical Jacobians and Hessians supplied.
U
prc
was chosen as U
prc
= 0.10, and bootstrapping considered N
s
= 500
re-samples. The two sets of active decision variables considered are
I
1
a
= I
A
I
B
{1, 19} and (7.50)
I
2
a
= {4, 11, 15, 16} . (7.51)
I
1
a
represents all modeled non-subsea wells which were not closed at the
time of optimization. Wells 1 and 19 were opened during the time interval
spanned by the tuning set, and therefore modeled, but these wells were closed
at the time of optimization. I
2
a
was determined by suggesting candidate sets
I
a
through observations of model uncertainty and assessing candidate sets I
a
with Algorithm 7.1. This procedure was repeated in a trial-and-error fashion
for a few iterations before the set I
2
a
was found, which obeys conditions 1
and 2.
A sampling interval of thirty minutes and a tuning set of two months
was considered. Total shutdowns were removed from the model by removing
all data points where total rates were below a threshold. The dataset was
1
2
7.3. Case study 129
manually inspected and some erroneous data removed. Measured rates from
multiphase meters have a seemingly random noisy element due to rapid
ow variations such as slugging. To avoid summing these rapid variations,
these measurements where smoothed. (q
l
o
, q
l
g
, q
l
w
, z
l
, q
l
gl
) was chosen as the
most recent allocated rates, supplied by the eld operator.
7.3.6 Results
The nominal parameter estimate are the parameters values that result from
solving (2.6) directly. The term nominal solution is used to refer to the
solution of (7.18)(7.27) which results from the nominal parameter estimate.
The nominal solution is shown for reference as it is the setpoint one would
likely implement if there were no uncertainty or if uncertainty were neglected.
The tuning set Z
N
is shown in Figure 7.5. Parameter estimation was
only able to nd weak local solutions, due to low information content in
the data set. Estimates of the tted model are compared with a section
of the tuning set in Figure 7.6, and with the validation set in Figure 7.7.
Estimated production is compared with measurements for the wells with
available multiphase meters in Figure 7.8 for the same time interval spanned
by Figure 7.7. The estimated models for wells of type A, B and C are shown
in Figures 7.9, 7.10 and 7.11, respectively. The distribution of estimates u()
and predicted prot and capacity utilization are shown for I
1
a
in Figures 7.12
and 7.13 and for I
2
a
in Figures 7.14 and 7.15, respectively. Fitted biases

o
,

g
and

w
were 2%, 15% and 10% of the total measured rate, respectively.
7.3.7 Discussion
The tted model represents the tuning set reasonably well, as seen in Figure
7.6, while a signicant bias between measurements and modeled rates is
observable in Figure 7.7. Parameter uncertainty as illustrated in Figures
7.9, 7.10 and 7.11 is signicant. These properties of the production model
are all as can be expected when the model is tted to a tuning set with low
information content. The model is able to correctly predict some responses in
production resulting from setpoint change in Figure 7.7, while other changes
in production cannot be described with the tted model. This indicates
that some tted parameters describe production well while others do not, as
expected. Some un-modeled changes in production may be a result of un-
modeled disturbances which can be compensated for by iteratively re-tting
production models in a moving-horizon manner.
As expected, values u()
i
I
1
a
found through Monte-Carlo simulation
and shown in Figure 7.12 vary signicantly as a result of low information
130
y
t
500 1000 1500 2000 2500
z
26
z
25
z
24
z
24
f
, z
25
f
, z
26
f
z
23
z
22
z
22
f
, z
23
f
z
21
, q
21
gl
z
20
, q
20
gl
z
19
, q
19
gl
z
18
, q
18
gl
z
17
, q
17
gl
z
16
z
15
z
14
z
13
z
12
z
11
z
10
z
9
z
8
z
7
z
6
z
5
z
4
z
3
z
2
z
1
0.6
0.8
1
1.2
Figure 7.5: Top graphs: total production of oil (solid), gas (dotted) and wa-
ter (dashdot) of modeled well and subsea wells. Middle graphs: production
valve openings (solid) and gas lift rates (dashed) of all modeled platform
wells. Bottom graphs: production valve openings (solid) of subsea wells and
owlines.
7.3. Case study 131
q
tot
o
q
tot
g
q
tot
w
t
2000 2200 2400 2600 2800
2000 2200 2400 2600 2800
2000 2200 2400 2600 2800
0.6
0.8
1
1.2
0.5
1
0.6
0.8
1
Figure 7.6: Measured production (solid) and modeled production (dashed)
for oil (top graphs), gas (middle graphs) and water (bottom graphs) for a
section of the tuning set. Plots down-sampled by a factor of 3 for legibility.
132
q
tot
o
q
tot
g
q
tot
w
t
2900 3000 3100 3200 3300 3400 3500
2900 3000 3100 3200 3300 3400 3500
2900 3000 3100 3200 3300 3400 3500
0.8
0.9
1
0.9
0.95
1
1.05
0.9
1
Figure 7.7: Validation data set spanning 1 month: Measured production
(solid) and modeled production (dashed) for oil (top graphs), gas (middle
graphs) and water (bottom graphs). Plots down-sampled by a factor of 3
for legibility.
7.3. Case study 133
i = 7
q
i
o
q
i
g
q
i
w
z
i
, q
i
gl
i = 8
i = 9
i = 14
i = 16
i = 17
i = 21
Figure 7.8: From left to right: measured oil rates, gas rates and water rates
(solid) compared with modeled rates (dashed) and most recent well test
(circles) for the validation data set spanning 1 month. Far right: relative
valve opening (solid) and gas-lift rate (dashed).
134
i = 1
q
i
o
q
i
g
q
i
w
i = 2
i = 3
i = 4
i = 5
i = 6
i = 7
i = 8
i = 9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
z
i
z
i
z
i
Figure 7.9: Models for oil (left), gas (center) and water (right) as a func-
tion of relative valve opening for production-valve throttled wells. Each
solid line illustrates the model plotted for a parameter value found through
bootstrapping. Dashed lines illustrate models plotted for upper and lower
bounds on parameter values. Most recent well tests shown as circles.
7.3. Case study 135
q
i
o
i = 17
q
i
g
q
i
w
i = 18
i = 19
i = 20
q
i
gl
i = 21
q
i
gl
q
i
gl
Figure 7.10: Models for oil (left), gas (centre) and water (right) as a function
of gas-lift rate for gas-lifted wells. Most recent well test (circle). Each
solid line illustrates the model plotted for a parameter value found through
bootstrapping.
136
q
i
o
i = 22
q
i
g
q
i
w
i = 23
i = 24
i = 25
z
i
i = 26
z
i
z
i
Figure 7.11: Models for oil (left), gas (center) and water (right) for subsea-
wells as a function of relative subsea valve opening. Each solid line illustrates
the model plotted for a parameter value found through bootstrapping.
7.3. Case study 137
z
2
z
3
z
4
z
5
z
6
z
7
z
8
z
9
z
10
z
11
z
12
z
13
z
14
z
15
z
16
q
17
gl
q
18
gl
q
20
gl
q
21
gl
Figure 7.12: Components of decision variable u for I
1
a
. Stems with squares
illustrate the most recent well tests, diamonds illustrate the currently imple-
mented value, dotted lines illustrate constraints, stems with circles illustrate
the nominal solution, bars illustrate the distribution of outcomes of Monte-
Carlo production optimization.
138
q
tot
o
q
tot
g
q
tot
w
0.98 0.99 1 1.01
0.85 0.9 0.95 1
1.04 1.06 1.08 1.1
Figure 7.13: Distribution of objective function and constraints of Monte-
Carlo production optimization with I
1
a
. Stems illustrate the nominal solu-
tion, dotted lines illustrate constraints, and bars illustrate the distribution
of outcomes from Monte-Carlo production optimization. Top graph: total
oil production, second graph: total gas production, third graph: total water
production. All values are normalized against currently measured values.
7.3. Case study 139
z
4
z
11
z
15
z
16
Figure 7.14: Components of decision variable u for I
2
a
. Stems with squares
illustrate the most recent well tests, diamonds illustrate the currently imple-
mented value, dotted lines illustrate constraints, stems with circles illustrate
the nominal solution, bars illustrate the distribution of outcomes of Monte-
Carlo production optimization.
140
q
tot
o
q
tot
g
q
tot
w
0.9995 0.9996 0.9997 0.9998 0.9999 1
0.97 0.975 0.98 0.985 0.99 0.995 1
1.03 1.035 1.04 1.045
Figure 7.15: Distribution of objective function and constraints of Monte-
Carlo production optimization with I
2
a
. Stems illustrate the nominal solu-
tion, dotted lines illustrate constraints, and bars illustrate the distribution
of outcomes from Monte-Carlo production optimization. Top graph: total
oil production, second graph: total gas production, third graph: total water
production. All values are normalized against currently measured values.
7.3. Case study 141
content. The estimated prot attainable from altering u
i
I
2
a
is signif-
icantly smaller than that attainable for u
i
I
1
a
, however the suggested
sign of setpoint change u
i
() u
i
0
, i I
2
a
is invariant to uncertainty f
),
and therefore one may be able to realize that prot by implementing deci-
sion variable change with an operational strategy. In contrast, the sign of
suggested setpoint change u
i
() u
i
0
, i I
1
a
is variant to uncertainty, and
therefore it may be dicult to nd a setpoint change involving all decision
variables which will increase prots with condence.
Figure 7.8 indicate the presence of signicant allocation error, and this
allocation error also reduces the ability of the model to correctly predict the
response of production to some changes in (u, d). Multiphase meters were
scheduled for re-calibrating, and measurement error may help explain the
deviation between measurements and estimates observed in Figure 7.8.
There is the possibility of improving model structure, parameter con-
straints and production constraints by including more physical and opera-
tional insight which was unavailable to the authors at the time of writing
this chapter, as much of this insight was not formally documented or docu-
mentation not up-to-date. A contribution of this chapter is to illustrate how
this form of insight can be included in the method suggested.
The analysis could have been much simplied by the addition of two ad-
ditional rate measurements. First, the lack of a dedicated measurement of
the total produced rate of gas meant that a large number of measurements
had to be aggregated, and with them measurement uncertainties, to estimate
this quantity. The large

g
relative to

o
is an indicator of large measure-
ment uncertainty for gas, and models may have been much improved by the
addition of a measurement of total produced rate of gas of scal quality,
similar to that of oil. Second, although our ambition was not to optimize
subsea wells, the malfunction of multiphase metering of those wells meant
that simple subsea models were still required to subtract the contribution of
those wells from total measured rates. Modeling of subsea wells could have
been avoided with the addition of a multiphase meter at each owline.
To simplify production optimization, the inclusion of pressures or tem-
peratures as inputs in the production model has not been considered, all
modeled disturbances d were assumed independent of u. Including mea-
sured temperatures and pressures can improve model performance and is
often done in models for other applications, but including these variables in
models for optimization introduces the additional problem of predicting how
pressures and temperatures will change with changes in u and d. It would
however be possible to extend the approach to include such measured vari-
ables only for those wells whose choke opening or gas-lift rate are considered
142
inactive decision variables and are unaltered by production optimization.
It may be possible to obtain larger prots than those indicated in Figure
7.15 by iteratively implementing changes in active decision variables, re-
modeling and re-optimizing. Also, the constraints used in simulations were
conservative.
An interesting avenue for further work is to design an algorithm which
removes the trial-and-error involved in determining I
a
with Algorithm 7.1.
7.4 Conclusions
Model uncertainty may prevent the determination of a setpoint change
through optimization which involves all decision variables and which will
increase prots with a high degree of condence. Active decision variables
methods may suggest change in selected decision variables which increase
prots with condence, which is practical for a human operator to imple-
ment, and which can be determined through analysis of data stemming from
normal operations. For these reasons active decision variables methods may
be of practical use in the oshore production of oil and gas.
Simulations indicated that a potential in the range of 4% could be re-
alized with high condence by altering four selected wells when a tentative
potential on the order of 4%10% was identied if all wells are considered.
Chapter 8
Conclusions and suggestions
for further work
In this chapter conclusions are drawn based on the work presented and
possible extensions of the work are suggested.
8.1 Conclusions
8.1.1 A data-driven approach to production modeling and
model updating
An approach to modeling of the oshore production of oil and gas which is
suitable for optimization has been presented. The design of models consid-
ered is motivated by the concepts of system identication. The modeling
approach was applied to two sets of real-world production data, and it was
suggested how physical knowledge can be included in parameter estima-
tion through constraints. Optimization has considered platform wells which
were assumed decoupled in modeling. Due to non-operational measurement
equipment, a rudimentary model of subsea wells was also suggested, which
was used for estimation in the second case study.
An advantage of the suggested modeling approach is that it requires lit-
tle manual maintenance, model parameters are found either from well-test
measurements or tted against data from normal operations. Simplied
maintenance may mean that the frequency at which new setpoints are imple-
mented can be increased in practice. The modeling approach suggested may
be especially well-suited when smaller changes in setpoints are implemented
in iterations by production optimization, as updating models between each
iteration may be feasible within practical time-constraints.
143
144 Conclusions and suggestions for further work
The ability of the suggested modeling approach to describe production
accurately will depend on the model structures chosen and on the prepro-
cessing of data. In both of these areas it may be possible to improve on the
results obtained by applying further process insight.
Based on the case studies, guidelines are presented for the ideal mea-
surement conguration on elds where this method is to be applied. Ideally
the total rates of separated oil, gas and water leaving each separation train
should be measured, where as conventionally total produced rates of gas and
water must frequently be estimated by aggregating several measurements.
A multiphase meter at each ow line transporting production from subsea
wells on elds with both platform and subsea wells would allow platform
wells to be optimized in isolation. The suggested methodology is more suit-
able when there are fewer wells produced from each separation train, as this
makes it easier to determine the contribution of each individual well from
total rate measurements.
The accuracy of the suggested modeling approach will depend on the
accuracy of measurements and the availability of a set of recent historical
data which is representative of current operations. The method may be
able to accommodate basic measurement errors, such as constant biases in
measured total rates. The method may be less suited for elds where wells
frequently change mode of operation, for instance elds where gas-lift is
frequently re-routed to dierent wells, as recent historical production data
may contain less representative data of current operation for such elds. If
un-modeled disturbances aect the production of wells at a frequency which
is high compared to the frequency of single-rate well tests, the suggested
method will also be less applicable, as this may introduce structural model
errors.
A simple model structure was used to demonstrate the idea of exploiting
information in measured total production rates using parameter estimation
in this thesis. This concept will also be applicable to more complex models,
such as those used in commercial simulators.
A challenge to the practical application of the methods suggested is mea-
surement quality. Feedback from measurements is central to the control
approach to production optimization considered in this thesis. Some of the
measurements required for this approach may currently be unreliable, un-
calibrated or unavailable. Improving measurement quality will ultimately
require industry awareness that improving measurement quality can enable
improved uncertainty handling in production optimization through feedback.
8.1. Conclusions 145
8.1.2 An explicit treatment of uncertainty in production op-
timization
This thesis has considered an explicit treatment of uncertainty in production
optimization. The estimates of uncertainty found for the two sets of real-
world production data in this thesis illustrate that the oshore production
of oil and gas is operated in a manner which introduces little excitation, and
therefore makes tted parameters in production models dicult to determine
uniquely. Although estimates of uncertainty are model-dependent in general,
the simple model structure considered combines the three basic elements
that would make up a more complex production model, namely it combines
information from single-rate well tests, information from measured total
rates and physical information. Therefore it is reasonable to expect that
more elaborate models would face similar issues with uncertainty to some
degree.
The key advantage of the explicit treatment of uncertainty is that it
enables analysis of production data as is without requiring initial experi-
ments. Analysis can for instance exploit data describing normal operations,
which is readily available. Another advantage of the explicit treatment of
uncertainty considered is that it associates quality tags with models, deci-
sion variables and prots, and providing a quantication of uncertainty and
its eects may itself be useful to practitioners.
The explicit treatment of uncertainty enabled the study the interactions
between the information content in production data, uncertainty in the tted
parameters of production models, and uncertainty in the setpoints resulting
from production optimization with uncertain models. An attempt has been
made to convey the consequences of uncertainty and these interactions in
monetary terms whenever possible, using the concepts of benet of excita-
tion, loss due to uncertainty or change in realized prot. These metrics,
although uncertain by their very nature, could provide a useful basis for
structured decision making.
The results of production optimization will depend on the constraints
considered. The emphasis of this thesis has not been to determine accurate
constraints, instead constraints were chosen conservatively to err on the side
of caution. It may be possible to improve on the results obtained by applying
further process insight to determine accurate non-conservative constraints.
A bootstrapping approach to estimating uncertainty in tted parame-
ters was considered in this thesis. This computational approach was chosen
for its conceptual simplicity, the method is simple to implement and intro-
duces few additional assumptions. The emphasis of this thesis has been to
gain an understanding of uncertainty in oshore oil and gas production, and
less emphasis has been placed on computational eciency or on comparing
dierent approaches to uncertainty estimation. An advantage of the com-
putational approach is that it can be applied regardless of model structure
or number or type of constraints considered, but computational time will
likely increase signicantly if more complex model structures were used. It
is interesting to note that in simulations results could be obtained on the
timescale of hours without particular attention to eciency, which indicates
that a computational approach may be suitable for production optimization
at intervals of days or weeks.
A limitation of the approach to describing uncertainty considered in this
thesis is that implicitly structural uncertainty is required to be small in
magnitude so that overall uncertainty can be expressed as uncertainty in
tted parameters. The uncertainty analysis considered cannot replace ac-
curate models, although the approach should be able to handle incomplete
knowledge of the value of physical model parameters.
Although treating uncertainty explicitly may be useful when production
data has low information content, production data must still contain some
minimum of information. If production data contains no information, pa-
rameter uncertainty approaches innity and proceeding with analysis will be
dicult. This problem may be faced in practice if some decision variables
are not altered at all in the production data set. A practical solution in
such cases is to exclude unvaried decision variables from optimization using
active decision variables, or to introduce planned excitation while accepting
that the benet of excitation cannot be estimated from data.
As long as production models are tted to recent historical data de-
scribing normal operations in some manner, the way in which the eld is
operated will have implications for the success of production optimization,
as illustrated in Figure 2.2. In general the elds studied are operated in a
manner which reveals less information than one would like, which can to
some extent be explained by operational considerations. The low informa-
tion content in data may also be an indication that model updating and
validation against measurements receives less emphasis within the current
engineering-approach to production modeling. Hopefully a contribution of
this thesis can be to raise awareness of the interaction between normal op-
erations, information content in data, and production optimization, which
may motivate operation of oshore oil and gas elds in a manner better
suited for a control-approach to modeling.
8.2. Directions for further work 147
8.1.3 An operational strategy approach to link optimization
with implementation and monitoring
A setpoint change cannot be implemented instantaneously, and the opera-
tional strategy presents a practical approach for adaptation to uncertainty
post-optimization which links production optimization with online imple-
mentation and monitoring. The suggested operational strategy exploits
measurements to ensure that production remains feasible, i.e. within ca-
pacity, without introducing conservativeness or equipment tripping due to
constraints being violated. The operational strategies considered are moti-
vated by how operators implement setpoint changes in practice.
The main limitation of the suggested approach is that it requires prots
and capacities to be measured and requires that the sign of the response in
capacity utilization and prots to a setpoint change is known. In the oshore
production of oil and gas these requirements are often met. In simulations
an operational strategy was considered which makes a number of small iter-
ative changes to decision variables in a manner which would be suitable for
automated implementation, but which may be too laborious for a human
operator to mimic. A human operator may need to make fewer and coarser
changes to setpoints, and this may diminish the eectiveness of the proposed
approach. Handling of transients explicitly was not considered explicitly in
the operational strategy, although it may be reasonable to assume that these
transients will be small when setpoint changes are small.
A contribution of this thesis is to include the operational strategy explic-
itly in analysis and consider it explicitly as an approach to handle uncertainty
in production optimization.
8.2 Directions for further work
8.2.1 Integration of production models with physics-based
simulators
Ideally a production model should express all available knowledge of produc-
tion, both empirical and physical. It is possible that predicted rates could be
found in simulation and expressed as constraints on parameters in parame-
ter estimation. The eect of this would be to integrate physical knowledge
and empirical knowledge, a form of tted proxy-model.
8.2.2 Active decision variables
Much further work is possible on combining the ideas of selecting active
decision variables and excitation planning. The fewer decision variables
are considered active and altered in each iteration, the more information
on the production of each well can be gained by observing responses in
total rates. A direction for further work may be to choose small groups of
active decision variables in each iteration of production optimization in a
manner which both reduces signicant parameter uncertainty and results in
increasing realized prots for each iteration.
8.2.3 Subsea wells
In this thesis only rudimentary modeling of subsea wells was considered,
which was used for estimation rather than optimization. Due to the pos-
sible coupling between the production of wells sharing a common owline,
and often the reduced availability of measurements, modeling subsea wells
may be more challenging than modeling platform wells but the rewards of
optimizing subsea wells may be greater.
8.2.4 Comparing alternative methods of estimating uncer-
tainty
Only bootstrapping residuals was considered in this thesis, but most or all
of the suggested methods can be combined with other methods for deter-
mining parameter uncertainty, some of which are mentioned in Section 1.2.
To reduce the computational eort required to produce estimates of param-
eter uncertainty, analytical methods or more ecient bootstrapping meth-
ods (Gigli, 1996) could be considered. The computational eort associated
with Monte-Carlo methods could be reduced by employing ecient sampling
techniques (Diwekar and Kalagnanam, 1997).
8.2.5 The benet of single-rate well tests
A well test can give two kinds of information. Firstly, the bias between
estimated and produced rates can be updated, and this information can be
gained form single-rate well tests. Secondly, the estimated response can be
updated against the measured response as decision variables or disturbances
change, and this information can be gained from multi-rate well tests.
In this thesis only the benet of performing a multi-rate well test was es-
timated. Performing a single-rate well test will also improve model accuracy,
and an avenue for further work is to estimate the benet of performing a
8.2. Directions for further work 149
single-rate well test. This thesis considers models which are local around an
operating point, and single-rate well tests could allow this operating point
to be updated. It may be possible to explicitly estimate the uncertainty
of the current operating point by observing what changes in the operating
point result in an equal or better t against the tuning set while reducing
estimated biases

. This estimate of uncertainty could then be exploited to
estimate the benet of single-rate well tests.
8.2.6 The value of reducing measurement uncertainty
As any model for production optimization must be tted to measurements in
some way, it may be that prots obtained from production optimization can
be signicantly increased through additional maintenance of existing instru-
ments or the installation of additional instruments. The main reason that
such measures have not already been performed is the lack of a documented
monetary benet. A structured approach to estimating the benet of re-
ducing measurement uncertainty using concepts similar to those presented
in this thesis is a direction for further work.
8.2.7 Constraint uncertainty
A challenge to the application of production optimization which has not
been considered explicitly is constraint uncertainty. Production optimiza-
tion requires constraints such as production capacities to be specied, these
specied capacities are themselves estimates which are subject to uncer-
tainty. If these estimates are conservative the production capacity may not
be fully utilized and lost prots may result. Combining the methods sug-
gested in this thesis with procedures for determining accurate capacities
without unnecessary conservativeness is a direction for further work.
8.2.8 Reservoir modeling
It may be possible to obtain production models which are accurate for longer
time intervals by including terms which provide a rudimentary description
of depletion or near-well dynamics. As such models may be able to describe
longer sets of historical production data, and longer sets of production data
generally include more excitation, extending production models in this man-
ner may reduce uncertainty in tted parameters.
8.2.9 Studying the link to reservoir management
Implementing production optimization without links to reservoir manage-
ment can result in increases in instantaneous prot at the expense of ulti-
mate recovery. In this thesis the topic of production optimization has been
studied in isolation, only implementing reservoir-related constraints in pro-
duction optimization that were specied by the operators of the studied
elds.
8.2.10 Managing transients in operational strategies
An operational strategy must implement setpoint changes slowly enough
to allow transients to settle and measured noise to be averaged out. In
instances where transients are signicant, an extension to the operational
strategy approach which predicts and manages transient dynamics could be
required to implement setpoint change within a reasonable time interval.
Such an extension could for instance be based on model-predictive control.
8.2.11 Exploiting shutdowns for model updating
It may be feasible to shut down wells sequentially during planned shutdowns
in a manner which allows the contribution of each well to be identied from
measured total rates, and possibly also allows responses to be identied.
Barring signicant un-modeled well or reservoir dynamics it could be possible
to calculate a new optimal setpoint using updated production models which
is implemented as production is re-started.
Appendix A
Data preprocessing
Before production data can be used to infer models or model parameters
it should be preprocessed to remove or adjust erroneous data. The quality
of preprocessing will impact the quality of inferred models and the quality
of production optimization. In this thesis a rudimentary pre-processing of
data has been performed in each of the case studies.
Through visual inspection of all measured variables spurious data was
removed, such as spikes in measured rates which could be due to measure-
ment malfunction. Measurements are often of dierent units, in which case
a unit conversion was performed. In some cases logged data were missing
for all variables for short time periods, in which case these intervals were
simply excluded from the tuning sets considered. When only a few mea-
surements were missing while others were present, missing values were de-
termined through interpolation. On shorter time-scales rates measured from
multiphase meters exhibit oscillations which may be attributed to slugging,
which were averaged out. The ranges of all measured variables were exam-
ined, to ensure that variables were within the ranges expected. In some cases
negative values were observed, for instance negative measured rates or nega-
tive choke openings. Negative values mostly occur during shutdowns, which
are excluded from the model anyway, but negative values were replaced
with zeros whenever observed. These negative values may be attributable
to production being outside the range for which measurement equipment
was calibrated, but may also be due to incorrect calibration.
Production chokes are often mounted in series with wing-valves. These
wing-valves are closed during shut-down. Measured settings of wing-valves
was not considered, in fact in the rst case-study performed the opening of
these wells where not logged, and this may have been a source of error.
Total shutdowns involving all wells were identied by observing measured
151
152 Data preprocessing
or estimated total rates of oil, gas and water in combination with produc-
tion choke openings and removed from the tuning sets considered. Time
intervals during which some wells were shut down while others were still
producing, partial shutdowns, were included in the tuning sets considered.
Care was taken to ensure that relative production choke openings were set
to zero during partial shutdowns. Gas-lift rates were applied in modeling
as supplied. It may be possible to reconcile gas-lift rate measurements by
installing equipment which estimates gas-lift rates by combining a calibrated
physical model with measured pressures, temperatures and rates.
Rates of total water and gas were aggregated from several measured
rates in the processing facilities, and either production specied in current
production and injection plans or wells tests summed for all wells were used
to validate aggregated total rates. No eort was devoted on ensuring that
total measured rates matched allocated rates precisely, as some discrepancy
should be expected due to disturbances since the times of the last well tests.
In some cases variation in measured total rates of oil, gas or water was ob-
served that could not be explained from variation in gas-lift rates or relative
choke openings modeled. These discrepancies may be due to un-modeled
disturbances in the reservoir or near-well areas, temporary measurement
malfunction or un-modeled change in the settings of processing facilities,
such as changes in the levels of a separator tank or the change in pressure
at an inlet separator. During intervals where such discrepancies are visible,
the residuals are large, and when these residuals are used to nd re-sampled
tuning sets during bootstrapping, larger estimates of parameter uncertainty
result. To err on the side of caution these residuals were included in when
designing re-sampled tuning sets, although more conservative estimates of
uncertainty may have been attainable by removing these residuals. Through
further analysis it may be possible either to explain these variations through
extensions to the models considered, or to identify these variations as spu-
rious data and remove them.
It may be possible to improve the quality of models identied in this
thesis by applying further eort toward the preprocessing of data, especially
for practitioners with more extensive eld-specic insight.
Bibliography
Almeida Netto, S. L., Schiozer, D. J., Ligero, E. L. and Maschio, C. (2003),
History matching using uncertainty analysis, in Petroleums Societys
Canadian International Petroleum Conference 2003, Calgary, Alberta.
Andersen, H. W., Rasmussen, K. H. and Jorgensen, S. B. (1991), Advances in
process identication, in Fourth International Conference on Chemical
Process Control, South Padre Island, Texas.
Awasthi, A., Sankaran, S. and Nikolau, M. (2008), Meeting the challenges
of real-time production optimization - a parametric model-based ap-
proach, in 2008 SPE Intelligent Energy Conference and Exhibition,
Amsterdam, The Netherlands. SPE 111853.
Barragn-Hernndez, V., Vzquez-Romn, R., Rosales-Marines, L. and
Garca-Snchez, F. (2005), A strategy for simulation and optimiza-
tion of gas and oil production, Computers & Chemical Engineering
30(2), 215227.
Becker, R., Hall, S. and Rustem, B. (1994), Robust optimal decisions with
stochastic nonlinear economic systems, Journal of Economic Dynamics
and Control 18, 125147.
Beggs, H. D. (2003), Production Optimization Using Nodal Analysis, 2nd
edn, Oil & Gas Consultants Intl., Tulsa, Oklahoma.
Bieker, H. P., Slupphaug, O. and Johansen, T. A. (2006), Optimal well-
testing strategy for production optimization: A monte carlo simulation
approach, in SPE Eastern Regional Meeting, Canton, Ohio. SPE
104535.
Bieker, H. P., Slupphaug, O. and Johansen, T. A. (2007a), Real time pro-
duction optimization of oshore oil and gas production systems: A.
technology survey, SPE Production & Operations 22(4), 382391. SPE
99446.
153
154 Bibliography
Bieker, H. P., Slupphaug, O. and Johansen, T. A. (2007b), Well manage-
ment under uncertain gas or water oil ratios, in SPE Digital Energy
Conference and Exhibition, Houston, Texas. SPE 106959.
Box, G. and Draper, N. (1969), Evolutionary operation; A statistical method
for process improvement, Wiley, New York, NY.
Branco, C. C. M., Pinto, A. C. C., Tinoco, P. M. B., Vieira, P. M. F.,
Sayd, A. M., Santos, R. L. A. and Prais, F. (2005), The role of the
value of information and long horizontal wells in the appraisal and
development studies of a brazilian oshore heavy-oil reservoir, in SPE
International Thermal Operations and Heavy Oil Symposium, Calgary,
Alberta, Canada. SPE 97846.
Campi, M. C. and Weyer, E. (2005), Guaranteed non-asymptotic condence
regions in system identication, Automatica 41(10), 17511764.
Chierici, G. L. (1995), Principles of petroleum reservoir engineering, Vol. 1,
Springer-Verlag, Berlin.
Costa, A. P. A., Schiozer, D. J. and Poletto, C. A. (2006), Use of un-
certainty analysis to improve production history matching and the
decision-making process, in SPE Europec/EAGE Annual Conference
and Exhibition, Vienna, Austria. SPE 99324.
Crowe, C. M. (1996), Data reconciliation progress and challenges, Jour-
nal of Process Control 6, 8998.
Darlington, J., Pantelides, C. C., Rustem, B. and Tanyi, B. A. (1999), An
algorithm for constrained nonlinear optimization under uncertainty,
Automatica 35(2), 217228.
Diwekar, U. M. and Kalagnanam, J. (1997), Ecient sampling technique
for optimization under uncertainty, AIChE Journal 43(2), 440449.
Efron, B. and Tibshirani, R. J. (1993), An Introduction to the Bootstrap,
Chapman & Hall.
Elgsaeter, S. M., Slupphaug, O. and Johansen, T. A. (2007), Challenges in
parameter estimation of models for oshore oil and gas production, in
2007 International Petroleum Technology Conference, Dubai,U.A.E.
Elgsaeter, S. M., Slupphaug, O. and Johansen, T. A. (2008a), Oil and gas
production optimization; lost potential due to uncertainty, in IFAC
World Congress 2008, Seoul, Korea.
Bibliography 155
Elgsaeter, S. M., Slupphaug, O. and Johansen, T. A. (2008b), Production op-
timization; system identication and uncertainty estimation, in Proc.
SPE Intelligent Energy Conference 2008, Amsterdam The Netherlands.
Elgsaeter, S. M., Slupphaug, O. and Johansen, T. A. (2009), A structured
approach to optimizing oshore oil and gas production with uncertain
models, Computers & Chemical Engineering? . submitted.
Engell, S. (2006), Feedback control for optimal process operation, Journal
of Process Control 17, 203219.
Feldbaum, A. A. (1961a), Dual control theory, parts i, Automation and
Remote Control 21(9), 874880.
Feldbaum, A. A. (1961b), Dual control theory, parts ii, Automation and
Remote Control 21(11), 10331039.
Forbes, J. F. and Marlin, T. E. (1994), Model accuracy for economic opti-
mizing controllers: the bias update case, Ind.Eng.Chem.Res. 33, 1919
1929.
Forbes, J. F., Marlin, T. E. and Yip, W. S. (2005), Real-time optimization:
Status, issues and opportunities, Encyclopedia of Chemical Processing
1(1), 25852598.
Foss, B. A. (1987), On parameter identication in reservoirs, PhD thesis,
The Norwegian Institute of Technology.
Franois, G., Srinivasan, B. and Bonvin, D. (2005), Use of measurements
for enforcing the necessary conditions of optimality in the presence of
constraints and uncertainty, Journal of Process Control 15(6), 701
712.
Frey, H. C. and Burmaster, D. E. (1999), Method for characterizing
variability and uncertainty: Comparison of bootstrap simulation and
likelihood-based approaches, Risk Analysis 19(1), 109130.
Friedman, Y. Z. (1995), What is wrong with closed-loop optimization?,
Hydrocarbon Processing Journal, October.
Friedman, Y. Z. (1997), Advanced process control - it takes eort to make
it work, Hydrocarbon Processing Journal, February.
156 Bibliography
Garatti, S., Campi, M. C. and Bittanti, S. (2004), Assessing the quality of
identied models through the asymptotic theory - when is the result
reliable?, Automatica 40(8), 13191332.
Gevers, M. (2006), A personal view of the development of system identi-
cation, IEEE Control Systems Magazine 26(6), 93105.
Gevers, M. and Ljung, L. (1986), Optimal experiment designs with respect
to the intended model application, Automatica 22(5), 543554.
Gigli, A. (1996), Ecient bootstrap methods: A review, J. ltal. Statist. Soc
5(1), 99127.
Griess, B. K., Diab, A. and Schulze-Riegert, R. (2006), Application of global
optimization techniques for model validation and prediction scenarios
of a north african oil eld, in SPE Europec/EAGE Annual Conference
and Exhibition, Vienna, Austria. SPE 100193.
Halvorsen, I. J., Skogestad, S., Morud, J. and Alstad, V. (2003), Optimal
selection of controlled variables, Ind. Eng. Chem. Res. 42(14), 3273
3284.
Huschens, J. (1994), On the use of product structure in secant methods
for nonlinear least squares problems, SIAM Journal of optimization
4(1), 108129.
Jalilova, N., Tautiyev, A., Forcadell, J., Rodriguez, J. C. and Sama, S.
(2008), Production optimization in an oil producing asset - the bp azeri
eld optimizer case, in SPE Gulf Coast Section 2008 Digital Energy
Conference and Exhibition, Houston, Texas. SPE 111853.
Johansen, T. A. (1997), On tikhonov regularization, bias and variance in
nonlinear system identication, Automatica 33(3), 441446.
Jones, D. R., Perttunen, C. D. and Stuckman, B. E. (1993), Lipschitzian
optimization without the lipschitz constant, Journal of Optimization
Theory and Applications 79(1), 157181.
Kall, P. and Wallace, S. W. (1994), Stochastic programming, Wiley, New
York, NY.
Krstic, M. (2000), Performance improvement and limitations in extremum
seeking control, Systems & Control Letters 39(5), 313326.
Bibliography 157
Little, A. J., Fincham, A. and Jutila, H. A. (2006), History matching
with production uncertainty eases transition into prediction, in SPE
Europec/EAGE Annual Conference and Exhibition, Vienna, Austria.
SPE 100206.
Ljung, L. (1999), System identication: Theory for the user, 2nd edn, Pren-
tice Hall, Cambridge.
Loeblein, C., Perkins, J. D., Srinivasan, B. and Bonvin, D. (1999), Eco-
nomic performance analysis in the design of on-line batch optimization
systems, Journal of Process Control 9(1), 6178.
Martinez, E. C. (2000), Batch process modeling for optimization using rein-
forcement learning, Computers and Chemical Engineering 24(2), 1187
1193.
Metropolis, N. and Ulam, S. (1949), The monte carlo method, Journal of
the American Statistical Association 44(247), 335341.
Miletic, I. and Marlin, T. E. (1998), On-line statistical result analysis in
real-time operations optimization, Ind. Eng. Chem. Res. 37, 3670
3684.
Mjaavatten, A., Aasheim, R., Saelid, S. and Gronning, O. (2006), A model
for gas coning and rate-dependent gas/oil ratio in an oil-rim reservoir,
in SPE Russian Oil and Gas Technical Conference and Exhibition,
Moscow, Russia. SPE 102390.
Mochizuki, S., Saputelli, L. A., Kabir, C. S., Cramer, R., Lochmann, M. J.,
Reese, R. D., Harms, L. K., Sisk, C. D., Hite, J. R. and Escorcia, A.
(2006), Real time optimization: Classication and assessment, SPE
Production & Operations 21(4), 455466. SPE paper 90213.
Mulvey, J. M. and Vanderbei, R. J. (1995), Robust optimization of large-
scale systems, Operations research 43(2), 264281.
Nadar, M. S., Schneider, T. S., Jackson, K. L., McKie, C. J. N. and Hamid,
J. (2008), Implementation of a total-system production optimization
model in a complex gas lifted oshore operation, SPE Productions &
Operations 23(1), 513.
Naus, M. M. J. J., Dolle, N. and Jansen, J. D. (2006), Optimization of
commingled production using innitely variable inow control valves,
SPE Production & Operations 21(2), 293301.
158 Bibliography
Nikolaou, M., Cullick, A. S. and Saputelli, L. (2006), Production optimiza-
tion - a moving-horizon approach, in SPE Intelligent Energy Confer-
ence and Exhibition, Amsterdam, The Netherlands. SPE 99358.
Ogunnaike, B. A. (1995), A contemporary industrial perspective on process
control theory and practice, A. Rev. Cont 20, 18.
Poulisse, H. (2008), Method and system for production metering of oil wells,
U.S. Patent. US2007295501.
Poulisse, H., van Overschee, P., Briers, J., Moncur, C. and Goh, K. C.
(2006), Continuous well production ow monitoring and surveillance,
in SPE Intelligent Energy Conference and Exhibition, Amsterdam,
The Netherlands. SPE 99963.
Pritchard, D. J. and Bacon, D. (1978), Prospects for reducing correlations
among parameter estimates in kinetic models, Chemical Engineering
Science 33, 15391543.
Qin, S. and Badgwell, T. (2003), A survey of industrial model predictive
control technology, Control Engineering Practice 11, 733764.
Rastoin, S., Schmidt, Z. and Doty, D. R. (1997), A review of multiphase ow
through chokes, Journal of Energy Resources Technology 119(1), 112.
Renard, G., Dembele, D., Lessi, J. and Mari, J. L. (1998), System iden-
tication approach applied to watercut analysis in waterooded lay-
ered reservoirs, in SPE/DOE Improved Oil Recovery Symposium,
Tulsa,OK. SPE 39606.
Roberts, P. D. and Williams, T. W. (1981), On an algorithm for com-
bined system optimization and parameter estimation, Automatica
17(1), 199209.
Rowan, G. and Clegg, M. W. (1963), The cybernetic approach to reservoir
engineering, in 38th Annual Fall Meeting of the Society of Petroleum
Engineers of AIME, New Orleans, Louisiana. SPE paper 727.
Saputelli, L. A., Mochizuki, S., Hutchins, L., Cramer, R., Anderson, M. B.,
Mueller, J. B., Escorcia, A., Harms, A. L., Sisk, C. D., Pennebaker,
S., Han, J. T., Brown, A., Kabir, C. S., Reese, R. D., Nuez, G. J.,
Landgren, K. M., McKie, C. J. and Airlie, C. (2003), Promoting real-
time optimization of hydrocarbon producing systems, in 2003 SPE
Oshore Europe Aberdeen, Aberdeen. SPE 83978.
Bibliography 159
Saputelli, L., Nikolaou, M. and Economides, M. J. (2005), Self-learning
reservoir management, SPE Reservoir Evaluation & Engineering
8(6), 534547.
Schittkowski, K. (1983), On the convergence of a sequential quadratic pro-
gramming method with an augmented lagrangian line search function,
Math. Operations sch. u. Statist., Ser. Optim. 14, 197216.
Schulze-Riegert, R. W., Haase, O. and Nekrassov, A. (2003), Combined
global and local optimization techniques applied to history matching,
in SPE Reservoir Simulation Symposium, Houston, Texas. SPE paper
79668.
Sjberg, J., Zhang, Q., Ljung, L., Benveniste, A., Delyon, B., Glorennec,
P. Y., Hjalmarsson, H. and Juditsky, A. (1995), Nonlinear black-box
modelling in system identication: a unied overview, Automatica
31(12), 16911724.
Skogestad, S. (2000), Self-optimizing control: the missing link between
steady-state optimization and control, Comput.Chem.Eng. 24(2), 569
575.
Sutton, R. S. and Barto, A. G. (1998), Reinforcement learning; an introduc-
tion, MIT, Cambridge, Ma.
Terwiesch, P., Agarwal, M. and Rippin, D. W. T. (1994), Batch unit opti-
mization with imperfect modeling - a survey, Journal of Process Con-
trol 4(4), 238258.
Tribbe, C. and Mller-Steinhagen, H. M. (2000), An evaluation of the per-
formance of phenomenological models for predicting pressure gradient
during gas-liquid ow in horizontal pipelines, International Journal of
Multiphase Flow 26(6), 10191036.
Ullmann, A. and Brauer, N. (2006), Closure relations for two-uid models
of two-phase stratied smooth and stratied wavy ows, International
Journal of Multiphase Flow 32(1), 82105.
Utvik, O. H., Rinde, T. and Valle, A. (2001), An experimental comparison
between a recombined hydrocarbon-water uid and a model uid sys-
tem in three-phase pipe ow, Journal of Energy Resources Technology
123(4), 253259.
160 Bibliography
van Dijk, F., Goh, K. C. and van Lienden, J. W. (2008), Closing the loop
for improved oil and gas production management, in 2008 SPE Intelli-
gent Energy Conference and Exhibition, Amsterdam, The Netherlands.
SPE 111997.
Wang, P. (2003), Development and application of production optimization
for petroleum elds. PhD thesis.
Wang, P., Litvak, M. and Aziz, K. (2002), Optimization of production op-
erations in petroleum elds, in SPE Annual Technical Conference and
Exhibition, San Antonio, Texas. SPE 77658.
Yang, C., Nghiem, L., Card, C. and Bremeier, M. (2007), Reservoir model
uncertainty quantication through computer-assisted history matching,
in 2007 SPE Annual Technical Conference and Exhibition, Anaheim,
California. SPE 109825.
Yip, W. S. and Marlin, T. E. (2003), Designing plant experiments for real-
time optimization systems, Contr. Eng. Practice 11, 837845.
Zangl, G., Graf, T. and Al-Kinani, A. (2006), Proxy modeling in production
optimization, in SPE Europec/EAGE Annual Conference and Exhibi-
tion, Vienna, Austria. SPE 100131.
Zhang, Y., Monder, D. and Fraser Forbes, J. (2002), Real-time optimiza-
tion under parametric uncertainty: a probability constrained approach,
Journal of Process Control 12(3), 373389.
Zhu, Y. (2006), System identication for process control: Recent experience
and outlook.

2008 Phd-Thesis: Modeling and Optimizing The Offshore Production of Oil and Gas Under Uncertainty (Steinar M. Elgsæter)

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

2008 Phd-Thesis: Modeling and Optimizing The Offshore Production of Oil and Gas Under Uncertainty (Steinar M. Elgsæter)

Hochgeladen von

Copyright:

Verfügbare Formate

Steinar M.

hard constraints on parameters

the covariance matrix of the asymptotic distri-

parameter values which are true in some sense

probability function for parameters

the set of indices of u which are to be decreased

() an optional constraint on . The components of y may

. The only modeled disturbance considered is

) that are infeasible, sub-optimal or may even reduce

are an expression of parameter un-

are the variance of each component of .

, d) and the cur-

) of (5.6) for a given parameter estimate

), d) M(x, u, d). (5.11)

) calculated with an uncertain

). When altering setpoints from u

for (5.13), and =

) are biases which

) should equal averaged observations over

) is included in the objective function (5.21) so that (5.11) yields

)) may dier from elements of

), for instance by bootstrapping as described in Chapter

be the set of indices corresponding the components of u

) in a Monte-Carlo manner. The approach is outlined in Algorithm 6.2

) found from recent histori-

have been determined

) and the ability of the operational strategy

). The benet of excitation will also

), an operational strategy, a process model which depends on parameters

). To limit the scope of this chapter, estimation

) can be found through

Das könnte Ihnen auch gefallen