This action might not be possible to undo. Are you sure you want to continue?
PRODUCTION OPTIMIZATION TECHNIQUES
FOR PETROLEUM FIELDS
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF PETROLEUM ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Pengju Wang
March 2003
ii
© Copyright by Pengju Wang 2003
All Rights Reserved
iii
I certify that I have read this thesis and that in my opinion it is
fully adequate, in scope and in quality, as a dissertation for the
degree of Doctor of Philosophy.
__________________________________
Dr. Khalid Aziz
(Principal Advisor)
I certify that I have read this thesis and that in my opinion it is
fully adequate, in scope and in quality, as a dissertation for the
degree of Doctor of Philosophy.
__________________________________
Dr. Roland N. Horne
I certify that I have read this thesis and that in my opinion it is
fully adequate, in scope and in quality, as a dissertation for the
degree of Doctor of Philosophy.
__________________________________
Dr. Anthony Kovscek
Approved for the University Committee on Graduate Studies:
_________________________________________
iv
Abstract
For some petroleum fields, optimization of production operations can be a major factor
on increasing production rates and reducing production cost. In this work we investigated
the formulations and solution methods for the following optimization problem:
determining the optimal production rates, lift gas rates, and well connections for a
gathering system with treelike structures to maximize daily operational objectives
subject to multiple flow rate and pressure constraints. While some aspects of this problem
have been studied by other investigators, existing methods are either inefficient or they
are based on significant simplifications that lead to suboptimal solutions. Hence, it is
necessary to develop approaches that can solve the problem efficiently and without
unreasonable assumptions.
We first investigated efficient procedures for simulating the production system and
computing sensitivity coefficients of production rates relative to system parameters.
These procedures have useful applications in simulation, sensitivity analysis, and
optimization of a petroleum field.
We then focused on the rate allocation problem, which refers to the determination of
the optimal production and lift gas rates subject to multiple constraints. When flow
interactions among different wells are not significant, the well performance can be
analyzed individually. Consequently, the rate allocation problem can be formulated as a
separable programming (SP) problem whose objective and constraint functions are sums
of functions of one variable. The SP problem is solved by various linear optimization
techniques. This method is very efficient. However, it may lead to bad solutions when the
flow interactions among wells are significant. In such cases, the rate allocation problem
v
should be appropriately formulated so that simulations capable of capturing such flow
interactions can be conducted in the optimization process. Several formulations were
investigated. The suggested formulation is able to handle well shutdown, avoid some
numerical difficulties, and is computationally efficient. Once formulated, the optimization
problem was solved by a sequential quadratic programming (SQP) algorithm.
A twolevel programming approach is developed to optimize the production rates,
lift gas rates, and well connections simultaneously. In this approach, the entire
optimization problem is solved in two levels. The upper level masters the overall solution
process and explicitly optimizes the well connections. For each new set of well
connections explored in the upper level, a lower level problem is formed to determine the
optimal set of production and lift gas rates. The lower level problem is a rate allocation
problem that can be solved by various optimization methods developed in this study. The
upper level problem is solved by a newly developed heuristic method. The method is both
efficient and robust, as verified by a genetic algorithm.
Production engineers often strive to achieve multiple conflicting goals when
operating a field. Several existing multiobjective optimization methods are used to
address this problem. Through an example, this study demonstrated that multiobjective
optimization methods can help decision makers identify the best tradeoffs.
The optimization tools were coupled with models for multiphase fluid flow in
reservoirs and surface pipeline networks through a commercial reservoir simulator. The
coupled procedure was applied to shortterm production optimization in the Prudhoe Bay
field of the NorthSlope of Alaska, and to longterm reservoir development studies for
two fields in the Gulf of Mexico. Results demonstrated the effectiveness of the approach.
vi
Acknowledgments
I am immensely grateful to my advisor Dr. Khalid Aziz for his sharing of my excitement
as well as frustrations in this research. Without his constant support, encouragement and
guidance, this work wouldn’t be possible. I also would like to thank Dr. Louis Durlofsky
for providing valuable advice on various occasions during this study. I sincerely thank Dr.
Roland Horne and Dr. Anthony Kovscek for reading the dissertation and providing
valuable comments.
Part of this work was conducted in BP. Dr. Michael Litvak provided tremendous help
in integrating the optimization tools into VIP (a commercial reservoir simulator of
Landmark Graphics Corporation) and testing the optimization methods on several field
models. I thank him for his help and guidance. Sincere thanks also go to other BP
colleagues who in one way or another contributed to my pleasant summer internship
experiences at BP. I also thank BP for providing me two summer internship opportunities
and the permission to put some field examples in this dissertation.
I thank the systems optimization laboratory at Stanford University for providing me
SNOPT 5.3, a general purpose nonlinear constrained optimization package. I thank
Mathematics and Computer Science Division at Argonne National Laboratory and the
Center for Research on Parallel Computation at Rice University for providing me
ADIFOR 2.0, an automatic differentiation software.
The reservoir simulation industrial affiliates program (SUPRIB) and the Department
of Petroleum Engineering have supported me financially. This support is gratefully
acknowledged.
vii
I am grateful for the friendship of and help from the faculty, staff, and fellow
students in the Petroleum Engineering Department. They contributed greatly to making
my stay at Stanford enjoyable. I especially thank my friend Tao Shu for supporting me in
many ways and bringing so much joy to my life for the last six years.
I deeply thank my parents, Jingmei Liu and Xingchun Wang. Their love and support
are a constant inspiration to me in striving to improve myself. This dissertation is
dedicated to you, Mom and Dad.
viii
Contents
Abstract .................................................................................................................. iv
Acknowledgments.................................................................................................. vi
Contents................................................................................................................viii
List of Tables........................................................................................................xiii
List of Figures ...................................................................................................... xiv
1. Introduction....................................................................................................... 1
1.1 Problem Statement................................................................................................. 1
1.2. Approach................................................................................................................ 4
1.3 Outline of Dissertation........................................................................................... 5
2. Literature Review ............................................................................................. 8
2.1 Overview of Optimization Techniques.................................................................. 8
2.1.1 Linear Programming.................................................................................... 9
2.1.2 Integer Programming................................................................................... 9
2.1.3 Nonlinear Programming............................................................................ 10
2.2 Applications of Optimization Techniques to Petroleum Fields........................... 14
2.2.1 Lift Gas and Production Rate Allocation .................................................. 14
2.2.2 Optimization of Production System Design and Operations..................... 17
2.2.3 Optimization of Reservoir Development and Planning ............................ 18
2.3 Concluding Remarks ........................................................................................... 20
3. Simulation of Gathering Systems .................................................................. 22
ix
3.1 Introduction.......................................................................................................... 22
3.2 Well Model .......................................................................................................... 24
3.3 Pipe Model........................................................................................................... 24
3.3.1 Pressure Gradient Equation....................................................................... 25
3.3.2 Liquid Holdup Evaluation......................................................................... 26
3.3.3 Upwards Twophase Flow Behavior......................................................... 27
3.4 Choke Model ....................................................................................................... 29
3.5 Network Simulation............................................................................................. 31
3.5.1 Governing Equations................................................................................. 32
3.5.2 The NewtonRaphson Method .................................................................. 34
3.5.3 Multiple Solutions..................................................................................... 36
3.6 Sensitivity Coefficients........................................................................................ 39
3.6.1 Finite Difference Method.......................................................................... 40
3.6.2 Jacobian Method ....................................................................................... 40
3.7 Examples.............................................................................................................. 42
3.7.1 Simulation Results .................................................................................... 43
3.7.2 Computation of Sensitivity Coefficients ................................................... 46
3.8 Concluding Remarks ........................................................................................... 49
4. Rate Allocation through Separable Programming ...................................... 51
4.1 Introduction.......................................................................................................... 51
4.2 Well Performance Estimation.............................................................................. 52
4.3 Rate Allocation through Separable Programming............................................... 54
4.3.1 Model LPI ................................................................................................ 56
4.3.2 Model MILPI ........................................................................................... 57
4.3.3 Model MILPII .......................................................................................... 60
4.3.4 Model LPII............................................................................................... 61
4.3.5 Discussions................................................................................................ 62
4.4 Application to Rate Allocation Problems ............................................................ 62
4.4.1 Handling Pressure and Velocity Constraints............................................. 63
4.4.2 Speedup Techniques.................................................................................. 64
x
4.4.3 A Gaslift Optimization Example ............................................................. 66
4.5 Concluding Remarks ........................................................................................... 68
5. Rate Allocation through Sequential Quadratic Programming................... 70
5.1 Introduction.......................................................................................................... 70
5.2 Sequential Quadratic Programming..................................................................... 71
5.3 Formulations of the Rate Allocation Problem..................................................... 73
5.3.1 Optimization Problem............................................................................... 73
5.3.2 Formulation P1.......................................................................................... 74
5.3.3 Formulation P2.......................................................................................... 75
5.3.4 Interpretation of Solutions of Problem P2................................................. 77
5.4 Solving the Optimization Problem...................................................................... 79
5.5 Examples.............................................................................................................. 80
5.5.1 Comparison of the MILPI Method and the SQP Method ........................ 80
5.5.2 Comparison of Formulation P1 and P2..................................................... 84
5.5.3 Efficiency of Formulation P2.................................................................... 86
5.6 Concluding Remarks ........................................................................................... 88
6. Optimization of Well Connections................................................................. 90
6.1 Introduction.......................................................................................................... 90
6.2 Twolevel Optimization Formulation.................................................................. 91
6.3 Solution Methods for the Optimization of Well Connections............................. 92
6.3.1 Computational Issues ................................................................................ 92
6.3.2 Partial Enumeration Method ..................................................................... 93
6.3.3 Genetic Algorithm..................................................................................... 94
6.4 Comparison of the Partial Enumeration Method and the Genetic Algorithm ..... 98
6.5 Concluding Remarks ......................................................................................... 101
7. Multiobjective Optimization of Production Operations............................ 102
7.1 Concept of Multiobjective Optimization........................................................... 102
7.2 Solution Techniques for Multiobjective Optimization Problems...................... 104
7.2.1 NoPreference Methods........................................................................... 105
xi
7.2.2 A Priori Methods..................................................................................... 105
7.2.3 Interactive Methods................................................................................. 107
7.2.4 A Posteriori Methods .............................................................................. 108
7.3 Multiobjective Optimization of Production Operations .................................... 110
7.3.1 Handling Soft Constraints ....................................................................... 111
7.3.2 Solution Methods .................................................................................... 111
7.4 Example ............................................................................................................. 113
7.5 Concluding Remarks ......................................................................................... 115
8. Coupling with a Reservoir Simulator ......................................................... 117
8.1 Introduction........................................................................................................ 117
8.2 Full Field Simulation......................................................................................... 118
8.2.1 Simulation of Petroleum Fields............................................................... 119
8.2.2 Coupling Reservoir and Production Network Solutions ......................... 122
8.3 Integration of Optimization Tools with VIPEXECUTIVE.............................. 124
8.3.1 A General Integration Procedure............................................................. 124
8.3.2 Integrated Optimization Tools................................................................. 126
8.4 Interfacing with an Independent Reservoir Simulator ....................................... 127
8.5 Applications....................................................................................................... 128
8.5.1 A Gulf of Mexico Oil Field Example...................................................... 128
8.5.2 Another Gulf of Mexico Field Example ................................................. 133
8.5.3 Application to the Prudhoe Bay Oil Field............................................... 139
8.6 Concluding Remarks ......................................................................................... 143
9. Conclusions.................................................................................................... 145
9.1 Conclusions........................................................................................................ 145
9.2 Summary of Contributions................................................................................. 146
9.3 Recommendations for Further Study................................................................. 148
Nomenclature...................................................................................................... 149
Bibliography........................................................................................................ 154
Appendix A. List of Programs........................................................................... 168
xii
A.1 Description of NETSO ...................................................................................... 168
A.1.1 Program Files .......................................................................................... 169
A.1.2 Imported Libraries and Software............................................................. 172
A.1.3 Compile and Run..................................................................................... 173
A.1.4 An Example Input File ............................................................................ 173
Appendix B. Problem Data for the GasLift Example in Section 4.4.3 ......... 178
xiii
List of Tables
Table 3.1: Slip model problem data .................................................................................. 28
Table 3.2: Network simulation results from different runs............................................... 45
Table 3.3: Comparision of sensitivity coefficients obtained from the Jacobian method and
the finite difference method ....................................................................................... 46
Table 4.1: Gaslift allocation results obtained from different methods ............................ 67
Table 4.2: Computational time of the MILPII method on various problems................... 68
Table 5.1: Well rates obtained in Step 1 of the MILPI method for Example 1 ............... 82
Table 5.2: Allocated lift gas and oil rates for Example 1.................................................. 83
Table 5.3: Convergence criteria for optimization problems in Example 3 ....................... 88
Table 5.4: Computational efficiency of Formulation P2 on various optimization problems
.................................................................................................................................... 88
Table 6.1: Performance data of the PE method................................................................. 99
Table 7.1: A classification of methods for multiobjective optimization......................... 104
Table 7.2: Pareto optimal solutions from the weighting function method...................... 114
Table B.1: Gas injection and oil production rates for a set of 56 wells ......................... 178
Table B.2: Gas injection and oil production rates for a set of 56 wells obtained from the
MILPII method ....................................................................................................... 181
xiv
List of Figures
Figure 3.1: Upstream pressure in a vertical well tubing string with a fixed downstream
(wellhead) pressure for different flow rates ............................................................... 29
Figure 3.2: A simple gathering system.............................................................................. 32
Figure 3.3: Illustration of multiple solutions for a single well system.............................. 38
Figure 3.4: Configuration of the gathering system............................................................ 43
Figure 3.5: Infinity norm of the governing equation residuals versus iteration number ... 44
Figure 3.6: Oil rate versus iteration number for Run 1 ..................................................... 44
Figure 3.7: Oil rate versus iteration number for Run 4 ..................................................... 45
Figure 3.8: Sensitivity coefficients of well oil rates with respect to system parameters (the
inner diameter of pipe 21 is 5.5 inches) ..................................................................... 48
Figure 3.9: Sensitivity coefficients of well oil rates with respect to system parameters (the
inner diameter of pipe 21 is 7.5 inches) ..................................................................... 48
Figure 3.10: Sensitivity coefficients of well oil rates with respect to system parameters
(the inner diameter of pipe 21 is 9.5 inches).............................................................. 49
Figure 4.1: Illustration of well performance curves .......................................................... 53
Figure 4.2: Illustration of gaslift performance curves...................................................... 61
Figure 4.3: The allocated lift gas and oil rates for well 47 from the MILPII method and
the LPI method.......................................................................................................... 68
Figure 5.1: Outflow performance curves with various choke settings.............................. 75
Figure 5.2: Impact of oil rate and lift gas rate on the well sandface pressure calculated
from the separator side............................................................................................... 79
Figure 5.3: Gaslift performance curves generated in Step 2 of the MILPI method........ 82
Figure 5.4: Convergence history of Formulation P1 and P2 for Scenario 1 ..................... 85
xv
Figure 5.5: Convergence history of Formulation P1 and P2 for Scenario 2 ..................... 86
Figure 5.6: Schematic illustration of gathering systems .................................................. 87
Figure 6.1: A simple GA flowchart................................................................................... 95
Figure 6.2: Encoding/decoding a well connection ............................................................ 96
Figure 6.3: Schematic illustration of singlepoint crossover............................................. 97
Figure 6.4: Schematic illustration of uniform crossover................................................... 97
Figure 6.5: Schematic illustration of mutation.................................................................. 97
Figure 6.6: Convergence history of the PE runs.............................................................. 100
Figure 6.7: Convergence history of the GA run.............................................................. 101
Figure 7.1: Graphical illustration of a Pareto optimal set ............................................... 103
Figure 7.2: The weighting functions method with convex set ........................................ 109
Figure 7.3: The weighting functions method with nonconvex set .................................. 110
Figure 7.4: The Pareto optimal sets obtained from the hierarchical method and the
weighting functions method..................................................................................... 115
Figure 8.1: Integration of optimization tools to VIPEXECUTIVE ............................... 125
Figure 8.2: The production system of a Gulf of Mexico oil field ................................... 129
Figure 8.3: Normalized daily total oil rate ...................................................................... 130
Figure 8.4: Normalized daily total gas rate ..................................................................... 131
Figure 8.5: Normalized daily total water rate.................................................................. 131
Figure 8.6: Normalized daily gas rate at HEADER_Y ................................................... 132
Figure 8.7: Normalized daily gas rate at HEADER_Z.................................................... 132
Figure 8.8: Normalized daily total oil production history for Case 1.............................. 136
Figure 8.9: Normalized cumulative oil production history for Case1............................. 136
Figure 8.10: Daily total gas rate history for Case 1......................................................... 137
Figure 8.11: Normalized cumulative lift gas injection for Case 1 .................................. 137
Figure 8.12: Normalized daily oil production history for Case 2.................................... 138
Figure 8.13: Normalized cumulative oil productive history for Case 2.......................... 138
Figure 8.14: Normalized cumulative lift gas injection for Case 2 .................................. 139
1
Chapter 1
Introduction
The ultimate goal of virtually all effort spent on modeling a petroleum field is to devise
an optimal strategy to develop, manage, and operate the field. For some petroleum fields,
optimization of production operations can be a major factor in increasing production rates
and reducing production costs. While for single wells or other small systems simple nodal
analysis may be adequate, large complex systems demand a much more sophisticated
approach to predict the response of a large complicated production system accurately and
to examine alternative operational scenarios efficiently.
As optimization algorithms and reservoir simulation techniques continue to develop
and computing power continues to increase, upstream oil and gas facilities previously
thought not to be candidates for advanced control or optimization are being given new
considerations (Clay et al., 1998). The objective of this study was to develop optimization
methods for solving a class of important production operation problems for petroleum
fields.
1.1 Problem Statement
In petroleum fields, hydrocarbon production is often constrained by reservoir conditions,
deliverability of the pipeline network, fluid handling capacity of facilities, safety and
economic considerations, or a combination of these considerations. The task of field
operators is to devise optimal operating strategies to achieve certain operational goals.
CHAPTER 1. INTRODUCTION 2
These goals can vary from field to field and with time. Typically one may wish to
maximize daily oil rates or minimize production costs. This research aims to develop
optimization methods that ease and automate the decision making of field operators for
certain operations. In this section, the major components (the objective, the control
variables, and the constraints) of the petroleum field optimization problem are described
in detail.
Objective. Usually, the objective is to maximize the profit from an oil field on a day
today basis. How to define the profit is a complicated issue that requires intensive study
in its own right. To keep this research focused, we restricted our attention to simple
objective functions such as maximizing weighted daily flow rates.
We emphasize that we only optimize production operations for a shortterm period.
The flow rate and other operational settings, once determined, are assumed to remain
fixed during that period. However the optimization procedures developed in this study
can be used repeatedly for longterm hydrocarbon recovery studies.
Control variables. The control variables are the production operation settings to be
optimized. In this study, the control variables included the lift gas rates, the production
rates, and the well connections to flow lines.
Continuous gaslift is a common artificial lift method used in the oil industry to
improve well performance. The mechanism of gaslift is fairly simple. Gas is injected
into the tubing string to lighten the liquid column and decrease the bottom hole pressure,
which allows the reservoir to push more fluids into the wellbore. At the same time,
increased flow rates in the tubing string and surface flow lines result in higher
backpressure on the well and adjacent wells that share a common flow line. This in turn
causes a reduction in well production rates. Therefore, lift gas has to be carefully
allocated to achieve maximum efficiency.
The production rate of a well is usually controlled by a choke. Adjusting production
rates is the most straightforward way to meet certain production targets and satisfy
certain operational constraints as described later.
CHAPTER 1. INTRODUCTION 3
In some petroleum fields such as the Prudhoe Bay oil field in Alaska, USA, a
production well can be connected to different flow lines that lead to different separation
units. In such fields, switching a well from one flow line to another flow line can be an
effective way to relieve the delivering/processing burden of one device/facility and
increase the delivering/processing efficiency of the overall system. Thus, well
connections were also considered as one type of decision variables in this study.
Constraints. Production operations in an oil field are usually subject to multiple
capacity, safety, and economic constraints.
Capacity constraints. It is not surprising to know that oil production in some
petroleum fields is constrained by the processing capacities of surface facilities, such as
the gas and water processing capacities of separators, and the gas compression capacity
of a central gas plant. For example, oil production in the Prudhoe Bay oil field (Barnes et
al., 1990) and the Kuparuk River field in Alaska (Stoisits et al., 1992, 1994, 1999) is
constrained by the gas handling capacities of surface facilities. The main reason for this is
that production systems are usually designed and installed at the very early stage of
reservoir development when information about the reservoir and future economics are
scarce or uncertain. Thus facility capacities may not always be able to meet the
production demand throughout the life of a reservoir. These constraints can be satisfied
and/or fully utilized by adjusting the lift gas and/or production rates and switching well
connections between flow lines.
Safety and economic constraints. For safety reasons, we may need a
maximum/minimum pressure constraint at the bottom of a well or on some surface
facility nodes. To operate the production system economically, we may put a
maximum/minimum flow rate constraints on certain production wells or facilities.
Furthermore, corrosion/erosion can lead to costly repairs. To avoid excessive
corrosion/erosion, fluid velocities may have to be limited (Svedeman and Arnold, 1994,
Kermani and Harrop, 1996). All these constraints can be satisfied by adjusting the
production rates with chokes.
In summary, our objective was to investigate optimization approaches to guide in
making operational decisions to enhance production subject to multiple capacity, safety,
CHAPTER 1. INTRODUCTION 4
and economic constraints. For example the engineer may be interested in the following
questions:
1. How to use chokes to control well rates?
2. How to distribute available liftgas among specified wells?
3. How to route fluids by switching well connections to flow lines?
1.2 Approach
The problem of interest requires simultaneous allocation of production rates, lift gas
rates, and well connections to flow lines. However, as discussed in Chapter 2, robust
procedures for such a task are not available. Either previous investigators have addressed
only a part of the problem, or ad hoc rules have been used that may lead to suboptimal
operations. Hence, it was necessary to develop approaches that optimize all control
variables simultaneously subject to all constraints. This dissertation investigated such
approaches.
The target problem is a nonlinearly constrained optimization problem with both
continuous and discrete decision variables. A twolevel programming approach was
developed to solve the problem. In this approach, optimization is conducted in two levels.
The upper level explicitly optimizes the well connections. For each set of well
connections, the upper level spawns a lower level problem to find the optimal production
and lift gas rates for that set of well connections. The upper level masters the overall
optimization procedure and the lower level can be viewed as a function evaluation
procedure for the upper level. The advantage of this twolevel programming approach is
flexibility. Because the continuous and discrete variables are optimized at different
levels, the approach offers us more freedom to use existing methods or develop new
methods for solving the overall problem.
The lower level problem optimizes the production and lift gas rates subject to
nonlinear constraints and is referred to in this study as the rate allocation problem. When
there are no flow interactions among different wells or when such interaction is not
significant, the performance of a well can be analyzed individually while ignoring the
impact of the rest of the wells in the system. As a consequence, the rate allocation
CHAPTER 1. INTRODUCTION 5
problem becomes a separable programming (SP) problem whose objective and constraint
functions are sums of functions of one variable (Gill et al., 1981). The SP problem can be
solved efficiently by linear optimization techniques.
The flow interaction among wells can play an important role in some rate allocation
problems. In such cases, the rate allocation problem is formulated as a general nonlinear
constrained optimization problem and solved by a Sequential Quadratic Programming
method (Gill et al., 2002). Different formulations have been investigated. The focus is on
how to handle the well chokes that control the production rates in the optimization
process. The preferred formulation chooses the production rates instead of the well
chokes as the decision variables and uses a set of constraints to guarantee that the optimal
set of production rates is feasible for the production system.
The upper level is an integer programming problem. Two methods were investigated.
The first is a heuristic method named in this study the partial enumeration method. This
method is an iterative procedure that sequentially finds the best connection for one well
while keeping other well connections fixed. The second method is a genetic algorithm
(Goldberg, 1989). This method encodes one set of well connections into one set of binary
strings, and evolves multiple sets of binary strings by means of selection, crossover, and
mutation to locate the best set of well connections.
The optimization approaches for the upper level and lower level were investigated
separately. The approaches can be used separately to optimize single types of operation
settings such as the lift gas distribution or combined together to optimize all types of
variables simultaneously.
1.3 Outline of Dissertation
General optimization techniques pertinent to this study are reviewed in Chapter 2, which
introduces the concept, solution algorithms, and applications of linear programming,
integer programming, and nonlinear programming. The chapter then describes the
applications of optimization techniques in the development and operations of petroleum
fields and sets the framework of this research.
CHAPTER 1. INTRODUCTION 6
Chapter 3 describes how the gathering system is simulated. In the oil industry, there
is no generally accepted method for simulating multiphase flow in general gathering
systems. The challenges lie in predicting the pressure drop across flow delivering
devices, determining the phase splitting ratios at flow junctions, finding the mathematical
solutions efficiently, and identifying the true physical solution from multiple
mathematical solutions. This research did not address all these challenges. Rather, the
study considered a simplified class of problems where the gathering system has a tree
like structure and flow directions are known. By exploiting the special properties of such
systems, an efficient solution method could be investigated. It was found that care should
be taken to identify the stable solution from multiple mathematical solutions. Further, an
efficient method was developed to compute sensitivity coefficients for a production
system. This method has useful applications in sensitivity analysis and optimization.
Chapter 4 describes how the rate allocation problem can be solved when flow
interaction among wells can be ignored. This section describes how to construct the well
performance information, how to formulate the rate allocation problem into a separable
problem, and how to solve the separable problem by linear optimization techniques. An
example is presented to demonstrate the advantages of the proposed method.
Chapter 5 describes how the rate allocation problem can be solved when flow
interaction can not be ignored. This chapter discusses the appropriate formulation of the
rate allocation problem under such cases. Once appropriately formulated, the problem
was solved by a Sequential Quadratic Programming (SQP) method. Application
examples demonstrated that the method is capable of handling rate allocation problems of
varying complexities and sizes.
Chapter 6 presents the two solution methods for the upper level problem: the partial
enumeration method and the genetic algorithm. Numerical experiments showed that the
partial enumeration method is more efficient than the genetic algorithm. The quality of
the solution by the simpler partial enumeration method was at least as good as that
obtained from the genetic algorithm.
Chapter 7 discusses possible applications of multiobjective optimization techniques
to production operations for a petroleum field. Often, production operations in a
CHAPTER 1. INTRODUCTION 7
petroleum field are subject to multiple conflicting goals, such as maximizing the daily oil
production while minimizing the water produced. Multiobjective optimization is a useful
tool to help decision makers identify the best tradeoffs.
Chapter 8 describes a procedure to integrate the optimization approaches with a
reservoir simulator. This procedure automates the use of the approaches developed for
both shortterm production optimization and longterm oil recovery studies. Chapter 8
presents several such applications. One example is the use of the optimization tools in the
EField Optimization System of Prudhoe Bay oil field, Alaska (Litvak et al., 2002).
Chapter 9 summarizes the work and recommends areas that need further study.
8
Chapter 2
Literature Review
Applications of optimization techniques in the upstream oil industry began in the early
1950s and have been flourishing since then. Applications have been reported for recovery
processes, planning, history matching, well placement and operation, drilling, facility
design and operation and so on. Optimization techniques employed in these applications
cover almost all subfields in mathematical programming, such as linear programming,
integer programming, and nonlinear programming. In this chapter, we first introduce
concepts, solution algorithms, and applications of some major subfields in mathematical
programming. Then we review their applications in areas that are pertinent to this study.
2.1 Overview of Optimization Techniques
Mathematical programming is a field born in the later 1940s (Lenstra et al., 1991).
Despite its short history, mathematical programming has developed into a sophisticated
field with deep specialization and great diversification. Mathematical programming
encompasses subfields such as linear programming, integer programming, nonlinear
programming, combinatorial optimization, stochastic programming, and so on.
Optimization problems in the most general form can be represented as
min ( ) ( ) { } m i u c l f
i i i
,..., 1 , : = ≤ ≤ x x (2.1)
CHAPTER 2. LITERATURE REVIEW 9
where the objective function f and the constraint functions { }
i
c are functions of control
variable x , and
i
l and
i
u are the lower and upper bounds for the th i constraint,
respectively. An optimization problem can be categorized according to the type of its
control variables, and objective and constraint functions.
2.1.1 Linear Programming
When the objective function f and constraint functions { }
i
c are linear functions of
control variable x , the problem described by Eq. 2.1 is a linear programming (LP)
problem. Various kinds of optimization problems can be formulated as LP problems,
such as transportation, production planning, resource allocation, and scheduling
problems.
In 1947, Dantzig proposed the first solution method for the LP problem, the simplex
algorithm. The simplex algorithm (Dantzig, 1963) finds the optimal solution by moving
along the vertices of the feasible region, thus its optimal solution is always a vertex (or an
extreme point) of the feasible region. In general, the simplex algorithm is very efficient.
However, for some problems, the simplex method can take an very large number of
iterations (Klee and Minty, 1972). Karmarkar (1984) introduced the first interior point
algorithm for the LP problem. The interior point algorithm approaches the optimal
solution from the interior of the feasible region. Thus the solution usually is not a vertex
of the feasible region. The interior point method is very efficient both theoretically and in
practice. Nowadays, LP problems with thousands or even millions of variables and
constraints can be solved efficiently by both the simplex and interior point algorithm
(Bertsimas and Tsitsiklis, 1997).
2.1.2 Integer Programming
When all components of the unknown x are discrete variables, the problem described by
Eq. 2.1 becomes an integer programming (IP) problem. When some but not all
components of x are discrete, the problem is a mixed integer programming (MIP)
problem. Discrete variables are useful to model indivisibility, logical requirements, and
on/off decisions. Integer programming finds its application in many real life problems,
CHAPTER 2. LITERATURE REVIEW 10
such as capital budgeting, airline crew scheduling, telecommunications, and production
planning. For the production optimization problem addressed in this study, the well
connections have to be modeled as discrete variables.
For linear integer programming (LIP) problems, the common solution techniques are
the cutting plane method and the branch and bound method. Both methods tackle the LIP
problem by solving a series of linear programming problems. The cutting plane method
was proposed by Gomory (1958). The fundamental idea is to add constraints to a series of
linear programming relaxations of the LIP problem until the optimal solution of a
relaxation problem takes integer values (a linear programming relaxation is formed by
allowing the integer variables in an LIP problem to take real values). In practice, the
cutting plane method is not effective. The most effective technique for solving the LIP
problem is the Branch and Bound method, which was first presented by Land and Doig
(1960). The Branch and Bound method uses a “divide and conquer” approach to explore
the set of feasible integer solutions. The method divides an optimization problem
recursively into multiple subproblems. Instead of enumerating all the subproblems, the
method uses bounds on the optimal objective value of a subproblem to avoid forming and
solving other subproblems.
Other methods for integer programming include dynamic programming (Bellman,
1957), simulated annealing (Kirkpatrick et al., 1983), genetic algorithms (Goldberg,
1989), and many special algorithms designed for particular problems. So far, there does
not exist a universal algorithm for integer programming: either the method takes a large
amount of time, or it only gives an approximate solution (Bertsimas and Tsitsiklis, 1997).
2.1.3 Nonlinear Programming
Eq. 2.1 becomes a nonlinear programming problem when its objective and/or constraint
functions are nonlinear. Nonlinear programming problems come in many different forms
and shapes. Algorithms have been developed for individual classes of problems. It is
impossible to give a thorough survey of all major optimization algorithms in a limited
space. Rather, this section gives a brief description of some of the optimization methods
encountered most frequently in the petroleum engineering literature. Some general
CHAPTER 2. LITERATURE REVIEW 11
references about nonlinear programming are Gill et al. (1981), Fletcher (1987), and
Bertsekas (1982).
Unconstrained optimization methods. An important class of methods for
unconstrained optimization problems is the socalled line search. Line search methods
approach a local minimum using the following iteration scheme:
k k k k
p x x α + =
+1
(2.2)
where
k
x and
1 + k
x are the current and next iterates,
k
p is a search direction along which
the function decreases, and
k
α is a step length that ensures “sufficient” progress toward
the solution.
A Newtonbased line search method (usually termed as Newton’s method) goes as
follows. At each iteration, the method approximates the objective function ( ) x f around
current iterate
k
x using a quadratic function defined as follows:
( ) ( ) ( )p x p p x x p
k
T T
k k k
f f f q
2
2
1
) ( ∇ + ∇ + = (2.3)
When the Hessian matrix, ( )
k
f x
2
∇ , is positivedefinite, the model function ( ) p
k
q has a
unique minimum which can be obtained by solving the following linear system
( ) ( )
k k k
f f x p x −∇ = ∇
2
(2.4)
The solution
k
p of Eq. 2.4 serves as the search direction of current iteration. The step
length
k
α can be obtained by a linear search procedure.
Newton’s method converges quadratically near a local optimum (Gill et al., 1981).
The key to this exceptional convergence rate is that the method uses the curvature
information of the objective function (the Hessian matrix) to compute the search
direction. However, sometimes it is impossible or timeconsuming to compute the
Hessian matrix.
QuasiNewton methods are based on the idea of gradually building up an
approximate Hessian matrix using the gradient information collected from previous
iterations. The initial Hessian approximation can be an identity matrix. In subsequent
iterations, the Hessian approximation is updated by using the gradient information to
mimic the behavior of the true Hessian matrix. QuasiNewton methods are attractive for
CHAPTER 2. LITERATURE REVIEW 12
practical use because their memory and computational requirements are low while they
have good convergence rate.
Constrained optimization methods. The optimum of a constrained optimization
problem is characterized by a certain set of conditions. These conditions were first
established by Kuhn and Tucker (1951) and they are often called the KuhnTucker
conditions or optimality conditions. Constrained optimization methods are developed to
find a point that satisfies these conditions. The major constrained optimization methods
include sequential quadratic and linear programming methods, reducedgradient methods,
and methods based on augmented Lagrangians, penalty, and barrier functions (Gill et al.,
1981).
The sequential quadratic programming (SQP) methods are widely regarded (Murray,
1997) as the most effective methods for nonlinearly constrained programming (NCP)
problem. The SQP methods have a structure of major and minor iterations. Each major
iteration involves formulating a quadratic programming (QP) subproblem to obtain the
search direction
k
p and using a line search procedure to obtain the steplength
k
α . The
QP subproblem is formulated in such a way that its objective function is a quadratic
approximation to the Lagrangian function of the original problem and its constraint
functions are the linearization of the original nonlinear constraints around the current
iterate. The minor iterations are used to solve a particular QP subproblem.
The sequential linearly constrained (SLC) methods (Murtagh and Saunders, 1982)
are another class of methods widely used for NCP problems. Similar to the SQP methods,
SLC methods also involve major and minor iterations. In contrast to SQP methods, which
formulate a QP subproblem in each major iteration, the SLC methods formulate a linearly
constrained subproblem whose objective function is a general approximation to the
Lagrangian function of the original problem. In general, an SLC method tends to require
fewer major iterations but more evaluations of the problem functions than an SQP
method (Murray, 2000).
The barrier methods solve a nonlinearly constrained problem by solving a sequence
of unconstrained optimization problems. A barrier method starts with a feasible point and
creates a sequence of unconstrained problems whose successive minimum stays feasible
CHAPTER 2. LITERATURE REVIEW 13
and converges to a minimum of the constrained problem. To achieve this, the objective
function of the unconstrained problem is constructed by adding a barrier function to the
original objective function of the constrained problem. The barrier function is a
modification of the constraint functions and becomes infinitely large when the iterate
approaches the boundary of the feasible region. The barrier method was first established
in the 1960s for general nonlinearly constrained optimization problems (Fiacco and
McCormick, 1968). Recently the method has been developed further to address
optimization problems whose objective and constraint functions are convex functions
(Boyd and Vandenberghe, 2001).
Separable programming is a special class of nonlinearly constrained optimization
problems whose objective and constraint functions are sums of functions of one variable.
Separable programming problems are usually solved by linear programming techniques
(Hillier and Lieberman, 2001).
Direct optimization methods refer to optimization methods that do not require
derivatives. When selecting an optimization method, a general rule is to choose a method
utilizing as much derivative information as possible (Gill et al., 1981). However, when
problem functions are not smooth or the derivatives are too expensive to compute, one
may choose a direct optimization method. While many direct optimization methods have
been proposed, here we only review a few that are popular in petroleum engineering
applications. A thorough review of the direct optimization methods is given by Powell
(1998).
The polytope method was first established by Spendley et al. (1962). The method
involves manipulation of 1 + n linearly independent data points, which can be viewed as
the vertices of a polytope, for a problem with dimension n . At each iteration, a new
polytope is generated by replacing the worst data point with a newly generated data point.
The data point is evaluated based on its objective function value. While the polytope
method is the most used method for unconstrained optimization in practice, some severe
cases of failure have been observed (Powell, 1998).
Genetic algorithms (GAs) are heuristic optimization algorithms introduced by
Holland (1975). GAs employ the idea of natural selection and genetics in the process of
CHAPTER 2. LITERATURE REVIEW 14
searching for the global optimum of a problem. In GAs, possible solutions are encoded as
chromosomes and modified by means of selection, crossover, and mutation. GAs have
several advantages over other optimization algorithms. Firstly, GAs are very general
optimization methods in that they handle both the discrete and continuous variables
naturally and require no domain knowledge of the optimization problem. Secondly, GAs
are stochastic search methods capable of avoiding local optima. Nevertheless, GAs have
certain disadvantages. First of all, convergence can be slow, especially at the later stages
of optimization. Secondly, GAs have been used primarily for unconstrained problems.
Although nonlinear constraints can be handled through penalty functions (Powell and
Sholnick, 1993; Michalewicz, 1996; Rasheed, 1998), the performance of such treatments
has not been impressive.
2.2 Applications of Optimization Techniques to
Petroleum Fields
Optimization techniques have been applied to virtually all aspects of the oil industry. In
this section, we review applications of optimization techniques in rate allocation,
production system design and operations, and reservoir development and management.
2.2.1 Lift Gas and Production Rate Allocation
Lift gas allocation. The mechanism of gaslift is elaborate. An appropriate amount of lift
gas increases the oil rate, while excessive lift gas injection will reduce the oil rate. To
determine the optimal lift gas rate, the usual practice is to allocate the lift gas to a well
according to a gaslift performance curve (Nishikiori et al., 1989). A gaslift performance
curve is a plot of oil rate versus lift gas rate for a gaslift well. When the gas supply is
unlimited, the optimal lift gas rate is the one corresponding to the maximum oil rate on
the performance curve. When the gas supply is limited, the lift gas is usually allocated
using some optimization algorithm. The earliest lift gas allocation method is a simple
heuristic method based on the concept of equalslope, which states that at the optimal
solution, the slope of the gaslift performance curves should be equal for all wells (Hong,
1975; Kanu et al., 1981). In addition to the equalslope method, Nishikiori et al. (1989)
CHAPTER 2. LITERATURE REVIEW 15
also applied a formal optimization algorithm, a QuasiNewton method, to the lift gas
allocation problem. Both the equalslope method and the QuasiNewton method rely on
derivative information to verify optimality, and therefore tend to get trapped in local
optima. This limitation can be serious when some gaslift performance curves are not
concave. Buitrago et al. (1996) addressed this problem by proposing a stochastic
algorithm that uses a heuristic method to calculate the descent direction. However, as
suggested by their results, their method is not good at handling constraints.
Gaslift optimization based on performance curves is simple and easy to implement.
However, this approach ignores flow interactions among wells in the optimization
process. DuttaRoy and Kattapuram (1997) analyzed a gaslift optimization problem with
two wells sharing a common flow line, and pointed out that when flow interactions are
significant, nonlinear optimization tools are needed to obtain satisfactory results. Dutta
Roy and Kattapuram (1997) applied a SQP method to a linearly constrained gaslift
optimization problem with 13 wells. Results showed that the SQP method does perform
better than methods based on performance curves.
General rate allocation. This problem refers to allocating production rates and lift
gas rate of single wells or the total production rates of reservoirs (in multireservoir cases)
to achieve certain operational goals. Linear programming seems to be the most popular
optimization algorithms for this kind of problems. Attra et al. (1961) applied linear
programming to maximize daily income from a multireservoir case subject to well
producing capacities, reservoir injection requirements, compressor capacity limits, gas
lift requirements, and sale contracts. Lo and Holden (1992) proposed a linear
programming model to maximize daily oil rate by allocating well rates subject to multiple
flow rate constraints. Fang and Lo (1996) applied LP to allocate both lift gas and
production rates. They first approximate gaslift performance curves by a piecewise
linear curve, and then formulate the rate allocation problem as an LP problem. The LP
method is very efficient for the rate allocation problem, however, suffers from the fact
that the objective and constraint functions in an LP problem have to be linear. Due to this
limitation, the gaslift performance curves in Fang and Lo’s method have to be concave.
CHAPTER 2. LITERATURE REVIEW 16
Options in commercial simulators. In some commercial reservoir simulators
(GeoQuest, 2000; Landmark, 2001), rate constraints on facilities are handled sequentially
by ad hoc rules. For example, first, gas constraints are considered at the lowest predictive
well management (PWM) level. Ad hoc rules based on the gasoil ratio of production
wells are applied to scale well rates and meet the gas constraint. Then, water, liquid, and
oil rate constraints at the lowest PWM level are handled sequentially and the procedure is
repeated for the next PWM level, etc. Furthermore, gaslift optimization is usually
considered separately from well rate optimization using, for example, the following
approach. First, liftgas rates are determined for specified wells based on their gaslift
performance curves or other information. Then, these liftgas rates are scaled using some
rules to match constraints on the total amount of available gas. A number of such
procedures are available in commercial reservoir simulators. In some cases, these
procedures may not yield the optimal solution. Also, using such approaches, it is difficult
to determine how much of the gas handling capacity should be used for gaslift and what
portion should be allocated for other purposes.
Field cases. Rate allocation problems are encountered frequently in mature fields
where production facilities cannot meet the field demand. For example, oil production in
the Prudhoe Bay oil field in Alaska is constrained by the gas handling capacity of the
separation units and the central gas plant. To utilize existing facilities fully, Barnes et al.
(1990) developed the Western Production Optimization Model (WPOM) for the Prudhoe
Bay field. This model allocates the oil rate and gas rate to surface facilities and wells in a
topbottom manner based on the “incremental GOR” concept that produces the next
incremental barrel of oil with the lowest GOR. Litvak et al. (1997) built an integrated
reservoir and gathering system model of the Prudhoe Bay field, and employed some
heuristic methods to optimize well connections to manifolds. Gaslift rates were allocated
based on gaslift tables of gasliquid ratio, liquid well rate and water cut. The
optimization algorithms in these applications are heuristic and they may lead to
suboptimal solutions.
The oil production in Kuparuk River oil field in Alaska is also constrained by the gas
processing capacities of surface facilities. Stoisits et al. (1992, 1994) applied a neural
network model to determine the optimal allocation of lift gas to wells subject to multiple
CHAPTER 2. LITERATURE REVIEW 17
gas constraints. The neural network was trained by results generated from single well
simulation. Stoisits et al. (1999) addressed the production optimization problem at
Kuparuk River field using a genetic algorithm. In their study, the decision variables were
the status of the production wells (close or open) and incremental GOR for drill sites. A
neural network model was developed to simulate the pressure drop through the surface
pipeline network. In all these applications (Stoisits et al., 1992, 1994, 1999) the neural
network is the central tool for speeding up the simulation/optimization process. However,
it is timeconsuming to train the neural network. Also, the accuracy of the neural
network, which depends on the accuracy of the training tools, the behavior of the
simulated system, and many other factors, is hard to control.
In summary, the general rate allocation problem has not been fully resolved yet,
especially when dealing with gigantic fields like the Prudhoe Bay and fields where flow
interactions cannot be ignored.
2.2.2 Optimization of Production System Design and Operations
Traditionally, optimization of production system design and operation in a petroleum
field has been performed by nodal analysis combined with trial and error (Beggs, 1991).
For example, by holding all other parameters fixed, a single variable is varied to see
which value of this variable gives the optimal objective function value. For multiple
decision variables, such a procedure fails because of the large number of function
evaluations required to cover the search space.
Carroll and Horne (1992) identified the need for formal optimization algorithms in
such applications. Carroll and Horne applied a Newtontype optimization algorithm and a
polytope method to optimize the tubing diameter and separator pressure of a singlewell
system and successfully demonstrated the usefulness of multivariate optimization
techniques. Ravindran (1992) followed the work of Carroll and Horne and allowed the
decision variables to vary with time. Fujii and Horne (1995) applied a genetic algorithm
to optimize the design and operations of a multiwell production system. Fujii and Horne
noted that multiple solutions may exist for the multiphase network flow problem.
However, no procedures were developed to select the legitimate solution. Palke and
CHAPTER 2. LITERATURE REVIEW 18
Horne (1997a) followed a similar line and applied a Newtontype algorithm, a polytope
method, and a genetic algorithm to a compositional model of a singlewell system. The
control variables included tubing diameter, separator pressure, and volume of gas
injected. A common observation of all these studies was as follows. When the tubing
diameter is involved, the surface of the objective function is highly irregular and
gradientbased algorithms often fail to solve the optimization problem. Palke and Horne
(1997a) concluded that genetic algorithms are the most robust algorithms for such cases.
NETOPT (SIMSCI, 1999) is a commercial network optimizer that combines a
general multiphase network simulator with a SQP method. NETOPT is a very general
tool, and can be interfaced with Eclipse (GeoQuest, 2000), a commercial reservoir
simulator, and be used to optimize various production design and operational problems.
Barua et al. (1997) discussed the general usage of NETOPT in the optimization of
production operations. DuttaRoy et al. (1997) applied NETOPT to analyze compressor
installation cost and operating strategy required to meet the lifetime production goals of a
gasfield. Heiba et al. (1997) applied NETOPT to devise optimal strategies for a cyclic
steam stimulation project. Because NETOPT employs SQP, a derivativebased
optimization algorithm, it locates only local optima and accepts only continuous decision
variables. The performance on individual optimization problems is not guaranteed and
care has to be taken to ensure reasonable results.
Zhang and Zhu (1996) minimized the total cost of a gas distribution network by
adjusting the diameters of pipelines of a fixed network layout. In their study, Zhang and
Zhu treated the optimization problem as a bilevel programming problem, and then
simplified the lower level and upper level problem by exploring their special properties.
2.2.3 Optimization of Reservoir Development and Planning
Reservoir development studies, such as drilling scheduling, well placement, and
production rate scheduling, have been an active area for optimization. The optimization
tools and simulation models employed in reservoir development studies have been
evolving.
CHAPTER 2. LITERATURE REVIEW 19
Early optimization studies of reservoir development featured simple reservoir models
and linear programming techniques. Aronofsky and Lee (1958) built a linear
programming model to maximize profit by scheduling production from multiple
homogeneous reservoirs. Bohannon (1970) presented a linear programming model to find
the optimum 15year development plan for a multireservoir pipeline system. The main
variables considered in their formulation were annual production rate from each
reservoir, number of wells to be drilled each year, and timing of major capital
investments such as initiating secondary recovery projects and expanding pipeline
facilities. Xiao et al. (1998) proposed a multiobjective linear programming model for an
oilfield injection recovery system. The main goal was to produce most oil with least
investment under the limitations of the available natural resource, equipment, and
manpower, etc. Aronofsky (1983) reviewed the application of linear programming in oil
and gas development.
Linear programming requires the objective and constraint functions to be linear. In
order to optimize the production system by linear programming, the nonlinear production
system has to be approximated by a system of linear functions. This greatly limits the
application of sophisticated reservoir simulators in the optimization of reservoir
developments. As the computing power increases and simulation techniques advance,
researchers are attempting to optimize reservoir development with more and more
complex reservoir models and sophisticated optimization techniques.
Various gradientbased optimization methods have been investigated. McFarland et
al. (1984) applied a generalized reduced gradient nonlinear programming method to
maximize present value of profits from a reservoir by deciding how many wells to drill in
each time period, the production rates, abandonment time, and platform size. They used
tanktype reservoir models to describe reservoir dynamics. Lasdon et al. (1986) studied
the problem of determining the optimum flows from a set of existing wells in each of
several time periods subject to certain demand schedules and operating constraints, and
combined numerical optimization techniques with a single phase, twodimensional
reservoir simulator. Zakirov et al. (1996) presented a method for optimizing net present
value of oil production with respect to well production rates as functions of time over the
whole life of the field. What distinguishes Zakirov et al.'s method from previous literature
CHAPTER 2. LITERATURE REVIEW 20
is that they used optimal control theory to obtain the gradients of the objective function in
an efficient way.
Most reservoir development problems have irregular surfaces for the objective
function. Therefore, stochastic optimization algorithms such as genetic algorithms have
gained popularity in such applications. Palke and Horne (1997b) used a genetic algorithm
to estimate the cost of uncertainty in reservoir data. In general, genetic algorithms are
slow. To speed up the search process, Bittencourt and Horne (1997) hybridized a genetic
algorithm with a polytope method, and applied the hybridized algorithm to a reservoir
development problem in which the decision variables are well locations and the objective
function is based on an economic model. Güyagüler et al. (2000) and Güyagüler and
Horne (2001) also looked at well placement problems by adapting the hybridized genetic
algorithm from Bittencourt and Horne (1997). In addition, they used a kriging method as
a proxy for the function evaluations to reduce the time for function evaluations. In the
optimization of reservoir development, a major concern is that we never possess true and
complete information about the reservoir. Güyagüler (2002) used the utility theory to
quantify the uncertainty in reservoir developments. Yeten et al. (2002) investigated the
problem of placing a multilateral well. The decision variables include the location of the
well, the number of laterals and the trajectory of the laterals. Yeten et al. (2002)
employed a genetic algorithm as the optimization driver and developed a hill climber and
a neural network model to speed up the optimization process. Results demonstrated that
the appropriate combinations of these tools can greatly speed up the optimization process.
2.3 Concluding Remarks
Optimization techniques have been applied to many aspects of the oil industry. Due to the
restrictions in computing power and simulation technologies, early applications favored
simple methods such as linear programming and heuristic methods. With the increase of
computing power and advances in simulation techniques, researchers and practitioners
are attempting to optimize reservoir development and operations with more and more
complex models and sophisticated optimization techniques. Frequently, the optimization
problems in the oil industry do not have smooth objective functions. Therefore, stochastic
CHAPTER 2. LITERATURE REVIEW 21
optimization algorithms such as genetic algorithms have been the major optimization
drivers in recent applications. Because reservoir simulation is a computationally intensive
process and the true reservoir information is unknown, it is worthwhile to investigate
ways of speeding up the simulation processes and combining uncertainty with
optimization.
The type of production optimization problem addressed in this study has not been
fully resolved. While extensive studies have been reported for unconstrained or linearly
constrained gaslift optimization problems (Kanu et al., 1981; Nishikiori et al., 1989;
Buitrago et al., 1996; DuttaRoy and Kattapuram, 1997), only few have been reported for
optimizing both the production rate and gaslift rate (Fang and Lo, 1996; Stoisits et al.,
1999). The method proposed by Fang and Lo (1996) ignores flow interactions among
wells and implicitly assumes the gaslift performance curves are concave and the well
water cut and the gas oil ratio (GOR) are constant with varying oil rates. The method
proposed by Stoisits et al. (1999) determines the optimum well shut down list and a drill
site IGOR (incremental gas oil ratio) list, not the production rates and lift gas rate for
individual wells. Further, no previous publication has addressed the problem of
optimizing production rates, lift gas rates, and well connections simultaneously subject to
nonlinear constraints. Therefore, it is necessary to explore new approaches to solve this
important class of production optimization problems.
This research extended the method proposed by Fang and Lo (1996) for the rate
allocation problem to cases where the gaslift performance curves can have any shape
and the well water cut and GOR can vary with oil rate (Chapter 4). A SQP method was
also applied to the rate allocation problem to account for flow interactions (Chapter 5). A
twolevel programming approach was developed to optimize all control variables
simultaneously (Chapter 6). Multiobjective optimization was introduced to handle
conflicting operational goals in a petroleum field (Chapter 7). Finally many of the
optimization approaches developed in this study were integrated with a reservoir
simulator so that the impact of the reservoir conditions could be taken into account in the
optimization process (Chapter 8). Some examples are presented throughout the
dissertation to demonstrate the advantages of the methods.
22
Chapter 3
Simulation of Gathering Systems
3.1 Introduction
A typical oil field contains a gathering system, a fluid distribution network, and an
injection network. The gathering system collects the fluids from production wells and
delivers them into separation units. The separated fluids are then distributed to different
destinations for storage, sale, disposal, injection, or further processing. The injection
network is used to inject fluids into the reservoir for enhanced oil recovery projects or for
fluid disposal/storage. The multiphase flow problems in the gathering, distribution and
injection network belong to network flow problems that have special properties. For the
production optimization problem considered in this work, a model is required to simulate
the multiphase flow in the gathering system. The distribution and injection networks are
ignored.
The network problem with singlephase flow is relatively easy to solve and has been
studied extensively in both the civil engineering and the petroleum engineering literature.
There are two major solution approaches. Both approaches formulate the network
problem according to the requirements of mass balance and Kirchoff’s law. The first
approach is to formulate the network problem as a set of system equations and solve the
system equations by the NewtonRapshon method or one of its variations (Donachie,
1973; Epp and Fowler, 1970; Liu, 1969). The second approach is to formulate the
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 23
network problem as an optimization problem and solve it using appropriate optimization
algorithms (Hall, 1976; Collins et al. 1978).
The multiphase network problem is also formulated according to the requirement of
mass balance and Kirchoff’s law. However, the multiphase network problem is much
more difficult than the single phase network problem for the following reasons: Firstly,
the system equations are more difficult to construct, because we need to determine the
fluid compositions in a link before the pressure loss along the link can be evaluated. This
is not easy in complex systems where the flow directions are unknown and phase splitting
ratios at junctions are hard to determine. Secondly, the NewtonRaphson method may not
converge or converge slowly due to the complex nature of the multiphase flow. Finally,
mathematically, multiple solutions exist for the multiphase network problem and not all
of them are physical. How to find the true solutions systematically is an unresolved
problem. To date, only a few researchers have reported solution methods for multiphase
network problems. Schiozer (1994) reported a solution method with pressure continuity
requirements for the system equations for networks with treelike structures. However,
Byer (2000) showed that Schiozer’s method can result in convergence problems due to
incorrect Jacobian terms. Byer (2000) proposed an alternative formulation with mass
balance requirements as the system equations for gathering systems with treelike
structures. Results show that even for a simple system with two wells, the solution
method can take more than ten Newton iterations to converge unless a good initial guess
is available. Fujii (1993) formulated the multiphase flow problem in a treelike structure
as an optimization problem and solved the problem using a GaussNewton method.
Several commercial software packages such as Eclipse (GeoQuest, 2000), VIP
(Landmark, 2001), and PIPEPHASE (SIMSCI, 2001) have options for the simulation of
multiphase network problems, however, their solution algorithms are not in the public
domain.
In this study, we consider gathering systems with treelike structures. The gathering
system can include such components as wells, tubing strings, chokes, and pipes. The
fluids in the system can include oil, gas, and water phases. Fluid properties are
represented by the blackoil model. This chapter first describes simulation models for the
individual components. The study investigated efficient procedures to solve the network
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 24
problem and to compute derivatives that are required in history matching, optimization,
and sensitivity analysis. An example is presented to demonstrate the performance of the
solution procedures that were found most suitable.
3.2 Well Model
In reservoir simulation, a well model is used to relate the flow rate of a well with its
bottomhole pressure and the reservoir pressure in the vicinity of the well (such as the grid
blocks in which the well is completed). The well model accounts for the geometric
characteristics of the well, the reservoir properties in the vicinity of the well, and the flow
behavior in the wellbore.
This study considered a vertical well completed in a single grid block and used a
basic well model proposed by Peaceman (1978). This model assumes steadystate radial
flow and can be expressed as
( )
∑
− =
p
w
p p c
w
c
p p WI q
,
λ (3.1)
where
w
c
q denotes the flow rate of component c , WI denotes the well index,
p c,
λ
represents the well mobility for component c in phase p ,
p
p denotes the reservoir
pressure of phase p , and
w
p denotes the well pressure. Expressions for WI and
p c,
λ can
be found in Aziz and Durlofsky (2002).
3.3 Pipe Model
In this work a homogenous model with slip was used to model multiphase pipe flow. In
the homogeneous model, different phases are lumped into one single pseudophase so that
pressure drop equations for singlephase flow can be utilized.
For the pseudophase, mixture density
m
ρ is based on in situ phase fractions:
g g l l m
E E ρ ρ ρ + = (3.2)
where
l
E and
g
E are the liquid and gas phase in situ fractions, respectively, and
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 25
1 = +
g l
E E (3.3)
Computation of gas volume fraction
g
E is presented in Section 3.3.2. Ouyang (1998)
suggested that the mixture velocity should be defined as
sg
m
g
sl
m
l
m
U U U
ρ
ρ
ρ
ρ
+ = (3.4)
where
sl
U and
sg
U are the superficial velocities of liquid and gas phases, respectively.
With these properties defined, we can proceed to compute the pressure loss for
multiphase flow in pipes.
3.3.1 Pressure Gradient Equation
Given a flow rate and the pressure at one end of a pipe, the pressure loss (or gain) in the
pipe can be computed by integrating the pressure gradient along the pipe. Consider
gas/liquid twophase flow in a pipe without perforations, the total pressure gradients
along the pipe can be decomposed into three parts
a g f
dx
dp
dx
dp
dx
dp
dx
dp

.

\

+

.

\

+

.

\

= (3.5)
where
•
f
dx
dp

.

\

is the pressure gradient due to wall friction,
•
g
dx
dp

.

\

is the pressure gradient due to gravity, and
•
a
dx
dp

.

\

is the pressure gradient due to acceleration.
Based on mass balance and momentum balance, it can be shown that (Brill and
Beggs, 1998)
θ ρ sin
m
g
g
dx
dp
− =

.

\

(3.6)
A
S
dx
dp
w
f
τ
− =

.

\

(3.7)
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 26
where θ is the pipe inclination angle,
w
τ is the wall friction shear stress, S is the pipe
perimeter, and A is the pipe area. The wall friction shear stress is computed as
2
2
1
m m m w
U f ρ τ = (3.8)
where the Fanning friction factor
m
f can be evaluated by the Colebrook equation
(Colebrook, 1939) or any of its approximation forms based on the relative pipe roughness
d
ε
and the mixture Reynolds number
m
m m
em
d U
R
µ
ρ
= (3.9)
where d is the pipe diameter and
m
µ is the mixture viscosity, which can be computed
from phase viscosities by (Brill and Beggs, 1998)
g g l l m
E E µ µ µ + = (3.10)
The acceleration pressure loss across a pipe segment depends on the difference
between the velocity heads of the flowing mixture at the two ends of the segment.
a
dx
dp

.

\

is usually small (Brill and Beggs, 1998) and it is ignored in this work.
3.3.2 Liquid Holdup Evaluation
A drift flux model (DFM) (GeoQuest, 2000; Chen, 2001) was used in this work to
calculate the liquid holdup. The DFM is used because of its continuity. Unlike
mechanistic models (Aziz and Petalas, 1994; Duns and Ros, 1963; Taitel and Dukler,
1976; Xiao et al., 1990) or correlations (Beggs and Brill, 1973), the DFM covers the
complete range of flowing conditions without any first order discontinuities. This
property makes the DFM more suitable for simulation and optimization studies than
mechanistic models.
In general, in all but downward flow, the gas phase has a tendency to flow faster than
the liquid phase. The DFM correlates the in situ gas velocity
g
U with the superficial
mixture velocity
sm
U (
sg sl sm
U U U + = ) using two parameters
0
C and
d
U
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 27
d sm g
U U C U + =
0
(3.11)
where
0
C is a profile parameter or distribution parameter that accounts for the effect of
velocity and gas concentration profiles.
d
U is the drift velocity of the gas, which
accounts for the local velocity differences between the two phases. Both
0
C and
d
U are
functions of the gas volumetric fraction
g
E
( )
g
E C C
0 0
= (3.12)
( )
g d d
E U U = (3.13)
The specific expression of Eqs. 3.12 and 3.13 can be found in GeoQuest (2000).
The superficial gas velocity and in situ gas velocity are related by:
g g sg
U E U = (3.14)
Substituting Eqs. 3.11, 3.12, and 3.13 into Eq. 3.14 we obtain a nonlinear equation
with
g
E as the unknown:
( ) ( ) 0
0
= + − =
d sm g sg g
U U C E U E f (3.15)
Solving Eq. 3.15 we can obtain
g
E , thus the liquid holdup
l
E .
Finally we mention that in certain cases, we found that the NewtonRaphson method
(Press et al., 1992) fails to converge for Eq. 3.15. The Van WijingaardenDekkerBrent
method (or Brent’s method) (Press et al., 1992), a method that combines root bracketing,
bisection, and inverse quadratic interpolation, proved to be both efficient and robust for
solving Eq. 3.15.
3.3.3 Upwards TwoPhase Flow Behavior
Under certain conditions, for upwards twophase flow, at low flow rates, increasing the
flow rate decreases the pressure drop across a pipe. Beyond a certain critical flow rate,
increasing the flow rate increases the pressure drop across the pipe. This phenomenon is
demonstrated in Figure 3.1. Table 3.1 presents the problem data used to generate Figure
3.1.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 28
The phenomenon described can be explained as follows. As shown in Eq. 3.5, the
pressure drop across a pipe segment is dominated by the gravity force and friction force.
The lower the flow rate, the smaller the frictional force. However, at low flow rates,
liquid phase tends to holdup (be held back) in the pipe causing bigger pressure drop due
to larger mixture density. Hence, at low flow rates, the gravity force dominates. As the
flow rate increases, the gravity force decreases due to reduced liquid holdup, the friction
forces increases but not large enough to compensate fully for the gravity force reduction.
Therefore, the overall pressure drop decreases as the flow rate increases. Since the
increase of frictional force is proportional to the square of the mixture velocity, beyond a
critical point, the friction force starts to dominant and the overall pressure drop increases
as the flow rate increases. Consequently, the typical pressure drop of multiphase flow
across a pipe demonstrates the behavior shown in Figure 3.1. A consequence of this flow
behavior is that multiple solutions may exist for the multiphase network flow problem.
This is discussed further in Section 3.5.3.
Table 3.1: Slip model problem data.
Tubing Parameter Fluid Property Flow Conditions
Diameter 3.5 in. Oil Density at
Standard Condition
30° API Downstream
Pressure
1000 psia
Roughness 0.0001 in. Specific Gas
Gravity
0.8 Downstream
Temperature
110 °F
Vertical Depth 8400 ft Water Density at
Standard Condition
62.37
lbm/cft
Upstream
Temperature
160 °F
Gas Oil Ratio 1000
SCF/STB
Water Oil Ratio 0.6
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 29
3600
3650
3700
3750
3800
3850
3900
3950
0 1000 2000 3000 4000 5000
OIL RATE, STB/DAY
U
P
S
T
R
E
A
M
P
R
E
S
S
U
R
E
,
P
S
I
A
Figure 3.1: Upstream pressure in a vertical well tubing string with a fixed
downstream (wellhead) pressure for different flow rates.
3.4 Choke Model
A good choke model is essential for accurately modeling multiphase flow in the
gathering system. The multiphase choke model developed by Sachdeva et al. (1986) has a
smooth transition between critical and subcritical flow and was used in this work. The
assumptions made when building this model are:
• incompressible liquid phase,
• equal phase velocities at the choke throat,
• no mass transfer between phases,
• onedimensional flow,
• accelerational pressure drop term dominates and the friction term is ignored,
• no distinction between free gas and solution gas, and
• liquid properties are obtained from the weighted average properties of oil and water.
The Sachdeva model can be expressed in the following context: given an upstream
pressure
1
p and a downstream pressure
2
p , determine the liquid flow rate across the
choke.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 30
The first step of the Sachdeva model is to determine the boundary of critical and
subcritical flow. Given an upstream pressure
1
p , the critical ratio of downstream to
upstream pressure is determined by the following equation:
( ) ( )
( ) ( )
1
2
1
2 1
1
2 1
1
1 1
1
2
1
2 1
1 1
1
−





.

\

(
¸
(
¸
−
+
−
+ +
−
− −
+
−
=
k
k
l
g
l
g
l
c g
c
x
x n
n
x
x n
n
k
k
x
y x
k
k
y
ρ
ρ
ρ
ρ
ρ
ρ
(3.16)
v p
C C k / = (3.17)
( )
( )
p v
v p
C x C x
C C x
n
1 1
1
1
1
− +
−
+ = (3.18)
where
c
y is the critical ratio of downstream pressure to upstream pressure, k is the ratio
of specific heat,
p
C is the specific heat capacity at constant pressure,
v
C is the specific
heat capacity at constant volume, x is the mass fraction of gas, n is the polytropic
exponent of gas, and the subscripts 1 and 2 represent upstream and downstream
conditions, respectively. Eq. 3.16 is a nonlinear equation with
c
y as the unknown.
Brent’s method (Press et al., 1992) is suitable for solving Eq. 3.16.
In the field units, the liquid rate across the choke is given as
( )( )
( )
5 . 0
1
1
1
1
1 2
2 1
2
2
1
1
1 1 525 . 0





.

\

(
(
(
(
(
¸
(
¸
−


.

\

−
− −
=
−
k
y k x
y x
p
C
d C
q
g
k
k
l
m
m
d
l
ρ ρ
ρ (3.19)
and


.

\

=
c
y
p
p
y , max
1
2
(3.20)
1
1
1
/ 1
1
1
2
1
−
(
(
¸
(
¸
−
+ =
l
k
g
m
x
y
x
ρ ρ
ρ (3.21)
( ) ( )
w w w o o o s o g m
B f B f R f GLR C ρ ρ γ + × + − × =
− − 5 7
2
10 5 . 6 10 84 . 8 (3.22)
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 31
where d is the choke inside diameter in inches,
d
C is the discharge coefficient,
o
f and
w
f are the flowing fractions of oil and water, B is the formation volume factor, γ is the
specific gravity,
s
R is the solution gasoil ratio, and GLR is the gas liquid ratio. Fluid
properties used to calculate
2 m
C are evaluated at the downstream conditions.
Eq. 3.20 states that if the actual value of
1 2
/ p p y
actual
= is greater than
c
y , the flow
is subcritical, the actual
2
p is used as the downstream pressure in Eqs. 3.21 and 3.22; if
actual
y is less than
c
y , the flow is critical, the value of
1
p y
c
is used as the downstream
pressure in Eqs. 3.21 and 3.22.
When the flow across a choke is critical, the downstream pressure can not be
determined uniquely. Therefore, when simulating the network flow, we formulated the
system equations in such a way that we only need to face the following problem: given a
downstream pressure and a flow rate, compute the upstream pressure of the choke. This
problem has a unique solution and can be obtained using appropriate root finding
procedures. Computational experience showed that bisection (Press et al., 1992) and
Brent’s method (Press et al., 1992) work well but the NewtonRaphson often fails to
converge.
3.5 Network Simulation
We considered a gathering system with a forestlike structure with multiple production
trees. A production tree includes wells, links, and nodes. A well refers to a wellbore and
its surrounding reservoir conditions (grid blocks in which the well is completed). Links
refer to any device or facility across which pressure changes. A link can be a tubing
string, a choke, or a pipeline. A node represents a flow junction or the terminal point of a
link. The top of the production tree has a fixed pressure and usually represents an inlet to
a separator.
In this section, we first formulate the governing equations of the network flow
problem. Then we describe a solution procedure. Finally we discuss possible outcomes of
the solution procedure.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 32
Node 4
Node 1
(separator)
Node 2
(manifold)
Node 3
Node 5 Node 6
Node 7 Node 8
Well 1 Well 2
Choke
Tubing
Pipe
Figure 3.2: A simple gathering system.
Example problem. We will use an example problem to facilitate our discussion. In
the example problem, two gaslifted wells are connected to a common manifold, then
they are connected to a separator through a common pipeline. The separator is operating
at a fixed pressure. The flow rate of each well is controlled by a choke between the tubing
string and the manifold. The tubing strings, chokes, and pipelines are modeled as links.
The junctions of two or more links are modeled as nodes. Figure 3.2 illustrates the
gathering system of this problem.
3.5.1 Governing Equations
The multiphase flow in a gathering system is described by mass conservation and
Kirchoff’s law. The mass conservation requires that every node should have zero net flow
rate unless it is a boundary node. Kirchoff’s law requires that the pressure for a node
should be the same no matter from which path it is computed.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 33
We choose the oil rates of wells as the independent variables and demonstrate how to
construct the governing equations for multiphase flow in the system of this example
problem.
Given the oil rate of a well, we obtain the water and gas flow rates of that well using
appropriate well models. Because the gathering system has a treelike structure and the
fluid flows from the bottom up through the production tree, according to mass
conservation, the flow rate in every node can be determined by summing up all the flow
streams entering that node:
w g o p q q
w
i
j
w
j p
n
i p
, , ,
, ,
= =
∑
Ω ∈
(3.23)
where
n
i p
q
,
denotes the flow rate of phase p in node i ,
w
j p
q
,
denotes the flow rate of phase
p for well j (phase g includes both the formation gas and liftgas, superscript n and w
denote nodes and wells, respectively), and
w
i
Ω denotes the set of wells whose flow
streams enter node i .
Because of the treelike structure of the gathering system, there is a unique path
connecting the separator with any well. Suppose node j receives fluids only from well
j . We can calculate the pressure of node j from two different pressure traverse
calculations. The first calculation is from the reservoir to node j . Denote the pressure for
node j obtained from this calculation as
r
j
p . The second calculation is from the
separator (we assume the separator is operating at a fixed pressure) to node j along the
path connecting the separator and node j . Denote the pressure for node j obtained from
this calculation as
s
j
p . Node j is called the solution node for well j . We choose a
solution node for every well and perform the described pressure traverse calculations.
According to Kirchoff’s law, the feasible set of oil flow rates is governed by:
0 = −
s
j
r
j
p p ,
n
s
j Ω ∈ ,
w
n
s
n = Ω (3.24)
where
n
s
Ω denotes the set of solution nodes,
w
n denotes the number of production wells.
In principle, for a particular well, we can choose its solution node as any node that
receives fluids solely from that well. For instance, in the example problem (see Figure
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 34
3.2), we can choose either node 3, 5, or 7 as the solution node for well 1, node 4, 6, or 8
as the solution node for well 2. In practice, we always chose the wellbore bottomholes,
i.e., node 7 and 8, as the solution nodes. This choice has certain computational
advantages. For instance, when computing the pressure drop across a surface choke, we
always compute its upstream pressure based on its downstream pressure and the flow rate
across it. This guarantees a unique upstream pressure no matter the flow through the
choke is critical or subcritical.
Eq. 3.24 has
w
n independent variables and
w
n equations, forming a closed system of
equations that can be solved by the NewtonRaphson method.
3.5.2 The NewtonRaphson Method
Let the oil flow rate of each well,
i o
q
,
, be the primary variable, and define
( ) ( ) ( )
o o o
q p q p q f
s r
− = (3.25)
Then Eq. 3.24 can be solved by the NewtonRaphson method with the following iterative
scheme:
( )( ) ( )
v
o
v
o
v
o
v
o
q f q q q J − = −
+1
(3.26)
where the Jacobian matrix ( )
v
o
q J is calculated as
o
s
o
r
q
p
q
p
J
∂
∂
−
∂
∂
= =
s
J J +
r
(3.27)
where
r
J denotes
o
r
q
p
∂
∂
and
s
J denotes
o
s
q
p
∂
∂
− .
The iteration terminates when the following convergence criteria are satisfied:
q
1 v
o
v
o
ε ≤ −
∞
−
q q (3.28)
and
( )
f
v
o
ε ≤
∞
q f (3.29)
where
q
ε and
f
ε are two userspecified tolerances.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 35
The major effort required in using the NewtonRaphson method is the construction
of the Jacobian matrix J , which is the summation of matrices
r
J and
s
J . In the current
network model, matrix
r
J is a diagonal matrix because pressure
r
j
p is a function of only
one primary variable
j o
q
,
. The element of
r
J can be evaluated analytically. Matrix
s
J
can be a dense matrix. The pattern of its nonzero elements depends on the connectivity of
the gathering system. Matrix
s
J can be constructed efficiently by the procedure
described below.
First we need to define some terminology. A well or a node can have one active
output node, which is the adjacent downstream node connected to that well or node. A
node can have multiple input nodes or wells, which are the adjacent upstream nodes or
wells connected to that node. Nodes are arranged in levels. The nodes with fixed
pressures are assigned to the first level. For a node in level l , all its input nodes or wells
are in level 1 + l . The pressure of all nodes and their derivatives are computed as follows:
1. For level 1, all nodes are boundary nodes. The node pressures are specified by the
user and the pressure derivatives are zero.
2. For node j or well j in level 1 + l , suppose we know the pressure and pressure
derivatives of its output node, say node i , in level l , then the pressure of node j can
be expressed as
( )
ij
i j
p f p α , = (3.30)
where vector
ij
α denotes the parameters that affect the pressure drop between node i
and node j . For the purpose of constructing matrix
s
J ,
ij
α should include the total
oil, gas, and water flow rate through the link between node i and node j .
The pressure derivatives of node j can be computed as
∑
∂
∂
•
∂
∂
+
∂
∂
•
∂
∂
=
∂
∂
m k
ij
m
ij
m
j
k
i
i
j
k
j
x
p
x
p
p
p
x
p
α
α
, n k ,..., 1 = (3.31)
where
k
x is the kth primary variable (i.e., the oil rate of a well),
ij
m
α is the mth
parameter that affects the pressure drop between node i and node j .
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 36
Eq. 3.31 is the key for efficient evaluation of the required derivatives because of
the following reasons:
• Derivative
k
i
x
p
∂
∂
is known at this stage.
• Derivative
k
ij
m
x ∂
∂α
can often be computed analytically with trivial computational
time.
• Derivatives
i
j
p
p
∂
∂
and
ij
m
j
p
α ∂
∂
can be approximated by the finite difference method
or computed using some automatic differentiation techniques (Berz et al., 1996).
Given a computer program for computing a function, automatic differentiation
software reads the computer program and produces an augmented computer
program capable of computing both the function and the derivatives of the
function. In this study we used such an automatic differentiation software,
ADIFOR (Bischof et al., 1998), to generate an augmented computer program for
computing
i
j
p
p
∂
∂
and
ij
m
j
p
α ∂
∂
.
• Derivatives
i
j
p
p
∂
∂
and
ij
m
j
p
α ∂
∂
are time consuming to compute. However, once they
are evaluated, they can be used to compute the derivatives of node pressure
j
p
with respect to all primary variables.
3. Repeat Step 2 for all nodes or wells in level 1 + l .
4. Let 1 + = l l . Repeat Step 2 and 3 until we obtain the pressure and pressure
derivatives for all nodes of the gathering system.
3.5.3 Multiple Solutions
Eq. 3.24 does not always have a unique solution. We use a single well system to illustrate
possible situations and discuss the complications in finding the true solutions of a
gathering system with multiple wells.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 37
Consider a single well system with a fixed well head pressure. For a fixed flow rate,
we can calculate the bottomhole pressure from two pressure traverse calculations. The
first calculation is from the reservoir to the bottomhole. Denote the bottomhole pressure
from this calculation as
r
p . The second path is from the surface (well head) to the
bottomhole. Denote the bottomhole pressure from this calculation as
s
p . We define the
relationship between
r
p and the oil flow rate as the inflow performance curve. We define
the relationship between
s
p and the oil flow rate as the outflow performance curve.
Several possibilities exist when we overlay the inflow and outflow performance curves:
1. The two curves have no intersections. In this case the corresponding governing
equation has no feasible solutions. This is the case when the reservoir pressure is too
low to support the fluid column in the tubing string.
2. The two curves have a unique intersection and the intersection represents the actual
flow rate of the single well system.
3. The two curves have two intersections. As demonstrated in Section 3.1, under certain
conditions, for upwards multiphase flow, at low flow rates, the pressure drop across
the tubing string decreases with increasing flow rates. Beyond a critical point, the
pressure drop across the tubing string increases with increasing flow rates.
Consequently,
s
p first decreases then increases as the flow rate increases. As a result
of this the inflow and outflow performance curves may have two intersections (Figure
3.3). However, the region of decreasing
s
p with increasing flow rate is unstable and
can cause intermittent or noflow condition (Lea and Tighe, 1983). Thus, the second
intersection represents the stable solution while the first intersection does not.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 38
3500
3550
3600
3650
3700
3750
3800
3850
3900
3950
0 1000 2000 3000 4000 5000
OIL RATE, STB/DAY
W
E
L
L
S
A
N
D
F
A
C
E
P
R
E
S
S
U
R
E
,
P
S
I
A
Outflow Curve
Inflow Curve
Figure 3.3: Illustration of multiple solutions for a single well system.
For a general gathering system with multiple wells, finding the true solutions to the
multiphase flow problem is more complicated. Several fundamental questions regarding
this problem are:
1. Is there a feasible solution? The governing equations for the network problem, Eq.
3.24, may have no feasible solutions because of the following reasons:
• The production system may be operated with some unrealistic operational
settings. For instance, in an optimization run, a well may have a large lift gas rate
but a small well choke diameter that is unable to deliver the large gas rate.
• Eq. 3.24 is formulated with the assumption that fluids always flow up a
production tree, which is not necessarily true for a particular set of operational
settings. The NewtonRaphson method can fail to converge no matter whether the
network problem has feasible solutions or not. For these reasons, it is difficult to
judge the feasibility of the network problem when the NewtonRaphson method
fails to solve the problem. One possible approach to address this problem is to
start the NewtonRaphson method from multiple starting points. If all runs fail to
obtain a feasible solution, then there is a large chance that the network problem is
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 39
infeasible. In such cases, production settings may be adjusted and the solution
process restarted.
2. How to determine if a feasible solution is stable? For naturally flowing wells, the
flow is stable if the outflow performance curve has a positive slope (Lea and Tighe,
1983). Therefore, the following inequality is used to determine if a solution to Eq.
3.24 is stable:
0
,
≥
∂
∂
j o
s
j
q
p
,
w
n j ,..., 1 = (3.32)
A solution is stable if it satisfies the inequality given by Eq. 3.32; otherwise it is not
stable. For wells on gaslift, the stability of the flow depends on many factors, and
several criteria are available to determine if the flow is stable (Fitremann and
Vedrines, 1985; Asheim, 1988). To keep this study focused, this work used Eq. 3.32
as the stability criteria for all wells.
3. For a fixed operational setting, is it possible for a network problem to have only
feasible but no stable solutions? Is it possible for a network problem to have multiple
stable solutions? If the answers to these questions are positive, how do we identify
and handle these scenarios in simulation and optimization? To date, available
literature does not offer a complete and systematic solution to those questions.
Therefore, postanalysis is required to verify the validity of a network solution.
3.6 Sensitivity Coefficients
Sensitivity coefficients refer to the derivatives of system response with respect to some
system parameters. Suppose the multiphase flow in a gathering system is governed by
( ) α u, f =0 (3.33)
where u denote the system response, i.e. the flow rates from each well, and α denote
the system parameters, e.g., the pipe diameters. Then
α
u
∂
∂
are the sensitivity coefficients
of system response u with respect to system parameters α.
Sensitivity coefficients are required in applications such as optimization, model
tuning, history matching, and sensitivity analysis. Sensitivity analysis can assist the
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 40
design of a new gathering system and potentially be an effective way to identify
bottlenecks in an existing system.
3.6.1 Finite Difference Method
To obtain sensitivity coefficients
α
u
∂
∂
, a conventional method is to use a finite difference
approximation:
1. Solve ( ) 0 = α u, f to obtain u.
2. For all
i
α , repeat the following steps:
a. Let
i i i
α ε α ) 1 ( + = ′ , where
i
ε is a small perturbation.
b. Solve ( ) 0 = ′ ′ α , u f to obtain u′ .
c. Compute
i i i
α ε α
u u u − ′
=
∂
∂
.
The finite difference approximation is straightforward to implement, however, it is
not efficient. For every perturbed parameter
i
u′ , we need to solve an updated set of
system equations. In addition, the accuracy of the sensitivity coefficients depends on the
value of the perturbation
i
ε and care should be taken to ensure that
i
ε is appropriate (Gill
et al., 1981).
3.6.2 Jacobian Method
Landa (1997) presented an efficient procedure for computing the derivatives of dynamic
production data (well pressure, flow rate, and reservoir saturation) with respect to static
reservoir properties (permeability and porosity in grid blocks) for a reservoir history
matching problem. A brief description of this method goes as follows.
Consider a discrete reservoir system
( ) 0 α, , u , u f
1
=
+
∆t
k k
(3.34)
where
k
u are the system responses (well pressure and flow rates) at time step k , α are
the reservoir properties (permeability and porosity), and t ∆ is the time step length. In
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 41
order to obtain sensitivity coefficients
α
u
∂
∂
+1 k
, we impose a small perturbation α δ on the
reservoir properties and expand Eq. 3.34 into a Taylor’s series as follows:
( ) 0
2
1
1
= +
∂
∂
+
∂
∂
+
∂
∂
+
+ +
+
+
+
+
δ O
1 k
k
k
k
1 k
1 k
1 k
k
δα
α
f
δu
u
f
δu
u
f
f (3.35)
The first term in the lefthand size is zero. If we drop the second order term Eq. 3.35 can
be expressed as
α
f
α
u
u
f
α
u
u
f
1
∂
∂
−
∂
∂
∂
∂
− =
∂
∂
∂
∂
+ + +
+
+ 1 k k
k
k 1 k
k
1 k 1
(3.36)
The implications of Eq. 3.36 are that we need to construct and invert matrix
1
1
+
+
∂
∂
k
k
u
f
only
once to obtain the derivatives of
1 + k
u with respect to all parameters α.
The same idea can be used to compute the sensitivity coefficients for a gathering
system governed by Eq. 3.33. To obtain
α
u
∂
∂
, we add a small perturbation
i
δα to
parameter
i
α
( ) 0 α α u, u f = + + δ δ (3.37)
where ( )
T
i
0 ,..., 0 , , 0 ,..., 0 δα δ = α . By expanding Eq. 3.37 around α and ignoring the
second order term, we obtain
( ) 0 =
∂
∂
+
∂
∂
+
i
i
δα
α
f
δu
u
f
α u, f (3.38)
Note ( ) 0 = α u, f . The Jacobian matrix is defined by
u
f
J
∂
∂
= (3.39)
and
i i
α α ∂
∂
− =
∂
∂ f u
J (3.40)
Again we need to construct and invert matrix J only once to obtain the derivatives
of vector u with respect to the whole parameter vector α. For the network problem,
constructing the Jacobian matrix requires extensive computational time. Thus the
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 42
Jacobian method is more efficient than the finite difference method, which needs to solve
an updated set of system equations whenever a parameter
i
α is perturbed.
3.7 Examples
In this section, we use an example to demonstrate various aspects of the procedures
discussed in the previous sections for simulating the network flow and computing the
sensitivity coefficients.
This is a synthetic example with ten vertical wells. The reservoir grid block pressures
around these wells vary between 3400 psi to 4900 psi. All wells are connected to the
same platform through tubing strings, chokes with a fixed diameter of 48/64 inches, and
3.5 inches diameter surface pipelines of lengths between 60 to 2000 ft. The platform is
connected to a separator through a pipeline with a diameter of 5.5 inches and a length of
4000 ft. The separator is operated at a fixed pressure of 160 psi. The oil gravity is 31
°
API. The specific relative gravity of gas is 0.8. The fluid temperature in the surface
pipeline is assumed to be 110
°
F. The fluid temperature at the bottomholes is assumed to
be 160
°
F. The fluid temperature in the tubing strings is linearly interpolated between the
bottomhole fluid temperature and the surface fluid temperature. Figure 3.4 shows the
configuration of the gathering system. Appendix A.1.4 shows the input file for this
example.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 43
Wells
Tubing
strings
Chokes
Platform
Separator
1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 6 7 8 9 10
11 12 13 14 15 16 17 18 19 20
21
Pipes
Figure 3.4: Configuration of the gathering system.
3.7.1 Simulation Results
We made five runs with different initial guesses. Runs 1, 2, 3, and 4 started the solution
procedure with 20, 200, 1000, 2000 STB/d oil rate for every well, respectively. Run 5
started the solution procedure with a different flow rate for each well. We deliberately
started Run 1 with a bad initial guess so that the solution procedure would be difficult to
converge.
All runs were able to reach the convergence criteria. Figure 3.5 shows how the
infinity norm of the residuals of the governing equations, ( )
∞
v
o
q f , decreased at each
iteration. For Run 3 and Run 4, ( )
∞
v
o
q f dropped below 1 psi within 3 iterations. As
expected, Run 1 had the slowest convergence rate. However, even for Run 1, ( )
∞
v
o
q f
dropped below 1 psi within 6 iterations. Figure 3.6 and 3.7 show the oil flow rate of each
well at every iteration for Run 1 and Run 4, respectively. In both runs, the oil flow rate
for each well approached its converged value within 24 iterations. Therefore, we
concluded that the proposed solution procedure is efficient.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 44
1.E11
1.E09
1.E07
1.E05
1.E03
1.E01
1.E+01
1.E+03
1.E+05
0 2 4 6 8 10
ITERATION
run1
run2
run3
run4
run5
P
S
I
,
∞
f
Figure 3.5: Infinity norm of the governing equation residuals versus
iteration number.
0
500
1000
1500
2000
2500
3000
3500
4000
4500
0 2 4 6 8 10
ITERATION
Q
o
,
S
T
B
/
D
Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 7
Well 8
Well 9
Well 10
Figure 3.6: Oil rate versus iteration number for Run 1.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 45
0
500
1000
1500
2000
2500
3000
3500
0 1 2 3 4 5 6
ITERATION
Q
o
,
S
T
B
/
D
Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 7
Well 8
Well 9
Well 10
Figure 3.7: Oil rate versus iteration number for Run 4.
Multiple solutions exist for this problem. As shown in Table 3.2, the five runs
converge to three sets of solutions. Evaluated by Eq. 3.32, the solutions of Run 1 and Run
2 are unstable, because for some wells
j o
s
j
q p
,
/ ∂ ∂ is negative. The solutions of Runs 3, 4,
and 5 are stable, because for all wells
j o
s
j
q p
,
/ ∂ ∂ is positive.
Table 3.2: Network simulation results from different runs.
Run 1 Run 2 Run 3, Run 4, and Run 5
Well
Oil Rate
(STB/d)
i o
s
i
q p
,
/ ∂ ∂
(psi•d/STB)
Oil Rate
(STB/d)
i o
s
i
q p
,
/ ∂ ∂
(psi•d/STB)
Oil Rate
(STB/d)
i o
s
i
q p
,
/ ∂ ∂
(psi•d/STB)
1 855.21 0.93 781.48 0.95 735.29 0.97
2 68.45 4.95 1026.53 1.63 988.84 1.57
3 2809.52 0.68 2677.97 0.67 2590.56 0.66
4 1044.75 4.95 913.44 1.73 828.05 0.86
5 1212.72 2.86 1186.67 2.79 1168.77 2.75
6 63.76 2.68 101.54 1.73 1142.99 1.05
7 616.05 0.36 256.14 0.58 144.35 0.78
8 933.19 6.24 933.19 6.24 933.19 6.24
9 759.87 4.39 748.94 4.25 740.97 4.17
10 3354.15 0.37 3088.45 0.36 2910.98 0.36
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 46
3.7.2 Computation of Sensitivity Coefficients
The purpose of this section is to demonstrate the use of the Jacobian method for
computing the sensitivity coefficients. We used both the finite difference and the
Jacobian method to compute the sensitivity coefficients of the oil flow rate of every well
with respect to the gaslift rate of every well, the diameter of every pipe (including the
tubing strings), and the diameter of every choke. Remember there are 10 gaslifted wells,
21 pipes, and 10 chokes in the example problem.
To obtain all the sensitivity coefficients, the finite difference method needs to run 42
network simulations. In contrast, the Jacobian method requires only one network
simulation and a little extra computational time. For this problem, on a Silicon Graphics
Origin 200 computer with four 270 MHz CPUs and 2304MB of RAM, a typical network
simulation costs about 470 ms of CPU time. The extra computational time required by
the Jacobian method is about 40ms. Table 3.3 compares several sensitivity coefficients
obtained from both the Jacobian method and the finite difference method. The results
from the finite difference method depend on the finite difference interval. Table 3.3
suggests that a finite difference interval of between 0.01% and 0.001% of the parameter
value is appropriate for this problem. The results from the Jacobian method are in good
agreement with the results from the finite difference method.
Table 3.3: Comparison of sensitivity coefficients obtained from the Jacobian
method and the finite difference method.
Finite Difference Approximation
Sensitivity
Coefficients
Jacobian
Method =
i
ε 1% =
i
ε 0.1%
=
i
ε
0.01%
=
i
ε
0.001%
=
i
ε
0.0001%
1 , 1 ,
/
tubing o
d q ∂ ∂
(STB/d/inch)
1.0654E+1 7.6811E+0 1.0356E+1 1.0627E+1 1.0654E+1 1.0657E+1
1 , 1 ,
/
choke o
d q ∂ ∂ ,
(64STB/d/inch)
4.5943E+0 4.4996E+0 4.5848E+1 4.5935E+1 4.5943E+1 4.5944E1
1 lg, 1 ,
/ q q
o
∂ ∂ ,
(STB/MSCF)
1.0079E1 1.0002E1 1.0072E1 1.0079E1 1.0080E1 1.0080E1
1 , 6 ,
/
tubing o
d q ∂ ∂ ,
(STB/d/inch)
4.3001E1 3.1003E1 4.1798E1 4.2892E1 4.3003E1 4.3020E1
1 , 6 ,
/
choke o
d q ∂ ∂ ,
(64STB/d/inch)
1.8543E1 1.8162E1 1.8505E1 1.8540E1 1.8544E1 1.8545E1
1 lg, 6 ,
/ q q
o
∂ ∂ ,
(STB/MSCF)
1.0819E2 1.0790E2 1.0817E2 1.0820E2 1.0820E2 1.0820E2
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 47
Sensitivity coefficients can give great insight about the flow behavior in a gathering
system and can be used to identify bottlenecks of the system. Figure 3.8 plots the
sensitivity coefficients of the well oil rates with respect to the pipe diameters (as shown
in Figure 3.4, pipe 110 refers to the tubing strings for well 110, respectively; pipe 1121
are surface pipelines). Figure 3.8 shows that the oil flow rate of all wells except well 8
are sensitive to the diameter of pipe 21. The sensitivity coefficients of the well oil rates
with respect to the diameters of other surface pipelines and tubing strings are close to
zero. This suggests that pipe 21 is the bottleneck of the gathering system. If we increase
the diameter of pipe 21,
21
d , the sensitivity coefficients with respect to
21
d decrease and
the sensitivity coefficients with respect to the diameters of tubing strings increase. As
shown in Figure 3.9, when
21
d increases to 7.5 inches, the diameters of the tubing strings
and pipe 21 has roughly the same impact on the well oil rates. When
21
d increases to 9.5
inches (Figure 3.10), the sensitivity coefficients with respect to
21
d become zero. In all
scenarios, the sensitivity coefficients with respect to the diameters of pipes 1120 are
approximately zero. This suggests pipes 1112 are not properly designed and may have
surplus capacity.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 48
1
2
3
4
5
6
7
8
9
10
5
10
15
20
500
0
500
1000
1500
S
T
B
/
D
/
I
N
C
H
P
IP
E
N
U
M
B
E
R
W
ELL
NUM
BER
Figure 3.8: Sensitivity coefficients of well oil rates with respect to system
parameters (the inner diameter of pipe 21 is 5.5 inches).
1
2
3
4
5
6
7
8
9 10
5
10
15
20
100
0
100
200
300
400
500
600
700
800
S
T
B
/
D
/
I
N
C
H
PIPE
N
U
M
B
E
R
WELL
NUMBER
Figure 3.9: Sensitivity coefficients of well oil rates with respect to system
parameters (the inner diameter of pipe 21 is 7.5 inches).
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 49
1
2
3
4
5
6
7
8
9
10
5
10
15
20
200
0
200
400
600
800
S
T
B
/
D
/
I
N
C
H
P
IP
E
N
U
M
B
E
R
WELL
NUMBER
Figure 3.10: Sensitivity coefficients of well oil rates with respect to system
parameters (the inner diameter of pipe 21 is 9.5 inches).
3.8 Concluding Remarks
The drift flux slip model for the multiphase pipe flow and the Sachdeva model for the
multiphase choke flow were presented in this chapter. Computational experience showed
that the NewtonRaphson method is not suitable for the rootfinding problems
encountered in these models. Brent’s method (Press et al., 1992) proved to be both
efficient and robust.
The multiphase network problem was formulated according to the mass balance
requirement and Krichoff’s law and solved by the NewtonRaphson method. The
Jacobian matrix was computed efficiently by exploring the treelike structure of the
network. An efficient method, the Jacobian method, was used to compute the sensitivity
coefficients. Results from an example showed that the NewtonRaphson method
converges fast for our formulation, and the Jacobian method is more efficient than the
finite difference method for computing sensitivity coefficients.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 50
It is straightforward to extend the formulation and solution methods presented here
to compositional models. Further research is required for the following problems:
1. How to build an efficient and robust solution procedure for multiphase flow problems
of gathering systems with loops.
2. How to build a procedure capable of automatically finding the stable solutions of a
multiphase flow problem.
51
Chapter 4
Rate Allocation through Separable
Programming
4.1 Introduction
Rate allocation refers to the problem of optimally adjusting production rates and lift gas
rates of production wells to achieve certain operational goals. These goals vary with the
field and time. In some petroleum fields, especially mature fields, oil production can be
constrained by fluid handling capacities of facilities. For such fields, rate allocation can
be an effective way to increase the oil rate or reduce the production cost. For example, if
the oil production in a field is constrained by the gas processing capacity of the
separation units, closing or reducing production rates of wells with the highest GOR will
increase the total oil rate; in addition, reducing the lift gas rate of certain wells may
increase the overall oil production by utilizing the gas processing capacity more
efficiently.
This study addressed the following rate allocation problem. The objective function is
the total oil rate. An oil, gas, water, or liquid flow rate constraint can be put on any
production well or network node. In abstract form, the optimization problem can be
expressed as
maximize { }
∑
=
w
n
i
w
i o
q
1
,
(4.1a)
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 52
subject to
n
j p
n
j p
Q q
, ,
≤ ,
n
j Ω ∈ , { } l w g o p , , , ∈ (4.1b)
where
w
i o
q
,
is the oil rate of well i ,
n
j p
q
,
is the flow rate of phase p in node j ,
n
j p
Q
,
is the
flow rate limit of phase p for node j ,
n
Ω is the set of all nodes. For gathering systems
with a treelike structure, the flow rate of network node j is the sum of the flow rates of
wells connected to node j .
Physically, the control variables for the rate allocation problem are the well chokes
and lift gas rates of wells. And the flow rate of each well,
w
j p
q
,
, is a nonlinear function of
the control variables. Fang and Lo (1996) studied a similar rate allocation problem. To
simplify the optimization problem, Fang and Lo (implicitly or explicitly) made the
following assumptions:
1. The well performance information can be evaluated individually for each well by
ignoring flow interactions among wells.
2. The GOR and water cut for a well remain constant for varying oil rate.
3. The gaslift performance curves are concave.
With the above assumptions, Fang and Lo (1996) reformulated the rate allocation
problem to a linear programming problem and solved it by the simplex algorithm. The
method was found to be very efficient.
In this part of the work we followed the line of Fang and Lo (1996) and developed
other solution methods for the rate allocation problem. It was our aim to develop
techniques that speed up the entire optimization process.
In Chapter 5, we will present a method that does not rely on assumption 1.
4.2 Well Performance Estimation
The method presented in this chapter required the performance estimation of individual
wells. Specifically, for unlifted wells, the method requires (1) the maximum oil rate that
is allowed or can be delivered from a well, and (2) the water and gas rates as functions of
the oil rate of a well. For lifted wells, the method requires (1) the oil rate as a function of
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 53
the lift gas rate (the gaslift performance curve), and (2) the water and (formation) gas
rates as functions of the oil rate of the well.
The well performance information can be described by a set of oil rate versus water,
formation gas, and lift gas rate curves (Figure 4.1), which are referred to in this study as
the performance curves. How to construct the performance curves is an extensively
studied area (Beggs, 1991; Lea and Tighe, 1983). Two common approaches are: (1)
derive the performance curves from well test information, or (2) estimate the
performance curves through simulation. This study adopted the second approach.
0
500
1000
1500
2000
2500
3000
3500
0 100 200 300 400 500 600 700
Qo (STB/D)
Q
g
(
M
S
C
F
/
D
)
0
500
1000
1500
2000
2500
3000
3500
Q
w
(
S
T
B
/
D
)
formation gas rate vs oil rate
lift gas rate versus oil rate
water rate verus oil rate
Figure 4.1: Illustration of well performance curves.
The production system is an integrated system in which the flow streams from
different wells influence each other. To establish the performance curves for individual
wells, a well has to be isolated from the whole system. This is done in two steps:
1. Isolate the well from the surface pipeline network by fixing the pressure of a network
node before the flow stream from that well joins the flow streams from other wells.
2. Ignore well interactions in the reservoir by developing an inflow performance model
for every well.
A well inflow performance model relates the oil rate of a well to its bottomhole
pressure, and its water and gas rates. An appropriate way to build this model is to run a
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 54
simulated well test on a numerical model of the reservoir. A simulated well test mimics
well tests performed in a real field. The test calculates the bottomhole pressure, and water
and gas rate of a well for various oil rates using a numerical model. A simulated well test
can be time consuming. To generate performance curves efficiently, this study used the
well model presented in Section 3.1 and assumed that the reservoir condition in well
blocks are constant for varying oil rates.
After a well is isolated from the production system, its performance curves can be
established as follows:
1. The oil rate versus water and formation gas rate curves can be solely determined by
the well inflow performance model.
2. The gaslift performance curve for a well can be constructed by determining the oil
rate of the singlewell system for a set of prescribed lift gas rates.
The well performance curves are represented by a set of discrete points. We assume
linearity between two adjacent points, thus the performance curves are piecewise linear
curves (Figure 4.1).
4.3 Rate Allocation through Separable Programming
The objective and constraint functions of the rate allocation problem (Eq. 4.1) are linear
combinations of the oil, water, and formation and lift gas rates of individual wells.
Further, when the well performance is approximated by piecewise linear performance
curves, the water, formation gas, and lift gas rates of a well can be regarded as functions
of the oil rate of that well. Therefore, if we regard the oil rate as the control variables,
denoted as x , Eq. 4.1 becomes an optimization problem whose objective and constraint
functions are the sums of functions of one variable, which can be expressed as
maximize ( )
∑
=
w
n
j
j j
x f
1
(4.2a)
subject to ( )
∑
≤
j
i j ij
b x h , m i ,..., 1 = (4.2b)
0 ≥
j
x ,
w
n j ,..., 1 = (4.2c)
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 55
where
j
f denotes the objective functions,
ij
h denotes the j th function involved in the
i th constraint,
i
b denotes the limit of the i th constraint, and m denotes the number of
constraints.
Optimization problems of the form of Eq. 4.2 are separable piecewise linear
problems, which can be solved by linear optimization techniques (Gill et al., 1981). This
is demonstrated in the following.
First, a point ( ) ( ) x f x, on a piecewise linear curve defined by a set of discrete points
( ) ( )
i i
x f x , , r i ,..., 0 = can be expressed as follows:
∑
=
=
r
j
j j
x x
0
λ (4.3a)
( ) ( )
∑
=
=
r
j
j j
x f x f
0
λ (4.3b)
1
r
0 j
=
∑
=
j
λ , 0 ≥
j
λ (4.3c)
no more than two
j
λ can be positive and they must be adjacent (4.3d)
Suppose for well j , each of its well performance curves are defined by
j
r +1 discrete
points with 0
0
=
j
x and
max
, j o jr
q x
j
= , where
max
, j o
q is the maximum oil rate for well j .
Then for all functions ( )
j j
x f and ( )
j ij
x g we can write
( )
∑
=
=
j
r
k
jk jk j j
f x f
0
λ , ( )
jk j jk
x f f = (4.4a)
( )
∑
=
=
j
r
k
ijk jk j ij
g x g
0
λ , ( )
jk ij ijk
x g g = , m i ,..., 1 = (4.4b)
∑
=
=
j
r
k
jk jk j
x x
0
λ (4.4c)
∑
=
=
j
r
k
jk
0
1 λ , 0 ≥
jk
λ for all j , k (4.4d)
for a given j , no more than two
jk
λ can be positive and
they must be adjacent (4.4e)
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 56
Substituting Eqs. 4.4a4.4e into Eq. 4.2, we obtain Problem 4.5:
maximize
∑∑
= =
=
w j n
j
r
k
jk jk
f z
1 0
λ (4.5a)
subject to
∑∑
= =
≤
w j n
j
r
k
i jk ijk
b g
1 0
λ , m i ,..., 1 = (4.5b)
∑
=
=
j
r
k
jk
0
1 λ ,
w
n j ,.., 1 = (4.5c)
0 ≥
jk
λ for all j , k (4.5d)
for a given j , no more than two
jk
λ can
be positive and they must be adjacent (4.5e)
Without constraint Eq. 4.5e, we would have a general linear programming problem.
It can be shown that if all ( )
j j
x f are concave functions and
ijk
g are convex functions,
then, the optimal solution of an LP problem defined by Eqs. 4.5a4.5d automatically
satisfies constraint Eq. 4.5e (Hartley, 1961). However, in general cases, constraint Eq.
4.5e has to be enforced explicitly.
Along this line, there are several models for the rate allocation problem depending
on how constraint Eq. 4.5e are treated. These models were investigated in this study and
are discussed below.
4.3.1 Model LPI
This model was proposed by Fang and Lo (1996).
In this model, constraint Eq. 4.5e is simply ignored. The rate allocation problem is
reformulated to a linear programming problem defined by 4.5a4.5d. This model is able
to solve correctly the rate allocation problem defined by Eq. 4.2 if the water and gas
versus oil rate curves involved in the constraints are strictly convex. Otherwise, as
demonstrated in Section 4.4.3, this model may allocate an unphysical point that is not on
the performance curves.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 57
4.3.2 Model MILPI
This model was applied to a rate allocation problem by Güyagüler and Byer (2001). This
model enforces constraint Eq. 4.5d explicitly and is suitable for rate allocation problems
with performance curves of arbitrary shapes. The disadvantage of this method is that even
for a rate allocation problem with moderate size, this model can contain a large number
of binary variables and constraints.
To enforce the constraint that for certain j at most two consecutive coefficients
jk
λ
are nonzero, we introduce a binary variable
jk
y , 1 ,..., 0 − =
j
r k , which can be equal to 1
only if
) 1 ( +
≤ ≤
k j j
k
j
x x x and 0 otherwise. Problem 4.5 is then formulated as a mixed
integer linear programming (MILP) problem (Bertsimas and Tsitsiklis, 1997) to obtain
Problem 4.6:
maximize
∑∑
= =
=
w j n
j
r
k
jk jk
f z
1 0
λ (4.6a)
subject to
∑∑
= =
≤
w j n
j
r
k
i jk ijk
b g
1 0
λ , m i ,..., 1 = (4.6b)
∑
=
=
j
r
k
jk
0
1 λ ,
w
n j ,.., 1 = (4.6c)
0 0 j j
y ≤ λ , for all j (4.6d)
jk k j jk
y y + ≤
− ) 1 (
λ , 1 ,..., 1 − =
j
r k , for all j (4.6e)
( ) 1 −
≤
j j
r j jr
y λ , for all j (4.6f)
∑
−
=
=
1
0
1
j
r
k
jk
y , for all j (4.6g)
0 ≥
jk
λ for all j , k (4.6h)
{ } 1 , 0 ∈
jk
y , 1 ,..., 1 − =
j
r k , for all j . (4.6i)
Several similar reformulations exist and their computational performances can be
different (Padberg, 2000).
In principle, a MILP problem can be solved by enumeration. However, complete
enumeration is computationally infeasible as soon as the number of integer variables in a
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 58
MILP problem exceeds 20 or 30 (Wolsey, 1998). So we need some strategies to cut the
number of necessary enumerations. An effective method for this purpose is the Branch
and Bound method (Wolsey, 1998).
Branch and Bound is a general search method for optimization problems over a
search space that can be represented as leaves of a tree. A fundamental idea behind the
Branch and Bound method is to divide and conquer. Consider a general optimization
problem
{ } S x x c ∈ = : max
T
z (4.7)
where x is the control variable and S denotes its feasible set. Let
K
S S S ...
1
∪ = be a
decomposition of S into smaller sets, and let { }
k
T k
z S x x c ∈ = : max for . ,..., 1 K k =
Then
k
k
z z max = (Wolsey, 1998). Based on this idea, Branch and Bound recursively
divides an optimization problem into several subproblems, and thereby forms an
enumeration tree, with each node in the tree representing a subproblem. The enumeration
tree is constructed implicitly, and the bound information on subproblems are used to
prune the tree. A complete description of the general Branch and Bound method can be
found in Chapter 7 of Wolsey (1998).
Next we present a Branch and Bound method adapted for solving Problem 4.6.
Algorithm 4.1: Branch & Bound for Problem 4.6. The algorithm maintains a
lower bound z , an upper bound z ( z z z ≤ ≤ ), and a stack of active search nodes
representing subproblems to be examined.
1. Initialization. Let −∞ = z and ∞ = z . Put Problem 4.6 into the stack of active search
nodes.
2. Choosing a node. If the stack of active nodes is empty, the entire tree has been
enumerated and the search ends. If the stack of active nodes contains several nodes,
pop out the node on top of the stack. This node represents a MILP subproblem
generated in Step 1 or Step 6.
3. Optimizing. Solve the LP relaxation of the subproblem selected in Step 2. Denote its
optimal value as zˆ . The definition of LP relaxation is discussed shortly.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 59
4. Bounding. If the node selected in Step 2 is the root node, update z z ˆ = . If the solution
in Step 3 is a feasible solution to Problem 4.6, update { } z z z ˆ , max = and store the
corresponding feasible solution.
5. Pruning. The following conditions allow us to prune the tree and thus enumerate a
large number of solutions implicitly.
(a) Pruning by optimality. If
z
z z ε ≤ − , where
z
ε is a prescribed tolerance, then the
feasible solution corresponding to z can be regarded as the optimal solution of
Problem 4.6. And the search ends.
(b) Pruning by feasibility. The solution from Step 3 is a feasible solution to Problem
4.6. There is no need to divide further the subproblem represented by current
node.
(c) Pruning by bound: z z ≤ ˆ . The upper bound of the subproblem represented by
current searching node is below the lower bound. There is no need to divide
further the subproblem represented by current node.
(d) Pruning by infeasibility. The problem examined has no feasible solution. There is
no need to divide further the subproblem represented by current node.
If condition (a) is met, stop. If condition (b), (c) or (d) is met, go to Step 2 to
backtrack. Otherwise, go to Step 6 to branch.
6. Branching. To reach this step, the solution from Step 3 must not be a feasible solution
to Problem 4.6. Some binary variable
jk
y has a fractional value. Suppose variable
jk
y ,
w
n j ≤ ≤ 1 , { } 1 ,..., 0 − ∈
j
r k has a fractional value. The MILP subproblem from
Step 2 (the parent subproblem) can be divided into two subproblems. The first
subproblem is comprised of the parent subproblem plus a constraint of 0 =
jk
y . The
second subproblem is comprised of the parent subproblem plus a constraint of
1 =
jk
y . Add the two subproblems on top of the stack of active nodes in a prescribed
order (such as the first subproblem goes first, and the second subproblem goes
second).
7. Continue the search by going to Step 2.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 60
The purpose of Step 3 is to compute an upper bound for the MILP subproblem from
Step 2. The bound information is used in Step 5 to prune the enumeration tree. The upper
bound of a MILP subproblem is often obtained by solving a relaxed problem, a problem
that has an optimal value no worse than that of the MILP subproblem. One such
relaxation for a MILP problem is its linear programming relaxation that allows the
integer variables in a MILP problem to take real values. For example, an LP relaxation of
Problem 4.6 is an LP problem that replaces constraint Eq. 4.6i with the following linear
constraints
1 0 ≤ ≤
jk
y , 1 ,..., 1 − =
j
r k for all j (4.8)
4.3.3 Model MILPII
This model was developed in this study. This model is motivated by and is suitable for
problems with the following properties:
1. For wells that can flow naturally, their gaslift performance curves are concave
(Curve I in Figure 4.2).
2. For wells that need a finite amount of lift gas to start flowing, their gaslift
performance curves are comprised of a horizontal line and a concave curve (Curve II
in Figure 4.2).
3. The oil versus water and formation gas curves are concave.
With these properties, constraint Eq. 4.5e can be automatically satisfied if we require the
coefficient
0 j
λ in Problem 4.5 to be a binary variable for wells whose gaslift
performance curves is nonconcave (Curve II in Figure 4.2).
This model has the same number of variables and constraints as Model LPI.
However, some of the decision variables are binary variables. Thus this model leads to a
MILP problem. The Branch and Bound method described in Section 4.3.2 can be used to
solve it. An example of the application of this method is presented in Section 4.4.3.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 61
A
0
LIFT GAS RATE
O
I
L
R
A
T
E
Curve I
Curve II
Figure 4.2: Illustration of gaslift performance curves.
4.3.4 Model LPII
This model was first proposed by Lo and Holden (1992). This is a simple model that does
not require the performance curves and therefore there is no need to satisfy constraint Eq.
4.5e. The model assumes the well water cut and GOR are constant for varying oil rates.
This model takes a set of flow streams (either from production wells or from satellite
reservoirs) as the input and scales them to meet the flow rate and velocity constraints in a
way that maximizes an objective function. A flow stream is represented by the GOR,
water cut, and a maximum oil rate of a well or a satellite reservoir. For example, suppose
we want to maximize the total oil rate of a field subject to a total gas rate constraint. The
problem can be formulated as
maximize
∑
=
w
n
i
i o i
q x
1
max
,
(4.9a)
subject to
∑
=
≤
w
n
i
g i i o i
Q GOR q x
1
max
,
(4.9b)
1 0 ≤ ≤
i
x ,
w
n i ,..., 1 =
(4.9c)
where
i
x denote the decision variable for Problem 4.9,
max
,i o
q denotes the maximum oil
rate for well i ,
g
Q denotes the limit of the total gas constraint. For gaslift wells, the
GOR includes both the formation gas and lift gas and is defined as
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 62
max
0 lg,
max
o
g
q
q q
GOR
+
= (4.10)
where
0 lg,
q denotes the lift gas rate that corresponds the maximum oil rate
max
o
q for a gas
lift well. In the optimal solution, 0 =
i
x indicates well i should be closed; 1 =
i
x
indicates well i should be fully open; ( ) 1 , 0 ∈
i
x indicates well i should be choked back.
4.3.5 Discussions
Model LPI and LPII are linear programming (LP) problems. Model MILPI and MILP
II are mixed integer linear programming (MILP) problems. These models are also called
the LPI, LPII, MILPI, and MILPII methods in this study, respectively. All these
methods require the objective and constraint functions be in separable form, thus they are
also referred to as the separable programming (SP) techniques.
When the computational time is not a concern, the MILPI method is the most
promising method because it can be applied to rate allocation problems with arbitrary
well performance curves. However, the MILPI method can generate a MILP problem
with a large number of binary variables and constraints even for a production system with
a moderate number of wells. For largescale problems, the other three methods may be
more appropriate depending on the characteristics of the performance of wells and how
fast we want to get the solution. For example, this study successfully applied the LPI and
LPII methods to the Prudhoe Bay oil field with appropriate speedup techniques (see
Section 4.4.2). The optimization model for Prudhoe Bay oil field has more than 1000
producers and hundreds of flowrate and velocity constraints. The MILPII method is
more appropriate for gaslift optimization problems where many wells cannot flow
without gaslift (see Section 4.4.3).
4.4 Application to Rate Allocation Problems
When the various methods are applied to a rate allocation problem, there are some
additional issues to be addressed:
1. How to handle nonlinear pressure and velocity constraints.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 63
2. How to further speed up the solution process.
This section discusses these issues and presents a gaslift optimization example. The
application of these methods to more complicated tasks (such as in well optimization and
longterm reservoir development studies) are presented in Chapter 8.
4.4.1 Handling Pressure and Velocity Constraints
For a production system, besides the flow rate constraints, a minimum pressure can be
specified for a node of the gathering system:
min , n
i
n
i
p p ≥ (4.11)
where
n
i
p denotes the pressure of node i . A maximum velocity constraint can be
specified for a tubing string or a surface flowline:
max
j j
v v ≤ (4.12)
where
j
v denote the in situ velocity of flowline j .
The SP techniques require that the optimization problem be in separable form.
However, the pressure constraints on junction nodes and velocity constraints for surface
flowlines are nonlinear functions of multiple well streams. The constraints do not satisfy
the requirement of the SP techniques, so special procedures are required.
In this study, the pressure constraints on network nodes are converted to flow rate
constraints on individual wells using a network simulation procedure proposed by Litvak
and Darlow (1995). In that procedure, whenever a pressure constraint is not honored at
some node, say node j , the flow rate of related production wells are recalculated subject
to the constraint:
min , n
j
n
j
p p = (4.13)
This procedure has been implemented in the VIP (Landmark, 2001) simulator.
The velocity constraints are incorporated into a separable problem by assuming that
the velocity of a flowline is a linear function of the flow rates of that flowline:
( )
w
v
w g
v
g o
v
o w g o
q c q c q c q q q v + + =
,
, (4.14)
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 64
where
o
q denotes the oil rate of that flowline, the weighting coefficient
o v
c
,
is a constant
defined as follows:
0 ,
max
0 ,
o
o v
o
q
v
c = (4.15)
where
0 , o
q is a specific oil rate of that flowline (such as the oil rate of that flowline before
optimization),
max
0 , o
v is the maximum in situ oil velocity along that flowline corresponding
to the reference oil rate
0 , o
q . Variables for the gas and water phases are defined similarly.
4.4.2 Speedup Techniques
Although the SP techniques are very fast for the rate allocation problems, their efficiency
is still of concern when it comes to largescale production systems with thousands of
production wells. In addition, in well connection optimization (Chapter 6) or long term
reservoir development (Chapter 8) studies, a rate allocation problem has to be solved
repeatedly hundreds or thousands of times. For these reasons, there is always a need to
speed up the solution process.
There are several potential ways to speed up the solution process. The idea presented
here is to use domain knowledge to reduce the problem size. Specifically, we want to
reduce the number of decision variables and the number of constraints of an optimization
problem, without sacrificing its accuracy. This can be done as follows.
Truncate the performance curves. The number of decision variables of Model LP
I, MILPI, and MILPII is proportional to the number of discrete points of the
performance curves. To save computational time, the performance curves can be
truncated at a production rate limit either specified in advance or converted from a
pressure constraint. In addition, a gaslift performance curve can be truncated at the point
at which the oil rate begins to decrease as more lift gas is injected.
Eliminate redundant physical constraints. For a rate allocation problem of a large
complex production system, a large number of flow rate and velocity constraints may be
specified. Often some constraints are redundant (impossible to be violated) given the
production limits of individual wells and other constraints specified for facilities and
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 65
flowlines. Some of the constraint redundancy can be eliminated using domain knowledge
(i.e., the connectivity of the production system). In this section we will describe such a
constraint elimination procedure.
First we need to define some terminology. The potential input of a node is an upper
bound on the oil, gas, and water flow rate that a node can receive from its input nodes.
The potential output of a node is an upper bound on the oil, gas, and water flow rate that
a node can deliver to its output node. The definition of input node and output node is
given in Section 3.5.2. Nodes are arranged in levels. A node belongs to the first level if
all its input nodes are wells. A node belongs to level l if the highest level of all its input
nodes is level 1 − l . With these definitions, the constraint elimination procedure goes as
follows.
1. Initialization. Let the potential output of all wells be their production limits. Let 1 = l .
2. Processing a node. For a node j in level l , perform the following steps.
2a. Calculate the potential input of node j by summing up the potential output of all
its input nodes.
2b. Determine whether a userspecified constraint for node j can be violated by the
potential input of node j .
• If the constraint can not be violated by the potential input, it is a redundant
constraint and can be eliminated. Let the potential output of node j equal to
the potential input of node j .
• If the constraint can be violated by the potential input, it is not a redundant
constraint. The potential output of node j is defined by the constraint limit.
2c. Repeat Step 2b for all userspecified constraints for node j .
3. Repeat Step 2 for all nodes in level l .
4. Let 1 + = l l . Repeat Step 2 and 3 until all nodes are processed.
As an example, this constraint elimination procedure has been applied to a rate
allocation problem in Prudhoe Bay oil field. In that problem, the user specified 190
velocity constraints and 53 flow rate constraints for facilities (the production limits on
individual wells are not counted). The speedup technique described here was able to
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 66
eliminate 226 constraints, leaving only 17 constraints. When this rate allocation problem
is solved by the LPII method, a speedup factor of more than five times was achieved as a
result of this constraint elimination procedure.
Avoid unnecessary variables and constraints in reformulation. The binary
variables { }
jk
y in Model MILPI are introduced to enforce constraint Eq. 4.5e. This is not
always necessary. When the objective and constraint functions of Problem 4.5 have
certain convex/concave properties, constraint Eq. 4.5e can be automatically satisfied
when the LP problem defined by 4.5a4.5d is solved by a general LP solver (Hartley,
1961). This rule can be used to reduce the number of decision variables of Model MILP
I. For example, consider a rate allocation problem whose objective is to maximize the
daily oil production, there is no need to introduce the binary variables { }
j jk
r k y ,..., 0 , =
for well j unless at least one of the following conditions is met:
1. The well is involved in a water or liquid flow rate constraint while its water rate
versus oil rate curve is not strictly convex.
2. The well is involved in a total gas (including formation and lift gas) flow rate
constraint while its total gas rate versus oil rate curve is not strictly convex.
3. The well is involved in a lift gas volume constraint while its lift gas rate versus oil
rate curve is not strictly convex.
4.4.3 A Gaslift Optimization Example
This gaslift optimization example has been presented in Wang et al. (2002a).
The problem of this example is from Buitrago et al. (1996). The problem is to
optimize oil production from a set of 56 wells with 22,500 MSCF/d of available gas. The
gaslift performance data are shown in Table B.1 of Appendix B. In this example, wells
4756 can not flow without gaslift. Buitrago et al. (1996) solved the problem by both an
equalslope method and the ExIn method, a stochastic algorithm that calculates the
descent direction heuristically. In this study, this problem was solved by the MILPII
method.
Table 4.1 compares the performance of the MILPII method with the equalslope
method and the ExIn method. Results for the equalslope method and the ExIn method
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 67
are taken from Buitrago et al. (1996). The MILPII method outperforms both the equal
slope method and the ExIn method for this example. Specifically, using the same
amount of lift gas, the MILPII method allocates 6.4% more oil than the equalslope
method. To allocate 21265 STB/d oil, the MILPII method requires 37.0% less lift gas
than the equalslope method. To allocate 21790 STB/d oil, the MILPII method requires
16.6% less lift gas than the ExIn method.
Table 4.1: Gaslift allocation results obtained from different methods.
MILPII EqualSlope ExIn
(1) (2) (3)
Lift Gas Rate (MSCF/d) 22,508 20,454 22,500 14,175 17,040
Oil Rate (STB/d) 21,265 21,790 22,632 21,265 21,790
(1) Allocate all available liftgas of 22500 MSCF/d.
(2) Minimize liftgas rate while keeping the oil rate at 21265 STB/d.
(3) Minimize liftgas rate while keeping the oil rate at 21790 STB/d.
To investigate the advantages of the MILPII method over the LPI method (Fang
and Lo, 1996), the LPI method was also used to solve this problem. One shortcoming of
the LPI method is that it may allocate gas from a point not lying on the gaslift
performance cuve. This is demonstrated in Figure 4.3, which presents the gaslift
performance curve and the allocated lift gas and oil rates from both the MILPII method
and the LPI method for well 47. For this well, the point selected by the MILPII method
is right on the gaslift performance curve while the point selected by the LPI method is
far above the gaslift performance curve. For the rest of wells, the data points allocated
by both methods are on the gaslift performance curves. Table B.2 of Appendix B shows
the optimal lift gas and oil rate allocated by the MILPII method for every well.
To investigate the efficiency of the MILPII method, we applied it to other similar
problems with different sizes. For example, in the second case, we have 560 wells and
225,000 MSCF/d lift gas available. All computations were performed on a Sun
Enterprise 5500 machine with 4GB of RAM and 2GB of swap space. The second row of
Table 4.2 shows the computational time (elapsed time). The third row of Table 4.2 shows
the number of LP problems solved in the solution process. Results demonstrated that the
MILPII method is very efficient for this problem.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 68
0
50
100
150
200
250
300
350
400
450
500
0 1000 2000 3000 4000
LIFT GAS RATE, MSCF/D
O
I
L
R
A
T
E
,
S
T
B
/
D
gaslift curve
LPI allocation
MILPII allocation
Figure 4.3: The allocated lift gas and oil rates for well 47 from the MILPII
method and the LPI method.
Table 4.2: Computational time of the MILPII method for various problems.
Number of Wells 56 560 3360
Elapsed Computational Time (Second) 0.27 3.06 140.29
Number of LP Problems Solved 5 6 40
4.5 Concluding Remarks
In this chapter we addressed the rate allocation problem for production systems where the
flow in surface flowlines has little impact on well performance. The approaches used here
follows the line of Fang and Lo (1996). The contributions of this study can be
summarized as follows:
1. Proposed the MILPII method, which is suitable for gaslift optimization problems in
which some wells need a finite amount of lift gas to start flowing.
2. Developed several effective speedup techniques. The speedup factor can be
significant for largescale systems.
3. The pressure and velocity constraints are handled through preprocessing and linear
approximations.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 69
The LPI and LPII methods have been coupled with a reservoir simulator so that the
impact of the reservoir can be updated at every iteration during a time step. The coupling
procedure and applications of these methods are presented in Chapter 8.
The quality of performance curves is crucial to the success of the optimization
procedure presented here. Therefore further research should be conducted to address the
problem of how to construct good quality performance curves efficiently.
70
Chapter 5
Rate Allocation through Sequential
Quadratic Programming
5.1 Introduction
Chapter 4 simplifies the rate allocation problem to a separable programming problem by
assuming that the flow interactions among production wells are not significant. However,
this assumption does not hold for many petroleum fields, especially for offshore
production systems, where the oil production is often constrained by the fluid delivering
capacities of some trunk pipelines. For such cases, rigorous network simulations have to
be employed in the optimization process to capture complex flow interactions among
different wells. Consequently the rate allocation problem has to be formulated as a
nonlinearly constrained optimization problem.
The first step in solving the nonlinear rate allocation problem is to select an
appropriate optimization algorithm. The major selection criteria are the efficiency and
robustness of the algorithm. Both genetic algorithms and gradientbased methods were
considered. The genetic algorithms can avoid being trapped in local optima, but they are
not good at handling nonlinear constraints and usually require many more function
evaluations than gradientbased methods. Because a network simulation is costly, a
gradientbased method is deemed appropriate for the rate allocation problem. Out of the
many gradient methods for constrained problems, a sequential quadratic programming
CHAPTER 5. RATE ALLOCATION THROUGH SQP 71
(SQP) method was selected. These methods are regarded as the best optimization method
for nonlinearly constrained problems whose number of decision variables is small or
moderate (Murray, 1997).
This chapter first presents a brief description of the SQP methods. Then it discusses
the appropriate formulation of the rate allocation problem and related practical issues.
Finally it presents some applications of the algorithm. Results showed that the method is
capable of handling rate allocation problems of varying complexities and sizes.
5.2 Sequential Quadratic Programming
SQP methods are widely considered the most effective algorithms for solving nonlinearly
constrained optimization problems (Murray, 1997). This section presents a brief
description of the SQP methods (Gill et al., 2002).
Consider an optimization problem, Problem 5.1, with nonlinear inequality
constraints
minimize ( ) x f (5.1a)
subject to ( ) 0 ≥ x
i
c , m i ,...., 1 = (5.1b)
where the objective function f and the constraints { }
i
c are functions of the control
variable x .
SQP methods seek a local minimum
*
x of the constrained optimization Problem 5.1.
The constraint ( ) 0 ≥ x
i
c is said to be active at
*
x if ( ) 0
*
= x
i
c , and inactive if ( ) 0
*
> x
i
c .
Let the vector ( )
*
ˆ x c denote the subset of t constraint functions that are active at
*
x , and
let ( )
*
x A
ˆ
be the matrix whose rows are the transposed gradient vectors of the active
constraints. The associated Lagrangian function for Problem 5.1 can be expressed in
terms of the objective function and active constraints (Gill et al., 1981)
( ) ( ) ( ) x c λ x λ x,
T
ˆ − ≡ f L (5.2)
where λ denote the vector of Lagrangian multipliers. The following conditions are
necessary for
*
x to be a local minimum of Problem 5.1 (Gill et al. 1981)
1. ( ) 0 ≥ x c , with ( ) 0 ˆ =
*
x c ,
CHAPTER 5. RATE ALLOCATION THROUGH SQP 72
2. ( ) ( ) 0 = ∇
*
T
*
x f x Z , or, equivalently, ( ) ( )
*
T
* *
λ x A x f
ˆ
= ∇ ,
3. 0
*
≥
i
λ , t i ,..., 1 = , and
4. ( ) ( ) ( )
* * *
T
*
x Z λ , x L x Z
xx
∇ is positive semidefinite
where ( )
*
x Z is a basis of the nullspace of ( )
*
x A
ˆ
. These optimality conditions for a
constrained optimization problem are called the second order KKT (KarushKuhn
Tucker) conditions.
SQP methods are based on these optimality conditions and they approach a local
minimum,
*
x , by using the following iteration scheme
k k k k
p x x α + =
+1
(5.3)
where
k
p is a search direction and
k
α is a step length. The search direction
k
p
approximates the error ( )
k
x x −
*
and the optimality conditions at
*
x should guide the
definition of
k
p . The step length
k
α is usually determined by a line search method.
These methods determine the search direction
k
p by solving a linearly constrained
quadratic programming (QP) subproblem
minimize p B p p g
k
T T
k
2
1
+ (5.4a)
subject to ( )
k k
c p x c − ≥ ∇ (5.4b)
where p denote the decision variables,
k
g denote ( )
k
f x ∇ , and the matrix
k
B represents
the Hessian of the Lagrangian function.
For constrained problems the choice of step length
k
α needs to ensure that the next
iterate will not only decrease f but it also will come closer to satisfying the constraints.
The most common approach is to choose
k
α in Eq. 5.3 to produce a “sufficient decrease”
in a merit function, which combines both the objective and constraint violations into one
function. The two most common choices of the merit function are the
1
l merit function
(Gill et al., 1981)
( ) ( ) ( ) ( )
∑
=
− =
m
i
i i
c f M
1
0 , min x x x ρ (5.5)
CHAPTER 5. RATE ALLOCATION THROUGH SQP 73
where 0 >
i
ρ are penalty parameters, and the augmented Lagrangian merit function (Gill
et al., 2002)
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) s x c s x c s x c x s λ, x,
T
− − + − − ≡
T
f M ρ λ ρ
2
1
, (5.6)
where s is a vector of slack variables to convert inequality constraint ( ) 0 ≥ x
i
c to an
equality constraint
( ) 0 ≥ x
i
c if and only if ( ) 0 = −
i i
s c x , 0 ≥
i
s (5.7)
The objective function of the QP subproblem and the merit function of the line
search step require estimates of the Lagrangian multipliers. These estimates can be
obtained by solving an auxiliary problem or by using the optimal multipliers for the QP
subproblem at the previous iteration (Gill et al., 1981).
5.3 Formulations of the Rate Allocation Problem
Formulation plays an important role in optimization. This section discusses the
appropriate formulation for the nonlinear rate allocation problem. The focus is on how to
handle the well chokes, one of the physical control variables of the rate allocation
problem.
5.3.1 Optimization Problem
We considered the rate allocation problems for gathering systems with treelike
structures. A gathering system includes wells, links, and nodes. The simulation model for
the gathering system is described in Chapter 3.
The objective function is the weighted sum of well rates. The physical decision
variables include lift gas rates and well chokes that control the production rate of
individual wells. The constraints include maximum/minimum flow rates and pressure
constraints imposed on production wells and/or network nodes, and maximum amount of
lift gas available for groups of gaslift wells.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 74
5.3.2 Formulation P1
A straightforward formulation of the well rate and liftgas rate allocation problem is to
choose the choke diameters and liftgas rate as the decision variables and formulate the
optimization problem as follows
maximize ( ) x f
(5.8a)
subject to ( ) u x F l ≤ ≤
(5.8b)
where ( ) x f denotes the objective function, the weighted sum of well rates, ( ) x F denotes
the constraint functions, the flow rates and/or pressures of wells and network nodes, and
l and u denote the lower and upper bounds of the constraints. This standard formulation
is named here Problem P1.
For Problem P1, given x , a particular set of choke diameters and liftgas rates, to
evaluate ( ) x f and ( ) x F , the flow rates and pressures of some wells and facilities, the
system equations of the gathering system, Eq. 3.24, have to be solved. In some cases, this
can cause severe computational difficulties. For instance, if some wells are so deep that
they cannot produce without an appropriate amount of liftgas, Eq. 3.24 will have no
feasible solutions unless appropriate amounts of liftgas are allocated to those wells. In
addition, Eq. 3.24 will have no feasible solutions if an excessive amount of gas is injected
into a well while the choke for that well is set to a small value in the optimization
process. Therefore, there are chances that the solution process for Problem P1 will fail
because feasible solutions for Eq. 3.24 do not exist.
Another significant drawback of Problem P1 is that it is not good at handling well
shutdown actions. When a well, say well j , is converted from open status to shutdown
status, we need to solve two sets of system equations with different dimensionalities (one
set without well j and one set with well j being slightly opened) to evaluate the partial
derivatives of the objective function and constraint functions with respect to the choke
diameter for well j . This procedure can be timeconsuming and the results can be
inaccurate, timeconsuming because we need to solve two sets of nonlinear equations to
obtain one partial derivative, inaccurate because we have to use finite differences to
approximate the derivatives of a potentially nonsmooth function.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 75
Some of these computational difficulties are inherent in the system of equations. As
described in Section 3.5, Eq. 3.24 is formulated under the assumption that the fluid flows
up the production trees. Replacing Eq. 3.24 with another system of equations may help
resolve some of the computational difficulties, but some difficulties will persist, such as
infeasible combination of liftgas rates and choke settings, and well shutdown handling.
Furthermore, as we will demonstrate by an example in Section 5.5.2, even when these
computational difficulties do not appear, the computational efficiency of Formulation P1
can be disappointing.
5.3.3 Formulation P2
Here we propose a new formulation to overcome limitations of Problem P1 discussed in
the previous section. The fundamental idea of the new formulation is to avoid solving the
system of equations when evaluating the objective and constraint functions.
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
0 500 1000 1500 2000
OIL RATE, STB/D
S
A
N
D
F
A
C
E
P
R
E
S
S
U
R
E
,
P
S
I
Dc=48/64 IN
DC=35/64 IN
Dc=25/64 IN
Dc=18/64 IN
Dc=13/64 IN
Dc=8.0/64 IN
Dc=4.9/64 IN
Figure 5.1: Outflow performance curves with various choke settings.
The new formulation chooses the oil rate and the liftgas rate for each well as the
decision variables and use the fact that for fixed flow rates the pressure drop across the
choke increases as the choke closes. Figure 5.1 illustrates this point. For a set of oil and
CHAPTER 5. RATE ALLOCATION THROUGH SQP 76
liftgas rates, denoted as x , we calculate the bottomhole pressures from two paths as
described in Section 3.5.1. One path is from the reservoir side to the bottomholes. The
other path is from a separator to the bottomholes. When performing the pressure
transverse calculation from the separator side, we set the well chokes fully open. The set
of oil and liftgas rate x thus computed is within the deliverability capacity of the
gathering system if we have
0 ≥ −
s
j
r
j
p p ,
w
n j ,..., 1 = (5.9)
where
r
j
p is the bottomhole pressure for well j calculated from the reservoir side,
s
j
p is
the bottomhole pressure for well j calculated from the separator side. In other words, if a
set of oil and liftgas rates satisfies constraint Eq. 5.9, this set of oil and lift gas rate can
be produced from the gathering system by, where appropriate, reducing choke diameters.
The well rate and gaslift optimization problem is now formulated as Problem P2
maximize ( ) x f
(5.10a)
subject to ( ) u x F l ≤ ≤
(5.10b)
0 ≥ −
s
j
r
j
p p ,
w
n j ,..., 1 =
(5.10c)
Constraint Eq. 5.10c is named as the deliverability constraint. This constraint should be
modified to 0 = −
s
j
r
j
p p if the choke diameter of well j is not a control variable. As in
Problem P1, ( ) x f and ( ) x F denote the flow rates and pressures of some wells or
facilities. However, they are evaluated in a different way in Problem P2 than in P1. In
Problem P2, given a set of oil and liftgas rates, x , the flow rate and bottom hole pressure
of a well are computed from the reservoir side using appropriate well models when
computing
r
j
p ; the flow rate and pressure in network nodes are computed from the
separator side when computing
s
j
p . Constraint 5.10c ensures that the optimal oil and lift
gas rate for Problem P2 are within the deliverability capacity of the gathering system.
Thus these wells can be produced from the gathering system by appropriately adjusting
the choke diameters.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 77
In the solution process for Problem P2, we do not solve the system equations (Eq.
3.24) to evaluate the objective and constraint functions, thus most computational
difficulties encountered in Formulation P1 can be avoided. After the optimization
problem is solved, the required choke diameters can be determined according to the value
of the control variables and the status of constraint 5.10c. This is explained further in
Section 5.3.4.
Problem P2 simplifies the constraints on individual wells. For a production well, its
water and gas rate, and the bottomhole pressure can be regarded as functions of its oil
rate. Because Problem P2 takes the oil rate as one of the control variables, it can convert
the rate and pressure constraints on individual wells to a bound constraint on the oil rate.
In contrast, Formulation P1 has to treat the rate and pressure constraints on individual
wells as nonlinear constraints. A nonlinear constraint is much more difficult to handle
than a simple bound constraint (Gill et al., 1981).
5.3.4 Interpretation of Solutions of Problem P2
The optimal solution of Problem P2 includes the oil rate and lift gas rate of each well, and
the status of deliverability constraint Eq. 5.10c (active or inactive). These results need
further interpretation so that the desired choke diameters can be determined. For an
arbitrary well j , its allocated oil rate, lift gas rate, and the status of corresponding
deliverability constraint in the optimal solution can be interpreted as follows:
1. A zero oil rate and lift gas rate indicate that well j should be closed. Note, a zero oil
rate really means that the oil rate for well j is at its lower bound, a small positive
value (i.e., 1 STB/d). The lower bound of the oil rate is not allowed to be zero,
because at zero oil rate, some partial derivatives required by the optimization process
can not be evaluated.
2. A positive oil rate but inactive deliverability constraint (
s
j
r
j
p p > ) indicates well j
should be partially closed (a positive oil rate means the oil rate for well j is greater
than its lower bound). The desired choke diameter can be computed according to the
desired pressure drop across the well choke for the optimal oil rate and lift gas rate of
well j .
CHAPTER 5. RATE ALLOCATION THROUGH SQP 78
3. A positive oil rate and active deliverability constraint (
s
j
r
j
p p = ) usually indicates that
well j should remain fully open. If the oil rate is close to its lower bound, then it may
indicate well j should be closed. See next item (item 4) for explanation.
4. A zero (or close to the lower bound) oil rate, a (usually small) positive lift gas rate,
and an active deliverability constraint indicate that well j should be closed. This is
because, for certain wells, say well j , reducing its production rate improves the
objective function. However, as its production rate reduces, the pressure drop in its
tubing string increases and the corresponding deliverability constraint becomes
active. To improve further the objective and keep the deliverability constraint
feasible, the optimization algorithm has to reduce its production rate but allocate
some lift gas to keep the deliverability constraint feasible. As shown in Figure 5.2, for
a production well, increasing the oil rate and lift gas rate a little bit at low oil (or
liquid) rate region can significantly reduce pressure drop in the tubing string (thus the
sandface pressure calculated from the surface side decreases).
5. Infeasible problem. If certain deliverability constraints are infeasible because the
corresponding wells are too weak to flow, the corresponding wells have to be
manually shut down, and the optimization problem has to be resolved.
Finally, it is necessary to check whether the optimal production rate is a stable
production rate for the production system. Condition 3.32 presented in Section 3.5.3 can
be used for this purpose.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 79
3000
3100
3200
3300
3400
3500
3600
3700
3800
3900
4000
0 10 20 30 40 50
OIL RATE, STB/D
S
A
N
D
F
A
C
E
P
R
E
S
S
U
R
E
,
P
S
I
QLIFT=0MSCF/D
QLIFT=20 MSCF/D
QLIFT=60MSCF/D
QLIFT=80 MSCF/D
QLIFT=100 MSCF/D
QLIFT=120 MSCF/D
QLIFT=200 MSCF/D
Figure 5.2 Impact of oil rate and lift gas rate on the well sandface pressure
calculated from the separator side. Major characteristics of this well are as
follows. The water cut is 0.26 STB/STB, GOR is 12.50 MSCF/d, tubing
diameter is 3.5 inches, well depth is 8000 ft, well head pressure is 1080 psi.
5.4 Solving the Optimization Problem
Problems P1 and P2 are nonlinearly constrained optimization problems. In this work,
they were solved by SNOPT (Gill et al., 1998), a generalpurpose system for solving
largescale optimization problems. SNOPT implements a SQP algorithm that obtains
search directions from a sequence of quadratic programming subproblems. SNOPT
requires relatively few evaluations of the problem functions, hence it is especially
effective if the objective or constraint functions are expensive to evaluate. The SQP
algorithm used in SNOPT is described fully in Gill et al. (2002).
SQP is a derivativebased optimization algorithm. Successful application of SQP
requires efficient and accurate evaluations of gradients of objective and constraint
functions. In this work, we were able to compute the gradient information efficiently and
accurately by exploring the treelike structure of the gathering system and utilizing
CHAPTER 5. RATE ALLOCATION THROUGH SQP 80
automatic differentiation techniques. Details of this procedure are presented in Section
3.5.2.
Experience with the application of SNOPT to the nonlinear rate allocation problem
showed that the number of required function evaluations is sensitive to the starting point,
the gathering system and the optimization problem, the convergence criteria, and the
model and derivative accuracy. Therefore, the performance of SNOPT should be tuned
properly whenever the condition of the production system or the optimization problem
changes significantly.
5.5 Examples
We present three examples in this section. The first example compares the performance
of the SQP method with the MILPI method (see Section 4.3.2) for a well and liftgas rate
allocation problem. The second example compares the computational efficiencies of
Formulations P1 and P2. The third example demonstrates the computational efficiency of
Formulation P2 on problems of various sizes and properties. These examples have been
presented in Wang et al. (2002b).
5.5.1 Comparison of the MILPI Method and the SQP Method
The production system of this example is described in Section 3.7. The problem is to
maximize the total oil production rates by allocating production rates and unlimited lift
gas to the ten production wells. The constraint is that the total water production rate can
not exceed 5000 STB/d.
The optimization problem is solved by the MILPI method presented in Section 4.3.2
as follows.
1. Choose a realistic flowing pressure at the platform. Suppose wells are producing with
a set of liftgas rate (not optimized) as shown in column 2 of Table 5.1. We simulated
the multiphase flow in the gathering system and obtained the flowing pressure at the
platform, which is 1545 psi. Table 5.1 also shows the oil flow rate, the water cut, and
the gas oil ratio (GOR) for each well obtained in this step.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 81
2. Construct gaslift performance curves. First we fixed the platform pressure at 1545
psi. Then for any well, say well j , given its gaslift rate, we could determine its oil
flow rate by solving the multiphase flow problem for the flow path connecting well j
and the platform. For each well, we computed the oil flow rate for different gaslift
rates, namely, 0, 200, 500, 1000, 2000, 3000, 4000, 5000, 6000, and 7000 MSCF/d,
and obtain its gaslift performance curve. Figure 5.3 plots the gaslift performance
curves we obtained in this step.
3. Allocate liftgas. First we approximated each gaslift performance curve using a
piecewise linear curve. Then we solved the rate allocation problem using the MILPI
method. The optimal solution predicted that the field will produce at an oil flow rate
of 12449 STB/d. The allocated liftgas rate and oil rate for each well are presented in
Table 5.2.
4. Update multiphase flow. We allocated the optimal set of liftgas rates from Step 3 to
each well, simulate the multiphase flow in the gathering system, and obtained the oil
flow rate of each well. Results indicated that the field can produce at a total oil flow
rate of 12045 STB/d and a total water flow rate of 4830 STB/d, which does not
violate the total water flow rate constraint. The flowing pressure at the platform is
1595 psi.
We then proceeded to solve the same optimization problem using the SQP method
with Formulation P2. We selected the oil rate and liftgas rate for each well as the
decision variables. Starting with an initial guess of 400 STB/d of oil rate and 3000
MSCF/d of lift gas rate for every well, the program converged to the optimal solution
using 27 function evaluations. One function evaluation refers to one complete evaluation
of the objective and constraint functions and their derivatives. The allocated oil rate and
liftgas rate for each well are shown in Table 5.2. Results indicate that the field will
produce at a total oil rate of 13019 STB/d and a total water rate of 5000 STB/d. The
flowing pressure at the platform is 1079 psi. Note that in this example the flowing
pressure at the platform is sensitive to the flow rates in the common flow line.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 82
Table 5.1: Well rates obtained in Step 1 of the MILPI method for Example 1.
Well Liftgas Rate
(MSCF/d)
Oil Rate
(STB/d)
Water Cut Gas Oil Ratio
(MSCF/STB)
1 3000.00 735.35 0.53 1.21
2 0 988.90 0.26 12.50
3 2000.00 2590.69 0.38 1.65
4 3000.00 828.18 0.70 1.02
5 0 1168.80 0.16 13.41
6 0 1143.09 0.07 6.84
7 2000.00 144.49 0.47 2.87
8 0 933.03 0.11 40.38
9 0 740.73 0.10 31.41
10 3000.00 2911.25 0.22 1.02
Total Flow Rate 13000.00 121842.52 6149.55
*
10618.42
**
*
total water flow rate, STB/d
**
total formation gas flow rate, MSCF/d
0
500
1000
1500
2000
2500
3000
3500
0 2000 4000 6000 8000
LIFTGAS RATE, MSCF/D
O
I
L
R
A
T
E
,
S
T
B
/
D
WELL 1
WELL 3
WELL 4
WELL 6
WELL 7
WELL 10
Figure 5.3: Gaslift performance curves generated in Step 2 of the MILPI
method. For clarity, curves for high GOR wells (like well 6) are not
shown.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 83
Table 5.2: Allocated lift gas and oil rates for Example 1.
Step 3 of the MILPI MethodStep 4 of the MILPI Method The SQP Method
Well Liftgas Rate
(MSCF/d)
Oil Rate
(STB/d)
Liftgas Rate
(MSCF/d)
Oil Rate
(STB/d)
Lift Gas Rate
(MSCF/d)
Oil Rate
(STB/d)
1 2674.00 668.36 2674.00 649.60 
*

*
2 0 988.64 0 947.07 0 1237.64
3 5000.00 2678.83 5000.00 2608.54 0 3233.68
4 
*

*

*

*

*

*
5 0 1168.09 0 1149.46 0 1293.56
6 506.22 1149.29 506.00 1082.15 0 1621.08
7 7000.00 927.29 7000.00 872.57 750.15 3108.30
8 0 932.62 0 933.03 
*

*
9 0 741.21 0 731.75 0
**
270.72
**
10 7000.00 3194.95 7000.00 3070.97 2769.25 4301.56
Total 22180.22 12449.24 22180.00 12045.15 3519.40 13019.66
*
shutin wells
**
partially closed wells
Some comments on the MILPI method and the SQP method are appropriate at this
stage:
• In Step 2 of the MILPI method, in order to construct the gaslift performance curves,
we needed to fix the platform pressure to a rather arbitrary value. The results of the
MILPI method will largely depend on this value. However, there is no simple way to
tell which value is appropriate to choose.
• In Step 3 of the MILPI method, the liftgas was allocated based on existing gaslift
performance curves. This ignores the fact that increasing the flow rate from one well
will decrease the flow rate from other wells sharing a common flowline. So the
MILPI method tends to overinject liftgas in Step 3.
• The SQP method takes the flow interaction into account through the deliverability
constraint Eq. 5.10c, and is able to make more informed decisions. For example, well
8 and 9 have high GORs. For the same amount of oil produced, these two wells will
cause a larger pressure drop through the common flowline than the rest of the wells,
consequently the optimization algorithm will reduce the flow rates from other wells.
The SQP method closed well 8 and partially closed well 9 to utilize fully the
deliverability capacity of the common flowline. As a result, the SQP method allocated
much less liftgas but produces 8% more oil than the MILPI method.
• The major computational effort of the MILPI method is in constructing the gaslift
performance curves. Thus the computational time of the MILPI method is
CHAPTER 5. RATE ALLOCATION THROUGH SQP 84
proportional to the number of wells in the system. The computational time of the SQP
method depends on many factors and is hard to predict. For this particular example,
the computational time required for the MILPI method and the SQP method were of
the same order.
Finally we mention that this problem cannot be solved by Formulation P1, because
some wells (i.e. well 7) are too weak to flow under certain conditions and well shutdown
actions (i.e. well 1, 4, and 8) must be performed in the optimization process.
5.5.2 Comparison of Formulation P1 and P2
Formulation P2 is easier to implement and is more robust than P1. However, Formulation
P2 turns an unconstrained optimization problem into a constrained one. We must ensure
that such a reformulation does not make an easy problem hard to solve. The purpose of
this example is to show that the performance of Formulation P2 is not necessarily worse
than P1 on problems that can be solved by both formulations.
In this example, the reservoir conditions and the configuration of the gathering
system are exactly the same as in the example of Section 5.5.1. The difference is that the
tubing strings and pipes were made shorter in this example so that all wells can produce
without gaslift.
We present two scenarios. The first scenario is an unconstrained gaslift optimization
problem. The objective is to maximize the total oil production by allocating liftgas. Both
Formulation P1 and P2 were used to solve this problem, and they yielded the same
optimal solution, which is zero liftgas rate for every well. However, Formulation P2
converged much faster than P1. Formulation P2 optimizes both the oil flow rate and lift
gas rate of every well. Starting with an initial guess of 400 STB/d oil flow rate and 5000
MSCF/d liftgas rate for every well, Formulation P2 converged to the optimal solution
with 21 function evaluations. Formulation P1 optimizes only the liftgas rate. Starting
with an initial guess of 5000 MSCF/d liftgas rate for every well, Formulation P1
converged to the optimal solution with 67 equivalent function evaluations. The
convergence histories Formulation of P1 and P2 are plotted in Figure 5.4.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 85
The second scenario is a constrained well rates and gaslift optimization problem.
The objective is to maximize the total oil production rate by allocating liftgas and
adjusting choke diameters. The total water flow rate can not exceed 20,000 STB/d.
Again, both formulations yielded the same optimal solution: zero liftgas rate for all
wells, and the well choke for well 4, the well with highest water cut, was partially closed.
However, Formulation P2 was much more efficient than P1. Starting with an initial guess
of 400 STB/d oil flow rate and 5000 MSCF/d liftgas rate for every well, Formulation P2
converged to the optimal solution with 20 function evaluations. Starting with an initial
guess of 5000 MSCF/d liftgas rate and fully open choke status, Formulation P1
converged to the optimal solution with 82 equivalent function evaluations. The
convergence history is plotted in Figure 5.5.
In both scenarios, Formulation P2 performed better than P1.
0
5000
10000
15000
20000
25000
30000
35000
40000
0 10 20 30 40 50 60 70
NUMBER OF FUNCTION EVALUATIONS
O
B
J
E
C
T
I
V
E
F
U
N
C
T
I
O
N
,
S
T
B
/
D
P1
P2
Figure 5.4: Convergence history of Formulation P1 and P2 for Scenario 1.
CHAPTER 5. RATE ALLOCATION THROUGH SQP 86
0
5000
10000
15000
20000
25000
30000
35000
40000
0 20 40 60 80 100
NUMBER OF FUNCTION EVALUATIONS
O
B
J
E
C
T
I
V
E
F
U
N
C
T
I
O
N
,
S
T
B
/
D
P1
P2
Figure 5.5: Convergence history of Formulation P1 and P2 for Scenario 2.
5.5.3 Efficiency of Formulation P2
This example demonstrates the computational efficiency of Formulation P2 on problems
of various sizes and properties.
We considered three gathering systems with 2, 10, and 50 production wells,
respectively. Figure 5.6 shows the configuration of these gathering systems. For each
gathering system, we considered three optimization problems: unconstrained, constrained
by total water flow rate, constrained by total gas and water flow rates. The objectives of
all problems were to maximize the daily oil production by allocating the well and liftgas
rates. We solved the nine optimization problems using Formulation P2. The initial guess
for every optimization problem was 400 STB/d oil flow rate and 3000 MSCF/d liftgas
rate for every well. The convergence criteria for the optimization process are presented in
Table 5.3.
Table 5.4 shows the number of function evaluations required to solve the
optimization problem and the total oil rate of the optimal solution for every problem.
Constraints can greatly influence problem difficulty. For example, problems 7, 8, and 9
have the same size but require 62, 50, and 28 function evaluations, respectively. The
CHAPTER 5. RATE ALLOCATION THROUGH SQP 87
number of function evaluations required is not sensitive to the problem size. All of the
problems discussed here can be solved within about 60 function evaluations (one function
evaluation is equivalent to one Newton iteration when Eq. 3.24 is solved by the Newton
Raphson method). In other words, all of the problems discussed here can be solved within
approximately 10 times the number of iterations normally required to solve the network
problem (defined by Eq. 3.24) by the NewtonRaphson method. Thus, we conclude that
the SQP method with Formulation P2 is computationally efficient.
Separator
Choke
Wells
Tubing
string
Pipe
(a) Gathering system with 2 wells (b) Gathering system with 10 wells
(c) Gathering system with 50 wells (with a choke installed right before the
manifold for every well)
Figure 5.6: Schematic illustration of gathering systems.
Wells
Tubing
strings
Chokes
Platform
Separator
Pipes
CHAPTER 5. RATE ALLOCATION THROUGH SQP 88
Table 5.3: Convergence criteria for optimization problems of Example 3.
Major Feasibility Tolerance
*
5.0e4
Major Optimality Tolerance
*
1.0e3
Minor Feasibility Tolerance
*
5.0e6
Minor Optimality Tolerance
*
1.0e6
*
These are the major convergence criteria used in SNOPT (Gill et al., 1998). The major
feasibility tolerance specifies how accurately the nonlinear constraints should be satisfied. The
major optimality tolerance is used to judge the optimality of the optimization problem. The
minor feasibility tolerance specifies how accurately the bound and linear constraints should be
satisfied. The minor optimality tolerance is used to judge the optimality of each QP subproblem.
Table 5.4: Computational efficiency of Formulation P2 on various
optimization problems.
Problem Number
of Wells
Constraints Number of Function
Evaluations
Total Oil Rate
(STB/d)
1 2 None 21 1976.52
2 2
1000 ≤
t
w
q STB/d
11 1480.38
3 2
1000 ≤
t
w
q STB/d
20000 ≤
t
g
q MSCF/d
14 1267.21
4 10 None 49 14884.96
5 10
8000 ≤
t
w
q STB/d
37 14585.92
6 10
8000 ≤
t
w
q STB/d
50000 ≤
t
g
q MSCF/d
45 14333.10
7 50 None 62 83423.14
8 50
35000 ≤
t
w
q STB/d
50 81698.21
9 50
35000 ≤
t
w
q STB/d
300000 ≤
t
g
q MSCF/d
28 81067.22
5.6 Concluding Remarks
We have investigated a new formulation P2 for the optimization problem of allocating
well rates and liftgas rates simultaneously subject to multiple flow rate and pressure
constraints. The optimization problem is solved by a SQP algorithm. Results
demonstrated that the SQP method performs much better on certain problems than the
MILPI method. This is due to the fact that the MILPI method optimizes the well rates
and liftgas rates based on gaslift performance curves and ignores the deliverability
CHAPTER 5. RATE ALLOCATION THROUGH SQP 89
constraint of the gathering system in the optimization process, while the SQP method
makes no such simplifications. We determined that Formulation P2 is appropriate for the
simultaneous optimization of well rates and liftgas rates, because with this formulation,
it is easier to handle well shutdown actions and avoid other computational difficulties.
Furthermore, with P2 the optimization problem can be solved efficiently.
However there are some limitations of Formulation P2. In the optimal solution of P2,
sometimes a well can have a zero oil rate and a positive liftgas rate, which is not
realistic. This positive liftgas rate is used to lighten the oil column so that constraint Eq.
5.10c is satisfied. Fortunately, our experience shows that the value of this liftgas rate
usually is small and has little impact on the solution of the optimization problem.
More research work could be done on this subject. We can incorporate more decision
variables into the optimization problem, such as pumps and subsurface chokes for
multilateral and multisegment wells. We can modify Formulation P2 to allow loops in the
network. It is also necessary to investigate the performance of the linear optimization
methods and the nonlinear optimization method under different conditions.
90
Chapter 6
Optimization of Well Connections
6.1 Introduction
In some petroleum fields, a well or a flow line can have several potential output
connections that join that well or flowline to different flowlines and facilities. Redirecting
the well connections is an effective way to debottleneck the production system. The
problem of well connection optimization is to identify the best set of well connections
that maximizes an operational objective. To achieve the best results, well connections
often need to be optimized simultaneously with production rates and lift gas rates (and
potentially other production operations). Thus, the production optimization problem
addressed in this section involves maximizing a certain operational objective (such as to
maximize the daily oil rate) by optimally allocating the production and lift gas rates and
well connections subject to multiple flow rate and pressure constraints. To differentiate
this problem from the rate allocation problem addressed in previous chapters, the
optimization problem addressed in this chapter is referred to as the global problem.
The global problem is a nonlinearly constrained optimization problem with both
continuous (production and lift gas rates) and integer (well connections) variables. No
existing formal optimization methods can be easily adapted to solve such a complicated
problem. Litvak et al. (1997) presented a heuristic optimization procedure to allocate the
production rates, lift gas rates, and well connections for the Prudhoe Bay oil field in
Alaska. However, the performance of that procedure can be unsatisfactory because (1)
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 91
the production rates, lift gas rates, and well connections are allocated sequentially rather
than simultaneously; (2) the multiple flow rate constraints are checked and satisfied
sequentially rather than simultaneously; and (3) the allocation procedures (either for well
connections, production rates, or gaslift rates) are too simple for a complex system like
the Prudhoe Bay field.
In this chapter, a new formulation of the optimization problem is presented. Two
solution methods are discussed. The performances of the solution methods were
investigated and will be illustrated here by an example.
6.2 TwoLevel Optimization Formulation
Although the physical decision variables for the global problem are the well connections,
lift gas rates, and well chokes, it was shown in Chapter 5 that optimizing on well chokes
is computationally inconvenient for the rate allocation problem. Because the rate
allocation problem is a part of the global problem, production rates instead of well chokes
are selected as decision variables for the global problem.
The rate allocation problem has been described extensively in Chapter 4 and 5. The
global problem is formulated in such a way that existing methods for rate allocation can
be utilized for the global problem. Let x denote a set of well connections, y denote lift
gas rates, and z denote production rates. The global problem is formulated as a twolevel
optimization problem:
z y, x,
maximize ( ) x
u
f (6.1a)
subject to
wc
Ω ∈ x (6.1b)
where
wc
Ω is the domain of feasible well connections, ( ) x
u
f is defined as the optimal
solution of a rate allocation problem
( ) = x
u
f
z y,
maximize ( ) z y, x,
l
f (6.2a)
subject to ( ) u z y, x, F l ≤ ≤ (6.2b)
where ( ) z y, x,
l
f is the objective function for the rate allocation problem parameterized
by x , ( ) z y, x, F are the constraint functions for the rate allocation problem parameterized
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 92
by x . l and u are the lower and upper bounds of the constraints, respectively.
l
f and F
are specified by the global problem. For example, if the global problem is to maximize
the total field oil rate subject to a total field gas rate constraint, then
l
f is the total field
oil rate and F is the total field gas rate.
In this formulation, Problem 6.1 is named the upper level problem, and Problem 6.2
is named the lower level problem. The upper level problem is an unconstrained integer
optimization problem. The lower level problem is a nonlinearly constrained rate
allocation problem. The lower level problem can be regarded as a function evaluation
procedure for the upper level problem.
This formulation separates the search on well connections and the search on
production and lift gas rates into two levels. The lower level problem has been described
extensively in Chapters 4 and 5. This chapter focuses on optimization methods for the
upper level problem – optimization of well connections. The upper level problem is also
referred to as the well connection optimization problem in this chapter.
6.3 Solution Methods for the Optimization of Well
Connections
This section first motivates the choice of solution methods for the upper level problem,
then describes the investigation of some possible solution methods.
6.3.1 Computational Issues
The upper level problem is a nonlinear integer programming (NIP) problem. In principle,
a NIP problem can be solved by enumeration. However, complete enumeration is
computationally impractical even for well connection optimization problems with a
moderate number of connections. For example, consider a production system with 20
wells, each well has three potential output connections. A combination of these well
connections represents 3
20
or 3486784401 distinctive configurations of the production
system. Suppose a computer can evaluate one configuration per second, it would take
about 110 years to evaluate all the possible configurations, which is not practical.
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 93
So far, there does not exist a universal algorithm for integer programming: existing
algorithms either require huge computational effort, or they only give approximate
solutions (Bertsimas and Tsitsiklis, 1997). Because efficiency is an important
consideration for realtime production optimization, we developed a partial enumeration
method that gives only an approximate solution but is efficient for the upper level
problem. A genetic algorithm was also used. The genetic algorithm has the ability to
avoid local optima and it is easy to implement. The genetic algorithm can be used to
crosscheck the performance of the partial enumeration method.
6.3.2 Partial Enumeration Method
The partial enumeration (PE) method is a heuristic method designed specially for the
optimization of well connections. The method reconnects a well to its best output
connection while keeping other well connections fixed. This process is repeated until no
further significant improvement can be made. The basic steps of this method can be
illustrated as follows.
Algorithm 6.1: The partial enumeration method.
1. Set iteration index 0 = k .
2. Select a production well, well j , according to some operational rules.
3. For well j , perform the following operations:
3a) Switch well j from its current output connection to another potential output
connection. This step generates a new configuration of the production system.
3b) Invoke optimization tools as described in Chapter 4 or Chapter 5 to reallocate
the production and lift gas rate subject to imposed constraints. The optimal
objective value from the rate allocation problem is regarded as the merit or
value of the new output connection of well j . (This step solves the lower level
problem.)
3c) Repeat Step 3a)3c) for all potential output connections of well j . Define the
output connection with the maximum merit/value as the best output connection
of well j .
3d) Switch well j to its best output connection as determined in step 3c).
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 94
4. Repeat steps 2 and 3 for all production wells in the field.
5. Increase the iteration index k by 1.
6. Repeat Step 1 to 5 until certain user defined convergence criteria are met. One such
criterion is: between iteration 1 − k and k , no well changes its connection or no
discernable increase of the objective function value is achieved.
The PE method mimics what a field operator facing the same problem would do. The
method attempts to find a better system configuration at each search step, but it does not
guarantee the global optimum.
The PE method is flexible. Different rate allocation algorithms can be employed in
this method. Fieldspecific operational rules can be incorporated easily. For example, the
actual implementation of Algorithm 6.1 can limit the maximum number of well
reconnections so that the configuration of the production system does not change
dramatically from day to day.
The performance of the PE method was evaluated and is illustrated in an example
presented in Section 6.4.
6.3.3 Genetic Algorithm
The PE method is a heuristic method that examines only a portion of the complete search
space. In addition, at each search step only one well is allowed to change its connection.
This may cause some concerns for the quality of the solutions obtained from such a
method. Genetic algorithm (GA) is a general, robust, and well developed optimization
method. We used this method to crosscheck the performance of the PE method.
A genetic algorithm is a stochastic optimization algorithm based on natural selection.
In GA, decision variables are encoded as a chromosome. Multiple chromosomes form a
population. A GA evolves a population by means of selection, crossover, and mutation,
etc. When the evolution terminates, the chromosome with the best fitness represents the
optimal solution of the optimization problem. A flowchart of a simple GA is illustrated in
Figure 6.1. A GA is applied to the well connection optimization problem in the following
manner.
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 95
Create an initial population
Decode and evaluate current population
Produce next generation by
selection, crossover, mutation, etc.
Meet stopping criteria?
Return best value and stop
No
Yes
Figure 6.1: A simple GA flowchart.
Encoding/decoding. In GA, a set of decision variables is usually, though not
necessarily, encoded as a binary string, with each single variable corresponding to a
segment of the binary string. In GA vocabulary, the binary string is called a chromosome,
and the segment of the binary string representing a variable is named a gene. With such
an encoding scheme, decision variables take discrete values.
This study used the following encoding/decoding scheme for well connections.
Suppose a well has
c
n potential (output) connections and they are labeled by a set of
consecutive numbers, [ ] 1 , 0 −
c
n . Then the encoding and decoding process for this well
goes as follows:
• Encoding. A specific well connection is encoded into a gene by expressing its
labeling number in binary format. The length of a gene is the minimum number of
bits required to cover all potential connections of a well:
{ }
c
m
m
n l ≥ = 2 min (6.3)
For example, suppose a well has three potential connections. Then its connections
can be encoded into 00, 01, 10 respectively (Figure 6.2).
• Decoding. The decoding process maps a gene to a well connection by converting its
binary expression to a natural number, k . If [ ] 1 , 0 − ∈
c
n k , the gene is mapped to the
well connection labeled as k . If [ ]
l
c
n k 2 , ∈ , the gene is mapped to the well
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 96
connection labeled as ) 1 ( − −
c
n k . Consider the above encoding example, the gene
expressions 00, 01, and 10 can be mapped to well connections labeled 0, 1, and 2,
respectively. A gene expression of 11 can be mapped to a well connection labeled 0.
This decoding scheme is illustrated in Figure 6.2.
The genes for all decision variables comprise a chromosome.
0 1 2
00 01 10 11
0 1 2
00 01 10 11
well connection
labeling
gene
encoding
decoding
Figure 6.2: Encoding/decoding a well connection.
Initializing a population. This step generates the initial population (multiple sets of
solutions or chromosomes) that GA begin to evolve with. One chromosome encodes the
current set of well connections to ensure a good starting point. The rest of the
chromosomes are randomly generated to ensure diversity. The number of chromosomes
in the population is denoted as n .
Evolving the population. This is the major part of a GA, and is an iterative
procedure with each iteration containing the following steps:
1. Evaluate the population. Decode each chromosome and evaluate its fitness (function
value). For each chromosome, a rate allocation problem is solved and the optimal
objective function value is assigned as the fitness of the corresponding chromosome.
2. Produce next generation. The current generation of the population is modified
through selection, crossover, mutation, and other operations to produce the next
generation of population.
2a) Selection. Replicate chromosomes for mating. This study implements a
tournament selection, in which a mating chromosome is selected as follows.
Randomly pick two chromosomes from the current population, and the one with
higher fitness value is selected as the mating chromosome. Repeat this procedure
until n mating chromosomes are selected.
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 97
2b) Crossover. Randomly pick two chromosomes from the replicated population.
Crossover the two chromosomes (the parent) to produce two new chromosomes
(the children). There are different ways to crossover two chromosomes. One type
of crossover is called singlepoint crossover, in which the two chromosomes swap
their bits at a randomly selected position with a probability
c
p . Figure 6.3 shows
an example of singlepoint swap. Another type of crossover is called uniform
crossover, in which the two chromosomes exchange their bits at every position
based on a probability
c
p . Figure 6.4 shows an example of uniform crossover.
This study implements both crossover schemes.
1 1 1 1 1 1 1
0 0 0 0 0 0 0
1 1 1 1 0 0 0
0 0 0 0 1 1 1
crossover
Figure 6.3: Schematic illustration of singlepoint crossover.
1 1 1 1 1 1 1
0 0 0 0 0 0 0
1 0 1 1 0 0 1
0 1 0 0 1 1 0
crossover
Figure 6.4: Schematic illustration of uniform crossover.
2c) Mutation. Mutation is another way to increase the diversity of the population. In
this step each bit in a chromosome flips according to some probability
m
p . A
mutation example is shown in Figure 6.5.
1 1 1 1 1 1 1
mutation
1 0 1 1 0 1 1
Figure 6.5: Schematic illustration of mutation.
2d) There are other techniques that can enhance the performance of the GA. This
study used the following two: (1) the best parent is always copied to the next
generation (elitism), and (2) if there is not enough diversity in a population (a
population is converged because all chromosomes are too similar), start a new
generation. The new generation copies the best chromosome from the converged
population and generates the rest of the chromosomes randomly.
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 98
3. Repeat Step 1 and 2 until stopping criteria is met. The stopping criterion used in this
study was a specified maximum number of generations.
The genetic algorithm used in this study was based on a genetic algorithm driver
developed by Carroll (2000).
6.4 Comparison of the PE Method and the GA
We analyzed the performance of the PE method and the genetic algorithm on a model of
the Western Operating Area of Prudhoe Bay oil field in Alaska.
Prudhoe Bay oil field is the largest oil field in north America. Currently there are 23
well pads in the Western Operating Area (WOA) and 16 drill sites in the Eastern
Operating Area (EOA). Each well pad or drill site contains about 3040 wells (Litvak et
al., 1997). Oil production from Prudhoe Bay is on decline and is constrained by the gas
and water handling limits of surface facilities and velocity constraints in flow lines. On
the other hand, Prudhoe Bay has a flexible, highly automated and complex surface
pipeline network. Production from many wells can be redirected to different flowlines
and facilities (Litvak et al., 2002). For these reasons, optimization of well connections
and production rates is a viable and economic way to increase the daily oil production in
the Prudhoe Bay field. Therefore, the production optimization problem for the Prudhoe
Bay field is to maximize the daily oil production by optimally allocating the well
connections and production rates subject to multiple flow rate and velocity constraints.
The field model of Western Operating Area contains about 500 hundred wells, each well
has two to three potential connections. Roughly 100 wells are considered for re
connection. Other wells’ connections remain fixed.
The production optimization problem was formulated as a twolevel optimization
problem. The lower level problem optimizes the production rates subject to all
constraints. The lower level problem was solved by the LPII method described in
Section 4.3.4. The upper level optimizes the well connections. Both the PE method and
the GA were used to solve the upper level problem.
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 99
The PE method is a deterministic heuristic technique. To see if the method was
sensitive to its starting point, we made five test runs. Each run started the PE method
from a randomly generated set of well connections. These runs found different sets of
solutions after 400 function evaluations (one function evaluation refers to solving an
instance of the lower level problem). However, the best objective values are close to each
other. Table 6.1 shows the starting and final objective values of all runs. Figure 6.6 plots
the convergence histories of the first three runs. The line labeled PE BEST in Figure 6.6
represents the best objective found by these runs. The objective value, which is the total
oil rate, is normalized by a selected value. Figure 6.6 shows that the PE method increased
the oil rate quickly at early stages, then slowed down at later stages. It is observed from
Table 6.1 that the final objective values from these runs were not sensitive to the starting
point. The biggest difference is 0.0013 of normalized oil rate.
Table 6.1: Performance data of the PE method.
PE1 PE2 PE3 PE4 PE5
Starting Normalized
Total Oil Rate
0.8911 0.9070 0.9352 0.8852 0.9202
Final Normalized
Total Oil Rate
0.9935 0.9935 0.9922 0.9921 0.9928
Number of Lower Level
Problems Solved
477 636 636 638 638
For the GA run, the following basic GA parameters were used: the number of
chromosomes, n , of a population was 10, uniform crossover probability,
c
p , was 0.5,
and mutation probability, 1 . 0 / 1 = = n p
m
. In this run, all initial chromosomes were
generated randomly, the current set of well connections were not encoded as one of the
initial chromosomes. In this way, the starting points for both the GA and the PE method
were generated randomly and the comparison between the GA and the PE method is
therefore more meaningful. After 1000 generations or 10,000 function evaluations (a
function evaluation is defined in the same way as defined for the PE method), the best
objective value obtained by the GA method was a normalized oil rate of 0.9922. The
objective function was normalized in the same way as in the PE method. Figure 6.7 plots
the convergence history of the GA method. The line labeled PE BEST represents the best
objective function value obtained from the PE method. The early convergence history of
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 100
the GA method is also plotted in Figure 6.6. The following observations are made from
Figure 6.6 and Figure 6.7:
• The GA was not as efficient as the PE method. The GA converged to a value no
larger than that from the PE method and it used 25 times more function evaluations
than the PE method.
• The PE method is robust because it was not sensitive to its starting points and it
achieved a better objective value than the GA.
The good performance of the PE method can be explained partially as follows. The
configuration of a real production system that has been operated for a long period is
already close to optimum. Such a system needs only fine adjustment in response to the
changing operational conditions. The PE method, which changes at most one well
connection at each step, is suitable for this purpose.
0.88
0.9
0.92
0.94
0.96
0.98
1
0 100 200 300 400
NUMBER OF FUNCTION EVALAUTIONS
N
O
R
M
A
L
I
Z
E
D
O
I
L
R
A
T
E
PE1
PE2
PE3
GENETIC ALG
PE BEST
Figure 6.6: Convergence history of the PE runs.
CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 101
0.94
0.95
0.96
0.97
0.98
0.99
1
0 2000 4000 6000 8000 10000
NUMBER OF FUNCTION EVALUATIONS
N
O
R
M
A
L
I
Z
E
D
O
I
L
R
A
T
E
GENETIC ALG
PE BEST
Figure 6.7: Convergence history of the GA run.
6.5 Concluding Remarks
The advantage of the twolevel optimization is flexibility. The approach transforms the
entire problem to an upper level problem and a lower level problem, and it allows the use
of existing or new optimization methods separately for each level. This advantage is
significant for the production optimization problem addressed in this chapter because no
existing optimization algorithms can solve the entire problem efficiently. The solution
quality of this approach depends on the properties of the global problem and which
optimization methods are used for the two levels.
For the upper level problem, if the number of wells in a production system is small
(say, less than 20), complete enumeration is possible to find the best set of well
connections. For moderate or largescale systems, approximate methods have to be used.
The partial enumeration method proved to be both robust and efficient for several fields.
However, as a specially designed heuristic method, the performance of the PE method
has to be examined carefully before it is applied to a new field. Genetic algorithm
provides an alternative solution method, but it is generally more expensive. The GA can,
however, be used to test the robustness of the PE method.
102
Chapter 7
Multiobjective Optimization of
Production Operations
7.1 Concept of Multiobjective Optimization
The optimization problems addressed in previous chapters assume a single objective
function. In practice, petroleum production engineers often strive to achieve more than
one goal when operating a field. Typical goals include maximizing the daily oil
production, minimizing the production cost, and minimizing safety and environmental
hazards. In most cases, it is unlikely that different objectives can be optimized by the
same set of decision variables. The problem of optimizing multiple and distinct
objectives is named multiobjective optimization problem (MOP) (Miettinen, 1999).
Mathematically, a MOP can be expressed as
minimize ( ) ( ) ( ) ( ) [ ] x x x x F
k
f f f ,..., ,
2 1
= (7.1a)
subject to ( ) { } m i c
i
,..., 1 , 0 : = ≤ = ∈ x x S x (7.1b)
where ( ) x
1
f , ( ) x
2
f ,..., ( ) x
k
f are the k objective functions or attributes, x is an n
dimensional vector of decision variables, S denotes the feasible set of x , and ( ) x
i
c
denotes a constraint. The space to which the objective vector ( ) x F belongs is called the
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 103
objective space or criterion space. The feasible set of ( ) x F is called the attained set,
which is denoted by
( ) { } S x x F Y ∈ = : (7.2)
In multiobjective optimization, no solution is optimal for all the k objectives
simultaneously. Thus the optimality criterion used for single objective optimization must
be replaced by a new one. One such replacement is the notion of Pareto optimality
(Miettinen, 1999). Consider Problem 7.1 and two solution vectors x and y . x is
dominant over y if (Andersson, 2001):
{ } ( ) ( ) y x
i i
f f k i ≤ ∈ ∀ : ,..., 2 , 1 and { } ( ) ( ) y x
j j
f f k j < ∈ ∃ : ,..., 2 , 1 (7.3)
A point S x
*
∈ is said to be Pareto optimal for a MOP if and only if there is no S x ∈ is
dominant over
*
x . All the Pareto optimal solutions form the Pareto Optimal set. The
solutions in the Pareto optimal set are not dominant over each other. Moving from one
point to another point within the set will deteriorate at least one objective or attribute. In
most cases, the Pareto optimal set is on the boundary of the feasible set Y (see Figure
7.1).
attained set Y
A
B
C
D
1
f
2
f
Pareto optimal set
attained set Y
A
B
C
D
1
f
2
f
Pareto optimal set
Figure 7.1: Graphical illustration of a Pareto optimal set. The problem is to
minimize both
1
f and
2
f . The elliptic area is the attained set of the problem.
The solid line is the Pareto optimal set. Points A and B are in the Pareto
optimal set, points C and D are not.
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 104
If a point x is not in the Pareto optimal set, there always exists a point S y ∈ that is
dominant over x . If a point x is in the Pareto optimal set, there is no solution that is
better than x in all attributes. Therefore, it is rational to choose the final solution for a
MOP from its Pareto optimal set.
7.2 Solution Techniques for Multiobjective
Optimization Problems
The solutions in a Pareto optimal set are not dominant over each other. In order to
determine the final solution for a MOP, subjective judgement or preference information
from the decision maker is required. The preference information is used to convert a
MOP to one or a series of scalar optimization problems. According to Hwang et al.
(1980), the solution methods for MOP can be classified into four categories depending on
when the preference information of the decision maker is articulated in the process of
multiobjective decision making: (1) nopreference methods, (2) a priori methods, (3)
interactive methods, and (4) a posteriori methods. Table 7.1 lists a few major methods
for each category.
Table 7.1: A classification of methods for multiobjective optimization
(after Huang et al., 1980).
NoPreference Methods Global Criterion Method (Salukvadze, 1974)
A Priori Methods
Utility Functions (Keeney and Raiffa, 1976)
Lexicographic Method (Fishburn, 1974)
Goal Programming (Charnes and Cooper, 1961)
Hierarchical Method (Azarm, 2002)
Interactive Methods
Method of ZiontsWallenius (Wallenius, 1975)
STEM method (Benayoun et al., 1971)
Interactive MOLP (Steuer, 1977)
A Posteriori Methods
Weighting Functions Method (Gal and Nedoma, 1972)
ε constraint method (Haimes et al., 1975)
Section 7.2.17.2.4 gives a brief description of each category. Methods investigated
in this study are described in detail where appropriate. The materials of Section 7.2.1
7.2.4 are based on Hwang et al. (1980) and Azarm (2002).
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 105
7.2.1 NoPreference Methods
Nopreference methods require no subjective preference information from the decision
maker. Multiple objectives are aggregated into one objective function without any
opinions of the decision maker. Then a scalar optimization problem is solved and the
solution is presented to the decision maker who may accept or reject the solution. The
advantage of this approach is that the decision maker is not disturbed in the solution
process and the method is easy to use. The disadvantage is that the decision maker has no
control over the solution quality.
For example, in the method of Global Criterion (Salukvadze, 1974), the objective
vector is aggregated into a scalar objective function in such a way that the deviation of
the objective vector from an ideal objective vector is minimized:
( )
( )
∑
=
−
=
k
i
p
i
i i
f
f f
f
1
*
*
x
x (7.4)
where
*
i
f is the ideal value of the i th objective and is usually obtained by minimizing
i
f
individually while ignoring the rest objectives. Boychuk and Ovchinnikov (1973)
suggests 1 = p while Salukvadze (1974) recommends 2 = p .
7.2.2 A Priori Methods
A priori methods require that the preference information from the decision maker is given
before the multiobjective optimization is conducted. Thus the decision maker must have
some a priori understanding of the optimization problem.
Utility function methods. In these methods, a utility function is used to express the
decision maker’s preference information over the objectives:
( ) ( ) ( ) ( ) ( ) ( ) x x x x F
3 2 1
,..., , f f f U U = (7.5)
where ( ) F U is a utility function of the multiple objectives. The advantage of the utility
function methods is that once the utility function is appropriately formulated, the most
satisfactory solution of the MOP can be obtained. However, even for a simple problem,
the utility function is very difficult to construct.
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 106
A special form of the utility function is to use weighting coefficients to indicate the
importance of each objective:
( ) ( ) ( )
∑
=
=
k
i
i i
f w U
1
x x F (7.6)
Methods using utility function given by Eq. 7.6 are also called the weighted sum methods.
Goal Programming. Goal Programming (GP) was first developed by Charnes and
Cooper (1961). For this method, the decision maker must construct a set of goals and
rank the importance of the goals. Problem 7.1 is formulated in GP as follows:
minimize ( ) ( ) ( ) [ ]
+ − + − + −
d , d d , d d , d
l l
h P h P h P ,..., ,
2 2 1 1
(7.7a)
subject to ( )
j j j j
b d d c = − +
+ −
x , . ,..., 2 , 1 m j = (7.7b)
( )
i i i i
b d d f = − +
+ −
x , k i ,..., 2 , 1 = (7.7c)
0 , ≥
+ −
i i
d d for all i (7.7d)
0 = •
+ −
i i
d d for all i (7.7e)
where
j
b is the aspiration level of the j th goal,
−
j
d is the negative deviation from the
j th goal,
+
j
d is the positive deviation from the j th goal, ( )
+ −
d , d
j
h is a function of the
deviation values called the th j achievement function, and
j
P is the weighting
coefficients for the j th achievement function. The achievement functions are absolutely
ordered, which means
1 +
>>
j j
P P and there does not exist a number N that makes
1 + j
NP
greater than
j
P .
The solution procedure for GP goes as follows:
1. Minimize the first achievement function
1
h , and let
*
1 1
min h h = . Let 1 = j .
2. Let 1 + = j j . Minimize the th j achievement function
j
h while requiring
*
i i
h h ≤ for
1 ,..., 1 − = j i . That is, when minimizing an achievement function, achievement
functions with higher rank can not be deteriorated. Denote the optimal value of the
th j achievement function as
*
j
h .
3. Repeat Step 2 until all achievement functions are minimized.
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 107
Hierarchical optimization method. This method (Azarm, 2002) allows the decision
maker to rank and minimize the objectives in descending order of importance. Suppose
the k objectives are ordered from
1
f (most important) to
k
f (least important), the
solution procedure can be illustrated as follows:
1. Find the optimum point
1 *,
x for
1
f , subject to the original set of constraints
minimize ( ) x
1
f (7.8a)
subject to ( ) 0 ≤ x
i
c , m i ,..., 1 = (7.8b)
Let ( )
1 *, *
1
x f denote the optimal objective function value for Problem 7.8. Let 1 = j .
2. Let 1 + = j j . Find the optimum point
j *,
x for the j th objective function
j
f subject
to the original and an additional set of constraints
minimize ( ) x
j
f (7.9a)
subject to ( ) 0 ≤ x
i
c , m i ,..., 1 = (7.9b)
( ) ( ) ( )
*,i
i i i
f f x x
*
1 ε + ≤ , 1 ,..., 1 − = j i (7.9c)
where 0 ≥
i
ε is the coefficient of function increment for
i
f . Denote the optimal
objective value of Problem 7.9 as ( )
j
j
f
*, *
x .
3. Repeat Step 2 until j reaches the total number of objectives k .
7.2.3 Interactive Methods
For interactive methods, the decision maker progressively articulates preference
information by interacting with the solution process. The decision maker controls the
solution process by updating his preference as he learns about the problem. The
advantage of these methods is that no a priori information is required from the decision
maker. As the solution process goes, the decision maker learns about the problem and
guides the solution process according to what has been learned. The solution process is a
learning procedure for the decision maker and the decision maker is likely to accept the
solution. The disadvantage of this method is that a great deal of effort is required from
the decision maker.
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 108
Interactive methods often have a calculation phase and a decision phase. The
calculation phase gathers the tradeoff information, and the decision phase uses the trade
off information to guide the next search step. For example, see the methods of Zionts
Wallenius and the STEM methods reviewed in Huang et al. (1980).
7.2.4 A Posteriori Methods
A posteriori methods require no preference information from the decision maker before
and during the solution process. The methods construct a subset of Pareto optimal from
which the decision maker makes a subjective judgement and selects the most satisfactory
solution. The advantage of this method is that the decision maker usually has a better
understanding of the tradeoffs because he/she has a discrete sampling of the Pareto
optimal set. However, the computational load of this method is heavy, especially when
there are more than two objective functions. Another disadvantage is that the decision
maker may have too many solutions to choose from.
To construct a discrete sampling of the Pareto optimal set, the common theme of this
class of methods is to vary systematically the algorithm parameters expressing the
decision maker’s preference information. For example, the ε constraint method (Haimes,
1975) chooses one objective for minimization and reformulates other objectives to
constraints whose righthand side is a usercontrolled variable ε . Solving the resulting
scalar optimization problems for various ε will generate a set of Pareto set of optimal
solutions. The Hierarchical method (Azarm, 2002) described in Section 7.2.2 is able to
generate an evenspread sampling of the Pareto optimal set by varying parameter
i
ε .
Evolutionary multiobjective optimization algorithms (Fonseca and Fleming, 1993) try to
evolve a set of solutions in such a way that the set of solutions spread to the Pareto
optimal front.
The weighting functions method. The weighting functions method (Gal and
Nedoma, 1972) samples the Pareto optimal set by changing the weighting coefficients of
the weighted sum method:
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 109
max ( ) ( )
∑
=
=
k
i
i i
f w f
1
x x (7.10a)
subject to S x ∈ (7.10b)
The weighting functions method can be interpreted geometrically. Consider a MOP
problem with two objectives
1
f and
2
f . The weighting function
2 2 1 1
f w f w f + = defines
a line L with slope
2 1
/ w w − in the objective space (Figure 7.2). Minimization of
function f can be seen as moving line L towards the origin while keeping the
intersection of line L and the feasible region Y. The optimum point is obtained when
line L is tangent to the boundary of the feasible region.
If the Pareto optimal set is not convex, some points in the set can not be allocated by
the weighting function method no matter what weights are used. This is illustrated in
Figure 7.3. The line tangent to point A can be moved further to points B and C. Thus
point A can not be allocated by the weighting function method.
Y
L
1
f
2
f
L
2 1
/ slope w w − =
Figure 7.2: The weighting functions method with convex set (based on
Huang et al. (1980) and Azarm et al. (2002)).
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 110
A
1
f
2
f
2 1
/ slope w w − =
B
C
L
L
Y
Figure 7.3: The weighting functions method with nonconvex set (based on
Huang et al. (1980) and Azarm et al. (2002)).
7.3 Multiobjective Optimization of Production
Operations
Production operations in a petroleum field aim to maximize profit from the field subject
to restrictions of facility capacities, commercial contract, safety and environmental
regulations, and other considerations. In optimization terminology, this task breaks down
to the following components:
• Objectives. For daytoday production operations, the objectives typically include
maximizing daily oil production rate to meet contract target, and minimizing lift gas
rate to minimize production cost, etc.
• Hard constraints. Hard constraints refer to constraints that are well defined and
cannot be violated, such as the fluid processing capacities of surface facilities.
• Soft constraints. Soft constraints refer to constraints that are not well defined and/or
can be violated, such as the maximum flow rate or pressure limit on individual wells
or the target of the total oil rate from the field. Soft constraints can be viewed as
objectives.
The soft constraints need to be formulated in mathematical language so that they can
be handled in the optimization process.
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 111
7.3.1 Handling Soft Constraints
In this study, the soft constraints were handled in a way similar to how the goals are
handled in goal programming (Charnes and Cooper, 1961). The soft constraints can be
regarded as one type of goals and are converted to a set of hard constraints and an
objective function by introducing some deviation variables. Consider the following set of
soft constraints
( )
i i
b c ≤ x , m i ,.., 1 = (7.11)
where ( ) x
i
c is a soft constraint function and
i
b is a limit value. Constraint Eq. 7.11 can
be converted to the following optimization problem, which we will call Problem 7.12:
minimize ( ) d h (7.12a)
subject to ( )
i i i
b d x c ≤ − , m i ,.., 1 = (7.12b)
0 ≥
i
d , m i ,.., 1 = (7.12c)
where
i
d is a deviation variable and h is a function of the deviation variables. A simple
example of h is the weighted sum of the deviation variables
( ) ( )
∑
=
i
i i
d w h d (7.13)
where
i
w is a usersupplied weighting coefficient.
Soft constraints with the following forms can be treated in a similar manner
( )
i i
b c ≥ x or ( )
i i
b c = x (7.14)
7.3.2 Solution Methods
This section presents a framework to formulate and solve the multiobjective production
optimization problems. The actual implementation is limited to the nonlinear rate
allocation problem described in Chapter 5.
Decision variables. It is necessary to specify which wells’ lift gas rates are to be
optimized. The oil flow rate of every production well and deviation variables for soft
constraints are automatically selected as decision variables. The lower and upper bounds
and the initial values for all decision variables are specified in advance.
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 112
Hard constraints. Multiple flow rate constraints can be specified for any well or
network node.
Soft constraint. Multiple soft (oil, gas, water flow rate) constraints can be specified
for any well or network node. The whole set of soft constraints can be divided into
several subsets. Each subset can be converted to a set of hard constraints and an objective
function defined by Eq. 7.13 (this objective function is considered as an objective
variable below).
Objectives. The problem may include multiple objectives. Each objective is a
weighted sum of a set of objective variables:
( )
∑
=
i
i i j
y w f y (7.15)
where
j
f denotes the j th objective,
i
y denotes an objective variable, and
i
w denotes a
weighting coefficient. An objective variable can be a single decision variable or a
function of multiple decision variables. For example, an objective variable can be oil, gas,
and water flow rates, or the pressure of a well or a network node, or a function of some
deviation variables defined by Eq. 7.13.
Solution method. The primary solution method investigated in this study was the
Hierarchical optimization method (Azarm, 2002) described in Section 7.2.2. To use this
method, the user needs to rank the k objectives in descending order of importance and
specify a relaxation parameter
i
ε for each of the first 1 − k objectives.
The Hierarchical method was adopted because
1. The method is easy to use and its logic is easy to understand for the decision maker.
2. The method is able to locate a Pareto optimal solution that resides in the nonconvex
part of the Pareto optimal set.
The proposed framework is flexible. The Hierarchical method can be converted to
other popular multiobjective methods by manipulating the algorithmic parameters. For
example, the method reduces to the Lexicographic method (Fishburn, 1974) if all
i
ε are
set to zero; the method reduces to the goal programming method (Charnes and Cooper,
1961) if all the objectives are defined as soft constraints and all
i
ε are set to zero; and the
method reduces to the weighting functions method if all objective variables are
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 113
aggregated into one objective function using Eq. 7.15. The Hierarchical method can be
used as an a priori method, an interactive method, or an a posteriori method depending
on how the adjustable algorithmic parameters (weighting coefficients
i
w and relaxation
parameters
i
ε ) are adjusted in the optimization process (no adjustment, adjusted
interactively by the decision maker, or adjusted by the analyst to generate a sampling of
the Pareto optimal set, respectively).
7.4 Example
The production system is the same as the one presented in Section 3.7. The configuration
of the gathering system is illustrated in Figure 3.4.
Suppose that the produced gas is not for sale and it is to be reinjected into the
reservoir for storage. Let us assume that the production engineer wants to operate the
field with the following goals:
1. Produce the field with a total oil production rate
t o
q
,
of exactly 14,000 STB/D.
2. Minimize the total water flow rate
t
w
q .
3. Minimize the total gas flow rate
t
g
q .
Since the field is able to produce at an oil flow rate higher than 14,000 STB/D, the first
goal is to formulate the total oil production target as a hard constraint. The other two
goals are formulated as two objectives.
The cost of each barrel of produced water and cubic feet of produced gas depends on
many factors and may be hard to evaluate quantitatively. Therefore it is difficult to come
up with utility functions to aggregate the last two objectives into one objective function.
An appropriate solution is to construct a subset of the Pareto optimal set for this problem
from which the production engineer can choose the most satisfactory solution.
Both the hierarchical method and the weighting functions method were used to
construct the Pareto optimal set. The solution procedure of the hierarchical method goes
as follows:
1. Minimize the total water flow rate of the field with the following constraint:
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 114
000 , 14 =
t
o
q STB/D (7.16)
Denote the minimum water flow rate value as
,* t
w
q . Let 1 = j .
2. Minimize the total gas flow rate of the field subject to following constraints:
000 , 14 =
t
o
q STB/D (7.17)
( )
,*
1 . 0 * 1
t
w
t
w
q j q + ≤ (7.18)
3. Repeat Step 3 for ,... 3 , 2 = j until the total gas flow rate no longer decreases as j
increases.
In the weighting functions method, the multiobjective optimization problem was
transformed to the following scalar optimization problem
minimize
t g g t w w
q w q w f
, ,
+ = (7.19a)
subject to 000 , 14 =
t
o
q (7.19b)
Here the weighting pair ( )
g w
w w , was not used to indicate cost of the produced water or
gas, but was varied to generate a subset of the Pareto optimal solutions. The weighting
pairs and the corresponding solutions generated for this problem are shown in Table 7.2.
Table 7.2: Pareto optimal solutions from the weighting functions method.
Weighting Pair (1,1) (3,1) (3.5,1) (5,1) (10,1) (15,1) (15.5,1) (20,1)
Total Water Rate
(STB/d)
10310 10074 8709 8653 8583 7246 6904 6354
Total Gas Rate
(MSCF/d)
26419 26829 31058 31295 31834 50192 55404 63923
Both methods obtained roughly the same Pareto optimal set (Figure 7.4). However, it
is much easier to control the sampling point with the hierarchical method than with the
weighting functions method, because for the weighting functions method, there is no
clear relation between the weighting pair and the Pareto optimal solution. Furthermore,
the weighting functions method is unable to locate the nonconvex parts of the Pareto
optimal set, such as the curve segment roughly between 9000 STB/D and 10,000 STB/D
of water flow rate (Figure 7.4). Thus, the hierarchical method is more appropriate for this
problem than the weighting functions method.
Once the complete set of Pareto optimal set is sampled, the production engineers can
look at the tradeoffs in the sampled set and make their operational decisions according to
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 115
their preference. Note that the Pareto optimal set only needs to be constructed once, no
matter what preferences the production engineers have. This is an advantage of the a
posteriori methods over the a priori methods.
20000
30000
40000
50000
60000
70000
6000 7000 8000 9000 10000 11000
TOTAL WATER RATE, STB/D
T
O
T
A
L
G
A
S
R
A
T
E
,
M
S
C
F
/
D
WEIGHTED SUM
HIERARICHAL
Figure 7.4: The Pareto optimal sets obtained from the Hierarchical method
and the weighting functions method.
7.5 Concluding Remarks
When the multiple objectives can be appropriately aggregated into one objective function,
i.e., the net present value (NPV), the utility function method is able to obtain the most
satisfactory solution efficiently. However, sometimes, the utility function is difficult to
construct. It appears that the weighted sum method described in Section 7.2.2 or the
weighting functions method described in Section 7.2.4 is often used in the upstream oil
industry to handle multiple objectives. However, as demonstrated in Section 7.4, certain
multiobjective optimization problems in a petroleum field can have nonconvex Pareto
optimal solutions, and the weighted sum method may fail to obtain solutions of interest.
The hierarchical method, on the other hand, does not require a utility function, and is
able to sample the complete Pareto optimal set. However, as the number of objectives in
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 116
an MOP increases, the computational demand of the hierarchical method may quickly
becomes unmanageable, and the multidimensional Pareto optimal set get hard to visualize.
Perhaps the more useful application of multiobjective optimization to the upstream
oil industry is to improve the design of a production system, because in this case there are
more tradeoffs involved and the impact of the tradeoffs is more significant in such
design problems than in production operation problems. The evolutionary multiobjective
optimization methods (Fonseca and Fleming, 1993) is a promising method for such tasks
because of the following reasons:
• The optimization problem of production system design has rough surfaces.
Evolutionary algorithms are more appropriate than the gradientbased methods for
such problems (Palke and Horne, 1997a).
• Evolutionary algorithms maintain multiple solutions and have the capability to
sample various Paretooptimal solutions in parallel (Oyama and Liou, 2002). `
117
Chapter 8
Coupling with a Reservoir Simulator
8.1 Introduction
The optimization methods investigated in this study require numerical tools to evaluate
objective and constraint functions. For research purposes (such as evaluating the
performance of the optimization algorithms without worrying about model accuracy), in
Chapter 4 and 5 the objective and constraint functions were evaluated by a network
simulator assuming that the required reservoir conditions around each well are given. For
practical use, the optimization methods can be implemented in the following ways:
1. Integrating into a field simulator. In this approach, the optimization methods can be
implemented in a field simulator that is capable of simulating both reservoirs and
gathering systems. The function evaluation procedure of the optimization method can
be implemented within the simulator. The advantage of this approach is that the
optimizer is easy to use. If the engineer has built a simulation model for a field, it
requires little extra effort to extend the simulation model for production optimization.
The disadvantage of this approach is that its implementation requires access to the
source code of the field simulator, which is not always possible.
2. Interfacing with a reservoir simulator. In this approach, the optimization methods are
implemented in a standalone optimizer as in Chapter 4 and 5. The optimizer is then
interfaced with a reservoir simulator, which feeds the optimizer the required reservoir
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 118
conditions. In contrast to the first approach, this approach is easy to implement, but
the coupled system may be cumbersome to use.
3. Integrating with inhouse software. For instance, the optimization methods proposed
as a result of this study can replace heuristic rules of an existing optimization system
developed for a specific field.
In the course of this study, some of the optimization tools developed in previous
chapters were integrated into VIPEXECUTIVE (Landmark, 2001), a commercial field
simulator. This allows the optimization tools that are found to be most suitable to be
applied to a large variety of petroleum fields. In particular, The VIPEXECUTIVE with
the optimization tools is an integral part of the EField Optimization System (EFOS)
developed at Prudhoe Bay oil field in Alaska (Litvak et al., 2002). EFOS is designed to
maximize the daily oil production in Prudhoe Bay field by optimally allocating the
production rates, well connections to flowlines and (potentially) lift gas subject to
multiple capacity constraints of facilities and velocity constraints in flowlines.
This chapter describes how the optimization tools can be integrated into/interfaced
with a reservoir simulator (using VIPEXECUTIVE as an example) and presents several
examples of the integrated system.
To facilitate our discussion of the integration/interfacing procedure, we first discuss
how an integrated reservoir and production network (tubing strings and surface pipeline
systems) model can be solved.
8.2 Full Field Simulation
Traditionally simulation models for reservoirs and production networks have been
developed independently. Various methods have been proposed to couple the reservoir
model and the network model for full field solutions (Emanuel and Ranney, 1981; Litvak
and Darlow, 1995; Schiozer, 1994; Hepguler et al., 1997; Byer, 2000). Full field
simulation offers increased accuracy in predicting the reservoir deliverability and is
critical for many reservoir development and management studies. In this section, we
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 119
describe first the governing equations and then the existing solution techniques for full
field simulations.
8.2.1 Simulation of Petroleum Fields
In the simulation of petroleum fields, fluid properties can be represented by a
compositional model or a blackoil model. Consequently, the variables and the governing
equations for the field simulation are different. The following discussion assumes a
compositional model, for the blackoil model can be regarded as a special case of a
compositional model.
The Field simulator models the multiphase flow in an integrated system of
reservoirs, wellbores, well tubing strings, surface pipeline systems, and possibly the fluid
separation units and other facilities. A field simulator typically is capable of obtaining the
following information (Litvak and Darlow, 1995):
1. In gridblocks of a discretized reservoir model we obtain
• pressures of each phases,
• oil, gas, and water saturations, and
• composition of hydrocarbon phases.
2. In production and injection wells we obtain
• bottomhole pressures (or flow rates) and
• molar rates of the hydrocarbon components.
3. In well tubing strings and surface pipeline systems
• pressure distribution,
• oil, water, and gas rates for every link and node (the notation of link and node
used in this chapter is defined in Chapter 3),
• molar rates of the hydrocarbon components for every link and node.
This information is determined from the simultaneous solution of the following
governing equations:
1. A system of equations that describe multiphase flow in porous media of an
underground reservoir. These equations include mass balance equations, phase
equilibrium relations, capillary pressure relations, saturation constraints, and phase
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 120
constraints. In reservoir simulation, these equations are classified into primary and
secondary equations and the variables are classified as primary and secondary
variables
( ) 0 , =
s p p
x x F (8.1)
( ) 0 , =
s p s
x x F (8.2)
where subscript p represents primary, s represents second. The classification is not
unique. As an example, in Coats model (Coats, 1980),
p
F represents the mass balance
equations for each component in each grid block.
p
x represents the gridblock
pressure, oil and water saturations, and the hydrocarbon component compositions of
each gridblock. The concrete form of Eqs. 8.1 and 8.2 for various simulation models
can be found in Aziz and Durlofsky (2002) and Cao (2002).
2. Wellbore flow equations that relate the well flow rate of component c ,
w
c
q , with the
bottomhole pressure
w
p
( )
w w
c
w
c
p q q = (8.3)
The concrete form of Eq. 8.3 for various well models can be found in Aziz and
Durlofsky (2002).
3. Multiphase flow models that relate the flow rate with the pressure drop across a
tubing string:
( )
w g o
q q q p p p , ,
, 1 2 2
= (8.4)
where
1
p and
2
p denote the upstream and downstream pressure of a tubing string,
and
o
q ,
g
q , and
w
q denotes the oil, gas, and water flow rate across that tubing string,
respectively. Eq. 8.4 can be in the form of numerical multiphase flow models or
hydraulic lookup tables.
4. The following system of equations that describe multiphase flow in a surface pipeline
network:
• Mass balance equation for each component at every node
∑
Ω ∈
=
n
i
j
cij
q 0 ,
n
n i ,..., 1 = ,
c
n c ,..., 1 = (8.5)
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 121
where
cij
q denotes the flow rate of component c between node i and j .
n
i
Ω
denotes all the adjacent nodes for node i ,
n
n is the number of node, and
c
n is the
number of components.
• Kirchoff’s law, which requires that the pressure for a node should be the same no
matter from which path it is computed.
• Flow equations that relate the flow rate and pressure drop through a link (i.e., a
pipe, a choke, or a pump, etc.)
( )
w g o
q q q p p p , ,
, 1 2 2
= (8.6)
where
1
p and
2
p denote the upstream and downstream pressure of a link.
A reservoir simulator usually models multiphase flow in a reservoir up to the
wellbores. The selected set of equations are solved by the NewtonRaphson method. The
basic steps in a reservoir simulation can be described as follows:
1. Initialize the reservoir model.
2. March forward a time step.
2a. Form a set of system equations by constructing Eq. 8.1 for every gridblock and
Eq. 8.3 for every well.
2b. Linearize the set of system equations.
2c. Solve the linearized system equations to obtain updated estimates of the primary
variables.
2d. Update the secondary variables using Eq. 8.2 for every gridblock and Eq. 8.3 for
every well.
2e. Repeat Step 2a2e until converge.
3. Repeat Step 2.
A network simulator simulates multiphase flow in well tubing strings and surface
pipeline systems by treating production wells as sources with fixed pressures or flow
rates. Sometimes a network model also contains wellbore models. A network simulation
involves solving a set of nonlinear equations constructed from Eqs. 8.38.6 and
Kirchoff’s law. Chapter 3 presents a solution procedure for network simulation.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 122
Full field simulators combine the reservoir model and the network model. To utilize
fully existing techniques developed for reservoir simulation and network simulation,
solution techniques developed for full field simulation focus on how to couple the
reservoir model and the network model at the timestep level. Various coupling methods
have been proposed in the literature and they are reviewed in the next section.
8.2.2 Coupling Reservoir and Production Network Solutions
The coupling methods proposed in the literature can be classified into implicit and
explicit procedures.
Algorithm 8.1: Standard implicit coupling. This procedure combines the system of
equations for the reservoir model and the network model and solves them simultaneously
using the NewtonRaphson method. Litvak and Darlow (1995) proposed such a
procedure for a compositional full field model. The steps in this procedure are described
below.
1. Solve linearized system of equations for the full field model. Linearize and solve the
flow equations in the reservoir, wellbores, and the surface pipeline system to obtain
an estimate of the primary variables.
2. Update reservoir conditions. Update pressure, saturations, and component
compositions in reservoir gridblocks using the solution from Step 1 and the secondary
equations.
3. Repeat Steps 12 until convergence.
4. March to the next time step.
This procedure requires that the reservoir model and the network model are
implemented in one simulator. Schiozer (1994) and Byer (2002) investigated a similar
implicit coupling procedure for a blackoil model.
Algorithm 8.2: Iterative coupling. This procedure alternately solves a network
problem and a reservoir problem until the system of equations for the full field model are
satisfied. The following procedure was proposed by Litvak and Darlow (1995) for a
compositional model (this method was named as an “explicit procedure” in Litvak and
Darlow (1995)):
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 123
1. Solve a network problem. Form a network problem by taking reservoir conditions in
well gridblocks from the previous iteration (or the previous time step at the first
iteration) as part of the boundary conditions for the network problem. Solve the
network problem and obtain well production rates and bottomhole pressures.
2. Take a Newton iteration for the reservoir model. Taking the bottomhole pressures or
production rates from Step 1 as the well constraints, linearize and solve the reservoir
flow equations. Update pressure, saturation, and component compositions in reservoir
gridblocks using the solution of the linearized flow equations.
3. Repeat Steps 12 until convergence. The convergence criterion is that the production
rates and bottomhole pressure obtained from Step 1 and Step 2 are within a specified
tolerance for every well.
4. March to the next time step.
The advantage of this procedure is that it can be used to couple a reservoir simulator
with an independent network simulator. However, for certain cases, this procedure may
not converge or converge to a wrong solution (Litvak and Darlow, 1995). Similar
procedures were also presented by Hepguler et al. (1997), Emanuel and Ranney (1981),
and Breaux et al. (1985).
Algorithm 8.3: Explicit coupling. This procedure performs network simulation only
once, based on reservoir conditions at the beginning of a time step. Consequently the
solution may not satisfy the system equations of the full field model. However, this
approach is the easiest to implement and the cheapest to run among all coupling methods.
1. Solve a network problem. Form a network problem by taking reservoir conditions in
well gridblocks from the previous time step as part of the boundary conditions of the
network problem. Solve the network problem and obtain well production rates and
bottomhole pressures.
2. Solve a reservoir problem. Taking the bottomhole pressures or production rates from
Step 1 as the well constraints, solve a reservoir simulation problem as described in
Section 8.2.1.
3. March to the next time step.
This procedure was investigated by Schiozer (1994) and Byer (2000), and was used
by Emanuel and Ranney (1981) to study an offshore reservoir.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 124
8.3 Integration of Optimization Tools with VIP
EXECUTIVE
Some of the optimization tools we investigated were integrated with a commercial field
simulator, VIPEXECUTIVE (Landmark, 2001). This section first presents a general
procedure for integrating production optimization tools into VIPEXECUTIVE, then
describes the integrated optimization tools and some related issues.
8.3.1 A General Integration Procedure
As a reservoir development and management tool, VIPEXECUTIVE contains certain
reservoir management features, such as well controls and gaslift allocation, etc. The role
of these management features in a field simulator is the same as those of our optimization
tools. Therefore, the optimization tools can be integrated into the solution procedure of
the field simulator in the same way as these reservoir management features are integrated.
Algorithm 8.4: Integration procedure. The optimization tools are integrated with
the standard fully implicit coupling procedure of VIPEXECUTIVE at the Newton
iteration level (Figure 8.1).
1. Perform well management routines (optional). Based on user specifications and
estimated field conditions (well water cut, GOR, total field flow rates, etc.), invoke
well management routines that handle well constraints and gaslift allocation. The
results of this step may include the following:
• Updated well status. Some wells are closed. Some wells are reopened. Some wells
are worked over. Some wells are reconnected to different pipeline systems, etc.
• Updated well constraints. For example, a well should not produce at an oil flow
rate higher than 3000 STB/D to satisfy a total field gas flow rate constraint.
• Updated lift gas rate for gaslift wells.
• Updated injection rate for injection wells.
2. Convert pressure constraints (optional). Based on current reservoir conditions,
invoke a procedure proposed by Litvak and Darlow (1995) to convert pressure and
flow rate constraints on network nodes to flow rate constraints on individual wells. If
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 125
Step 3 is to be invoked, only pressure constraints are converted. The flow rate
constraints on network nodes are handled in Step 3.
3. Perform production optimizations (optional). Perform userspecified production
optimizations with current reservoir conditions. The results of this step may include
• Updated well constraints.
• Updated lift gas rate for gaslift wells.
• Updated well connections.
4. Solve linearized system equations for the full field. Linearize and solve the flow
equations in the reservoir, wellbore, and surface pipeline systems to obtain an
estimate of the primary variables. The updated operation settings from Step 1 and 2
are incorporated in this step.
5. Update reservoir conditions. Update pressure, saturations, and component
compositions in reservoir gridblocks using the solution from Step 1.
6. Repeat Steps 15 until convergence.
7. March to the next time step.
Perform well management
routines
Optimize Production
Operation Settings
Converged?
Optimize?
Previous time step
Next time step
Yes
No
Determine active
constraints
Form and solve full field
system equations
Update reservoir conditions
and network conditions
Yes
Convert pressure constraints to
flow rate constraints
Perform well management
routines
Optimize Production
Operation Settings
Converged?
Optimize?
Previous time step
Next time step
Yes
No
Determine active
constraints
Form and solve full field
system equations
Update reservoir conditions
and network conditions
Yes
No
Convert pressure constraints to
flow rate constraints
Perform well management
routines
Optimize Production
Operation Settings
Converged?
Optimize?
Previous time step
Next time step
Yes
No
Determine active
constraints
Form and solve full field
system equations
Update reservoir conditions
and network conditions
Yes
Convert pressure constraints to
flow rate constraints
Perform well management
routines
Optimize Production
Operation Settings
Converged?
Optimize?
Previous time step
Next time step
Yes
No
Determine active
constraints
Form and solve full field
system equations
Update reservoir conditions
and network conditions
Yes
No
Convert pressure constraints to
flow rate constraints
Figure 8.1: Integration of optimization tools to VIPEXECUTIVE.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 126
8.3.2 Integrated Optimization Tools
The following optimization tools were integrated into VIPEXECUTIVE:
1. the LPI method for rate allocation (Section 4.3.1),
2. the LPII method for rate allocation (Section 4.3.4), and
3. the partial enumeration method and the genetic algorithm for well connection
optimization (Chapter 6).
Specific integration issues for each optimization tool are discussed below.
The LPII method for rate allocation. To ensure that the constraints will not be
violated at the end of a time step, a rate allocation problem is solved by the LPII method
at every Newton iteration of the solution procedure for the full field model. The
following procedure corresponds to Step 2 and 3 of Algorithm 8.4.
1. Convert pressure constraints on wells and network nodes to flow rate constraints on
individual wells using a procedure proposed by Litvak and Darlow (1995). The
weighting coefficients
v
o
c ,
v
g
c , and
v
w
c for each velocity constraint (see Section 4.4.1)
are determined during this procedure.
2. Use the following information to build the LPII model.
• the oil rate, GOR, and water cut for each well from Step 1,
• the weighting coefficients
v
o
c ,
v
g
c , and
v
w
c from Step 1, and
• the userspecified constraint limits and weighting coefficients for the objective
function.
The speedup techniques described in Section 4.4.2 are used to reduce the size of the
LPII model.
3. Solve the LP problem formulated in Step 2.
4. Update the well constraints for the next Newton iteration. For every well i , if the
scaling factor 0 =
i
x , well i is closed in the next iteration; if 1 =
i
x , well i remains
fully open in the next iteration, and its oil rate is allowed to increase at the next
iteration; if ( ) 1 , 0 ∈
i
x , the oil rate of well i is not allowed to exceed
max
,i o i
q x at the next
iteration, where
max
,i o
q is the maximum oil rate of well i used to build the LP model in
Step 2.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 127
In Step 4, once a well rate is scaled, it is not allowed to exceed its scaled value in the
next iteration. Consequently in later iterations fewer constraints may be violated and the
speedup techniques (the constraint elimination procedure described in Section 4.3.2) will
be more effective.
At each Newton iteration, the LPII method optimizes rate allocation based on
updated reservoir conditions. Therefore, in a time step, the nonlinear interactions in the
reservoir are partially taken into account by solving a rate allocation problem at every
Newton iteration.
The LPI method for rate allocation. This method is coupled into VIP
EXECUTIVE in the same manner as the LPII method except
1. to build Model LPI, the well performance curves have to be constructed, and
2. to reduce the computational load, the method is not invoked at all Newton iterations.
The user can specify at which Newton iterations to invoke the LPI method.
Well connection optimization. Both the partial enumeration method and the genetic
algorithm described in Chapter 6 are implemented. The rate allocation problem within
both algorithms are solved by the LPII method.
Well reconnections can significantly change the flow behavior in the system. Thus
performing well connections between iterations can hinder convergence of the solution
procedure of the field simulator. In addition, well connection optimizations are time
consuming. For these reasons, in the current implementation, the user can specify in
which Newton iterations well connection optimization will be performed.
8.4 Interfacing with an Independent Reservoir
Simulator
When the optimization tools are developed as a standalone package without reservoir
models, the optimization package can be interfaced with a reservoir simulator to do full
field simulation and production optimization. The optimization package and the reservoir
simulator can be coupled together using the iterative coupling procedure or the explicit
coupling procedure described in Section 8.2.2.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 128
8.5 Applications
The optimization system described in Section 8.3 can be used for reservoir development
studies or for onsite real time production control and optimization. This section presents
three examples: two for longterm reservoir prediction/development studies and one for
onsite production optimization.
The first example predicted the oil production of an oil field in the Gulf Of Mexico
(GOM) for five years. At each time step, the daily oil production from the field was
maximized by allocating the production rates and the well connections. The second
example predicted the oil production of another oil field in the GOM for the whole life of
the field. At each time step, the daily oil production from the field was maximized by
optimally allocating production and lift gas. The third example describes how the
optimization tools can be applied to the giant Prudhoe Bay oil field in Alaska.
8.5.1 A Gulf of Mexico Oil Field Example
An integrated reservoir and surface pipeline network model had been built for one of the
Gulf of Mexico oil fields. Our test example was based on this reservoir. This example has
been presented in Wang, Litvak, and Aziz (2002a).
Five production wells, A1A5, are tied to a 40slot conventional fixedleg platform.
Ten production wells were drilled from a 10slot subsea template and connected to the
platform by two 22000 foot long pipelines. On the platform, a threestage separator
processes the produced fluids. Wells A1A5 can be tied to any stage of the separator.
Wells B1B10 can be connected to either the lowpressure stage or the highpressure
stage of the separator. Figure 8.2 illustrates the production system.
For evaluation purposes, we imposed four hypothetical constraints on the production
system. The separator has a maximum gas handling capacity of 30 MMSCF/D and a
maximum of water handling capacity of 5000 STB/D. Moreover, the separator cannot
receive gas from either pipeline at a rate of more than 12 MMSCF/D. Our job was to
maximize the oil production by selecting appropriate well rates and well connections.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 129
Oil
Gas
Water
B01
Separator
lp hp lp ip hp
TEMPLATE
HEADER_Y HEADER_Z
B02 B03
B04
B05
B06
B07
B09
B10
A5
A1
A2
A3
A4
B08
Figure 8.2: The production system of a Gulf Of Mexico oil field.
We made three predictive runs using VIPEXECUTIVE. Each run predicted the oil
production for about five years. In Run 1, at every Newton iteration of each time step, we
checked flow rate constraints sequentially and scaled well rates if necessary: first we
checked the two gas flow rate constraints on HEADER_Y and HEADER_Z, if either
constraint was violated, the production rates of the highest GOR wells connecting to
HEADER_Y (or HEADER_Z depending on which constraint was violated) were scaled;
then we checked the gas flow rate constraint on the separator, if it was violated, the
highest GOR wells of the whole field were scaled to meet the gas constraint; finally we
checked the water constraint on the separator, if it was violated, the highest water cut
wells of the whole field were scaled to meet the water constraint. This procedure is a well
management option available in VIPEXECUTIVE (Landmark, 2001). In Run 2, we
applied the LPII method described in Section 4.3.4 to select the well rates. In Run 3, we
applied the partial enumeration method with the LPII method to optimize the well rates
and well connections simultaneously. The results are plotted in Figure 8.38.7.
The well rates in Figure 8.38.7 were normalized by some fixed oil, gas, and water
rates. Figure 8.3 shows that Run 2 produced at a much higher rate than Run 1. Run 3 was
even better. Over five years, Run 2 predicted 11.2% more oil than Run 1, and Run 3
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 130
predicted 18.9% more oil than Run 1. The excellent performance of Run 2 compared to
Run 1 can be explained as follows. The rate allocation procedure in Run 1 first scaled the
highest GOR wells to meet gas constraints, and then scaled the highest water cut wells to
meet the water constraint. This resulted in suboptimal solutions. On the other hand, Run 2
used a rigorous optimization procedure (the LPII method) and scaled the well rates
differently. For instance, most of the time Run 2 had been producing at a lower total
water rate (Figure 8.5) but a higher gas flow rate (Figure 8.4) than Run 1. And most of
the time Run 2 predicted a much higher gas rate for HEADER_Z than Run 1. In other
words, Run 1 did not utilize the processing capacity of HEADER_Z as efficiently as Run
2. The fact that Run 3 outperformed Run 2 can be explained as follows. At later times in
Run 2, without well reconnection, most of the gas processing capacity of HEADER_Y
was wasted (Figure 8.6), while HEADER_Z was constrained by its gas processing
capacity (Figure 8.7). The algorithm in Run 3 identified this imbalance and switched a
high GOR well from HEADER_Z to HEADER_Y.
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0 500 1000 1500 2000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
O
I
L
R
A
T
E
RUN 1
RUN 2
RUN 3
Figure 8.3: Normalized daily total oil rate.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 131
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0 500 1000 1500 2000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
G
A
S
R
A
T
E
RUN 1
RUN 2
RUN 3
Figure 8.4: Normalized daily total gas rate.
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0 500 1000 1500 2000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
W
A
T
E
R
R
A
T
E
RUN 1
RUN 2
RUN 3
Figure 8.5: Normalized daily total water rate.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 132
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0 500 1000 1500 2000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
G
A
S
R
A
T
E
RUN 1
RUN 2
RUN 3
Figure 8.6: Normalized daily gas rate at HEADER_Y.
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0 500 1000 1500 2000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
G
A
S
R
A
T
E
RUN 1
RUN 2
RUN 3
Figure 8.7: Normalized daily gas rate at HEADER_Z.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 133
8.5.2 Another Gulf of Mexico Field Example
This is a reservoir development model for a newly discovered deepwater oil field in the
Gulf of Mexico. In this model, 26 wells are scheduled to be drilled at different times. All
wells are connected to a common offshore platform, on which resides a separator and
possibly other surface facilities. The vertical distances between the separator and wells
vary between 10,000 ft and 14,000 ft. There are no flow interactions among different
wells.
The purpose of this example was to demonstrate that the application of the developed
shortterm optimization algorithms in longterm prediction can improve facility design.
In this example, two cases were considered, each with a different gas handling capacity
for the separator.
Case 1. In this case, the separator has a gas handling capacity of 98 MMSCF/D. The
total field oil production rate cannot exceed a certain target rate. The lift gas rate for a
gaslift well cannot exceed 6000 MSCF/D. The prediction period is about 22 years.
During the prediction period, some wells will be recompleted, placed on gas lift,
converted from producers to injectors, closed for maintenance, etc. To account for the
downtime between two reporting dates in simulation, an ONTIME factor with a value
between 0 and 1 may be defined for a well. For example, if a producer is expected to
produce equivalent of one month during a two month reporting period, then an ONTIME
factor of 0.5 will be specified for that producer. To facilitate later discussions, we make
the following definitions: instantaneous flow rate is the actual flow rate of a producer
when it is flowing, discounted flow rate is the average flow rate of a well for a reporting
period. For a well, its discounted flow rates for a time step are the instantaneous flow
rates time its ONTIME factor for that time step.
For evaluation purposes, we made two runs: a BASE run and a LPI run. The LPI
run uses Mode LPI described in Section 4.3.1 to allocate the lift gas and production rates
subject to the total oil rate constraint of the field and the total gas rate constraint of the
separator. The BASE run allocated the lift gas and production rates using a
“conventional” method available in VIPEXECUTIVE (Landmark, 2001). This method
proceeds as follows:
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 134
1. Liftgas rates are determined for specified wells based on some gaslift tables. These
tables correlate gasliquid ratio for a production well to its water cut and liquid
production rate. If the formation gasliquid ratio for a production well is higher than
the table value, gaslift is not implemented for the well. Otherwise, the well’s lift gas
rate is calculated to meet the gasliquid ratio specified in the table (Litvak et al.,
1997).
2. Update the production rates of all wells with allocated gaslift rates.
3. Check the total gas rate constraint. If it is violated, scale both the production rates and
lift gas rates of wells with the highest GOR to meet the total gas rate constraint.
4. Check the oil rate constraint. If the constraint is violated, scale the production rates of
certain wells to meet the oil rate constraint.
We emphasize that when checking the flow rate constraints, the LPI run used
instantaneous production rates from each well but the BASE run used discounted
production rates (these options are implemented in this way in VIPEXECUTIVE). Thus,
the same flow rate constraint appeared more strict in the LPI run than in the BASE run.
Figure 8.8 plots the daily field oil production history for both runs (the daily oil rates
in both runs are normalized by the field oil target rate). It appears that the oil production
histories of the two runs are the same except when the oil rate constraint in the BASE run
was active. When the oil rate constraint in the BASE run was active, the LPI run
produced at a lower oil rate. This is because the LPI method scales instantaneous
production rate to meet the flow rate constraint and VIPEXECUTIVE reports discounted
flow rate. Figure 8.9 plots the normalized cumulative oil production for both runs. The
trend of Figure 8.9 agrees with what we observe in Figure 8.8.
Figure 8.10 shows the daily total gas rate for both runs. The total gas constraint was
not active in both runs for the whole prediction period. Thus the gas handling capacity of
the separator is oversized for this reservoir development scenario. Figure 8.11 plots the
normalized cumulative lift gas injection for both runs. It shows that the BASE run
injected about 20% more lift gas than the LPI run for the whole prediction period. This
can be explained as follows. When allocating lift gas, the LPI method needs to construct
the gaslift performance curve. In this procedure, a minimum gaslift efficiency can be
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 135
specified for each well, thus avoiding overinjection of lift gas. On the other hand, the
liftgas table does not have this feature and can overinject lift gas.
Case 2. The reservoir model and the development scenario are exactly the same as in
Case 1, except that the gas handling capacity of the separator is lowered to 60
MMSCF/D. Again, we made the same two runs as in case 1: a BASE run and a LPI run.
Figure 8.12 shows the normalized daily field oil production history for both runs.
Figure 8.13 plots the normalized cumulative field oil production history for both runs.
Figure 8.14 plots the normalized cumulative lift gas rate for both runs. From Figure 8.12
we can see that at about 1900 days the oil production in the BASE run dropped sharply to
a lower rate. This can be explained as follows. Around 1900 days, some wells were
placed on gaslift. The BASE run allocated the lift gas to these wells without the
consideration of the total gas constraint. However, certain careful considerations will
reveal that no lift gas should be injected (Figure 8.14 shows that the LPI run did not
inject any lift gas until about 3000 days). As a result, injecting lift gas decreased the total
oil production instead of increasing it. This demonstrates the danger of allocating lift gas
and production rates separately. Figure 8.12 shows that the BASE run continued to
produce at a lower oil rate than the LPI run until 4500 days, after which the BASE run
produced at a higher oil rate than the LPI run. The lower oil rate in the LPI run at later
stages of the prediction period is probably due to the fact that there was less oil in the
reservoir of the LPI run than in the reservoir of the BASE run at later stages of
prediction. Figure 8.13 shows that over the whole prediction period the LPI run
produced slightly more oil than the BASE run. Figure 8.14 shows that over the whole
prediction period the LPI run allocated about 40% less lift gas than the BASE run. This
is due to two facts: 1) the LPI run optimized the production and lift gas rates
simultaneously while the approach in the BASE run did not; 2) the LPI method can
control the gaslift efficiency.
We can conclude that facility design depends on the well management tools (the
“conventional” approach or the LPI method) used to evaluate the sizes of the separator
and the gas compressor for lift gas. If used appropriately, the LPI method can save
significant amount of construction and operational cost.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 136
0.0E+00
2.0E01
4.0E01
6.0E01
8.0E01
1.0E+00
1.2E+00
0 2000 4000 6000 8000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
O
I
L
R
A
T
E
BASE
LPI
CONSTRAINT
Figure 8.8: Normalized daily total oil production history for Case 1.
0.0E+00
2.0E01
4.0E01
6.0E01
8.0E01
1.0E+00
1.2E+00
0 2000 4000 6000 8000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
C
U
M
U
L
A
T
I
V
E
O
I
L
P
R
O
D
U
C
T
I
O
N
BASE
LPI
Figure 8.9: Normalized cumulative oil production history for Case 1.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 137
0.0E+00
2.0E+04
4.0E+04
6.0E+04
8.0E+04
1.0E+05
1.2E+05
0 2000 4000 6000 8000
TIME, DAYS
D
A
I
L
Y
T
O
T
A
L
G
A
S
R
A
T
E
,
M
S
C
F
/
D
BASE
LPI
CONSTRAINT
Figure 8.10: Daily total gas rate history for Case 1.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2000 4000 6000 8000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
C
U
M
U
L
A
T
I
V
E
L
I
F
T
G
A
S
I
N
J
E
C
T
I
O
N
BASE
LPI
Figure 8.11: Normalized cumulative lift gas injection for Case 1.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 138
0.0E+00
1.0E01
2.0E01
3.0E01
4.0E01
5.0E01
6.0E01
7.0E01
8.0E01
9.0E01
1.0E+00
0 2000 4000 6000 8000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
O
I
L
R
A
T
E
BASE
LPI
Figure 8.12: Normalized daily total oil production history for Case 2.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2000 4000 6000 8000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
C
U
M
U
L
A
T
I
V
E
O
I
L
P
R
O
D
U
C
T
I
O
N
BASE
SP
Figure 8.13: Normalized cumulative oil production history for Case 2.
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 139
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2000 4000 6000 8000
TIME, DAYS
N
O
R
M
A
L
I
Z
E
D
C
U
M
U
L
A
T
I
V
E
L
I
F
T
G
A
S
I
N
J
E
C
T
I
O
N
BASE
LPI
Figure 8.14: Normalized cumulative lift gas injection for Case 2.
8.5.3 Application to the Prudhoe Bay Oil Field
Prudhoe Bay oil field is located on the North Slope of Alaska and operated by BP
Exploration (Alaska) Inc. Oil production from Prudhoe Bay is on decline and constrained
by the gas and water handling limits of surface facilities and velocity constraints in flow
lines. The Efield optimization system (EFOS) was developed at Prudhoe Bay oil field to
automate and optimize the production operations of the field. EFOS contains four major
modules: automatic data preparation, automatic model tuning, production optimization,
and a userfriendly interface. The production optimization module uses the optimization
system described in Section 8.3 to maximize the daily oil production and debottleneck the
production system. EFOS shows how our optimization tools can be used for onsite
production optimization in a real petroleum field. EFOS is described fully by Litvak et al.
(2002).
The major purpose of this section is to demonstrate possible use of the optimization
tools for Prudhoe Bay field and their high efficiency even for such a giant field. To
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 140
facilitate later discussions, a brief description of the production system of Prudhoe Bay
field is presented below.
Production system description. Prudhoe bay has a flexible, highly automated and
complex surface pipe network and processing facilities. The following description of the
production system of Prudhoe Bay is extracted from Litvak et al. (1997).
“Production wells are tied to 22 well pads in the Western Operating Area (WOA) and
16 drill sites in the Eastern Operating Area (EOA). Most production wells can be
switched between high and lowpressure headers at the well pad manifolds. Gaslift is
used in some production wells with high water cut. Gaslift wells can be tied only to
lowpressure headers. Well chokes are installed in well pad manifolds.
The production stream from individual well pads and drill sites is sent without any
separation to three gathering centers in the WOA and to three flow stations in EOA.
Well pads and drill sites are tied to separator banks through high pressure and/or low
pressure flow lines. Each gathering center and flow station has four threestage
separator banks. Separator banks can operate in high pressure or lowpressure mode
with first separation stage pressure of about 670 psig and 150 psig, respectively.
Separator offgas is routed from the gathering centers and flow stations to the central
gas facility where natural gas liquids and miscible injectant are extracted. The natural
gas liquids are then blended with the oil stream coming from the separators. The
miscible injectant is used in the enhanced oil recovery (EOR) project. It is injected, as
part of a WAG (water alternating gas) process in some 120 EOR patterns. The lean
gas, after compression in the central compression plant, is reinjected into the gas cap
for pressure maintenance. The produced water is either reinjected in the EOR pattern
or into water disposal wells.”
Possible usages of the optimization tools. For Prudhoe Bay oil field, the possible
usage of the optimization tools can be summarized as follows.
1. Optimization of production rate and well connections in response to changing
operational conditions. Oil production in Prudhoe Bay field is constrained by
capacity constraints of facilities and velocity constraints of flowlines. In such a
constrained environment, the maximum possible field oil rate is sensitive to the
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 141
changing operational conditions. Also the operational conditions at Prudhoe Bay can
change significantly in a short period. For example, field gas compression capacity
can be reduced by 1.2 BCF/D when the ambient temperature increases from 0
°
F to 40
°
F, which can happen during a 24hour period in Prudhoe Bay (Litvak et al., 2002).
Production capacity can also be affected by equipment maintenance and repair, and
other planned or unplanned activities. Therefore, it is necessary to optimize the
production rate and well connections from time to time to maximize the total field oil
rate in response to the changing operational conditions.
One contribution of this study is that we developed optimization tools that can
simultaneously optimize the production rate and the well connections. To
demonstrate the potential advantage of the simultaneous optimization, we conducted
two optimization runs for the whole Prudhoe Bay field. In the first run, the production
rates were optimized using the LPII method (see Section 4.3.4). In the second run,
the production rates and well connections were optimized simultaneously by the
combined use of the partial enumeration (PE) method and the LPII method. The
number of well reconnections was not allowed to exceed 50. In both runs, the choke
settings for all production wells were not allowed to increase. Results showed that the
second run improved oil production by 2.55% compared to the first run, which is
significant for this large field. On an IBM RS6000 computer, the first run took 0.07
seconds, the second run took about 850 seconds (in the second run, the major portion
of the computational time was spent on solving the network problem after each well
reconnection). This demonstrates the high efficiency of the developed optimization
methods.
2. Debottlenecking studies. Another important use of the optimization tools is to identify
bottlenecks in the production system and evaluate debottlenecking opportunities. This
is illustrated here by the LPII method for rate allocation.
The bottlenecks of the production system can be identified from the optimal
solution of the LPII model for the rate allocation problem. Suppose the rate
allocation problem is formulated as a LP problem, Problem 8.7, as follows
maximzie ( ) x c x
T
f = (8.7a)
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 142
subject to
i
T
i
b x a ≤ , m i ,..., 1 = (8.7b)
1 0 ≤ ≤
i
x , n i ,..., 1 = (8.7c)
where x denotes the scaling factors for well rates, b denotes the maximum fluid
processing capacity of facilities and velocity limits for flowlines, ( ) x f is the
objective function. The optimal solution of a LP problem contains a dual variable
i
λ
for each constraint and
i
i
b
f
∂
∂
− =
*
λ (8.8)
where
*
f denotes the optimal objective function of Problem 8.7. Eq. 8.8 states that at
the optimal solution of a LP problem, the dual variable indicates how sensitive the
optimal objective function is relative to the righthand side of the constraints. 0 =
i
λ
indicates that increasing
i
b has no impact on the optimal objective function value.
The more negative the
i
λ , the more sensitive the objective function to
i
b , and the
more likely it is that facility capacity or velocity constraint
i
b is the bottleneck of the
production system.
The rate allocation problem for Prudhoe Bay field contains mostly multiple gas,
water constraints of facilities and velocity constraints. The dual variables for water,
gas, and velocity constraints have different units. To avoid missing debottlenecking
opportunities, several constraints with the most negative dual variables from each
category (gas, water, or velocity constraints) can be identified as debottlenecking
opportunities.
After debottlenecking opportunities are identified, a series of sensitivity runs can
be performed to evaluate the selected opportunities. For example, the gas
compression capacity of the most loaded compressor may be increased and the rate
allocation problem resolved to see how much oil increase can be obtained. The
debottlenecking opportunity with the best economic return can be implemented.
3. Gaslift optimization. In the Prudhoe Bay field model, the gaslift performance is
evaluated by the combined use of inflow performance curves and well tubing
hydraulic models. However, the tubing hydraulic model is not accurate enough to
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 143
represent real physics because of the large diameters of the tubing strings (existing
multiphase flow models do not perform well for large diameter tubing strings (Litvak
et al., 2002)). Therefore, the LPI method has not been applied to the real field.
However, the LPI method can be used to identify potential benefits of doing gaslift
optimization for the Prudhoe Bay field.
We conducted two optimization runs for the whole Prudhoe Bay field. In the first
run, the production rates were optimized using the LPII method. In the second run,
the production rates and lift gas rates were optimized simultaneously by the LPI
method. Results showed that the second run improves daily oil production by 6%
compared to the first run. This study demonstrated that the optimal allocation of lift
gas is a potential effective way to increase the daily oil production. On an IBM
RS6000 computer, the second optimization run took about 10 seconds.
8.6 Concluding Remarks
For practical use, the optimization tools can be implemented in different ways. The
methods can be integrated into or interfaced with a commercial simulator. They can also
be implemented in an inhouse software. The use of a commercial simulator has several
advantages: (1) the integrated system can be applied to a variety of fields; (2) advances in
simulation technology can be easily incorporated and transferred to other fields (Litvak et
al., 2002); (3) the user needs less effort to develop and maintain the simulation and
optimization models, etc.
In a full field study, the network model and the reservoir model can be coupled
together using either the implicit procedures or the explicit procedure. If possible, the
implicit coupling procedures are recommended, because implicit coupling is able to
obtain simulation results consistent with the full field model. If the optimization tools are
applied for longterm prediction, their use should be managed carefully to avoid severe
convergence issues of the solution procedure and to ensure no constraints violations at
the end of a time step.
Application examples presented in this chapter show that optimization methods
developed in this study have distinctive advantages over the heuristic methods currently
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 144
employed in the oil industry. In addition, they are efficient enough to be used for both
realtime optimization and longterm prediction studies for largescale reservoirs.
The optimization tools developed in this study were designed to optimize the short
term production operation based on static reservoir conditions. The reservoir dynamics
were not considered in the optimization step. Although we did apply these optimization
tools to longterm predictions, the impact of decisions based on this approach on long
term oil recovery needs further investigation.
145
Chapter 9
Conclusions
This study addressed the problem of optimizing the production rates, lift gas rates, and
well connections to flowlines subject to multiple flow rate and pressure constraints to
achieve certain shortterm operational goals. This problem is being faced in many mature
fields and is an important element to consider in planning the development of a new field.
In this research we developed, compared and tested optimization methods for petroleum
production problems. The major contributions of this work and areas for further study are
summarized in this section.
9.1 Conclusions
1. The optimization methods investigated and developed here are effective for problems
of varying complexities and sizes. The methods can be used for both shortterm
production optimizations and longterm reservoir development studies. For instance,
they can be used to optimize well connections and well rates simultaneously, even for
large production systems such as those in the Prudhoe Bay field.
2. For the rate allocation problem, when flow interactions among wells are not
significant, the various separable programming methods (the LPI, LPII, MILPI,
MILPII methods described in Chapter 4) proved to be better than other conventional
approaches used in the oil industry (i.e., the equalslope method for gaslift
optimization, and the rulebased methods available in commercial simulators).
CHAPTER 9. CONCLUSIONS 146
Specifically, the MILPI method can be used for problems with performance curves
of arbitrary shapes. The MILPI, LPI, and LPII methods are suitable for largescale
rate allocation problems. The MILPII method is especially suitable for gaslift
optimization problems with many nonconcave gaslift performance curves.
3. For the rate allocation problem, when flow interactions have significant impact on the
optimal solution, simulation models capable of capturing such flow interactions
should be used in the optimization process. Formulation P2 (Chapter 5) proved to be
most appropriate for this problem. The formulated nonlinearly constrained problem
(NCP) can be solved efficiently by a sequential quadratic programming (SQP)
algorithm.
4. For the optimization of well connections, the partial enumeration (PE) algorithm
proved to be both efficient and robust compared to a genetic algorithm (GA).
5. The twolevel programming approach developed for the entire production
optimization problem is flexible. Different optimization algorithms can be used for
the upper level problem and the lower level problem.
6. For a production optimization problem with multiple objectives, the weightedsum
method may fail to obtain solutions of interest to us.
7. Multiple solutions may exist for the multiphase flow problem of a gathering system.
This may bring computational difficulties to full field simulations and production
optimizations. Postanalysis should be performed to ensure the validity of a solution
from a simulation or an optimization run.
9.2 Summary of Contributions
The major contributions of this study can be summarized as follows:
1. Optimization methods were developed and tested on various synthetic examples and
applied successfully to a real petroleum field, the giant Prudhoe Bay oil field in
Alaska. Results demonstrated that the methods developed have distinct advantages
over conventional production optimization approaches.
2. A twolevel programming approach was developed to optimize the production rates,
lift gas rates, and well connections simultaneously subject to multiple flow rate and
pressure constraints.
CHAPTER 9. CONCLUSIONS 147
3. An appropriate formulation, Formulation P2 (Chapter 5), was proposed for the
nonlinear rate allocation problem. The optimization problem formulated in this
manner was solved by a SQP algorithm.
4. Several existing separable programming models (Model LPI, LPII, and MILPI)
were investigated and a new model (Model MILPII) was proposed for the rate
allocation problem. Several techniques were developed to speed up the optimization
process, and they can be very effective for largescale problems.
5. A partial enumeration (PE) method was developed to optimize well connections. The
PE method is more efficient than a genetic algorithm (GA). Solutions obtained from
the PE method were at least as good as those from the GA.
6. An efficient solution procedure was developed to simulate multiphase flow in a
gathering system.
7. An efficient method, the Jacobian method, was proposed to compute the sensitivity
coefficients of a gathering system. Sensitivity coefficients are required in applications
such as optimization, model tuning, and sensitivity analysis.
8. A framework was developed to formulate and solve multiobjective production
optimization problems. This framework unifies several existing multiobjective
optimization algorithms. It was demonstrated through an example that the weighted
sum method may fail to obtain solutions of interest to us.
9. Some of the optimization tools were coupled with models for multiphase flow in the
reservoir and surface pipeline network in a commercial reservoir simulator. This
automates the application of optimization tools to shortterm production optimization
and longterm reservoir development studies.
10. An automatic differentiation technique was applied to compute derivatives required in
network simulation and production optimization. Derivatives obtained from
automatic differentiation are more accurate than those obtained from finite difference
approximations. In certain cases, automatic differentiation is more efficient than finite
difference approximations.
CHAPTER 9. CONCLUSIONS 148
9.3 Recommendations for Further Study
1. Investigate whether well interactions within the reservoir are important for shortterm
production optimization problems. And if so, how to handle them efficiently.
2. Investigate the applicability of genetic algorithms on constrained production
optimization problems. Genetic algorithms are slow, however, they are flexible and
robust enough to handle optimization problems with complex simulation models and
rough objective surfaces.
3. Investigate how the accuracy of the numerical models for multiphase pipe flow and
choke flow impacts the optimal solution.
4. Investigate the impact of decisions obtained from shortterm production optimizations
on longterm oil recovery.
5. Develop optimization methods for gathering systems with loops.
6. Investigate multiobjective optimization methods for surface facility design problems.
7. Develop methods to find stable solutions of the multiphase network flow problem.
8. Develop methods to construct quality well performance curves efficiently.
149
Nomenclature
A = pipe area
B = formation volume factor
i
b = righthand side of the th i constraint
0
C = profile parameter in the drift flux model
d
C = discharge coefficient in the Sachdeva choke model (Sachdeva et al.,
1986)
p
C ,
v
C = specific heat capacity at constant pressure and volume, respectively
c = constraint functions;
cost coefficients for an objective function
v
p
c
= velocity coefficient for flow rate of phase p (Section 4.4.1)
cˆ = subset of constraints that are active at the optimal solution
d = choke inside diameter (Section 3.4);
deviational variable (Chapter 7)
l
E ,
g
E = liquid and gas phase in situ fractions, respectively
F = constraint functions;
vector of objective functions (Chapter 7)
f = function
m
f = Fanning friction factor
o
f ,
w
f = flowing fractions of oil and water
g
= gravity;
gradient (Section 5.2)
h = achievement function (Chapter 7)
ij
h = the th j function involved in th i constraint of a separable problem
J = Jacobian matrix
k = ratio of specific heat capacity
L = Lagrangian function
NOMENCLATURE 150
l = lower bound of a constraint
M = merit function
m = number of constraints
n = number of decision variables,
polytropic exponent of gas (Section 3.4)
c
n = number of connections of a well
n
n = number of nodes
w
n = number of wells
p
= pressure;
search direction (Section 2.1.3, Section 5.2)
n
j p
Q
,
= flow rate limit of phase p for node j
q
= flow rate
cij
q = flow rate of component c between node i and j
n
j p
q
,
= flow rate of phase p for node j
w
j p
q
,
= flow rate of phase p for well j
em
R = Reynolds number of a mixture
s
R = solution gasoil ratio
r = number of discrete points of a piecewise linear well performance
curve
S = feasible set of an optimization problem
S = pipe perimeter
s = slack variable
U = velocity (Chapter 3);
utility function (Chapter 7)
d
U = drift velocity of the gas in the drift flux model
sg
U = superficial velocity of the gas phase
sl
U = superficial velocity of the liquid phase
sm
U = superficial velocity of the mixture
u = upper bound of a constraint;
system response (Section 3.6)
v = velocity
max
0 , p
v
= the maximum in situ velocity along a flowline for a reference flow
rate of phase p (Section 4.4.1)
w = weighting coefficients
NOMENCLATURE 151
x = decision variable, independent variable;
mass fraction of gas (Section 3.4);
well connections (Section 6.2)
Y = feasible set of a vector of objective functions
y
= lift gas rates (Section 6.2);
ratio of downstream pressure to upstream pressure (Section 3.4);
binary decision variable (Section 4.3.2)
c
y = critical ratio of downstream pressure to upstream pressure
z = value of an objective function;
production rates (Section 6.2)
z = lower bound for an objective function
z = upper bound for an objective function
Acronyms
ADIFOR = an automatic differentiation software (Bischof et al., 1998)
BHP = bottomhole pressure
DFM = drift flux model for multiphase pipe flow
EFOS = Efield optimization system for the Prudhoe Bay oil field
EOA = Eastern Operating Area of the Prudhoe Bay oil field
GA = genetic algorithm
GOR = gas oil ratio
GLR = gas liquid ratio
GOM = Gulf of Mexico
GP = goal programming
IP = integer programming
LIP = linear integer programming
LP = linear programming
LPI = a model for solving the rate allocation problem (Section 4.3.1)
LPII = a model for solving the rate allocation problem (Section 4.3.4)
MILP = mixed integer linear programming
MILPI = a model for solving the rate allocation problem (Section 4.3.2)
MILPII = a model for solving the rate allocation problem (Section 4.3.3)
MIP = mixed integer programming
MOP = multiple objective optimization problem
NCP = nonlinearly constrained programming
P1 = formulation P1, defined in Section 5.3.1
P2 = formulation P2, defined in Section 5.3.2
NOMENCLATURE 152
PE = partial enumeration method for well connections optimization,
defined in Section 6.3.2
SLC = sequential linearly constrained
SNOPT = a generalpurpose system for solving largescale optimization
problems (Gill et al., 1998)
SP = separable programming
SQP = sequential quadratic programming
WAG = water alternating gas
WI = well index
WOA = Western Operating Area of the Prudhoe Bay oil field
Symbols
α = system parameters;
steplength of a line search optimization method (Section 2.1.3,
Section 5.2)
δ = delta (change)
ε = absolute pipe roughness (Section 3.3.1);
tolerances (Section 3.5.2);
perturbations (Section 3.6.1)
γ
= specific gravity
λ = decision variable (Section 4.3);
Lagrangian variable (Section 5.2);
dual variable (Section 8.5.3)
p c,
λ = mobility of component c in phase p (Section 3.2)
µ
= viscosities
θ = angle
ρ
= density;
penalty coefficients (Section 5.2)
w
τ = wall friction shear stress
t ∆ = time step length in reservoir simulation
n
Ω
= set of all network nodes
n
i
Ω
= set of nodes that directly connected to node i
w
i
Ω
= set of wells whose flow streams enter node i
n
s
Ω
= set of solution nodes
∀ = arbitrary
∃ = exist
NOMENCLATURE 153
∇ = gradient
∞
= Linfinite norm
Subscript
1 = upstream
2 = downstream
c = component
g
= gas
l = liquid
lg = lift gas
m = mixture
o = oil
p
= phase
w = water
Superscript
max = maximum
min = minimum
n = node
r = reservoir
s = surface
t = total
v = iteration index;
velocity (Section 4.4.1)
w = well
154
Bibliography
ANDERSON, E., BAI, Z., BISCHOF, C., BLACKFORD, S., DEMMEL, J.,
DONGARRA, J., DU CROZ, J., GREENBAUM, A., HAMMARLING, S.,
MCKENNY, A., AND SORENSEN, D., 1999. LAPACK User’s Guide, Third
Edition, Society for Industrial Applied Mathematics.
ANDERSSON, J., 2001. Multiobjective Optimization in Engineering Design –
Applications to Fluid Power Systems, dissertation No. 675, Linköping University,
Linköping, Sweden.
ARONOFSKY, J.S., 1983. “Optimization Methods in Oil and Gas Development”, paper
SPE 12295.
ARONOFSKY, J.S., AND LEE, A.S., 1958. “The Use of Linear Programming Model for
Scheduling Crude Oil Production”, Trans., AIME 213, 5154.
ASHEIM, H., 1988. “Criteria for Gas Lift Stability”, JPT, November, 14521456.
ATTRA, H.D., WISE, W.B., AND BLACK, W.M., 1961. “Application of Optimizing
Techniques for Studying Field Producing Operations”, Journal of Petroleum
Technology, January, 8286.
AZARM, S., 2002. Class Notes for Multiobjective Optimization, Retrieved in October
2002 from http://www.glue.umd.edu/~azarm/optimum_notes/multi/multi.html.
AZIZ, K., AND DURLOFSKY, L., 2002. Fundamentals of Reservoir Simulation,
Stanford University, Course material for PE223, winter 20012002.
BIBLIOGRAPHY 155
AZIZ, K., AND PETALAS, N., 1994. “New PCBased Software for Multiphase Flow
Calculations”, paper SPE 28249 presented at the SPE Petroleum Computer
Conference, Dallas, 31 July3 August.
BARNES, D.A., HUMPHREY, K., AND MUELLENBERG, L., 1990. “A Production
Optimization System for Western Prudhoe Bay Field, Alaska”, paper SPE 20653
presented at the 65
th
Annual Technical Conference and Exhibition of the Society of
Petroleum Engineers held in New Orleans, LA, September 2326.
BARUA, S., PROBST, A., AND DUTTAROY, K., 1997. “Application of a General
Purpose Network Optimizer to Oil and Gas Production”, paper SPE 38838
presented at the 1997
th
SPE Annual Technical Conference and Exhibition held in
San Antonio, Texas, 58 October.
BEGGS, H.D., 1991. Production Optimization, OGCI Publications, Oil and Gas
Consultants International Inc., Tulsa.
BEGGS, H.D., AND BRILL, J.P., 1973. “A Study of Two Phase Flow in Inclined Pipes,”
Trans. of Pet. Soc. of AIME, vol. 256, 607.
BELLMAN, R., 1957. Dynamic Programming, Princeton University Press.
BENAYOUN, R., DE MONTGOLFIER, J., TERGNY, J., AND LARICHEV, O., 1971.
“Linear Programming with Multiple Objective Functions: STEP Method (STEM)”,
Mathematical Programming 1(3), 366375.
BERTSEKAS, D.P., 1982. Constrained Optimization and Lagrange Multiplier Methods,
Academic Press, New York.
BERTSIMAS, D., AND TSITSIKLIS, J., 1997. Introduction to Linear Optimization,
Athena Scientific, Belmont, Massachusetts.
BERZ, M., BISCHOF, C., CORLISS, G., AND GRIEWANK, EDITORS, 1996.
Computational Differentiation: Techniques, Applications, and Tools, SIAM,
Philadelphia.
BISCHOF, C., CARLE, A., HOVLAND, P., KHADEMI, P., AND MAUER, A., 1998.
ADIFOR 2.0 User’s Guide Technical Report (Revision D), Mathematics and
BIBLIOGRAPHY 156
Computer Science Division Technical Memorandum No. 192 and Center for
Research on Parallel Computation Technical Report CRPC95516S.
BITTENCOURT, A.C., AND HORNE, R.N., 1997. “Reservoir Development and Design
Optimization”, paper SPE 38895 presented at the 1997 SPE Annual Technical
Conference and Exhibition held in Antonio, Texas, 58 October.
BOHANNON, J.M., 1970. “A Linear Programming Model for Optimum Development of
MultiReservoir Pipeline Systems”, Journal of Petroleum Technology, November,
14291436.
BOYCHUK, L.M. AND OVCHINNIKOV, V.O., 1973. “Principal Methods of Solution
Multicriterial Optimization Problems”, Soviet Automatic Control, 6, 14.
BOYD, S., AND VANDENBERGHE, L., 2001. Convex Optimization, Course Reader for
Class EE364: Convex Optimization with Engineering Applications. Stanford, CA.
To be published in 2003.
BREAUX, E.J., MONROE, S.A., BLANK, L.S., YARBERRY, D.W., AND AL
UMRAN, S.A., 1985. “Application of a Reservoir Simulator Interfaced with a
Surface Facility Network: A Case History”, Society of Petroleum Engineering
Journal, 397404, June.
BRILL, J.P., AND BEGGS, H.D., 1998. TwoPhase Flow in Pipes, Class Notes, U of
Tulsa.
BUITRAGO, S., AND RODRIGUEZ, E., AND ESPIN, D., 1996. “Global Optimization
Techniques in Gas Allocation for Continuous Flow Gas Lift Systems”, paper SPE
35616 presented at the Gas Technology Conference held in Calgary, Alberta,
Canada 28 April  1 May.
BYER, T.J., 2000. Preconditioned Newton Methods for Simulation of Reservoirs with
Surface Facilities, Ph.D. dissertation, Stanford University.
CARROLL, D.L., 2000. Fortran Genetic Algorithm Driver, Retrieved in August 2000
from http://cuaerospace.com/carroll/ga.html.
BIBLIOGRAPHY 157
CARROLL, J.A.III, AND HORNE, R.N., 1992. “Multivariate Optimization of
Production Systems”, Journal of Petroleum Technology, July, 782831.
CAO, H., 2002. Development of Techniques for General Purpose Simulators, Ph.D.
dissertation, Stanford University, CA.
CHARNES, A., AND COOPER, W.W., 1961. Management Models and Industrial
Applications of Linear Programming, Vol. I, Wiley, New York.
CHEN, Y.G., 2001. Modeling GasLiquid Flow in Pipes: Flow Pattern Transitions and
DriftFlux Modeling, MS thesis, Stanford University, CA.
CLAY, R.M., STOISITS, R.F., PRITCHETT, M.D., ROOD, R.C., AND COLOGGI,
J.R., 1998. “An Approach to RealTime Optimization of the Central Gas Facility at
the Prudhoe Bay Field”, paper SPE 49123 presented at the SPE Annual Technical
Conference and Exhibition held in New Orleans, Louisiana, 2730 September.
COATS, K.H., 1980. “An Equation of State Compositional Model”, SPEJ, October, pp.
363376.
COLEBROOK, C.F., 1939. “Turbulent Flow in Pipes with Particular Reference to the
Transition Region Between the Smooth and RoughPipe Laws”, J. Inst. Civil
Engrs., 11, 133.
COLLINS, M., COOPER, L., HELGASON, R., KENNINGTON J., AND LEBLANC, L.,
1978. “Solving The Pipe Network Analysis Problem Using Optimization
Techniques”, Management Science, Vol. 24, No. 7., March. pp. 747760.
DANTZIG, G. B., 1963. Linear Programming and Extensions, Princeton University
Press.
DONACHIE, R.P., 1973. “Digital Program for Water Network Analysis”, Jo. Hydraulics
Div., Proc. Amer. Soc. Civil Engineers, Vol. 100, pp. 393403.
DUNS, H.JR., AND ROS, N.C.J., 1963. “Vertical Flow of Gas and Liquid Mixtures in
Wells”, Sixth World Petroleum Congress, Section 2, Paper No. 22, Frankfurt.
BIBLIOGRAPHY 158
DUTTAROY, K., BARUA, S., AND HEIBA, A., 1997. “ComputerAided Gas Field
Planning and Optimization”, paper SPE 37447 presented at the 1997 SPE
Production Operations Symposium, held in Oklahoma City, Okalahoma, 911
March.
DUTTAROY, K., AND KATTAPURAM, J., 1997. “A New Approach to GasLift
Allocation Optimization”, paper SPE 38333 presented at the 1997 SPE Western
Regional Meeting held in Long Beach, California, 2527 June.
EMANUEL, A.S., AND RANNEY, J.C., 1981. “Studies of Offshore Reservoir with an
Interfaced Reservoir/Piping Network Simulator”, Journal of Petroleum Technology,
399406, March.
EPP, R., AND FOWLER, A.G., 1970. “Efficient Code for SteadyState Flows in
Networks”, Jo. Hydraulics Div., Proc. Amer. Soc. Civil Engineers, Vol. 96, pp. 43
56.
FANG, W.Y., AND LO, K.K., 1996. “A Generalized WellManagement Scheme for
Reservoir Simulation”, SPE Reservoir Engineering, May, 116120.
FIACCO, A.V., AND MCCORMICK, G.P., 1968. Nonlinear Programming: Sequential
Unconstrained Minimization Techniques, John Wiley and Sons, New York and
Toronto.
FISHBURN, P.C., 1974. “Lexicographic Orders, Utilities and Decision Rules: A
Survey”, Mgmt Sci. 20 (11), 14421471.
FITREMANN, J.M., AND VEDRINES, P., 1985. “Non Steady Gas Liquid Flow in Pipes
and Gaslifted Wells”, Proceedings of the Second International Conference on
Multiphase Flow, London, June 1921, 245262.
FLETCHER, R., 1987. Practical Methods of Optimization, 2nd ed. John Wiley & Sons,
Inc., New York.
FONSECA, C.M., AND FLEMING, P.J., 1993. “Genetic Algorithms for Multiobjective
Optimization: Formulation, Discussion and Generalization”, Proceedings of the
BIBLIOGRAPHY 159
Fifth International Conference on Genetic Algorithms, Morgan Kaufmann
Publishers, Inc., San Mateo, CA, pp. 416423.
FUJII, H., 1993. Multivafuate Production Systems Optimization in Pipeline Networks,
M.S. Thesis, Stanford University, CA.
FUJII, H., AND HORNE, R., 1995. “Multivariate Optimization of Networked Production
Systems”, SPE Production & Facilities, August, 165171.
GAL, T., AND NEDOMA, J., 1972. “Multiparametric Linear Programming”, Mgmt Sci.
18 (7), 406421.
GEOQUEST, 2000. Eclipse 300 Technical Description 2000A, GeoQuest, Schlumberger.
GILL, P.E., MURRAY, W., AND SAUNDERS, M.A, 1998. User's Guide for SNOPT
5.3: A Fortran Package for Largescale Nonlinear Programming, Retrieved in
October 2002 from http://www.sbsisoloptimize.com/manuals.htm.
GILL, P.E., MURRAY, W., AND SAUNDERS, M.A., 2002. “SNOPT: An SQP
Algorithm for Largescale Constrained Optimization”, SIAM J. Optimization, Vol.
19, No. 4, pp. 9791006.
GILL, P.E., MURRAY, W., AND WRIGHT, M.H., 1981. Practical Optimization,
Academic Press,San Diego, CA
GOLDBERG, D.E., 1989. Genetic Algorithms in Search, Optimization, and Machine
Learning, AddisonWesley, Reading, MA.
GOMORY, R.E, 1958. “Outline of an algorithm for integer solutions to linear programs”,
Bulletin of the American Mathematical Society, 64, 275278.
GREEN, A.S., JOHNSON, B.V., AND CHOI, H.J., 1993. “FlowRelated Corrosion in
LargeDiameter Multiphase Flowlines”, SPE Production & Facilities, May, 97100.
GÜYAGÜLER, B., 2002. Optimization of Well Placement and Assessment of
Uncertainty, Ph.D. Disertation, Stanford Unvierstiy, CA, USA.
GÜYAGÜLER, B., AND BYER, T., 2001. Production Optimization. A Presentation
Given at a Stanford SUPRIO Seminar, December.
BIBLIOGRAPHY 160
GÜYAGÜLER, B. AND HORNE, R.N., 2001. “Uncertainty Assesment of Well
Placement Optimization”, paper SPE 71265 presented at the 2001 SPE Annual
Technical Conference and Exhibition, New Orleans, Louisiana, September 30
October 3.
GÜYAGÜLER, B., HORNE, R.N., ROGERS, L., AND ROSENZWEIG, J.J., 2000.
“Optimization of Well Placement in a Gulf of Mexico Waterflooding Project”,
paper SPE 63221 presented at the 2000 SPE Annual Technical Conference and
Exhibition held in Dallas, Texas, 14 October.
HAIMES, Y.Y., HALL, W.A., AND FREEDMAN, H.T., 1975. Multiobjective
Optimization in Water Resources Systems, The Surrogate Worth Tradeoff Method,
Elsevier Scientific, New York.
HALL, M.A., 1976. “Hydraulic Network Analysis Using (Generalized) Geometric
Programming”, Networks, Vol. 6, pp. 105130.
HARTLEY, H.O., 1961. “Nonlinear Programming by the Simplex Method”,
Econometrica 29, pp. 223237.
HEIBA, A., BARUA, S., AND DUTTAROY, K., 1997. “Surface Facilities Management
for Thermal Recovery Processes”, paper SPE 37529 presented at SPE International
Thermal and Heavy Oil Symposium, Bakersfield, CA, February.
HEPGULER, G., BARUA, S., AND BARD, W., 1997. “Integration of a Field Surface
and Production Newtork with a Reservoir Simulator”, SPE Computer Applications,
8893, June.
HILLIER, F.S., AND LIEBERMAN, G.J., 2001. Introduction to Operations Research, 7
th
Edition, McCrawHill Companies, Inc., New York, NY,
HOLLAND, J.H., 1975. Adaptation in Natural and Artificial Systems, The University of
Michigan Press, Ann Harbor.
HONG, T.T., 1975. Effect of the Variables on Optimization of Continuous Gas Lift
System, M.S. thesis, U. of Tulsa.
BIBLIOGRAPHY 161
HWANG, C.L., PAIDY, S.R., AND YOON, K., 1980. “Mathematical Programming with
Multiple Objectives: A Tutorial”, Computing and Operational Research 7, 531.
KANU, E.P., MACH, J., AND BROWN, K.B., 1981. “Economic Approach to Oil
Production and Gas Allocation in Continuous Gas Lift”, Journal of Petroleum
Technology, October, 18871892.
KARMARKAR, N., 1984. “A New Polynomialtime Algorithm for Linear
Programming”, Combinatorica, 4, 373395.
KEENEY, R.L., AND RAIFFA, H., 1976. Decision with Multiple Objectives:
Preferences and Value Tradeoffs, Wiley, New York.
KERMANI, M.B., AND HARROP, D., 1996. “The Impact of Corrosion on the Oil and
Gas Industry”, SPE Production & Facilities, August, 186190.
KIRKPATRICK, S., GELATT, C.D.JR., AND VECCHI, M.P., 1983. “Optimization by
Simulated Annealing”, Science, 220, 4598, 671680.
KLEE, V., AND MINTY, D.W., 1972. “How good is the simplex algorithm?”, in
Inequalities  III, O. Shisha (ed.), Academic Press, New York, NY, 159175.
KUHN, H.W., AND TUCKER, A.W., 1951. “Nonlinear Programming”, in Proceedings
of the Second Berkeley Symposium on Mathematical Statistics and Probability ( J.
Neyman, ed. ) pp. 481492, Berkeley, University of California Press.
LAND, A.H., AND DOIG, A.G., 1960. “An Automatic Method for Solving Discrete
Programming Problems”, Econometrica 28, 497520.
LANDA, J.L., 1997. Reservoir Parameter Estimation Constrained to Pressure Transient,
Performance History and Distributed Saturation Data, Ph.D. dissertation, Stanford
University, CA.
LANDMARK, 2001. VIPEXECUTIVE Technical Reference, Landmark Graphics
Corporation, September.
BIBLIOGRAPHY 162
LASDON, L., COFFMAN, P.E., MACDONALD, R., MCFARLAND, J.W., AND
SEPEHRNOORI, K., 1986. “Optimal Hydrocarbon Reservoir Production Policies”,
Operations Research, Vol. 34, No. 1, JanuaryFebruary, 4054.
LEA, J.F.JR., AND TIGHE, R.E., 1983. “Gas Well Operation with Liquid Production”,
paper SPE 11583 prsented at the 1983 Production Operation Symposium held at
Oklahoma City, Oklahoma, February 27 – March 1.
LENSTRA, J.K., RINOOY KAN, A.H.G., AND SCHRIJVER, A. (EDITORS), 1991.
History of Mathematical Programming: A Collection of Personal Reminiscences,
North Holland.
LITVAK, M.L., CLARK, A.J., FAIRCHILD, J.W., FOSSUM, M.P., MACDONALD,
C.D., AND WOOD, A.R.O., 1997. “Integration of Prudhoe Bay Surface Pipeline
Network and Full Field Reservoir Models”, paper SPE 38885 presented at the 1997
SPE Annual Technical Conference and Exhibition held in Antonio, Texas, 58
October.
LITVAK, M.L., AND DARLOW, B.L., 1995. “Surface Network and Well Tubinghead
Pressure Constraints in Compositional Simulation”, paper SPE 29125 presented at
the 13th SPE Symposium on Reservoir Simulation held in San Antonio, TX, 1215
February.
LITVAK, M.L., HUTCHINS, L.A., SKINNER, R.C., DARLOW, B.L., WOOD, R.C.,
AND KUEST L.J., 2002. “Prudhoe Bay EField Production Optimization System
Based on Integrated Reservoir and Facility Simulation”, paper SPE 77643 presented
at the SPE Annual Technical Conference and Exhibition held in San Antonio,
Texas, 29 September – 2 October.
LIU, K.T., 1969. “The Numerical Analyses of Water Supply Networks by Digital
Computer”, Thirteenth Congress of the International Association for Hydraulic
Research, Vol. 1 (Aug.Sept.), pp. 3643.
BIBLIOGRAPHY 163
LO, K.K., AND HOLDEN, C.W., 1992. “Use of Well Management Schemes for Rate
Forecasts”, paper SPE 24071 presented at the Western Regional Meeting held in
Bakersfield, California, March 30April 1.
LO, K.K., STARLEY, G.P., AND HOLDEN, C.W., 1995. “Application of Linear
Programming to Reservoir Development Evaluations”, SPE Reservoir Engineering,
February, 5258.
MCFARLAND, J.W., LASDON, L., LOOSE, V., 1984. “Development Planning and
Management of Petroleum Reservoirs Using Tank Models and Nonlinear
Programming”, Operations Research, Vol. 32, No. 2, MarchApril, 270289.
MICHALEWICZ, Z., 1996. Genetic Algorithms+Data Structure = Evoluation Programs,
3
rd
Edition, SpringerVerlag, Berlin.
MIETTINEN, K., 1999. Nonlinear Multiobjective Optimization. Kluwer Academic
Publishers, Boston.
MILLER, C.E., 1963. “The Simplex Method for Local Separable Programming”, in
Recent Advances in Mathematical Programming (R. L. Graves and P. Wolfe, eds.),
pp. 89100, McGrawHill Book Co., New York.
MURRAY, W., 1997. “Sequential Quadratic Programming Methods for LargeScale
Problems”, J. Comput. Optim. Appl., 7, pp. 127142.
MURRAY, W., 2000. Class Notes for MS&E 312: Optimization Algorithms, Stanford
University, CA.
MURTAGH, B.A., 1981. Advanced Linear Programming: Computation and Practice.
McGrawHill International Book Co., New York.
MURTAGH, B.A., AND SAUNDERS, M.A., 1982. “A Projected Lagrangian Algorithm
and Its Implementation for Sparse Nonlinear Constraints”, Mathematical
Programming Study 16, 84117.
NISHIKIORI, N., REDNER, R.A., DOTY, D.R., AND SCHMIDT, Z., 1989. “An
Improved Method for Gas Lift Allocation Optimization”, paper SPE 19711
BIBLIOGRAPHY 164
presented at the 64th Annual Technical Conference and Exhibition of the Society of
Petroleum Engineering held in San Antonio, TX, October 811.
OUYANG, L., 1998. Single Phase and Multiphase Fluid Flow in Horizontal Wells, Ph.D.
dissertation, Stanford University, CA.
OYAMA, A., AND LIOU, M., 2002. “Multiobjective Optimization of A MultiStage
Compressor Using Evolutionary Algorithm”, paper AIAA 20023535.
PADBERG, M., 2000. “Approximating Separable Nonlinear Functions Via Mixed Zero
One Programs”, Operations Research Letters, 27 (200), 15.
PALKE, M.R., AND HORNE, R.N., 1997a. “Nonlinear Optimization of Well Production
Considering Gas Lift and Phase Behavior”, paper SPE 37428 presented at the 1997
SPE Production Operations Symposium held in Oklahoma City, Oklahoma, 911
March.
PALKE, M.R., AND HORNE, R.N., 1997b. “Determining the Value of Reservoir Data
by Using Nonlinear Production Optimization Techniques”, paper SPE 28047
presented at the 1997 SPE Asia Pacific Oil & Gas Conference and Exhibition, held
in Kuala Lumpur, Malaysia, 1416 April.
PEACEMAN, D.W., 1978. “Interpretation of WellBlock Pressures in Numerical
Reservoir Simulation”, Soc. Pet. Eng. J. (June) 183194; Trans. AIME, 253.
POWELL, M.J.D., 1998. “Direct Search Algorithms for Optimization Calculations”, in
Acta Numerica, Vol. 7, Cambridge University Press.
POWELL, D. AND SHOLNICK, M., 1993. “Using Genetic Algorithms in Engineering
Design Optimization with Nonlinear Constraints”, Proceedings of the Fifth
International Conference on Genetic Algorithms, Morgan Kaufmann. pp. 424431.
PRESS, W.H., TEUKOLSKY, S.A., VETTERLING, W.T., AND FLANNERY, B.P.,
1992. Numerical Recipes in Fortran 77, Cambridge University Press.
RASHEED, K., 1998. “An Adaptive Penaly Approach for Constrained Genetic
Algortihm Optimization”, The Third Annual Conference on Genetic Programming
(GP98): Symposium on Genetic Algorithms (SGA98).
BIBLIOGRAPHY 165
RAVINDRAN, N., 1992. Multivariate Optimization of Production Systems  the Time
Dimension, MS thesis, Stanford University, CA, USA.
SACHDEVA, R., SCHMIDT, Z., BRILL, J.P., BLAIS, R.M., 1986. “TwoPhase Flow
Through Chokes”, paper SPE 15657 presented at the 61
st
Annual Technical
Conference and Exhibition of SPE. New Orleans, LA October 58.
SALUKVADZE, M.E., 1974. “On the Existence of Solution in Problems of Optimization
under Vector Valued Criteria”, Journal of Optimization Theory and Applications,
12, 2, 203217.
SCHIOZER, D.J., 1994. Simultaneous Simulation of Reservoir and Surface Facilities.
Ph.D. dissertation, Stanford University, CA.
SIMSCI, 1999. Input Manual for NETOPT, Simulation Sciences, Brea, California, USA.
SIMSCI, 2001. PIPEPHASE User’s Guide, Simulation Sciences, Brea, California, USA.
SPENDLEY, W., HEXT, G.R., AND HIMSWORTH, F.R., 1962. “Sequential
Application of Simplex Designs in Optimisation and Evolutionary Operation”,
Technometrics, Vol. 4, pp. 441461.
STEUER, R.E., 1977. “An Interactive Multiple Objective Linear Programming
Procedure”, in Multiple Criteria Decision Making, M. K. Starr and M. Zeleny
(Eds), pp. 225239, North Holland, New York.
STOISITS, R.F., BATESOLE, E.C., CHAMPION, J.H., AND PARK, D.H., 1992.
“Application of Nonlinear Adaptive Modeling for Rigorous Representation of
Production Facilities in Reservoir Simulation”, paper SPE 24898 presented at the
67th Annual Technical Conference and Exhibition of the Society of Petroleum
Engineers held in Washington, DC, October 47.
STOISITS, R.F., CRAWFORD, K.D., MACALLISTER, D.J., MCCORMACK, M.D.,
LAWAL, A.S., AND OGBE, D.O., 1999. “Production Optimization at the Kuparuk
River Field Utilizing Neural networks and Genetic Algorithms”, paper SPE 52177
presented at the 1999 MidContinent Operations Symposium held in Oklahoma
City, Oklahoma, 2831 March.
BIBLIOGRAPHY 166
STOISITS, R.F., SCHERER, P.W., AND SCHMIDT, S.E., 1994. “Gas Optimization at
the Kuparuk River Field”, paper SPE 28467 presented at the 69th Annual Technical
Conference and Exhibition held in New Orleans, LA, 2528 September.
SVEDEMAN, S.J., AND ARNOLD, K.E., 1994. “Criteria for Sizing Multiphase
Flowlines for Erosive/Corrosive Service”, SPE Production & Facilities, February,
7480.
TAITEL, Y., AND DUKLER, A.E., 1976. “A Model for Predicting Flow Regime
Transitions in Horizontal and Near Horizontal GasLiquid Flow”, AIChE Journal,
22, 47.
WALLENIUS, J., 1975. “Interactive Multiple Criteria Decision Methods: An
Investigation and An Approach”, Acta Academiae Oecnonomicae Helsingiensis,
Series A: 14, Helsinki School of Economics, Finland.
WANG, P., LITVAK, M.L., AND AZIZ, K., 2002a. “Optimization Production from
Mature Fields”, paper presented at the 17
th
World Petroleum Congress, Rio De
Janerio, Brazil, 15 September.
WANG, P., LITVAK, M.L., AND AZIZ, K., 2002b. “Optimization of Production
Operations in Petroleum Fields”, paper SPE 77658 presented at the SPE Annual
Technical Conference and Exhibition held in San Antonio, Texas, 29 September – 2
October.
WOLSEY, L.A., 1998. Integer Programming, WileyInterscience, USA.
XIAO, W., LIU, Z., JIANG, M., AND SHI, Y., 1998. “Multiobjective Linear
Programming Model on Injection Oilfield Recovery System”, Computers Math.
Applic., Vol. 36, No. 5, pp. 127135.
XIAO, J.J., SHOHAM, O., BRILL, J.P., 1990. “A Comprehensive Mechanistic Model for
TwoPhase Flow in Pipelines”, paper SPE 20631, 65
th
ATCE of SPE, New Orleans,
September 2326.
YETEN, B., DURLOFSKY, L.J., AND AZIZ, K., 2002. “Optimization of
Nonconventional Well Type, Location and Trajectory”, paper SPE 77565 presented
BIBLIOGRAPHY 167
at the 2002 SPE Annual Technical Conference and Exhibition, San Antonio, Texas,
September 29October 2.
ZAKIROV, I., AANONSEN, S., ZAKIROV, E., AND PALATNIK, B.M., 1996.
“Optimizing Reservoir Performance by Automatic Allocation of Well Rates”, paper
presented at the 5th European Conference on the Mathematics of Oil Recovery,
Leoben, Austria, 35 Sept.
ZHANG J., AND ZHU, D., 1996. “A Bilevel Programming Method for Pipe Network
Optimization”, SIAM J. Optimization, Vol. 6, No. 3, pp. 838857, August.
168
Appendix A
Lists of Programs
Two computer programs were developed in the course of this study: a standalone package
called NETwork Simulation and Optimization (NETSO) and a package of subroutines
called Production Optimization Package (POP). Both packages are written in Fortran 77.
NETSO was developed at Stanford University and implements the following simulation
and optimization methods:
• The network simulation and sensitivity coefficient computation procedure described
in Chapter 3.
• The MILPI method for rate allocation described in Section 4.3.2 (The MILPII
method was implemented in another standalone package. The LPI and LPII
methods were implemented in VIPEXECUTIVE (Landmark, 2001)).
• The SQP nonlinear rate allocation method described in Chapter 5.
• The framework for multiobjective optimization of production operations described in
Chapter 7.
POP was developed at BP and implements the following optimization methods:
• The LPI and LPII methods for rate allocation as described in Chapter 4. Pressure
and velocity constraints described in Section 4.4.1 can be handled in this package.
The speedup techniques described in Section 4.4.2 are implemented in this package.
• The partial enumeration method and the genetic algorithm for well connection
optimization described in Chapter 6.
APPENDIX A. LIST OF PROGRAMS 169
• The coupling procedure described in Chapter 8.
POP is a collection of subroutines that heavily depend on the data structure and
subroutines of VIPEXECUTIVE (Landmark, 2001) and is not available for use as a
standalone package. Thus POP is not described further.
In this appendix, we describe NETSO in more detail.
A.1 Description of NETSO
NETSO is a standalone package written in Fortran 77 language. This section presents a
brief description of program files and libraries used in NETSO, simple instructions about
how to compile and run NETSO, and an example input file for NETSO.
A.1.1 Program Files
The program files can be classified roughly into the following categories:
• Simulation and sensitivity coefficient calculation.
• The SQP method for rate allocation.
• The separable programming methods for rate allocation.
• Miscellaneous files that define the global data structure, initialize the data structure,
allocate memory, and perform read and write operations.
A brief description of each group of program files is given below.
Simulation and sensitivity coefficient calculation. Chapter 4 describes the
simulation and sensitivity coefficient calculation procedures implemented in these
program files.
netsim.f the driver for network simulation and sensitivity calculation.
simsol.f uses the NewtonRaphson method to solve the network problem.
calpd.f calculate the BHP and its derivatives from the surface side, called from
simsol.f.
calpu.f calculates the BHP and its derivatives from the reservoir side.
APPENDIX A. LIST OF PROGRAMS 170
fmseqn.f calculates the Jacobian matrix and righthand side of Eq. 3.28.
isconv.f checks convergence of the NewtonRaphson method.
simsen.f computes the sensitivity coefficients.
calpds.f computes the BHP and its derivatives from the surface side, called from
simsen.f.
fluid.f implements functions for calculating fluid properties.
g_fluid.f generated from fluid.f by ADIFOR.
pipe.f implements the drift flux model for multipphase pipe flows.
g_pipe.f generated from pipe.f by ADIFOR.
s_pipe.f computes certain derivatives for sensitivity coefficient calculation.
choke.f implements the Sachdeva et al. two phase choke model.
g_choke.f generated from choke.f by ADIFOR.
s_choke.f computes certain derivatives for sensitivity calculation.
well.f implements a basic well model.
Separable programming method for rate allocation. Chapter 4 describes the
optimization methods implemented in these program files.
ntgpc.f generates well performance curves.
rdpc.f reads performance curves from an input file.
genmps.f generates model MILPI for the rate allocation problem. The model is
written to a file in MPS format (Murtagh, 1981) so that they can be solved
by other standalone mixed integer programming solvers.
sndrv.f implements the MILPI method described in Section 4.3.2. The linear
programming problems are solved by SNOPT. SNOPT implements a
simplex method for solving an LP problem.
fmsnmilp.f initializes the data structure for SNOPT.
snbbsol.f implements the Branch and Bound method described in Section 4.3.2.
APPENDIX A. LIST OF PROGRAMS 171
SQP method for Rate Allocation. Chapter 5 describes the optimization methods
implemented in these program files.
netsof2.f driver for nonlinear optimization of rate allocation problems with
Formulation P2.
ntdata.f setups the data structure for SNOPT with Formulation P2 (Section 5.3.3).
It calls subroutines in files getdim.f, setopa.f, setbnd,f and setidat.f.
getdim.f calculates the sizes of a rate allocation problem . Used for Formulation P2.
setopa.f setups the structure of the Jacobian matrix for an optimization problem.
Initialize the Jacobian matrix with appropriate values. Used for
Formulation P2.
setbnd.f setups the lower and upper bounds for decision variables and constraints.
Used for Formulation P2.
setidat.f initializes certain variables required by SNOPT. Used for Formulation P2.
netobj.f dummy subroutines required by SNOPT. Used for Formulation P2.
netcon.f computes the nonlinear constraint functions and their derivatives (the
Jacobian matrix). Used for Formulation P2.
fmscst.f computes the derivatives of nonlinear constraint functions. Used for
Formulation P2.
setgcon.f transposes the Jacobian matrix. Used for Formulation P2.
setnpar.f numbers the decision variables of an optimization problem (Formulation
P2) or the system parameters for sensitivity coefficient calculation.
netsof1.dat driver for nonlinear optimization of rate allocation problems with
Formulation P1.
fntdata.f same use as ntdata.f except it is for Formulation P1 (Chapter 5).
fgetdim.f same use as getdim.f except it is for Formulation P1.
fsetopa.f same use as setopa.f except it is for Formulation P1.
fsetidat.f same use as fsetidat except it is for Formulation P1.
fgcon.f somputes the objective and constraint functions and their derivatives.
Used for Formulation P1.
APPENDIX A. LIST OF PROGRAMS 172
Miscellaneous files. These program files are used to define the global data structure,
initialize the data structure, allocate memory, and perform read and write.
main.f the main program.
inpobj.h a header file, defines some parameters, variables, and common blocks.
net.h a header file, defines the index variables.
ntbldt.f data block file that initializes the index variables defined in net.h.
pnt.h a header file, defines memory pointers for working arrays.
rd*.f90 a set of Fortran 90 files used to process the input file.
ntinit.f initializes some variables.
chekinp.f checks if the input variables have been properly assigned.
allocm.f allocates memory for working arrays.
ntcnst.f initializes the network structure.
A.1.2 Imported Libraries and Software
NETSO uses several libraries and programs from other sources. These are described
below:
• SNOPT. SNOPT (Gill et al., 1998) is a generalpurpose system for solving large
scale optimization problems. SNOPT implements a sequential quadratic programming
(SQP) method that obtains search directions from a sequence of quadratic
programming subproblems. SNOPT is most efficient if only some of the variables
enter nonlinearly, or if the number of active constraints is nearly as large as number of
variables. SNOPT requires relatively few evaluations of the problem functions, hence
it is especially effective if the objective or constraint functions are expensive to
evaluate. SNOPT was developed at the Stanford Systems Optimization Laboratory
(SOL). More information on SNOPT can be found at http://www.sbsisol
optimize.com/SNOPT.htm.
• ADIFOR. ADIFOR (Bischof et al., 1998) is a tool for the automatic differentiation of
functions written in Fortran 77. Given a Fortran 77 source code and a set of user
specified dependent and independent variables, ADIFOR will generate an augmented
derivative code that computes the partial derivatives of all of the specified output
APPENDIX A. LIST OF PROGRAMS 173
variables with respect to all of the specified independent variables in addition to the
original results. ADIFOR is a collaborative project between the Mathematics and
Computer Science division at Argonne National Laboratory and the Center for
Research on Parallel Computation at Rice University. More information on ADIFOR
can be found at http://wwwunix.mcs.anl.gov/autodiff/ADIFOR/.
• LAPACK. LAPACK (Anderson et al., 1999) is a standard linear algebra package,
which can be downloaded from http://www.netlib.org/.
• The PNNL Fortran Library. The Pacific Northwest National Laboratory (PNNL)
Fortran library is a set of general purpose Fortran 90 routines, which can be freely
downloaded from http://www.pnl.gov/berc/flib/index.htm. This package was used in
NETSO for processing input files.
A.1.3 Compile and Run
NETSO is written in Fortran77 language and was developed on a Silicon Graphics
Origin 200 workstation. All the source files reside in a subdirectory with a compile file
named makefile. The libraries reside in other subdirectories that are specified in the
makefile.
To compile NETSO, simply key in the gmake command to execute the instructions
specified in file makefile. The created executable file is named netso.
To run NETSO, an input file named input.dat is required. An example input file is
presented in the next section. If the run is to solve a rate allocation optimization using
SNOPT, a file named netso.spc can be supplied, though it is not required by SNOPT. The
keywords that can be used in file netso.spc are defined in Gill et al. (1998).
A.1.4 Example Input File
The following is the input file for the example problem presented in Section 3.7.
input.dat
C ***************************************************
C Fluid and rock property
APPENDIX A. LIST OF PROGRAMS 174
C ***************************************************
OIL
API 31.0
GAS
sgpg 0.8
WATER
SGW 1.0
KROWTAB
SW KRW KROW PCOW
0.12000 0.00000 1.00000 0.00000E+00
0.13000 0.00000 1.00000 0.00000E+00
0.92000 1.00000 0.00000 0.00000E+00
KROGTAB
SG KRG KROG PCOG
0.00000 0.00000 1.00000 0.00000E+00
0.00100 0.00000 1.00000 0.00000E+00
0.02000 0.00000 0.99700 0.00000E+00
0.05000 0.00500 0.98000 0.00000E+00
0.12000 0.02500 0.70000 0.00000E+00
0.20000 0.07500 0.35000 0.00000E+00
0.25000 0.12500 0.20000 0.00000E+00
0.30000 0.19000 0.09000 0.00000E+00
0.40000 0.41000 0.02100 0.00000E+00
0.45000 0.60000 0.01000 0.00000E+00
0.50000 0.72000 0.00100 0.00000E+00
0.60000 0.87000 0.00010 0.00000E+00
0.70000 0.94000 0.00000 0.00000E+00
0.85000 0.98000 0.00000 0.00000E+00
1.00000 1.00000 0.00000 0.00000E+00
C ***************************************************
C Gathering system components
C ***************************************************
PIPES
NN NAME DIAMET LENGTH DEPTH TEMPUP TEMPDW ROUGHN
C inch foot foot F F in
1 TUB1 3.0 8000.0 8000.0 160 110 0.0003
2 TUB2 3.5 8000.0 8000.0 160 110 0.0003
3 TUB2 3.5 7500.0 7500.0 160 110 0.0003
4 TUB2 3.5 7800.0 7800.0 160 110 0.0003
5 TUB2 3.5 8000.0 8000.0 160 110 0.0003
6 TUB2 3.5 7700.0 7700.0 160 110 0.0003
7 TUB2 3.5 8500.0 8000.0 160 110 0.0003
8 TUB2 3.5 7000.0 7000.0 160 110 0.0003
9 TUB2 3.5 8000.0 8000.0 160 110 0.0003
10 TUB10 3.5 8000.0 8000.0 160 110 0.0003
11 P1 3.0 500.0 0.0 110 110 0.0003
12 P2 3.5 60.0 0.0 110 110 0.0003
13 P3 3.5 1200.0 0.0 110 110 0.0003
14 P4 3.5 300.0 0.0 110 110 0.0003
15 P5 3.5 400.0 0.0 110 110 0.0003
16 P6 3.5 700.0 0.0 110 110 0.0003
17 P7 3.5 1800.0 0.0 110 110 0.0003
18 P8 3.5 2000.0 0.0 110 110 0.0003
19 P9 3.5 1000.0 0.0 110 110 0.0003
20 P10 3.5 600.0 0.0 110 110 0.0003
21 P11 5.5 4000.0 0.0 110 110 0.0003
C
VALVE
NN NAME DIAMET DISC TEMPUP TEMPDW CTRTYP
C 1/64th inch deg F deg F DIAM/DP/FIXED
1 C1 48 0.85 110 110 DIAM
APPENDIX A. LIST OF PROGRAMS 175
2 C2 48 0.85 110 110 DIAM
3 C3 48 0.85 110 110 DIAM
4 C4 48 0.85 110 110 DIAM
5 C5 48 0.85 110 110 DIAM
6 C6 48 0.85 110 110 DIAM
7 C7 48 0.85 110 110 DIAM
8 C8 48 0.85 110 110 DIAM
9 C9 48 0.85 110 110 DIAM
10 C10 48 0.85 110 110 DIAM
C
C Reservoir well grid blocks
BLOCK
NN NAME DX DY DZ KX KY SW SG PRES TEMP
1 B1 100 100 30 50 50 0.6 0.1 3800. 160.0
2 B2 100 100 30 100 100 0.3 0.4 3400. 160.0
3 B3 200 100 50 70 70 0.5 0.15 4500 160.0
4 B4 100 150 40 500 500 0.7 0.0 3900 160.0
5 B1 100 100 30 50 50 0.25 0.4 4800. 160.0
6 B2 100 100 30 100 100 0.2 0.3 3400. 160.0
7 B3 200 100 50 70 70 0.5 0.25 3500 160.0
8 B4 100 150 40 500 500 0.2 0.5 4900 160.0
9 B3 200 100 50 70 70 0.2 0.45 4200 160.0
10 B4 100 150 40 500 500 0.4 0. 3900 160.0
C
C Well flow rate or pressure constraints can be imposed here
C LFTTYP has following choices
C AUTO: used in option OPTF1, automatic allocation of lift gas.
C QLFT is the upper limit
C FIXED: QLFT is the fixed amount
WELL
NN NAME IBLOC DIAM SKIN PMIN QOM QGM QWM QLM LFTTYP QLFT
1 W1 1 0.400 1 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 3000
2 W2 2 0.475 0 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 0
3 W1 3 0.400 1 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 2000
4 W2 4 0.475 0 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 3000
5 W1 5 0.400 1 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 0
6 W2 6 0.475 0 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 0
7 W1 7 0.400 1 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 2000
8 W2 8 0.475 0 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 0
9 W1 9 0.400 1 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 0
10 W2 10 0.475 0 1000 1E23 1.0E15 1.0E15 1.0E16 AUTO 3000
C
C Nodes number, node name, and constraints
C nodes flowrate and pressure constraints can be imposed here
NODES
NN NAME PMIN QOM QGM QWM QLM
1 H1 50 1.0E15 1.0E15 1.0E15 1.0E15
2 H2 50 1.0E15 1.0E15 1.0E15 1.0E15
3 H3 50 1.0E15 1.0E15 1.0E15 1.0E15
4 H4 50 1.0E15 1.0E15 1.0E15 1.0E15
5 H5 50 1.0E15 1.0E15 1.0E15 1.0E15
6 H6 50 1.0E15 1.0E15 1.0E15 1.0E15
7 H7 50 1.0E15 1.0E15 1.0E15 1.0E15
8 H8 50 1.0E15 1.0E15 1.0E15 1.0E15
9 H9 50 1.0E15 1.0E15 1.0E15 1.0E15
10 H10 50 1.0E15 1.0E15 1.0E15 1.0E15
11 I1 50 1.0E15 1.0E15 1.0E15 1.0E15
12 I2 50 1.0E15 1.0E15 1.0E15 1.0E15
13 I3 50 1.0E15 1.0E15 1.0E15 1.0E15
14 I4 50 1.0E15 1.0E15 1.0E15 1.0E15
15 I5 50 1.0E15 1.0E15 1.0E15 1.0E15
APPENDIX A. LIST OF PROGRAMS 176
16 I6 50 1.0E15 1.0E15 1.0E15 1.0E15
17 I7 50 1.0E15 1.0E15 1.0E15 1.0E15
18 I8 50 1.0E15 1.0E15 1.0E15 1.0E15
19 I9 50 1.0E15 1.0E15 1.0E15 1.0E15
20 I10 50 1.0E15 1.0E15 1.0E15 1.0E15
21 M1 50 1.0E15 1.0E15 1.0E15 1.0E15
22 SEP 160 1.0E15 7.0E24 8.0E23 1.0E15
C ***************************************************
C Gathering system structure
C ***************************************************
C For wells there are three types of STYP
C SEQ: fixed open well choke
C SIEQ: adjustable well choke
C SHUT: well is shutdown
C for nodes, there are two types of STYP
C BND: boundary nodes, pressure are specified
C this must be the root of a production tree
C CON: connection/junction node, pressure are continuous
C at this node
C ===================================================
WELCON
NN OUTNOD OUTCON OUTCNT STYP
1 1 1 PIPE SIEQ
2 2 2 PIPE SIEQ
3 3 3 PIPE SIEQ
4 4 4 PIPE SIEQ
5 5 5 PIPE SIEQ
6 6 6 PIPE SIEQ
7 7 7 PIPE SIEQ
8 8 8 PIPE SIEQ
9 9 9 PIPE SIEQ
10 10 10 PIPE SIEQ
NODCON
NN OUTNOD OUTCON OUTCNT STYP
1 11 1 CHOKE CON
2 12 2 CHOKE CON
3 13 3 CHOKE CON
4 14 4 CHOKE CON
5 15 5 CHOKE CON
6 16 6 CHOKE CON
7 17 7 CHOKE CON
8 18 8 CHOKE CON
9 19 9 CHOKE CON
10 20 10 CHOKE CON
11 21 11 PIPE CON
12 21 12 PIPE CON
13 21 13 PIPE CON
14 21 14 PIPE CON
15 21 15 PIPE CON
16 21 16 PIPE CON
17 21 17 PIPE CON
18 21 18 PIPE CON
19 21 19 PIPE CON
20 21 20 PIPE CON
21 22 21 PIPE HEAD
C
C Define boundary nodes
NODCON
NN STYP
22 BND
C
APPENDIX A. LIST OF PROGRAMS 177
C****************************************************
C Total gaslift volume, only used in option OPTF1
C****************************************************
LFTVOL 3e23
C ***************************************************
C Optimization according to operation policies
C ***************************************************
INTVAR
NVAR LBND UBND XINIT
1 0 4000 300
2 0 4000 400
3 0 4000 1200
4 0 4000 400
5 0 4000 500
6 0 4000 1000
7 0 4000 20
8 0 2000 400
9 0 2000 400
10 0 5000 1500
DVAR
NVAR DEVTYP IDEV PAR LBND UBND XINIT
1 WELL 1 GLFT 0 6000 3000
2 WELL 2 GLFT 0 6000 3000
3 WELL 3 GLFT 0 6000 3000
4 WELL 4 GLFT 0 6000 3000
5 WELL 5 GLFT 0 6000 3000
6 WELL 6 GLFT 0 6000 3000
7 WELL 7 GLFT 0 6000 3000
8 WELL 8 GLFT 0 6000 3000
9 WELL 9 GLFT 0 6000 3000
10 WELL 10 GLFT 0 6000 3000
C
HCON
NCON DEVTYP IDEV PAR LBND UBND
c 1 NODE 22 OIL 14000 14000
1 NODE 22 WAT 0 6400
2 NODE 22 GAS 0 900000
C
GOAL 1
NOBJ DEVTYP IDEV PAR OBJ TARG COEF
1 FIELD 0 OIL MAX 2000 1
GOAL 2
NOBJ DEVTYP IDEV PAR OBJ TARG COEF
1 FIELD 0 GAS MAX 2000 1
GOAL 3
NOBJ DEVTYP IDEV PAR OBJ TARG COEF
1 FIELD 0 WAT MIN 2000 1
C
C define the jobs
JOBS
RANK IGOAL MAX RELF
1 1 YES 0.8
c 2 3 NO 0.1
C
SOLVER SEN
178
Appendix B
Problem Data for the Gaslift Example in
Section 4.4.3
Table B.1: Gas injection and oil production rates for a set of 56 wells
(from Buitrago et al. (1996)).
Well Lift Gas Rate, MSCF/d
Oil Rate, STB/d
1 0. 225. 314. 672. 768. 1485.
0. 290. 324. 386. 391. 433.
2 0. 68. 144. 266. 450. 693. 861. 1035.
487. 530. 560. 591. 626. 652. 665. 676.
3 0. 60. 129. 210. 348. 521. 748. 792. 1027.
441. 481. 516. 541. 575. 605. 629. 632. 651.
4 0. 42. 86. 542. 945.
280. 284. 287. 307. 315.
5 0. 35. 71. 464. 1164.
281. 285. 289. 306. 315.
6 0. 74. 157. 294. 505. 890. 926. 1114.
287. 315. 333. 352. 366. 376. 376. 379.
7 0. 114. 235. 386. 577. 1279. 1507.
790. 815. 836. 857. 876. 908. 914.
8 0. 58. 97. 268. 651. 1725.
0. 209. 233. 276. 297. 311.
9 0. 551. 716. 1295. 1706.
0. 1038. 1199. 1568. 1424.
10 0. 66. 135. 215. 308. 413. 553. 756. 975. 1053.
233. 240. 246. 252. 258. 263. 268. 274. 279. 281.
11 0. 91. 201. 397. 675. 1054. 1378. 1740.
548. 656. 727. 814. 896. 958. 987. 1021.
12 0. 617. 800. 1345. 1513. 2880. 6765.
0. 459. 510. 590. 614. 679. 687.
APPENDIX B. PROBLEM DATA FOR THE GASLIFT EXAMPLE 179
Table B.1: Gas injection and oil production rates for a set of 56 wells (cont).
Well Lift Gas Rate, MSCF/d
Oil Rate, STB/d
13 0. 16. 33. 181. 1218.
108. 109. 111. 118. 137.
14 0. 128. 186. 413. 1105. 1111.
0. 277. 302. 336. 361. 360.
15 0. 65. 153. 252. 381. 598. 884. 1112. 1392.
351. 420. 493. 549. 594. 648. 688. 709. 735.
16 0. 40. 69. 253. 460. 926. 982. 907.
0. 170. 198. 310. 361. 415. 417. 455.
17 0. 128. 262. 370. 494. 647. 823. 1081. 2174.
892. 909. 931. 946. 960. 975. 986. 1001. 1029.
18 0. 138. 282. 468. 705. 1027. 2297.
1151. 1185. 1213. 1240. 1266. 1290. 1317.
19 0. 86. 175. 295. 469. 686. 943. 1362.
310. 317. 323. 329. 335. 342. 347. 353.
20 0. 136. 215. 465. 541. 975. 1948. 4225. 12161.
0. 162. 213. 303. 316. 391. 458. 496. 476.
21 0. 189. 289. 772. 894. 1479.
0. 251. 307. 455. 466. 523.
22 0. 270. 370. 727. 816. 1929.
0. 195. 214. 234. 234. 237.
23 0. 110. 221. 1669.
944. 948. 951. 957.
24 0. 49. 105. 182. 283. 399. 556. 725. 767. 1030. 3621.
956. 1062. 1149. 1240. 1338. 1420. 1505. 1578. 1594. 1680. 1955.
25 0. 51. 103. 160. 225. 551. 839.
487. 495. 502. 509. 516. 536. 548.
26 0. 54. 120. 200. 319. 579. 734. 1023. 1972. 3317.
82. 94. 105. 113. 119. 127. 129. 133. 141. 139.
27 0. 54. 109. 590. 1100.
353. 354. 355. 358. 357.
28 0. 105. 211. 1618.
1044. 1051. 1056. 1079.
29 0. 69. 141. 224. 322. 439. 579. 754. 760.
184. 188. 191. 195. 198. 201. 204. 206. 206.
30 0. 56. 112. 545. 561.
308. 309. 310. 309. 308.
31 0. 37. 75.
354. 354. 354.
32 0. 64. 131. 217. 1093.
618. 638. 654. 664. 728.
33 0. 184. 280. 682. 897. 1646. 4485.
0. 185. 211. 257. 265. 291. 308.
34 0. 31. 63. 366. 1299.
209. 210. 210. 215. 221.
35 0. 40. 86. 144. 195. 283. 389. 515. 717. 801. 1273.
162. 177. 190. 207. 216. 229. 242. 251. 263. 267. 281.
36 0. 28. 59. 108. 166. 235. 330. 404. 1145.
179. 189. 195. 204. 213. 220. 227. 232. 253.
APPENDIX B. PROBLEM DATA FOR THE GASLIFT EXAMPLE 180
Table B.1: Gas injection and oil production rates for a set of 56 wells (cont).
Well Lift Gas Rate, MSCF/d
Oil Rate, STB/d
37 0. 34. 71. 126. 199. 292. 409. 564. 772. 1346.
64. 67. 69. 72. 74. 77. 79. 82. 84. 87.
38 0. 112. 157. 345. 690. 1891.
0. 270. 282. 310. 326. 339.
39 0. 105. 168. 301. 800. 860.
0. 131. 174. 207. 246. 251.
40 0. 98. 116. 186. 330. 757. 1149.
0. 27. 28. 29. 31. 35. 38.
41 0. 63. 126. 636. 801.
372. 373. 373. 374. 374.
42 0. 28. 57. 361. 1701.
200. 201. 202. 207. 201.
43 0. 538. 797. 976. 1300. 2920. 3650.
0. 256. 337. 362. 407. 467. 198.
44 0. 61. 123. 666. 1303.
397. 403. 406. 422. 427.
45 0. 11. 21. 137. 712.
83. 83. 83. 87. 93.
46 0. 7. 14. 22. 31. 40. 50. 61. 91. 700.
47. 48. 50. 51. 52. 53. 55. 56. 59. 80.
47 0. 1558. 2653. 2849. 2982. 3042.
0. 0. 0. 0. 394. 441.
48 0. 1450. 1633. 2466. 5096.
0. 0. 197. 483. 490.
49 0. 1228. 1268. 1418. 1690.
0. 0. 0. 232. 248.
50 0. 1135. 1441. 2538. 3980.
0. 0. 146. 188. 188.
51 0. 162. 369. 2009. 2224. 3488. 4899.
0. 0. 0. 0. 223. 232. 237.
52 0. 1512. 2536. 2671. 2750. 2830.
0. 0. 0. 194. 260. 317.
53 0. 1174. 1304. 1484.
0. 0. 186. 267.
54 0. 1559. 2340. 2594. 2650. 3342.
0. 0. 0. 278. 300. 328.
55 0. 465. 1333. 2169. 2317. 2445. 2596. 3100.
0. 0. 0. 0. 152. 160. 167. 184.
56 0. 761. 1247. 1590. 1655. 1730. 1770. 2300. 5360.
0. 0. 0. 350. 403. 444. 452. 505. 658.
APPENDIX B. PROBLEM DATA FOR THE GASLIFT EXAMPLE 181
Table B.2: Gas injection and oil production rates for a set of 56 wells obtained
from the MILPII method.
Well Oil Rate
STB/d
Gas Rate
MSCF/d
Well Oil Rate
STB/d
Gas Rate
MSCF/d
Well Oil Rate
STB/d
Gas Rate
MSCF/d
1 386 672 20 391 975 39 207 301
2 626 450 21 455 772 40 27 98
3 605 521 22 214 370 41 372 0
4 280 0 23 944 0 42 200 0
5 281 0 24 1680 1030 43 337 797
6 333 157 25 487 0 44 397 0
7 836 235 26 105 120 45 83 0
8 276 268 27 353 0 46 50 14
9 1568 1295 28 1044 0 47 441 3042
10 233 0 29 184 0 48 483 2466
11 957 1048 30 308 0 49 232 1418
12 510 800 31 354 0 50 0 0
13 108 0 32 654 131 51 0 0
14 302 186 33 211 280 52 0 0
15 648 598 34 209 0 53 267 1484
16 361 460 35 216 195 54 0 0
17 892 0 36 204 108 55 0 0
18 1213 282 37 64 0 56 452 1770
19 310 0 38 282 157
© Copyright by Pengju Wang 2003 All Rights Reserved
ii
I certify that I have read this thesis and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy.
__________________________________
Dr. Khalid Aziz (Principal Advisor)
I certify that I have read this thesis and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy.
__________________________________
Dr. Roland N. Horne
I certify that I have read this thesis and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy.
__________________________________
Dr. Anthony Kovscek
Approved for the University Committee on Graduate Studies:
_________________________________________
iii
While some aspects of this problem have been studied by other investigators. it may lead to bad solutions when the flow interactions among wells are significant. the rate allocation problem can be formulated as a separable programming (SP) problem whose objective and constraint functions are sums of functions of one variable. The SP problem is solved by various linear optimization techniques. We then focused on the rate allocation problem. and optimization of a petroleum field. the well performance can be analyzed individually. However.Abstract For some petroleum fields. the rate allocation problem iv . sensitivity analysis. In this work we investigated the formulations and solution methods for the following optimization problem: determining the optimal production rates. optimization of production operations can be a major factor on increasing production rates and reducing production cost. When flow interactions among different wells are not significant. which refers to the determination of the optimal production and lift gas rates subject to multiple constraints. Consequently. This method is very efficient. it is necessary to develop approaches that can solve the problem efficiently and without unreasonable assumptions. Hence. In such cases. and well connections for a gathering system with treelike structures to maximize daily operational objectives subject to multiple flow rate and pressure constraints. existing methods are either inefficient or they are based on significant simplifications that lead to suboptimal solutions. lift gas rates. We first investigated efficient procedures for simulating the production system and computing sensitivity coefficients of production rates relative to system parameters. These procedures have useful applications in simulation.
this study demonstrated that multiobjective optimization methods can help decision makers identify the best tradeoffs. lift gas rates. Results demonstrated the effectiveness of the approach. Production engineers often strive to achieve multiple conflicting goals when operating a field. The suggested formulation is able to handle well shutdown. and is computationally efficient. v . a lower level problem is formed to determine the optimal set of production and lift gas rates. The upper level masters the overall solution process and explicitly optimizes the well connections. The coupled procedure was applied to shortterm production optimization in the Prudhoe Bay field of the NorthSlope of Alaska. and to longterm reservoir development studies for two fields in the Gulf of Mexico.should be appropriately formulated so that simulations capable of capturing such flow interactions can be conducted in the optimization process. as verified by a genetic algorithm. A twolevel programming approach is developed to optimize the production rates. avoid some numerical difficulties. The lower level problem is a rate allocation problem that can be solved by various optimization methods developed in this study. Several existing multiobjective optimization methods are used to address this problem. the optimization problem was solved by a sequential quadratic programming (SQP) algorithm. The optimization tools were coupled with models for multiphase fluid flow in reservoirs and surface pipeline networks through a commercial reservoir simulator. For each new set of well connections explored in the upper level. and well connections simultaneously. In this approach. The method is both efficient and robust. the entire optimization problem is solved in two levels. Once formulated. The upper level problem is solved by a newly developed heuristic method. Several formulations were investigated. Through an example.
Khalid Aziz for his sharing of my excitement as well as frustrations in this research.3. encouragement and guidance. Sincere thanks also go to other BP colleagues who in one way or another contributed to my pleasant summer internship experiences at BP. I thank him for his help and guidance. a general purpose nonlinear constrained optimization package. Louis Durlofsky for providing valuable advice on various occasions during this study. Roland Horne and Dr. This support is gratefully acknowledged. I sincerely thank Dr. Part of this work was conducted in BP. I thank the systems optimization laboratory at Stanford University for providing me SNOPT 5. vi . The reservoir simulation industrial affiliates program (SUPRIB) and the Department of Petroleum Engineering have supported me financially. an automatic differentiation software. I thank Mathematics and Computer Science Division at Argonne National Laboratory and the Center for Research on Parallel Computation at Rice University for providing me ADIFOR 2. I also would like to thank Dr. Without his constant support. Michael Litvak provided tremendous help in integrating the optimization tools into VIP (a commercial reservoir simulator of Landmark Graphics Corporation) and testing the optimization methods on several field models. Anthony Kovscek for reading the dissertation and providing valuable comments. Dr. I also thank BP for providing me two summer internship opportunities and the permission to put some field examples in this dissertation.0. this work wouldn’t be possible.Acknowledgments I am immensely grateful to my advisor Dr.
I deeply thank my parents. This dissertation is dedicated to you. I especially thank my friend Tao Shu for supporting me in many ways and bringing so much joy to my life for the last six years. staff. Jingmei Liu and Xingchun Wang. vii . Mom and Dad. They contributed greatly to making my stay at Stanford enjoyable. and fellow students in the Petroleum Engineering Department.I am grateful for the friendship of and help from the faculty. Their love and support are a constant inspiration to me in striving to improve myself.
.....1 Overview of Optimization Techniques.....................3 Outline of Dissertation............ 9 2........... Introduction.......................2...............................1 Linear Programming.......Contents Abstract ............................ 5 2...................................................................................................................................................... 10 2.........................3 Optimization of Reservoir Development and Planning ............. 22 viii .................................... 14 2...3 Concluding Remarks ......................................................... Literature Review ....... 1 1.................................................................................................2 Applications of Optimization Techniques to Petroleum Fields........................................... 1 1....................1.................................................. Simulation of Gathering Systems .................................... 20 3................................................................................................................................................................1 Problem Statement..................3 Nonlinear Programming...................2.......................................................................... 8 2.......................................xiv 1............ vi Contents......... 14 2.......................... 9 2..................... 8 2........................................................................viii List of Tables..........1............. iv Acknowledgments....................................xiii List of Figures ...................................... 17 2.........1 Lift Gas and Production Rate Allocation ............2 Integer Programming...........................................................2 Optimization of Production System Design and Operations................................................ 4 1........................... 18 2.............. Approach.................1........................................................2...........2......................................................................................................................................................................
....................................................... 64 ix ...........................1 Governing Equations..... 62 4.............................................1 Handling Pressure and Velocity Constraints.....................2 Model MILPI .......8 Concluding Remarks ..........................................1 Pressure Gradient Equation..........4......... 60 4......................................................4 Application to Rate Allocation Problems ..................... 22 3............................................ 63 4............................1 Simulation Results ..........7......... 25 3..6 Sensitivity Coefficients........ 40 3.................................................... 27 3................................3 Pipe Model..................3 Model MILPII ......................................2 Speedup Techniques........................................7..... 39 3...........................2 Well Model ..................................2 Well Performance Estimation........................................5........................................................................................3 Upwards Twophase Flow Behavior.........................3..............1 Introduction...................................................................... 31 3.....................................3......... 40 3.......... 51 4....................1 Finite Difference Method ...................4..7 Examples.. 42 3.....................................................................................................3.....................................4 Choke Model ............................................................................................5...............................5 Discussions................................................................................. 62 4........................................................................................................................ 56 4.........................................................1 Model LPI .........................4 Model LPII................3...................................5 Network Simulation......................................................................................2 The NewtonRaphson Method ..............................................................3 Multiple Solutions.................................3......1 Introduction............................................................3... 52 4............ 34 3..........................3 Rate Allocation through Separable Programming .................................................. Rate Allocation through Separable Programming ................................................... 51 4......... 24 3..........................6... 24 3......................... 32 3......................5.....2 Liquid Holdup Evaluation .................... 54 4........ 49 4.....3......................... 29 3... 43 3.............. 26 3.................................. 57 4.............................................................3.......3.............................. 36 3..........................................................2 Computation of Sensitivity Coefficients ... 61 4............................. 46 3...............................................................2 Jacobian Method ....................................................................................6........
................1 Comparison of the MILPI Method and the SQP Method ...2 Partial Enumeration Method ........1 NoPreference Methods............................ 88 6.5................................ 70 5........... 77 5.......................... 86 5............. 71 5..... 105 x .........4 Solving the Optimization Problem .....3.........4.1 Computational Issues ............................................ Rate Allocation through Sequential Quadratic Programming.............................................................................. 68 5................................ 70 5........................... 73 5.......... 102 7..........................2.......... 80 5.............................................5.......................... Multiobjective Optimization of Production Operations.................................................................................... 90 6...5 Concluding Remarks ...........................................1 Concept of Multiobjective Optimization ......................................................................3 Formulation P2..................4......................................................................................... 66 4.......................5 Examples.................3........2 Twolevel Optimization Formulation .................3..........................................................................................................................................................2 Comparison of Formulation P1 and P2 .2 Formulation P1..........5.................1 Introduction....1 Optimization Problem ......... 90 6.......................... 92 6. 84 5................3 Solution Methods for the Optimization of Well Connections ...................1 Introduction.............................................3 Formulations of the Rate Allocation Problem .........................5 Concluding Remarks ................................................................................... 74 5......2 Sequential Quadratic Programming... 79 5.................3..........3 A Gaslift Optimization Example .................................... 98 6..................2 Solution Techniques for Multiobjective Optimization Problems............... 102 7.................................3 Genetic Algorithm................................. 93 6........................ Optimization of Well Connections....................... 94 6........................................... 91 6.................... 101 7....4 Interpretation of Solutions of Problem P2.............................................................3...................................... 80 5........... 92 6.......................... 75 5........................................................................................................3.................................................6 Concluding Remarks .....................4 Comparison of the Partial Enumeration Method and the Genetic Algorithm .....3............................. 73 5............................3 Efficiency of Formulation P2 ... 104 7....................................
...........................3................4 Example .....4 Interfacing with an Independent Reservoir Simulator .......... 118 8...................................2...... 146 9..2 Solution Methods ......................... 111 7..................................................... 117 8.................................................5..........2 Full Field Simulation ...............1 Introduction.................................................................................................. 149 Bibliography........................................................3 Recommendations for Further Study........................................1 A General Integration Procedure............1 A Gulf of Mexico Oil Field Example............................. 108 7........................3 Integration of Optimization Tools with VIPEXECUTIVE .................. 117 8..............1 Simulation of Petroleum Fields...................... 107 7................................................ 128 8.... 127 8........................................................................................................................... 119 8........................ 122 8....................................................... 126 8..............1 Handling Soft Constraints . 105 7....... 145 9...2 A Priori Methods.............................................................................................. Conclusions ................. 124 8........2..... Coupling with a Reservoir Simulator ...................1 Conclusions....................3 Interactive Methods.....2......5 Concluding Remarks ... 148 Nomenclature .................................... List of Programs............................................................................2 Integrated Optimization Tools. 154 Appendix A.............................................................................................. 113 7...................................................2 Summary of Contributions....... 133 8............................3 Application to the Prudhoe Bay Oil Field ............................................................. 128 8........................................2.............. 124 8........................................................................................................................................... 168 xi ..............................................................................3..................................................................................2 Coupling Reservoir and Production Network Solutions ................................................ 110 7.......................................... 143 9................................................5.3....... 139 8...2 Another Gulf of Mexico Field Example ..........................................7...................... 111 7...............5...............6 Concluding Remarks ..................................................... 145 9............... 115 8....3 Multiobjective Optimization of Production Operations .....5 Applications......2....3.4 A Posteriori Methods .......................................................................................................................
......................1....................1...............1 Description of NETSO ..................4. 178 xii ........1 Program Files .......................3 ...................................................... Problem Data for the GasLift Example in Section 4... 169 A..................A...........................................3 Compile and Run................1... 173 Appendix B....................................... 172 A......2 Imported Libraries and Software...... 173 A........................................................................................... 168 A..............1......4 An Example Input File ..............
............................ 46 Table 4....................... 114 Table B...................3: Convergence criteria for optimization problems in Example 3 ...................................................................................................... 181 xiii ........... 28 Table 3.........................................................................2: Network simulation results from different runs ........................1: Slip model problem data ........ 104 Table 7...2: Allocated lift gas and oil rates for Example 1.........................................1: Gas injection and oil production rates for a set of 56 wells ... 82 Table 5...... 99 Table 7................. 83 Table 5................ 67 Table 4................................... 88 Table 5...........2: Computational time of the MILPII method on various problems...................................... 178 Table B.....4: Computational efficiency of Formulation P2 on various optimization problems .......................1: Performance data of the PE method............................................... 88 Table 6.................................1: A classification of methods for multiobjective optimization.......... 45 Table 3................................ 68 Table 5.1: Gaslift allocation results obtained from different methods ........List of Tables Table 3..................................2: Pareto optimal solutions from the weighting function method............1: Well rates obtained in Step 1 of the MILPI method for Example 1 ...............................................................2: Gas injection and oil production rates for a set of 56 wells obtained from the MILPII method ......3: Comparision of sensitivity coefficients obtained from the Jacobian method and the finite difference method ......
....................................................................2: Impact of oil rate and lift gas rate on the well sandface pressure calculated from the separator side ................8: Sensitivity coefficients of well oil rates with respect to system parameters (the inner diameter of pipe 21 is 5..... 38 Figure 3.........................................................................................................................................3: The allocated lift gas and oil rates for well 47 from the MILPII method and the LPI method...9: Sensitivity coefficients of well oil rates with respect to system parameters (the inner diameter of pipe 21 is 7.................... 32 Figure 3........ 85 xiv .... 75 Figure 5.......................... 79 Figure 5............... 68 Figure 5..........................5: Infinity norm of the governing equation residuals versus iteration number .........7: Oil rate versus iteration number for Run 4 ...........4: Configuration of the gathering system....................1: Outflow performance curves with various choke settings ..........................5 inches) ......................................... 61 Figure 4.......... 82 Figure 5........................................ 48 Figure 3.................................. 48 Figure 3... 45 Figure 3..............................................5 inches)...............6: Oil rate versus iteration number for Run 1 ...............5 inches) .3: Illustration of multiple solutions for a single well system .............4: Convergence history of Formulation P1 and P2 for Scenario 1 .......................10: Sensitivity coefficients of well oil rates with respect to system parameters (the inner diameter of pipe 21 is 9.........List of Figures Figure 3..............1: Upstream pressure in a vertical well tubing string with a fixed downstream (wellhead) pressure for different flow rates .........................2: Illustration of gaslift performance curves..................................... 29 Figure 3...............1: Illustration of well performance curves .................................................... 53 Figure 4....................... 44 Figure 3.............................3: Gaslift performance curves generated in Step 2 of the MILPI method......... 44 Figure 3...................................... 49 Figure 4.................... 43 Figure 3....................2: A simple gathering system.......
..7: Normalized daily gas rate at HEADER_Z................... 115 Figure 8............................ 97 Figure 6...........4: Normalized daily total gas rate ................7: Convergence history of the GA run ......................6: Convergence history of the PE runs......... 131 Figure 8................................... 136 Figure 8.................... 132 Figure 8.. 137 Figure 8............... 130 Figure 8.4: Schematic illustration of uniform crossover..............................11: Normalized cumulative lift gas injection for Case 1 ...................................................6: Schematic illustration of gathering systems ............................................. 136 Figure 8.......14: Normalized cumulative lift gas injection for Case 2 ..................3: The weighting functions method with nonconvex set ...... 131 Figure 8..................... 101 Figure 7................ 86 Figure 5......5: Normalized daily total water rate..... 103 Figure 7........1: Integration of optimization tools to VIPEXECUTIVE ............................... 95 Figure 6......................9: Normalized cumulative oil production history for Case1................................ 125 Figure 8................. 100 Figure 6.......4: The Pareto optimal sets obtained from the hierarchical method and the weighting functions method..................................... 137 Figure 8.... 139 xv .......................................... 110 Figure 7...........................................1: A simple GA flowchart.............................. 129 Figure 8....................... 97 Figure 6.........................................8: Normalized daily total oil production history for Case 1................. 96 Figure 6. 138 Figure 8.......2: Encoding/decoding a well connection ...................... 109 Figure 7..................13: Normalized cumulative oil productive history for Case 2 ..............................6: Normalized daily gas rate at HEADER_Y ..........................................3: Normalized daily total oil rate ..............2: The production system of a Gulf of Mexico oil field .........5: Schematic illustration of mutation..................................5: Convergence history of Formulation P1 and P2 for Scenario 2 .....................................................................1: Graphical illustration of a Pareto optimal set .............................12: Normalized daily oil production history for Case 2.........................................................2: The weighting functions method with convex set ...................10: Daily total gas rate history for Case 1...........Figure 5...................................................................................3: Schematic illustration of singlepoint crossover....................................................... 97 Figure 6................... 87 Figure 6..................................... 138 Figure 8................ 132 Figure 8...........................................
fluid handling capacity of facilities. hydrocarbon production is often constrained by reservoir conditions. 1. While for single wells or other small systems simple nodal analysis may be adequate. 1 . deliverability of the pipeline network.. or a combination of these considerations. The objective of this study was to develop optimization methods for solving a class of important production operation problems for petroleum fields. upstream oil and gas facilities previously thought not to be candidates for advanced control or optimization are being given new considerations (Clay et al.Chapter 1 Introduction The ultimate goal of virtually all effort spent on modeling a petroleum field is to devise an optimal strategy to develop. For some petroleum fields. and operate the field. optimization of production operations can be a major factor in increasing production rates and reducing production costs. safety and economic considerations. The task of field operators is to devise optimal operating strategies to achieve certain operational goals. 1998).1 Problem Statement In petroleum fields. manage. As optimization algorithms and reservoir simulation techniques continue to develop and computing power continues to increase. large complex systems demand a much more sophisticated approach to predict the response of a large complicated production system accurately and to examine alternative operational scenarios efficiently.
the production rates. the control variables. . once determined. Control variables. However the optimization procedures developed in this study can be used repeatedly for longterm hydrocarbon recovery studies. INTRODUCTION 2 These goals can vary from field to field and with time. Usually. How to define the profit is a complicated issue that requires intensive study in its own right. In this study. The flow rate and other operational settings. the objective is to maximize the profit from an oil field on a daytoday basis. the control variables included the lift gas rates. Gas is injected into the tubing string to lighten the liquid column and decrease the bottom hole pressure. are assumed to remain fixed during that period. which allows the reservoir to push more fluids into the wellbore. increased flow rates in the tubing string and surface flow lines result in higher backpressure on the well and adjacent wells that share a common flow line. lift gas has to be carefully allocated to achieve maximum efficiency. At the same time. and the constraints) of the petroleum field optimization problem are described in detail. Therefore. Continuous gaslift is a common artificial lift method used in the oil industry to improve well performance. This in turn causes a reduction in well production rates.CHAPTER 1. we restricted our attention to simple objective functions such as maximizing weighted daily flow rates. the major components (the objective. To keep this research focused. The production rate of a well is usually controlled by a choke. The mechanism of gaslift is fairly simple. This research aims to develop optimization methods that ease and automate the decision making of field operators for certain operations. We emphasize that we only optimize production operations for a shortterm period. The control variables are the production operation settings to be optimized. In this section. and the well connections to flow lines. Objective. Adjusting production rates is the most straightforward way to meet certain production targets and satisfy certain operational constraints as described later. Typically one may wish to maximize daily oil rates or minimize production costs.
In summary. corrosion/erosion can lead to costly repairs. fluid velocities may have to be limited (Svedeman and Arnold. In such fields. 1992. Safety and economic constraints. safety.CHAPTER 1. oil production in the Prudhoe Bay oil field (Barnes et al. Constraints. 1999) is constrained by the gas handling capacities of surface facilities. switching a well from one flow line to another flow line can be an effective way to relieve the delivering/processing burden of one device/facility and increase the delivering/processing efficiency of the overall system. The main reason for this is that production systems are usually designed and installed at the very early stage of reservoir development when information about the reservoir and future economics are scarce or uncertain. All these constraints can be satisfied by adjusting the production rates with chokes. we may need a maximum/minimum pressure constraint at the bottom of a well or on some surface facility nodes. To operate the production system economically. INTRODUCTION 3 In some petroleum fields such as the Prudhoe Bay oil field in Alaska. such as the gas and water processing capacities of separators. we may put a maximum/minimum flow rate constraints on certain production wells or facilities. Kermani and Harrop. To avoid excessive corrosion/erosion. Production operations in an oil field are usually subject to multiple capacity. It is not surprising to know that oil production in some petroleum fields is constrained by the processing capacities of surface facilities. .. 1996). Thus. For safety reasons. safety. 1994. For example. a production well can be connected to different flow lines that lead to different separation units. 1990) and the Kuparuk River field in Alaska (Stoisits et al.. Thus facility capacities may not always be able to meet the production demand throughout the life of a reservoir. 1994. and the gas compression capacity of a central gas plant. Capacity constraints. USA. Furthermore. well connections were also considered as one type of decision variables in this study. These constraints can be satisfied and/or fully utilized by adjusting the lift gas and/or production rates and switching well connections between flow lines. our objective was to investigate optimization approaches to guide in making operational decisions to enhance production subject to multiple capacity. and economic constraints.
As a consequence. The target problem is a nonlinearly constrained optimization problem with both continuous and discrete decision variables. Because the continuous and discrete variables are optimized at different levels. In this approach. as discussed in Chapter 2. The upper level explicitly optimizes the well connections. For example the engineer may be interested in the following questions: 1. A twolevel programming approach was developed to solve the problem. When there are no flow interactions among different wells or when such interaction is not significant. However. Hence. This dissertation investigated such approaches. lift gas rates. the upper level spawns a lower level problem to find the optimal production and lift gas rates for that set of well connections. How to route fluids by switching well connections to flow lines? 1.CHAPTER 1.2 Approach The problem of interest requires simultaneous allocation of production rates. For each set of well connections. or ad hoc rules have been used that may lead to suboptimal operations. The upper level masters the overall optimization procedure and the lower level can be viewed as a function evaluation procedure for the upper level. robust procedures for such a task are not available. The advantage of this twolevel programming approach is flexibility. How to distribute available liftgas among specified wells? 3. and well connections to flow lines. the rate allocation . The lower level problem optimizes the production and lift gas rates subject to nonlinear constraints and is referred to in this study as the rate allocation problem. it was necessary to develop approaches that optimize all control variables simultaneously subject to all constraints. the approach offers us more freedom to use existing methods or develop new methods for solving the overall problem. How to use chokes to control well rates? 2. the performance of a well can be analyzed individually while ignoring the impact of the rest of the wells in the system. Either previous investigators have addressed only a part of the problem. INTRODUCTION 4 and economic constraints. optimization is conducted in two levels.
. integer programming.CHAPTER 1. and nonlinear programming. This method is an iterative procedure that sequentially finds the best connection for one well while keeping other well connections fixed. The upper level is an integer programming problem. 1. The optimization approaches for the upper level and lower level were investigated separately. The flow interaction among wells can play an important role in some rate allocation problems. The chapter then describes the applications of optimization techniques in the development and operations of petroleum fields and sets the framework of this research. The preferred formulation chooses the production rates instead of the well chokes as the decision variables and uses a set of constraints to guarantee that the optimal set of production rates is feasible for the production system. solution algorithms. which introduces the concept. and evolves multiple sets of binary strings by means of selection. The SP problem can be solved efficiently by linear optimization techniques.. The first is a heuristic method named in this study the partial enumeration method. 1981). INTRODUCTION 5 problem becomes a separable programming (SP) problem whose objective and constraint functions are sums of functions of one variable (Gill et al. Two methods were investigated. crossover. The second method is a genetic algorithm (Goldberg. 2002).. In such cases.3 Outline of Dissertation General optimization techniques pertinent to this study are reviewed in Chapter 2. The approaches can be used separately to optimize single types of operation settings such as the lift gas distribution or combined together to optimize all types of variables simultaneously. The focus is on how to handle the well chokes that control the production rates in the optimization process. 1989). and mutation to locate the best set of well connections. This method encodes one set of well connections into one set of binary strings. and applications of linear programming. the rate allocation problem is formulated as a general nonlinear constrained optimization problem and solved by a Sequential Quadratic Programming method (Gill et al. Different formulations have been investigated.
there is no generally accepted method for simulating multiphase flow in general gathering systems. an efficient solution method could be investigated. how to formulate the rate allocation problem into a separable problem. and identifying the true physical solution from multiple mathematical solutions. It was found that care should be taken to identify the stable solution from multiple mathematical solutions. the study considered a simplified class of problems where the gathering system has a treelike structure and flow directions are known. Once appropriately formulated. This chapter discusses the appropriate formulation of the rate allocation problem under such cases. the problem was solved by a Sequential Quadratic Programming (SQP) method. Chapter 5 describes how the rate allocation problem can be solved when flow interaction can not be ignored. Chapter 7 discusses possible applications of multiobjective optimization techniques to production operations for a petroleum field. By exploiting the special properties of such systems. An example is presented to demonstrate the advantages of the proposed method. Rather. The quality of the solution by the simpler partial enumeration method was at least as good as that obtained from the genetic algorithm. determining the phase splitting ratios at flow junctions. finding the mathematical solutions efficiently. In the oil industry. Further. an efficient method was developed to compute sensitivity coefficients for a production system. production operations in a .CHAPTER 1. Numerical experiments showed that the partial enumeration method is more efficient than the genetic algorithm. This method has useful applications in sensitivity analysis and optimization. INTRODUCTION 6 Chapter 3 describes how the gathering system is simulated. Often. Chapter 4 describes how the rate allocation problem can be solved when flow interaction among wells can be ignored. This section describes how to construct the well performance information. Chapter 6 presents the two solution methods for the upper level problem: the partial enumeration method and the genetic algorithm. This research did not address all these challenges. Application examples demonstrated that the method is capable of handling rate allocation problems of varying complexities and sizes. The challenges lie in predicting the pressure drop across flow delivering devices. and how to solve the separable problem by linear optimization techniques.
Chapter 8 describes a procedure to integrate the optimization approaches with a reservoir simulator.CHAPTER 1. 2002).. This procedure automates the use of the approaches developed for both shortterm production optimization and longterm oil recovery studies. Multiobjective optimization is a useful tool to help decision makers identify the best tradeoffs. One example is the use of the optimization tools in the EField Optimization System of Prudhoe Bay oil field. INTRODUCTION 7 petroleum field are subject to multiple conflicting goals. Chapter 9 summarizes the work and recommends areas that need further study. Alaska (Litvak et al. Chapter 8 presents several such applications. such as maximizing the daily oil production while minimizing the water produced. .
. integer programming. Applications have been reported for recovery processes. nonlinear programming. well placement and operation. integer programming.1 Overview of Optimization Techniques Mathematical programming is a field born in the later 1940s (Lenstra et al. 2. mathematical programming has developed into a sophisticated field with deep specialization and great diversification.1) 8 .. history matching. In this chapter. stochastic programming. solution algorithms... facility design and operation and so on. i = 1. Despite its short history. m} (2. and nonlinear programming. we first introduce concepts. and so on. such as linear programming. combinatorial optimization. 1991). drilling. Optimization techniques employed in these applications cover almost all subfields in mathematical programming. Then we review their applications in areas that are pertinent to this study. planning.Chapter 2 Literature Review Applications of optimization techniques in the upstream oil industry began in the early 1950s and have been flourishing since then. Mathematical programming encompasses subfields such as linear programming. and applications of some major subfields in mathematical programming.. Optimization problems in the most general form can be represented as min { f (x ) : li ≤ ci (x ) ≤ ui .
the simplex algorithm. In 1947.1 becomes an integer programming (IP) problem. Karmarkar (1984) introduced the first interior point algorithm for the LP problem. Various kinds of optimization problems can be formulated as LP problems. 2. However. logical requirements. the problem described by Eq.1. and scheduling problems. 2. such as transportation. the simplex method can take an very large number of iterations (Klee and Minty. Dantzig proposed the first solution method for the LP problem.1. and li and ui are the lower and upper bounds for the ith constraint. The interior point algorithm approaches the optimal solution from the interior of the feasible region. 1997). LITERATURE REVIEW 9 where the objective function f and the constraint functions {ci } are functions of control variable x .CHAPTER 2. Thus the solution usually is not a vertex of the feasible region. the simplex algorithm is very efficient. When some but not all components of x are discrete. Nowadays. . Discrete variables are useful to model indivisibility. the problem described by Eq. An optimization problem can be categorized according to the type of its control variables.1 Linear Programming When the objective function f and constraint functions {ci } are linear functions of control variable x . production planning. and on/off decisions. Integer programming finds its application in many real life problems. In general. thus its optimal solution is always a vertex (or an extreme point) of the feasible region.1 is a linear programming (LP) problem. 2. The simplex algorithm (Dantzig. 1972). the problem is a mixed integer programming (MIP) problem. for some problems. LP problems with thousands or even millions of variables and constraints can be solved efficiently by both the simplex and interior point algorithm (Bertsimas and Tsitsiklis. 1963) finds the optimal solution by moving along the vertices of the feasible region. respectively. The interior point method is very efficient both theoretically and in practice. 2.2 Integer Programming When all components of the unknown x are discrete variables. resource allocation. and objective and constraint functions.
airline crew scheduling. In practice. which was first presented by Land and Doig (1960).3 Nonlinear Programming Eq. For the production optimization problem addressed in this study. 1983). genetic algorithms (Goldberg. or it only gives an approximate solution (Bertsimas and Tsitsiklis.1. and many special algorithms designed for particular problems. The fundamental idea is to add constraints to a series of linear programming relaxations of the LIP problem until the optimal solution of a relaxation problem takes integer values (a linear programming relaxation is formed by allowing the integer variables in an LIP problem to take real values). 2. Both methods tackle the LIP problem by solving a series of linear programming problems. Other methods for integer programming include dynamic programming (Bellman. 1989).1 becomes a nonlinear programming problem when its objective and/or constraint functions are nonlinear. So far. the well connections have to be modeled as discrete variables. the method uses bounds on the optimal objective value of a subproblem to avoid forming and solving other subproblems. 1997). It is impossible to give a thorough survey of all major optimization algorithms in a limited space. The cutting plane method was proposed by Gomory (1958).CHAPTER 2. telecommunications.. The Branch and Bound method uses a “divide and conquer” approach to explore the set of feasible integer solutions. Algorithms have been developed for individual classes of problems. Instead of enumerating all the subproblems. LITERATURE REVIEW 10 such as capital budgeting. simulated annealing (Kirkpatrick et al. the common solution techniques are the cutting plane method and the branch and bound method. there does not exist a universal algorithm for integer programming: either the method takes a large amount of time. Some general . 2. The method divides an optimization problem recursively into multiple subproblems. For linear integer programming (LIP) problems. Nonlinear programming problems come in many different forms and shapes. the cutting plane method is not effective. Rather. this section gives a brief description of some of the optimization methods encountered most frequently in the petroleum engineering literature. The most effective technique for solving the LIP problem is the Branch and Bound method. and production planning. 1957).
Line search methods approach a local minimum using the following iteration scheme: x k +1 = x k + α k p k (2.CHAPTER 2. 1981). p k is a search direction along which the function decreases. is positivedefinite. LITERATURE REVIEW 11 references about nonlinear programming are Gill et al. An important class of methods for unconstrained optimization problems is the socalled line search. Fletcher (1987). At each iteration. the model function qk (p ) has a unique minimum which can be obtained by solving the following linear system ∇ 2 f (x k )p k = −∇f (x k ) (2. and α k is a step length that ensures “sufficient” progress toward the solution. sometimes it is impossible or timeconsuming to compute the Hessian matrix. The key to this exceptional convergence rate is that the method uses the curvature information of the objective function (the Hessian matrix) to compute the search direction.2) where x k and x k +1 are the current and next iterates. Newton’s method converges quadratically near a local optimum (Gill et al. The step length α k can be obtained by a linear search procedure. However. QuasiNewton methods are attractive for . the Hessian approximation is updated by using the gradient information to mimic the behavior of the true Hessian matrix. In subsequent iterations.4) The solution p k of Eq. 2.4 serves as the search direction of current iteration. (1981). and Bertsekas (1982). the method approximates the objective function f (x ) around current iterate x k using a quadratic function defined as follows: qk ( p) = f (x k ) + ∇f (x k ) p + T 1 T 2 p ∇ f (x k )p 2 (2. The initial Hessian approximation can be an identity matrix. Unconstrained optimization methods.. ∇ 2 f (x k ) . A Newtonbased line search method (usually termed as Newton’s method) goes as follows.3) When the Hessian matrix. QuasiNewton methods are based on the idea of gradually building up an approximate Hessian matrix using the gradient information collected from previous iterations.
A barrier method starts with a feasible point and creates a sequence of unconstrained problems whose successive minimum stays feasible . Constrained optimization methods. The SQP methods have a structure of major and minor iterations. SLC methods also involve major and minor iterations. These conditions were first established by Kuhn and Tucker (1951) and they are often called the KuhnTucker conditions or optimality conditions. 1982) are another class of methods widely used for NCP problems. Each major iteration involves formulating a quadratic programming (QP) subproblem to obtain the search direction p k and using a line search procedure to obtain the steplength α k . In contrast to SQP methods. Similar to the SQP methods. the SLC methods formulate a linearly constrained subproblem whose objective function is a general approximation to the Lagrangian function of the original problem. and barrier functions (Gill et al. 1997) as the most effective methods for nonlinearly constrained programming (NCP) problem. which formulate a QP subproblem in each major iteration. and methods based on augmented Lagrangians. 1981). The optimum of a constrained optimization problem is characterized by a certain set of conditions. The sequential linearly constrained (SLC) methods (Murtagh and Saunders. The major constrained optimization methods include sequential quadratic and linear programming methods. 2000).. penalty. The QP subproblem is formulated in such a way that its objective function is a quadratic approximation to the Lagrangian function of the original problem and its constraint functions are the linearization of the original nonlinear constraints around the current iterate. LITERATURE REVIEW 12 practical use because their memory and computational requirements are low while they have good convergence rate.CHAPTER 2. an SLC method tends to require fewer major iterations but more evaluations of the problem functions than an SQP method (Murray. The sequential quadratic programming (SQP) methods are widely regarded (Murray. Constrained optimization methods are developed to find a point that satisfies these conditions. reducedgradient methods. In general. The barrier methods solve a nonlinearly constrained problem by solving a sequence of unconstrained optimization problems. The minor iterations are used to solve a particular QP subproblem.
GAs employ the idea of natural selection and genetics in the process of . which can be viewed as the vertices of a polytope. some severe cases of failure have been observed (Powell. While the polytope method is the most used method for unconstrained optimization in practice. when problem functions are not smooth or the derivatives are too expensive to compute. one may choose a direct optimization method. LITERATURE REVIEW 13 and converges to a minimum of the constrained problem. the objective function of the unconstrained problem is constructed by adding a barrier function to the original objective function of the constrained problem. To achieve this. A thorough review of the direct optimization methods is given by Powell (1998). 1981). (1962). Recently the method has been developed further to address optimization problems whose objective and constraint functions are convex functions (Boyd and Vandenberghe. 2001). The data point is evaluated based on its objective function value. a general rule is to choose a method utilizing as much derivative information as possible (Gill et al. a new polytope is generated by replacing the worst data point with a newly generated data point. Genetic algorithms (GAs) are heuristic optimization algorithms introduced by Holland (1975). 1998). At each iteration. The polytope method was first established by Spendley et al. When selecting an optimization method. The barrier function is a modification of the constraint functions and becomes infinitely large when the iterate approaches the boundary of the feasible region.CHAPTER 2.. While many direct optimization methods have been proposed. 1968). for a problem with dimension n . However. Direct optimization methods refer to optimization methods that do not require derivatives. here we only review a few that are popular in petroleum engineering applications. The method involves manipulation of n + 1 linearly independent data points. Separable programming problems are usually solved by linear programming techniques (Hillier and Lieberman. The barrier method was first established in the 1960s for general nonlinearly constrained optimization problems (Fiacco and McCormick. Separable programming is a special class of nonlinearly constrained optimization problems whose objective and constraint functions are sums of functions of one variable. 2001).
1996. 1998).. Rasheed. the slope of the gaslift performance curves should be equal for all wells (Hong. Although nonlinear constraints can be handled through penalty functions (Powell and Sholnick. In addition to the equalslope method. 2. we review applications of optimization techniques in rate allocation.2. In GAs. the lift gas is usually allocated using some optimization algorithm. crossover. 1993.2 Applications of Optimization Techniques to Petroleum Fields Optimization techniques have been applied to virtually all aspects of the oil industry. GAs have several advantages over other optimization algorithms. In this section. GAs have certain disadvantages. LITERATURE REVIEW 14 searching for the global optimum of a problem. (1989) . The mechanism of gaslift is elaborate. To determine the optimal lift gas rate. Secondly. Kanu et al.CHAPTER 2. First of all. 2. and mutation. GAs are very general optimization methods in that they handle both the discrete and continuous variables naturally and require no domain knowledge of the optimization problem. the optimal lift gas rate is the one corresponding to the maximum oil rate on the performance curve. the performance of such treatments has not been impressive. possible solutions are encoded as chromosomes and modified by means of selection. while excessive lift gas injection will reduce the oil rate. 1989). Michalewicz. production system design and operations. GAs have been used primarily for unconstrained problems. and reservoir development and management. Nevertheless. Secondly. GAs are stochastic search methods capable of avoiding local optima. A gaslift performance curve is a plot of oil rate versus lift gas rate for a gaslift well. Firstly. 1981). convergence can be slow.. The earliest lift gas allocation method is a simple heuristic method based on the concept of equalslope. When the gas supply is limited. When the gas supply is unlimited. especially at the later stages of optimization. which states that at the optimal solution. Nishikiori et al. 1975. the usual practice is to allocate the lift gas to a well according to a gaslift performance curve (Nishikiori et al.1 Lift Gas and Production Rate Allocation Lift gas allocation. An appropriate amount of lift gas increases the oil rate.
and sale contracts. (1996) addressed this problem by proposing a stochastic algorithm that uses a heuristic method to calculate the descent direction. However. DuttaRoy and Kattapuram (1997) applied a SQP method to a linearly constrained gaslift optimization problem with 13 wells. Buitrago et al. however. Both the equalslope method and the QuasiNewton method rely on derivative information to verify optimality.CHAPTER 2. and therefore tend to get trapped in local optima. The LP method is very efficient for the rate allocation problem. Fang and Lo (1996) applied LP to allocate both lift gas and production rates. compressor capacity limits. However. and pointed out that when flow interactions are significant. Gaslift optimization based on performance curves is simple and easy to implement. reservoir injection requirements. (1961) applied linear programming to maximize daily income from a multireservoir case subject to well producing capacities. LITERATURE REVIEW 15 also applied a formal optimization algorithm. a QuasiNewton method. General rate allocation. suffers from the fact that the objective and constraint functions in an LP problem have to be linear. Due to this limitation. this approach ignores flow interactions among wells in the optimization process. They first approximate gaslift performance curves by a piecewise linear curve. This limitation can be serious when some gaslift performance curves are not concave. gaslift requirements. to the lift gas allocation problem. and then formulate the rate allocation problem as an LP problem. Results showed that the SQP method does perform better than methods based on performance curves. . as suggested by their results. This problem refers to allocating production rates and lift gas rate of single wells or the total production rates of reservoirs (in multireservoir cases) to achieve certain operational goals. Lo and Holden (1992) proposed a linear programming model to maximize daily oil rate by allocating well rates subject to multiple flow rate constraints. the gaslift performance curves in Fang and Lo’s method have to be concave. DuttaRoy and Kattapuram (1997) analyzed a gaslift optimization problem with two wells sharing a common flow line. Linear programming seems to be the most popular optimization algorithms for this kind of problems. their method is not good at handling constraints. nonlinear optimization tools are needed to obtain satisfactory results. Attra et al.
(1990) developed the Western Production Optimization Model (WPOM) for the Prudhoe Bay field. water. Litvak et al. gas constraints are considered at the lowest predictive well management (PWM) level. for example. First. using such approaches. 2001). The oil production in Kuparuk River oil field in Alaska is also constrained by the gas processing capacities of surface facilities.CHAPTER 2. and oil rate constraints at the lowest PWM level are handled sequentially and the procedure is repeated for the next PWM level. rate constraints on facilities are handled sequentially by ad hoc rules. For example. In some cases. Then. A number of such procedures are available in commercial reservoir simulators. Rate allocation problems are encountered frequently in mature fields where production facilities cannot meet the field demand. This model allocates the oil rate and gas rate to surface facilities and wells in a topbottom manner based on the “incremental GOR” concept that produces the next incremental barrel of oil with the lowest GOR. gaslift optimization is usually considered separately from well rate optimization using. Gaslift rates were allocated based on gaslift tables of gasliquid ratio. For example. first. In some commercial reservoir simulators (GeoQuest. (1992. Barnes et al. these procedures may not yield the optimal solution. etc. it is difficult to determine how much of the gas handling capacity should be used for gaslift and what portion should be allocated for other purposes. liquid. 2000. (1997) built an integrated reservoir and gathering system model of the Prudhoe Bay field. Landmark. and employed some heuristic methods to optimize well connections to manifolds. 1994) applied a neural network model to determine the optimal allocation of lift gas to wells subject to multiple . the following approach. Also. LITERATURE REVIEW 16 Options in commercial simulators. Field cases. The optimization algorithms in these applications are heuristic and they may lead to suboptimal solutions. Then. Ad hoc rules based on the gasoil ratio of production wells are applied to scale well rates and meet the gas constraint. liquid well rate and water cut. Stoisits et al. To utilize existing facilities fully. these liftgas rates are scaled using some rules to match constraints on the total amount of available gas. oil production in the Prudhoe Bay oil field in Alaska is constrained by the gas handling capacity of the separation units and the central gas plant. liftgas rates are determined for specified wells based on their gaslift performance curves or other information. Furthermore.
CHAPTER 2. LITERATURE REVIEW
17
gas constraints. The neural network was trained by results generated from single well simulation. Stoisits et al. (1999) addressed the production optimization problem at Kuparuk River field using a genetic algorithm. In their study, the decision variables were the status of the production wells (close or open) and incremental GOR for drill sites. A neural network model was developed to simulate the pressure drop through the surface pipeline network. In all these applications (Stoisits et al., 1992, 1994, 1999) the neural network is the central tool for speeding up the simulation/optimization process. However, it is timeconsuming to train the neural network. Also, the accuracy of the neural network, which depends on the accuracy of the training tools, the behavior of the simulated system, and many other factors, is hard to control. In summary, the general rate allocation problem has not been fully resolved yet, especially when dealing with gigantic fields like the Prudhoe Bay and fields where flow interactions cannot be ignored.
2.2.2 Optimization of Production System Design and Operations
Traditionally, optimization of production system design and operation in a petroleum field has been performed by nodal analysis combined with trial and error (Beggs, 1991). For example, by holding all other parameters fixed, a single variable is varied to see which value of this variable gives the optimal objective function value. For multiple decision variables, such a procedure fails because of the large number of function evaluations required to cover the search space. Carroll and Horne (1992) identified the need for formal optimization algorithms in such applications. Carroll and Horne applied a Newtontype optimization algorithm and a polytope method to optimize the tubing diameter and separator pressure of a singlewell system and successfully demonstrated the usefulness of multivariate optimization techniques. Ravindran (1992) followed the work of Carroll and Horne and allowed the decision variables to vary with time. Fujii and Horne (1995) applied a genetic algorithm to optimize the design and operations of a multiwell production system. Fujii and Horne noted that multiple solutions may exist for the multiphase network flow problem. However, no procedures were developed to select the legitimate solution. Palke and
CHAPTER 2. LITERATURE REVIEW
18
Horne (1997a) followed a similar line and applied a Newtontype algorithm, a polytope method, and a genetic algorithm to a compositional model of a singlewell system. The control variables included tubing diameter, separator pressure, and volume of gas injected. A common observation of all these studies was as follows. When the tubing diameter is involved, the surface of the objective function is highly irregular and gradientbased algorithms often fail to solve the optimization problem. Palke and Horne (1997a) concluded that genetic algorithms are the most robust algorithms for such cases. NETOPT (SIMSCI, 1999) is a commercial network optimizer that combines a general multiphase network simulator with a SQP method. NETOPT is a very general tool, and can be interfaced with Eclipse (GeoQuest, 2000), a commercial reservoir simulator, and be used to optimize various production design and operational problems. Barua et al. (1997) discussed the general usage of NETOPT in the optimization of production operations. DuttaRoy et al. (1997) applied NETOPT to analyze compressor installation cost and operating strategy required to meet the lifetime production goals of a gasfield. Heiba et al. (1997) applied NETOPT to devise optimal strategies for a cyclic steam stimulation project. Because NETOPT employs SQP, a derivativebased optimization algorithm, it locates only local optima and accepts only continuous decision variables. The performance on individual optimization problems is not guaranteed and care has to be taken to ensure reasonable results. Zhang and Zhu (1996) minimized the total cost of a gas distribution network by adjusting the diameters of pipelines of a fixed network layout. In their study, Zhang and Zhu treated the optimization problem as a bilevel programming problem, and then simplified the lower level and upper level problem by exploring their special properties.
2.2.3 Optimization of Reservoir Development and Planning
Reservoir development studies, such as drilling scheduling, well placement, and production rate scheduling, have been an active area for optimization. The optimization tools and simulation models employed in reservoir development studies have been evolving.
CHAPTER 2. LITERATURE REVIEW
19
Early optimization studies of reservoir development featured simple reservoir models and linear programming techniques. Aronofsky and Lee (1958) built a linear programming model to maximize profit by scheduling production from multiple homogeneous reservoirs. Bohannon (1970) presented a linear programming model to find the optimum 15year development plan for a multireservoir pipeline system. The main variables considered in their formulation were annual production rate from each reservoir, number of wells to be drilled each year, and timing of major capital investments such as initiating secondary recovery projects and expanding pipeline facilities. Xiao et al. (1998) proposed a multiobjective linear programming model for an oilfield injection recovery system. The main goal was to produce most oil with least investment under the limitations of the available natural resource, equipment, and manpower, etc. Aronofsky (1983) reviewed the application of linear programming in oil and gas development. Linear programming requires the objective and constraint functions to be linear. In order to optimize the production system by linear programming, the nonlinear production system has to be approximated by a system of linear functions. This greatly limits the application of sophisticated reservoir simulators in the optimization of reservoir developments. As the computing power increases and simulation techniques advance, researchers are attempting to optimize reservoir development with more and more complex reservoir models and sophisticated optimization techniques. Various gradientbased optimization methods have been investigated. McFarland et al. (1984) applied a generalized reduced gradient nonlinear programming method to maximize present value of profits from a reservoir by deciding how many wells to drill in each time period, the production rates, abandonment time, and platform size. They used tanktype reservoir models to describe reservoir dynamics. Lasdon et al. (1986) studied the problem of determining the optimum flows from a set of existing wells in each of several time periods subject to certain demand schedules and operating constraints, and combined numerical optimization techniques with a single phase, twodimensional reservoir simulator. Zakirov et al. (1996) presented a method for optimizing net present value of oil production with respect to well production rates as functions of time over the whole life of the field. What distinguishes Zakirov et al.'s method from previous literature
Frequently. Therefore.CHAPTER 2. Therefore. stochastic optimization algorithms such as genetic algorithms have gained popularity in such applications. stochastic . LITERATURE REVIEW 20 is that they used optimal control theory to obtain the gradients of the objective function in an efficient way. 2. the number of laterals and the trajectory of the laterals. In addition. a major concern is that we never possess true and complete information about the reservoir. The decision variables include the location of the well. To speed up the search process. genetic algorithms are slow. Güyagüler (2002) used the utility theory to quantify the uncertainty in reservoir developments.3 Concluding Remarks Optimization techniques have been applied to many aspects of the oil industry. Bittencourt and Horne (1997) hybridized a genetic algorithm with a polytope method. Results demonstrated that the appropriate combinations of these tools can greatly speed up the optimization process. Due to the restrictions in computing power and simulation technologies. Yeten et al. With the increase of computing power and advances in simulation techniques. (2000) and Güyagüler and Horne (2001) also looked at well placement problems by adapting the hybridized genetic algorithm from Bittencourt and Horne (1997). In general. early applications favored simple methods such as linear programming and heuristic methods. (2002) employed a genetic algorithm as the optimization driver and developed a hill climber and a neural network model to speed up the optimization process. In the optimization of reservoir development. Most reservoir development problems have irregular surfaces for the objective function. they used a kriging method as a proxy for the function evaluations to reduce the time for function evaluations. Yeten et al. researchers and practitioners are attempting to optimize reservoir development and operations with more and more complex models and sophisticated optimization techniques. Palke and Horne (1997b) used a genetic algorithm to estimate the cost of uncertainty in reservoir data. Güyagüler et al. and applied the hybridized algorithm to a reservoir development problem in which the decision variables are well locations and the objective function is based on an economic model. (2002) investigated the problem of placing a multilateral well. the optimization problems in the oil industry do not have smooth objective functions.
lift gas rates. 1981.CHAPTER 2. Multiobjective optimization was introduced to handle conflicting operational goals in a petroleum field (Chapter 7). LITERATURE REVIEW 21 optimization algorithms such as genetic algorithms have been the major optimization drivers in recent applications. This research extended the method proposed by Fang and Lo (1996) for the rate allocation problem to cases where the gaslift performance curves can have any shape and the well water cut and GOR can vary with oil rate (Chapter 4). Buitrago et al. it is worthwhile to investigate ways of speeding up the simulation processes and combining uncertainty with optimization. not the production rates and lift gas rate for individual wells. Some examples are presented throughout the dissertation to demonstrate the advantages of the methods. Because reservoir simulation is a computationally intensive process and the true reservoir information is unknown. The method proposed by Fang and Lo (1996) ignores flow interactions among wells and implicitly assumes the gaslift performance curves are concave and the well water cut and the gas oil ratio (GOR) are constant with varying oil rates. 1996. A twolevel programming approach was developed to optimize all control variables simultaneously (Chapter 6).. .. Further.. and well connections simultaneously subject to nonlinear constraints. Nishikiori et al. Finally many of the optimization approaches developed in this study were integrated with a reservoir simulator so that the impact of the reservoir conditions could be taken into account in the optimization process (Chapter 8). only few have been reported for optimizing both the production rate and gaslift rate (Fang and Lo. DuttaRoy and Kattapuram. While extensive studies have been reported for unconstrained or linearly constrained gaslift optimization problems (Kanu et al. 1996.. no previous publication has addressed the problem of optimizing production rates. 1989. Therefore. Stoisits et al. 1999). The type of production optimization problem addressed in this study has not been fully resolved. (1999) determines the optimum well shut down list and a drill site IGOR (incremental gas oil ratio) list. A SQP method was also applied to the rate allocation problem to account for flow interactions (Chapter 5). 1997). The method proposed by Stoisits et al. it is necessary to explore new approaches to solve this important class of production optimization problems.
distribution and injection network belong to network flow problems that have special properties. a model is required to simulate the multiphase flow in the gathering system. 1969). disposal. The injection network is used to inject fluids into the reservoir for enhanced oil recovery projects or for fluid disposal/storage.1 Introduction A typical oil field contains a gathering system. There are two major solution approaches. 1973. Both approaches formulate the network problem according to the requirements of mass balance and Kirchoff’s law. or further processing. a fluid distribution network. and an injection network. For the production optimization problem considered in this work. injection. The distribution and injection networks are ignored. sale. The second approach is to formulate the 22 . The multiphase flow problems in the gathering. Liu. The network problem with singlephase flow is relatively easy to solve and has been studied extensively in both the civil engineering and the petroleum engineering literature. 1970. The separated fluids are then distributed to different destinations for storage. Epp and Fowler. The first approach is to formulate the network problem as a set of system equations and solve the system equations by the NewtonRapshon method or one of its variations (Donachie.Chapter 3 Simulation of Gathering Systems 3. The gathering system collects the fluids from production wells and delivers them into separation units.
and PIPEPHASE (SIMSCI. Byer (2000) proposed an alternative formulation with mass balance requirements as the system equations for gathering systems with treelike structures. 2000). mathematically. 2001) have options for the simulation of multiphase network problems. only a few researchers have reported solution methods for multiphase network problems. gas. In this study. However. How to find the true solutions systematically is an unresolved problem. Fujii (1993) formulated the multiphase flow problem in a treelike structure as an optimization problem and solved the problem using a GaussNewton method. the NewtonRaphson method may not converge or converge slowly due to the complex nature of the multiphase flow. The multiphase network problem is also formulated according to the requirement of mass balance and Kirchoff’s law. Secondly. however. This chapter first describes simulation models for the individual components. VIP (Landmark. because we need to determine the fluid compositions in a link before the pressure loss along the link can be evaluated. Results show that even for a simple system with two wells. Several commercial software packages such as Eclipse (GeoQuest. chokes. The gathering system can include such components as wells. the multiphase network problem is much more difficult than the single phase network problem for the following reasons: Firstly. the system equations are more difficult to construct. Collins et al. Byer (2000) showed that Schiozer’s method can result in convergence problems due to incorrect Jacobian terms. the solution method can take more than ten Newton iterations to converge unless a good initial guess is available. Schiozer (1994) reported a solution method with pressure continuity requirements for the system equations for networks with treelike structures. 1976. Fluid properties are represented by the blackoil model. However. 1978). The fluids in the system can include oil. tubing strings. 2001).CHAPTER 3. Finally. The study investigated efficient procedures to solve the network . and pipes. SIMULATION OF GATHERING SYSTEMS 23 network problem as an optimization problem and solve it using appropriate optimization algorithms (Hall. multiple solutions exist for the multiphase network problem and not all of them are physical. and water phases. This is not easy in complex systems where the flow directions are unknown and phase splitting ratios at junctions are hard to determine. we consider gathering systems with treelike structures. their solution algorithms are not in the public domain. To date.
In the homogeneous model. 3.2) . the reservoir properties in the vicinity of the well. For the pseudophase. and the flow behavior in the wellbore. p p p − pw ) (3.2 Well Model In reservoir simulation. and p w denotes the well pressure. p p denotes the reservoir pressure of phase p . mixture density ρ m is based on in situ phase fractions: ρ m = ρ l El + ρ g E g where El and E g are the liquid and gas phase in situ fractions. different phases are lumped into one single pseudophase so that pressure drop equations for singlephase flow can be utilized. and (3. This model assumes steadystate radial flow and can be expressed as q cw = ∑ WI λ ( p c. SIMULATION OF GATHERING SYSTEMS 24 problem and to compute derivatives that are required in history matching. An example is presented to demonstrate the performance of the solution procedures that were found most suitable. Expressions for WI and λc . p represents the well mobility for component c in phase p . λc. and sensitivity analysis.3 Pipe Model In this work a homogenous model with slip was used to model multiphase pipe flow.CHAPTER 3. WI denotes the well index. p can be found in Aziz and Durlofsky (2002). This study considered a vertical well completed in a single grid block and used a basic well model proposed by Peaceman (1978). 3. optimization. respectively. The well model accounts for the geometric characteristics of the well. a well model is used to relate the flow rate of a well with its bottomhole pressure and the reservoir pressure in the vicinity of the well (such as the grid blocks in which the well is completed).1) where qcw denotes the flow rate of component c .
3. it can be shown that (Brill and Beggs.5) dp is the pressure gradient due to acceleration.6) τ S dp =− w A dx f (3. dx f dp is the pressure gradient due to gravity. 3. With these properties defined. SIMULATION OF GATHERING SYSTEMS El + E g = 1 25 (3. Consider gas/liquid twophase flow in a pipe without perforations. we can proceed to compute the pressure loss for multiphase flow in pipes. Ouyang (1998) suggested that the mixture velocity should be defined as Um = ρ ρl U sl + g U sg ρm ρm (3.4) where U sl and U sg are the superficial velocities of liquid and gas phases. the total pressure gradients along the pipe can be decomposed into three parts dp dp dp dp = + + dx dx f dx g dx a where • • • dp is the pressure gradient due to wall friction. 1998) dp = − gρ m sin θ dx g (3.7) . dx a Based on mass balance and momentum balance. and dx g (3.3. respectively. the pressure loss (or gain) in the pipe can be computed by integrating the pressure gradient along the pipe.CHAPTER 3.2.3) Computation of gas volume fraction E g is presented in Section 3.1 Pressure Gradient Equation Given a flow rate and the pressure at one end of a pipe.
τ w is the wall friction shear stress. dx a is usually small (Brill and Beggs.3. Chen. The DFM correlates the in situ gas velocity U g with the superficial mixture velocity U sm ( U sm = U sl + U sg ) using two parameters C0 and U d . Duns and Ros. The DFM is used because of its continuity. the DFM covers the complete range of flowing conditions without any first order discontinuities.CHAPTER 3. the gas phase has a tendency to flow faster than the liquid phase.10) The acceleration pressure loss across a pipe segment depends on the difference dp between the velocity heads of the flowing mixture at the two ends of the segment. 2001) was used in this work to calculate the liquid holdup. SIMULATION OF GATHERING SYSTEMS 26 where θ is the pipe inclination angle. which can be computed from phase viscosities by (Brill and Beggs. 2000.. 1939) or any of its approximation forms based on the relative pipe roughness ε d and the mixture Reynolds number Rem = ρ mU m d µm (3.9) where d is the pipe diameter and µ m is the mixture viscosity. 1994. S is the pipe perimeter. 1973). In general. 1963.8) where the Fanning friction factor f m can be evaluated by the Colebrook equation (Colebrook. 1998) and it is ignored in this work. 1998) µ m = µ l El + µ g E g (3.2 Liquid Holdup Evaluation A drift flux model (DFM) (GeoQuest. Xiao et al. Taitel and Dukler. 3. 1990) or correlations (Beggs and Brill. in all but downward flow. 1976. This property makes the DFM more suitable for simulation and optimization studies than mechanistic models. The wall friction shear stress is computed as τw = 1 2 f m ρ mU m 2 (3. and A is the pipe area. Unlike mechanistic models (Aziz and Petalas.
. Table 3. at low flow rates.14) Substituting Eqs. This phenomenon is demonstrated in Figure 3. 3.1 presents the problem data used to generate Figure 3.3. SIMULATION OF GATHERING SYSTEMS U g = C0U sm + U d 27 (3. and 3. and inverse quadratic interpolation. Beyond a certain critical flow rate. thus the liquid holdup El . which accounts for the local velocity differences between the two phases.11) where C0 is a profile parameter or distribution parameter that accounts for the effect of velocity and gas concentration profiles. proved to be both efficient and robust for solving Eq. 3.CHAPTER 3. 1992). The superficial gas velocity and in situ gas velocity are related by: U sg = E gU g (3. 3. we found that the NewtonRaphson method (Press et al. Both C0 and U d are functions of the gas volumetric fraction E g C0 = C0 (E g ) U d = U d (E g ) (3. 3.13 into Eq. The Van WijingaardenDekkerBrent method (or Brent’s method) (Press et al.12 and 3..13) The specific expression of Eqs. bisection. a method that combines root bracketing.1. 3.13 can be found in GeoQuest (2000).12) (3.15) Solving Eq.15. 3. Finally we mention that in certain cases. increasing the flow rate increases the pressure drop across the pipe. 3. 1992) fails to converge for Eq. 3.1.15.15 we can obtain E g .14 we obtain a nonlinear equation with E g as the unknown: f (E g ) = U sg − E g (C0U sm + U d ) = 0 (3. increasing the flow rate decreases the pressure drop across a pipe.12..3 Upwards TwoPhase Flow Behavior Under certain conditions. for upwards twophase flow.11. U d is the drift velocity of the gas.
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS
28
The phenomenon described can be explained as follows. As shown in Eq. 3.5, the pressure drop across a pipe segment is dominated by the gravity force and friction force. The lower the flow rate, the smaller the frictional force. However, at low flow rates, liquid phase tends to holdup (be held back) in the pipe causing bigger pressure drop due to larger mixture density. Hence, at low flow rates, the gravity force dominates. As the flow rate increases, the gravity force decreases due to reduced liquid holdup, the friction forces increases but not large enough to compensate fully for the gravity force reduction. Therefore, the overall pressure drop decreases as the flow rate increases. Since the increase of frictional force is proportional to the square of the mixture velocity, beyond a critical point, the friction force starts to dominant and the overall pressure drop increases as the flow rate increases. Consequently, the typical pressure drop of multiphase flow across a pipe demonstrates the behavior shown in Figure 3.1. A consequence of this flow behavior is that multiple solutions may exist for the multiphase network flow problem. This is discussed further in Section 3.5.3. Table 3.1: Slip model problem data.
Tubing Parameter 3.5 in. Diameter Roughness Vertical Depth 0.0001 in. 8400 ft Fluid Property Oil Density at 30° API Standard Condition 0.8 Specific Gas Gravity 62.37 Water Density at lbm/cft Standard Condition 1000 Gas Oil Ratio SCF/STB 0.6 Water Oil Ratio Flow Conditions 1000 psia Downstream Pressure Downstream 110 °F Temperature Upstream 160 °F Temperature
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS
29
3950 3900 3850 3800 3750 3700 3650 3600 0 1000 2000 3000 4000 5000 OIL RATE, STB/DAY
Figure 3.1: Upstream pressure in a vertical well tubing string with a fixed downstream (wellhead) pressure for different flow rates.
3.4 Choke Model
A good choke model is essential for accurately modeling multiphase flow in the gathering system. The multiphase choke model developed by Sachdeva et al. (1986) has a smooth transition between critical and subcritical flow and was used in this work. The assumptions made when building this model are: • • • • • • • incompressible liquid phase, equal phase velocities at the choke throat, no mass transfer between phases, onedimensional flow, accelerational pressure drop term dominates and the friction term is ignored, no distinction between free gas and solution gas, and liquid properties are obtained from the weighted average properties of oil and water. The Sachdeva model can be expressed in the following context: given an upstream pressure p1 and a downstream pressure p2 , determine the liquid flow rate across the choke.
UPSTREAM PRESSURE, PSIA
CHAPTER 3. SIMULATION OF GATHERING SYSTEMS
30
The first step of the Sachdeva model is to determine the boundary of critical and subcritical flow. Given an upstream pressure p1 , the critical ratio of downstream to upstream pressure is determined by the following equation:
k −1 (1 − x1 )ρ g1 (1 − y c ) k + k −1 x1 ρ l yc = 2 n n (1 − x1 )ρ g 2 n n (1 − x1 )ρ g 2 k + + + k −1 2 2 x1 ρ l x1 ρ l
k = C p / Cv
k
(3.16)
(3.17) (3.18)
n =1+
x1C v + (1 − x1 )C p
x1 (C p − C v )
where y c is the critical ratio of downstream pressure to upstream pressure, k is the ratio of specific heat, C p is the specific heat capacity at constant pressure, C v is the specific heat capacity at constant volume, x is the mass fraction of gas, n is the polytropic exponent of gas, and the subscripts 1 and 2 represent upstream and downstream conditions, respectively. Eq. 3.16 is a nonlinear equation with y c as the unknown. Brent’s method (Press et al., 1992) is suitable for solving Eq. 3.16. In the field units, the liquid rate across the choke is given as
k −1 x1k 1 − y k 0.525C d d 2 2 (1 − x1 )(1 − y ) ql = p1 ρ m 2 Cm 2 ρ l1 ρ g1 (k − 1) 0.5
(3.19)
and
p y = max 2 , y c p 1
(3.20)
−1
ρm2
x1 1 − x1 = + 1/ k ρ l1 ρ g1 y
(3.21)
C m 2 = 8.84 × 10 −7 γ g (GLR − f o Rs ) + 6.5 × 10 −5 ( f o ρ o Bo + f w ρ w Bw ) (3.22)
This problem has a unique solution and can be obtained using appropriate root finding procedures. The top of the production tree has a fixed pressure and usually represents an inlet to a separator. and nodes. A node represents a flow junction or the terminal point of a link. and GLR is the gas liquid ratio.20 states that if the actual value of y actual = p2 / p1 is greater than y c . or a pipeline.CHAPTER 3. a choke. In this section. When the flow across a choke is critical. 1992) and Brent’s method (Press et al. . the flow is subcritical.22. 3. if y actual is less than y c . when simulating the network flow. the actual p2 is used as the downstream pressure in Eqs.. 1992) work well but the NewtonRaphson often fails to converge. f o and f w are the flowing fractions of oil and water. links. 3. Computational experience showed that bisection (Press et al. Rs is the solution gasoil ratio.21 and 3. A link can be a tubing string.5 Network Simulation We considered a gathering system with a forestlike structure with multiple production trees. γ is the specific gravity. Fluid properties used to calculate C m 2 are evaluated at the downstream conditions. the value of y c p1 is used as the downstream pressure in Eqs.21 and 3. 3.. SIMULATION OF GATHERING SYSTEMS 31 where d is the choke inside diameter in inches. Eq. A well refers to a wellbore and its surrounding reservoir conditions (grid blocks in which the well is completed). Finally we discuss possible outcomes of the solution procedure. we first formulate the governing equations of the network flow problem. the downstream pressure can not be determined uniquely. A production tree includes wells. Therefore. Links refer to any device or facility across which pressure changes. Then we describe a solution procedure. compute the upstream pressure of the choke. B is the formation volume factor. 3. the flow is critical.22. we formulated the system equations in such a way that we only need to face the following problem: given a downstream pressure and a flow rate. Cd is the discharge coefficient.
1 Governing Equations The multiphase flow in a gathering system is described by mass conservation and Kirchoff’s law. chokes. then they are connected to a separator through a common pipeline. The separator is operating at a fixed pressure. two gaslifted wells are connected to a common manifold. SIMULATION OF GATHERING SYSTEMS 32 Node 1 (separator) Pipe Node 2 (manifold) Node 3 Choke Node 5 Tubing Node 4 Node 6 Node 7 Node 8 Well 1 Well 2 Figure 3. The flow rate of each well is controlled by a choke between the tubing string and the manifold.2: A simple gathering system. The tubing strings.5. In the example problem. Kirchoff’s law requires that the pressure for a node should be the same no matter from which path it is computed. The junctions of two or more links are modeled as nodes.CHAPTER 3. Example problem. and pipelines are modeled as links. .2 illustrates the gathering system of this problem. Figure 3. 3. The mass conservation requires that every node should have zero net flow rate unless it is a boundary node. We will use an example problem to facilitate our discussion.
j . For instance. Because the gathering system has a treelike structure and the fluid flows from the bottom up through the production tree. respectively). for a particular well.i = p j∈Ωiw ∑q w p. We choose a solution node for every well and perform the described pressure traverse calculations. there is a unique path connecting the separator with any well. the flow rate in every node can be determined by summing up all the flow streams entering that node: q n . and Ω iw denotes the set of wells whose flow streams enter node i . According to Kirchoff’s law. Suppose node j receives fluids only from well j . the feasible set of oil flow rates is governed by: p rj − p sj = 0 . we can choose its solution node as any node that receives fluids solely from that well. we obtain the water and gas flow rates of that well using appropriate well models. p = o. The second calculation is from the separator (we assume the separator is operating at a fixed pressure) to node j along the path connecting the separator and node j . j denotes the flow rate of phase p p p for well j (phase g includes both the formation gas and liftgas. We can calculate the pressure of node j from two different pressure traverse calculations.i denotes the flow rate of phase p in node i . n w denotes the number of production wells. Ω n = n w s s (3. Node j is called the solution node for well j . superscript n and w denote nodes and wells.23) where q n . w (3. s In principle. Because of the treelike structure of the gathering system. g .CHAPTER 3. Denote the pressure for node j obtained from this calculation as p sj . Denote the pressure for node j obtained from this calculation as p rj .24) where Ω n denotes the set of solution nodes. j ∈ Ω n . Given the oil rate of a well. The first calculation is from the reservoir to node j . in the example problem (see Figure . SIMULATION OF GATHERING SYSTEMS 33 We choose the oil rates of wells as the independent variables and demonstrate how to construct the governing equations for multiphase flow in the system of this example problem. q w. according to mass conservation.
5. forming a closed system of equations that can be solved by the NewtonRaphson method.27) where J r denotes ∂p r ∂p s and J s denotes − . ∂q o ∂q o The iteration terminates when the following convergence criteria are satisfied: v v q o − q o−1 ∞ ≤ εq (3.29) where ε q and ε f are two userspecified tolerances. SIMULATION OF GATHERING SYSTEMS 34 3.24 has n w independent variables and n w equations. or 7 as the solution node for well 1.26) J= ∂p r ∂p s − = Jr + Js ∂q o ∂q o (3. node 7 and 8. we can choose either node 3. This choice has certain computational advantages. qo.28) and v f (q o ) ≤ ε f ∞ (3. and define f (q o ) = p r (q o ) − p s (q o ) (3.5. 3. be the primary variable. 3.CHAPTER 3. we always compute its upstream pressure based on its downstream pressure and the flow rate across it. In practice. we always chose the wellbore bottomholes.. 6. This guarantees a unique upstream pressure no matter the flow through the choke is critical or subcritical.25) Then Eq. i.e.2 The NewtonRaphson Method Let the oil flow rate of each well. or 8 as the solution node for well 2. when computing the pressure drop across a surface choke.24 can be solved by the NewtonRaphson method with the following iterative scheme: v v v v J (q o )(q o+1 − q o ) = −f (q o ) v where the Jacobian matrix J (q o ) is calculated as (3.2). node 4. For instance. 3.i . Eq. as the solution nodes. .
k = 1. A node can have multiple input nodes or wells. The pressure of all nodes and their derivatives are computed as follows: 1. Matrix J s can be a dense matrix. j .. say node i . n ∂xk ∂xk m ∂α m (3. For a node in level l . Matrix J s can be constructed efficiently by the procedure described below.CHAPTER 3. which is the adjacent downstream node connected to that well or node. . First we need to define some terminology. α m is the mth parameter that affects the pressure drop between node i and node j . the oil rate of a well). For level 1.. 2. Nodes are arranged in levels. The pressure derivatives of node j can be computed as ∂p j ∂xk = ∂p j ∂pi • ∂p j ∂α ij ∂pi + ∑ ij • m . gas. For node j or well j in level l + 1 .. then the pressure of node j can be expressed as p j = f ( pi . The node pressures are specified by the user and the pressure derivatives are zero. For the purpose of constructing matrix J s . in level l . The pattern of its nonzero elements depends on the connectivity of the gathering system. all its input nodes or wells are in level l + 1 .e. In the current network model. The nodes with fixed pressures are assigned to the first level.. which is the summation of matrices J r and J s . The element of J r can be evaluated analytically. matrix J r is a diagonal matrix because pressure p rj is a function of only one primary variable qo.31) ij where xk is the kth primary variable (i. α ij ) (3. α ij should include the total oil. and water flow rate through the link between node i and node j . SIMULATION OF GATHERING SYSTEMS 35 The major effort required in using the NewtonRaphson method is the construction of the Jacobian matrix J . all nodes are boundary nodes. which are the adjacent upstream nodes or wells connected to that node. A well or a node can have one active output node..30) where vector α ij denotes the parameters that affect the pressure drop between node i and node j . suppose we know the pressure and pressure derivatives of its output node.
3 Multiple Solutions Eq. ADIFOR (Bischof et al. automatic differentiation software reads the computer program and produces an augmented computer program capable of computing both the function and the derivatives of the function.. . We use a single well system to illustrate possible situations and discuss the complications in finding the true solutions of a gathering system with multiple wells. SIMULATION OF GATHERING SYSTEMS 36 Eq. • Derivatives and ∂p j ij ∂α m are time consuming to compute.24 does not always have a unique solution. 3. Given a computer program for computing a function.. 4. 3. 3. In this study we used such an automatic differentiation software.5. Repeat Step 2 for all nodes or wells in level l + 1 .CHAPTER 3. • Derivatives ∂p j ∂pi and ∂p j ij ∂α m can be approximated by the finite difference method or computed using some automatic differentiation techniques (Berz et al. 1996). once they are evaluated. 3. Let l = l + 1 . ∂xk ij ∂α m can often be computed analytically with trivial computational ∂xk • • Derivative Derivative time. Repeat Step 2 and 3 until we obtain the pressure and pressure derivatives for all nodes of the gathering system. to generate an augmented computer program for computing ∂p j ∂pi ∂p j ∂pi and ∂p j ij ∂α m . they can be used to compute the derivatives of node pressure p j with respect to all primary variables.31 is the key for efficient evaluation of the required derivatives because of the following reasons: ∂pi is known at this stage. However. 1998).
Consequently. The two curves have two intersections. Thus. we can calculate the bottomhole pressure from two pressure traverse calculations. at low flow rates. SIMULATION OF GATHERING SYSTEMS 37 Consider a single well system with a fixed well head pressure. For a fixed flow rate. p s first decreases then increases as the flow rate increases. We define the relationship between p r and the oil flow rate as the inflow performance curve. the region of decreasing p s with increasing flow rate is unstable and can cause intermittent or noflow condition (Lea and Tighe. The first calculation is from the reservoir to the bottomhole.1. The two curves have a unique intersection and the intersection represents the actual flow rate of the single well system. As a result of this the inflow and outflow performance curves may have two intersections (Figure 3. the second intersection represents the stable solution while the first intersection does not. under certain conditions. The two curves have no intersections. the pressure drop across the tubing string decreases with increasing flow rates.3). the pressure drop across the tubing string increases with increasing flow rates. Denote the bottomhole pressure from this calculation as p r . In this case the corresponding governing equation has no feasible solutions. This is the case when the reservoir pressure is too low to support the fluid column in the tubing string. Several possibilities exist when we overlay the inflow and outflow performance curves: 1.CHAPTER 3. The second path is from the surface (well head) to the bottomhole. 3. 1983). Denote the bottomhole pressure from this calculation as p s . Beyond a critical point. . As demonstrated in Section 3. However. 2. for upwards multiphase flow. We define the relationship between p s and the oil flow rate as the outflow performance curve.
The NewtonRaphson method can fail to converge no matter whether the network problem has feasible solutions or not. For a general gathering system with multiple wells.24. Several fundamental questions regarding this problem are: 1. in an optimization run.CHAPTER 3. If all runs fail to obtain a feasible solution. For these reasons. One possible approach to address this problem is to start the NewtonRaphson method from multiple starting points. PSIA 3900 3850 3800 3750 3700 3650 3600 3550 3500 0 1000 2000 3000 4000 5000 Outflow Curve Inflow Curve OIL RATE. Eq.24 is formulated with the assumption that fluids always flow up a production tree. which is not necessarily true for a particular set of operational settings.3: Illustration of multiple solutions for a single well system. then there is a large chance that the network problem is . finding the true solutions to the multiphase flow problem is more complicated. may have no feasible solutions because of the following reasons: • The production system may be operated with some unrealistic operational settings. STB/DAY Figure 3. 3. it is difficult to judge the feasibility of the network problem when the NewtonRaphson method fails to solve the problem. SIMULATION OF GATHERING SYSTEMS 38 3950 WELL SANDFACE PRESSURE. • Eq. For instance. 3. Is there a feasible solution? The governing equations for the network problem. a well may have a large lift gas rate but a small well choke diameter that is unable to deliver the large gas rate.
32. this work used Eq. j = 1. Asheim. j ≥ 0 . 3. For a fixed operational setting. production settings may be adjusted and the solution process restarted... otherwise it is not stable. How to determine if a feasible solution is stable? For naturally flowing wells. In such cases. To keep this study focused. and α denote the system parameters. the pipe diameters.6 Sensitivity Coefficients Sensitivity coefficients refer to the derivatives of system response with respect to some system parameters. SIMULATION OF GATHERING SYSTEMS 39 infeasible.32 as the stability criteria for all wells. 3.32) A solution is stable if it satisfies the inequality given by Eq. the stability of the flow depends on many factors.. For wells on gaslift. and several criteria are available to determine if the flow is stable (Fitremann and Vedrines. Therefore. 1988). model tuning.33) where u denote the system response.g. Sensitivity coefficients are required in applications such as optimization. the following inequality is used to determine if a solution to Eq. Sensitivity analysis can assist the . Suppose the multiphase flow in a gathering system is governed by f (u.e.CHAPTER 3.. Therefore. e. 3. 1983). available literature does not offer a complete and systematic solution to those questions.24 is stable: ∂p sj ∂qo . postanalysis is required to verify the validity of a network solution. i. the flow is stable if the outflow performance curve has a positive slope (Lea and Tighe. is it possible for a network problem to have only feasible but no stable solutions? Is it possible for a network problem to have multiple stable solutions? If the answers to these questions are positive. 2. history matching. 3. α ) =0 (3. 3. 1985. nw (3. Then ∂u are the sensitivity coefficients ∂α of system response u with respect to system parameters α .. the flow rates from each well. how do we identify and handle these scenarios in simulation and optimization? To date. and sensitivity analysis.
∆t ) = 0 (3. repeat the following steps: a. ∂u . 3. A brief description of this method goes as follows. In addition. flow rate. α are the reservoir properties (permeability and porosity). we need to solve an updated set of system equations. Solve f (u′. 1981).1 Finite Difference Method To obtain sensitivity coefficients approximation: 1. 3. and ∆t is the time step length. α ) = 0 to obtain u . For every perturbed parameter ui′ . Compute ∂u u′ − u . Let α i′ = (1 + ε i )α i .. Solve f (u. Consider a discrete reservoir system f (u k +1 .CHAPTER 3. u k .34) where u k are the system responses (well pressure and flow rates) at time step k . α′) = 0 to obtain u′ .2 Jacobian Method Landa (1997) presented an efficient procedure for computing the derivatives of dynamic production data (well pressure. For all α i . the accuracy of the sensitivity coefficients depends on the value of the perturbation ε i and care should be taken to ensure that ε i is appropriate (Gill et al. a conventional method is to use a finite difference ∂α 2.6. and reservoir saturation) with respect to static reservoir properties (permeability and porosity in grid blocks) for a reservoir history matching problem. it is not efficient. b. where ε i is a small perturbation.6. = ε iα i ∂α i The finite difference approximation is straightforward to implement. α. however. SIMULATION OF GATHERING SYSTEMS 40 design of a new gathering system and potentially be an effective way to identify bottlenecks in an existing system. c. In .
we add a small perturbation δα i to ∂α (3.CHAPTER 3..33.40) Again we need to construct and invert matrix J only once to obtain the derivatives of vector u with respect to the whole parameter vector α . Thus the . α + δα ) = 0 T (3. 3. we obtain f (u. α ) = 0 ...35 can be expressed as ∂f k +1 ∂u k ∂f k +1 ∂f k +1 ∂u k +1 − =− ∂α ∂u k ∂α ∂u k +1 ∂α The implications of Eq.36 are that we need to construct and invert matrix once to obtain the derivatives of u k +1 with respect to all parameters α . α ) + ∂f ∂f δα i = 0 δu + ∂u ∂α i (3. By expanding Eq.39) and J ∂u ∂f =− ∂α i ∂α i (3. δα i .38) Note f (u. The Jacobian matrix is defined by J= ∂f ∂u (3.0..36) ∂f k +1 only ∂u k +1 f (u + δu. 3. 3. we impose a small perturbation δα on the ∂α reservoir properties and expand Eq.0) ..37) where δα = (0. To obtain parameter α i ∂u . The same idea can be used to compute the sensitivity coefficients for a gathering system governed by Eq. 3. constructing the Jacobian matrix requires extensive computational time...0. 3.35) The first term in the lefthand size is zero. For the network problem.37 around α and ignoring the second order term.34 into a Taylor’s series as follows: f k +1 + ∂f k +1 k +1 ∂f k +1 k ∂f k +1 δu + δu + δα + O (δ 2 ) = 0 ∂u k +1 ∂u k ∂α (3. If we drop the second order term Eq.. SIMULATION OF GATHERING SYSTEMS 41 order to obtain sensitivity coefficients ∂u k +1 .
8.5 inches and a length of 4000 ft. chokes with a fixed diameter of 48/64 inches. The reservoir grid block pressures around these wells vary between 3400 psi to 4900 psi. The specific relative gravity of gas is 0. All wells are connected to the same platform through tubing strings. The fluid temperature in the tubing strings is linearly interpolated between the bottomhole fluid temperature and the surface fluid temperature. The fluid temperature in the surface pipeline is assumed to be 110 °F. we use an example to demonstrate various aspects of the procedures discussed in the previous sections for simulating the network flow and computing the sensitivity coefficients. The platform is connected to a separator through a pipeline with a diameter of 5. SIMULATION OF GATHERING SYSTEMS 42 Jacobian method is more efficient than the finite difference method. This is a synthetic example with ten vertical wells.7 Examples In this section.5 inches diameter surface pipelines of lengths between 60 to 2000 ft.1. Appendix A. 3.4 shows the configuration of the gathering system. which needs to solve an updated set of system equations whenever a parameter α i is perturbed.CHAPTER 3.4 shows the input file for this example. and 3. The separator is operated at a fixed pressure of 160 psi. . The oil gravity is 31° API. Figure 3. The fluid temperature at the bottomholes is assumed to be 160 °F.
we concluded that the proposed solution procedure is efficient. Run 1 had the slowest convergence rate. For Run 3 and Run 4. 3. respectively. In both runs.6 and 3. . f (q o ) ∞ dropped below 1 psi within 6 iterations. 1000. However. All runs were able to reach the convergence criteria. We deliberately started Run 1 with a bad initial guess so that the solution procedure would be difficult to converge. f (q o ) . f (q o ) dropped below 1 psi within 3 iterations.4: Configuration of the gathering system.7. 2000 STB/d oil rate for every well. 2. respectively.CHAPTER 3. the oil flow rate for each well approached its converged value within 24 iterations. Figure 3. decreased at each ∞ v iteration. Run 5 started the solution procedure with a different flow rate for each well. 200. SIMULATION OF GATHERING SYSTEMS Separator 21 Platform 43 11 12 13 14 15 16 17 18 19 20 Pipes Chokes 1 2 3 4 5 6 7 8 9 10 Tubing strings 1 2 3 4 5 6 7 8 9 10 Wells Figure 3.1 Simulation Results We made five runs with different initial guesses.7 show the oil flow rate of each well at every iteration for Run 1 and Run 4.5 shows how the v infinity norm of the residuals of the governing equations. Figure 3. As ∞ v expected. even for Run 1. 3. Therefore. Runs 1. and 4 started the solution procedure with 20.
E11 0 run2 run3 run4 run5 2 4 6 8 10 ITERATION Figure 3. STB/D 2500 2000 1500 1000 500 0 0 2 4 6 8 10 ITERATION W ell 1 W ell 2 W ell 3 W ell 4 W ell 5 W ell 6 W ell 7 W ell 8 W ell 9 W ell 10 Figure 3.E+01 f ∞ .E03 1.E09 1.E07 1.E05 1.E01 run1 1.E+03 1. SIMULATION OF GATHERING SYSTEMS 44 1.6: Oil rate versus iteration number for Run 1. 4500 4000 3500 3000 Qo.E+05 1.CHAPTER 3. PSI 1.5: Infinity norm of the governing equation residuals versus iteration number. .
86 2. 4.25 3088.48 0.54 1.32.84 2590. because for all wells ∂p sj / ∂qo . j is negative.99 144.63 2677.87 4.45 0. Evaluated by Eq.72 2. Well Run 1 ∂pis / ∂qo.37 Run 2 ∂pis / ∂qo .14 0. j is positive. SIMULATION OF GATHERING SYSTEMS 45 3500 3000 2500 Qo.68 616. and 5 are stable. STB/D 2000 1500 1000 500 0 0 1 2 3 ITERATION 4 5 6 W ell 1 W ell 2 W ell 3 W ell 4 W ell 5 W ell 6 W ell 7 W ell 8 W ell 9 W ell 10 Figure 3.58 933.2: Network simulation results from different runs. and Run 5 ∂pis / ∂qo .66 0.19 6.i Oil Rate (psi•d/STB) (STB/d) 781.CHAPTER 3.19 740.75 4.79 101.36 Run 3. 3.93 68. Run 4.95 1212.73 256.78 6.2.35 933.36 933.45 4. the solutions of Run 1 and Run 2 are unstable. The solutions of Runs 3.73 1186.97 0.i Oil Rate 1 2 3 4 5 6 7 8 9 10 (STB/d) 735.19 6.86 63.68 1044.57 0.98 (psi•d/STB) 0.05 0.05 1168. because for some wells ∂p sj / ∂qo .17 0.67 2.29 988.94 4.24 4.77 1142.44 1.05 0.67 913.56 828.24 748.52 0.36 .7: Oil rate versus iteration number for Run 4.i Oil Rate (psi•d/STB) (STB/d) 855.95 2809. As shown in Table 3.76 2.75 1. the five runs converge to three sets of solutions.97 1. Multiple solutions exist for this problem.53 1.21 0.95 1026.24 759. Table 3.39 3354.15 0.97 2910.
0080E1 4. a typical network simulation costs about 470 ms of CPU time. the Jacobian method requires only one network simulation and a little extra computational time.0820E2 εi = 0.8540E1 1. The extra computational time required by the Jacobian method is about 40ms.5943E+1 1.0001% 1.4996E+0 1. on a Silicon Graphics Origin 200 computer with four 270 MHz CPUs and 2304MB of RAM. and 10 chokes in the example problem.0817E2 εi = 0.0627E+1 4.1 / ∂qlg. Table 3.8545E1 1.8543E1 1. 21 pipes.6 / ∂d choke.5944E1 1. The results from the Jacobian method are in good agreement with the results from the finite difference method.8544E1 1.1 / ∂d choke.01% 1.001% of the parameter value is appropriate for this problem.2 Computation of Sensitivity Coefficients The purpose of this section is to demonstrate the use of the Jacobian method for computing the sensitivity coefficients.3020E1 1.6 / ∂qlg. (64STB/d/inch) ∂qo. (STB/MSCF) 1.1 . (STB/d/inch) ∂qo.8162E1 1.0820E2 ∂qo.3003E1 1.5943E+0 1. In contrast.01% and 0. SIMULATION OF GATHERING SYSTEMS 46 3.0080E1 4.1 .5935E+1 1.7.1 .0079E1 4.0820E2 εi = 0.5848E+1 1. (STB/MSCF) ∂qo.CHAPTER 3.0072E1 4.8505E1 1.0657E+1 4.3 compares several sensitivity coefficients obtained from both the Jacobian method and the finite difference method. Table 3.0790E2 ε i = 0.001% 1.3: Comparison of sensitivity coefficients obtained from the Jacobian method and the finite difference method.1 .1 (STB/d/inch) ∂qo. We used both the finite difference and the Jacobian method to compute the sensitivity coefficients of the oil flow rate of every well with respect to the gaslift rate of every well.0002E1 3.0819E2 .1% 1. (64STB/d/inch) ∂qo.0356E+1 4.0654E+1 4. Finite Difference Approximation Sensitivity Coefficients Jacobian Method ε i = 1% 7.3 suggests that a finite difference interval of between 0.0654E+1 4.1003E1 1. Table 3. Remember there are 10 gaslifted wells.6811E+0 4.1798E1 1.0079E1 4. the diameter of every pipe (including the tubing strings).3001E1 1.1 .1 / ∂d tubing . To obtain all the sensitivity coefficients. For this problem. the finite difference method needs to run 42 network simulations. The results from the finite difference method depend on the finite difference interval.2892E1 1.6 / ∂d tubing . and the diameter of every choke.
4.9. The sensitivity coefficients of the well oil rates with respect to the diameters of other surface pipelines and tubing strings are close to zero. the sensitivity coefficients with respect to d 21 decrease and the sensitivity coefficients with respect to the diameters of tubing strings increase. Figure 3. when d 21 increases to 7. d 21 .10). Figure 3. pipe 1121 are surface pipelines). As shown in Figure 3.CHAPTER 3. This suggests pipes 1112 are not properly designed and may have surplus capacity. the sensitivity coefficients with respect to the diameters of pipes 1120 are approximately zero. If we increase the diameter of pipe 21. .5 inches. the diameters of the tubing strings and pipe 21 has roughly the same impact on the well oil rates. respectively. pipe 110 refers to the tubing strings for well 110. This suggests that pipe 21 is the bottleneck of the gathering system.8 plots the sensitivity coefficients of the well oil rates with respect to the pipe diameters (as shown in Figure 3. the sensitivity coefficients with respect to d 21 become zero.8 shows that the oil flow rate of all wells except well 8 are sensitive to the diameter of pipe 21. SIMULATION OF GATHERING SYSTEMS 47 Sensitivity coefficients can give great insight about the flow behavior in a gathering system and can be used to identify bottlenecks of the system. In all scenarios. When d 21 increases to 9.5 inches (Figure 3.
8: Sensitivity coefficients of well oil rates with respect to system parameters (the inner diameter of pipe 21 is 5. 8 00 7 00 6 00 STB/D/INCH 5 00 4 00 3 00 2 00 1 00 0 1 00 5 10 PI PE 15 NUM BE R 20 9 10 5 6 7 8 1 2 3 4 M BE R W E LL N U Figure 3. SIMULATION OF GATHERING SYSTEMS 48 150 0 100 0 STB/D/INCH 5 00 0 5 00 5 10 PI PE N UM 15 BE R 20 9 10 6 7 8 3 4 5 1 2 W EL L NUM BE R Figure 3.5 inches).9: Sensitivity coefficients of well oil rates with respect to system parameters (the inner diameter of pipe 21 is 7. .5 inches).CHAPTER 3.
Brent’s method (Press et al..8 Concluding Remarks The drift flux slip model for the multiphase pipe flow and the Sachdeva model for the multiphase choke flow were presented in this chapter. was used to compute the sensitivity coefficients. SIMULATION OF GATHERING SYSTEMS 49 8 00 6 00 STB/D/INCH 4 00 2 00 0 200 5 10 PI PE NUM BE R 15 20 10 6 7 8 9 2 3 4 5 1 M B ER W EL L N U Figure 3. Computational experience showed that the NewtonRaphson method is not suitable for the rootfinding problems encountered in these models.CHAPTER 3. 3.5 inches). An efficient method. the Jacobian method. 1992) proved to be both efficient and robust. and the Jacobian method is more efficient than the finite difference method for computing sensitivity coefficients.10: Sensitivity coefficients of well oil rates with respect to system parameters (the inner diameter of pipe 21 is 9. . Results from an example showed that the NewtonRaphson method converges fast for our formulation. The Jacobian matrix was computed efficiently by exploring the treelike structure of the network. The multiphase network problem was formulated according to the mass balance requirement and Krichoff’s law and solved by the NewtonRaphson method.
How to build a procedure capable of automatically finding the stable solutions of a multiphase flow problem. How to build an efficient and robust solution procedure for multiphase flow problems of gathering systems with loops. Further research is required for the following problems: 1. . 2.CHAPTER 3. SIMULATION OF GATHERING SYSTEMS 50 It is straightforward to extend the formulation and solution methods presented here to compositional models.
1 Introduction Rate allocation refers to the problem of optimally adjusting production rates and lift gas rates of production wells to achieve certain operational goals. or liquid flow rate constraint can be put on any production well or network node. These goals vary with the field and time. if the oil production in a field is constrained by the gas processing capacity of the separation units. oil production can be constrained by fluid handling capacities of facilities. water. In some petroleum fields. in addition. the optimization problem can be expressed as maximize ∑ {q } nw w o . gas. An oil. closing or reducing production rates of wells with the highest GOR will increase the total oil rate. especially mature fields. reducing the lift gas rate of certain wells may increase the overall oil production by utilizing the gas processing capacity more efficiently. rate allocation can be an effective way to increase the oil rate or reduce the production cost.1a) 51 .Chapter 4 Rate Allocation through Separable Programming 4. In abstract form.i i =1 (4. For example. This study addressed the following rate allocation problem. For such fields. The objective function is the total oil rate.
j . q w. p ∈ {o. Fang and Lo (1996) studied a similar rate allocation problem. Specifically. With the above assumptions. The gaslift performance curves are concave. j . l } p 52 (4. j ∈ Ω n . Fang and Lo (implicitly or explicitly) made the following assumptions: 1. 4. and (2) the water and gas rates as functions of the oil rate of a well.1b) w n where qo. j is the flow rate of phase p in node j .CHAPTER 4. the method requires (1) the oil rate as a function of . the method requires (1) the maximum oil rate that is allowed or can be delivered from a well. j is the p flow rate limit of phase p for node j . The method was found to be very efficient. g . w. For gathering systems with a treelike structure. Q p . we will present a method that does not rely on assumption 1.2 Well Performance Estimation The method presented in this chapter required the performance estimation of individual wells. for unlifted wells. is a nonlinear function of p the control variables. j ≤ Q p . Ω n is the set of all nodes. 2. the flow rate of network node j is the sum of the flow rates of wells connected to node j . The GOR and water cut for a well remain constant for varying oil rate. Physically. q n . And the flow rate of each well. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING n subject to q n . Fang and Lo (1996) reformulated the rate allocation problem to a linear programming problem and solved it by the simplex algorithm. For lifted wells. To simplify the optimization problem. 3. the control variables for the rate allocation problem are the well chokes and lift gas rates of wells.i is the oil rate of well i . The well performance information can be evaluated individually for each well by ignoring flow interactions among wells. In Chapter 5. It was our aim to develop techniques that speed up the entire optimization process. In this part of the work we followed the line of Fang and Lo (1996) and developed other solution methods for the rate allocation problem.
Lea and Tighe. and its water and gas rates.CHAPTER 4. This is done in two steps: 1. Isolate the well from the surface pipeline network by fixing the pressure of a network node before the flow stream from that well joins the flow streams from other wells.1). A well inflow performance model relates the oil rate of a well to its bottomhole pressure. This study adopted the second approach. which are referred to in this study as the performance curves. Ignore well interactions in the reservoir by developing an inflow performance model for every well. Two common approaches are: (1) derive the performance curves from well test information. and (2) the water and (formation) gas rates as functions of the oil rate of the well. a well has to be isolated from the whole system. The production system is an integrated system in which the flow streams from different wells influence each other. To establish the performance curves for individual wells. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 53 the lift gas rate (the gaslift performance curve).1: Illustration of well performance curves. or (2) estimate the performance curves through simulation. 1991. An appropriate way to build this model is to run a . 1983). 2. How to construct the performance curves is an extensively studied area (Beggs. and lift gas rate curves (Figure 4. formation gas. The well performance information can be described by a set of oil rate versus water. 3500 3000 2500 Qg (MSCF/D) 2000 1500 1000 500 0 0 100 200 300 400 500 600 Qo (STB/D) formation gas rate vs oil rate lift gas rate versus oil rate water rate verus oil rate 3500 3000 2500 2000 1500 1000 500 0 700 Qw (STB/D) Figure 4.
4. The gaslift performance curve for a well can be constructed by determining the oil rate of the singlewell system for a set of prescribed lift gas rates.. Eq.. nw . m ij j i j x j ≥ 0 . 4. the water. 4. i = 1. thus the performance curves are piecewise linear curves (Figure 4.2b) (4. if we regard the oil rate as the control variables.2c) ∑ h (x ) ≤ b .3 Rate Allocation through Separable Programming The objective and constraint functions of the rate allocation problem (Eq. formation gas. We assume linearity between two adjacent points.CHAPTER 4..1) are linear combinations of the oil. 2. water.. To generate performance curves efficiently..1 and assumed that the reservoir condition in well blocks are constant for varying oil rates. and formation and lift gas rates of individual wells.2a) (4.. and lift gas rates of a well can be regarded as functions of the oil rate of that well. A simulated well test can be time consuming. and water and gas rate of a well for various oil rates using a numerical model. Further. The well performance curves are represented by a set of discrete points. The oil rate versus water and formation gas rate curves can be solely determined by the well inflow performance model. its performance curves can be established as follows: 1. j = 1. After a well is isolated from the production system. which can be expressed as maximize subject to ∑ f (x ) j j j =1 nw (4.. when the well performance is approximated by piecewise linear performance curves.1). denoted as x . Therefore. The test calculates the bottomhole pressure. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 54 simulated well test on a numerical model of the reservoir. this study used the well model presented in Section 3. A simulated well test mimics well tests performed in a real field..1 becomes an optimization problem whose objective and constraint functions are the sums of functions of one variable.
. j is the maximum oil rate for well j . a point (x.4d) for a given j .CHAPTER 4. λ jk ≥ 0 for all j . RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 55 where f j denotes the objective functions. bi denotes the limit of the i th constraint. no more than two λ jk can be positive and they must be adjacent (4.3b) ∑λ j=0 r j = 1. each of its well performance curves are defined by r j +1 discrete max max points with x j 0 = 0 and x jrj = qo . i = 1. First...4e) . Then for all functions f j (x j ) and g ij (x j ) we can write f j (x j ) = ∑ λ jk f jk . which can be solved by linear optimization techniques (Gill et al.3d) no more than two λ j can be positive and they must be adjacent Suppose for well j .. i = 0. m k =0 (4. Optimization problems of the form of Eq.. f (xi )) .. where qo .3a) f ( x ) = ∑ λ j f (x j ) r j =0 (4. 4. f (x )) on a piecewise linear curve defined by a set of discrete points (xi ..2 are separable piecewise linear problems.4c) ∑λ k =0 rj jk = 1 . k (4. hij denotes the j th function involved in the i th constraint.. λj ≥ 0 (4. f jk = f j (x jk ) k =0 rj rj (4.4a) g ij (x j ) = ∑ λ jk g ijk . g ijk = g ij (x jk ) .. r can be expressed as follows: r x = ∑λjxj j =0 (4. and m denotes the number of constraints. This is demonstrated in the following.4b) x j = ∑ λ jk x jk k =0 rj (4. j . 1981).3c) (4.
4.4.5: maximize z = ∑∑ f jk λ jk j =1 k = 0 nw rj 56 (4. 4.5e are treated. this model may allocate an unphysical point that is not on the performance curves.3. in general cases.5a) subject to ∑∑ g j =1 k = 0 rj nw rj ijk λ jk ≤ bi . then. 4.1 Model LPI This model was proposed by Fang and Lo (1996). constraint Eq.5a4..2.. 4. 1961). we would have a general linear programming problem. j = 1. there are several models for the rate allocation problem depending on how constraint Eq. These models were investigated in this study and are discussed below. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING Substituting Eqs. However. 4. nw (4.2 if the water and gas versus oil rate curves involved in the constraints are strictly convex.5c) (4.5b) ∑λ k =0 jk = 1 . Along this line.5d) λ jk ≥ 0 for all j .CHAPTER 4.5d automatically satisfies constraint Eq. 4. no more than two λ jk can be positive and they must be adjacent (4. It can be shown that if all f j (x j ) are concave functions and g ijk are convex functions.5e has to be enforced explicitly. The rate allocation problem is reformulated to a linear programming problem defined by 4. .5a4.5e is simply ignored.5e (Hartley.5d.... we obtain Problem 4.3. as demonstrated in Section 4. 4. In this model.4a4. Otherwise. constraint Eq. the optimal solution of an LP problem defined by Eqs.5e) Without constraint Eq. 4. k for a given j . 4. 4..5e. This model is able to solve correctly the rate allocation problem defined by Eq. m (4.4e into Eq.. i = 1.
6c) (4. for all j .....6: maximize z = ∑∑ f jk λ jk j =1 k = 0 nw rj (4. rj − 1 ..6g) (4.6f) (4.. 4.6a) subject to ∑∑ g j =1 k = 0 nw rj ijk λ jk ≤ bi .3. This model enforces constraint Eq. nw (4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 57 4. 1997) to obtain Problem 4. rj − 1 . for all j λ jr ≤ y j (r −1) . which can be equal to 1 only if x j k ≤ x j ≤ x j ( k +1) and 0 otherwise. for all j λ jk ≤ y j ( k −1) + y jk . In principle. i = 1. a MILP problem can be solved by enumeration.5 is then formulated as a mixed integer linear programming (MILP) problem (Bertsimas and Tsitsiklis.. 2000). k = 0..2 Model MILPI This model was applied to a rate allocation problem by Güyagüler and Byer (2001).6h) (4.6d) (4..6b) ∑λ k =0 rj jk = 1 . we introduce a binary variable y jk ..6e) (4. j = 1.. this model can contain a large number of binary variables and constraints. k y jk ∈ {0. m (4.1}.. k = 1. k = 1.5d explicitly and is suitable for rate allocation problems with performance curves of arbitrary shapes..CHAPTER 4. Several similar reformulations exist and their computational performances can be different (Padberg... However. for all j j j ∑y k =0 r j −1 jk = 1 . complete enumeration is computationally infeasible as soon as the number of integer variables in a . Problem 4. r j − 1 .6i) λ j 0 ≤ y j 0 . To enforce the constraint that for certain j at most two consecutive coefficients λ jk are nonzero.. The disadvantage of this method is that even for a rate allocation problem with moderate size. for all j λ jk ≥ 0 for all j ....
1998). This node represents a MILP subproblem generated in Step 1 or Step 6. An effective method for this purpose is the Branch and Bound method (Wolsey. the entire tree has been enumerated and the search ends. 3. Based on this idea. If the stack of active nodes is empty. Initialization... If the stack of active nodes contains several nodes.6 into the stack of active search nodes. and let z k = max{ T x : x ∈ S k } for k = 1. Consider a general optimization problem z = max{ T x : x ∈ S} c (4. Put Problem 4. Choosing a node. with each node in the tree representing a subproblem. Algorithm 4. an upper bound z ( z ≤ z ≤ z ).7) where x is the control variable and S denotes its feasible set. and the bound information on subproblems are used to prune the tree. 1998). A complete description of the general Branch and Bound method can be found in Chapter 7 of Wolsey (1998). Optimizing. Let z = −∞ and z = ∞ . Next we present a Branch and Bound method adapted for solving Problem 4..S K be a c decomposition of S into smaller sets.. and thereby forms an enumeration tree. .6. The algorithm maintains a lower bound z . Denote its ˆ optimal value as z .. and a stack of active search nodes representing subproblems to be examined. 2.CHAPTER 4. So we need some strategies to cut the number of necessary enumerations. A fundamental idea behind the Branch and Bound method is to divide and conquer. pop out the node on top of the stack. Branch and Bound is a general search method for optimization problems over a search space that can be represented as leaves of a tree. 1998). Solve the LP relaxation of the subproblem selected in Step 2.6. Branch and Bound recursively divides an optimization problem into several subproblems. The enumeration tree is constructed implicitly. 1.. Then z = max k z k (Wolsey. Let S = S1 ∪ . K . RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 58 MILP problem exceeds 20 or 30 (Wolsey. The definition of LP relaxation is discussed shortly.1: Branch & Bound for Problem 4.
RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 59 ˆ 4. z} and store the corresponding feasible solution. Otherwise. Add the two subproblems on top of the stack of active nodes in a prescribed order (such as the first subproblem goes first. Continue the search by going to Step 2. and the second subproblem goes second). (d) Pruning by infeasibility. go to Step 2 to backtrack.6. the solution from Step 3 must not be a feasible solution to Problem 4. k ∈ { . Some binary variable y jk has a fractional value. Pruning. where ε z is a prescribed tolerance.6. There is no need to divide further the subproblem represented by current node. rj − 1} has a fractional value. update z = z .6. If condition (a) is met. And the search ends. then the feasible solution corresponding to z can be regarded as the optimal solution of Problem 4. Bounding. Suppose variable y jk . 5.. If the node selected in Step 2 is the root node. (b) Pruning by feasibility. If condition (b). If z − z ≤ ε z .. 1 ≤ j ≤ nw .. 6. (c) or (d) is met. (a) Pruning by optimality. The second subproblem is comprised of the parent subproblem plus a constraint of y jk = 1 . The following conditions allow us to prune the tree and thus enumerate a large number of solutions implicitly. stop. There is no need to divide further the subproblem represented by current node. update z = max{z. 7. The upper bound of the subproblem represented by current searching node is below the lower bound. The solution from Step 3 is a feasible solution to Problem 4. The first subproblem is comprised of the parent subproblem plus a constraint of y jk = 0 .CHAPTER 4. ˆ (c) Pruning by bound: z ≤ z .6. If the solution ˆ in Step 3 is a feasible solution to Problem 4. The MILP subproblem from 0 Step 2 (the parent subproblem) can be divided into two subproblems. There is no need to divide further the subproblem represented by current node. .. Branching. go to Step 6 to branch. To reach this step. The problem examined has no feasible solution.
. a problem that has an optimal value no worse than that of the MILP subproblem. For wells that need a finite amount of lift gas to start flowing. an LP relaxation of Problem 4.2). 4.4..6i with the following linear constraints 0 ≤ y jk ≤ 1 . The bound information is used in Step 5 to prune the enumeration tree. 3. 2. rj − 1 for all j (4.CHAPTER 4. This model has the same number of variables and constraints as Model LPI. With these properties. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 60 The purpose of Step 3 is to compute an upper bound for the MILP subproblem from Step 2.3 Model MILPII This model was developed in this study.3. For wells that can flow naturally.2 can be used to solve it. The Branch and Bound method described in Section 4.. The oil versus water and formation gas curves are concave. The upper bound of a MILP subproblem is often obtained by solving a relaxed problem.2). An example of the application of this method is presented in Section 4. For example. constraint Eq.. their gaslift performance curves are comprised of a horizontal line and a concave curve (Curve II in Figure 4.. their gaslift performance curves are concave (Curve I in Figure 4.5e can be automatically satisfied if we require the coefficient λ j 0 in Problem 4.2). One such relaxation for a MILP problem is its linear programming relaxation that allows the integer variables in a MILP problem to take real values.6 is an LP problem that replaces constraint Eq. some of the decision variables are binary variables. This model is motivated by and is suitable for problems with the following properties: 1. 4. However.3. Thus this model leads to a MILP problem.8) 4.5 to be a binary variable for wells whose gaslift performance curves is nonconcave (Curve II in Figure 4.3. k = 1.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING
61
Curve I
OIL RATE
Curve II
A 0 LIFT GAS RATE Figure 4.2: Illustration of gaslift performance curves.
4.3.4 Model LPII
This model was first proposed by Lo and Holden (1992). This is a simple model that does not require the performance curves and therefore there is no need to satisfy constraint Eq. 4.5e. The model assumes the well water cut and GOR are constant for varying oil rates. This model takes a set of flow streams (either from production wells or from satellite reservoirs) as the input and scales them to meet the flow rate and velocity constraints in a way that maximizes an objective function. A flow stream is represented by the GOR, water cut, and a maximum oil rate of a well or a satellite reservoir. For example, suppose we want to maximize the total oil rate of a field subject to a total gas rate constraint. The problem can be formulated as maximize subject to
∑ xi qomax ,i
i =1 nw
nw
(4.9a) (4.9b)
∑x q
i i =1
max o ,i
GORi ≤ Q g
0 ≤ xi ≤ 1 , i = 1,..., n w
(4.9c)
max where xi denote the decision variable for Problem 4.9, qo ,i denotes the maximum oil
rate for well i , Q g denotes the limit of the total gas constraint. For gaslift wells, the GOR includes both the formation gas and lift gas and is defined as
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING
max q g + qlg, 0 max qo
62
GOR =
(4.10)
max where qlg, 0 denotes the lift gas rate that corresponds the maximum oil rate qo for a gas
lift well. In the optimal solution, xi = 0 indicates well i should be closed; xi = 1 indicates well i should be fully open; xi ∈ (0,1) indicates well i should be choked back.
4.3.5 Discussions
Model LPI and LPII are linear programming (LP) problems. Model MILPI and MILPII are mixed integer linear programming (MILP) problems. These models are also called the LPI, LPII, MILPI, and MILPII methods in this study, respectively. All these methods require the objective and constraint functions be in separable form, thus they are also referred to as the separable programming (SP) techniques. When the computational time is not a concern, the MILPI method is the most promising method because it can be applied to rate allocation problems with arbitrary well performance curves. However, the MILPI method can generate a MILP problem with a large number of binary variables and constraints even for a production system with a moderate number of wells. For largescale problems, the other three methods may be more appropriate depending on the characteristics of the performance of wells and how fast we want to get the solution. For example, this study successfully applied the LPI and LPII methods to the Prudhoe Bay oil field with appropriate speedup techniques (see Section 4.4.2). The optimization model for Prudhoe Bay oil field has more than 1000 producers and hundreds of flowrate and velocity constraints. The MILPII method is more appropriate for gaslift optimization problems where many wells cannot flow without gaslift (see Section 4.4.3).
4.4 Application to Rate Allocation Problems
When the various methods are applied to a rate allocation problem, there are some additional issues to be addressed: 1. How to handle nonlinear pressure and velocity constraints.
CHAPTER 4. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING
2. How to further speed up the solution process.
63
This section discusses these issues and presents a gaslift optimization example. The application of these methods to more complicated tasks (such as in well optimization and longterm reservoir development studies) are presented in Chapter 8.
4.4.1 Handling Pressure and Velocity Constraints
For a production system, besides the flow rate constraints, a minimum pressure can be specified for a node of the gathering system:
pin ≥ pin ,min
(4.11)
where pin denotes the pressure of node i . A maximum velocity constraint can be specified for a tubing string or a surface flowline:
v j ≤ v max j
where v j denote the in situ velocity of flowline j .
(4.12)
The SP techniques require that the optimization problem be in separable form. However, the pressure constraints on junction nodes and velocity constraints for surface flowlines are nonlinear functions of multiple well streams. The constraints do not satisfy the requirement of the SP techniques, so special procedures are required. In this study, the pressure constraints on network nodes are converted to flow rate constraints on individual wells using a network simulation procedure proposed by Litvak and Darlow (1995). In that procedure, whenever a pressure constraint is not honored at some node, say node j , the flow rate of related production wells are recalculated subject to the constraint:
p n = p n ,min j j
This procedure has been implemented in the VIP (Landmark, 2001) simulator.
(4.13)
The velocity constraints are incorporated into a separable problem by assuming that the velocity of a flowline is a linear function of the flow rates of that flowline:
v v v v (qo , q g , qw ) = co qo + c g q g + cw qw
(4.14)
a large number of flow rate and velocity constraints may be specified. Specifically.0 . RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 64 where qo denotes the oil rate of that flowline. 4.2 Speedup Techniques Although the SP techniques are very fast for the rate allocation problems. For these reasons. The idea presented here is to use domain knowledge to reduce the problem size.15) where qo. there is always a need to speed up the solution process. In addition. their efficiency is still of concern when it comes to largescale production systems with thousands of production wells.0 is a specific oil rate of that flowline (such as the oil rate of that flowline before max optimization). a rate allocation problem has to be solved repeatedly hundreds or thousands of times. For a rate allocation problem of a large complex production system. in well connection optimization (Chapter 6) or long term reservoir development (Chapter 8) studies. This can be done as follows. the performance curves can be truncated at a production rate limit either specified in advance or converted from a pressure constraint. MILPI. 0 v o (4. In addition.CHAPTER 4. The number of decision variables of Model LPI.0 is the maximum in situ oil velocity along that flowline corresponding to the reference oil rate qo. we want to reduce the number of decision variables and the number of constraints of an optimization problem.4. the weighting coefficient cv .o is a constant defined as follows: max vo . without sacrificing its accuracy. a gaslift performance curve can be truncated at the point at which the oil rate begins to decrease as more lift gas is injected. Variables for the gas and water phases are defined similarly. vo . 0 c = qo. There are several potential ways to speed up the solution process. and MILPII is proportional to the number of discrete points of the performance curves. Truncate the performance curves. To save computational time. Often some constraints are redundant (impossible to be violated) given the production limits of individual wells and other constraints specified for facilities and . Eliminate redundant physical constraints.
A node belongs to level l if the highest level of all its input nodes is level l − 1 . First we need to define some terminology. Determine whether a userspecified constraint for node j can be violated by the potential input of node j .CHAPTER 4. The definition of input node and output node is given in Section 3. 3. In that problem. 2b. it is a redundant constraint and can be eliminated. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 65 flowlines. and water flow rate that a node can deliver to its output node.. and water flow rate that a node can receive from its input nodes. Processing a node. 2. For a node j in level l . 1. perform the following steps. Initialization. 2c. gas. Let l = l + 1 . Let l = 1 . With these definitions. Repeat Step 2 for all nodes in level l . it is not a redundant constraint. The speedup technique described here was able to . The potential output of node j is defined by the constraint limit. Repeat Step 2 and 3 until all nodes are processed. The potential output of a node is an upper bound on the oil. Nodes are arranged in levels. the connectivity of the production system). gas. Repeat Step 2b for all userspecified constraints for node j . Some of the constraint redundancy can be eliminated using domain knowledge (i. Let the potential output of node j equal to the potential input of node j . the user specified 190 velocity constraints and 53 flow rate constraints for facilities (the production limits on individual wells are not counted). In this section we will describe such a constraint elimination procedure. • If the constraint can be violated by the potential input.5. this constraint elimination procedure has been applied to a rate allocation problem in Prudhoe Bay oil field. 2a. the constraint elimination procedure goes as follows. 4. Calculate the potential input of node j by summing up the potential output of all its input nodes.2. The potential input of a node is an upper bound on the oil. A node belongs to the first level if all its input nodes are wells. Let the potential output of all wells be their production limits. • If the constraint can not be violated by the potential input.e. As an example.
In this example. The problem is to optimize oil production from a set of 56 wells with 22. The binary variables {y jk } in Model MILPI are introduced to enforce constraint Eq. Results for the equalslope method and the ExIn method . Buitrago et al. 4. When the objective and constraint functions of Problem 4. (1996) solved the problem by both an equalslope method and the ExIn method. When this rate allocation problem is solved by the LPII method. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 66 eliminate 226 constraints.. wells 4756 can not flow without gaslift..500 MSCF/d of available gas. Avoid unnecessary variables and constraints in reformulation.5 have certain convex/concave properties.5e can be automatically satisfied when the LP problem defined by 4.3 A Gaslift Optimization Example This gaslift optimization example has been presented in Wang et al.1 compares the performance of the MILPII method with the equalslope method and the ExIn method. 4. This is not always necessary. 2.4.1 of Appendix B.5a4. The well is involved in a lift gas volume constraint while its lift gas rate versus oil rate curve is not strictly convex. this problem was solved by the MILPII method.5d is solved by a general LP solver (Hartley. consider a rate allocation problem whose objective is to maximize the daily oil production. (1996). The problem of this example is from Buitrago et al. (2002a). Table 4. a speedup factor of more than five times was achieved as a result of this constraint elimination procedure.5e. there is no need to introduce the binary variables {y jk . The gaslift performance data are shown in Table B. leaving only 17 constraints. constraint Eq. rj } for well j unless at least one of the following conditions is met: 1. In this study. 3. The well is involved in a total gas (including formation and lift gas) flow rate constraint while its total gas rate versus oil rate curve is not strictly convex. k = 0. The well is involved in a water or liquid flow rate constraint while its water rate versus oil rate curve is not strictly convex. 4. 1961).. This rule can be used to reduce the number of decision variables of Model MILPI.CHAPTER 4. a stochastic algorithm that calculates the descent direction heuristically.. For example.
For example. the data points allocated by both methods are on the gaslift performance curves. To allocate 21265 STB/d oil. Results demonstrated that the MILPII method is very efficient for this problem. (1996). One shortcoming of the LPI method is that it may allocate gas from a point not lying on the gaslift performance cuve.454 Lift Gas Rate (MSCF/d) 21.1: Gaslift allocation results obtained from different methods. in the second case.CHAPTER 4. The third row of Table 4.2 shows the computational time (elapsed time). the LPI method was also used to solve this problem. To investigate the efficiency of the MILPII method.500 22.3. To investigate the advantages of the MILPII method over the LPI method (Fang and Lo.790 Oil Rate (STB/d) (1) Allocate all available liftgas of 22500 MSCF/d.175 21.0% less lift gas than the equalslope method.265 (3) 17.2 of Appendix B shows the optimal lift gas and oil rate allocated by the MILPII method for every well. the MILPII method requires 37. EqualSlope ExIn (1) 22. The MILPII method outperforms both the equalslope method and the ExIn method for this example. To allocate 21790 STB/d oil. .508 20. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 67 are taken from Buitrago et al.000 MSCF/d lift gas available. All computations were performed on a SunEnterprise 5500 machine with 4GB of RAM and 2GB of swap space.6% less lift gas than the ExIn method.040 21. the MILPII method requires 16. (3) Minimize liftgas rate while keeping the oil rate at 21790 STB/d. The second row of Table 4. Table B.2 shows the number of LP problems solved in the solution process. Specifically.632 MILPII (2) 14. This is demonstrated in Figure 4. (2) Minimize liftgas rate while keeping the oil rate at 21265 STB/d. the MILPII method allocates 6.4% more oil than the equalslope method. Table 4. 1996). For this well. For the rest of wells.790 22. the point selected by the MILPII method is right on the gaslift performance curve while the point selected by the LPI method is far above the gaslift performance curve. which presents the gaslift performance curve and the allocated lift gas and oil rates from both the MILPII method and the LPI method for well 47. we have 560 wells and 225.265 21. using the same amount of lift gas. we applied it to other similar problems with different sizes.
Table 4. 3. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 68 500 450 400 OIL RATE. The speedup factor can be significant for largescale systems.3: The allocated lift gas and oil rates for well 47 from the MILPII method and the LPI method.5 Concluding Remarks In this chapter we addressed the rate allocation problem for production systems where the flow in surface flowlines has little impact on well performance. Number of Wells Elapsed Computational Time (Second) Number of LP Problems Solved 56 0.29 40 4.2: Computational time of the MILPII method for various problems. MSCF/D gaslift curve LPI allocation MILPII allocation Figure 4. .06 6 3360 140. 2.27 5 560 3.CHAPTER 4. STB/D 350 300 250 200 150 100 50 0 0 1000 2000 3000 4000 LIFT GAS RATE. which is suitable for gaslift optimization problems in which some wells need a finite amount of lift gas to start flowing. The contributions of this study can be summarized as follows: 1. Developed several effective speedup techniques. The pressure and velocity constraints are handled through preprocessing and linear approximations. The approaches used here follows the line of Fang and Lo (1996). Proposed the MILPII method.
CHAPTER 4. The coupling procedure and applications of these methods are presented in Chapter 8. RATE ALLOCATION THROUGH SEPARABLE PROGRAMMING 69 The LPI and LPII methods have been coupled with a reservoir simulator so that the impact of the reservoir can be updated at every iteration during a time step. . The quality of performance curves is crucial to the success of the optimization procedure presented here. Therefore further research should be conducted to address the problem of how to construct good quality performance curves efficiently.
For such cases. a sequential quadratic programming 70 . especially for offshore production systems.Chapter 5 Rate Allocation through Sequential Quadratic Programming 5. Both genetic algorithms and gradientbased methods were considered. The major selection criteria are the efficiency and robustness of the algorithm. Consequently the rate allocation problem has to be formulated as a nonlinearly constrained optimization problem. Out of the many gradient methods for constrained problems. rigorous network simulations have to be employed in the optimization process to capture complex flow interactions among different wells. a gradientbased method is deemed appropriate for the rate allocation problem. Because a network simulation is costly.1 Introduction Chapter 4 simplifies the rate allocation problem to a separable programming problem by assuming that the flow interactions among production wells are not significant. where the oil production is often constrained by the fluid delivering capacities of some trunk pipelines. but they are not good at handling nonlinear constraints and usually require many more function evaluations than gradientbased methods. The first step in solving the nonlinear rate allocation problem is to select an appropriate optimization algorithm. However. The genetic algorithms can avoid being trapped in local optima. this assumption does not hold for many petroleum fields.
m (5. Consider an optimization problem. Results showed that the method is capable of handling rate allocation problems of varying complexities and sizes. This section presents a brief description of the SQP methods (Gill et al. 1997).. Finally it presents some applications of the algorithm. Problem 5.1b) where the objective function f and the constraints {ci } are functions of the control variable x .CHAPTER 5. SQP methods seek a local minimum x * of the constrained optimization Problem 5.. i = 1. Then it discusses the appropriate formulation of the rate allocation problem and related practical issues.. .2 Sequential Quadratic Programming SQP methods are widely considered the most effective algorithms for solving nonlinearly constrained optimization problems (Murray. λ ) ≡ f (x ) − λ T c (x ) (5. This chapter first presents a brief description of the SQP methods.. These methods are regarded as the best optimization method for nonlinearly constrained problems whose number of decision variables is small or moderate (Murray. ˆ Let the vector c (x * ) denote the subset of t constraint functions that are active at x * .1 can be expressed in terms of the objective function and active constraints (Gill et al. c (x ) ≥ 0 . 2002).1. with c (x * ) = 0 . 1997)...1.1a) (5. The following conditions are necessary for x * to be a local minimum of Problem 5.1 (Gill et al..2) where λ denote the vector of Lagrangian multipliers. and ˆ let A (x * ) be the matrix whose rows are the transposed gradient vectors of the active constraints. 5. RATE ALLOCATION THROUGH SQP 71 (SQP) method was selected. with nonlinear inequality constraints minimize f (x ) subject to ci (x ) ≥ 0 . The constraint ci (x ) ≥ 0 is said to be active at x * if ci (x * ) = 0 . and inactive if ci (x * ) > 0 . The associated Lagrangian function for Problem 5. 1981) ˆ L(x. 1981) ˆ 1.
i = 1.3) where p k is a search direction and α k is a step length.. equivalently. by using the following iteration scheme x k +1 = x k + α k p k (5. which combines both the objective and constraint violations into one function. The search direction p k approximates the error (x * − x k ) and the optimality conditions at x * should guide the definition of p k .4b) subject to ∇c(x k )p ≥ −c k where p denote the decision variables. or. These optimality conditions for a constrained optimization problem are called the second order KKT (KarushKuhnTucker) conditions. and the matrix B k represents the Hessian of the Lagrangian function.5) . RATE ALLOCATION THROUGH SQP ˆ 2. 1981) M (x ) = f (x ) − ∑ ρ i min (ci (x ).. The two most common choices of the merit function are the l1 merit function (Gill et al. For constrained problems the choice of step length α k needs to ensure that the next iterate will not only decrease f but it also will come closer to satisfying the constraints. λ * )Z (x * ) is positive semidefinite T ˆ where Z(x * ) is a basis of the nullspace of A (x * ) . and i 4.4a) (5. x * .. The most common approach is to choose α k in Eq.CHAPTER 5. These methods determine the search direction p k by solving a linearly constrained quadratic programming (QP) subproblem 1 minimize g T p + p T B k p k 2 (5. SQP methods are based on these optimality conditions and they approach a local minimum. Z (x * ) ∇ xx L(x * . Z (x * ) ∇f (x * ) = 0 . 5.0 ) i =1 m (5.3 to produce a “sufficient decrease” in a merit function. t . ∇f (x * ) = A (x * ) λ * . g k denote ∇f (x k ) . The step length α k is usually determined by a line search method.. T T 72 3.. λ* ≥ 0 .
RATE ALLOCATION THROUGH SQP 73 where ρ i > 0 are penalty parameters. links. The focus is on how to handle the well chokes. 5.7) The objective function of the QP subproblem and the merit function of the line search step require estimates of the Lagrangian multipliers.. one of the physical control variables of the rate allocation problem.6) where s is a vector of slack variables to convert inequality constraint ci (x ) ≥ 0 to an equality constraint ci (x ) ≥ 0 if and only if ci (x ) − si = 0 .3 Formulations of the Rate Allocation Problem Formulation plays an important role in optimization. The simulation model for the gathering system is described in Chapter 3. and maximum amount of lift gas available for groups of gaslift wells. The constraints include maximum/minimum flow rates and pressure constraints imposed on production wells and/or network nodes. A gathering system includes wells. This section discusses the appropriate formulation for the nonlinear rate allocation problem.3.1 Optimization Problem We considered the rate allocation problems for gathering systems with treelike structures.. 2002) M (x. ρ ) ≡ f (x ) − λT (c(x ) − s ) + 1 ρ (c(x ) − s )T (c (x ) − s ) 2 (5. and the augmented Lagrangian merit function (Gill et al.CHAPTER 5. 5. The physical decision variables include lift gas rates and well chokes that control the production rate of individual wells. The objective function is the weighted sum of well rates. s. These estimates can be obtained by solving an auxiliary problem or by using the optimal multipliers for the QP subproblem at the previous iteration (Gill et al. . si ≥ 0 (5. λ. 1981). and nodes.
Eq. given x . In addition.24. if some wells are so deep that they cannot produce without an appropriate amount of liftgas. and l and u denote the lower and upper bounds of the constraints.8b) where f (x ) denotes the objective function. timeconsuming because we need to solve two sets of nonlinear equations to obtain one partial derivative. we need to solve two sets of system equations with different dimensionalities (one set without well j and one set with well j being slightly opened) to evaluate the partial derivatives of the objective function and constraint functions with respect to the choke diameter for well j .24 will have no feasible solutions unless appropriate amounts of liftgas are allocated to those wells.8a) (5. This standard formulation is named here Problem P1. inaccurate because we have to use finite differences to approximate the derivatives of a potentially nonsmooth function. have to be solved. Another significant drawback of Problem P1 is that it is not good at handling well shutdown actions. Eq.24 will have no feasible solutions if an excessive amount of gas is injected into a well while the choke for that well is set to a small value in the optimization process. a particular set of choke diameters and liftgas rates. Eq. the weighted sum of well rates. the flow rates and/or pressures of wells and network nodes. the flow rates and pressures of some wells and facilities. When a well.2 Formulation P1 A straightforward formulation of the well rate and liftgas rate allocation problem is to choose the choke diameters and liftgas rate as the decision variables and formulate the optimization problem as follows maximize f (x ) subject to l ≤ F(x ) ≤ u (5. 3. to evaluate f (x ) and F(x ) .CHAPTER 5. this can cause severe computational difficulties. 3. say well j . RATE ALLOCATION THROUGH SQP 74 5. is converted from open status to shutdown status. 3. This procedure can be timeconsuming and the results can be inaccurate. In some cases. For Problem P1. F(x ) denotes the constraint functions. the system equations of the gathering system.24 do not exist.3. Therefore. there are chances that the solution process for Problem P1 will fail because feasible solutions for Eq. 3. . For instance.
even when these computational difficulties do not appear. RATE ALLOCATION THROUGH SQP 75 Some of these computational difficulties are inherent in the system of equations. and well shutdown handling. as we will demonstrate by an example in Section 5. 6000 5500 SANDFACE PRESSURE.5. Replacing Eq. The fundamental idea of the new formulation is to avoid solving the system of equations when evaluating the objective and constraint functions.2.1 illustrates this point.1: Outflow performance curves with various choke settings. Eq.CHAPTER 5. 3.3 Formulation P2 Here we propose a new formulation to overcome limitations of Problem P1 discussed in the previous section. For a set of oil and . 5.3. As described in Section 3.24 is formulated under the assumption that the fluid flows up the production trees.5.9/64 IN Dc=25/64 IN Dc=18/64 IN Dc=13/64 IN Dc=48/64 IN DC=35/64 IN Figure 5. the computational efficiency of Formulation P1 can be disappointing. The new formulation chooses the oil rate and the liftgas rate for each well as the decision variables and use the fact that for fixed flow rates the pressure drop across the choke increases as the choke closes. such as infeasible combination of liftgas rates and choke settings. STB/D 1500 2000 Dc=8.0/64 IN Dc=4. but some difficulties will persist. PSI 5000 4500 4000 3500 3000 2500 2000 1500 1000 0 500 1000 OIL RATE. 3.24 with another system of equations may help resolve some of the computational difficulties. Furthermore. Figure 5.
the flow rate and bottom hole pressure of a well are computed from the reservoir side using appropriate well models when computing p rj . they are evaluated in a different way in Problem P2 than in P1... The other path is from a separator to the bottomholes.. x . the flow rate and pressure in network nodes are computed from the separator side when computing p sj . However. This constraint should be modified to p rj − p sj = 0 if the choke diameter of well j is not a control variable. As in Problem P1.5. reducing choke diameters. f (x ) and F(x ) denote the flow rates and pressures of some wells or facilities. we calculate the bottomhole pressures from two paths as described in Section 3. One path is from the reservoir side to the bottomholes..10c ensures that the optimal oil and liftgas rate for Problem P2 are within the deliverability capacity of the gathering system. Thus these wells can be produced from the gathering system by appropriately adjusting the choke diameters. 5. n w (5.. denoted as x .10a) (5. nw (5.9.CHAPTER 5.10b) (5.. When performing the pressure transverse calculation from the separator side.1. The set of oil and liftgas rate x thus computed is within the deliverability capacity of the gathering system if we have p rj − p sj ≥ 0 . The well rate and gaslift optimization problem is now formulated as Problem P2 maximize f (x ) subject to l ≤ F(x ) ≤ u p rj − p sj ≥ 0 .10c) Constraint Eq. . In other words. given a set of oil and liftgas rates. where appropriate. j = 1.10c is named as the deliverability constraint. p sj is the bottomhole pressure for well j calculated from the separator side. j = 1. Constraint 5. this set of oil and lift gas rate can be produced from the gathering system by.9) where p rj is the bottomhole pressure for well j calculated from the reservoir side. 5. we set the well chokes fully open. if a set of oil and liftgas rates satisfies constraint Eq. RATE ALLOCATION THROUGH SQP 76 liftgas rates. In Problem P2...
These results need further interpretation so that the desired choke diameters can be determined.10c.4. it can convert the rate and pressure constraints on individual wells to a bound constraint on the oil rate. A zero oil rate and lift gas rate indicate that well j should be closed. some partial derivatives required by the optimization process can not be evaluated. its allocated oil rate.e. A nonlinear constraint is much more difficult to handle than a simple bound constraint (Gill et al.24) to evaluate the objective and constraint functions. a small positive value (i. Because Problem P2 takes the oil rate as one of the control variables. 3. we do not solve the system equations (Eq.. RATE ALLOCATION THROUGH SQP 77 In the solution process for Problem P2. The lower bound of the oil rate is not allowed to be zero. and the bottomhole pressure can be regarded as functions of its oil rate. . and the status of deliverability constraint Eq. For an arbitrary well j .3.. because at zero oil rate. In contrast.4 Interpretation of Solutions of Problem P2 The optimal solution of Problem P2 includes the oil rate and lift gas rate of each well. Formulation P1 has to treat the rate and pressure constraints on individual wells as nonlinear constraints. This is explained further in Section 5. 5. and the status of corresponding deliverability constraint in the optimal solution can be interpreted as follows: 1. For a production well. 1 STB/d). lift gas rate. the required choke diameters can be determined according to the value of the control variables and the status of constraint 5. The desired choke diameter can be computed according to the desired pressure drop across the well choke for the optimal oil rate and lift gas rate of well j . a zero oil rate really means that the oil rate for well j is at its lower bound. Problem P2 simplifies the constraints on individual wells. 1981). After the optimization problem is solved. A positive oil rate but inactive deliverability constraint ( p rj > p sj ) indicates well j should be partially closed (a positive oil rate means the oil rate for well j is greater than its lower bound). 5. thus most computational difficulties encountered in Formulation P1 can be avoided.10c (active or inactive). its water and gas rate.3. 2. Note.CHAPTER 5.
and an active deliverability constraint indicate that well j should be closed. If the oil rate is close to its lower bound. for certain wells. it is necessary to check whether the optimal production rate is a stable production rate for the production system. a (usually small) positive lift gas rate. then it may indicate well j should be closed.CHAPTER 5. However. and the optimization problem has to be resolved. If certain deliverability constraints are infeasible because the corresponding wells are too weak to flow. A zero (or close to the lower bound) oil rate. To improve further the objective and keep the deliverability constraint feasible.32 presented in Section 3. 4. Finally. 5. the optimization algorithm has to reduce its production rate but allocate some lift gas to keep the deliverability constraint feasible. the corresponding wells have to be manually shut down. This is because.3 can be used for this purpose. say well j . . for a production well. A positive oil rate and active deliverability constraint ( p rj = p sj ) usually indicates that well j should remain fully open.5. increasing the oil rate and lift gas rate a little bit at low oil (or liquid) rate region can significantly reduce pressure drop in the tubing string (thus the sandface pressure calculated from the surface side decreases). Condition 3. Infeasible problem. As shown in Figure 5. as its production rate reduces.2. reducing its production rate improves the objective function. the pressure drop in its tubing string increases and the corresponding deliverability constraint becomes active. RATE ALLOCATION THROUGH SQP 78 3. See next item (item 4) for explanation.
SNOPT implements a SQP algorithm that obtains search directions from a sequence of quadratic programming subproblems. SNOPT requires relatively few evaluations of the problem functions.5 inches. they were solved by SNOPT (Gill et al.CHAPTER 5. STB/D QLIFT=120 MSCF/D QLIFT=200 MSCF/D QLIFT=60MSCF/D QLIFT=80 MSCF/D QLIFT=100 MSCF/D QLIFT=0MSCF/D QLIFT=20 MSCF/D Figure 5. (2002). RATE ALLOCATION THROUGH SQP 79 4000 3900 SANDFACE PRESSURE. PSI 3800 3700 3600 3500 3400 3300 3200 3100 3000 0 10 20 30 40 50 OIL RATE.26 STB/STB. Successful application of SQP requires efficient and accurate evaluations of gradients of objective and constraint functions. The water cut is 0. well depth is 8000 ft. well head pressure is 1080 psi. tubing diameter is 3. 1998). SQP is a derivativebased optimization algorithm.50 MSCF/d.4 Solving the Optimization Problem Problems P1 and P2 are nonlinearly constrained optimization problems. GOR is 12. a generalpurpose system for solving largescale optimization problems. hence it is especially effective if the objective or constraint functions are expensive to evaluate. The SQP algorithm used in SNOPT is described fully in Gill et al. we were able to compute the gradient information efficiently and accurately by exploring the treelike structure of the gathering system and utilizing . In this work.2 Impact of oil rate and lift gas rate on the well sandface pressure calculated from the separator side. 5. Major characteristics of this well are as follows. In this work..
The constraint is that the total water production rate can not exceed 5000 STB/d. which is 1545 psi.5. the water cut.2) for a well and liftgas rate allocation problem. These examples have been presented in Wang et al. The first example compares the performance of the SQP method with the MILPI method (see Section 4. . The second example compares the computational efficiencies of Formulations P1 and P2. (2002b). and the gas oil ratio (GOR) for each well obtained in this step. The problem is to maximize the total oil production rates by allocating production rates and unlimited liftgas to the ten production wells. the convergence criteria. 1.1. The optimization problem is solved by the MILPI method presented in Section 4.3. Suppose wells are producing with a set of liftgas rate (not optimized) as shown in column 2 of Table 5. the gathering system and the optimization problem. Therefore.1 Comparison of the MILPI Method and the SQP Method The production system of this example is described in Section 3. and the model and derivative accuracy.5.1 also shows the oil flow rate. Experience with the application of SNOPT to the nonlinear rate allocation problem showed that the number of required function evaluations is sensitive to the starting point.7. The third example demonstrates the computational efficiency of Formulation P2 on problems of various sizes and properties. Choose a realistic flowing pressure at the platform.5 Examples We present three examples in this section. We simulated the multiphase flow in the gathering system and obtained the flowing pressure at the platform.CHAPTER 5. 5.2. 5.3. Table 5. Details of this procedure are presented in Section 3. the performance of SNOPT should be tuned properly whenever the condition of the production system or the optimization problem changes significantly. RATE ALLOCATION THROUGH SQP 80 automatic differentiation techniques.2 as follows.
3000. The allocated oil rate and liftgas rate for each well are shown in Table 5. The flowing pressure at the platform is 1595 psi. we could determine its oil flow rate by solving the multiphase flow problem for the flow path connecting well j and the platform. 3. 6000. simulate the multiphase flow in the gathering system. 4. the program converged to the optimal solution using 27 function evaluations. Results indicate that the field will produce at a total oil rate of 13019 STB/d and a total water rate of 5000 STB/d. Construct gaslift performance curves. Then for any well. Starting with an initial guess of 400 STB/d of oil rate and 3000 MSCF/d of lift gas rate for every well. and 7000 MSCF/d. RATE ALLOCATION THROUGH SQP 81 2. The allocated liftgas rate and oil rate for each well are presented in Table 5. given its gaslift rate. and obtain its gaslift performance curve. 500. and obtained the oil flow rate of each well. say well j . Update multiphase flow. 5000. Then we solved the rate allocation problem using the MILPI method. We selected the oil rate and liftgas rate for each well as the decision variables.2. 200. The optimal solution predicted that the field will produce at an oil flow rate of 12449 STB/d. 4000. namely. The flowing pressure at the platform is 1079 psi. One function evaluation refers to one complete evaluation of the objective and constraint functions and their derivatives. 2000. For each well. First we fixed the platform pressure at 1545 psi. 0. We then proceeded to solve the same optimization problem using the SQP method with Formulation P2. which does not violate the total water flow rate constraint. First we approximated each gaslift performance curve using a piecewise linear curve. We allocated the optimal set of liftgas rates from Step 3 to each well. Results indicated that the field can produce at a total oil flow rate of 12045 STB/d and a total water flow rate of 4830 STB/d.CHAPTER 5. Note that in this example the flowing pressure at the platform is sensitive to the flow rates in the common flow line. Allocate liftgas. we computed the oil flow rate for different gaslift rates.2.3 plots the gaslift performance curves we obtained in this step. 1000. . Figure 5.
00 0 0 2000.25 121842. STB/D .21 12.50 1.26 0.38 31.38 0. MSCF/d 3500 3000 2500 2000 1500 1000 500 0 0 2000 4000 6000 8000 LIFTGAS RATE.00 0 2000. STB/d total formation gas flow rate.00 0 0 3000. RATE ALLOCATION THROUGH SQP 82 Table 5.52 Water Cut 0.CHAPTER 5.49 933.02 13.47 0.02 10618.65 1.1: Well rates obtained in Step 1 of the MILPI method for Example 1.35 988. Well 1 2 3 4 5 6 7 8 9 10 Total Flow Rate * ** Liftgas Rate (MSCF/d) 3000.41 1. For clarity.10 0.69 828.87 40. MSCF/D W ELL 1 W ELL 3 W ELL 4 W ELL 6 W ELL 7 W ELL 10 Figure 5.84 2.00 13000.00 Oil Rate (STB/d) 735. curves for high GOR wells (like well 6) are not shown.18 1168.41 6.07 0.11 0.80 1143.3: Gaslift performance curves generated in Step 2 of the MILPI method.42** total water flow rate.09 144. OIL RATE.16 0.73 2911.53 0.03 740.55* Gas Oil Ratio (MSCF/STB) 1.70 0.22 6149.00 3000.90 2590.
24 22180. • The major computational effort of the MILPI method is in constructing the gaslift performance curves. 5.00 2678. So the MILPI method tends to overinject liftgas in Step 3.64 5000.00 649. well 8 and 9 have high GORs.36 2674. we needed to fix the platform pressure to a rather arbitrary value.68 * * * * * * 0 1168.2: Allocated lift gas and oil rates for Example 1.15 3108.21 0 731. • In Step 3 of the MILPI method. However. consequently the optimization algorithm will reduce the flow rates from other wells.03 * * ** 0 741.30 0 932.60 * * 0 988.00 668.46 0 1293.22 1149.00 2608. As a result.00 12045. the liftgas was allocated based on existing gaslift performance curves.56 506.CHAPTER 5.97 2769.64 0 947.62 0 933.00 1082.00 3070. • The SQP method takes the flow interaction into account through the deliverability constraint Eq. For example. The SQP method closed well 8 and partially closed well 9 to utilize fully the deliverability capacity of the common flowline.07 0 1237. and is able to make more informed decisions.54 0 3233.00 872. The results of the MILPI method will largely depend on this value.00 927. there is no simple way to tell which value is appropriate to choose.29 7000.40 13019. Thus the computational time of the MILPI method is .57 750. the SQP method allocated much less liftgas but produces 8% more oil than the MILPI method. RATE ALLOCATION THROUGH SQP 83 Table 5.29 506.22 12449. This ignores the fact that increasing the flow rate from one well will decrease the flow rate from other wells sharing a common flowline.95 7000. in order to construct the gaslift performance curves.10c.00 3194.15 0 1621.15 3519.75 0 270.25 4301.56 22180. these two wells will cause a larger pressure drop through the common flowline than the rest of the wells.09 0 1149.08 7000. For the same amount of oil produced. Well 1 2 3 4 5 6 7 8 9 10 Total * ** Step 3 of the MILPI Method Step 4 of the MILPI Method The SQP Method Liftgas Rate Oil Rate Liftgas Rate Oil Rate Lift Gas Rate Oil Rate (MSCF/d) (STB/d) (MSCF/d) (STB/d) (MSCF/d) (STB/d) 2674.72** 7000.83 5000.66 shutin wells partially closed wells Some comments on the MILPI method and the SQP method are appropriate at this stage: • In Step 2 of the MILPI method.
The convergence histories Formulation of P1 and P2 are plotted in Figure 5.2 Comparison of Formulation P1 and P2 Formulation P2 is easier to implement and is more robust than P1. Formulation P1 optimizes only the liftgas rate. 4. The objective is to maximize the total oil production by allocating liftgas.1. In this example.5. which is zero liftgas rate for every well.5. because some wells (i. Starting with an initial guess of 5000 MSCF/d liftgas rate for every well. the reservoir conditions and the configuration of the gathering system are exactly the same as in the example of Section 5. Formulation P2 turns an unconstrained optimization problem into a constrained one. For this particular example. RATE ALLOCATION THROUGH SQP 84 proportional to the number of wells in the system. the computational time required for the MILPI method and the SQP method were of the same order. Formulation P2 converged to the optimal solution with 21 function evaluations.4. The computational time of the SQP method depends on many factors and is hard to predict. The difference is that the tubing strings and pipes were made shorter in this example so that all wells can produce without gaslift. well 1. The purpose of this example is to show that the performance of Formulation P2 is not necessarily worse than P1 on problems that can be solved by both formulations.e. However. However. and they yielded the same optimal solution. We must ensure that such a reformulation does not make an easy problem hard to solve. and 8) must be performed in the optimization process. Finally we mention that this problem cannot be solved by Formulation P1. Starting with an initial guess of 400 STB/d oil flow rate and 5000 MSCF/d liftgas rate for every well. Formulation P2 converged much faster than P1. Formulation P1 converged to the optimal solution with 67 equivalent function evaluations. well 7) are too weak to flow under certain conditions and well shutdown actions (i. We present two scenarios. Formulation P2 optimizes both the oil flow rate and liftgas rate of every well. 5. .e.CHAPTER 5. Both Formulation P1 and P2 were used to solve this problem. The first scenario is an unconstrained gaslift optimization problem.
40000 35000 30000 P1 25000 20000 15000 10000 5000 0 0 10 20 30 40 50 60 70 P2 OBJECTIVE FUNCTION.CHAPTER 5. STB/D NUMBER OF FUNCTION EVALUATIONS Figure 5. Starting with an initial guess of 5000 MSCF/d liftgas rate and fully open choke status. was partially closed. Formulation P2 converged to the optimal solution with 20 function evaluations. Starting with an initial guess of 400 STB/d oil flow rate and 5000 MSCF/d liftgas rate for every well.000 STB/d. Formulation P2 performed better than P1. and the well choke for well 4. RATE ALLOCATION THROUGH SQP 85 The second scenario is a constrained well rates and gaslift optimization problem. the well with highest water cut. Formulation P2 was much more efficient than P1. However. Formulation P1 converged to the optimal solution with 82 equivalent function evaluations. both formulations yielded the same optimal solution: zero liftgas rate for all wells. The convergence history is plotted in Figure 5. The total water flow rate can not exceed 20.4: Convergence history of Formulation P1 and P2 for Scenario 1. In both scenarios.5. Again. The objective is to maximize the total oil production rate by allocating liftgas and adjusting choke diameters. .
10. RATE ALLOCATION THROUGH SQP 86 40000 OBJECTIVE FUNCTION. The objectives of all problems were to maximize the daily oil production by allocating the well and liftgas rates. The convergence criteria for the optimization process are presented in Table 5. The initial guess for every optimization problem was 400 STB/d oil flow rate and 3000 MSCF/d liftgas rate for every well. and 50 production wells. 5.6 shows the configuration of these gathering systems.5. constrained by total gas and water flow rates. For each gathering system. Constraints can greatly influence problem difficulty. constrained by total water flow rate.3 Efficiency of Formulation P2 This example demonstrates the computational efficiency of Formulation P2 on problems of various sizes and properties. and 28 function evaluations.5: Convergence history of Formulation P1 and P2 for Scenario 2. For example. STB/D 35000 30000 25000 20000 15000 10000 5000 0 0 20 40 60 80 100 P1 P2 NUMBER OF FUNCTION EVALUATIONS Figure 5. Table 5.CHAPTER 5. 50. respectively. Figure 5. We considered three gathering systems with 2. The . problems 7.3.4 shows the number of function evaluations required to solve the optimization problem and the total oil rate of the optimal solution for every problem. We solved the nine optimization problems using Formulation P2. we considered three optimization problems: unconstrained. 8. respectively. and 9 have the same size but require 62.
Separator Pipe Choke Separator Platform Pipes Chokes Tubing string Wells Tubing strings Wells (a) Gathering system with 2 wells (b) Gathering system with 10 wells (c) Gathering system with 50 wells (with a choke installed right before the manifold for every well) Figure 5. 3. All of the problems discussed here can be solved within about 60 function evaluations (one function evaluation is equivalent to one Newton iteration when Eq. Thus.24) by the NewtonRaphson method.CHAPTER 5. . 3. all of the problems discussed here can be solved within approximately 10 times the number of iterations normally required to solve the network problem (defined by Eq.6: Schematic illustration of gathering systems.24 is solved by the NewtonRaphson method). RATE ALLOCATION THROUGH SQP 87 number of function evaluations required is not sensitive to the problem size. we conclude that the SQP method with Formulation P2 is computationally efficient. In other words.
0e6 1.4: Computational efficiency of Formulation P2 on various optimization problems. Table 5.0e6 These are the major convergence criteria used in SNOPT (Gill et al.14 81698. Problem 1 2 Number of Wells 2 2 Constraints None t q w ≤ 1000 STB/d t qw ≤ 1000 STB/d t qg ≤ 20000 MSCF/d Number of Function Total Oil Rate Evaluations (STB/d) 21 1976.3: Convergence criteria for optimization problems of Example 3. The major feasibility tolerance specifies how accurately the nonlinear constraints should be satisfied. The major optimality tolerance is used to judge the optimality of the optimization problem.10 7 8 9 50 50 50 t qw ≤ 35000 STB/d 62 50 28 83423.52 11 1480.38 3 2 14 1267.6 Concluding Remarks We have investigated a new formulation P2 for the optimization problem of allocating well rates and liftgas rates simultaneously subject to multiple flow rate and pressure constraints. 1998).CHAPTER 5.21 4 5 6 10 10 10 None t qw t qw t qg ≤ ≤ 8000 STB/d ≤ 8000 STB/d 50000 MSCF/d None 49 37 45 14884. Major Feasibility Tolerance* Major Optimality Tolerance* Minor Feasibility Tolerance* Minor Optimality Tolerance* * 5. The minor optimality tolerance is used to judge the optimality of each QP subproblem.92 14333. The minor feasibility tolerance specifies how accurately the bound and linear constraints should be satisfied. This is due to the fact that the MILPI method optimizes the well rates and liftgas rates based on gaslift performance curves and ignores the deliverability ..96 14585.21 81067.22 t qw ≤ 35000 STB/d t qg ≤ 300000 MSCF/d 5. The optimization problem is solved by a SQP algorithm.0e4 1.0e3 5. RATE ALLOCATION THROUGH SQP 88 Table 5. Results demonstrated that the SQP method performs much better on certain problems than the MILPI method.
sometimes a well can have a zero oil rate and a positive liftgas rate. it is easier to handle well shutdown actions and avoid other computational difficulties. It is also necessary to investigate the performance of the linear optimization methods and the nonlinear optimization method under different conditions.CHAPTER 5. RATE ALLOCATION THROUGH SQP 89 constraint of the gathering system in the optimization process.10c is satisfied. which is not realistic. We determined that Formulation P2 is appropriate for the simultaneous optimization of well rates and liftgas rates. because with this formulation. This positive liftgas rate is used to lighten the oil column so that constraint Eq. . while the SQP method makes no such simplifications. Fortunately. 5. such as pumps and subsurface chokes for multilateral and multisegment wells. with P2 the optimization problem can be solved efficiently. our experience shows that the value of this liftgas rate usually is small and has little impact on the solution of the optimization problem. More research work could be done on this subject. Furthermore. We can modify Formulation P2 to allow loops in the network. We can incorporate more decision variables into the optimization problem. However there are some limitations of Formulation P2. In the optimal solution of P2.
The global problem is a nonlinearly constrained optimization problem with both continuous (production and lift gas rates) and integer (well connections) variables. and well connections for the Prudhoe Bay oil field in Alaska. To achieve the best results. the optimization problem addressed in this chapter is referred to as the global problem. However. (1997) presented a heuristic optimization procedure to allocate the production rates. lift gas rates. a well or a flow line can have several potential output connections that join that well or flowline to different flowlines and facilities. Redirecting the well connections is an effective way to debottleneck the production system. The problem of well connection optimization is to identify the best set of well connections that maximizes an operational objective.Chapter 6 Optimization of Well Connections 6. well connections often need to be optimized simultaneously with production rates and lift gas rates (and potentially other production operations). the production optimization problem addressed in this section involves maximizing a certain operational objective (such as to maximize the daily oil rate) by optimally allocating the production and lift gas rates and well connections subject to multiple flow rate and pressure constraints. Litvak et al. No existing formal optimization methods can be easily adapted to solve such a complicated problem. To differentiate this problem from the rate allocation problem addressed in previous chapters. the performance of that procedure can be unsatisfactory because (1) 90 . Thus.1 Introduction In some petroleum fields.
1a) (6. z ) are the constraint functions for the rate allocation problem parameterized . f u (x ) is defined as the optimal solution of a rate allocation problem f u (x ) = maximizey. The global problem is formulated as a twolevel optimization problem: maximizex. y denote lift gas rates. or gaslift rates) are too simple for a complex system like the Prudhoe Bay field. The performances of the solution methods were investigated and will be illustrated here by an example. Two solution methods are discussed.2 TwoLevel Optimization Formulation Although the physical decision variables for the global problem are the well connections. z f l (x. The global problem is formulated in such a way that existing methods for rate allocation can be utilized for the global problem. Because the rate allocation problem is a part of the global problem. Let x denote a set of well connections. F(x.2a) (6. and z denote production rates. y.CHAPTER 6. OPTIMIZATION OF WELL CONNECTIONS 91 the production rates. y. z ) subject to l ≤ F(x. z ) is the objective function for the rate allocation problem parameterized by x . 6. lift gas rates.2b) where f l (x. In this chapter. y. and well connections are allocated sequentially rather than simultaneously. lift gas rates. and well chokes. production rates. a new formulation of the optimization problem is presented. production rates instead of well chokes are selected as decision variables for the global problem.1b) subject to x ∈ Ω wc where Ω wc is the domain of feasible well connections. z f u (x ) (6. y. The rate allocation problem has been described extensively in Chapter 4 and 5. (2) the multiple flow rate constraints are checked and satisfied sequentially rather than simultaneously. and (3) the allocation procedures (either for well connections. it was shown in Chapter 5 that optimizing on well chokes is computationally inconvenient for the rate allocation problem. y. z ) ≤ u (6.
1 is named the upper level problem. The upper level problem is an unconstrained integer optimization problem. The lower level problem has been described extensively in Chapters 4 and 5.3. 6. A combination of these well connections represents 320 or 3486784401 distinctive configurations of the production system. This formulation separates the search on well connections and the search on production and lift gas rates into two levels. However. OPTIMIZATION OF WELL CONNECTIONS 92 by x . f l and F are specified by the global problem.CHAPTER 6. which is not practical. For example. In principle. Suppose a computer can evaluate one configuration per second. then f l is the total field oil rate and F is the total field gas rate. The lower level problem can be regarded as a function evaluation procedure for the upper level problem. Problem 6. The upper level problem is also referred to as the well connection optimization problem in this chapter. it would take about 110 years to evaluate all the possible configurations. and Problem 6. consider a production system with 20 wells. The lower level problem is a nonlinearly constrained rate allocation problem. In this formulation. respectively.1 Computational Issues The upper level problem is a nonlinear integer programming (NIP) problem. if the global problem is to maximize the total field oil rate subject to a total field gas rate constraint. complete enumeration is computationally impractical even for well connection optimization problems with a moderate number of connections. For example. . then describes the investigation of some possible solution methods.3 Solution Methods for the Optimization of Well Connections This section first motivates the choice of solution methods for the upper level problem. l and u are the lower and upper bounds of the constraints. each well has three potential output connections.2 is named the lower level problem. a NIP problem can be solved by enumeration. This chapter focuses on optimization methods for the upper level problem – optimization of well connections. 6.
there does not exist a universal algorithm for integer programming: existing algorithms either require huge computational effort. This process is repeated until no further significant improvement can be made.) 3c) Repeat Step 3a)3c) for all potential output connections of well j . Algorithm 6. The optimal objective value from the rate allocation problem is regarded as the merit or value of the new output connection of well j . according to some operational rules. The basic steps of this method can be illustrated as follows. A genetic algorithm was also used. OPTIMIZATION OF WELL CONNECTIONS 93 So far. For well j . or they only give approximate solutions (Bertsimas and Tsitsiklis. The genetic algorithm has the ability to avoid local optima and it is easy to implement. The genetic algorithm can be used to crosscheck the performance of the partial enumeration method. 3. Select a production well. The method reconnects a well to its best output connection while keeping other well connections fixed. Define the output connection with the maximum merit/value as the best output connection of well j . perform the following operations: 3a) Switch well j from its current output connection to another potential output connection. 3b) Invoke optimization tools as described in Chapter 4 or Chapter 5 to reallocate the production and lift gas rate subject to imposed constraints. This step generates a new configuration of the production system. 1. well j . 2.3.1: The partial enumeration method. Set iteration index k = 0 .CHAPTER 6. 3d) Switch well j to its best output connection as determined in step 3c).2 Partial Enumeration Method The partial enumeration (PE) method is a heuristic method designed specially for the optimization of well connections. Because efficiency is an important consideration for realtime production optimization. (This step solves the lower level problem. 6. . 1997). we developed a partial enumeration method that gives only an approximate solution but is efficient for the upper level problem.
4. and well developed optimization method.1 can limit the maximum number of well reconnections so that the configuration of the production system does not change dramatically from day to day. The method attempts to find a better system configuration at each search step. etc. For example. no well changes its connection or no discernable increase of the objective function value is achieved. Increase the iteration index k by 1. Fieldspecific operational rules can be incorporated easily. crossover. the actual implementation of Algorithm 6.3 Genetic Algorithm The PE method is a heuristic method that examines only a portion of the complete search space.CHAPTER 6. This may cause some concerns for the quality of the solutions obtained from such a method. Genetic algorithm (GA) is a general.3. robust. 94 6. the chromosome with the best fitness represents the optimal solution of the optimization problem. We used this method to crosscheck the performance of the PE method. When the evolution terminates. Different rate allocation algorithms can be employed in this method. 6. A genetic algorithm is a stochastic optimization algorithm based on natural selection. The performance of the PE method was evaluated and is illustrated in an example presented in Section 6. Repeat Step 1 to 5 until certain user defined convergence criteria are met. 5.1. . A flowchart of a simple GA is illustrated in Figure 6. A GA is applied to the well connection optimization problem in the following manner. at each search step only one well is allowed to change its connection. Repeat steps 2 and 3 for all production wells in the field. Multiple chromosomes form a population. decision variables are encoded as a chromosome. The PE method is flexible. The PE method mimics what a field operator facing the same problem would do. In GA. In addition. and mutation. A GA evolves a population by means of selection. but it does not guarantee the global optimum. OPTIMIZATION OF WELL CONNECTIONS 4. One such criterion is: between iteration k − 1 and k .
2). etc. This study used the following encoding/decoding scheme for well connections.2 l . The length of a gene is the minimum number of bits required to cover all potential connections of a well: l = min m {2 m ≥ nc } (6. a set of decision variables is usually. The decoding process maps a gene to a well connection by converting its binary expression to a natural number. 01. If k ∈ n c . k .3) For example. nc − 1] .CHAPTER 6. the gene is mapped to the well connection labeled as k . • Decoding. though not necessarily. with each single variable corresponding to a segment of the binary string. [0. nc − 1] . A specific well connection is encoded into a gene by expressing its labeling number in binary format. encoded as a binary string. OPTIMIZATION OF WELL CONNECTIONS 95 Create an initial population Decode and evaluate current population Produce next generation by selection. the binary string is called a chromosome. Then its connections can be encoded into 00. Suppose a well has nc potential (output) connections and they are labeled by a set of consecutive numbers. If k ∈ [0. the gene is mapped to the well [ ] .1: A simple GA flowchart. In GA. 10 respectively (Figure 6. mutation. With such an encoding scheme. In GA vocabulary. No Meet stopping criteria? Yes Return best value and stop Figure 6. and the segment of the binary string representing a variable is named a gene. Encoding/decoding. decision variables take discrete values. Then the encoding and decoding process for this well goes as follows: • Encoding. crossover. suppose a well has three potential connections.
OPTIMIZATION OF WELL CONNECTIONS 96 connection labeled as k − ( nc − 1) . mutation. well connection labeling 0 1 2 encoding 00 01 10 11 00 01 10 0 1 2 decoding 11 gene Figure 6. The current generation of the population is modified through selection. the gene expressions 00. in which a mating chromosome is selected as follows. For each chromosome. and other operations to produce the next generation of population. One chromosome encodes the current set of well connections to ensure a good starting point. a rate allocation problem is solved and the optimal objective function value is assigned as the fitness of the corresponding chromosome. Evaluate the population.2. and is an iterative procedure with each iteration containing the following steps: 1. The rest of the chromosomes are randomly generated to ensure diversity. respectively. The genes for all decision variables comprise a chromosome. This step generates the initial population (multiple sets of solutions or chromosomes) that GA begin to evolve with. Evolving the population. Produce next generation. and the one with higher fitness value is selected as the mating chromosome. This decoding scheme is illustrated in Figure 6. 1. This is the major part of a GA. Initializing a population. Randomly pick two chromosomes from the current population. The number of chromosomes in the population is denoted as n . 2. and 2. Replicate chromosomes for mating. 01.CHAPTER 6.2: Encoding/decoding a well connection. . 2a) Selection. Decode each chromosome and evaluate its fitness (function value). Consider the above encoding example. and 10 can be mapped to well connections labeled 0. Repeat this procedure until n mating chromosomes are selected. crossover. This study implements a tournament selection. A gene expression of 11 can be mapped to a well connection labeled 0.
CHAPTER 6. 2c) Mutation.5. in which the two chromosomes swap their bits at a randomly selected position with a probability pc . A mutation example is shown in Figure 6. 1111111 mutation 1011011 Figure 6. in which the two chromosomes exchange their bits at every position based on a probability pc . OPTIMIZATION OF WELL CONNECTIONS 97 2b) Crossover.3: Schematic illustration of singlepoint crossover. In this step each bit in a chromosome flips according to some probability pm . One type of crossover is called singlepoint crossover.4: Schematic illustration of uniform crossover. Crossover the two chromosomes (the parent) to produce two new chromosomes (the children). Randomly pick two chromosomes from the replicated population.3 shows an example of singlepoint swap. The new generation copies the best chromosome from the converged population and generates the rest of the chromosomes randomly. There are different ways to crossover two chromosomes. 2d) There are other techniques that can enhance the performance of the GA. 1111111 0000000 1111000 0000111 crossover Figure 6. and (2) if there is not enough diversity in a population (a population is converged because all chromosomes are too similar). Mutation is another way to increase the diversity of the population. Figure 6. This study implements both crossover schemes. 1111111 0000000 1011001 0100110 crossover Figure 6. . This study used the following two: (1) the best parent is always copied to the next generation (elitism). Another type of crossover is called uniform crossover.5: Schematic illustration of mutation. Figure 6.4 shows an example of uniform crossover. start a new generation.
Both the PE method and the GA were used to solve the upper level problem. The upper level optimizes the well connections.. Oil production from Prudhoe Bay is on decline and is constrained by the gas and water handling limits of surface facilities and velocity constraints in flow lines. highly automated and complex surface pipeline network. Production from many wells can be redirected to different flowlines and facilities (Litvak et al. 2002). the production optimization problem for the Prudhoe Bay field is to maximize the daily oil production by optimally allocating the well connections and production rates subject to multiple flow rate and velocity constraints.3. Therefore. optimization of well connections and production rates is a viable and economic way to increase the daily oil production in the Prudhoe Bay field. The lower level problem was solved by the LPII method described in Section 4. On the other hand. For these reasons. OPTIMIZATION OF WELL CONNECTIONS 98 3. Roughly 100 wells are considered for reconnection. The stopping criterion used in this study was a specified maximum number of generations.4. Currently there are 23 well pads in the Western Operating Area (WOA) and 16 drill sites in the Eastern Operating Area (EOA). 6. each well has two to three potential connections. 1997). Repeat Step 1 and 2 until stopping criteria is met. Other wells’ connections remain fixed.CHAPTER 6. The field model of Western Operating Area contains about 500 hundred wells.. Prudhoe Bay has a flexible. Prudhoe Bay oil field is the largest oil field in north America. The lower level problem optimizes the production rates subject to all constraints.4 Comparison of the PE Method and the GA We analyzed the performance of the PE method and the genetic algorithm on a model of the Western Operating Area of Prudhoe Bay oil field in Alaska. The production optimization problem was formulated as a twolevel optimization problem. The genetic algorithm used in this study was based on a genetic algorithm driver developed by Carroll (2000). Each well pad or drill site contains about 3040 wells (Litvak et al. .
It is observed from Table 6.9202 0.6 plots the convergence histories of the first three runs. Starting Normalized Total Oil Rate Final Normalized Total Oil Rate Number of Lower Level Problems Solved PE1 0. However.9352 0. After 1000 generations or 10.9935 636 PE3 0. the best objective values are close to each other. The objective value.9070 0. In this way.CHAPTER 6.9921 638 PE5 0.9922. pm = 1 / n = 0. The line labeled PE BEST represents the best objective function value obtained from the PE method. Figure 6. the best objective value obtained by the GA method was a normalized oil rate of 0.6 represents the best objective found by these runs. is normalized by a selected value. Figure 6. OPTIMIZATION OF WELL CONNECTIONS 99 The PE method is a deterministic heuristic technique.1 that the final objective values from these runs were not sensitive to the starting point. and mutation probability. The line labeled PE BEST in Figure 6. the following basic GA parameters were used: the number of chromosomes. the starting points for both the GA and the PE method were generated randomly and the comparison between the GA and the PE method is therefore more meaningful. all initial chromosomes were generated randomly. n .5.8852 0. uniform crossover probability. then slowed down at later stages.1 .1 shows the starting and final objective values of all runs. The objective function was normalized in the same way as in the PE method.9928 638 For the GA run. Each run started the PE method from a randomly generated set of well connections. which is the total oil rate. The early convergence history of .7 plots the convergence history of the GA method. In this run.9922 636 PE4 0. pc .8911 0. These runs found different sets of solutions after 400 function evaluations (one function evaluation refers to solving an instance of the lower level problem).000 function evaluations (a function evaluation is defined in the same way as defined for the PE method).0013 of normalized oil rate. Table 6. The biggest difference is 0. Figure 6. of a population was 10.9935 477 PE2 0.6 shows that the PE method increased the oil rate quickly at early stages. we made five test runs. To see if the method was sensitive to its starting point. was 0. the current set of well connections were not encoded as one of the initial chromosomes.1: Performance data of the PE method. Table 6.
which changes at most one well connection at each step.98 NORMALIZED OIL RATE 0. The configuration of a real production system that has been operated for a long period is already close to optimum. The GA converged to a value no larger than that from the PE method and it used 25 times more function evaluations than the PE method.88 0 100 200 300 400 NUMBER OF FUNCTION EVALAUTIONS PE1 PE2 PE3 GENETIC ALG PE BEST Figure 6.CHAPTER 6. • The PE method is robust because it was not sensitive to its starting points and it achieved a better objective value than the GA. The good performance of the PE method can be explained partially as follows. OPTIMIZATION OF WELL CONNECTIONS 100 the GA method is also plotted in Figure 6.96 0.6.92 0. .7: • The GA was not as efficient as the PE method.6: Convergence history of the PE runs.6 and Figure 6. is suitable for this purpose. The following observations are made from Figure 6.94 0. The PE method. 1 0. Such a system needs only fine adjustment in response to the changing operational conditions.9 0.
The GA can.95 0.CHAPTER 6. For moderate or largescale systems. and it allows the use of existing or new optimization methods separately for each level. the performance of the PE method has to be examined carefully before it is applied to a new field. if the number of wells in a production system is small (say. This advantage is significant for the production optimization problem addressed in this chapter because no existing optimization algorithms can solve the entire problem efficiently. The solution quality of this approach depends on the properties of the global problem and which optimization methods are used for the two levels.5 Concluding Remarks The advantage of the twolevel optimization is flexibility.94 0 2000 4000 6000 8000 10000 NUMBER OF FUNCTION EVALUATIONS Figure 6.97 0. However. 6. be used to test the robustness of the PE method. OPTIMIZATION OF WELL CONNECTIONS 101 1 0.99 NORMALIZED OIL RATE 0. The approach transforms the entire problem to an upper level problem and a lower level problem. approximate methods have to be used. as a specially designed heuristic method. . but it is generally more expensive.96 0. Genetic algorithm provides an alternative solution method. less than 20).7: Convergence history of the GA run. The partial enumeration method proved to be both robust and efficient for several fields. however.98 GENETIC ALG PE BEST 0. For the upper level problem. complete enumeration is possible to find the best set of well connections.
.1a) (7. The problem of optimizing multiple and distinct objectives is named multiobjective optimization problem (MOP) (Miettinen. Typical goals include maximizing the daily oil production. Mathematically. petroleum production engineers often strive to achieve more than one goal when operating a field.. a MOP can be expressed as minimize F(x ) = [ f 1 (x ).. and minimizing safety and environmental hazards.1 Concept of Multiobjective Optimization The optimization problems addressed in previous chapters assume a single objective function. it is unlikely that different objectives can be optimized by the same set of decision variables. The space to which the objective vector F(x ) belongs is called the 102 . f 2 (x ) . x is an n dimensional vector of decision variables. m} (7.. f k (x ) are the k objective functions or attributes..Chapter 7 Multiobjective Optimization of Production Operations 7.. minimizing the production cost. S denotes the feasible set of x .1b) where f 1 (x ) . f 2 (x). In most cases.. i = 1. 1999)..... f k (x )] subject to x ∈ S = {x : ci (x ) ≤ 0. and c i (x ) denotes a constraint.. In practice.
3) A point x * ∈ S is said to be Pareto optimal for a MOP if and only if there is no x ∈ S is dominant over x * . no solution is optimal for all the k objectives simultaneously. 1999). The solutions in the Pareto optimal set are not dominant over each other.. Points A and B are in the Pareto optimal set.1). The problem is to minimize both f 1 and f 2 . All the Pareto optimal solutions form the Pareto Optimal set. f2 D attained set Y A C Pareto optimal set B f1 Figure 7.2. k } : f i (x ) ≤ f i (y ) and ∃j ∈ {1.1 and two solution vectors x and y ...CHAPTER 7. . Moving from one point to another point within the set will deteriorate at least one objective or attribute.1: Graphical illustration of a Pareto optimal set. points C and D are not. One such replacement is the notion of Pareto optimality (Miettinen. In most cases. the Pareto optimal set is on the boundary of the feasible set Y (see Figure 7. The feasible set of F(x ) is called the attained set. which is denoted by Y = {F(x ) : x ∈ S} (7. 2001): ∀i ∈ {1. The elliptic area is the attained set of the problem.2) In multiobjective optimization. x is dominant over y if (Andersson. The solid line is the Pareto optimal set. Consider Problem 7.. k } : f j (x ) < f j (y ) (7.2... MULTIOBJECTIVE OPTIMIZATION 103 objective space or criterion space.. Thus the optimality criterion used for single objective optimization must be replaced by a new one..
Table 7. . According to Hwang et al.2 Solution Techniques for Multiobjective Optimization Problems The solutions in a Pareto optimal set are not dominant over each other. subjective judgement or preference information from the decision maker is required. 1976) Lexicographic Method (Fishburn. 1971) Interactive MOLP (Steuer. The materials of Section 7.1: A classification of methods for multiobjective optimization (after Huang et al.. 2002) Method of ZiontsWallenius (Wallenius. NoPreference Methods Global Criterion Method (Salukvadze. the solution methods for MOP can be classified into four categories depending on when the preference information of the decision maker is articulated in the process of multiobjective decision making: (1) nopreference methods. and (4) a posteriori methods.2. 1975) STEM method (Benayoun et al.17. there is no solution that is better than x in all attributes. 1980). it is rational to choose the final solution for a MOP from its Pareto optimal set. 1974) Goal Programming (Charnes and Cooper. If a point x is in the Pareto optimal set.2. Table 7.CHAPTER 7. Methods investigated in this study are described in detail where appropriate. In order to determine the final solution for a MOP. (2) a priori methods.1 lists a few major methods for each category. there always exists a point y ∈ S that is dominant over x .4 gives a brief description of each category. 1972) ε constraint method (Haimes et al. (1980). MULTIOBJECTIVE OPTIMIZATION 104 If a point x is not in the Pareto optimal set.. 1977) Weighting Functions Method (Gal and Nedoma. (1980) and Azarm (2002). 1975) A Priori Methods Interactive Methods A Posteriori Methods Section 7.17.4 are based on Hwang et al. The preference information is used to convert a MOP to one or a series of scalar optimization problems.2.. 7.2. 1961) Hierarchical Method (Azarm. (3) interactive methods. Therefore. 1974) Utility Functions (Keeney and Raiffa.
1 NoPreference Methods Nopreference methods require no subjective preference information from the decision maker.4) where f i * is the ideal value of the i th objective and is usually obtained by minimizing f i individually while ignoring the rest objectives.CHAPTER 7. f 3 (x )) (7. For example.2. the objective vector is aggregated into a scalar objective function in such a way that the deviation of the objective vector from an ideal objective vector is minimized: f (x ) − f i * f (x) = ∑ i f i* i =1 k p (7. the most satisfactory solution of the MOP can be obtained. MULTIOBJECTIVE OPTIMIZATION 105 7. Boychuk and Ovchinnikov (1973) suggests p = 1 while Salukvadze (1974) recommends p = 2 . The advantage of this approach is that the decision maker is not disturbed in the solution process and the method is easy to use. The disadvantage is that the decision maker has no control over the solution quality. However. 7. Multiple objectives are aggregated into one objective function without any opinions of the decision maker.2 A Priori Methods A priori methods require that the preference information from the decision maker is given before the multiobjective optimization is conducted.2.. even for a simple problem. in the method of Global Criterion (Salukvadze. a utility function is used to express the decision maker’s preference information over the objectives: U (F (x)) = U ( f 1 (x). The advantage of the utility function methods is that once the utility function is appropriately formulated. the utility function is very difficult to construct. In these methods.5) where U (F ) is a utility function of the multiple objectives. f 2 (x ). Utility function methods. Thus the decision maker must have some a priori understanding of the optimization problem. Then a scalar optimization problem is solved and the solution is presented to the decision maker who may accept or reject the solution. 1974).. ...
Pl hl (d − . d + is the positive deviation from the j th goal. j 3. d + ). Minimize the j th achievement function h j while requiring hi ≤ hi* for i = 1... The achievement functions are absolutely ordered. i = 1. Problem 7.... when minimizing an achievement function. ( ) . Denote the optimal value of the j th achievement function as h * . Let j = 1 . j − 1 . and Pj is the weighting coefficients for the j th achievement function.2. Goal Programming. d + ) subject to c j (x ) + d − − d + = b j . the decision maker must construct a set of goals and rank the importance of the goals.6) Methods using utility function given by Eq.7b) (7.7e) where b j is the aspiration level of the j th goal. j = 1. d i+ ≥ 0 for all i d i− • d i+ = 0 for all i [ ] (7.6 are also called the weighted sum methods.7d) (7... and let min h1 = h1* . Repeat Step 2 until all achievement functions are minimized. That is.. m. 7... Goal Programming (GP) was first developed by Charnes and Cooper (1961). j j f i (x ) + d i− − d i+ = bi .. achievement functions with higher rank can not be deteriorated. The solution procedure for GP goes as follows: 1. h j d − . Let j = j + 1 .. Minimize the first achievement function h1 .1 is formulated in GP as follows: minimize P1 h1 (d − .. For this method... P2 h2 (d − .7c) (7. 2.CHAPTER 7. which means Pj >> Pj +1 and there does not exist a number N that makes NPj +1 greater than Pj .. d + ). 2. d + is a function of the j deviation values called the j th achievement function. k d i− .7a) (7. d − is the negative deviation from the j j th goal. MULTIOBJECTIVE OPTIMIZATION 106 A special form of the utility function is to use weighting coefficients to indicate the importance of each objective: U (F(x )) = ∑ wi f i (x ) i =1 k (7.
i ) .. ( ) 7.9b) (7.9a) (7. i = 1. subject to the original set of constraints minimize f 1 (x ) subject to c i (x ) ≤ 0 . i = 1. the decision maker learns about the problem and guides the solution process according to what has been learned. the decision maker progressively articulates preference information by interacting with the solution process. . Let j = 1 . The disadvantage of this method is that a great deal of effort is required from the decision maker.1 for f 1 .8b) Let f 1* (x*. m f i (x ) ≤ (1 + εi ) f i * (x*. 2.. Let j = j + 1 . 3. Denote the optimal objective value of Problem 7.CHAPTER 7. As the solution process goes. the solution procedure can be illustrated as follows: 1.9c) where εi ≥ 0 is the coefficient of function increment for f i .. The solution process is a learning procedure for the decision maker and the decision maker is likely to accept the solution....9 as f j* x *. j − 1 (7. The advantage of these methods is that no a priori information is required from the decision maker. 2002) allows the decision maker to rank and minimize the objectives in descending order of importance.. j ..8.. The decision maker controls the solution process by updating his preference as he learns about the problem. m (7. This method (Azarm. Find the optimum point x *..2.3 Interactive Methods For interactive methods.1 ) denote the optimal objective function value for Problem 7.. Find the optimum point x *. MULTIOBJECTIVE OPTIMIZATION 107 Hierarchical optimization method.. j for the j th objective function f j subject to the original and an additional set of constraints minimize f j (x) subject to c i (x ) ≤ 0 . i = 1.8a) (7. Suppose the k objectives are ordered from f 1 (most important) to f k (least important). Repeat Step 2 until j reaches the total number of objectives k .
see the methods of ZiontsWallenius and the STEM methods reviewed in Huang et al. the common theme of this class of methods is to vary systematically the algorithm parameters expressing the decision maker’s preference information. To construct a discrete sampling of the Pareto optimal set. The weighting functions method.2. The weighting functions method (Gal and Nedoma. MULTIOBJECTIVE OPTIMIZATION 108 Interactive methods often have a calculation phase and a decision phase.2.4 A Posteriori Methods A posteriori methods require no preference information from the decision maker before and during the solution process.2 is able to generate an evenspread sampling of the Pareto optimal set by varying parameter εi . especially when there are more than two objective functions. (1980). the computational load of this method is heavy. 1975) chooses one objective for minimization and reformulates other objectives to constraints whose righthand side is a usercontrolled variable ε . Solving the resulting scalar optimization problems for various ε will generate a set of Pareto set of optimal solutions. However. Another disadvantage is that the decision maker may have too many solutions to choose from. The methods construct a subset of Pareto optimal from which the decision maker makes a subjective judgement and selects the most satisfactory solution. For example. 2002) described in Section 7.CHAPTER 7. and the decision phase uses the tradeoff information to guide the next search step. 1993) try to evolve a set of solutions in such a way that the set of solutions spread to the Pareto optimal front. the ε constraint method (Haimes. The calculation phase gathers the tradeoff information. 1972) samples the Pareto optimal set by changing the weighting coefficients of the weighted sum method: . 7. The Hierarchical method (Azarm. Evolutionary multiobjective optimization algorithms (Fonseca and Fleming. For example. The advantage of this method is that the decision maker usually has a better understanding of the tradeoffs because he/she has a discrete sampling of the Pareto optimal set.
The line tangent to point A can be moved further to points B and C.10a) (7. (1980) and Azarm et al. MULTIOBJECTIVE OPTIMIZATION max f (x ) = ∑ wi f i (x ) i =1 k 109 (7. some points in the set can not be allocated by the weighting function method no matter what weights are used. (2002)). Thus point A can not be allocated by the weighting function method.CHAPTER 7.10b) subject to x ∈ S The weighting functions method can be interpreted geometrically. f2 Y L slope = − w1 / w2 L f1 Figure 7. . The weighting function f = w1 f 1 + w2 f 2 defines a line L with slope − w1 / w2 in the objective space (Figure 7. Minimization of function f can be seen as moving line L towards the origin while keeping the intersection of line L and the feasible region Y .2).3.2: The weighting functions method with convex set (based on Huang et al. If the Pareto optimal set is not convex. This is illustrated in Figure 7. The optimum point is obtained when line L is tangent to the boundary of the feasible region. Consider a MOP problem with two objectives f 1 and f 2 .
(2002)). MULTIOBJECTIVE OPTIMIZATION f2 Y 110 B slope = − w1 / w2 A C L L f1 Figure 7. . The soft constraints need to be formulated in mathematical language so that they can be handled in the optimization process.3 Multiobjective Optimization of Production Operations Production operations in a petroleum field aim to maximize profit from the field subject to restrictions of facility capacities. For daytoday production operations. this task breaks down to the following components: • Objectives. Soft constraints refer to constraints that are not well defined and/or can be violated.3: The weighting functions method with nonconvex set (based on Huang et al. Hard constraints refer to constraints that are well defined and cannot be violated. such as the maximum flow rate or pressure limit on individual wells or the target of the total oil rate from the field. Soft constraints can be viewed as objectives. the objectives typically include maximizing daily oil production rate to meet contract target. • • Hard constraints. safety and environmental regulations. 7. In optimization terminology.CHAPTER 7. and minimizing lift gas rate to minimize production cost. Soft constraints. such as the fluid processing capacities of surface facilities. etc. commercial contract. and other considerations. (1980) and Azarm et al.
. which we will call Problem 7.CHAPTER 7..2 Solution Methods This section presents a framework to formulate and solve the multiobjective production optimization problems.12b) (7. It is necessary to specify which wells’ lift gas rates are to be optimized. MULTIOBJECTIVE OPTIMIZATION 111 7. Decision variables. A simple example of h is the weighted sum of the deviation variables h (d ) = ∑ (wi d i ) i (7.12a) (7.. .11) where c i (x ) is a soft constraint function and bi is a limit value.. The lower and upper bounds and the initial values for all decision variables are specified in advance.13) where wi is a usersupplied weighting coefficient.. i = 1..12c) where d i is a deviation variable and h is a function of the deviation variables. 1961). Soft constraints with the following forms can be treated in a similar manner c i (x ) ≥ bi or c i (x ) = bi (7.. m (7.3. the soft constraints were handled in a way similar to how the goals are handled in goal programming (Charnes and Cooper.. Constraint Eq.12: minimize h (d ) subject to ci ( x ) − d i ≤ bi . i = 1. m (7. The oil flow rate of every production well and deviation variables for soft constraints are automatically selected as decision variables. 7. The soft constraints can be regarded as one type of goals and are converted to a set of hard constraints and an objective function by introducing some deviation variables. Consider the following set of soft constraints c i (x ) ≤ bi .14) 7.1 Handling Soft Constraints In this study.. i = 1. The actual imple mentation is limited to the nonlinear rate allocation problem described in Chapter 5.11 can be converted to the following optimization problem.3. m d i ≥ 0 .
15) where f j denotes the j th objective. The problem may include multiple objectives. Multiple flow rate constraints can be specified for any well or network node.2. an objective variable can be oil.13 (this objective function is considered as an objective variable below). or a function of some deviation variables defined by Eq. and the method reduces to the weighting functions method if all objective variables are . Objectives.CHAPTER 7. An objective variable can be a single decision variable or a function of multiple decision variables. The primary solution method investigated in this study was the Hierarchical optimization method (Azarm. and water flow rates. Solution method. the method reduces to the Lexicographic method (Fishburn. yi denotes an objective variable. The method is able to locate a Pareto optimal solution that resides in the nonconvex part of the Pareto optimal set. The proposed framework is flexible. 1974) if all εi are set to zero.2. 7. water flow rate) constraints can be specified for any well or network node. 2. The Hierarchical method was adopted because 1. 1961) if all the objectives are defined as soft constraints and all εi are set to zero. and wi denotes a weighting coefficient. Each subset can be converted to a set of hard constraints and an objective function defined by Eq. For example. Soft constraint. gas. Multiple soft (oil.13. For example. 2002) described in Section 7. 7. MULTIOBJECTIVE OPTIMIZATION 112 Hard constraints. The whole set of soft constraints can be divided into several subsets. the user needs to rank the k objectives in descending order of importance and specify a relaxation parameter εi for each of the first k − 1 objectives. the method reduces to the goal programming method (Charnes and Cooper. Each objective is a weighted sum of a set of objective variables: f j (y ) = ∑ wi y i i (7. The method is easy to use and its logic is easy to understand for the decision maker. The Hierarchical method can be converted to other popular multiobjective methods by manipulating the algorithmic parameters. or the pressure of a well or a network node. To use this method. gas.
4 Example The production system is the same as the one presented in Section 3.000 STB/D.4. Since the field is able to produce at an oil flow rate higher than 14. The cost of each barrel of produced water and cubic feet of produced gas depends on many factors and may be hard to evaluate quantitatively. The solution procedure of the hierarchical method goes as follows: 1. Therefore it is difficult to come up with utility functions to aggregate the last two objectives into one objective function. The configuration of the gathering system is illustrated in Figure 3. an interactive method. An appropriate solution is to construct a subset of the Pareto optimal set for this problem from which the production engineer can choose the most satisfactory solution. Produce the field with a total oil production rate qo . The Hierarchical method can be used as an a priori method.000 STB/D. or an a posteriori method depending on how the adjustable algorithmic parameters (weighting coefficients wi and relaxation parameters εi ) are adjusted in the optimization process (no adjustment. Suppose that the produced gas is not for sale and it is to be reinjected into the reservoir for storage. Let us assume that the production engineer wants to operate the field with the following goals: 1. The other two goals are formulated as two objectives. Both the hierarchical method and the weighting functions method were used to construct the Pareto optimal set. adjusted interactively by the decision maker. 7. respectively). 2. Minimize the total water flow rate q tw .t of exactly 14.15. the first goal is to formulate the total oil production target as a hard constraint.CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION 113 aggregated into one objective function using Eq. t 3. 7. Minimize the total water flow rate of the field with the following constraint: .7. Minimize the total gas flow rate qg . or adjusted by the analyst to generate a sampling of the Pareto optimal set.
wg ) was not used to indicate cost of the produced water or gas.16) Denote the minimum water flow rate value as q tw. Thus. In the weighting functions method.CHAPTER 7. Table 7. the production engineers can look at the tradeoffs in the sampled set and make their operational decisions according to . Once the complete set of Pareto optimal set is sampled.17) (7. Repeat Step 3 for j = 2.* .* 3.000 (7.1) 10310 26419 (3..19a) (7.2: Pareto optimal solutions from the weighting functions method.1) 8709 31058 (5.19b) Here the weighting pair (ww . Let j = 1 .4).5.18) q tw ≤ (1 + j * 0.1)qtw. until the total gas flow rate no longer decreases as j increases.1) 10074 26829 (3. However..2. it is much easier to control the sampling point with the hierarchical method than with the weighting functions m ethod. because for the weighting functions method.1) 6904 55404 (20. Weighting Pair Total Water Rate (STB/d) Total Gas Rate (MSCF/d) (1.5. the multiobjective optimization problem was transformed to the following scalar optimization problem minimize f = ww q w.1) 7246 50192 (15. Minimize the total gas flow rate of the field subject to following constraints: t qo = 14.4). t t subject to qo = 14.1) 8583 31834 (15.3. The weighting pairs and the corresponding solutions generated for this problem are shown in Table 7. but was varied to generate a subset of the Pareto optimal solutions. such as the curve segment roughly between 9000 STB/D and 10.000 STB/D (7. MULTIOBJECTIVE OPTIMIZATION t qo = 14. there is no clear relation between the weighting pair and the Pareto optimal solution. the hierarchical method is more appropriate for this problem than the weighting functions method.. the weighting functions method is unable to locate the nonconvex parts of the Pareto optimal set.1) 6354 63923 Both methods obtained roughly the same Pareto optimal set (Figure 7. Furthermore. t + wg q g .000 STB/D of water flow rate (Figure 7. 2.1) 8653 31295 (10.000 STB/D 114 (7.
the net present value (NPV)..2. 70000 TOTAL GAS RATE.4: The Pareto optimal sets obtained from the Hierarchical method and the weighting functions method.2 or the weighting functions method described in Section 7.4. and is able to sample the complete Pareto optimal set. However. However.CHAPTER 7. as demonstrated in Section 7. sometimes. 7. on the other hand. does not require a utility function. i. The hierarchical method.4 is often used in the upstream oil industry to handle multiple objectives. the utility function is difficult to construct. Note that the Pareto optimal set only needs to be constructed once.5 Concluding Remarks When the multiple objectives can be appropriately aggregated into one objective function. no matter what preferences the production engine ers have. certain multiobjective optimization problems in a petroleum field can have nonconvex Pareto optimal solutions. However. MSCF/D 60000 WEIGHTED SUM 50000 HIERARICHAL 40000 30000 20000 6000 7000 8000 9000 10000 11000 TOTAL WATER RATE. as the number of objectives in .2. STB/D Figure 7. and the weighted sum method may fail to obtain solutions of interest.e. the utility function method is able to obtain the most satisfactory solution efficiently. It appears that the weighted sum method described in Section 7. MULTIOBJECTIVE OPTIMIZATION 115 their preference. This is an advantage of the a posteriori methods over the a priori methods.
CHAPTER 7. MULTIOBJECTIVE OPTIMIZATION
116
an MOP increases, the computational demand of the hierarchical method may quickly becomes unmanageable, and the multidimensional Pareto optimal set get hard to visualize. Perhaps the more useful application of multiobjective optimization to the upstream oil industry is to improve the design of a production system, because in this case there are more tradeoffs involved and the impact of the tradeoffs is more significant in such design problems than in production operation problems. The evolutionary multiobjective optimization methods (Fonseca and Fleming, 1993) is a promising method for such tasks because of the following reasons: • The optimization problem of production system design has rough surfaces. Evolutionary algorithms are more appropriate than the gradientbased methods for such problems (Palke and Horne, 1997a). • Evolutionary algorithms maintain multiple solutions and have the capability to sample various Paretooptimal solutions in parallel (Oyama and Liou, 2002). `
Chapter 8 Coupling with a Reservoir Simulator
8.1 Introduction
The optimization methods investigated in this study require numerical tools to evaluate objective and constraint functions. For research purposes (such as evaluating the performance of the optimization algorithms without worrying about model accuracy), in Chapter 4 and 5 the objective and constraint functions were evaluated by a network simulator assuming that the required reservoir conditions around each well are given. For practical use, the optimization methods can be implemented in the following ways: 1. Integrating into a field simulator. In this approach, the optimization methods can be implemented in a field simulator that is capable of simulating both reservoirs and gathering systems. The function evaluation procedure of the optimization method can be implemented within the simulator. The advantage of this approach is that the optimizer is easy to use. If the engineer has built a simulation model for a field, it requires little extra effort to extend the simulation model for production optimization. The disadvantage of this approach is that its implementation requires access to the source code of the field simulator, which is not always possible. 2. Interfacing with a reservoir simulator. In this approach, the optimization methods are implemented in a standalone optimizer as in Chapter 4 and 5. The optimizer is then interfaced with a reservoir simulator, which feeds the optimizer the required reservoir
117
CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR
118
conditions. In contrast to the first approach, this approach is easy to implement, but the coupled system may be cumbersome to use. 3. Integrating with inhouse software. For instance, the optimization methods proposed as a result of this study can replace heuristic rules of an existing optimization system developed for a specific field. In the course of this study, some of the optimization tools developed in previous chapters were integrated into VIPEXECUTIVE (Landmark, 2001), a commercial field simulator. This allows the optimization tools that are found to be most suitable to be applied to a large variety of petroleum fields. In particular, The VIPEXECUTIVE with the optimization tools is an integral part of the EField Optimization System (EFOS) developed at Prudhoe Bay oil field in Alaska (Litvak et al., 2002). EFOS is designed to maximize the daily oil production in Prudhoe Bay field by optimally allocating the production rates, well connections to flowlines and (potentially) lift gas subject to multiple capacity constraints of facilities and velocity constraints in flowlines. This chapter describes how the optimization tools can be integrated into/interfaced with a reservoir simulator (using VIPEXECUTIVE as an example) and presents several examples of the integrated system. To facilitate our discussion of the integration/interfacing procedure, we first discuss how an integrated reservoir and production network (tubing strings and surface pipeline systems) model can be solved.
8.2 Full Field Simulation
Traditionally simulation models for reservoirs and production networks have been developed independently. Various methods have been proposed to couple the reservoir model and the network model for full field solutions (Emanuel and Ranney, 1981; Litvak and Darlow, 1995; Schiozer, 1994; Hepguler et al., 1997; Byer, 2000). Full field simulation offers increased accuracy in predicting the reservoir deliverability and is critical for many reservoir development and management studies. In this section, we
oil. In production and injection wells we obtain bottomhole pressures (or flow rates) and molar rates of the hydrocarbon components. 2. 1995): 1.1 Simulation of Petroleum Fields In the simulation of petroleum fields. and gas rates for every link and node (the notation of link and node used in this chapter is defined in Chapter 3). The Field simulator models the multiphase flow in an integrated system of reservoirs. gas. molar rates of the hydrocarbon components for every link and node. and water saturations. capillary pressure relations. surface pipeline systems. Consequently. 3. In well tubing strings and surface pipeline systems pressure distribution.2. A system of equations that describe multiphase flow in porous media of an underground reservoir. well tubing strings. the variables and the governing equations for the field simulation are different. This information is determined from the simultaneous solution of the following governing equations: 1. The following discussion assumes a compositional model. water. and composition of hydrocarbon phases. wellbores. saturation constraints.CHAPTER 8. A field simulator typically is capable of obtaining the following information (Litvak and Darlow. for the blackoil model can be regarded as a special case of a compositional model. COUPLING WITH A RESERVOIR SIMULATOR 119 describe first the governing equations and then the existing solution techniques for full field simulations. and phase . fluid properties can be represented by a compositional model or a blackoil model. oil. These equations include mass balance equations. phase equilibrium relations. In gridblocks of a discretized reservoir model we obtain • • • • • • • • pressures of each phases. 8. and possibly the fluid separation units and other facilities.
3) The concrete form of Eq. Fp represents the mass balance equations for each component in each grid block. nc (8. in Coats model (Coats. i = 1.. 8.. COUPLING WITH A RESERVOIR SIMULATOR 120 constraints.1) (8.2) where subscript p represents primary.. Multiphase flow models that relate the flow rate with the pressure drop across a tubing string: p2 = p2 ( p1. Eq. and qw denotes the oil.3 for various well models can be found in Aziz and Durlofsky (2002). x p represents the gridblock pressure..2 for various simulation models can be found in Aziz and Durlofsky (2002) and Cao (2002). Wellbore flow equations that relate the well flow rate of component c .CHAPTER 8. q g . oil and water saturations.5) .. 8. qw ) (8. these equations are classified into primary and secondary equations and the variables are classified as primary and secondary variables Fp (x p . 4.. gas. The concrete form of Eqs. The following system of equations that describe multiphase flow in a surface pipeline network: • Mass balance equation for each component at every node j∈Ωin ∑q cij = 0 .4) where p1 and p2 denote the upstream and downstream pressure of a tubing string. and the hydrocarbon component compositions of each gridblock. x s ) = 0 Fs (x p . q g . qo . 3. x s ) = 0 (8. As an example. and qo . qcw . respectively. c = 1.4 can be in the form of numerical multiphase flow models or hydraulic lookup tables. The classification is not unique.. and water flow rate across that tubing string. 1980). In reservoir simulation. 8. with the bottomhole pressure p w qcw = qcw ( p w ) (8..1 and 8. nn . s represents second. 2.
2a. which requires that the pressure for a node should be the same no matter from which path it is computed. A reservoir simulator usually models multiphase flow in a reservoir up to the wellbores. 8. Solve the linearized system equations to obtain updated estimates of the primary variables.1 for every gridblock and Eq. Flow equations that relate the flow rate and pressure drop through a link (i.3 for every well.e. qw ) (8. A network simulation involves solving a set of nonlinear equations constructed from Eqs. Update the secondary variables using Eq. 2. The selected set of equations are solved by the NewtonRaphson method. . Repeat Step 2a2e until converge. 2d. or a pump. 3. Repeat Step 2. qo . 8. Linearize the set of system equations. 8. 2b. etc. Sometimes a network model also contains wellbore models.. q g .CHAPTER 8. Chapter 3 presents a solution procedure for network simulation. 2c. 8. COUPLING WITH A RESERVOIR SIMULATOR 121 where qcij denotes the flow rate of component c between node i and j . March forward a time step. a pipe.) p2 = p2 ( p1.6) where p1 and p2 denote the upstream and downstream pressure of a link. A network simulator simulates multiphase flow in well tubing strings and surface pipeline systems by treating production wells as sources with fixed pressures or flow rates.2 for every gridblock and Eq. Form a set of system equations by constructing Eq. and nc is the number of components.3 for every well. 8. Ω in denotes all the adjacent nodes for node i . The basic steps in a reservoir simulation can be described as follows: 1. 2e.6 and Kirchoff’s law. nn is the number of node.38. • • Kirchoff’s law. a choke. Initialize the reservoir model.
Litvak and Darlow (1995) proposed such a procedure for a compositional full field model. solution techniques developed for full field simulation focus on how to couple the reservoir model and the network model at the timestep level. Update reservoir conditions. March to the next time step. Various coupling methods have been proposed in the literature and they are reviewed in the next section. COUPLING WITH A RESERVOIR SIMULATOR 122 Full field simulators combine the reservoir model and the network model. and the surface pipeline system to obtain an estimate of the primary variables. Algorithm 8. This procedure combines the system of equations for the reservoir model and the network model and solves them simultaneously using the NewtonRaphson method. Repeat Steps 12 until convergence. saturations. Algorithm 8. Update pressure. This procedure requires that the reservoir model and the network model are implemented in one simulator.2: Iterative coupling. and component compositions in reservoir gridblocks using the solution from Step 1 and the secondary equations. 2. wellbores. This procedure alternately solves a network problem and a reservoir problem until the system of equations for the full field model are satisfied. Schiozer (1994) and Byer (2002) investigated a similar implicit coupling procedure for a blackoil model.CHAPTER 8. To utilize fully existing techniques developed for reservoir simulation and network simulation.2 Coupling Reservoir and Production Network Solutions The coupling methods proposed in the literature can be classified into implicit and explicit procedures.2. 3. 1. The following procedure was proposed by Litvak and Darlow (1995) for a compositional model (this method was named as an “explicit procedure” in Litvak and Darlow (1995)): . 8. Solve linearized system of equations for the full field model. The steps in this procedure are described below. Linearize and solve the flow equations in the reservoir. 4.1: Standard implicit coupling.
saturation. This procedure was investigated by Schiozer (1994) and Byer (2000). Form a network problem by taking reservoir conditions in well gridblocks from the previous time step as part of the boundary conditions of the network problem. . March to the next time step. Take a Newton iteration for the reservoir model. this procedure may not converge or converge to a wrong solution (Litvak and Darlow. (1985). and component compositions in reservoir gridblocks using the solution of the linearized flow equations. Solve the network problem and obtain well production rates and bottomhole pressures. March to the next time step. Form a network problem by taking reservoir conditions in well gridblocks from the previous iteration (or the previous time step at the first iteration) as part of the boundary conditions for the network problem. The advantage of this procedure is that it can be used to couple a reservoir simulator with an independent network simulator. 4. However. Similar procedures were also presented by Hepguler et al. Solve a network problem. 1. Consequently the solution may not satisfy the system equations of the full field model. and Breaux et al. This procedure performs network simulation only once. 2. linearize and solve the reservoir flow equations. 1995). solve a reservoir simulation problem as described in Section 8. 2. 3. However. Solve a reservoir problem. COUPLING WITH A RESERVOIR SIMULATOR 123 1. (1997). The convergence criterion is that the production rates and bottomhole pressure obtained from Step 1 and Step 2 are within a specified tolerance for every well. Algorithm 8. and was used by Emanuel and Ranney (1981) to study an offshore reservoir. Update pressure. Taking the bottomhole pressures or production rates from Step 1 as the well constraints.3: Explicit coupling. for certain cases.1. Repeat Steps 12 until convergence. based on reservoir conditions at the beginning of a time step. Emanuel and Ranney (1981).2. Solve a network problem. Taking the bottomhole pressures or production rates from Step 1 as the well constraints.CHAPTER 8. 3. this approach is the easiest to implement and the cheapest to run among all coupling methods. Solve the network problem and obtain well production rates and bottomhole pressures.
This section first presents a general procedure for integrating production optimization tools into VIPEXECUTIVE.). invoke a procedure proposed by Litvak and Darlow (1995) to convert pressure and flow rate constraints on network nodes to flow rate constraints on individual wells. 1. The optimization tools are integrated with the standard fully implicit coupling procedure of VIPEXECUTIVE at the Newton iteration level (Figure 8. such as well controls and gaslift allocation.1 A General Integration Procedure As a reservoir development and management tool. etc. VIPEXECUTIVE (Landmark. COUPLING WITH A RESERVOIR SIMULATOR 124 8. Some wells are closed. 2.1). etc. Therefore. total field flow rates. invoke well management routines that handle well constraints and gaslift allocation.3.CHAPTER 8. For example. If . the optimization tools can be integrated into the solution procedure of the field simulator in the same way as these reservoir management features are integrated. Updated well constraints. etc. Perform well management routines (optional). The results of this step may include the following: • • • • Updated well status. Based on current reservoir conditions.4: Integration procedure. GOR. Some wells are reopened. 8. Algorithm 8.3 Integration of Optimization Tools with VIPEXECUTIVE Some of the optimization tools we investigated were integrated with a commercial field simulator. The role of these management features in a field simulator is the same as those of our optimization tools. VIPEXECUTIVE contains certain reservoir management features. Updated lift gas rate for gaslift wells. Convert pressure constraints (optional). Based on user specifications and estimated field conditions (well water cut. Some wells are worked over. Updated injection rate for injection wells. 2001). a well should not produce at an oil flow rate higher than 3000 STB/D to satisfy a total field gas flow rate constraint. Some wells are reconnected to different pipeline systems. then describes the integrated optimization tools and some related issues.
The flow rate constraints on network nodes are handled in Step 3. 4. wellbore. saturations. Perform production optimizations (optional). Perform userspecified production optimizations with current reservoir conditions. COUPLING WITH A RESERVOIR SIMULATOR 125 Step 3 is to be invoked. only pressure constraints are converted. Repeat Steps 15 until convergence. and component compositions in reservoir gridblocks using the solution from Step 1. The results of this step may include • • • Updated well constraints. Update pressure.1: Integration of optimization tools to VIPEXECUTIVE. Previous time step Perform well management routines Convert pressure constraints to flow rate constraints Optimize Production Operation Settings Optimize? No Determine active constraints Yes Form and solve full field system equations No Converged? Yes Next time step Update reservoir conditions and network conditions Figure 8. March to the next time step. 3. 7. Update reservoir conditions. Updated well connections. 6. Updated lift gas rate for gaslift wells. Solve linearized system equations for the full field. The updated operation settings from Step 1 and 2 are incorporated in this step. and surface pipeline systems to obtain an estimate of the primary variables. 5. Linearize and solve the flow equations in the reservoir. .CHAPTER 8.
3. The following procedure corresponds to Step 2 and 3 of Algorithm 8.1) . For every well i . The speedup techniques described in Section 4. well i remains fully open in the next iteration. the oil rate of well i is not allowed to exceed xi q o. and water cut for each well from Step 1. Update the well constraints for the next Newton iteration. COUPLING WITH A RESERVOIR SIMULATOR 126 8.4. the LPI method for rate allocation (Section 4. where qo . if xi ∈ (0. Solve the LP problem formulated in Step 2. GOR. .1). and cw from Step 1. 2. the LPII method for rate allocation (Section 4. v v v the weighting coefficients co .CHAPTER 8. and the userspecified constraint limits and weighting coefficients for the objective function.4. 3. Convert pressure constraints on wells and network nodes to flow rate constraints on individual wells using a procedure proposed by Litvak and Darlow (1995). if xi = 1 .3. Use the following information to build the LPII model. well i is closed in the next iteration.1) are determined during this procedure. c g .i at the next max iteration.i is the maximum oil rate of well i used to build the LP model in Step 2.2 are used to reduce the size of the LPII model. and 3.4). and cw for each velocity constraint (see Section 4. 4.4. • • • the oil rate. if the scaling factor xi = 0 . The LPII method for rate allocation. 2. the partial enumeration method and the genetic algorithm for well connection optimization (Chapter 6). 1.2 Integrated Optimization Tools The following optimization tools were integrated into VIPEXECUTIVE: 1. To ensure that the constraints will not be violated at the end of a time step. Specific integration issues for each optimization tool are discussed below. c g . a rate allocation problem is solved by the LPII method at every Newton iteration of the solution procedure for the full field model. The v v v weighting coefficients co . and its oil rate is allowed to increase at the next max iteration.3.
well connection optimizations are timeconsuming. Well connection optimization. In addition.2) will be more effective. to reduce the computational load. the LPII method optimizes rate allocation based on updated reservoir conditions. Both the partial enumeration method and the genetic algorithm described in Chapter 6 are implemented. 8. For these reasons. in a time step. the user can specify in which Newton iterations well connection optimization will be performed.2.2. The rate allocation problem within both algorithms are solved by the LPII method. . the method is not invoked at all Newton iterations. Consequently in later iterations fewer constraints may be violated and the speedup techniques (the constraint elimination procedure described in Section 4. it is not allowed to exceed its scaled value in the next iteration. COUPLING WITH A RESERVOIR SIMULATOR 127 In Step 4. to build Model LPI. once a well rate is scaled. The LPI method for rate allocation. the nonlinear interactions in the reservoir are partially taken into account by solving a rate allocation problem at every Newton iteration. Thus performing well connections between iterations can hinder convergence of the solution procedure of the field simulator. The optimization package and the reservoir simulator can be coupled together using the iterative coupling procedure or the explicit coupling procedure described in Section 8. Well reconnections can significantly change the flow behavior in the system. The user can specify at which Newton iterations to invoke the LPI method. This method is coupled into VIP EXECUTIVE in the same manner as the LPII method except 1. in the current implementation.3. the optimization package can be interfaced with a reservoir simulator to do full field simulation and production optimization. Therefore. and 2.4 Interfacing with an Independent Reservoir Simulator When the optimization tools are developed as a standalone package without reservoir models. the well performance curves have to be constructed. At each Newton iteration.CHAPTER 8.
Wells B1B10 can be connected to either the lowpressure stage or the highpressure stage of the separator. . For evaluation purposes. we imposed four hypothetical constraints on the production system. This section presents three examples: two for longterm reservoir prediction/development studies and one for onsite production optimization. The third example describes how the optimization tools can be applied to the giant Prudhoe Bay oil field in Alaska. The second example predicted the oil production of another oil field in the GOM for the whole life of the field.2 illustrates the production system. A1A5. This example has been presented in Wang. the separator cannot receive gas from either pipeline at a rate of more than 12 MMSCF/D. are tied to a 40slot conventional fixedleg platform. a threestage separator processes the produced fluids. the daily oil production from the field was maximized by optimally allocating production and lift gas. On the platform. The separator has a maximum gas handling capacity of 30 MMSCF/D and a maximum of water handling capacity of 5000 STB/D. and Aziz (2002a). Litvak. COUPLING WITH A RESERVOIR SIMULATOR 128 8. Our test example was based on this reservoir. the daily oil production from the field was maximized by allocating the production rates and the well connections. Five production wells.5. Wells A1A5 can be tied to any stage of the separator. Our job was to maximize the oil production by selecting appropriate well rates and well connections. The first example predicted the oil production of an oil field in the Gulf Of Mexico (GOM) for five years.5 Applications The optimization system described in Section 8. Moreover. At each time step. Ten production wells were drilled from a 10slot subsea template and connected to the platform by two 22000 foot long pipelines.1 A Gulf of Mexico Oil Field Example An integrated reservoir and surface pipeline network model had been built for one of the Gulf of Mexico oil fields.3 can be used for reservoir development studies or for onsite real time production control and optimization. 8. At each time step. Figure 8.CHAPTER 8.
then we checked the gas flow rate constraint on the separator. Run 2 predicted 11. Each run predicted the oil production for about five years. if it was violated. gas. 2001). The well rates in Figure 8. We made three predictive runs using VIPEXECUTIVE. The results are plotted in Figure 8.CHAPTER 8. This procedure is a well management option available in VIPEXECUTIVE (Landmark. In Run 1. we checked flow rate constraints sequentially and scaled well rates if necessary: first we checked the two gas flow rate constraints on HEADER_Y and HEADER_Z. we applied the partial enumeration method with the LPII method to optimize the well rates and well connections simultaneously. Figure 8. COUPLING WITH A RESERVOIR SIMULATOR Separator lp hp lp ip hp Oil Gas Water 129 TEMPLATE HEADER_Y HEADER_Z B05 B01 B02 B03 B04 B06 A1 B10 B07 B09 B08 A2 A3 A4 A5 Figure 8. and Run 3 . Over five years.2% more oil than Run 1.4 to select the well rates.2: The production system of a Gulf Of Mexico oil field. the production rates of the highest GOR wells connecting to HEADER_Y (or HEADER_Z depending on which constraint was violated) were scaled.38. Run 3 was even better. the highest water cut wells of the whole field were scaled to meet the water constraint. if either constraint was violated. In Run 3. In Run 2.3 shows that Run 2 produced at a much higher rate than Run 1. finally we checked the water constraint on the separator.3. and water rates.7 were normalized by some fixed oil. we applied the LPII method described in Section 4. the highest GOR wells of the whole field were scaled to meet the gas constraint. if it was violated. at every Newton iteration of each time step.7.38.
Run 1 did not utilize the processing capacity of HEADER_Z as efficiently as Run 2. .60 0.20 0. without well reconnection. The rate allocation procedure in Run 1 first scaled the highest GOR wells to meet gas constraints. This resulted in suboptimal solutions.3: Normalized daily total oil rate. 1.6). and then scaled the highest water cut wells to meet the water constraint.7).9% more oil than Run 1. most of the time Run 2 had been producing at a lower total water rate (Figure 8.40 0. The fact that Run 3 outperformed Run 2 can be explained as follows. COUPLING WITH A RESERVOIR SIMULATOR 130 predicted 18. On the other hand. most of the gas processing capacity of HEADER_Y was wasted (Figure 8. In other words.00 NORMALIZED OIL RATE 0. The excellent performance of Run 2 compared to Run 1 can be explained as follows. while HEADER_Z was constrained by its gas processing capacity (Figure 8.20 1. At later times in Run 2.00 0 500 1000 TIME.CHAPTER 8. The algorithm in Run 3 identified this imbalance and switched a high GOR well from HEADER_Z to HEADER_Y. And most of the time Run 2 predicted a much higher gas rate for HEADER_Z than Run 1.4) than Run 1.80 0. For instance.5) but a higher gas flow rate (Figure 8. Run 2 used a rigorous optimization procedure (the LPII method) and scaled the well rates differently. DAYS 1500 2000 RUN 1 RUN 2 RUN 3 Figure 8.
80 0. DAYS 1500 2000 NORMALIZED GAS RATE RUN 1 RUN 2 RUN 3 Figure 8.40 0.40 0.4: Normalized daily total gas rate.00 0.20 NORMALIZED WATER RATE 1.20 0.5: Normalized daily total water rate.00 0.00 0 500 1000 TIME.00 0 500 1000 TIME. .CHAPTER 8. 1. DAYS 1500 2000 RUN 1 RUN 2 RUN 3 Figure 8.80 0.60 0.60 0.20 1.20 0. COUPLING WITH A RESERVOIR SIMULATOR 131 1.
40 0.60 0.00 0 500 1000 TIME.20 1.20 0.20 0. .80 0.00 0.7: Normalized daily gas rate at HEADER_Z. COUPLING WITH A RESERVOIR SIMULATOR 132 1.6: Normalized daily gas rate at HEADER_Y.80 0.00 0 500 1000 TIME.20 1.60 0.00 0. DAYS 1500 2000 RUN 1 RUN 2 RUN 3 Figure 8.40 0.CHAPTER 8. DAYS 1500 2000 RUN 1 RUN 2 RUN 3 NORMALIZED GAS RATE NORMALIZED GAS RATE Figure 8. 1.
The vertical distances between the separator and wells vary between 10. two cases were considered. In this model. This method proceeds as follows: .5 will be specified for that producer. For evaluation purposes. etc. The purpose of this example was to demonstrate that the application of the developed shortterm optimization algorithms in longterm prediction can improve facility design. we made two runs: a BASE run and a LPI run.1 to allocate the lift gas and production rates subject to the total oil rate constraint of the field and the total gas rate constraint of the separator. 26 wells are scheduled to be drilled at different times. 2001). The lift gas rate for a gaslift well cannot exceed 6000 MSCF/D. an ONTIME factor with a value between 0 and 1 may be defined for a well. its discounted flow rates for a time step are the instantaneous flow rates time its ONTIME factor for that time step. closed for maintenance. some wells will be recompleted. The BASE run allocated the lift gas and production rates using a “conventional” method available in VIPEXECUTIVE (Landmark. discounted flow rate is the average flow rate of a well for a reporting period. converted from producers to injectors. All wells are connected to a common offshore platform. Case 1. For a well. There are no flow interactions among different wells. During the prediction period. placed on gas lift.3. For example.2 Another Gulf of Mexico Field Example This is a reservoir development model for a newly discovered deepwater oil field in the Gulf of Mexico.000 ft and 14. COUPLING WITH A RESERVOIR SIMULATOR 133 8. The total field oil production rate cannot exceed a certain target rate. then an ONTIME factor of 0.000 ft. on which resides a separator and possibly other surface facilities.CHAPTER 8.5. In this example. In this case. To account for the downtime between two reporting dates in simulation. each with a different gas handling capacity for the separator. if a producer is expected to produce equivalent of one month during a two month reporting period. the separator has a gas handling capacity of 98 MMSCF/D. The LPI run uses Mode LPI described in Section 4. we make the following definitions: instantaneous flow rate is the actual flow rate of a producer when it is flowing. The prediction period is about 22 years. To facilitate later discussions.
In this procedure. the well’s lift gas rate is calculated to meet the gasliquid ratio specified in the table (Litvak et al. the LPI run produced at a lower oil rate. a minimum gaslift efficiency can be .10 shows the daily total gas rate for both runs. 3. the same flow rate constraint appeared more strict in the LPI run than in the BASE run. It shows that the BASE run injected about 20% more lift gas than the LPI run for the whole prediction period.CHAPTER 8. Thus. These tables correlate gasliquid ratio for a production well to its water cut and liquid production rate. This can be explained as follows. Liftgas rates are determined for specified wells based on some gaslift tables. Figure 8. Check the oil rate constraint.9 agrees with what we observe in Figure 8. Figure 8. gaslift is not implemented for the well. COUPLING WITH A RESERVOIR SIMULATOR 134 1. If it is violated. The total gas constraint was not active in both runs for the whole prediction period. It appears that the oil production histories of the two runs are the same except when the oil rate constraint in the BASE run was active. 4. When the oil rate constraint in the BASE run was active. When allocating lift gas. This is because the LPI method scales instantaneous production rate to meet the flow rate constraint and VIPEXECUTIVE reports discounted flow rate. Update the production rates of all wells with allocated gaslift rates.8 plots the daily field oil production history for both runs (the daily oil rates in both runs are normalized by the field oil target rate).9 plots the normalized cumulative oil production for both runs. 2. Check the total gas rate constraint.8. 1997). Otherwise. scale the production rates of certain wells to meet the oil rate constraint..11 plots the normalized cumulative lift gas injection for both runs. the LPI method needs to construct the gaslift performance curve. Figure 8. The trend of Figure 8. If the formation gasliquid ratio for a production well is higher than the table value. If the constraint is violated. Thus the gas handling capacity of the separator is oversized for this reservoir development scenario. scale both the production rates and lift gas rates of wells with the highest GOR to meet the total gas rate constraint. the LPI run used instantaneous production rates from each well but the BASE run used discounted production rates (these options are implemented in this way in VIPEXECUTIVE). We emphasize that when checking the flow rate constraints. Figure 8.
Figure 8.14 shows that over the whole prediction period the LPI run allocated about 40% less lift gas than the BASE run. certain careful considerations will reveal that no lift gas should be injected (Figure 8. From Figure 8. The reservoir model and the development scenario are exactly the same as in Case 1. If used appropriately.12 shows that the BASE run continued to produce at a lower oil rate than the LPI run until 4500 days. . This is due to two facts: 1) the LPI run optimized the production and lift gas rates simultaneously while the approach in the BASE run did not. The lower oil rate in the LPI run at later stages of the prediction period is probably due to the fact that there was less oil in the reservoir of the LPI run than in the reservoir of the BASE run at later stages of prediction.14 shows that the LPI run did not inject any lift gas until about 3000 days). after which the BASE run produced at a higher oil rate than the LPI run. we made the same two runs as in case 1: a BASE run and a LPI run. some wells were placed on gaslift.CHAPTER 8.14 plots the normalized cumulative lift gas rate for both runs. the LPI method can save significant amount of construction and operational cost. Again. Figure 8. Figure 8. the liftgas table does not have this feature and can overinject lift gas. Around 1900 days. Figure 8. 2) the LPI method can control the gaslift efficiency. We can conclude that facility design depends on the well management tools (the “conventional” approach or the LPI method) used to evaluate the sizes of the separator and the gas compressor for lift gas. Figure 8.13 plots the normalized cumulative field oil production history for both runs. Figure 8. Case 2.12 we can see that at about 1900 days the oil production in the BASE run dropped sharply to a lower rate.12 shows the normalized daily field oil production history for both runs. However. This demonstrates the danger of allocating lift gas and production rates separately. This can be explained as follows. The BASE run allocated the lift gas to these wells without the consideration of the total gas constraint. COUPLING WITH A RESERVOIR SIMULATOR 135 specified for each well. As a result. injecting lift gas decreased the total oil production instead of increasing it. thus avoiding overinjection of lift gas.13 shows that over the whole prediction period the LPI run produced slightly more oil than the BASE run. except that the gas handling capacity of the separator is lowered to 60 MMSCF/D. On the other hand.
1. COUPLING WITH A RESERVOIR SIMULATOR 136 1.0E+00 NORMALIZED OIL RATE 8.2E+00 1.0E01 BASE 4.8: Normalized daily total oil production history for Case 1.0E+00 0 2000 4000 TIME. .0E01 0. DAYS 8000 Figure 8.9: Normalized cumulative oil production history for Case 1.2E+00 NORMALIZED CUMULATIVE OIL PRODUCTION 1.0E01 4. DAYS 6000 8000 BASE LPI CONSTRAINT Figure 8.0E01 6.0E+00 8.0E01 6.0E01 2.0E01 0.0E+00 LPI 0 2000 4000 6000 TIME.CHAPTER 8.0E01 2.
2E+05 DAILY TOTAL GAS RATE. .0E+05 8.0E+04 4.10: Daily total gas rate history for Case 1.2 0.1 0 0 2000 4000 TIME.0E+00 0 2000 4000 TIME.CHAPTER 8.0E+04 0.11: Normalized cumulative lift gas injection for Case 1.9 BASE 0. COUPLING WITH A RESERVOIR SIMULATOR 137 1.0E+04 6. 1 NORMALIZED CUMULATIVE LIFT GAS INJECTION 0.5 0. MSCF/D 1.4 0.7 0.3 0. DAYS 6000 8000 BASE LPI CONSTRAINT Figure 8.0E+04 2.6 0.8 0. DAYS 6000 8000 LPI Figure 8.
12: Normalized daily total oil production history for Case 2. DAYS 6000 8000 Figure 8.2 0.1 0 SP BASE 0 2000 4000 TIME.0E01 4. DAYS 6000 8000 LPI BASE Figure 8.0E+00 9.13: Normalized cumulative oil production history for Case 2. COUPLING WITH A RESERVOIR SIMULATOR 138 1.CHAPTER 8. 1 NORMALIZED CUMULATIVE OIL PRODUCTION 0. .0E01 1.0E01 7.7 0.0E+00 0 2000 4000 TIME.0E01 2.4 0.0E01 0.9 0.0E01 3.3 0.5 0.0E01 5.0E01 NORMALIZED OIL RATE 8.6 0.8 0.0E01 6.
7 0.3 0.6 0.8 0. production optimization. EFOS shows how our optimization tools can be used for onsite production optimization in a real petroleum field. The major purpose of this section is to demonstrate possible use of the optimization tools for Prudhoe Bay field and their high efficiency even for such a giant field.3 Application to the Prudhoe Bay Oil Field Prudhoe Bay oil field is located on the North Slope of Alaska and operated by BP Exploration (Alaska) Inc. The production optimization module uses the optimization system described in Section 8.CHAPTER 8.1 0 BASE LPI 0 2000 4000 TIME. EFOS contains four major modules: automatic data preparation.4 0. and a userfriendly interface.3 to maximize the daily oil production and debottleneck the production system.5 0. EFOS is described fully by Litvak et al. (2002). 8.9 0. automatic model tuning. DAYS 6000 8000 Figure 8.2 0. Oil production from Prudhoe Bay is on decline and constrained by the gas and water handling limits of surface facilities and velocity constraints in flow lines. To .5.14: Normalized cumulative lift gas injection for Case 2. The Efield optimization system (EFOS) was developed at Prudhoe Bay oil field to automate and optimize the production operations of the field. COUPLING WITH A RESERVOIR SIMULATOR 139 1 NORMALIZED CUMULATIVE LIFT GAS INJECTION 0.
“Production wells are tied to 22 well pads in the Western Operating Area (WOA) and 16 drill sites in the Eastern Operating Area (EOA). The produced water is either reinjected in the EOR pattern or into water disposal wells. Production system description. as part of a WAG (water alternating gas) process in some 120 EOR patterns. a brief description of the production system of Prudhoe Bay field is presented below. For Prudhoe Bay oil field. In such a constrained environment. after compression in the central compression plant. Separator offgas is routed from the gathering centers and flow stations to the central gas facility where natural gas liquids and miscible injectant are extracted. Each gathering center and flow station has four threestage separator banks. It is injected. Well chokes are installed in well pad manifolds. Well pads and drill sites are tied to separator banks through high pressure and/or lowpressure flow lines. the possible usage of the optimization tools can be summarized as follows. Oil production in Prudhoe Bay field is constrained by capacity constraints of facilities and velocity constraints of flowlines. 1. Prudhoe bay has a flexible. highly automated and complex surface pipe network and processing facilities. The miscible injectant is used in the enhanced oil recovery (EOR) project. Gaslift is used in some production wells with high water cut. Separator banks can operate in high pressure or lowpressure mode with first separation stage pressure of about 670 psig and 150 psig. (1997). COUPLING WITH A RESERVOIR SIMULATOR 140 facilitate later discussions. Optimization of production rate and well connections in response to changing operational conditions. is reinjected into the gas cap for pressure maintenance. The natural gas liquids are then blended with the oil stream coming from the separators. The lean gas.” Possible usages of the optimization tools. the maximum possible field oil rate is sensitive to the . Gaslift wells can be tied only to lowpressure headers. The following description of the production system of Prudhoe Bay is extracted from Litvak et al. respectively. The production stream from individual well pads and drill sites is sent without any separation to three gathering centers in the WOA and to three flow stations in EOA.CHAPTER 8. Most production wells can be switched between high and lowpressure headers at the well pad manifolds.
For example.7. COUPLING WITH A RESERVOIR SIMULATOR 141 changing operational conditions. Suppose the rate allocation problem is formulated as a LP problem. Debottlenecking studies. In the first run. 2. Production capacity can also be affected by equipment maintenance and repair.. 2002). Results showed that the second run improved oil production by 2. which can happen during a 24hour period in Prudhoe Bay (Litvak et al. In the second run. This is illustrated here by the LPII method for rate allocation.55% compared to the first run. the choke settings for all production wells were not allowed to increase. the second run took about 850 seconds (in the second run. which is significant for this large field. One contribution of this study is that we developed optimization tools that can simultaneously optimize the production rate and the well connections. the first run took 0. On an IBM RS6000 computer.4). as follows maximzie f (x ) = c T x (8. and other planned or unplanned activities. Therefore. the production rates were optimized using the LPII method (see Section 4.2 BCF/D when the ambient temperature increases from 0 °F to 40 ° F. the major portion of the computational time was spent on solving the network problem after each well reconnection). Also the operational conditions at Prudhoe Bay can change significantly in a short period. This demonstrates the high efficiency of the developed optimization methods.7a) .CHAPTER 8.07 seconds. we conducted two optimization runs for the whole Prudhoe Bay field. Another important use of the optimization tools is to identify bottlenecks in the production system and evaluate debottlenecking opportunities. Problem 8.3. it is necessary to optimize the production rate and well connections from time to time to maximize the total field oil rate in response to the changing operational conditions. In both runs. To demonstrate the potential advantage of the simultaneous optimization. the production rates and well connections were optimized simultaneously by the combined use of the partial enumeration (PE) method and the LPII method. The number of well reconnections was not allowed to exceed 50. field gas compression capacity can be reduced by 1. The bottlenecks of the production system can be identified from the optimal solution of the LPII model for the rate allocation problem.
The rate allocation problem for Prudhoe Bay field contains mostly multiple gas. 8. the gaslift performance is evaluated by the combined use of inflow performance curves and well tubing hydraulic models.8) where f * denotes the optimal objective function of Problem 8. and velocity constraints have different units.8 states that at the optimal solution of a LP problem.. i = 1. the tubing hydraulic model is not accurate enough to . several constraints with the most negative dual variables from each category (gas. b denotes the maximum fluid processing capacity of facilities and velocity limits for flowlines. In the Prudhoe Bay field model. m i 0 ≤ xi ≤ 1 .. λi = 0 indicates that increasing bi has no impact on the optimal objective function value... water.. 3. The dual variables for water.CHAPTER 8. i = 1..7c) where x denotes the scaling factors for well rates. n 142 (8. gas. COUPLING WITH A RESERVOIR SIMULATOR subject to a T x ≤ b i . f (x ) is the objective function. The debottlenecking opportunity with the best economic return can be implemented.. To avoid missing debottlenecking opportunities. After debottlenecking opportunities are identified. The optimal solution of a LP problem contains a dual variable λi for each constraint and λi = − ∂f * ∂bi (8. and the more likely it is that facility capacity or velocity constraint bi is the bottleneck of the production system.. Gaslift optimization. the gas compression capacity of the most loaded compressor may be increased and the rate allocation problem resolved to see how much oil increase can be obtained. or velocity constraints) can be identified as debottlenecking opportunities.7b) (8. The more negative the λi . the dual variable indicates how sensitive the optimal objective function is relative to the righthand side of the constraints.7. However. For example. Eq. the more sensitive the objective function to bi . water constraints of facilities and velocity constraints. a series of sensitivity runs can be performed to evaluate the selected opportunities.
We conducted two optimization runs for the whole Prudhoe Bay field.. Results showed that the second run improves daily oil production by 6% compared to the first run. This study demonstrated that the optimal allocation of lift gas is a potential effective way to increase the daily oil production. (3) the user needs less effort to develop and maintain the simulation and optimization models. They can also be implemented in an inhouse software. Application examples presented in this chapter show that optimization methods developed in this study have distinctive advantages over the heuristic methods currently .6 Concluding Remarks For practical use. Therefore. COUPLING WITH A RESERVOIR SIMULATOR 143 represent real physics because of the large diameters of the tubing strings (existing multiphase flow models do not perform well for large diameter tubing strings (Litvak et al. the LPI method has not been applied to the real field. the optimization tools can be implemented in different ways. their use should be managed carefully to avoid severe convergence issues of the solution procedure and to ensure no constraints violations at the end of a time step. the second optimization run took about 10 seconds. If the optimization tools are applied for longterm prediction. etc. The use of a commercial simulator has several advantages: (1) the integrated system can be applied to a variety of fields. If possible. 2002). 2002)). The methods can be integrated into or interfaced with a commercial simulator. In a full field study.. because implicit coupling is able to obtain simulation results consistent with the full field model. However.CHAPTER 8. In the second run. the production rates and lift gas rates were optimized simultaneously by the LPI method. (2) advances in simulation technology can be easily incorporated and transferred to other fields (Litvak et al. the production rates were optimized using the LPII method. the network model and the reservoir model can be coupled together using either the implicit procedures or the explicit procedure. 8. the LPI method can be used to identify potential benefits of doing gaslift optimization for the Prudhoe Bay field. the implicit coupling procedures are recommended. In the first run. On an IBM RS6000 computer.
In addition. they are efficient enough to be used for both realtime optimization and longterm prediction studies for largescale reservoirs. The optimization tools developed in this study were designed to optimize the shortterm production operation based on static reservoir conditions.CHAPTER 8. COUPLING WITH A RESERVOIR SIMULATOR 144 employed in the oil industry. The reservoir dynamics were not considered in the optimization step. the impact of decisions based on this approach on long term oil recovery needs further investigation. . Although we did apply these optimization tools to longterm predictions.
The major contributions of this work and areas for further study are summarized in this section. even for large production systems such as those in the Prudhoe Bay field. the equalslope method for gaslift optimization. For instance. The optimization methods investigated and developed here are effective for problems of varying complexities and sizes. LPII. This problem is being faced in many mature fields and is an important element to consider in planning the development of a new field. MILPI. and well connections to flowlines subject to multiple flow rate and pressure constraints to achieve certain shortterm operational goals. 2.1 Conclusions 1. For the rate allocation problem. compared and tested optimization methods for petroleum production problems. lift gas rates.e. The methods can be used for both shortterm production optimizations and longterm reservoir development studies. In this research we developed. they can be used to optimize well connections and well rates simultaneously. when flow interactions among wells are not significant. and the rulebased methods available in commercial simulators).Chapter 9 Conclusions This study addressed the problem of optimizing the production rates. the various separable programming methods (the LPI.. 9. 145 . MILPII methods described in Chapter 4) proved to be better than other conventional approaches used in the oil industry (i.
The MILPII method is especially suitable for gaslift optimization problems with many nonconcave gaslift performance curves. . CONCLUSIONS 146 Specifically. 4. Formulation P2 (Chapter 5) proved to be most appropriate for this problem. 5. and well connections simultaneously subject to multiple flow rate and pressure constraints. For the optimization of well connections. 9. Postanalysis should be performed to ensure the validity of a solution from a simulation or an optimization run. A twolevel programming approach was developed to optimize the production rates. the partial enumeration (PE) algorithm proved to be both efficient and robust compared to a genetic algorithm (GA). This may bring computational difficulties to full field simulations and production optimizations. For a production optimization problem with multiple objectives. For the rate allocation problem. and LPII methods are suitable for largescale rate allocation problems. the giant Prudhoe Bay oil field in Alaska. 6. Multiple solutions may exist for the multiphase flow problem of a gathering system.2 Summary of Contributions The major contributions of this study can be summarized as follows: 1. Optimization methods were developed and tested on various synthetic examples and applied successfully to a real petroleum field. 7. The twolevel programming approach developed for the entire production optimization problem is flexible. Different optimization algorithms can be used for the upper level problem and the lower level problem. the weightedsum method may fail to obtain solutions of interest to us. when flow interactions have significant impact on the optimal solution. The MILPI. 2.CHAPTER 9. The formulated nonlinearly constrained problem (NCP) can be solved efficiently by a sequential quadratic programming (SQP) algorithm. the MILPI method can be used for problems with performance curves of arbitrary shapes. 3. simulation models capable of capturing such flow interactions should be used in the optimization process. LPI. lift gas rates. Results demonstrated that the methods developed have distinct advantages over conventional production optimization approaches.
Several techniques were developed to speed up the optimization process. An appropriate formulation. and they can be very effective for largescale problems. LPII. the Jacobian method. Derivatives obtained from automatic differentiation are more accurate than those obtained from finite difference approximations. In certain cases. 10. A partial enumeration (PE) method was developed to optimize well connections. Solutions obtained from the PE method were at least as good as those from the GA. and sensitivity analysis. .CHAPTER 9. was proposed to compute the sensitivity coefficients of a gathering system. The PE method is more efficient than a genetic algorithm (GA). The optimization problem formulated in this manner was solved by a SQP algorithm. CONCLUSIONS 147 3. Some of the optimization tools were coupled with models for multiphase flow in the reservoir and surface pipeline network in a commercial reservoir simulator. 8. Several existing separable programming models (Model LPI. Sensitivity coefficients are required in applications such as optimization. An efficient solution procedure was developed to simulate multiphase flow in a gathering system. 5. An efficient method. 7. automatic differentiation is more efficient than finite difference approximations. This automates the application of optimization tools to shortterm production optimization and longterm reservoir development studies. A framework was developed to formulate and solve multiobjective production optimization problems. It was demonstrated through an example that the weightedsum method may fail to obtain solutions of interest to us. An automatic differentiation technique was applied to compute derivatives required in network simulation and production optimization. Formulation P2 (Chapter 5). model tuning. and MILPI) were investigated and a new model (Model MILPII) was proposed for the rate allocation problem. 9. was proposed for the nonlinear rate allocation problem. 6. This framework unifies several existing multiobjective optimization algorithms. 4.
Develop methods to construct quality well performance curves efficiently. 3.3 Recommendations for Further Study 1. 2. Develop optimization methods for gathering systems with loops. And if so. CONCLUSIONS 148 9. 4. 5. Investigate the impact of decisions obtained from shortterm production optimizations on longterm oil recovery. how to handle them efficiently. Investigate whether well interactions within the reservoir are important for shortterm production optimization problems. Investigate multiobjective optimization methods for surface facility design problems. Investigate how the accuracy of the numerical models for multiphase pipe flow and choke flow impacts the optimal solution.CHAPTER 9. Investigate the applicability of genetic algorithms on constrained production optimization problems. Genetic algorithms are slow. 6. they are flexible and robust enough to handle optimization problems with complex simulation models and rough objective surfaces. Develop methods to find stable solutions of the multiphase network flow problem. . 7. however. 8.
cost coefficients for an objective function velocity coefficient for flow rate of phase p (Section 4.Nomenclature A B bi = = = = = = = = = = = = = = = = = = = = = pipe area formation volume factor righthand side of the ith constraint profile parameter in the drift flux model discharge coefficient in the Sachdeva choke model (Sachdeva et al.1) subset of constraints that are active at the optimal solution choke inside diameter (Section 3. 1986) specific heat capacity at constant pressure and volume. deviational variable (Chapter 7) liquid and gas phase in situ fractions. vector of objective functions (Chapter 7) function Fanning friction factor flowing fractions of oil and water gravity. Cv c cv p ˆ c d El . f w g h hij J k L 149 . E g F f fm fo ..4). respectively constraint functions.2) achievement function (Chapter 7) the jth function involved in ith constraint of a separable problem Jacobian matrix ratio of specific heat capacity Lagrangian function C0 Cd C p .4. respectively constraint functions. gradient (Section 5.
Section 5.0 w .4) number of connections of a well number of nodes number of wells pressure. j lower bound of a constraint merit function number of constraints number of decision variables. polytropic exponent of gas (Section 3.NOMENCLATURE l 150 M m n = = = = = = = = = = = = = = = = = = = = = = = = = = = = nc nn nw p n Q p.6) velocity the maximum in situ velocity along a flowline for a reference flow rate of phase p (Section 4.3.2) flow rate limit of phase p for node j flow rate flow rate of component c between node i and j flow rate of phase p for node j flow rate of phase p for well j Reynolds number of a mixture solution gasoil ratio number of discrete points of a piecewise linear well performance curve feasible set of an optimization problem pipe perimeter slack variable velocity (Chapter 3).1) weighting coefficients q qcij qn. j p Rem Rs r S S s U Ud U sg U sl U sm u v v max p. utility function (Chapter 7) drift velocity of the gas in the drift flux model superficial velocity of the gas phase superficial velocity of the liquid phase superficial velocity of the mixture upper bound of a constraint.1. search direction (Section 2. j p q w. system response (Section 3.4.
well connections (Section 6. defined in Section 5. independent variable.1 formulation P2.3. 1998) bottomhole pressure drift flux model for multiphase pipe flow Efield optimization system for the Prudhoe Bay oil field Eastern Operating Area of the Prudhoe Bay oil field genetic algorithm gas oil ratio gas liquid ratio Gulf of Mexico goal programming integer programming linear integer programming linear programming a model for solving the rate allocation problem (Section 4.1) a model for solving the rate allocation problem (Section 4.4) mixed integer linear programming a model for solving the rate allocation problem (Section 4.3.. mass fraction of gas (Section 3. binary decision variable (Section 4.2 . production rates (Section 6.3.3.3.2) critical ratio of downstream pressure to upstream pressure value of an objective function.2) feasible set of a vector of objective functions lift gas rates (Section 6. ratio of downstream pressure to upstream pressure (Section 3.3.2) a model for solving the rate allocation problem (Section 4.NOMENCLATURE x = decision variable.2).4).4). defined in Section 5.3.2) lower bound for an objective function upper bound for an objective function 151 Y y = = yc z = = = = z z Acronyms ADIFOR BHP DFM EFOS EOA GA GOR GLR GOM GP IP LIP LP LPI LPII MILP MILPI MILPII MIP MOP NCP P1 P2 = = = = = = = = = = = = = = = = = = = = = = = an automatic differentiation software (Bischof et al.3) mixed integer programming multiple objective optimization problem nonlinearly constrained programming formulation P1.
2) delta (change) absolute pipe roughness (Section 3.3. Section 5.1.. 1998) separable programming sequential quadratic programming water alternating gas well index Western Operating Area of the Prudhoe Bay oil field 152 Symbols α = system parameters.3.2). Lagrangian variable (Section 5.3).NOMENCLATURE PE SLC SNOPT SP SQP WAG WI WOA = = = = = = = = partial enumeration method for well connections optimization. p µ θ ρ = = = = = = = = = = = = τw ∆t Ωn Ω in Ω iw Ωn s ∀ ∃ . penalty coefficients (Section 5.1).2) wall friction shear stress time step length in reservoir simulation set of all network nodes set of nodes that directly connected to node i set of wells whose flow streams enter node i set of solution nodes arbitrary exist δ ε = = γ λ = = λc . steplength of a line search optimization method (Section 2.2) viscosities angle density.5.6.1) specific gravity decision variable (Section 4.2). tolerances (Section 3.2 sequential linearly constrained a generalpurpose system for solving largescale optimization problems (Gill et al. defined in Section 6.5.3) mobility of component c in phase p (Section 3. dual variable (Section 8. perturbations (Section 3.3.
NOMENCLATURE ∇ ∞ 153 = = gradient Linfinite norm Subscript 1 2 c g l lg = = = = = = = = = = upstream downstream component gas liquid lift gas mixture oil phase water m o p w Superscript max min n r s t v = = = = = = = = w maximum minimum node reservoir surface total iteration index. velocity (Section 4.1) well .4.
S. K.. A. Third Edition. S. “Criteria for Gas Lift Stability”. AND SORENSEN.html. Z.S. L. J. W.D. paper SPE 12295.B.... ARONOFSKY. Society for Industrial Applied Mathematics. winter 20012002. BLACKFORD.. J. HAMMARLING. November. 2002. Linköping. 2002. “The Use of Linear Programming Model for Scheduling Crude Oil Production”. 1961.. Sweden.. J. GREENBAUM.glue. ARONOFSKY. ANDERSSON. LAPACK User’s Guide.. AZARM. A...edu/~azarm/optimum_notes/multi/multi. 675. Trans. JPT... H. AND LEE. AND DURLOFSKY. Stanford University. S.. ATTRA. 14521456. 1983. W. BISCHOF.. 1988.. Multiobjective Optimization in Engineering Design – Applications to Fluid Power Systems..Bibliography ANDERSON. J.S. 154 . DU CROZ. D. 8286.. C. ASHEIM. “Application of Optimizing Techniques for Studying Field Producing Operations”. Course material for PE223. Linköping University. Fundamentals of Reservoir Simulation. DONGARRA. DEMMEL.. MCKENNY. 5154.M.. S. E.. 2001. 1999. AZIZ. H. January.umd. J. “Optimization Methods in Oil and Gas Development”. Class Notes for Multiobjective Optimization.. J... WISE. A. 1958. AIME 213. AND BLACK. BAI. Retrieved in October 2002 from http://www. Journal of Petroleum Technology. dissertation No.
C. BISCHOF. C. BARNES. AND GRIEWANK... D. 1982. DE MONTGOLFIER.P.. 1994. AND MUELLENBERG. K. CORLISS. Constrained Optimization and Lagrange Multiplier Methods. “A Study of Two Phase Flow in Inclined Pipes. R.P. AND MAUER. 1991. Dallas. and Tools. HOVLAND. 1997. HUMPHREY. paper SPE 28249 presented at the SPE Petroleum Computer Conference. A. J. SIAM.. J. Dynamic Programming. LA.A. 1971. O. Texas.. OGCI Publications. Athena Scientific. BENAYOUN. Production Optimization. Introduction to Linear Optimization.. BELLMAN. M. K. S. New York. Soc. BISCHOF. vol. P. “Application of a GeneralPurpose Network Optimizer to Oil and Gas Production”. H.0 User’s Guide Technical Report (Revision D). TERGNY. AND PETALAS. 607. of Pet.. 1973. BEGGS.. D... R. “Linear Programming with Multiple Objective Functions: STEP Method (STEM)”. Alaska”. CARLE. paper SPE 38838 presented at the 1997th SPE Annual Technical Conference and Exhibition held in San Antonio..BIBLIOGRAPHY 155 AZIZ.. P. Philadelphia. L.. N. Massachusetts. EDITORS.. Applications. “A Production Optimization System for Western Prudhoe Bay Field. A. Mathematical Programming 1(3)... of AIME. 256.D. 31 July3 August.. ADIFOR 2... J. 1998.D. 1990. Belmont.. paper SPE 20653 presented at the 65th Annual Technical Conference and Exhibition of the Society of Petroleum Engineers held in New Orleans. D. 1996. Mathematics and . J. 1957. AND DUTTAROY. AND LARICHEV. Princeton University Press. BARUA.. AND BRILL. Oil and Gas Consultants International Inc. PROBST. 58 October. K. BERZ. 366375. Computational Differentiation: Techniques. BERTSIMAS.. 1997. BEGGS.” Trans. September 2326. H.. Academic Press. BERTSEKAS... G.. AND TSITSIKLIS. Tulsa. “New PCBased Software for Multiphase Flow Calculations”. KHADEMI.. A..
. CARROLL.L.P. BLANK.C. 1998. paper SPE 35616 presented at the Gas Technology Conference held in Calgary.BIBLIOGRAPHY 156 Computer Science Division Technical Memorandum No. BOYCHUK... 14291436.N. AND ALUMRAN. dissertation. 192 and Center for Research on Parallel Computation Technical Report CRPC95516S. T. BOYD. S. BUITRAGO.. Society of Petroleum Engineering Journal. D..O. Retrieved in August 2000 from http://cuaerospace. AND BEGGS. Stanford University. 1997.. L. AND ESPIN.. S. Convex Optimization. YARBERRY. “Principal Methods of Solution Multicriterial Optimization Problems”. 2000.A. November. Preconditioned Newton Methods for Simulation of Reservoirs with Surface Facilities. E. BRILL. S. BOHANNON.. Stanford. J. 1985. 14.com/carroll/ga. .. Course Reader for Class EE364: Convex Optimization with Engineering Applications. “Global Optimization Techniques in Gas Allocation for Continuous Flow Gas Lift Systems”. To be published in 2003. D.. 1970. TwoPhase Flow in Pipes... June. AND OVCHINNIKOV. R. V.1 May. E. 397404. L.J. “Reservoir Development and Design Optimization”. S.D. 1996. Texas. Fortran Genetic Algorithm Driver.. Canada 28 April .. 58 October. Soviet Automatic Control. MONROE. 1973.J. 2000.. Alberta. A. AND RODRIGUEZ. BYER. CA. Journal of Petroleum Technology.M. 6..D. 2001.S. AND VANDENBERGHE. “A Linear Programming Model for Optimum Development of MultiReservoir Pipeline Systems”. L. BREAUX... paper SPE 38895 presented at the 1997 SPE Annual Technical Conference and Exhibition held in Antonio. D.html. AND HORNE.M. Class Notes.A. BITTENCOURT. Ph. U of Tulsa.W. “Application of a Reservoir Simulator Interfaced with a Surface Facility Network: A Case History”. J. H.
D.. dissertation.. A. HELGASON.. Y. Vol. L.D. Sixth World Petroleum Congress.G.R. COLEBROOK. 133. H. CHEN. paper SPE 49123 presented at the SPE Annual Technical Conference and Exhibition held in New Orleans. 747760. 100. “An Equation of State Compositional Model”.. SPEJ. DUNS. J. CHARNES.. 1963.. Civil Engineers.A. “An Approach to RealTime Optimization of the Central Gas Facility at the Prudhoe Bay Field”. 24. Frankfurt.. DANTZIG. AND ROS.. Stanford University.N. 11.. 2730 September. AND COLOGGI.BIBLIOGRAPHY 157 CARROLL. 393403.. “Multivariate Optimization of Production Systems”. Stanford University. G. R. M.. Paper No. 1961. Princeton University Press.JR. October. R. Vol. AND HORNE. CA. R.. L. 1978. 1973. 1939. R. pp. Development of Techniques for General Purpose Simulators.M. Ph. 2001. DONACHIE. Management Models and Industrial Applications of Linear Programming. 1992. “Solving The Pipe Network Analysis Problem Using Optimization Techniques”.. pp. KENNINGTON J. Jo..III. Proc.H. Journal of Petroleum Technology. CLAY. pp. March.. W. Wiley. “Digital Program for Water Network Analysis”. J. C. ROOD.. Amer. Linear Programming and Extensions. COLLINS. 22. STOISITS. New York.F. AND LEBLANC. CAO. K.F.W. M. 7.... CA. Inst. Hydraulics Div. “Vertical Flow of Gas and Liquid Mixtures in Wells”. R. COOPER. COATS. No.. N. July. AND COOPER. J. I. 2002. H.C. 1998. “Turbulent Flow in Pipes with Particular Reference to the Transition Region Between the Smooth and RoughPipe Laws”. Louisiana. PRITCHETT. Civil Engrs. 1980. Management Science. Modeling GasLiquid Flow in Pipes: Flow Pattern Transitions and DriftFlux Modeling. R..C. Section 2. 782831.P. .. 363376.J. Soc.. 1963. MS thesis. Vol. B..
A. AND LO. 14421471. P. Inc.. FIACCO. BARUA. 245262. FITREMANN. Jo. P. Amer. 2527 June. DUTTAROY. 1981. London. 4356.. R. John Wiley & Sons. AND KATTAPURAM. Hydraulics Div. California. Civil Engineers. “A New Approach to GasLift Allocation Optimization”.. 20 (11). paper SPE 37447 presented at the 1997 SPE Production Operations Symposium. 1970. New York and Toronto. 2nd ed. paper SPE 38333 presented at the 1997 SPE Western Regional Meeting held in Long Beach. AND RANNEY.P. pp.BIBLIOGRAPHY 158 DUTTAROY. AND FOWLER. June 1921.. 1996..M.. AND MCCORMICK. 1974. Mgmt Sci. “Genetic Algorithms for Multiobjective Optimization: Formulation.K. SPE Reservoir Engineering.C. “Lexicographic Orders. Practical Methods of Optimization. P. March... Utilities and Decision Rules: A Survey”.Y.J.. Soc. EPP. “Non Steady Gas Liquid Flow in Pipes and Gaslifted Wells”. 1968. W. J.. “Studies of Offshore Reservoir with an Interfaced Reservoir/Piping Network Simulator”. “Efficient Code for SteadyState Flows in Networks”.. FANG. FONSECA. Nonlinear Programming: Sequential Unconstrained Minimization Techniques. 1997. Proc. FISHBURN.C. AND VEDRINES. “A Generalized WellManagement Scheme for Reservoir Simulation”. FLETCHER. “ComputerAided Gas Field Planning and Optimization”. 1985... Vol. held in Oklahoma City.. G. Okalahoma. J. 1993. R. May. Journal of Petroleum Technology. J.G. C... Proceedings of the Second International Conference on Multiphase Flow. A.. A. Discussion and Generalization”. New York... A. 96. John Wiley and Sons. K. K. 911 March.. AND FLEMING. K..V. S. 116120. 399406.S. 1997. AND HEIBA.M. EMANUEL. 1987. Proceedings of the .
“Multiparametric Linear Programming”. A. 64. Inc. Genetic Algorithms in Search. 416423. Optimization. P.S. 4. 2000. Multivafuate Production Systems Optimization in Pipeline Networks. SIAM J. W.. AND HORNE. T. August. pp. “Outline of an algorithm for integer solutions to linear programs”. . B. GILL. GILL.. May. 1989.E. 165171.H. MURRAY. AND SAUNDERS.. Production Optimization. AND WRIGHT. Optimization of Well Placement and Assessment of Uncertainty. B. User's Guide for SNOPT 5. No. 1958.. 1998. 2002. CA. Vol.com/manuals.. Mgmt Sci.3: A Fortran Package for Largescale Nonlinear Programming. M. GÜYAGÜLER.sbsisoloptimize. J. H.. 1993. GREEN. CA GOLDBERG.. MA.D.BIBLIOGRAPHY 159 Fifth International Conference on Genetic Algorithms. R. W. USA. Stanford Unvierstiy...E.. H. “FlowRelated Corrosion in LargeDiameter Multiphase Flowlines”. JOHNSON. 2001.htm.A.A. 1993. Practical Optimization. 18 (7). FUJII. and Machine Learning.San Diego. Schlumberger. M. GEOQUEST. AND NEDOMA. P.. M..V. 2002. B. Ph. Retrieved in October 2002 from http://www. Stanford University. W. Reading. San Mateo. MURRAY. 1972.E.. Disertation.E. pp. CA. Morgan Kaufmann Publishers. AND CHOI.E. Eclipse 300 Technical Description 2000A.S. 9791006. GÜYAGÜLER. December. GeoQuest. H. SPE Production & Facilities. 1981. 275278. T... AND SAUNDERS. SPE Production & Facilities. R. MURRAY.. A Presentation Given at a Stanford SUPRIO Seminar. GAL. FUJII. CA. 97100. Academic Press. 406421... 19. “Multivariate Optimization of Networked Production Systems”. GOMORY. AddisonWesley.. AND BYER. M. “SNOPT: An SQP Algorithm for Largescale Constrained Optimization”.J. Optimization... GILL. D. Thesis. P. 1995. Bulletin of the American Mathematical Society.
New Orleans. 1975. Texas. paper SPE 37529 presented at SPE International Thermal and Heavy Oil Symposium. M. K. 105130. CA.BIBLIOGRAPHY 160 GÜYAGÜLER. H.S.. HOLLAND. F. Networks.N. T. 2001. Bakersfield. . W.H.. AND LIEBERMAN. June. “Optimization of Well Placement in a Gulf of Mexico Waterflooding Project”. AND DUTTAROY.. of Tulsa. 1976.A. February. HALL. AND BARD. The Surrogate Worth Tradeoff Method. SPE Computer Applications... Multiobjective Optimization in Water Resources Systems. Elsevier Scientific. HEIBA.. Ann Harbor. 6. 1997. 2001. W. thesis. BARUA. 7th Edition. Y. 8893. paper SPE 63221 presented at the 2000 SPE Annual Technical Conference and Exhibition held in Dallas. Econometrica 29. HAIMES.. BARUA. H.. R. J. R. pp.. Louisiana. “Integration of a Field Surface and Production Newtork with a Reservoir Simulator”. New York.Y..N. Inc. HALL. September 30October 3.. McCrawHill Companies. 1961. pp. HONG.T. AND FREEDMAN. G. U. L. B.O.. GÜYAGÜLER.. S. NY. “Uncertainty Assesment of Well Placement Optimization”. AND HORNE. The University of Michigan Press.. 223237. “Hydraulic Network Analysis Using (Generalized) Geometric Programming”. M.A. Vol. 2000. 1975.J.. 1975. AND ROSENZWEIG. HORNE. S.... HARTLEY. Introduction to Operations Research. HEPGULER.. J.. A. B.J. “Nonlinear Programming by the Simplex Method”. HILLIER. New York. “Surface Facilities Management for Thermal Recovery Processes”. 1997.S. G. ROGERS.. paper SPE 71265 presented at the 2001 SPE Annual Technical Conference and Exhibition. Adaptation in Natural and Artificial Systems. 14 October. Effect of the Variables on Optimization of Continuous Gas Lift System.T.
New York.L. . S. AND VECCHI. Combinatorica. in Inequalities . 4.W. 1996. 18871892.B. 4598.. 481492. 220. 671680.B. AND BROWN. AND DOIG. 1976.H. KIRKPATRICK..D... J. 1981. Berkeley. October. “Mathematical Programming with Multiple Objectives: A Tutorial”. C..P. V. 1972..D. 497520. Journal of Petroleum Technology. AND HARROP. 186190. Academic Press. VIPEXECUTIVE Technical Reference. Econometrica 28. C. University of California Press. September. R. PAIDY. August. MACH. A. “The Impact of Corrosion on the Oil and Gas Industry”. dissertation. J. Wiley. NY. ed. A. H. 1984.. 1980. SPE Production & Facilities. S. Shisha (ed.. AND TUCKER. “A New Polynomialtime Algorithm for Linear Programming”. A.JR.). AND RAIFFA. “How good is the simplex algorithm?”.. KERMANI. D. E. Reservoir Parameter Estimation Constrained to Pressure Transient. 1960.W. LAND. KEENEY. CA. KARMARKAR. KANU. “An Automatic Method for Solving Discrete Programming Problems”.III. 159175. Stanford University.. Science.. M.. O.L. KLEE. in Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability ( J. KUHN.BIBLIOGRAPHY 161 HWANG. 531.. H. 1997.R. 1983.. LANDMARK.. N.. ) pp.P.G. 2001.L. Ph.. 1951. Performance History and Distributed Saturation Data. “Nonlinear Programming”. Landmark Graphics Corporation. 373395. Computing and Operational Research 7... New York. Neyman.W. M.. Decision with Multiple Objectives: Preferences and Value Tradeoffs.. “Optimization by Simulated Annealing”. K. LANDA. D. K. AND MINTY. “Economic Approach to Oil Production and Gas Allocation in Continuous Gas Lift”. GELATT. AND YOON.
1995. LITVAK.). Operations Research... HUTCHINS. LIU..P. Vol. . AND WOOD.D. M.R. “Optimal Hydrocarbon Reservoir Production Policies”.W. “Prudhoe Bay EField Production Optimization System Based on Integrated Reservoir and Facility Simulation”. LITVAK.. 1 (Aug. Texas. No. 58 October... COFFMAN. RINOOY KAN. Oklahoma.. L.. LEA.. JanuaryFebruary.E. J. AND SEPEHRNOORI. paper SPE 29125 presented at the 13th SPE Symposium on Reservoir Simulation held in San Antonio. L.. A. A. CLARK. (EDITORS). K. MACDONALD. R..O. 1.. R. FAIRCHILD.A. TX. 2002.BIBLIOGRAPHY 162 LASDON.E. 3643. P. J. MACDONALD. History of Mathematical Programming: A Collection of Personal Reminiscences. MCFARLAND.. M. paper SPE 38885 presented at the 1997 SPE Annual Technical Conference and Exhibition held in Antonio. Thirteenth Congress of the International Association for Hydraulic Research.JR. R. FOSSUM. LITVAK. M... 34. pp. paper SPE 77643 presented at the SPE Annual Technical Conference and Exhibition held in San Antonio.H. M.K. paper SPE 11583 prsented at the 1983 Production Operation Symposium held at Oklahoma City.T. “Integration of Prudhoe Bay Surface Pipeline Network and Full Field Reservoir Models”. B. R. 1991.L.G. “Gas Well Operation with Liquid Production”. AND SCHRIJVER..L.F.. LENSTRA. A. AND DARLOW.. North Holland. 4054. J.. “Surface Network and Well Tubinghead Pressure Constraints in Compositional Simulation”. 1215 February. K. A. Vol.C.W.L. AND KUEST L. DARLOW. 29 September – 2 October.J.... AND TIGHE..L. B. 1986.L. SKINNER. Texas. WOOD. “The Numerical Analyses of Water Supply Networks by Digital Computer”. C. J. February 27 – March 1.Sept..J. 1969. 1997.C. 1983.
DOTY. 1995. 1981. 1989..A. AND HOLDEN.W. J.). G. 2000.. “Development Planning and Management of Petroleum Reservoirs Using Tank Models and Nonlinear Programming”. AND SCHMIDT. 7. LASDON. MIETTINEN. M. AND HOLDEN. 5258. SPE Reservoir Engineering. Berlin. K. MURRAY. Optim.. Z. Comput. C. McGrawHill Book Co. 127142.. 2. in Recent Advances in Mathematical Programming (R. LOOSE.. REDNER. 3rd Edition. 1982. 270289. MURRAY. McGrawHill International Book Co. STARLEY. eds.. 1963.. Nonlinear Multiobjective Optimization. CA. W. MILLER. NISHIKIORI.W. J.K. K. D. MURTAGH. “A Projected Lagrangian Algorithm and Its Implementation for Sparse Nonlinear Constraints”. paper SPE 19711 . Appl.A.W. MarchApril. 1997. L. Advanced Linear Programming: Computation and Practice. 1996. “An Improved Method for Gas Lift Allocation Optimization”. “Use of Well Management Schemes for Rate Forecasts”... New York.. SpringerVerlag. Vol... MCFARLAND. B. 84117. N. “Sequential Quadratic Programming Methods for LargeScale Problems”. March 30April 1. New York. Boston. February... Wolfe...A. L. W. C. MICHALEWICZ. No.. Graves and P. R. Class Notes for MS&E 312: Optimization Algorithms. paper SPE 24071 presented at the Western Regional Meeting held in Bakersfield.. AND SAUNDERS.P. K.. MURTAGH.E.BIBLIOGRAPHY 163 LO. C.A. 32.. 1992. B.R. “Application of Linear Programming to Reservoir Development Evaluations”. “The Simplex Method for Local Separable Programming”.. Genetic Algorithms+Data Structure = Evoluation Programs. V. Kluwer Academic Publishers.. 89100. Mathematical Programming Study 16. pp. Z. 1984. California. pp. 1999. Stanford University.. LO.K. Operations Research.
Operations Research Letters. AND FLANNERY. 2002. PRESS. Eng. TX. paper AIAA 20023535. Stanford University. 1978. OYAMA. 2000. AND HORNE. 424431.. CA. J.D. L. Vol. RASHEED.. Single Phase and Multiphase Fluid Flow in Horizontal Wells. Numerical Recipes in Fortran 77.. Pet. Morgan Kaufmann. 1997a.. 27 (200). 7.. Trans. . Ph. dissertation. POWELL.. Cambridge University Press.. 1998. Cambridge University Press. TEUKOLSKY. M. B. held in Kuala Lumpur. PEACEMAN. “Determining the Value of Reservoir Data by Using Nonlinear Production Optimization Techniques”. 1998.D. POWELL.. M. PALKE. AIME. W.N. W. VETTERLING. Malaysia. “Interpretation of WellBlock Pressures in Numerical Reservoir Simulation”. D. 1416 April. Soc. AND SHOLNICK. 1997b. “Using Genetic Algorithms in Engineering Design Optimization with Nonlinear Constraints”.. M.T. D.N. AND LIOU. 1993.H. R. paper SPE 37428 presented at the 1997 SPE Production Operations Symposium held in Oklahoma City. (June) 183194.P. K. M. in Acta Numerica. “Multiobjective Optimization of A MultiStage Compressor Using Evolutionary Algorithm”. PADBERG. 1992. A.A. OUYANG. Proceedings of the Fifth International Conference on Genetic Algorithms. M.R. M. “An Adaptive Penaly Approach for Constrained Genetic Algortihm Optimization”. “Approximating Separable Nonlinear Functions Via Mixed ZeroOne Programs”. “Nonlinear Optimization of Well Production Considering Gas Lift and Phase Behavior”. R. 1998. Oklahoma. 911 March.BIBLIOGRAPHY 164 presented at the 64th Annual Technical Conference and Exhibition of the Society of Petroleum Engineering held in San Antonio. paper SPE 28047 presented at the 1997 SPE Asia Pacific Oil & Gas Conference and Exhibition. AND HORNE. The Third Annual Conference on Genetic Programming (GP98): Symposium on Genetic Algorithms (SGA98).. 15.R. October 811.W. “Direct Search Algorithms for Optimization Calculations”.. PALKE. pp. 253...J.... S.
. R.M. DC. STOISITS. R. MCCORMACK. dissertation. D.. paper SPE 52177 presented at the 1999 MidContinent Operations Symposium held in Oklahoma City. 4. paper SPE 24898 presented at the 67th Annual Technical Conference and Exhibition of the Society of Petroleum Engineers held in Washington.H. LAWAL. 1992. SCHIOZER. pp. Simulation Sciences. Z. 2.. 1962.. October 47. MACALLISTER. CA. 1986. pp. Journal of Optimization Theory and Applications. in Multiple Criteria Decision Making. “Application of Nonlinear Adaptive Modeling for Rigorous Representation of Production Facilities in Reservoir Simulation”. R. 1999. W.. 225239.. CA.E.. paper SPE 15657 presented at the 61st Annual Technical Conference and Exhibition of SPE..H.. N..J. 203217. Brea. 1994. SPENDLEY. SCHMIDT. “TwoPhase Flow Through Chokes”. 2001. . California. BRILL. F. STEUER. Simultaneous Simulation of Reservoir and Surface Facilities.. Stanford University.. North Holland.. STOISITS. R. E. HEXT. 441461. BLAIS.O. J. D. California.. Technometrics. USA. SIMSCI. “An Interactive Multiple Objective Linear Programming Procedure”. K.D. “Sequential Application of Simplex Designs in Optimisation and Evolutionary Operation”. Multivariate Optimization of Production Systems . USA. D. 1974. New York.. M. Stanford University. SALUKVADZE. AND PARK. CHAMPION. J.F. MS thesis. 1992.. Simulation Sciences. M.P. K.D. AND HIMSWORTH. 1977.C. PIPEPHASE User’s Guide.BIBLIOGRAPHY 165 RAVINDRAN. BATESOLE. Starr and M.E. G.R. Ph..J.. Oklahoma. New Orleans. A. Brea. Input Manual for NETOPT. SACHDEVA.R.F.the Time Dimension. 12. AND OGBE..S. “On the Existence of Solution in Problems of Optimization under Vector Valued Criteria”. Zeleny (Eds).. Vol. M. D. CRAWFORD. 1999. USA. SIMSCI. “Production Optimization at the Kuparuk River Field Utilizing Neural networks and Genetic Algorithms”. R. LA October 58.. 2831 March.D.
W. Helsinki School of Economics..E. SPE Production & Facilities. A.. Location and Trajectory”.. P. 1998. 2002b. Rio De Janerio. “A Model for Predicting Flow Regime Transitions in Horizontal and Near Horizontal GasLiquid Flow”. paper SPE 28467 presented at the 69th Annual Technical Conference and Exhibition held in New Orleans. New Orleans. Z.A. BRILL. Finland. WANG.. 127135. AND ARNOLD. P. LIU. USA. AND AZIZ. 1998. No. “Criteria for Sizing Multiphase Flowlines for Erosive/Corrosive Service”..J... 2002a. “Optimization of Nonconventional Well Type. AND DUKLER.E. S. Series A: 14.. XIAO.. 22. paper presented at the 17th World Petroleum Congress. September 2326.. Integer Programming.. paper SPE 77658 presented at the SPE Annual Technical Conference and Exhibition held in San Antonio. WOLSEY. M. Brazil. L.. K.. 29 September – 2 October. 1990.J. YETEN. J.. LA. “Multiobjective Linear Programming Model on Injection Oilfield Recovery System”. 2528 September... Texas. 7480. W. L.J. Vol. 47. K. AND AZIZ. Applic. 1994. K. 1994. AND SCHMIDT. J. M. WileyInterscience.. Computers Math... XIAO. AND SHI. DURLOFSKY. pp. LITVAK.BIBLIOGRAPHY 166 STOISITS. 2002. “Gas Optimization at the Kuparuk River Field”. R. O. P. “Optimization of Production Operations in Petroleum Fields”. 5.L. 65th ATCE of SPE. WANG. AIChE Journal.. SVEDEMAN. JIANG.F.P.. Y. Acta Academiae Oecnonomicae Helsingiensis.. 15 September. February. TAITEL. M.. paper SPE 77565 presented . B. K. “A Comprehensive Mechanistic Model for TwoPhase Flow in Pipelines”. 1976. WALLENIUS. LITVAK. “Optimization Production from Mature Fields”..L. paper SPE 20631. “Interactive Multiple Criteria Decision Methods: An Investigation and An Approach”.. 36. AND AZIZ.. Y. SHOHAM. SCHERER.E. S. J. 1975.
1996. 838857. Optimization. AND PALATNIK. San Antonio.. AANONSEN. AND ZHU..M. Vol. . I. ZHANG J.BIBLIOGRAPHY 167 at the 2002 SPE Annual Technical Conference and Exhibition. No. 1996. 3. B. Texas. SIAM J. 35 Sept. “Optimizing Reservoir Performance by Automatic Allocation of Well Rates”. ZAKIROV.. September 29October 2.. D. E. ZAKIROV. “A Bilevel Programming Method for Pipe Network Optimization”. Austria.. 6. Leoben. pp.. August. paper presented at the 5th European Conference on the Mathematics of Oil Recovery. S.
2 (The MILPII method was implemented in another standalone package. The LPI and LPII methods were implemented in VIPEXECUTIVE (Landmark.4. The speedup techniques described in Section 4.1 can be handled in this package. 168 . • • The SQP nonlinear rate allocation method described in Chapter 5.4. NETSO was developed at Stanford University and implements the following simulation and optimization methods: • • The network simulation and sensitivity coefficient computation procedure described in Chapter 3. The framework for multiobjective optimization of production operations described in Chapter 7. Pressure and velocity constraints described in Section 4. POP was developed at BP and implements the following optimization methods: • The LPI and LPII methods for rate allocation as described in Chapter 4. Both packages are written in Fortran 77. The MILPI method for rate allocation described in Section 4. 2001)). • The partial enumeration method and the genetic algorithm for well connection optimization described in Chapter 6.3.2 are implemented in this package.Appendix A Lists of Programs Two computer programs were developed in the course of this study: a standalone package called NETwork Simulation and Optimization (NETSO) and a package of subroutines called Production Optimization Package (POP).
Chapter 4 describes the simulation and sensitivity coefficient calculation procedures implemented in these program files. LIST OF PROGRAMS • The coupling procedure described in Chapter 8. In this appendix. 2001) and is not available for use as a standalone package. and an example input file for NETSO. The separable programming methods for rate allocation. Simulation and sensitivity coefficient calculation.f calpd.1 Description of NETSO NETSO is a standalone package written in Fortran 77 language.f simsol. Miscellaneous files that define the global data structure. allocate memory. and perform read and write operations.f. A. uses the NewtonRaphson method to solve the network problem. called from simsol. . initialize the data structure.APPENDIX A. The SQP method for rate allocation.1 Program Files The program files can be classified roughly into the following categories: • • • • Simulation and sensitivity coefficient calculation. A brief description of each group of program files is given below.f the driver for network simulation and sensitivity calculation. A. Thus POP is not described further.1.f calculates the BHP and its derivatives from the reservoir side. netsim. calculate the BHP and its derivatives from the surface side. This section presents a brief description of program files and libraries used in NETSO. 169 POP is a collection of subroutines that heavily depend on the data structure and subroutines of VIPEXECUTIVE (Landmark. calpu. we describe NETSO in more detail. simple instructions about how to compile and run NETSO.
Separable programming method for rate allocation.f by ADIFOR.f s_choke.f calpds. generated from choke. 1981) so that they can be solved by other standalone mixed integer programming solvers.f s_pipe. ntgpc. computes certain derivatives for sensitivity coefficient calculation.2.f g_fluid. generated from pipe.f simsen.f genmps.f g_choke.3.3.f by ADIFOR. fluid. implements the Sachdeva et al. reads performance curves from an input file. SNOPT implements a simplex method for solving an LP problem. implements the drift flux model for multipphase pipe flows.f isconv. sndrv. The linear programming problems are solved by SNOPT.f implements functions for calculating fluid properties.f rdpc. 170 computes the BHP and its derivatives from the surface side. Chapter 4 describes the optimization methods implemented in these program files.f calculates the Jacobian matrix and righthand side of Eq. checks convergence of the NewtonRaphson method. 3.f well. called from simsen.f snbbsol.f choke. implements a basic well model. .f by ADIFOR.f.28.APPENDIX A.f pipe. generated from fluid. LIST OF PROGRAMS fmseqn. computes the sensitivity coefficients.f implements the MILPI method described in Section 4. implements the Branch and Bound method described in Section 4. The model is written to a file in MPS format (Murtagh. generates model MILPI for the rate allocation problem.2. two phase choke model.f generates well performance curves. computes certain derivatives for sensitivity calculation. fmsnmilp.f g_pipe.f initializes the data structure for SNOPT.
netsof1. computes the nonlinear constraint functions and their derivatives (the Jacobian matrix).f except it is for Formulation P1.APPENDIX A. netsof2.f same use as ntdata. getdim. ntdata. Used for Formulation P2.3). somputes the objective and constraint functions and their derivatives. numbers the decision variables of an optimization problem (Formulation P2) or the system parameters for sensitivity coefficient calculation.f setups the lower and upper bounds for decision variables and constraints. setups the structure of the Jacobian matrix for an optimization problem. Chapter 5 describes the optimization methods implemented in these program files. Used for Formulation P2.3.f driver for nonlinear optimization of rate allocation problems with Formulation P2. Initialize the Jacobian matrix with appropriate values. setidat. Used for Formulation P2.f netobj. setopa. Used for Formulation P2. setbnd. same use as fsetidat except it is for Formulation P1.dat driver for nonlinear optimization of rate allocation problems with Formulation P1. Used for Formulation P2. same use as setopa.f initializes certain variables required by SNOPT. setgcon.f computes the derivatives of nonlinear constraint functions. It calls subroutines in files getdim.f fgcon.f fgetdim. fmscst. Used for Formulation P1.f setopa. LIST OF PROGRAMS 171 SQP method for Rate Allocation.f fsetopa. .f setups the data structure for SNOPT with Formulation P2 (Section 5.f fsetidat. same use as getdim.f netcon.f except it is for Formulation P1.f except it is for Formulation P1 (Chapter 5).f. Used for Formulation P2.f transposes the Jacobian matrix. setbnd.f and setidat. Used for Formulation P2. dummy subroutines required by SNOPT. Used for Formulation P2.f. fntdata.f setnpar.f calculates the sizes of a rate allocation problem .f.
hence it is especially effective if the objective or constraint functions are expensive to evaluate. a header file. defines some parameters.f pnt.2 Imported Libraries and Software NETSO uses several libraries and programs from other sources. a header file. or if the number of active constraints is nearly as large as number of variables.APPENDIX A. a header file. SNOPT is most efficient if only some of the variables enter nonlinearly. and perform read and write. a set of Fortran 90 files used to process the input file. initializes some variables. and common blocks. SNOPT requires relatively few evaluations of the problem functions..com/SNOPT.f ntcnst. 1998) is a tool for the automatic differentiation of functions written in Fortran 77.h.h ntbldt. main.sbsisoloptimize.. SNOPT was developed at the Stanford Systems Optimization Laboratory (SOL). SNOPT implements a sequential quadratic programming (SQP) method that obtains search directions from a sequence of quadratic programming subproblems. 1998) is a generalpurpose system for solving largescale optimization problems. checks if the input variables have been properly assigned.htm. More information on SNOPT can be found at http://www. defines the index variables. A. initializes the network structure. variables. data block file that initializes the index variables defined in net. ADIFOR will generate an augmented derivative code that computes the partial derivatives of all of the specified output . SNOPT (Gill et al.f allocm.f90 ntinit. defines memory pointers for working arrays. Given a Fortran 77 source code and a set of user specified dependent and independent variables. allocates memory for working arrays. • ADIFOR. ADIFOR (Bischof et al.h net.1. allocate memory.f inpobj.f chekinp.f the main program. These are described below: • SNOPT. These program files are used to define the global data structure. LIST OF PROGRAMS 172 Miscellaneous files.h rd*. initialize the data structure.
spc can be supplied.org/. a file named netso. The libraries reside in other subdirectories that are specified in the makefile.3 Compile and Run NETSO is written in Fortran77 language and was developed on a Silicon Graphics Origin 200 workstation.anl. 1999) is a standard linear algebra package. which can be freely downloaded from http://www. The Pacific Northwest National Laboratory (PNNL) Fortran library is a set of general purpose Fortran 90 routines. ADIFOR is a collaborative project between the Mathematics and Computer Science division at Argonne National Laboratory and the Center for Research on Parallel Computation at Rice University. which can be downloaded from http://www. though it is not required by SNOPT. The created executable file is named netso.1. An example input file is presented in the next section.htm. • LAPACK.dat is required. To run NETSO..spc are defined in Gill et al.pnl. All the source files reside in a subdirectory with a compile file named makefile. A.1.4 Example Input File The following is the input file for the example problem presented in Section 3. If the run is to solve a rate allocation optimization using SNOPT. an input file named input. input. More information on ADIFOR can be found at http://wwwunix.netlib.APPENDIX A. (1998). The keywords that can be used in file netso. To compile NETSO.mcs.gov/berc/flib/index. This package was used in NETSO for processing input files.gov/autodiff/ADIFOR/.7.dat C *************************************************** C Fluid and rock property . LAPACK (Anderson et al. LIST OF PROGRAMS 173 variables with respect to all of the specified independent variables in addition to the original results. simply key in the gmake command to execute the instructions specified in file makefile. A. • The PNNL Fortran Library.
00000E+00 0.0 0.0003 C VALVE NN NAME DIAMET DISC TEMPUP TEMPDW CTRTYP C 1/64th inch deg F deg F DIAM/DP/FIXED 1 C1 48 0.0 160 110 0.0003 12 P2 3.0 0.0003 16 P6 3.00000E+00 0.0 8000.0 110 110 0.99700 0.5 1800.00000 1.12000 0.00000E+00 0.0 0.87000 0.94000 0.8 WATER SGW 1.0 GAS sgpg 0.0003 7 TUB2 3.5 600.5 7500.00000E+00 0.5 300.60000 0.5 8000.5 1200.98000 0.00000 0.0003 3 TUB2 3.5 8000.60000 0.00100 0.00000 0.0 110 110 0.0 7800.00000E+00 0.41000 0.0003 9 TUB2 3.50000 0.40000 0.00000 0.0 110 110 0.0 110 110 0.20000 0.02100 0.19000 0.0003 15 P5 3.0 160 110 0.09000 0.00000E+00 C *************************************************** C Gathering system components C *************************************************** PIPES NN NAME DIAMET LENGTH DEPTH TEMPUP TEMPDW ROUGHN C inch foot foot F F in 1 TUB1 3.0 0.98000 0.00010 0.0003 5 TUB2 3.00000 0.0 8000.00000E+00 0.00000E+00 0.0 8000.0 110 110 0.0 0.00000E+00 0.0003 8 TUB2 3.5 7000.00100 0.0 110 110 0.00500 0.02000 0.0 110 110 0.5 8000.0 0.0 160 110 0.0003 10 TUB10 3.0 0.00000E+00 0.5 700.85000 0.0 160 110 0.0 110 110 0.5 8500.0 7000.00000 0.35000 0.5 4000.72000 0.0003 14 P4 3.13000 0.0 110 110 0.00000 1.07500 0.00000E+00 0.0 110 110 0.0 0.20000 0.0003 17 P7 3.0 160 110 0.0003 19 P9 3.0003 4 TUB2 3.0 8000.APPENDIX A.0 160 110 0.00000 0.00000 0.00000 1. LIST OF PROGRAMS C *************************************************** OIL API 31.92000 1.00000 0.70000 0.0003 13 P3 3.25000 0.0 0.0003 6 TUB2 3.00000E+00 0.00000 1.00000E+00 KROGTAB SG KRG KROG PCOG 0.5 60.0003 2 TUB2 3.0 160 110 0.0 7500.00000 0.0 8000.0 160 110 0.05000 0.0 KROWTAB SW KRW KROW PCOW 0.00000E+00 0.01000 0.00000 0.02500 0.00000E+00 0.30000 0.12500 0.70000 0.12000 0.0 8000.5 7700.00000 1.45000 0.0 7700.5 1000.5 7800.0003 20 P10 3.0 160 110 0.0003 21 P11 5.00000E+00 0.5 2000.00000 0.0 110 110 0.0 8000.0 160 110 0.00000 0.5 8000.0 0.00000E+00 1.85 110 110 DIAM 174 .0003 11 P1 3.0003 18 P8 3.0 0.5 400.00000E+00 0.0 500.
0E16 AUTO 3000 C C Nodes number.0 3900 160.0E15 4 H4 50 1.0E15 1.0E15 12 I2 50 1.0E15 1.0E15 1.0E15 1.0E15 1.0E15 1.0E15 1.85 0.15 4500 160.0E15 1.0E16 AUTO 2000 8 W2 8 0.0E15 1.0E15 1.4 0.0E15 1.0 8 B4 100 150 40 500 500 0. node name.0E15 5 H5 50 1.4 4800.3 3400. 160.0E15 1.0E16 AUTO 3000 2 W2 2 0.0E15 1.0E15 1.0E16 AUTO 0 6 W2 6 0.0E15 1.0E15 1.0E15 1.0E16 AUTO 0 3 W1 3 0.0E15 1.0E15 1.4 3400.0 5 B1 100 100 30 50 50 0.0E15 1.1 3800.0E15 15 I5 50 1.2 0.0E15 1.0E15 1.0E16 AUTO 0 7 W1 7 0.0E15 1.0E15 1.0E15 1.0E15 1.0E15 1.0E15 1.0E15 3 H3 50 1.0E16 AUTO 3000 5 W1 5 0.0E15 1.0E15 1.45 4200 160.0E16 AUTO 2000 4 W2 4 0.400 1 1000 1E23 1.6 0.0E15 1.0E15 1.0 3 B3 200 100 50 70 70 0.0E15 1.475 0 1000 1E23 1.0E15 1.0E15 1.0 9 B3 200 100 50 70 70 0.0E15 1.400 1 1000 1E23 1.475 0 1000 1E23 1.2 0.0E15 8 H8 50 1.0E15 1.0E15 7 H7 50 1.0E15 1.475 0 1000 1E23 1.85 110 110 110 110 110 110 110 110 110 110 110 110 110 110 110 110 110 110 DIAM DIAM DIAM DIAM DIAM DIAM DIAM DIAM DIAM 175 C C Reservoir well grid blocks BLOCK NN NAME DX DY DZ KX KY SW SG PRES TEMP 1 B1 100 100 30 50 50 0.0 10 B4 100 150 40 500 500 0.0E15 1.0E15 1.7 0.0 C C Well flow rate or pressure constraints can be imposed here C LFTTYP has following choices C AUTO: used in option OPTF1.0E15 14 I4 50 1. C QLFT is the upper limit C FIXED: QLFT is the fixed amount WELL NN NAME IBLOC DIAM SKIN PMIN QOM QGM QWM QLM LFTTYP QLFT 1 W1 1 0.0 6 B2 100 100 30 100 100 0.0E15 1.85 0.5 0.85 0.0E16 AUTO 0 9 W1 9 0.0E15 13 I3 50 1. automatic allocation of lift gas.0E15 1.0E15 9 H9 50 1.0E15 2 H2 50 1.0 4 B4 100 150 40 500 500 0.0E15 11 I1 50 1.3 0.0E15 1.0E15 1.400 1 1000 1E23 1.0 2 B2 100 100 30 100 100 0.0E15 1. and constraints C nodes flowrate and pressure constraints can be imposed here NODES NN NAME PMIN QOM QGM QWM QLM 1 H1 50 1.85 0.0E15 1.0E15 1.85 0.APPENDIX A.0E15 1.25 3500 160.0E15 1.0E15 1. LIST OF PROGRAMS 2 3 4 5 6 7 8 9 10 C2 C3 C4 C5 C6 C7 C8 C9 C10 48 48 48 48 48 48 48 48 48 0.0E15 1.85 0.0E15 1.400 1 1000 1E23 1.0E15 .0E15 1. 160. 160.475 0 1000 1E23 1.5 4900 160.85 0.0E15 1.400 1 1000 1E23 1.0E15 1.0 7 B3 200 100 50 70 70 0. 160.0E15 1.0E15 1.0E15 1.2 0.0E15 6 H6 50 1.0E15 1.85 0.0E15 10 H10 50 1.0E15 1.0E15 1.475 0 1000 1E23 1.25 0.0E16 AUTO 0 10 W2 10 0. 3900 160.0E15 1.0E15 1.0E15 1.5 0.0E15 1.
0E15 20 I10 50 1.0E15 1. pressure are continuous C at this node C =================================================== WELCON NN OUTNOD OUTCON OUTCNT STYP 1 1 1 PIPE SIEQ 2 2 2 PIPE SIEQ 3 3 3 PIPE SIEQ 4 4 4 PIPE SIEQ 5 5 5 PIPE SIEQ 6 6 6 PIPE SIEQ 7 7 7 PIPE SIEQ 8 8 8 PIPE SIEQ 9 9 9 PIPE SIEQ 10 10 10 PIPE SIEQ NODCON NN OUTNOD OUTCON OUTCNT STYP 1 11 1 CHOKE CON 2 12 2 CHOKE CON 3 13 3 CHOKE CON 4 14 4 CHOKE CON 5 15 5 CHOKE CON 6 16 6 CHOKE CON 7 17 7 CHOKE CON 8 18 8 CHOKE CON 9 19 9 CHOKE CON 10 20 10 CHOKE CON 11 21 11 PIPE CON 12 21 12 PIPE CON 13 21 13 PIPE CON 14 21 14 PIPE CON 15 21 15 PIPE CON 16 21 16 PIPE CON 17 21 17 PIPE CON 18 21 18 PIPE CON 19 21 19 PIPE CON 20 21 20 PIPE CON 21 22 21 PIPE HEAD C C Define boundary nodes NODCON NN STYP 22 BND C 176 .0E15 1.0E15 1.0E15 1.APPENDIX A.0E15 1.0E15 1.0E15 7.0E15 1.0E15 1.0E15 22 SEP 160 1. pressure are specified C this must be the root of a production tree C CON: connection/junction node.0E15 19 I9 50 1.0E15 1.0E15 1. there are two types of STYP C BND: boundary nodes.0E15 21 M1 50 1.0E23 1.0E15 1.0E15 1.0E15 17 I7 50 1.0E15 1. LIST OF PROGRAMS 16 I6 50 1.0E15 1.0E15 1.0E24 8.0E15 18 I8 50 1.0E15 1.0E15 1.0E15 C *************************************************** C Gathering system structure C *************************************************** C For wells there are three types of STYP C SEQ: fixed open well choke C SIEQ: adjustable well choke C SHUT: well is shutdown C for nodes.0E15 1.
1 C SOLVER SEN 177 .APPENDIX A. only used in option OPTF1 C**************************************************** LFTVOL 3e23 C *************************************************** C Optimization according to operation policies C *************************************************** INTVAR NVAR LBND UBND XINIT 1 0 4000 300 2 0 4000 400 3 0 4000 1200 4 0 4000 400 5 0 4000 500 6 0 4000 1000 7 0 4000 20 8 0 2000 400 9 0 2000 400 10 0 5000 1500 DVAR NVAR DEVTYP IDEV PAR LBND UBND XINIT 1 WELL 1 GLFT 0 6000 3000 2 WELL 2 GLFT 0 6000 3000 3 WELL 3 GLFT 0 6000 3000 4 WELL 4 GLFT 0 6000 3000 5 WELL 5 GLFT 0 6000 3000 6 WELL 6 GLFT 0 6000 3000 7 WELL 7 GLFT 0 6000 3000 8 WELL 8 GLFT 0 6000 3000 9 WELL 9 GLFT 0 6000 3000 10 WELL 10 GLFT 0 6000 3000 C HCON NCON DEVTYP IDEV PAR LBND UBND c 1 NODE 22 OIL 14000 14000 1 NODE 22 WAT 0 6400 2 NODE 22 GAS 0 900000 C GOAL 1 NOBJ DEVTYP IDEV PAR OBJ TARG COEF 1 FIELD 0 OIL MAX 2000 1 GOAL 2 NOBJ DEVTYP IDEV PAR OBJ TARG COEF 1 FIELD 0 GAS MAX 2000 1 GOAL 3 NOBJ DEVTYP IDEV PAR OBJ TARG COEF 1 FIELD 0 WAT MIN 2000 1 C C define the jobs JOBS RANK IGOAL MAX RELF 1 1 YES 0. LIST OF PROGRAMS C**************************************************** C Total gaslift volume.8 c 2 3 NO 0.
0. 6765. 629. 246. STB/d 0. 1513. 1053. 308. 91. 0. 314. 987. 487. 315. 284. 1725. 0. 294. 896. 876. 542. 252. 315. 2880. 1021. 376. 716. 748. 258. 651. 1345. 114. 575. 441. 505. 577. 632. 665. 42. 0. 86. 591. 324. 297. 235. 413. 201. 58. 768. 560. 276. 0. 590. 551. 274. 926. 861. 945. 727. 233. 890. 391. 481. 35. 1485. 306. 553.Appendix B Problem Data for the Gaslift Example in Section 4. 1740. 129. 225. 333.4. 675. 975. 240. 1038. 836. 1114. 548. 676. 315. 1507. 397. 0. 1568. 693. 0. 215. 157. 144. 285. 672.1: Gas injection and oil production rates for a set of 56 wells (from Buitrago et al. 1279. 348. 178 . 617. MSCF/d Oil Rate. 459. 790. 210. 97. 679. 857. 268.3 Table B. 1035. Well 1 2 3 4 5 6 7 8 9 10 11 12 Lift Gas Rate. 135. 0. 386. 290. 376. 281. 800. 656. 516. 66. 233. 60. 0. 914. 1295. 281. 279. 433. 307. 280. 268. 366. 450. 311. 1378. 1164. 1706. 1424. 908. 266. 71. 0. 605. 386. 510. 0. 792. 209. 756. 289. 530. 0. 651. 0. 687. 1054. 614. 541. 626. 958. 74. 287. (1996)). 652. 68. 287. 464. 0. 521. 379. 0. 1199. 814. 815. 352. 1027. 263.
2174. 1105. 413. 139. 111. 179. 323. 960. 1185. 105. 0. 469. 51. 975. 0. 215. 175. 944. 417. 86. 37. 986. 638. 1240. 354. 109. 0. 138. 144. 189. 210. 103. 86. 166. 1680. 0. 64. 127. 354. 210. 3317. 0. 909. 162. 594. 113. 119. 1578. 404. 590. 516. 1362. 224. 177. 1218. 110. 54. 1479. 342. 415. 3621. 1111. 495. 330. 0. 1669. 1062. 253. 237. 664. 105. 54. 839. 1112. 515. 1030. 136. 466. 94. 310. 1955. 232. 141. 129. 137. 63. 1001. 252. 181. 579. 476. PROBLEM DATA FOR THE GASLIFT EXAMPLE 179 Table B. 892. 265. 128. 458. 502. 204. 439. 303. 0. 289. 468. 186. 59. 133. 69. 207. 734. 0. 370. 0. 884. 705. 1646. 217. 12161. 801. 1618. 200. 1266. 1023. 56. 242. 420. 211. 728. MSCF/d Oil Rate. 40. 760. 263. 0. 1093.1: Gas injection and oil production rates for a set of 56 wells (cont). 309. 688. 209. 184. 682. 556. 185. 1972. 206. 0. 353. 1027. 1029. 282. 216. 1505. 277. 399. 948. 523. 0. 49. 1299. 0. 160. 686. 291. 894. 647. 455. 1213. 308. 943. 355. 317.APPENDIX B. 108. 128. 336. 509. 0. 109. 0. 1149. 0. 545. 648. 310. 0. 213. 0. 211. 251. 1420. 225. 1948. 0. 33. 496. 1079. 4485. 220. 31. 455. 1290. 295. 361. 354. 16. 310. 316. 270. 141. 0. 487. 235. 951. 709. 195. 267. 0. 198. 1044. 735. 118. 465. 1392. 370. 170. 0. 182. 201. 354. 1056. 897. 957. 541. 204. 1100. 221. 162. 4225. 618. 257. 281. 494. 0. 579. 1594. 184. 956. 0. 561. 189. 1273. 1240. 823. 907. 120. 308. 358. . 1081. 214. 0. 360. 725. 493. 112. 816. 322. 227. Well 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 Lift Gas Rate. 280. 251. 75. 234. 0. 389. 191. 982. 0. 754. 1929. STB/d 0. 1145. 153. 105. 361. 108. 549. 190. 307. 229. 308. 195. 460. 309. 391. 536. 551. 283. 357. 1151. 931. 69. 262. 1338. 213. 1317. 946. 283. 2297. 221. 319. 598. 234. 40. 347. 548. 353. 0. 253. 206. 195. 926. 329. 381. 65. 215. 772. 82. 767. 351. 195. 654. 335. 198. 975. 28. 188. 727. 717. 366. 131. 0. 1051. 302.
483. 0. 3980. 772. 63. 79. 1690. 198. 0. 797. 345. 0. 0. 0. 0. 83. 93. 0. 202. 3650. 757. 61. 1558. 197.1: Gas injection and oil production rates for a set of 56 wells (cont). 0. 1228. 1450. 74. 490. 1441. 67. 0. 28. 374. 61. 69. 251. 0. 186. 444. 51. 339. 2920. 1770. 0. 2009. 105. 2538. 361. 1418. 5096. 403. 14. 369. 27. 700. 0. 1633. 860. 1303. 0. 1135. 0. 1701. 2445. 0. 188. 0. 350. 55. 162. 465. 362. 87. 427. 282. 1149. 2169. 2224. 0. 505. 3100. 1304. STB/d 0. 98. 201. 0. 538. 0. 564. 84. 317. 0. 976. 0. 2650. 0. 1512. 0. 28. 330. 160. 666. 0. 31. 409. 116. 441. 0. 407. 422. 126. 56. 0. 0. 256. 0. 0. 22. 0. 201. 0.APPENDIX B. 35. 373. 0. 48. 199. 800. 0. 112. 207. 3488. 0. 397. 0. 373. 71. 2750. 131. 1590. 3342. 2340. 326. 146. 0. 38. 801. 0. 301. 2300. 0. 80. 59. 2466. 2671. 761. 0. 0. MSCF/d Oil Rate. 0. 246. 0. 232. Well 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 Lift Gas Rate. 91. 152. 310. 11. 300. 0. 83. 126. 87. 712. 34. 47. 1333. 184. 267. 21. 1891. 57. 0. 4899. 0. 53. 232. 248. 1268. 82. 186. 1484. 64. 636. 406. 207. 2317. 123. 0. 1559. 83. 167. 188. 72. 40. 0. 1346. 0. 2594. 1174. 200. 2849. 2536. 260. 0. 5360. 77. 394. 1300. 2596. 50. 2653. 0. 168. 278. 2982. 157. 337. 1655. 452. 372. 270. 50. 2830. 3042. 658. 403. 137. PROBLEM DATA FOR THE GASLIFT EXAMPLE 180 Table B. 31. 374. 29. 223. 194. 0. 7. 237. . 467. 0. 1247. 328. 292. 0. 174. 690. 1730. 52.
2: Gas injection and oil production rates for a set of 56 wells obtained from the MILPII method. Well 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Oil Rate STB/d 386 626 605 280 281 333 836 276 1568 233 957 510 108 302 648 361 892 1213 310 Gas Rate MSCF/d 672 450 521 0 0 157 235 268 1295 0 1048 800 0 186 598 460 0 282 0 Well 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 Oil Rate STB/d 391 455 214 944 1680 487 105 353 1044 184 308 354 654 211 209 216 204 64 282 Gas Rate MSCF/d 975 772 370 0 1030 0 120 0 0 0 0 0 131 280 0 195 108 0 157 Well 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 Oil Rate STB/d 207 27 372 200 337 397 83 50 441 483 232 0 0 0 267 0 0 452 Gas Rate MSCF/d 301 98 0 0 797 0 0 14 3042 2466 1418 0 0 0 1484 0 0 1770 .APPENDIX B. PROBLEM DATA FOR THE GASLIFT EXAMPLE 181 Table B.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.