Multivariate Optimization

of Production Systems
James A. Carroll III, * SPE, and Roland N. Horne, SPE, Stanford U.

Summary. Traditional analysis of Introduction not adopted by the petroleum industry until
production systems treats individual Mathematically, optimization involves find- the early 1950's. The bulk of the literature
ing the extreme values of a function. Given since then discusses linear programming
nodes one at a time, calculating a techniques applied to reservoir management
a function of several variables,
feasible, but not necessarily optimal, on a macrolevel. Aronofsky 2 provides a
Z=f(xt ,Xi,X;,· .. ,.t;:), .......... (1) detailed treatment of optimization methods
solution. Multivariate optimization
an optimization scheme will find the com- in petroleum engineering.
enables determination of the most
bination of these variables that produces an Aronofsky and Lee 3 developed a linear
profitable configuration, including all extreme value in the function, whether it is programming model to maximize profit by
variables simultaneously. The optimi- a minimum or a maximum value. Many ex- scheduling production from multiple single-
zation can also find the optimal re- amples of optimization exist. For instance, well reservoirs. Aronofsky and Williams 4
if a function gives an investor's expected extended that model to investigate the prob-
covery over a period of time, rather lems of scheduling production for a fixed
return on the basis of different investments,
than just at a single instant, as in tradi- numerical optimization of the function will drilling program and scheduling drilling for
tional methods. determine the mix of investments that will a fixed production schedule. Charnes and
yield the maximum expected return. This is Cooper 5 used linear programming to devel-
the basis of modem portfolio theory. If a op a reservoir model that minimized the cost
function gives the difference between a set of wells and facilities subject to a constant
of data and a model of the data, numerical production schedule. Attra et al. 6 devel-
optimization of the function will produce the oped a linear programming model to max-
best fit of the model to the data. This is the imize flow rate subject to several production
basis for nonlinear parameter estimation. constraints. All these models used linear
Similar examples can be given for network reservoir models that were based on mate-
analysis, queueing theory, decision analy- rial-balance considerations, and the reser-
sis, etc. voirs generally were assumed to be uniform
Historically, the petroleum industry has and single phase.
used optimization to allocate production Rowan and Warren 7 demonstrated how
through pipeline networks, to schedule to formulate the reservoir management prob-
transoceanic shipments of petroleum from lem in terms of optimal control theory.
supply sites to demand sites, to model Bohannon 8 used a mixed-integer linear
refinery throughput, and to determine the programming model to optimize a pipeline
best use of limited capital. Lasdon et al. 1 network. O'Dell et al. 9 developed a linear
point out that the production sector of the programming model to optimize production
petroleum industry has seen few successful scheduling from a multireservoir system.
applications of optimization methods. Huppler 10 developed a dynamic program-
The objective of this study is to demon- ming model to optimize well and facility de-
strate the effectiveness of multivariate op- sign, given the delivery schedule and using
timization techniques applied to the a material-balance reservoir model. Kuller
performance of hydrocarbon wells. The and Cummings 11 developed an economic
study consists of two primary phases: the linear programming model of production
development of a well model that determines and investment for petroleum reservoirs.
the economic benefit of a well and the op- Several investigators attempted to couple
timization of a well's economic benefit with numerical reservoir simulation with linear
multivariate optimization techniques. programming models. The idea is to use the
reservoir simulator to generate a linearized
Optimization Studies unit-response matrix to use with linear
In Petroleum Engineering programming models . Wattenbarger 12
Optimization methodologies are the focus of used this technique to develop a linear
operations research, a field that began in the programming model to schedule production
1940's. Operations-research concepts were from a gas storage reservoir. Rosenwald and
Green 13 used influence functions in a
• Now with Conoco Inc. mixed-integer linear programming model to
Copyright 1992 Society of Petroleum Engineers optimize well placement. Gajdica et al. 14

782 July 1992 • JPT

 ___Performance

Upstream!
Downstream
Pressure Pressure
at ~ Differential o
If:
,,
Solution at
Node Solution
Node Root of,'
Function

Flow Rate Q* Flow Rate Q*

FIg. i-Graphic examples of nodal analysis.

used aquifer influence functions to model tion that causes the pressure differential at jective function that uses all the decision
performance of gas reservoirs with water in- the solution node to vanish. This root will variables as input and determines the objec-
flux. Murray and Edgar 15 used influence be the stabilized flow rate that the well will tive variable. Examples of different objec-
functions in a mixed-integer linear program- achieve for the given well conditions. The tive functions are numerous: to maximize
ming model to optimize well placement and confusion with production optimization the present value of the production stream,
production scheduling. Extending previous arises because people typically combine no- to maximize the net present value of the
work, 16,17 See and Home, 18 demonstrated dal analysis with univariate optimization. well, to maximize the cumulative recovery
how to refine the unit-response matrix with For example, by holding all other parame- on an equivalent barrel basis, to minimize
nonlinear regression techniques. Lang and ters constant, a single parameter (e.g., tub- the cumulative GOR, to minimize the cu-
Home 19 expanded this work to consider ing diameter) is varied to see where the mulative WOR, to minimize the total invest-
dynamic programming techniques. extreme value of the objective function ment per equivalent barrel produced, to
Asheim 20 studied petroleum develop- (e.g., maximum stabilized flow rate) will maximize the rate of return of the well, etc.
ments in the North Sea by coupling reservoir occur. Having found the extreme value in
simulation and optimization. McFarland et the objective function, the parameter of Model Development
aPl used multivariate optimization to op- interest is optimized. Several investiga-
The first step of this study was to develop
timize reservoir production scheduling. tors 22 -27 provide more information on no-
a model of well performance. The model is
Note that virtually all previous research dal analysis, also known as production
systems optimization. composed of several major components,
attempted to model reservoir performance briefly described here. Carro1l 28 gives
by linearizing the reservoir performance and To optimize a function of a single varia-
ble, simply plot the variable vs. the objec- more details on the various components.
entering these data into some variation of
linear programming. The main focus in tive criterion and take the extreme value.
Only a single performance curve is required. The Reservoir Component. To replicate
every case was to model the reservoir per- the reservoir dynamics and inflow perform-
formance. None of the models endeavored Complications arise if a function of several
variables is to be optimized, particularly if ance in the well model, this study adopted
to optimize the well performance. a reservoir model developed by Borthne. 29
the variables are interrelated. To optimize
a function of two variables, a performance This model was originally designed to pro-
Modeling Well Performance vide a simple but functional reservoir com-
With Multivariate Optimization curve ofthe first variable plotted vs. the ob-
jective criterion is required for every dis- ponent for inclusion in a larger model. The
In the current literature, the terms "nodal crete value of the second variable. The model provides an accurate representation
analysis" and "production optimization" single performance curve required to op- of constrained reservoir performance that re-
are sometimes used interchangeably. This timize one variable balloons into a family quires minimal computer processing.
terminology can be somewhat confusing un- of performance curves when two variables Borthne's model is a black-oil model that
til one understands the distinction: nodal are optimized. If three variables are op- performs a generalized material-balance cal-
analysis finds the root of a function that timized, a family of performance curves is culation in concert with an inflow perform-
yields the stabilized flow rate of a well (Fig. required for every discrete value of the third ance relationship that is based on pseudo-
1); production optimization finds the root of variable. This manner of optimization rapid- pressure. The reservoir model is constrained
the gradient of a function that yields the ly becomes intractable because it suffers by both pressure and flow rate. In addition
maximum or minimum value the function from the curse of dimensionality-the work to specifying a minimum flowing well pres-
can achieve. To restate this important dis- involved increases rapidly as additional di- sure, two flow rates must be specified: a
tinction, nodal analysis finds a solution by mensions are added. It is at this point, when minimum flow rate and a maximum target
locating the root of a function; optimization the practical limits to sequential univariate flow rate. The model attempts to satisfy the
finds the optimal solution by locating the optimization are encountered, that the vigor target flow rate without violating the mini-
root of the gradient of a function. and potency of multivariate optimization can mum flowing well pressure.
The mathematical basis for nodal analysis be fully appreciated.
is relatively straightforward. Several sub- With multivariate optimization tech- The Tubing Component. The next stage in
models are coupled to provide a working niques, there is no limit to the number of the model development was to design and
model of a well (Fig. 2). The procedure of decision variables that can be optimized con- develop a component to model the mul-
nodal analysis is to find the root of the func- currently. The only requirement is an ob- tiphase flow through the vertical flowstring.

JPT • July 1992 783

 Input
Choke
Size Output
Oil Flow Rate
"The study consists of
Input Input
two primary phases: the Flowline Separator
development of a well Diameter Pressure

model that determines Input

Tubing
the economic benefit Diameter
of a well and the
optimization of a
well's economic benefit
with multivariate Reservoir Model
optimization Input Input
Reservoir
techniques." Perforation
Parameters
Parameters

Fig. 2-Schematic of a production-system model.

Three multiphase flow correlations were the recommended flow map superimposed. ary. Because the well model was for itera-
selected for the model: the modified Note that the map appears to be continuous tive use, the surface-choke component had
Hagedorn and Brown 30 correlation, the everywhere except at the boundary between to be applicable to both critical and sub-
Orkiszewski 31 correlation, and the Aziz et bubble flow and slug flow. Fig. 5 shows the critical flow conditions. Unfortunately, al-
al. 32 correlation. All three correlations are Orkiszewski gradient map. Here, the bound- though good correlations are available for
used extensively in the petroleum industry. aries between all four flow regions are very single-phase flow across chokes, good cor-
The Hagedorn and Brown 30 correlation is distinct. relations for multiphase flow across chokes
one of the most empirical correlations avail- are rare. Most of the available correlations
able; the Aziz et al. 32 correlation has the The Choke Component. After the reser- are strictly for critical flow.
strongest theoretical basis. voir and flowstring components were com- Nodal analysis handles the surface-choke
Figs. 3 through 5 show gradient maps of pleted, work began on a component to model discontinuity by avoiding it. In nodal anal-
the three correlations. Fig 3 shows the gra- the surface choke. We encountered a sur- ysis, the separator pressure is specified and
dient surface generated by the modified prising degree of difficulty at this stage. A then related to the pressure downstream of
Hagedorn and Brown correlation. The tran- surface choke is a binary device that oper- the choke by a horizontal flow correlation.
sition from the original Hagedorn and ates in either critical or subcritical flow. Then the pressure drop across the choke is
Brown correlation, representing slug flow, Also recall that, in critical flow, the flow obtained with the assumption that the choke
to the Griffith and Wallis 33 correlation, rate through the choke is independent of the is always in critical flow and with an em-
representing bubble flow, is very distinct. downstream pressure. Thus, a discontinuity pirical correlation. Note that by specifying
Fig. 4 shows the Aziz et al. correlation with occurs at the critical/subcritical flow bound- the downstream pressure and then calculat-
ing the upstream pressure, this procedure
manages to determine both the upstream and
downstream pressures in critical flow. If the
upstream pressure is specified in an attempt
to determine the downstream pressure dur-
ing critical flow, the most that can be deter-
mined is the maximum downstream
pressure. This is because flow rate through
the choke is independent of downstream
0'9
pressure during critical flow.
We rejected the common assumption that
0" the choke is always in critical flow as an op-
L;: tion for this well model and tried to devise
'S- 03 3
a method to handle both critical and subcrit-
Q'
ical flow. The Sachdeva et al. 34 choke
0.<<1.
model is one of the few correlations that at-
tempts to model both critical and subcriti-
cal flow. With a few modifications to the
algorithm, we incorporated this model 34
into our well model. During critical flow
through the choke, the downstream pressure
was determined by assuming that the choke
was at the critical/subcritical boundary. This
assumption enabled the critical pressure ratio
to determine the downstream pressure.
The Sachdeva et al. 34 model was devel-
oped and tested with the other well-model
components. Fig. 6 shows the present-value
surface generated by the Sachdeva et al.
Fig. 3-Hagedorn and Brown 30 gradient map. model as a function of choke and tubing di-

784 July 1992 • JPT

 " .. virtually all
previous research
attempted to model
reservoir performance
by linearizing the
reservoir performance
and entering these
da~a into some
variation of linear
programming. None of
the models endeavored
to optimize the well
performance. II

Fig. 4-Aziz et al. 32 gradient map.

ameters. Note that the criticallsubcritical duction profile for a given time step and a ent value of the revenue stream generated
flow boundary occurs at a choke diameter given project life. In all examples cited here, over the life of the project. The present value
of about 10 cm. Below this critical size, the we used a 6-month timestep and a IS-year of the revenue stream is obtained by dis-
surface is completely a function of the choke project life. counting the revenue stream over the life of
diameter. Above the critical choke diameter, the project back to the present. The discount-
the surface is completely a function of the Optimization of the ing scheme used in this model is relatively
tubing diameter. However, a physical con- Objective Function simple and easy to follow. The model ac-
straint is required: the choke diameter can- counts for both the company's cost of capi-
To perform optimization on the well model,
not be allowed to exceed the tubing
the production profile had to be transformed tal and the expected inflation rate. Both these
diameter. The line superimposed on the sur-
into an objective criterion. The objective parameters are assumed to be constant over
face in Fig. 6 represents this constraint.
Although a choke component that criterion used in this research was the pres- the life of the project.
modeled both critical and subcritical flows
was successfully incorporated into the well
model, the success was at great computa-
tional expense. In addition, several assump-
tions in the choke model (e.g., no mass
transfer occurring between phases) were
considered to be highly suspect. For these
reasons, we decided not to include a choke
component in the well model.

The Separation Component. The next

stage was to develop a component to model
surface facilities. We decided to model the
surface facilities with a two-stage separation
process. This process involves flashing the
well stream initially at the separator operat-
ing conditions and then flashing the resulting
liquid stream at stock-tank conditions. The
flash calculation implemented in this com-
ponent is based on the Soave-Redlich-
Kwong 35 equation of state. Note that the
multivariate optimization approach dis-
cussed in this work is not limited to two-
stage separation; it can be applied to any
number of stages.
Thus, in its final state, the well model con-
sists of a reservoir component, a production-
string component, and a surface phase-
separation component. For any combination
of variables, the model will give the pro- Fig. 5-0rkiszewski 31 gradient map.

JPT • July 1992 785

"With multivariate
optimization
techniques, there is no
limit to the number of
decision variables that 120.
can be optimized
concurrently. The only $million 100.
requirement is an -~
ec .
objective function that .~

uses all the decision EC. ~- I

l
, Choke
variables as input and 0.0 ::::==--t55E /Diameter
(cm)
8.0
determines the 10.0 12.0 14.0 .
16.0
18.0 . 20.0 22.0
objective variable." Tubing Diameter (cm)
24.0

Fig. 6-Constrained present-value surface of choke diameter vs. tubing diameter.

A company has two primary means of At the end of each time step , the well may include capital expenditure, tax pay-
securing capital from outside sources: the model gives the cumulative production that ments, royalty payments, labor cost, and
issuance of equity in the company and the occurred during a timestep. Specifically, the corporate overhead contribution. Because
assumption of debt. A company's cost of well model will yield a stream of n produc- the purpose of our research was to demon-
capital is typically determined as a weighted tion quantities, C 1 C2 , . .. ,Cn> where n is strate the effective application of optimiza-
average of the company's cost of debt and the number of timesteps occurring over the tion techniques and not to concoct an
cost of equity, expressed as life of the project. These production quan- elaborate economic model, we made pres-
tities are discounted back to the present by ent value the objective criterion. To use this
D E assuming constant production over the technique to design or analyze an actual well
C e =i D (1-TR ) - - + r E - - ,
D+E D+E timestep and discounting from the midpoint system, however, the objective criterion
of the timestep (Fig. 7). should be made to reflect the financial
................. (2)
For instance, if the timestep is 1 year, then ramifications as accurately as possible;
where iD = interest rate of debt, rE = return the cumulative production produced over the therefore, a criterion of net present value
on equity expected by shareholders, d= first year will be discounted as if it were sold should be used.
amount of long-term debt, E=amount of in one discrete quanitity at 6 months. Thus,
outstanding equity, and TR = corporate tax the present value of the production stream Results
rate. for n timesteps of length llt is
Typically, a company's executive man- Surface of the Well Model. Once the well
agement specifies the weighted-average cost
of capital to use in financial calculations.
v
P
= E\" CjPj
j=1 ((1+R)[(2j-I)/2]JlI
J, ... (4)
model was completed and the objective
criterion was decided upon, we generated
Knowing the cost of capital, C e , and the a surface map of the objective criterion as
inflation rate, i, we can express the discount where Pj is the real price of the production a function of two decision variables. Spe-
rate in real terms as at the j th timestep. cifically, the present value of the produc-
The present-value calculation determines tion stream was plotted as a function of the
(1+ Cd only the positive effects of the decision vari- separator pressure and the tubing diameter
R= -1. ............... (3) (Fig. 8). The surface appears to be a text-
(1+i) ables, namely the revenue stream. The con-
cept of present value is easy to conceptualize book example of an optimization surface:
Thus, the effects of both the cost of capital because it is simply the gross value of the smooth, elliptical contour lines bounding the
and inflation have been coupled into one dis- revenue stream. To change the objective extreme value (the bull's eye). Closer in-
count factor. Note that this discount factor, criterion from present value to net present spection of the surface, however, reveals a
the real discount factor R, may be used only value, the negative effects of the decision different story. As Fig. 9 shows, the view
with cash flows expressed in real terms. variables must be included. Negative effects at a closer level reveals a surface that is very

t=O t=T
at at at

Cl C2 C3 Cn-l Cn

Fig. 7-Revenue stream discounted from center of timestep.

786 July 1992 • JPT

 7.0

1. 80

6.0

0'
IL
2- 5.0

~
~
" . . the well model
VJ
VJ
~
4.0 consists of a reservoir
IL
c
component, a
.8
~ 3.0 production-string
0

""
U1
component, and
2.0 a surface phase-
separation component."
1.0

5.0 6.0 7.0 8.0 9.0 10.0 11.0 12.0 13.0 14.0 15.':

Tubing Diameter (em)

Fig. 8-Present-value surface of separator pressure vs. tubing diameter.

rough. The noisy data were masked by the shows, significant improvement was ical to the performance of both methods.
graphics software in Fig. 8 but are plainly achieved in the convergence to the maxi- With relatively large finite-difference inter-
visible in Fig. 9. mum value. Even when the starting points vals, the adverse effects of the noisy sur-
A better understanding of the surface is were not near the solution, the method al- face diminish. We conclude that finite dif-
obtained by examination of several uni- ways converged to the optimum solution. ferences one-tenth the size of the decision
dimensional profiles. Figs. 10 and 11 show Because analytic derivatives were unavail- variables produce good results.
two different profiles ofthe surface. Fig. 10 able, both the Newton-based methods deter- As an alternative to derivative-based
shows a constant separator pressure with mined the gradient and Hessian matrix with methods, we investigated the performance
varied tubing diameters. Clearly, the sur- finite-difference approximations. Because of the polytope method. The polytope al-
face is very rough and discontinuous along the surface of the function is noisy, obtain- gorithm is a direct-search method that uses
this dimension, and noise is present at the ing meaningful derivatives was difficult. The function comparisons instead of derivative
surface maximum. Fig. 11 is a profile of the size of the finite-difference interval was crit- information. Fig. 13 shows the performance
other dimension, obtained by holding the
tubing diameter constant and varying the
separator pressure. Fig. 11 shows that, in
this dimension, the surface is a very smooth
and continuous function. Thus, the surface
is noisy as a function of tubing diameter and
smooth as a function of separator pressure.

Performing Optimization on the Well

Model. Having developed the well model
and analyzed the surface properties of the
decision variables, we used numerical op-
timization techniques to maximize the func-
tion. Knowing that the surface is not smooth cv
'-
with respect to one of the variables, we did ::J
VI
not expect derivative-based optimization VI
methods to perform very well on the func- cv
'-
tion. We found that the unmodified New- Q..
ton method converged only when initial '-
o
estimates were within certain well-defined o
regions. See Carroll 28 for details. .....
o
Next, we tried to optimize the function
with a modified form of Newton's method,
obtained from Walter Murray, Stanford U.
Optimization Laboratory. Gill et al. 36
describe the algorithm, which is based on
modified Cholesky factorization. The mod-
ification is very similar to the Greenstadt 37
modification because, after it factors the
Hessian matrix to obtain the diagonal, it
forces the diagonal elements to be positive
and greater than some small value. The al- Tubing Diameter (em)
gorithm also uses a line-search procedure
to ensure that the optimal step length is taken
for a given search direction. As Fig. 12 Fig. 9-Rough features of present-value surface.

JPT • July 1992 787

 of the polytope algorithm. Measured strictly
on the number of function evaluations re-
quired to find the solution, the polytope
method took the equivalent of five iterations
of Newton's method. Specifically, the con-
2.031
vergence path shown in Fig. 13 required 45
function evaluations, which is equivalent to
five iterations of a Newton-based method
2.030 with a nine-point finite-differencing scheme.

Conclusions
~ The authors hope to impress on the reader
< 2.029

-
o
~
that multivariate optimization techniques can
be successfully applied to optimize produc-
tion systems and that multivariate optimi-
i5: 2.028 zation is an enlightened alternative to
Z sequential univariate optimization.
Optimization of hydrocarbon production
systems has several useful features.
2.027 1. Optimization may be used to optimize
various objective criteria, including max-
imizing the present value of the production
2.026 stream, the net present value of the well, the
cumulative recovery on an equivalent bar-
rel basis, and the rate of return of the well
and minimizing the cumulative GOR, the
cumulative WOR, and the total investment
7.4 7.6 7.8 8.0 per equivalent barrel produced.
Tubing Diameter (em) 2. Optimization facilitates the analysis of
parameters that vary in time. This is best
described by considering an example. For
a lO-year time span, multivariate optimiza-
Fig. i0-Closeup of present value vs. tubing diameter at optimum. Separator pres- tion will determine the optimal tubing di-
sure kept constant at 2921.11 kPa.
ameter to use each year-simply add 10
decision variables, one for each tubing size
for each year. As another example, if only
three tubing changes are allowed over the
lO-year span, multivariate optimization will
determine when the changes should be made
and what size the tubing should be. Attempt-
ing this with sequential univariate optimi-
2.0
zation is ill-advised.
Use of multivariate optimization in lieu
of sequential univariate optimization has sig-
nificant advantages.
~ 1.9 1. Multivariate optimization is not limit-
<
0 ed by the number of decision variables. An
...... unlimited number of decision variables may
iI"7
....... be optimized simultaneously. Sequential
CI)
:I 1.8 univariate optimization is effectively limit-
C; ed to one or two decision variables and be-
>... comes intractable when three or more
=
~
variables are optimized, especially when the
variables are interrelated.
~
t:l-. 1.7 2. Multivariate optimization converges to
the optimal values extremely quickly, ap-
proaching quadratic convergence.
3. Multivariate optimization avoids the
1.6 human input required by sequential univar-
iate optimization.
This study investigated several different
multivariate optimization methods. The sig-
nificant findings are below.
2 345 6 7 1. The unmodified Newton's method is
Separator Pressure (MPa) not viable. This technique is highly sensi-
tive to the initial guess of the decision vari-
ables. Instead, Newton's method should be
used in conjunction with a line-search proce-
Fig. ii-Present value vs. separator pressure for constant tubing diameter of 7.55226 dure and a modification to ensure a direc-
cm. tion of descent.
788 July 1992 • JPT
 2. For noisy functions, the polytope al-
gorithm provides an effective alternative to
7.0
derivative-based methods.
3. For noisy functions, the finite-differ-
ence approximations are greatly affected by 6.0
the size of the finite-difference interval. We
found a finite-difference interval of one- 0'
(L

tenth the variable size to be advisable. 2- 5.0

~
:J
Nomenclature '"'"
Q) 4.0
Cc = cost of capital, fraction G:
Ci = production quantities .sE' 3.0
D = long-term debt, dollars 0
"-
E = outstanding equity Q)
(f)

f = function 2.0
i = inflation rate, fraction
iD = interest rate of debt, fraction
1.0
j = timestep
n = number of timesteps 5.0 6.0 7.0 8.0 9.0 10.0 11.0 12.0 13.0 14.0 15.0

Nx,Ny = Aziz et al. 32 correlating Tubing Diameter (em)

parameters
Pj = real price of production at jth
time step
Q* = stabilized flow rate, BID Fig. 12-Convergence path of present-value map with modified Newton's method.
rE = return on equity expected by
9. O'Dell, P.M., Steubing, N.W., and Gray, Original Gas in Place," SPERE (Aug. 1988)
shareholders, fraction J.W.: "Optimization of Gas Field Opera- 985-94.
R = discount rate tion," JPT (April 1973) 419-25. 15. Murray, J.E. ill and Edgar, T.F.: "Schedul-
t = time, years 10. Huppler, J.D.: "Scheduling Gas Field Pro- ing of Production and Compression in Gas
tr = total time, years duction for Maximum Profit," SPEl (June Fields," JPT (Jan. 1979) 109-16.
At = time step size, years 1974) 274-79; Trans., AIME, 257. 16. Coats, K.H.: "An Approach to Locating New
11. Kuller, R.G. and Cummings, R.G.: "An Ec- Wells in Heterogeneous, Gas Producing
TR = corporate tax rate Fields," JPT (May 1969) 549-58.
onomic Model for Investment and Production
J;; = present value for Petroleum Reservoirs, " Am. Econ. Rev. 17. Crichlow, H.B.: Modem Reservoir Engineer-
x> = n vector of decision variables (1974) 64, 66-79. ing-A Simulation Approach, Prentice-Hall,
Z = objective variable 12. Wattenbarger, R.A.: "Maximizing Seasonal Englewood Cliffs, NJ (1977).
18. See, B.A. and Home, R.N.: "Optimal Reser-
Withdrawals From Gas Storage Reservoirs,"
Acknowledgments voir Production Scheduling by Using Reser-
JPT (Aug. 1970) 994-98.
voir Simulations," SPEl (Oct. 1983) 717-26.
We thank the Natl. Science Foundation and 13. Rosenwald, G.W. and Green, D.W.: "A 19. Lang, Z.X. and Home, R.N.: "Optimal Pro-
the Stanford U. Petroleum Research Inst. for Method for Determining the Optimal Loca- duction Scheduling Using Reservoir Simula-
tion of Wells in a Reservoir Using Mixed- tors: A Comparison of Linear Programming
providing the financial support and re-
Integer Programming," SPEl (Feb. 1974) and Dynamic Programming Techniques,"
sources to undertake this research. 44-54; Trans., AIME, 257. paper SPE 12159 presented at the 1983 SPE
14. Gajdica, R.J., Wattenbarger, R.A., and
References Startzman, R.A.: "A New Method of Match-
I. Lasdon, 1. et al.: "Optimal Hydrocarbon ing Aquifer Performance and Determining (To Page 831)
Reservoir Production Policies," Operations
Research (Jan. 1986) 34, 40-54.
2. Aronofsky, J .S.: "Optimization Methods in
Oil and Gas Development, " paper SPE 12295
available at SPE, Richardson, TX.
3. Aronofsky, J.S. and Lee, A.S.: "Linear
Programming Model for Scheduling Crude
Oil Production," Trans., AIME (1958) 213,
51-54.
4. Aronofsky, J.S. and Williams, A.C.: "The
Use of Linear Programming and Mathemat-
ical Models in Underground Oil Production, "
Management Sci. (1962) 8, 394-407.
5. Chames, A. and Cooper, W.W.: Manage-
ment Models and Industrial Applications of
Linear Programming, John Wiley & Sons,
New York City (1961) II.
6. Attra, H.D., Wise, H.B., and Black, W.M.: ,.0
"Application of Optimizing Techniques for
Studying Field Producing Operations," JPT
(Jan. 1961) 82-86; Trans., AIME, 222.
7. Rowan, G. and Warren, J.E.: "A Systems
Approach to Reservoir Engineering, Opti-
mum Development Planning," J. Cdn. Pet. ".0
Tech. (July 1967) 84-94. , •. 0
8. Bohannon, J.M.: "A Linear Programming
Model for Optimum Development of Multi-
Reservoir Pipeline Systems," JPT (Nov.
1970) 1429-36; Trans., AIME, 249. Fig. 13-Convergence path of polytope method.

JPT • July 1992 789

Multivariate 30. Hagedorn, A.R. and Brown, K.E.: "Ex-
perimental Study of Pressure Gradients Oc- Authors
Optimization of curring During Continuous Two-Phase Flow
Production Systems in Small-Diameter Vertical Conduits," JPT
(April 1965) 475-84.
(From Page 789) 31. Orkiszewski, J.: "Predicting Two-Phase
Pressure Drops in Vertical Pipes," JPT(June
Annual Technical Conference and Exhibition, 1967) 829-38; Trans., AIME, 240.
San Francisco, Oct. 5-8. 32. Aziz, K., Govier, G.W., and Fogarasi, M.:
20. Asheim, H.A.: "Offshore Petroleum Explo- "Pressure Drop in Wells Producing Oil and
ration Planning by Numerical Simulation and Gas," J. Cdn. Pet. Tech. (July 1972) 38-48.
Optimization," PhD dissertation, U. of 33. Griffith, P. and Wallis, G.B.: "Two-Phase
Texas, Austin (Dec. 1978). Slug Flow," J. Heat Transfer, ASME (Aug.
21. McFarland, J.W., Lasdon, L., and Loose, 1961) 307-20.
V.: "Development Planning and Manage- 34. Sachdeva, R. et al.: "Two-Phase Flow
ment of Petroleum Reservoirs Using Tank Through Chokes," paper SPE 15657
Models and Nonlinear Programming," Op- presented at the 1986 SPE Annual Technical
erations Research (March 1984) 32, 270-89.
Conference and Exhibition, New Orleans,
Oct. 5-8.
22. Brown, K.E.: The Technology ofArtificial Lift
35. Soave, G.: "Equilibrium Constants From a
Methods, PennWell Publishing Co., Tulsa
Modified Redlich-Kwong Equation of State,"
(1977) Chaps. 1 and 4.
Chern. Eng. Sci. (1972) 27, No.6, 1197-
23. Chu, M.H. and Evans, R.D.: "Production
1203.
System Optimization for Natural Flowing 36. Gill, P.E., Murray, W., and Wright, M.H.:
Water Drive Wells," paper SPE 11582 Practical Optimization, Academic Press, New
presented at the 1983 SPE Production Oper- York City (1981).
ation Symposium, Oklahoma City, Feb. 37. Greenstadt, J.L.: "On the Relative Efficien-
27-March I. cies of Gradient Methods," Mathematics of
24. Brown, K.E .. and Lea, J.F.: "Nodal Systems Computation (1967) 21, 360-67.
Analysis of Oil and Gas Wells," JPT (Oct.
1985) 1751-63. 51 Metric Conversion Factors
25. Golan, M. and Whitson, C.H.: Well Peifonn-
ance, IHRDC, Boston (1986). ft x 3.048' E-OI = m
26. Lea, J.F. and Brown, K.E.: "Production Op- in. x 2.54' E+OO = em
psi x 6.894757 E+OO = kPa
timization Using a Computerized Well
Model," paper SPE 14121 presented at the 'Conversion factor is exact.
1986 SPE Inti. Meeting on Petroleum Engi-
neering, Beijing, March 17-20. Provenance
27. Hunt, J.L.: "Production-Systems Analysis
for Fractured Wells," SPEPE (Nov. 1988) Original SPE manuscript, Multivariate Op-
608-14. timization of Production Systems, re-
28. Carroll, J .A.: "Multivariate Production Sys- ceived for review Dec. 9, 1991. Revised
tem Optimization, " MS thesis, Stanford U., manuscript received Apri1lO, 1992. Paper
Stanford, CA (1990). accepted for publication May 4, 1992. Paper
29. Borthne, G.: A Simulation Modelfor Oil and (SPE 22847) first presented at the 1991 SPE
Gas-Condensate Production Based on Mate- Annual Technical Conference' and Exhibi-
rial Balance and Infiow Peiformance Calcu- tion held in Dallas, Oct. 6-9.
lations, Norwegian Inst. of Technology,
Trondheim (1986). JPT

JPT • July 1992 831