IPTC-22155-MS

Production Optimization of a Large Network of Oil Wells with Electrical

Submersible Pumps as the Artificial Lift System

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

Arjun Roy, Senthilkumar Datchanamoorthy, Nagappa Sharad, and Sangeeta Nundy, Baker Hughes

Copyright 2022, International Petroleum Technology Conference DOI 10.2523/IPTC-22155-MS

This paper was prepared for presentation at the International Petroleum Technology Conference held in Riyadh, Saudi Arabia, 21-23 February 2022.

This paper was selected for presentation by an IPTC Programme Committee following review of information contained in an abstract submitted by the author(s).
Contents of the paper, as presented, have not been reviewed by the International Petroleum Technology Conference and are subject to correction by the author(s). The
material, as presented, does not necessarily reflect any position of the International Petroleum Technology Conference, its officers, or members. Papers presented at
IPTC are subject to publication review by Sponsor Society Committees of IPTC. Electronic reproduction, distribution, or storage of any part of this paper for commercial
purposes without the written consent of the International Petroleum Technology Conference is prohibited. Permission to reproduce in print is restricted to an abstract of
not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of where and by whom the paper was presented.
Write Librarian, IPTC, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax +1-972-952-9435.

Abstract
One of the primary tasks of a production engineer at an oil field is to maximize oil production from a field
comprising multiple wells, while respecting a multitude of constraints related to operational limitations
of components, component reliability, interaction between wells, environmental concerns, and operational
costs. This is a multi-objective multi-constraint problem involving multiple physics models that interact
with each other. Further, the total number of optimization parameters and constraints grows linearly with
the size of the field. This makes the problem computationally intensive for oil fields with hundreds of
wells and thus the direct use of a standard optimization algorithm will be inefficient. This paper describes
a computationally tractable and scalable approach to solve this problem.

Introduction
Oil production optimization from an individual artificially lifted well typically involves maximizing
production from the well. When using an electrical submersible pump (ESP), this could involve a
combination of running the pump at a higher frequency when possible or reducing the well head pressure
to increase production. Optimizing the production involves estimating the set points of a well model while
complying with any system- or operator-imposed constraints. When expanding the problem to a network of
wells, set points of each well model need to be determined which satisfy local as well as global constraints.
Production optimization for a network of ESP wells could target different objectives like maximizing
the oil production, minimizing the power consumption, and/or minimizing the water production. Recent
approaches for optimizing ESP wells consider different scenarios and follow different techniques. The
approach in [Kristoffersen 2017] creates a piece-wise linear approximation for oil production rate as a
function of frequency of the ESP and solves a network of two wells that finds the well-head pressure
for each well to match the inlet pressure of the flowline. In [Hoffman 2017], the optimization problem is
formulated as a mixed-integer linear problem by approximating the network model using piecewise linear
functions and applied on a network with 15 wells. With the help of reduced order model, the study in [Epelle
2019] implements a Real-Time Production Optimization (RTPO) approach for maximizing production from
naturally flowing, gas-lifted and Electrical Submersible Pump (ESP) wells.
2 IPTC-22155-MS

The complexity of the multi-well optimization problem considered here grows with the number of wells.
As the size of the field grows to hundreds or thousands of wells, the limiting factor becomes the speed with
which individual well models can be evaluated. The approach proposed in this paper uses a surrogate model
which is derived from the physics model of the ESP and the other sub-systems of the well. The surrogate
models are then used for optimizing the wells at a large-scale involving thousands of parameters and
constraints. The proposed approach also discusses some exploitable structures of the optimization problem
which can be used to easily solve the problem with hundreds or thousands of variables and constraints
on a small desktop computer in seconds or minutes. This is demonstrated in Section 6 where production

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

optimization is applied to a field of 500 ESP wells using a variety of constraints.

Problem Statement
A typical oil-field comprises of several wells that produce oil, gas and water with or without any
artificial lift. One of the primary tasks of a production engineer at an oil field is to maximize oil
production from a given field while respecting a multitude of constraints imposed by manufacturer
specifications, component reliability, interaction between wells, environmental concerns, cost of production,
etc. Traditional approaches to solve this problem have focused on using computational tools to determine
optimal operating points that maximize production of each well, while relying on the experience of the
production engineer to appropriately combine the answers from the individual wells to ensure that all
constraints are respected. It is clear though that multi-well optimization (MWO) is a multi-objective multi-
constraint problem involving multiple physics models that interact with each other. The number of variables
and constraints in the optimization problem increases linearly with the number of wells and there are
additional constraints related to limitations on storage and processing of produced oil, water and gas. The
additional constraints can be applied to clusters of wells connected to each other or to the entire field. This
makes the problem computationally intensive for fields with many wells organized into multiple clusters
and use of brute force approaches for this problem will be inefficient. This paper discusses an approach that
can scale with the number of wells and demonstrates that it can be used to compute optimal production for a
large field of 500 wells equipped with electrical submersible pumps (ESPs) as their artificial lift mechanism
within 2 minutes.
This paper focuses on oilfields with wells that have ESPs as their artificial lift mechanism. The wells in
the field can be grouped together into clusters. The definition of a cluster is flexible and needs to be decided
based upon the constraints imposed by the physical infrastructure connecting the wells and customer needs.
As a starting point, this paper considers a scenario wherein the wells belonging to a single cluster share
common topside-flow infrastructure and draw power from a common power source (as shown in the figure
below).
IPTC-22155-MS 3

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

Figure 1—Problem set-up

Thus, if there are wells in a cluster and the oil and water produced and the power consumed by the i-th
well are and respectively then the net oil and water produced, and the power consumed
by the cluster ( and respectively) is, simply, given by:

In a similar fashion, the oil and water produced, and the power consumed by a field ( and
respectively), comprising of m such clusters is given by:

The objective of MWO is to maximize , while minimizing Pfield and adhering to all imposed
constraints. Typical constraints in this case arise from the water handling/disposal capacity and power
consumption limits of a well, cluster or the entire field. In addition to these there are additional well-level
constraints arising from manufacturing specifications and well integrity (as detailed in the next section).
Finally, for any well, a given production rate can be achieved at multiple well head pressure and pump
frequency settings. We want to ensure that we achieve the maximum possible production from the field at
the least power consumption settings. Thus, the MWO problem can be loosely framed as:
4 IPTC-22155-MS

max and min subject to:

a. Field constraints
b. Cluster constraints
c. Well constraints

The exact nature of the constraints would be specific to a field. Since, each ESP well typically has two

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

control parameters, viz. the well head pressure (WHP) and pump operating frequency (ƒ), the number of
independent variables for the problem is twice the number of wells. Thus, it is evident that for large fields
with hundreds of wells the dimensionality of the problem is large. Thus, to ensure convergence of the
optimizer in a reasonable amount of time, we must ensure that the models of the individual wells run fast.
Our first step towards building a solution for production optimization at the field level involves developing
surrogate models for each individual well that can substitute for the physics based digital twins for each
well and reduce the time required to simulate the conditions in each well by orders of magnitude.

ESP well model

Figure 2. shows the setup of an ESP well. The well model comprises separate models for each sub-
component: reservoir, tubing, and ESP system. The component-level models are used to predict the change
in pressure and temperature across the component for a given flow rate, fluid and component properties and
boundary conditions (input/output pressure and temperature). The output conditions of each model serve as
the input conditions for the following model. In this manner, given the reservoir conditions, it is possible
to predict the conditions at the well head (or any intermediate point in the system). Equivalently, given the
conditions at the well head, it is possible to invert the model to predict the conditions at the reservoir, or
any intermediate point in the system.

Figure 2—ESP model

To simulate the entire well model, models that provide acceptable accuracy without being
computationally expensive are selected. The list of models used (not exhaustive) includes [Guo 2007]:

• Reservoir: Vogel's IPR for well inflow performance,

• Tubing: Hagedorn & Brown, and Beggs & Brill correlations,

• Pump: Pump curves describing pump performance (head, efficiency, and BHP, as a function of
flow rate and frequency), with corrections for viscosity and gas handling.
IPTC-22155-MS 5

The ESP model must also evaluate compliance with the pump constraints at a given operating point.
These constraints could include Dunbar/Turpin correlations for gas interference (Dunbar 1989 and Lea
1986), and flow rate between the rated limits. An additional requirement is placed on the fluid speed past
the motor to ensure adequate cooling of the motor.
Fluid production from an ESP well is typically controlled by two parameters, the well head pressure
(WHP) and the operating frequency of the variable speed drive. Increasing (decreasing) the WHP, increases
(decreases) the back pressure in the tubing and this effect propagates back to all the points in the well
thereby resulting in a decrease (increase) in drawdown. Increasing the frequency of the variable speed drive

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

increases the pump speed and thus results in an increase in production. Thus, production from an ESP
can be increased by decreasing the WHP and/or increasing the frequency of the variable speed drive. In
cases where the wells are not equipped with a variable speed drive, the only parameter that controls fluid
production is the WHP. This paper does not focus on these wells since they constitute a simpler subset of
wells with two control parameters.
From a modeling perspective, the independent variables for the ESP well are the liquid flow rate and
the pump operating frequency. Other parameters can be calculated by solving the physics models of the
different well components – these could comprise the WHP that the operator needs to set for achieving the
specified flow rate or the pump power consumed at this flow rate.
The aim of production optimization here is to determine the individual WHP and operating frequency for
a field of calibrated wells, subject to arbitrary user-specified well-level, pad-level, and field-level operational
constraints. Some of these are derived from manufacturer specifications, while others are needed to preserve
structural integrity of the wells and ensure downstream flow assurance. Some of the constraints that need
to be satisfied could include

• Bottom hole pressure (BHP) > minimum BHP limit

• WHP > minimum WHP limit

• Minimum pump inflow rate < Pump intake flow rate < maximum pump inflow rate (manufacturer
specs.)
• Dunbar and/or Turpin constraints (related to the maximum gas allowed at the intake of the ESP)

• Pump head < maximum allowable pump head

• Power consumed by the motor < maximum motor power limit (manufacturer specs.)

• Fluid velocity past motor > minimum fluid velocity past motor

• Motor temperature < maximum allowable motor temperature

A naïve approach to production optimization is to use the physics models listed earlier to estimate the
WHP and operating frequency of each well in a field, subject to all operational constraints. However, in
the current implementation of these models, the typical computational time required for solving all the
component models (for a moderately deep well) is approximately 100ms. Thus, for a field with hundreds
of wells, if the optimizer takes (say) 1000 iterations to converge, then the computational time required for
MWEO is on the order of hours. This can be reduced by parallelizing some steps in the optimization process,
or by compiling the models to highly optimized machine code.
In addition to this, the total number of constraints that the optimizer must deal with grows linearly with the
size of the field. As an example, for a field with 100 wells, the number of constraints can exceed 1000. This
adds to the complexity of the problem. In summary, large-scale MWEO is an optimization problem with
(typically) thousands of parameters and constraints, and in which the objective function and the constraints
are computationally expensive to calculate. To make the problem more tractable, surrogate models are
6 IPTC-22155-MS

constructed for the ESP well. Based on the above discussion, there are two desirable characteristics of such
surrogate models:

• running the surrogate models should be much faster than solving the physics models, and

• the surrogate models should be able to inform the optimizer if the chosen operating condition (at
any iteration) satisfies all the intrinsic well-level constraints.

Surrogate Models

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

The first step in surrogate modelling is calculating the boundary of feasible operation in the operator control
(well head pressure – pump operating frequency) space as shown in Fig. 3. This is a closed curve that is
determined by the intersection of the individual curves beyond which each constraint is violated. Any point
inside this boundary is a valid operating point that does not violate any of the constraints that the operator
must satisfy while operating the ESP well. Points on the boundary are values of operating parameters at
which one or more of the constraints are violated (Fig. 3). If a factor of safety is used for any of the
constraints, then a smaller feasible operating region is obtained.

Figure 3—Heat map of liquid flow rate at the surface as a function of

operating parameters – well head pressure and pump operating frequency.

Thereafter surrogate models are built for different output parameters, such as liquid flow rate, gas
production and power consumed at different operation points that will be used by the optimizer to compute
the objective function and user-specified constraints. This is done by running the well model for a random
sample of points inside the feasible operating region and then building a surrogate model for the output
parameters. The MWEO problem, thus, now becomes more tractable using the feasible region boundary
and the surrogate models built in it. In every iteration of the optimizer, we can now use a simple point-in-
polygon algorithm to evaluate if the choice of operating condition is feasible or not and assign a penalty
(based on its distance from the boundary) if it is not. We also use the surrogate models for faster computation
of objective function and the user-specified constraints.
Figure 3 shows a heat map generated using surrogate models built for different output parameters. The
value of these parameters (eg. liquid flow rate at the surface, pump power consumed, etc.) are shown in the
dialog boxes as the user hovers over each point. On the boundaries the violated constraints are also shown.
IPTC-22155-MS 7

The blue dot shows the current operating point, the green dot shows the maximum flow-rate operating point
that can be obtained by just increasing the frequency (and keeping the well head pressure constant), the
yellow dot shows the maximum flow-rate operating point by only reducing well head pressure and the black
dot shows the global maximum flow-rate operating point. The heat map shows the entire mathematically
feasible operating space of the well. Some parts of the feasible operating region might not be operationally
desirable (eg. high well head pressure regions). The user can limit the optimizer search space further by
putting in additional constraints.

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

Multi-well ESP Optimization
A brief description of the optimization problem is discussed in Section 2. The primary objective can
be defined as maximizing the summation of the oil production from all the wells. However, a given oil
production from any well can be achieved via different combinations of well-head pressure and pump
frequency. An additional objective function thus considered is to minimize total power consumption by the
field.
The manufacturing constraints at the well-level are discussed in section 3 and these are used to determine
the feasible operating region in the independent parameter space. Additional operational constraints at the
well, cluster and field levels are handled by the optimizer. The constraints or the limits arise from this
problem may be related to (i) oil production, (ii) water production and (iii) power consumption. All these
three parameters need to be within the lower and upper limit defined by the user or based on the physical
system. Also, these constraints can arise at field level and/or cluster level and/or at individual well level and
the optimal solution needs to satisfy all these constraints.
In this problem, the decision variables are surface liquid flow, pump frequency and pump power for
all the wells. The given well constraints on water, pump power and frequency can be merged with the oil
constraints by updating the lower and upper limits. Each of these constraints can be translated into set of
constraints involving only oil constraints and then applying union operator on these constraints will reduce
the number of constraints. For example, if we consider the pump power constraints, we can find the lower
and upper limit of the oil flow from the wells when the pump can operate within the given power limits.
Thereby, we can eliminate the well level constraints on water production, power consumption and pump
operating frequency. By setting the search limit of the decision variable (oil flow or surface liquid flow),
we can also eliminate the oil level constraints at the well level. Now the problem needs to be solved only
with field and cluster level constraints. The oil constraints and water constraints at field and cluster level are
linear relations and hence they satisfy the property of convexity. These constraints can be easily handled by
any standard solver. For including the power constraints in the problem, the boundary of the operating space
is used. The set of points on the boundary of the operating region is function of surface liquid flow and well-
head pressure as explained in Section 4. From the boundary plot, we observed that the relation between
surface liquid flow and the pump power is convex and hence the overall problem is a convex problem.
Since, our objective is to maximize the total oil flow rate, the optimal solution lies at minimum well-head
pressure. The power consumption at minimum well-head pressure changes the optimization problem to
finding minimum power for given flow. We can find the minimum power by interpolation operation on the
set of points on the boundary discussed in previous section. With the above discussed approach, the original
optimization is reformulated to standard form which can be solved used any optimization packages such as
Scipy [Virtanen 2020], Pyomo [Hart 2011] or CVXPY [Diamond 2016].
If the Jacobian matrix is not defined, then the solver will try to compute it numerically using the
forward difference. This might not be efficient, and it increases the overall time for finding the optimal
solution. Hence, defining the Jacobian for both objective and constraints helps the solver to find the
solution quickly. In our problem, since the objective is summation of the oil flow from individual wells
the Jacobian is a vector of length n and it is obtained by
8 IPTC-22155-MS

where qi and wci are the surface liquid flow and the water-cut for different wells and n is the total number of
wells. If the total number of constraints is nc, then the dimension of the Jacobian matrix for the constraints
will be (nc × n). The elements of the Jacobian corresponding to the oil constraints and water constraints can
be obtained directly using the water cut. For the Jacobian corresponding to the power constraints, computing
first-order partial derivatives of the different elements of the matrix is not easy as the relationship between
power, pump operating frequency and well-head pressure is non-linear and complex. Hence, we use the first-
order differences by choosing a small value of delta for flow and compute the elements of Jacobian. This is

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

an approximation, and its value may not be valid for all the regions of the operating space. The optimizer in
Scipy [Virtanen 2020] updates the Jacobian matrix automatically by recomputing its elements while moving
from one solution to another solution in the solution space. Since changing the flow in a particular well does
not affect the pump power of other wells, a lot of elements in Jacobian matrix will become zero. Hence, the
number of computations will reduce to n, instead of nc × n. When we have substantial number of wells, this
technique gives a significant advantage in computation time. By exploiting these properties of the system,
we can quickly get the Jacobian matrix and solve the optimization problem in a brief time.

Sample Results
The first set of results, shown in Fig. 4(a), (b) and (c) corresponds to a field comprising of 9 wells divided
into three clusters (three wells in each cluster). Each row represents a cluster of wells. The x-axis in each
plot represents the surface liquid flow rate (total liquid production) and the y-axis represents the power
consumed by each well.

Figure 4—Optimization results for different field level constraints for a 9-well 3-cluster field.

Each plot is a heat map in surface liquid flow rate and the red star represents the optimal operating point for
different field-level constraints on water produced and power consumed by the field. Fig 4(a) corresponds
to the case when there is no upper limit on the water produced and the power consumed by the field. Thus,
in this case each well is producing at its maximum limit (that does not violate any of the constraints in each
ESP well). In Fig. 4(b) the water produced by the field is constrained to almost half of the allowable value.
In this case, the wells with high water-cut are penalized the most and their production almost reduces to
zero. Production in wells with low water-cut is not impacted as much. Finally, Fig. 4(c) corresponds to the
case in which the power consumption by the field is also constrained. This reduces production in other wells
too. However, wells that require less power are not constrained as much.
IPTC-22155-MS 9

The next set of results corresponds to a large field comprising of 500 wells. The wells are organized
into 50 clusters of 20 wells each. This is thus a 1000 parameter optimization problem with 5000 well-level
constraints and potential 153 cluster-level and field-level constraints. Fig. 5 shows the current operating
condition of the field, such as oil, water and gas produced, and power consumed. It also shows the same
parameters for each cluster and well. As can be seen, the production potential of the field is currently higher
than the current production and thus there is scope for an increase in production.

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

Figure 5—Initial configuration of a 500 well 50 cluster field. The left panel shows production detail
of each cluster while the strip on the top shows production details of the field. Constraints imposed
by the user prior to each optimization run are set using the slide-bar. The bar-charts and pie-charts
show the production details graphically. Similar webpages for cluster and well-level constraints exist.

Fig. 6 is the result of optimization when there are additional constraints at the field level on the water
produced and power consumed at the field level. In this case the constraint on the power consumed is more
limiting and thus the optimizer hits this constraint first. In Fig, 7, the constraint on the water consumption
is more restrictive and that constraint is hit first. In either case, the optimizer takes less than a minute to
converge to the optimal solution. Thus, with this approach large scale field-level optimization problems can
be solved under a minute.
10 IPTC-22155-MS

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

Figure 6—Optimized production subject to field level constraints on
water consumption and power production (simulation time: 56 seconds)

Figure 7—Optimized production subject to field level constraints on

water consumption and power production (simulation time: 33 seconds)

Conclusion
The paper discusses a novel approach to optimize production in large oil-fields with ESP wells. This is
large-scale multi-objective multi-constraint problem involving multiple physics models that interact with
each other. As an example, for a field comprising 500 wells the number of independent variables is 1000
and the number of constraints to be accounted for is more than 5000. The approach used in this paper relies
on building multiple surrogate models for each well along with determining the feasible operating region
for each well. This reduces the optimizer search space, the number of constraints that the optimizer needs
to satisfy, and the time required to run the physics models during each iteration of the optimizer. The paper
IPTC-22155-MS 11

demonstrates that using this approach production optimization for large fields can be performed under a
minute using standard optimization libraries.

Acknowledgements
The authors would like to thank Bhaskerrao Keely for his contribution towards building physics based ESP
models.

Downloaded from http://onepetro.org/IPTCONF/proceedings-pdf/22IPTC/2-22IPTC/D021S040R003/2619381/iptc-22155-ms.pdf/1 by Curtin University user on 17 October 2022

References
Diamond S. and Boyd S., "CVXPY: A Python-embedded modeling language for convex optimization", Journal of
Machine Learning Research, Vol. - 17, Num. – 83, 1–5, 2016.
Dunbar, C. E. 1989. Determination of Proper Type of Gas Separator. SPE Microcomputer Users Group Meeting, Long
Beach, California, USA, 15–17 October
Epelle E. I. and Gerogiorgis D.I., Mixed-Integer Nonlinear Programming (MINLP) for production optimisation of
naturally flowing and artificial lift wells with routing constraints, Chemical Engineering Research and Design, Vol.
152, 2019, pp 134–148, ISSN 0263-8762.
Guo, B., Lyons, W. C., and Ghalambor, A., "Petroleum production engineering: A computer-assisted approach",
Burlington, MA: Gulf Professional Pub, 2007.
Hart, W. E., Jean-Paul W., and David L. W. "Pyomo: modeling and solving mathematical programs in Python"
Mathematical Programming Computation 3, no. 3 (2011): 219–260.
Hoffmann A. and Stanko M., Short-term model-based production optimization of a surface production network with
electric submersible pumps using piecewise-linear functions, Journal of Petroleum Science and Engineering, Vol. 158,
2017, pp 570–584, ISSN 0920-4105.
Kristoffersen Ø., "Production Optimization of a field with ESP lifted wells", Norwegian University of Science and
Technology, MS Thesis, Jun 2017.
Lea, J. F., Turpin, J. L., and Bearden, J. L. 1986. Correlation of Performance Data for Electric Submersible Pumps With
Gas-Liquid Flow. Proc., 33rd Annual Southwestern Petroleum Short Course, Lubbock, Texas, USA April, 267–281.
Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M., Reddy, T., Cournapeau,, "SciPy 1.0: Fundamental
Algorithms for Scientific Computing in Python", Nature Methods (2020), 17, 261–272. https://doi.org/10.1038/
s41592-019-0686-2.