Journal of Petroleum Science and Engineering 209 (2022) 109755

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

journal homepage: www.elsevier.com/locate/petrol

Stochastic simplex approximation gradient for reservoir production

optimization: Algorithm testing and parameter analysis
Jianchun Xu a, b, *, Wenxin Zhou a, b, Hangyu Li a, b, Xiaopu Wang a, b, Shuyang Liu a, b, Ling Fan a, b
a
Key Laboratory of Unconventional Oil & Gas Development (China University of Petroleum (East China)), Ministry of Education, Qingdao 266580, PR China
b
School of Petroleum Engineering, China University of Petroleum (East China), Qingdao 266580, PR China

A R T I C L E I N F O A B S T R A C T

Keywords: Production optimization is an effective technique to maximize the oil recovery or the net present value in
Stochastic simplex approximation gradient reservoir development. Recently, the stochastic simplex approximation gradient (StoSAG) optimization algo­
Production optimization rithm draws significant attention in the optimization algorithm family. It shows high searching quality in large-
Algorithm testing
scale engineering problems. However, its optimization performance and features are not fully understood. This
Computational cost
Reservoir numerical simulation
study evaluated and analyzed the influence of some key parameters related to the optimization process of StoSAG
including the ensemble size to estimate the approximation gradient, the step size, the cut number, the pertur­
bation size, and the initial position by using 47 mathematical benchmark functions. Statistical analysis was
employed to diminish the randomness of the algorithm. The quality of the optimization results, the convergence,
and the computational time consuming were analyzed and compared. The parameter selection strategy was
presented. The results showed that a larger ensemble size was not always favorable to obtain better optimization
results. The increase of the search step size was favorable to escape from the local optimum. A large step size
needed to match a large cut number. The increase of cut number was beneficial to increase the local search­
ability, but also made the algorithm more easily fall into the local optimum. The random initial position was
beneficial to find the global optimal point. Moreover, the effectiveness of the parameter selection strategy was
tested by a classical reservoir production optimization example. The final net present value (NPV) for water
flooding reservoir production optimization substantially increased, which indicated the excellent performance of
StoSAG by adjusting the key parameters.

1. Introduction et al., 2016; Lerlertpakdee et al., 2014; Van Essen et al., 2011; Liu and
Forouzanfar, 2018; Liu et al., 2018; Liu and Reynolds, 2020; Liu and
It is a broad consensus that the chance to discover new oilfields Reynolds, 2021).
remarkably decreases and further development of mature fields is The optimization process with uncertainty and unconstrained deci­
becoming increasingly attractive. Mature fields are defined as the oil sion variables can lead to very complex and large solution space,
fields, which reach the production peak or are in their declining mode, resulting in many possible solutions. Due to a large number of possible
after a certain production period (Babadagli, 2007). In order to maxi­ combinations of variables, it is not sufficient to determine the optimal
mize the economic benefit or reservoir recovery of the production life of set of variables based on the intuitive engineering judgement alone (Al
mature fields, reservoir engineers must plan optimally for production Dossary and Nasrabadi, 2016). Optimization algorithms are usually
parameters such as well location, well type, flowing bottom hole pres­ combined with reservoir simulators to solve this optimization problems.
sure (BHP) and interval control valves (ICV) (Yang et al., 2017).Reser­ The optimization target is usually to obtain the maximum expected net
voir production optimization is influenced by multiple factors such as present value (NPV) or maximum recovery ratio. (Liu and Reynolds,
geology, engineering and economics. It involves multiple decision var­ 2016). Numerical reservoir simulation is used to calculate objective
iables and constraints. Therefore, it is a complex and challenging functions such as cumulative oil production and recovery ratio (Chen
problem(Nasrabadi et al., 2012; Al Dossary and Nasrabadi, 2016; Wang et al., 2017b). Optimization algorithms are used to implement the

* Corresponding author. Key Laboratory of Unconventional Oil & Gas Development (China University of Petroleum (East China)), Ministry of Education, Qingdao
266580, PR China.
E-mail addresses: [email protected], [email protected] (J. Xu).

https://doi.org/10.1016/j.petrol.2021.109755
Received 11 August 2021; Received in revised form 7 October 2021; Accepted 29 October 2021
Available online 3 November 2021
0920-4105/© 2021 Elsevier B.V. All rights reserved.
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

optimization process. In order to achieve efficient optimization, a great robust optimization. This approach treats the reservoir numerical
deal of work has been carried out by previous researchers. Many opti­ simulator as a black box and calculates the approximate gradient by
mization algorithms have been proposed. These algorithms fall into two simulating the response of a single (deterministic) reservoir model to a
broad categories: derivative-free optimization algorithms and randomly generated well control vector, where each control vector
gradient-based optimization algorithms. The derivative-free optimiza­ contains the settings of all wells in all time steps. As the research with
tion algorithms used in the field of reservoir development optimization the EnOpt algorithm progressed, Fonseca et al. (2017) proposed an
include: genetic algorithms (GA) (Tabatabaei Nejad et al., 2007), par­ improved EnOpt algorithm with better performance, namely the sto­
ticle swarm algorithms (PSO) (Lee and Stephen, 2019), simulated chastic simplex approximation gradient formulation (StoSAG). Subse­
annealing algorithms (Tukur et al., 2019), neural network algorithms quently, the StoSAG algorithm was widely used in reservoir production
(Ali et al., 2015), covariance matrix adaptive-evolutionary strategy al­ optimization processes such as well location optimization. Hanea et al.
gorithms (Forouzanfar et al., 2016), imperialist competition algorithms (2017) used multiple model implementations to represent uncertainties
(Hosseini-Moghari et al., 2015), etc. These methods have the better in reservoir structure and phase distribution to account for geological
global searching ability. However, they require extensive reservoir uncertainty constraints, and used the StoSAG algorithm to optimize the
simulation runs and are not suitable for optimization of large reservoir target well and borehole trajectory. Chen et al. (2017a) developed a
models. The gradient-based optimization algorithms include: adjoint framework based on the lexicographic method. They used the
gradient algorithm (Zhang et al., 2010), ensemble optimization (Leeu­ stochastic-simple-approximation-gradient algorithm to maximize the
wenburgh et al., 2010), and steepest descent method (Liu and Reynolds, expected NPV and minimize the associated risk or uncertainty for robust
2021), etc. Gradient-based methods converge faster, but have the poor life-cycle production optimization. The channelized reservoir model and
global searching ability. They always fall into local optimum. the Brugge reservoir model indicated the effectiveness of risk measure
The adjoint method, which is subordinate to the gradient-based al­ selection. Lu et al. (2017a) proposed an efficient robust optimization
gorithm, is widely used in the calculation of model-based life-cycle algorithm using the steepest ascent method with StoSAG, where a large
optimization problems in which the gradients are obtained through number of representative realizations were considered. The results
adjoint techniques. The algorithm is highly accurate and computation­ showed that the algorithm not only improved the speed of convergence,
ally efficient, but it is difficult to apply in practice because it is an but also achieved a higher optimal NPV. Lu et al. (2017b) used a slightly
invasive method that requires access to the reservoir simulator source modified version of StoSAG for the bi-objective optimization of optimal
code and also requires running a large number of simulations (Fonseca well trajectories and optimal control settings for injection and produc­
et al., 2015a). To address this problem, researchers have worked to find tion wells to maximize the expected value of life-cycle NPV and mini­
alternative algorithms, one of which is a non-invasive approximate mize risk. They also presented the first graphical solution for the joint
gradient method known as integrated optimization method-ensemble well location and well control optimization of the StoSAG method
based optimization (EnOpt). It was firstly proposed by Lorentzen et al. considering the minimum well spacing constraint, and discussed the
(2006) and Nwaozo (2006). Then, Chen et al. (2009) used EnOpt for implementation of the computationally efficient decoupled StoSAG

Table 1
Many Local Minima functions.
No. Function Dim Range fmin
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ⎞
⎛
1 √ d
√1 ∑
(
d
) 2 xi ∈ [ − 32.768, 32.768] 0
1∑
f(x) = − a exp⎝ − b√ x2i ⎠ − exp cos(cxi ) + a + exp(1)
d i=1 d i=1

a = 20, b =√0.2 and c = 2π̅

̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
2 ⃒
f(x) = 100 ⃒x2 − 0.01x21 ⃒ + 0.01|x1 + 10|
⃒ 2 x1 ∈ [− 15, − 5] 0
x2 ∈ [− 3, 3]
3 ⎛⃒ ⎛⃒ √⃒̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅⎞⃒ ⎞0.1 2 xi ∈ [ − 10, 10] − 2.0626
⃒
⃒
⃒
⃒ x21 + x22 ⃒⃒ ⃒⃒
f(x) = − 0.0001 ⃒sin(x1 )sin(x2 )exp ⃒100 −
⎝ ⃒ ⎝ ⃒ ⃒⎠⃒ + 1⎠
⃒ ⃒
⃒ ⃒ π ⃒ ⃒
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
4 1 + cos(12 x21 + x22 ) 2 xi ∈ [ − 5.12, 5.12] − 1
f(x) = −
0.5(x21 + x( 2) + 2
2 √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
5 ⃒ x1 ⃒) √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ 2 xi ∈ [ − 512, 512] − 959.64
⃒ ⃒
f(x) = − (x2 + 47)sin ⃒x2 + + 47⃒ − x1 sin( |x1 − (x2 + 47)|)
2
6 sin(10πx) 1 x ∈ [0.5, 2.5] − 0.86901
f(x) = + (x − 1)4
2x
7 ∑ x2 ∏ d ( )
xi 2 xi ∈ [ − 600, 600] 0
f(x) = di=1 i − cos √̅ + 1, d = 2
4000 i=1 i
⃒ ⎛⃒ √⃒̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅⎞⃒
8 ⃒ ⃒ x21 + x22 ⃒⃒ ⃒⃒ 2 xi ∈ [ − 10, 10] − 19.209
⃒ ⃒
f(x) = − ⃒sin(x1 )cos(x2 )exp ⃒1 −
⃒ ⎝ ⃒ ⃒⎠⃒
⃒ ⃒
⃒ ⃒ π ⃒ ⃒
( )
9 ∑ m
1 ∑ d ∑ d 2 xi ∈ [0, 10] − 4.1558
2
f(x) = ci exp − (xj − Aij ) cos(π (xj − Aij )2 )
i=1
π j=1 j=1

( )T
3 5 2 1 7
d = 2 , m = 5 , c = (1, 2, 5, 2, 3) and A =
5 2 1 4 9
10 ∑d− xi − 1 2 xi ∈ [ − 10, 10] 0
f(x) = sin 2 (πw1 ) + 1
i=1 (wi − 1)2 [1 + 10 sin 2 (πwi + 1)] + (wd − 1)2 [1 + sin 2 (2πwd )] wi = 1 + ,for i = 1, ...,d
4
11 2 2 2
f(x) = sin (3πx1 ) + (x1 − 1) [1 + sin (3πx2 )] + (x2 − 1) [1 + sin (2πx2 )] 2 2 2 xi ∈ [ − 10, 10] 0
12 ∑ 2 xi ∈ [ − 5.12, 5.12] 0
f(x) = 10d + di=1 [x2i − 10 cos(2πxi )] , d = 2
13 sin 2 (x21 − x22 ) − 0.5 2 xi ∈ [ − 100, 100] 0
f(x) = 0.5 + 2
[1 + 0.001(x
⃒ 1 + x⃒2 )]
2 2
14 cos 2 (sin(⃒x21 − x22 ⃒)) − 0.5 2 xi ∈ [ − 100, 100] 0.29258
f(x) = 0.5 + 2
[1 + 0.001(x
∑d
2 + x2 )]
1 √2̅̅̅̅̅̅̅
15 f(x) = 418.9829d − 2 xi ∈ [ − 500, 500] 0
i=1 xi sin( |xi |) ， d =2
∑5 ∑
16 f(x) = ( i=1 icos((i + 1)x1 + i))( 5i=1 icos((i + 1)x2 + i)) 2 xi ∈ [ − 5.12, 5.12] − 186.7309

2
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

method for solving bi-objective optimization problems. Chen and Rey­ Table 2
nolds (2018) proposed an optimization framework based on the Bowl-Shaped functions.
augmented Lagrangian method and the newly developed StoSAG algo­ No. Function Dim Range fmin
rithm, and applied it to the simultaneous optimization of well control
17 f(x) = + x21 2x22
− 0.3 cos(3πx1 ) − 2 xi ∈ [ − 100, 100] 0
and WAG half-cycle lengths during CO2 water-alternating-gas injection 0.4 cos(4πx2 ) + 0.7
for the enhanced oil recovery process. Chen and Xu (2019) provided a 18 2 xi ∈ [ − 2, 2] 0
theoretical demonstration of the superiority of StoSAG over EnOpt, and f(x) =
provided real case numerical examples. The optimization results illus­ ( ( ))2
trated the advantages of the StoSAG algorithm over EnOpt algorithm. ∑d ∑d 1
+ β) xij − i ,d =
i=1 j=1 (j
Based on our previous research (Chen and Xu, 2019), the main work j

of this study was to test the influence of key parameters on the optimi­ 2 , β = 10
∑d ∑i
zation performance of the StoSAG algorithm. Based on the analysis 19 f(x) = 2 2 xi ∈ [ − 0
i=1 j=1 xj ， d =2
65.536, 65.536]
obtained from the work above, some parameters were selected and ∑d
20 f(x) = 2 2 xi ∈ [ − 0
applied to the optimization problem of a real reservoir case, and the i=1 xi ， d =2
5.12, 5.12]
optimization results were discussed. In this paper, the basic StoSAG al­ ∑d
21 f(x) = i+1
d =2 2 xi ∈ [ − 1, 1] 0
gorithm was shown in section 2, the algorithm testing results were i=1 |xi | ，
∑d
22 f(x) = 2 2 xi ∈ [ − 10, 10] 0
presented in section 3, and a reservoir production optimization example i=1 i xi
d =2 ，
∑ ∑d
was shown to indicate how to choose reasonable parameters to find a 23 f(x) = di=1 (xi − 1)2 − 2 xi ∈ [ − 4, 4] − 2
i=2 xi xi− 1 ，
better global solution in section 4. d =2

2. Optimization algorithm Table 3

Plate-Shaped functions.
As mentioned previously (Chen and Xu, 2019), when the real No. Function Dim Range fmin
gradient is not available, ensemble-based methods provide a convenient
24 f(x) = (x1 + 2x2 − 7)2 + 2 xi ∈ [ − 10, 10] 0
alternative to find the approximate gradient. Here, we consider a gen­
(2x1 + x2 − 5)2
eral optimization problem. The problem is to find u, which maximizes
25 f(x) = 0.26(x21 + x22 ) − 0.48x1 x2 2 xi ∈ [ − 10, 10] 0
the expectation over m for a nonlinear function J. u is a vector of the
26 f(x) = sin(x1 + x2 ) + (x1 − x2 ) − 2 2 x1 ∈ [ − 1.5, 4], − 1.9133
system inputs. m is the given model. The general formulation for StoSAG
1.5x1 + 2.5x2 + 1 x2 ∈ [ − 3, 4]
search direction is given by:
27 ∑ ∑ 2
f(x) = di=1 [( dj=1 xij ) − bi ] , d = 2, 2 xi ∈ [0, 2] 116
1 ∑Ne ( ( ) ) ( ( ) )
b = (8, 18)
dl = u l,j T + δ̂
u l,j δ̂
δ̂ u l,j J ̂u l,j , m − J(ul , m) (1)
Ne j=1
28 ∑ ∑ 2 xi ∈ [ − 5, 10] 0
f(x) = di=1 x2i + ( di=1 0.5ixi )2 +
∑d 4
( i=1 0.5ixi ) , d = 2
Where Ne represents the number of control perturbations. Control
perturbation ̂u l,j ,j = 1, 2, ...Ne at iteration l is generated from the distri­
bution N(ul , CU ). The superscript sign “+” on a matrix denotes the Table 4
Moore-Penrose pseudo-inverse, where δ ̂ u l,j − ul . CU is a predefined
u l,j = ̂ Valley-Shaped functions.

covariance matrix which is defined as: No. Function Dim Range fmin

⎛ ⎞ 29 x6 2 xi ∈ [ − 5, 5] 0
f(x) = 2x21 − 1.05x41 + 1 + x1 x2 +
CU1 0 ... 0 6
⎜ ⎟ x22
⎜ 0 C2 ... 0 ⎟
⎜ ⎟
CU = ⎜ U
⎟ (2) 30
(
x4
)
2 x1 ∈ [ − 3, 3], − 1.0316
⎜ ... ... ... ... ⎟ f(x) = 4 − 2.1x21 + 1 x21 + x1 x2 +
⎝ ⎠ 3 x2 ∈ [ − 2, 2]
0 0 ... CUn ( − 4 + 4x22 )x22
31 f(x) = (x1 − 1)2 + 2 xi ∈ [ − 10, 10] 0
∑d 2
Where CwU , w
= 1, 2, ..., n is the covariance matrix which is used to 2
i=2 i (2xi − xi− 1 ) , d = 2

impose a degree of smoothness; CU is a diagonal matrix and the diagonal 32 ∑d− 1

f(x) = i=1 [100(xi+1 − x2i ) +
2 2 xi ∈ [ − 5, 10] 0
elements are Cx × Cx in the following discussion where Cx is the (xi − 1)2 ], d = 2
perturbation size; ndenotes the number of the total well variables.
When using the StoSAG, the foundational StoSAG search direction Ridges/Drops and 6-Other. These test functions are listed in Tables 1–6.
can be used as follows: More details can be found from the website (http://www.sfu.ca/
~ssurjano/optimization.html). Dim means the dimensionality of the
1 ∑ Ne
( )( ( ) )
df − sto = u l,j − ul J ̂
̂ u l,j , m − J(ul , m) (3) function. The Range is the search region. The fmini is the optimal
Ne j=1
(minimum) value of the test function. The images of the test functions
are shown in Figure A1 in Appendix A. Since some of the test functions
3. Algorithm testing have more than two independent variables, the images of those func­
tions are not given.
In this section, the effect of five key parameters for StoSAG on Since the StoSAG algorithm calculates the approximate gradient by
optimization performance was tested by 47 test functions. These key randomly generated perturbations around the search point, randomness
parameters include the ensemble size to estimate the approximation exists for the optimization results. The search might fall into different
gradient (the number of control samples), the step size (the initial step local optimums. This is observed in the functional testing process,
size in steepest ascent method), the cut number (the maximum number especially for the function which has multiple local optimums, as shown
of step size cuts), the perturbation size, and the initial position. The in Fig. 1. On the other hand, when the function is a single-peaked
stochastic characteristic of the StoSAG algorithm was discussed. The test function, it eventually converges to the global optimum through
functions consist of six different kinds of functions:1-Many Local different search paths. This could be observed by the function value at
Minima, 2-Bowl-Shaped, 3-Plate-Shaped, 4-Valley-Shaped, 5-Steep different iteration, as shown in Fig. 2. Therefore, in the subsequent

3
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table 5
Steep Ridges/Drops functions.
No. Function Dim Range fmin
( )−
33 25 1 2 xi ∈ [ − 65.536, 65.536] 0.9980
∑ 1
f(x) = 0.002 +
i=1 i + (x1 − a1i )6 + (x2 − a2i )6
( )
− 32 − 16 0 16 32 − 32 ... 0 16 32
a=
− 32 − 32 − 32 − 32 − 32 − 16 ... 32 32 32
34 f(x) = − cos(x1 )cos(x2 )exp( − (x1 − π)2 − (x2 − π)2 ) 2 xi ∈ [ − 100, 100] − 1
( 2)
35 ∑d 2m ixi
2 xi ∈ [0, π] − 1.8013
f(x) = − i=1 sin (xi )sin ， m = 10 , d = 2
π

Table 6
Other functions.
No. Function Dim Range fmin

36 f(x) = (1.5 − x1 + x1 x2 )2 + (2.25 − x1 + x1 x22 ) + (2.625 − x1 + x1 x32 )

2 2 2 xi ∈ [ − 0
4.5, 4.5]
37 2
f(x) = a(x2 − bx21 + cx1 − r) + s(1 − t)cos(x1 ) + s 2 x1 ∈ [ − 5, 10] 0.39789
2
a = 1 , b = 5.1/(4π ) , c = 5/π , r = 6 , s = 10 and t = 1/(8π)
38 2 2
f(x) = 100(x21 − x2 ) + (x1 − 1)2 + (x3 − 1)2 + 90(x23 − x4 ) + 10.1((x2 − 1)2 + (x4 − 1)2 ) + 19.8(x2 − 1)(x4 − 1) 4 xi ∈ [ − 10, 10] 0
39 2
f(x) = (6x − 2) sin(12x − 4) 1 x ∈ [0, 1] − 6.0207
40 f(x) = [1 + (x1 + x2 + 1)2 (19 − 14x1 + 3x21 − 14x2 + 6x1 x2 + 3x22 )] × [30 + (2x1 − 3x2 )2 (18 − 32x1 + 12x21 + 48x2 − 36x1 x2 + 2 xi ∈ [ − 2, 2] 3
27x22 )]
41 4
∑ 3
∑ 3 xi ∈ (0, 1) − 3.8628
f(x) = − αi exp(− Aij (xj − Pij )2 )
i=1 j=1
⎛ ⎞ ⎛ ⎞
3.0 10 30 3689 1170 2673
⎜ ⎟ ⎜ ⎟
⎜ 0.1 10 35 ⎟ ⎜ 4699 4387 7470 ⎟
⎜ ⎟ ⎜ ⎟
α = (1.0, 1.2, 3.0, 3.2)T , A = ⎜ ⎜
⎟ , P = 10−
⎟
4⎜
⎜
⎟
⎟
⎜ 3.0 10 30 ⎟ ⎜ 1091 8732 5547 ⎟
⎝ ⎠ ⎝ ⎠
0.1 10 35 381 5743 8828
42 1 ∑4 ∑4 2 4 xi ∈ [0, 1] − 3.1345
f(x) = [1.1 − i=1 αi exp( − j=1 Aij (xj − Pij ) )]
0.839
43 ∑ 4 ∑6 6 xi ∈ (0, 1) − 3.3224
f(x) = − αi exp(− Aij (xj − Pij )2 )
i=1 j=1
⎛ ⎞
10 3 17 3.50 1.7 8
⎜ ⎟
⎜ 0.05 10 17 0.1 8 14 ⎟
⎜ ⎟
α = (1.0, 1.2, 3.0, 3.2)T , A = ⎜
⎜
⎟
⎟
⎜ 3 3.5 1.7 10 17 8 ⎟
⎝ ⎠
17 8 0.05 10 0.1 14
⎛ ⎞
1312 1696 5569 124 8283 5586
⎜ ⎟
⎜ 2329 4135 8307 3736 1004 9991 ⎟
⎜
− 4⎜
⎟
P = 10 ⎜ ⎟
⎟
⎜ 2348 1451 3522 2883 3047 6650 ⎟
⎝ ⎠
4047
( 8828 ( 8732 5743))1091 381
44 ∑ ∑d i
( )i
xj
2 2 xi ∈ [ − 2, 2] 0
f(x) = di=1 j=1 (j + β) − 1 , d = 2 , β = 0.5
j
45 d/4
∑ 4 xi ∈ [ − 4, 5] 0
f(x) = [(x4i− 3 + 10x4i− 2 )2 + 5(x4i− 1 − x4i )2 + (x4i− 2 − 2x4i− 1 )4 + 10(x4i− 3 − x4i )4 ]
i=1
d=4
46 m ∑
∑ 4
1 4 xi ∈ [0, 10] − 10.5364
f(x) = − ( (xj − Cji )2 + βi )− 1
, m = 10 , β= (1, 2, 2, 4, 4, 6, 3, 7, 5, 5)T
i=1 j=1
10
⎛ ⎞
4 1 8 6 3 2 5 8 6 7
⎜ ⎟
⎜4 1 8 6 7 9 3 1 2 3.6 ⎟
⎜ ⎟
C=⎜ ⎜
⎟
⎟
⎜4 1 8 6 3 2 5 8 6 7 ⎟
⎝ ⎠
4 1 8 6 7 9 3 1 2 3.6
47 1 ∑d 2 xi ∈ [ − 5, 5] − 78.3323
f(x) = (x4 − 16x2i + 5xi ) , d = 2
i=1 i
2

sensitivity analysis for each parameter, the statistical analysis was domain, we fix the initial search position at the midpoint; otherwise, the
employed. We control the parameters as constant and 50 repetitions of initial search position is set at the midpoint of the first quadrant of the
optimization were performed. The optimal value, average value, and coordinate system in the function search domain. In addition, a loga­
computational time are recorded. All the algorithms are programmed in rithm transformation is applied to each variable (Chen and Reynolds,
Matlab (2012a). Simulations are performed by a Core i5 PC with 3 GHz 2016). The variables will be truncated to the interval [− 7, 7] in the
processing frequency of CPU and 16 GB of RAM. We set Ne = 10, cut optimization process.
number = 5, step size = 1, Cx = 0.1 as basic values. The following
principle for the selection of the initial search position is adopted: if the
global optimum is not located at the midpoint of the function search

4
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Fig. 1. Optimization curves of functions falling into different local optimums in algorithm tests.

Fig. 2. Optimization curves for cases that converge to the global optimum by different paths in algorithm tests.

Table 7
Ne sensitivity-partial test data I.
Ne F3 Ne F4
best ave time(s) best ave time(s)

5 − 2.06250 − 2.0614 11.05 5 − 0.93625 − 0.93624 10.891

10 − 2.0624 − 2.0611 18.49 10 − 0.93625 − 0.93624 18.138
15 − 2.0625 − 2.0611 26.16 15 − 0.9375 − 0.93627 25.521
20 − 2.0625 − 2.0614 33.39 20 − 0.93625 − 0.93624 32.907

Ne F23 Ne F24
best ave time(s) best ave time(s)

5 − 1.9999 − 1.9988 12.813 5 1.27E-03 3.13E-02 10.664

10 − 1.9998 − 1.9987 21.09 10 1.49E-03 3.16E-02 17.974
15 − 1.9999 − 1.9988 29.184 15 3.73E-04 3.33E-02 25.286
20 − 1.9999 − 1.9989 37.474 20 3.18E-04 3.02E-02 32.647

Ne F30 Ne F33
best ave time(s) best ave time(s)

5 − 1.0315 − 1.027 10.874 5 1.0007 2.4639 21.688

10 − 1.0308 − 1.0267 18.174 10 0.99812 2.1308 33.583
15 − 1.0316 − 1.0272 25.689 15 1.0052 2.1713 45.645
20 − 1.0314 − 1.0271 33.002 20 0.99918 2.1718 57.681

Ne F39 Ne F40
best ave time(s) best ave time(s)

5 − 6.0179 − 1.1868 8.185 5 3.0213 3.1462 10.717

10 − 0.98554 − 0.98554 14.175 10 3.0071 3.14 18.19
15 − 6.0179 − 1.2875 20.289 15 3.0008 3.1589 25.357
20 − 0.98554 − 0.98554 26.246 20 3.008 3.1583 32.643

3.1. Sensitivity analysis of the ensemble size (Ne) an overall gradient of comparable quality to the exact gradient. They
showed that a larger ensemble size could improve the quality of the
A previous study is done to address the effect of the ensemble size approximate gradient. The effect of the ensemble size on the optimiza­
(number of perturbations) on optimization results by reservoir produc­ tion results was also discussed by Chen and Reynolds (2016). The results
tion optimization problem. Fonseca et al. (2015b) used the principle of showed that a large ensemble size was not favorable to obtain better
statistical hypothesis testing to quantify the ensemble size required for optimization results (a higher NPV at the final time). On the basis of

5
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table 8 optimum. The fluctuation happens due to the nature of the test function,
Ne sensitivity-partial test data II. which will be specifically analyzed in Section 3.3.
Ne F6 Ne F22 Table 8 gives the data for the second case of the test results. The
best ave time best ave time complete data are detailed in Table B2 in Appendix B. The optimization
(s) (s) results of these functions show that the global optimization ability de­
5 − 0.52336 − 0.01885 8.855 5 3.27E- 2.09E- 12.743 creases as Ne increased. In addition, over 60% of the listed functions are
03 02 valley functions. Here we take function F6 as an example. Fig. 4 shows
10 − 0.52336 0.014048 14.522 10 7.11E- 2.21E- 20.879 the function graph and the test curve of F6. From the figure we can see
04 02
15 0.025015 0.025015 20.573 15 1.63E- 2.23E- 29.099
that the algorithm also finds the global optimum when Ne equals 10,
04 02 while the average value increases compared to the results of Ne = 5.
20 0.025015 0.025015 26.786 20 1.98E- 2.26E- 37.19 When Ne is equal to 15 and 20, the optimal value and the average values
03 02 overlap and the optimal value is much larger than the global optimum.
Ne F25 Ne F29 And this indicated that the results of 50 repetitions of the optimization
best ave time best ave time all converge on the same local optimum. It seems that the increase of Ne
(s) (s) in this case enhances the local optimization-seeking ability of StoSAG,
5 1.07E-04 9.28E-04 10.662 5 1.37E- 1.07E- 11.188 while weakening the global search capability of it. Meanwhile, the
03 01 running time greatly increases due to the increase of Ne.
10 1.74E-05 8.53E-04 17.901 10 4.42E- 1.07E- 18.444
As can be seen, the impact of ensemble size (Ne) on optimization
04 01
15 4.10E-05 7.92E-04 25.176 15 5.52E- 1.07E- 25.996 mainly includes the following two aspects: under the same test condi­
04 01 tions, the size of Ne has a great impact on running time. The larger the
20 5.80E-05 7.89E-04 32.442 20 7.35E- 9.51E- 33.412 Ne, optimization consumes more time. Large size of Ne does not promote
05 02 the global optimum ability of the algorithm. The large size of Ne might
Ne F34 Ne F36 enhance the local search ability of the algorithm and increase the pos­
best ave time best ave time sibility of falling into the local optimum.
(s) (s)

5 − 0.99908 − 0.32614 10.674 5 7.89E- 2.52E- 10.831 3.2. Sensitivity analysis of the step size
06 04
10 − 0.99949 − 0.35067 18.019 10 1.60E- 2.52E- 18.28
05 04 In this section, the impact of the step size on the optimization per­
15 − 0.98902 − 0.35469 25.314 15 3.26E- 3.17E- 25.767 formance of StoSAG algorithm is analyzed. Similar to the sensitivity
05 04 analysis of ensemble size (Ne) in Section 3.1, we have presented partial
20 − 0.98748 − 0.29531 32.683 20 5.30E- 3.09E- 33.04 data for two cases of the test results in Tables 9–10, respectively. The
05 04
complete data are detailed in Tables B3-B4 in Appendix B. The functions
in Table 9 and Table B3 show the first case, i.e., the optimal value
these investigations, this section focuses on the verification of the
fluctuated with the increase of the step size. As a result, the average
findings above by test function examples. We evaluate each of the 47 test
value increases or fluctuated. As the step size increases, the probability
functions and then list the test results in Tables 7–8. Note that the time
of falling into local optimum increases, and the global optimization
given in the table is the total time required to perform 50 trials in the
capability could not be guaranteed.
optimization process.
From the test data in Table 9 and Table B3, it could be observed that
Table 7 gives partial data for the first case of the test results. The
in this case, the algorithm has better performance in seeking the opti­
complete data are detailed in Table B1 in Appendix B. These functions
mum when the step size is 1. That is, when the step size is 1, the algo­
present similar characteristics at different ensemble size (Ne). It could be
rithm obtains the smallest optimal value or the average value of the
observed that here Ne had no significant effect on the optimization re­
result is smaller when the optimal value is close. While for other step
sults. The optimal and average values in the tests fluctuated as Ne
sizes, the optimal value is difficult to find. The reason is that the cut
increased. From Table 7 and Table B1, we can see that except for F39,
number do not appropriately increase with the increase of step size. The
which fluctuated significantly with the change of Ne, the results of all
minimum search step length will increase as the step size increased,
other test functions fluctuate slightly around the optimization value
which could lead to a significant decrease in the local search ability of
obtained at Ne = 5. We take function F39 as an example for analysis. The
the algorithm. Therefore, it is necessary to set a suitable cut number for
function graph and test data curve of F39 are shown in Fig. 3. When Ne
large step sizes in practical optimization applications. Meanwhile, the
= 5 or Ne = 15, the algorithm partially converges to the global optimum.
benefit is found by increasing the step size in another case shown in
While Ne = 10 or Ne = 20, the algorithm converges all to the same local
Table 10 and Table B4. If the step size is greater than 1, the results may

Fig. 3. The function graph of test function F39 and its test results about Ne.

6
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Fig. 4. The function graph of test function F6 and its test results about Ne.

Table 9
Step size sensitivity-partial test data I.
Step size F1 Step size F18
best ave time(s) best ave time(s)

1 0.087224 2.0666 18.395 1 0.00049 0.023542 20.982

2 1.5073 5.9954 18.408 2 0.003099 0.067347 21.028
3 0.46192 9.3157 18.376 3 0.001592 0.17387 21.069
4 3.0127 9.7673 18.415 4 0.008042 0.10282 21.082
5 3.0533 13.4945 18.435 5 0.00147 0.17458 21.029

Step size F24 Step size F31

best ave time(s) best ave time(s)

1 0.000235 0.030078 17.95 1 0.000674 0.021456 20.869

2 0.009303 0.13313 18.031 2 0.002439 0.042005 20.829
3 0.036692 0.27693 17.966 3 0.004258 0.17475 20.886
4 0.013859 0.50313 18.001 4 0.000446 0.094642 20.869
5 0.008696 0.87561 18 5 0.001111 0.13032 20.879

Step size F33 Step size F36

best ave time(s) best ave time(s)

1 1.0054 2.0447 33.592 1 1.89E-05 2.60E-04 18.27

2 1.3538 2.3193 33.646 2 6.03E-05 1.14E-03 18.381
3 12.6705 12.6705 33.675 3 9.53E-05 4.56E-03 18.323
4 1.4637 2.3518 33.633 4 6.49E-04 8.16E-03 18.31
5 12.6705 12.6705 33.571 5 2.18E-04 9.41E-03 18.26

Table 10
Step size sensitivity-partial test data II.
Step size F5 Step size F27
best ave time(s) best ave time(s)

1 − 894.529 − 648.42 18.366 1 116.3208 116.3208 21.132

2 − 956.893 − 882.42 18.429 2 116.3208 116.3208 21.088
3 − 956.754 − 906.78 18.553 3 116.3208 116.3208 21.129
4 − 956.34 − 934.19 18.545 4 116.3208 116.3208 21.152
5 − 956.34 − 900.43 18.43 5 116.3208 116.3208 21.099

Step size F30 Step size F34

best ave time(s) best ave time(s)

1 − 1.0311 − 1.0271 18.255 1 -0.99775 -0.97976 17.314

2 − 1.0311 − 1.0157 18.14 2 -0.99437 -0.9344 17.228
3 − 1.0315 − 1.0169 18.11 3 -0.99742 -0.83284 17.251
4 − 1.0315 − 0.98223 18.132 4 -0.97626 -0.71415 17.245
5 − 1.0246 − 0.89889 18.04 5 -0.99884 -0.65933 17.796

Step size F39 Step size F41

best ave time(s) best ave time(s)

1 − 6.0179 − 1.0862 14.132 1 − 3.8628 − 3.8318 18.85

2 − 6.0179 − 1.0826 14.204 2 − 3.8628 − 3.8624 18.813
3 − 6.0179 − 6.0179 14.172 3 − 3.8628 − 3.862 18.84
4 − 5.7948 − 1.1744 14.202 4 − 3.8627 − 3.8614 18.86
5 − 5.7948 − 5.7948 14.193 5 − 3.8627 − 3.8603 18.822

be better than those obtained when the step size is equal to 1. Test to 1. The mean values for other step sizes are significantly lower
function F5 is most obvious. All four tests with step size larger than 1 compared to the test results for step size 1. This indicates that the in­
have significantly smaller optimal values than those with step size equal crease of step size is beneficial to get rid of local optimum near the

7
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table 11 size. Table 12 gives the test results before and after the improvement,
Step length and cut number. respectively. The symbol * indicates the optimized result after
Step size 1 2 3 4 5 improvement.
Cut number 5 6 6 7 7 Comparing the optimization results before and after the algorithm
improvement in Table 12, it is easy to see that the optimization results of
current search region. most functions become better. Moreover, in most cases with the improved
To verify the conjecture that the matching relationship between the strategy, the mean value of step size equal to 3 and 5 increases compared to
cut number and the step size affects the optimization results, we select step size equal to 2 and 4, respectively. The relatively large step size in­
some functions in Table B3 for testing. The initial search position and Ne dicates the large minimum search step size. This makes the local search
are kept constant, while the cut number is allowed to increase with the ability weak. In addition, the results in Table 12 are better for step size
step size to ensure that the minimum search step size is the same. Note greater than 1 for F17* and F24* than when the step size is equal to 1.
that when the above method to increase the cut number is used, the cut
number corresponding to step lengths equal to 3 and 5 is a decimal. We
3.3. Sensitivity analysis of cut number
set the cut number equal to 3 and 5 to be the same as step sizes 2 and 4,
respectively. Table 11 gives the cut number corresponding to each step
When testing the cut number, we continue to organize them

Table 12
Data of function test before and after algorithm improvement.
Step size F1 Step size F1*
best ave time(s) best ave time(s)

1 0.087224 2.0666 18.395 1 0.20025 2.7473 18.05

2 1.5073 5.9954 18.408 2 0.27274 4.2786 17.88
3 0.46192 9.3157 18.376 3 0.55118 3.4435 17.894
4 3.0127 9.7673 18.415 4 0.072014 1.6765 18.2
5 3.0533 13.4945 18.435 5 0.085571 2.4461 18.22

Step size F11 Step size F11*

best ave time(s) best ave time(s)

1 0.003303 0.052541 18.22 1 0.001382 0.05591 17.79

2 0.007225 0.13996 18.213 2 0.004348 0.067049 17.846
3 0.007556 0.44365 18.12 3 0.004949 0.10742 17.76
4 0.28896 0.62927 18.164 4 0.001404 0.06679 18.06
5 0.33254 0.81832 18.141 5 0.010617 0.097902 18.071

Step size F17 Step size F17*

best ave time(s) best ave time(s)

1 0.68083 2.8049 18.021 1 1.4458 3.0594 17.774

2 1.7986 8.4306 17.981 2 0.58138 2.8717 17.991
3 1.2217 18.6282 18.003 3 0.52637 5.7542 18.03
4 3.2538 38.4565 18.02 4 0.40823 2.8959 18.35
5 5.3777 72.7337 18.091 5 0.26082 4.0662 18.336

Step size F24 Step size F24*

best ave time(s) best ave time(s)

1 0.000235 0.030078 17.95 1 0.001933 0.030027 17.59

2 0.009303 0.13313 18.031 2 0.000168 0.035967 17.771
3 0.036692 0.27693 17.966 3 9.45E-05 0.07452 17.773
4 0.013859 0.50313 18.001 4 0.000635 0.037626 18.08
5 0.008696 0.87561 18 5 0.005305 0.043127 18.071

Table 13
Cut number sensitivity analysis-partial test data I.
Cut number F2 Cut number F17
best Ave time(s) best ave time(s)

5 0.044256 0.71038 18.135 5 0.53143 2.6568 18.09

10 0.044755 0.18078 19.696 10 0.00014566 0.0345 19.57
15 0.01493 0.039875 21.226 15 1.60E-06 2.83E-05 21.079
20 0.0023962 0.013621 22.867 20 4.11E-10 3.06E-08 22.548

Cut number F24 Cut number F30

best ave time(s) best ave time(s)

5 0.00042783 0.029729 17.973 5 − 1.0314 − 1.0271 18.011

10 2.74E-06 2.91E-05 19.45 10 − 1.0316 − 1.0316 19.619
15 2.54E-09 3.04E-08 20.959 15 − 1.0316 − 1.0316 21.062
20 3.50E-13 2.98E-11 22.475 20 − 1.0316 − 1.0316 22.59

Cut number F34 Cut number F39

best ave time(s) best ave time(s)

5 − 0.9994 − 0.23102 18.074 5 − 0.98554 − 0.98554 14.135

10 − 0.99986 − 0.99758 19.5 10 − 0.98632 − 0.98632 15.002
15 − 1 − 1 21.054 15 − 0.98633 − 0.98633 15.791
20 − 1 − 1 23.091 20 − 6.0207 − 1.2884 16.591

8
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table 14
Cut number sensitivity analysis-partial test data II.
Cut number F1 Cut number F12
best ave time(s) best ave time(s)

5 0.31725 3.0778 18.568 5 1.7353 11.5042 21.089

10 0.010722 0.43841 19.96 10 0.99553 13.2731 23.482
15 0.00048292 2.3867 21.516 15 0.99496 12.3374 25.706
20 1.48E-05 1.9887 23.044 20 0.99496 11.4221 28.105

Cut number F33 Cut number F38

best ave time(s) best ave time(s)

5 1.0402 2.0524 31.898 5 0.15487 0.81164 18.57

10 1.992 1.992 36.851 10 0.0004495 0.34399 20.279
15 0.998 1.9722 42.149 15 1.05E-05 0.21568 21.822
20 0.998 1.9324 47.464 20 0.00012523 0.4094 23.057

Cut number F41 Cut number F46

best ave time(s) best ave time(s)

5 − 3.8628 − 3.7545 18.86 5 − 10.4517 − 9.8149 19.19

10 − 3.8628 − 3.7082 20.659 10 − 10.5363 − 9.6396 21.168
15 − 3.8628 − 3.7855 21.978 15 − 10.5364 − 10.3831 23.091
20 − 3.8628 − 3.8009 23.149 20 − 10.5364 − 9.9805 24.897

according to 3.1. The partial test results for the first case are shown in of falling into local optimum.
Table 13, and the complete test data is detailed in Table B5 in The optimal values for most test cases listed in Table 14 and Table B6
Appendix B. In this case, the optimal and average values decrease are obtained before the cut number equals 20. Some functions are
further as cut number increases. continuously decreased as the cut number increases, while the average
As the cut number increased, the optimal and average values of most value fluctuates in the process. Take the function F6 as an example to
test functions are decreased. However, the results of test function F39 analyze. From the function image of F6 shown in Fig. 4, we can see that
show a significant decrease in the optimal value when the cut number the function is a multi-peaked function with a denser peak distribution.
increased from 15 to 20. From the average values corresponding to the During the testing, the optimal value is decreased when the cut number
different cut numbers, we could see that all the optimization processes increases from 5 to 10. The mean value of the results shows a fluctuating
fell into the local optimum except for the cut number equal to 20. trend. The algorithm does not find the global optimum during the test
Combined with the function images shown in Fig. 3, we believe that it is about the cut number, which may be caused by the effect of the step size,
mainly due to the enhanced local search capability of the algorithm as as described in Section 3.2. The increase in the cut number increases the
the cut number increased. The global optimum of function F39 is very chance of falling into local optimum due to the dense distribution of
close to the search starting point x = 0.5. The local optimum is close to peaks. As each local optimum is different, the mean value fluctuates
the starting point, which easily induces the algorithm to fall into the across the iterations. This implies that increasing the cut number
local optimum. When the cut number is small, it is easy to skip the region indiscriminately is not an appropriate strategy to increase the optimi­
that closes to the global optimal point in the searching process. This zation performance of the algorithm.
leads the algorithm easily to fail into the local optimum. Therefore, we
can see that the increase of cut number will improve the local search 3.4. Sensitivity analysis of the perturbation size
performance of the algorithm. It is beneficial for the global optimum of
some functions with small peak ranges but high peaks, although it takes Because the StoSAG algorithm calculates the approximate gradient
more time. based on randomly generated perturbations around the current search
The partial test results for the second case are given in Table 14, and point, the effect of the perturbation size on the accuracy of the random
the complete test data is detailed in Table B6 in Appendix B. In this case, approximate gradient is non-negligible. In this section, we analyze the
the algorithm may find the global optimum, but increases the possibility effect of the perturbation size on the optimization process by different

Table 15
Perturbation step sensitivity analysis-partial test data.
Cx F1 Cx F6
best ave time(s) best ave time(s)

0.01 0.20025 2.7473 20.218 0.01 0.025015 0.025015 14.885

0.1 0.42358 1.8433 19.558 0.1 − 0.52336 − 0.46852 14.979
0.5 0.1201 1.6905 19.305 0.5 − 0.52336 − 0.46852 14.956
1 0.44138 1.8136 19.299 1 − 0.52336 − 0.49045 15.389

Cx F28 Cx F29
best ave time(s) best ave time(s)

0.01 0.000103 0.01173 18.8 0.01 0.000108 0.071607 18.736

0.1 9.44E-05 0.011194 18.835 0.1 0.000101 0.016966 18.725
0.5 0.002282 0.011364 18.952 0.5 0.000106 0.035504 18.837
1 0.001585 0.011204 18.976 1 0.000383 0.034957 18.963

Cx F41 Cx F43
best ave time(s) best ave time(s)

0.01 − 3.8628 − 3.8163 19.792 0.01 − 3.0423 − 3.0421 21.173

0.1 − 3.8628 − 3.8318 19.546 0.1 − 3.0423 − 3.0421 21.121
0.5 − 3.8628 − 3.8627 19.616 0.5 − 3.0423 − 3.0384 21.294
1 − 3.8628 − 3.8627 19.604 1 − 3.0423 − 3.0372 21.167

9
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table 16
Initial position sensitivity analysis-partial test data.
Initial search location F5 Initial search location F15
best ave time(s) best ave time(s)

Fixed − 894.4586 − 657.4742 17.904 Fixed 118.8354 310.6765 21.099

Random − 956.8444 − 712.2101 18.564 Random 0.0052736 247.9856 21.335
Initial search location F41 Initial search location F43
best ave time(s) best ave time(s)

Fixed − 3.8628 − 3.7854 18.99 Fixed − 3.0423 − 3.0421 20.838

Random − 3.8628 − 2.6606 19 Random − 3.0423 − 3.0421 20.739

types of test functions. The partial test results are given in Table 15
below, and the complete test data is detailed in Table B7 in Appendix B.
In general, we believe that a smaller perturbation step corresponds to
a better gradient approximation accuracy. This is observed in the test
results of some functions in Table 15, such as test functions F10, F13, etc.
However, more results present that the smallest perturbation step does
not necessarily correspond to the best optimization result. The effect of
randomness is unavoidable. In the experiments, we conducted several
tests on function F6. The results indicated that better results might be
achieved for a larger Cx. Combined with the function image of F6, we
can see that a small Cx is easy to fall into the local optimum points near
the initial search location. A larger Cx is easier to escape from these local
optimum points and find the global optimum.
From the above results, it could be stated that the smaller the Cx is,
the stronger the local search ability of the algorithm has. Enhancing the
local search performance of the algorithm by increasing the cut number
requires more time cost, while the change of Cx has almost no particular
effect on the running time.

3.5. Effect of initial search location on search results

Fig. 5. The horizontal log permeability distribution and well location distri­
From the previous discussion, it is clear that the results of the optimal bution. (unit: mD).
solution for complex multi-peak functions are affected by various factors.
For an unfamiliar objective function, the optimal values of these param­ optimization and the test function optimization, we repeated the above
eters are difficult to be determined. Specifically, when the searching test process for the actual reservoir case to observe the effect of different
domain of the test function is large, the global optimum may be far from parameters on optimization results. The reservoir model is a single-layer
the initial search position. So by changing the initial position randomly is a reservoir with 25 × 25 × 1 grids. The grid block size is Δx = Δy = Δz =
feasible way. It is difficult to find the global optimum due to the step size 40 m. The reservoir contains four injectors and nine producers. The
limitation. By moving the initial search position to a position closer to the horizontal log permeability distribution and well location distribution
optimal point, the step length limitation can be diminished. To verify this, are shown in Fig. 5. The red dots indicate the producing wells and blue
we select several multi-peaked complex functions from the 47 test func­ dots indicate the injection wells. The porosity is 0.1. The compressibility
tions. The optimization process uses the basic values. In 50 repetitions of of water and rock is set to 1.2 × 10− 5 and 6.103 × 10− 5 psi− 1, respec­
the optimization, the initial position varied randomly within the definition tively. The initial reservoir pressure is 4351.13 psi. The total production
domain. The partial test data is given in Table 16 and the complete test time is equal to 3000 days.
data is detailed in Table B8 in Appendix B. The objective function commonly used in well placement optimiza­
According to the results shown in Table 16 and Table B8, it can be tion problem is the NPV, which is defined as:
seen that the random initial position is beneficial to find the global { [ N ]}
∑Nt ∑ p ( ) ∑ NI ( )
Δtn
optimum. For several functions, the optimal value does not change J(mi , u) = tn ro ⋅ qno,j − cw qnw,j − cwi ⋅ qnwi,k (4)
n=1 (1 + b)
because the global optimum is already found. Therefore, the randomized
365
j=1 k=1

initial position facilitates the algorithm to escape from the local opti­
where u is a Nu -dimensional column vector which contains all well
mum and enhances the global optimal seeking ability. The mean value of
random initial search positions increases compared to the mean value of control information; ndenotes the nth time step for the reservoir simu­
fixed positions for some specific functions, which means that random lation; Nt is the total number of time steps; the time at the end of the nth
initial positions also increase the chance of the algorithm falling into time step is denoted by tn ; tn is the nth time step size; b is the annual
other local optimum. discount rate; NP and NI denote the number of producers and injectors,
In summary, to find the optimal value of complex functions, we still respectively; ro is the oil revenue, in $/STB; cw , cwi , denote the disposal
recommend using random initial search positions to increase the prob­ cost of produced water and the cost of water injection in units of $/STB
ability of finding the global optimum. The running time of each test case respectively; qo,j and qw,j , respectively, denote the average oil produc­
slightly increases. This is mainly due to the process of generating the tion rate and the average water production rate at the ith producer for the
initial random positions. The increase in running time is acceptable in nth time step, in units of STB/day; qwi,k denote the average water injec­
order to find the global optimum. tion rate at the kth injector for the nth time step, in units of STB/day.
The objective is to obtain the maximum NPV shown in Eq. (4). The
4. Reservoir production optimization example oil price is set to be $60/STB. The treatment cost of both injected and
produced water is set to be $5/STB. The annual discount rate is set at
Considering the difference between the actual reservoir production

10
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Fig. 6. Optimization process corresponding to different Ne. Fig. 8. Optimization process corresponding to different cut number.

Fig. 7. Optimization process corresponding to different step size. Fig. 9. Optimization process corresponding to different Cx.

0.1. The optimized parameters are the bottom hole pressure with an
upper bound of 4351.13 psi and a lower bound of 2175.57 psi for pro­
duction wells and an upper bound of 5801.51 psi and a lower bound of
4351.13 psi for injection wells. The control variables approach is used
for sensitivity analysis. In which we set the parameters of the control
group as follows: Ne = 20, cut number = 6, step size = 1, Cx = 0.01. The
initialized BHP for production and injection wells were 3263.35 psi and
5076.32 psi, respectively. The maximum number of iterations is 100.
The optimization results are shown in Figs. 6–11.
The optimization curves for Ne are given in Fig. 6. The best opti­
mization result is obtained when Ne = 10 under the same test condi­
tions. The final optimization result is 262214500 $. This result is
consistent with the results mentioned by Chen and Reynolds (2016).
That is, a larger Ne does not ensure a better optimization result. Table 17
gives the running time. Large Ne needs more calculating times.
Fig. 7 shows the searching curves with different step sizes. We set
different cut numbers for different step sizes to ensure the same mini­
mum search step size to avoid the reduction of the local search capa­ Fig. 10. Optimization process corresponding to different initial search location.
bility. From Fig. 7, a large step size at the beginning of the iteration
The optimization curves for different cut numbers are given in Fig. 8.
yields better optimization results. As the iteration progresses, the dif­
Large cut number obtain relatively slightly better optimization results.
ference among different step sizes becomes small. At the end of the
The increase is not big. The step size is kept constant in the test, so the
iteration, the optimization curves overlap. It shows that the increase in
increase of cut number can only improve the local search capability, and
step size does not significantly affect the final optimization results.
does not improve the global search capability. Meanwhile, the runtimes
However, step size 6 obtains the optimal value with the least iterations.
of the groups with different cut number settings are given in Table 19,
The runtimes with different step sizes are shown in Table 18. The run­
and the impact of the change of cut number on the runtimes is not
time with a step size of 6 is much larger than that with a step size of 1.
obvious. Therefore, it is a relatively suboptimal strategy to improve
Alternatively, we can use a variable-step optimization strategy, i.e., use
reservoir development optimization by increasing the number of cuts
a large step size in the early stage and a small step size in the later stage
only without changing other conditions.
to increase the optimization efficiency.

11
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table 20
The optimization time corresponding to different Cx.
Cx 0.01 0.1 0.5 1
Time(s) 11026.49 9621.92 9744.74 8979.88

Table 21
Optimization time corresponding to different initial
positions.
Case Time(s)

Experimental group 1 14023.90

Experimental group 2 13231.32
Experimental group 3 13817.44
Experimental group 4 14305.15
Control group 10311.64

Fig. 11. Optimization curves before and after parameter improvement. final NPV is 2.62 × 10^8 $, while using the parameters from Chen et al.
(2016), the final NPV is 2.59 × 10^8 $. The running time of the first case
which parameters from Chen et al. (2016) is 4743.38 s, while the second
Table 17 is 4558.24 s. Therefore, the optimization result and optimization effi­
The optimization time corresponding to different Ne. ciency are improved.
Ne 5 10 15 20
Time(s) 3019.52 5881.57 8100.14 9701.56 5. Conclusion

In this work, we test the influence of key parameters on the opti­

Table 18 mization performance of the StoSAG algorithm. Based on the analysis
The optimization time corresponding to different step size. obtained from the above work, suitable parameters can be selected and
applied to the reservoir production optimization problem.
Step size 1 2 4 6
Time(s) 9600.80 12815.27 16162.51 16153.46 The ensemble size (Ne) has a significant impact on the computational
time consuming. Meanwhile, the ensemble size (Ne) has little effect on
the global optimization ability for most test functions. In terms of the
mathematical benchmark functions, increasing the step size is beneficial
Table 19
to escape the local optimum, while appropriate cut number should be
The optimization time corresponding to different cut number.
matched to ensure enough local search capability. However, in practical
Cut number 5 10 15 20
reservoir optimization, the effect of step size is mainly reflected in faster
Time(s) 10817.97 10093.40 10567.41 10226.03
convergence and does not have much influence on the optimization
results. A large cut number will increase the local search capability and
The optimization curves for different Cx are shown in Fig. 9. Overall, yield better search results for the functions with small peak ranges and
the smaller the Cx, the better the optimization results. The final opti­ high peaks in the test about mathematical benchmark functions, and in
mization result is achieved by 50 iterations for Cx = 0.01. While large Cx terms of actual reservoir optimization, the cut number is not as signifi­
tends to require more iterations. Combining the running times listed in cant as it is in the function test. The initial position has a large impact on
Table 20, it can be said that Cx = 0.01 is a reasonable setting. search results in the function test, but in practical reservoir optimiza­
When analyzing the effect of the initial positions on the optimization tion, the effect of it is merely in the cost of time for optimization. The
results, we used a random function to obtain four random initial posi­ random initial position may introduce large uncertainty of optimization
tions. The test results are given in Fig. 10, we can see that the initial process, but it does not contribute to the results. In terms of perturbation
position has strong influences on the initial process of optimization. The size (Cx), the function tests and the real case tests are consistent in
degree of influence decreases as the iteration increases. All case con­ concluding that a smaller Cx is beneficial for algorithm search. By
verges to the same final NPV. As shown in Table 21, the randomization sensitivity analysis and adjusting the key parameters, higher NPV can be
of the initial location leads to a large time cost, while the time required obtained at a lower cost of time.
for the control group is small. Therefore, in the optimization of the
actual reservoir production problem. The midpoint of the function Declaration of competing interest
definition domain is recommended.
To validate the above parameter selection strategy, we consider two The authors declared that they have no conflicts of interest to this
different cases: one refers to Chen et al. (2016) with Ne = 10, cut work.
number = 5, step size = 1, Cx = 1; another with step size = 2, Ne = 5, Cx
= 0.01, and cut number = 7. In case two, the optimization time cost and Acknowledgement
global seeking ability are both considered. The test results are given in
Fig. 11. It can be seen that better optimization results can be achieved by This study was supported by the Natural Science Foundation of
using the optimization strategy. By using the improved parameters, the Shandong Province of China (ZR2019BEE030).

12
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Appendix A. Function images of partial test functions

Fig. A.1. Function images of partial test functions.

13
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Fig. A.1. (continued).

14
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Appendix B. Algorithm testing results of different parameters

Table B.1
Ne sensitivity test data I

Ne F3 Ne F4
best ave time(s) best ave time(s)

5 − 2.06250 − 2.0614 11.05 5 − 0.93625 − 0.93624 10.891

10 − 2.0624 − 2.0611 18.49 10 − 0.93625 − 0.93624 18.138
15 − 2.0625 − 2.0611 26.16 15 − 0.9375 − 0.93627 25.521
20 − 2.0625 − 2.0614 33.39 20 − 0.93625 − 0.93624 32.907

Ne F5 Ne F8
best ave time(s) best ave time(s)

5 − 894.565 − 635.656 10.959 5 − 19.2084 − 15.9093 10.944

10 − 894.570 − 645.269 18.374 10 − 19.2084 − 16.104 18.333
15 − 894.466 − 674.630 25.841 15 − 19.2084 − 15.5234 25.74
20 − 894.528 − 654.987 33.328 20 − 19.2082 − 14.942 33.188

Ne F9 Ne F11
best ave time(s) best ave time(s)

5 − 4.1472 − 3.5533 13.203 5 5.54E-03 8.11E-02 10.871

10 − 4.1485 − 3.5798 21.661 10 2.55E-03 5.07E-02 18.376
15 − 4.1448 − 3.6838 30.02 15 4.01E-04 6.26E-02 25.656
20 − 4.1486 − 3.649 38.479 20 6.05E-03 5.58E-02 33.119

Ne F12 Ne F13
best ave time(s) best ave time(s)

5 1.0585 13.1699 12.854 5 3.47E-05 2.73E-03 10.941

10 1.0413 12.6009 21.028 10 1.40E-04 3.14E-03 18.114
15 0.82262 13.2615 29.273 15 3.49E-05 2.80E-03 25.556
20 0.97424 12.7887 37.402 20 1.41E-05 2.78E-03 32.673

Ne F16 Ne F17
best ave time(s) best ave time(s)

5 − 186.725 − 171.521 11.014 5 0.094209 2.8613 10.757

10 − 186.712 − 177.099 18.421 10 0.68682 2.6835 17.99
15 − 186.707 − 166.816 25.824 15 0.11131 2.5527 25.317
20 − 186.730 − 177.740 33.242 20 0.64805 2.7053 32.663

Ne F18 Ne F19
best ave time(s) best ave time(s)

5 8.79E-04 2.45E-02 13.02 5 4.90E-02 8.63E-01 12.766

10 5.05E-04 2.46E-02 21.199 10 1.12E-01 1.04E+00 21.069
15 7.56E-04 2.25E-02 29.65 15 4.76E-02 9.84E-01 29.254
20 2.28E-03 2.29E-02 37.39 20 4.52E-02 8.79E-01 37.203

Ne F20 Ne F21
best ave time(s) best ave time(s)

5 5.25E-05 4.30E-03 12.733 5 8.79E-09 4.64E-06 12.823

10 1.91E-04 3.97E-03 21.019 10 1.13E-07 5.88E-06 21.019
15 3.22E-04 4.24E-03 29.097 15 7.77E-09 5.27E-06 29.351
20 6.10E-05 3.98E-03 37.27 20 1.61E-07 5.63E-06 37.35

Ne F23 Ne F24
best ave time(s) best ave time(s)

5 − 1.9999 − 1.9988 12.813 5 1.27E-03 3.13E-02 10.664

10 − 1.9998 − 1.9987 21.09 10 1.49E-03 3.16E-02 17.974
15 − 1.9999 − 1.9988 29.184 15 3.73E-04 3.33E-02 25.286
20 − 1.9999 − 1.9989 37.474 20 3.18E-04 3.02E-02 32.647

Ne F27 Ne F28
best ave time(s) best ave time(s)

5 116.3208 116.3208 12.983 5 0.0026084 0.010851 12.833

10 116.3208 116.3208 21.078 10 7.64E-05 0.011072 21.049
15 116.3208 116.3208 29.338 15 0.00075073 0.010982 29.268
20 116.3208 116.3208 37.51 20 0.00028322 0.01019 37.47

Ne F30 Ne F33
best ave time(s) best ave time(s)

5 − 1.0315 − 1.027 10.874 5 1.0007 2.4639 21.688

10 − 1.0308 − 1.0267 18.174 10 0.99812 2.1308 33.583
15 − 1.0316 − 1.0272 25.689 15 1.0052 2.1713 45.645
(continued on next page)

15
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.1 (continued )

20 − 1.0314 − 1.0271 33.002 20 0.99918 2.1718 57.681

Ne F35 Ne F37
best ave time(s) best ave time(s)

5 − 1.8007 − 1.7941 13.123 5 0.398 0.40136 10.864

10 − 1.8007 − 1.794 21.463 10 0.39832 0.40104 18.232
15 − 1.801 − 1.7936 29.822 15 0.39816 0.4012 25.647
20 − 1.8009 − 1.7958 38.164 20 0.39868 0.40147 32.892

Ne F39 Ne F40
best ave time(s) best ave time(s)

5 − 6.0179 − 1.1868 8.185 5 3.0213 3.1462 10.717

10 − 0.98554 − 0.98554 14.175 10 3.0071 3.14 18.19
15 − 6.0179 − 1.2875 20.289 15 3.0008 3.1589 25.357
20 − 0.98554 − 0.98554 26.246 20 3.008 3.1583 32.643

Ne F41 Ne F42
best ave time(s) best ave time(s)

5 − 3.8628 − 3.739 11.324 5 − 3.1343 − 3.0532 11.534

10 − 3.8628 − 3.739 18.81 10 − 3.1344 − 3.0721 19.05
15 − 3.8628 − 3.8318 26.42 15 − 3.1343 − 3.0912 26.757
20 − 3.8628 − 3.8318 33.922 20 − 3.1344 − 3.0579 34.364

Ne F45 Ne F46
best ave time(s) best ave time(s)

5 3.77E-04 4.69E-03 11.704 5 − 10.3442 − 9.7274 11.594

10 2.71E-04 3.90E-03 19.366 10 − 10.2714 − 9.3249 19.309
15 4.30E-04 3.89E-03 27.105 15 − 10.4242 − 9.4693 26.919
20 1.83E-04 4.92E-03 34.755 20 − 10.3475 − 9.6065 34.541
Ne F47

best ave time(s)

5 − 78.3317 − 71.518 12.903

10 − 78.3292 − 74.3427 21.219
15 − 78.3317 − 75.7585 29.52
20 − 78.3314 − 75.1983 37.769

Table B.2
Ne sensitivity test data II

Ne F1 Ne F2
best ave time(s) best ave time(s)

5 0.11799 2.8519 10.92 5 0.07472 0.6607 10.77

10 0.37145 2.4892 18.45 10 0.15099 0.73597 18.13
15 0.27347 2.0590 25.82 15 0.1979 0.71518 25.52
20 0.77549 2.7229 33.33 20 0.09250 0.65414 32.82

Ne F6 Ne F7
best ave time(s) best ave time(s)

5 − 0.52336 − 0.01885 8.855 5 0.0091619 0.069056 13.783

10 − 0.52336 0.014048 14.522 10 0.014155 0.080197 22.318
15 0.025015 0.025015 20.573 15 0.012916 0.078689 30.977
20 0.025015 0.025015 26.786 20 0.0024606 0.080722 39.449

Ne F10 Ne F14
best ave time(s) best ave time(s)

5 1.18E-05 1.71E-03 14.995 5 0.29259 0.29319 10.824

10 3.34E-05 1.79E-03 23.977 10 0.29264 0.29324 18.103
15 3.58E-05 1.72E-03 32.988 15 0.29259 0.29325 25.407
20 1.82E-04 1.95E-03 42.119 20 0.29267 0.29331 32.803

Ne F15 Ne F22
best ave time(s) best ave time(s)

5 0.41861 320.9278 12.854 5 3.27E-03 2.09E-02 12.743

10 0.055702 311.0557 21.128 10 7.11E-04 2.21E-02 20.879
15 119.0084 285.9178 29.338 15 1.63E-04 2.23E-02 29.099
20 118.8036 301.2568 37.635 20 1.98E-03 2.26E-02 37.19

Ne F25 Ne F26
best ave time(s) best ave time(s)

5 1.07E-04 9.28E-04 10.662 5 − 1.9132 − 1.9125 10.745

10 1.74E-05 8.53E-04 17.901 10 − 1.9132 − 1.9125 18.08
15 4.10E-05 7.92E-04 25.176 15 − 1.9132 − 1.9124 25.367
(continued on next page)

16
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.2 (continued )

20 5.80E-05 7.89E-04 32.442 20 − 1.9131 − 1.9124 32.675

Ne F29 Ne F31
best ave time(s) best ave time(s)

5 1.37E-03 1.07E-01 11.188 5 5.77E-04 2.14E-02 12.703

10 4.42E-04 1.07E-01 18.444 10 8.43E-04 2.40E-02 20.822
15 5.52E-04 1.07E-01 25.996 15 2.83E-03 2.47E-02 29.015
20 7.35E-05 9.51E-02 33.412 20 4.27E-03 2.24E-02 37.167

Ne F32 Ne F34
best ave time(s) best ave time(s)

5 1.20E-04 3.73E-03 12.884 5 − 0.99908 − 0.32614 10.674

10 3.10E-04 5.74E-03 21.008 10 − 0.99949 − 0.35067 18.019
15 1.52E-04 9.51E-03 29.187 15 − 0.98902 − 0.35469 25.314
20 2.78E-04 1.28E-02 37.411 20 − 0.98748 − 0.29531 32.683

Ne F36 Ne F38
best ave time(s) best ave time(s)

5 7.89E-06 2.52E-04 10.831 5 0.12941 0.87389 15.072

10 1.60E-05 2.52E-04 18.28 10 0.030443 0.98311 18.693
15 3.26E-05 3.17E-04 25.767 15 0.15542 1.0857 26.106
20 5.30E-05 3.09E-04 33.04 20 0.30188 1.1266 33.582

Ne F43 Ne F44
best ave time(s) best ave time(s)

5 − 3.0423 − 3.0421 12.356 5 1.74E-06 6.57E-05 12.816

10 − 3.0424 − 3.0421 20.34 10 3.17E-06 5.50E-05 21.039
15 − 3.0422 − 3.0421 28.374 15 8.76E-06 5.46E-05 29.324
20 − 3.0423 − 3.0421 36.514 20 4.18E-06 4.28E-05 37.529

Table B.3
Step size sensitivity test data I

Step size F1 Step size F2

best ave time(s) best ave time(s)

1 0.087224 2.0666 18.395 1 0.21928 0.8274 18.107

2 1.5073 5.9954 18.408 2 0.2132 0.99272 18.147
3 0.46192 9.3157 18.376 3 0.36969 1.0483 18.13
4 3.0127 9.7673 18.415 4 0.1334 1.3328 18.179
5 3.0533 13.4945 18.435 5 0.44196 1.5151 18.195

Step size F3 Step size F4

best ave time(s) best ave time(s)

1 − 2.0626 − 2.0612 18.539 1 − 0.9363 − 0.9363 18.168

2 − 2.0626 − 2.06 18.491 2 − 0.9363 − 0.9362 18.243
3 − 2.0622 − 1.9805 18.495 3 − 0.9991 − 0.9639 18.239
4 − 2.0624 − 1.7971 18.628 4 − 0.9363 − 0.9362 18.165
5 − 2.0618 − 1.8898 18.584 5 − 0.9363 − 0.9362 18.201

Step size F7 Step size F10

best ave time(s) best ave time(s)

1 0.008213 0.075017 22.38 1 0.000246 0.001408 24.01

2 0.067685 0.25198 22.391 2 0.000609 0.008461 23.987
3 0.11066 0.43163 22.368 3 0.000778 0.018091 23.958
4 0.12858 0.65604 22.356 4 0.002351 0.053601 24.022
5 0.37306 0.95752 22.348 5 0.003445 0.054491 24.087

Step size F11 Step size F12

best ave time(s) best ave time(s)

1 0.003303 0.052541 18.22 1 0.55967 9.4484 21.074

2 0.007225 0.13996 18.213 2 0.98498 3.6658 20.959
3 0.007556 0.44365 18.12 3 0.002136 1.8605 21.069
4 0.28896 0.62927 18.164 4 0.52727 3.7378 21.111
5 0.33254 0.81832 18.141 5 1.1403 3.1091 21.089

Step size F13 Step size F14

best ave time(s) best ave time(s)

1 2.72E-07 2.56E-03 18.003 1 0.29259 0.29319 18.041

2 2.17E-03 1.08E-02 18.017 2 0.29338 0.29669 18.11
3 1.21E-04 2.21E-02 18.04 3 0.29395 0.30092 18.07
4 1.75E-02 3.87E-02 18.013 4 0.29499 0.30824 18.065
5 3.55E-05 5.17E-02 18.071 5 0.29279 0.31382 18.071

Step size F16 Step size F17

(continued on next page)

17
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.3 (continued )

best ave time(s) best ave time(s)

1 − 186.725 − 173.178 18.361 1 0.68083 2.8049 18.021

2 − 186.558 − 124.908 18.437 2 1.7986 8.4306 17.981
3 − 186.592 − 141.254 18.43 3 1.2217 18.6282 18.003
4 − 186.684 − 166.435 18.31 4 3.2538 38.4565 18.02
5 − 186.351 − 129.505 18.283 5 5.3777 72.7337 18.091

Step size F18 Step size F19

best ave time(s) best ave time(s)

1 0.00049 0.023542 20.982 1 0.008776 0.93768 20.829

2 0.003099 0.067347 21.028 2 0.62392 3.8349 20.835
3 0.001592 0.17387 21.069 3 0.078021 7.1493 20.899
4 0.008042 0.10282 21.082 4 1.2253 15.147 20.807
5 0.00147 0.17458 21.029 5 0.88114 32.5198 20.792

Step size F20 Step size F21

best ave time(s) best ave time(s)

1 9.02E-05 0.005062 20.892 1 4.13E-07 5.93E-06 21.013

2 0.000733 0.016885 20.919 2 3.73E-06 5.94E-05 21.009
3 0.002943 0.030923 20.909 3 1.65E-06 0.00016 21.078
4 0.002809 0.070351 20.93 4 5.24E-06 0.000295 21.029
5 0.01343 0.13112 20.89 5 2.45E-05 0.000771 21.089

Step size F22 Step size F23

best ave time(s) best ave time(s)

1 0.003615 0.022662 20.929 1 − 1.9999 − 1.9989 21.018

2 0.006012 0.087864 20.921 2 − 1.9997 − 1.9947 21.019
3 0.009605 0.19709 20.899 3 − 1.9997 − 1.9889 20.992
4 0.029516 0.37453 20.979 4 − 1.9999 − 1.9798 20.969
5 0.11383 0.68955 20.942 5 − 1.9999 − 1.9698 21.009

Step size F24 Step size F25

best ave time(s) best ave time(s)

1 0.000235 0.030078 17.95 1 3.09E-05 0.000842 17.891

2 0.009303 0.13313 18.031 2 0.000345 0.003243 18.01
3 0.036692 0.27693 17.966 3 0.00046 0.008597 17.864
4 0.013859 0.50313 18.001 4 0.000352 0.013957 17.94
5 0.008696 0.87561 18 5 0.001522 0.019537 17.937

Step size F26 Step size F29

best ave time(s) best ave time(s)

1 − 1.9131 − 1.9124 18.05 1 7.33E-05 0.088796 18.418

2 − 1.9132 − 1.9104 18.04 2 3.38E-05 0.077076 18.541
3 − 1.9127 − 1.9063 18.087 3 0.008389 0.051942 18.382
4 − 1.9128 − 1.9048 18.021 4 0.002652 0.11259 18.364
5 − 1.9104 − 1.8944 18.04 5 0.028166 0.12971 18.29

Step size F31 Step size F32

best ave time(s) best ave time(s)

1 0.000674 0.021456 20.869 1 0.000343 0.005447 20.973

2 0.002439 0.042005 20.829 2 0.000584 0.10138 20.988
3 0.004258 0.17475 20.886 3 0.000575 0.041011 21.009
4 0.000446 0.094642 20.869 4 0.008132 0.079684 21.042
5 0.001111 0.13032 20.879 5 0.000905 0.24809 20.969

Step size F33 Step size F35

best ave time(s) best ave time(s)

1 1.0054 2.0447 33.592 1 − 1.8011 − 1.7942 21.658

2 1.3538 2.3193 33.646 2 − 1.8001 − 1.7732 21.452
3 12.6705 12.6705 33.675 3 − 1.8007 − 1.7514 21.438
4 1.4637 2.3518 33.633 4 − 1.8 − 1.591 21.449
5 12.6705 12.6705 33.571 5 − 1.7868 − 1.6389 21.418

Step size F36 Step size F37

best ave time(s) best ave time(s)

1 1.89E-05 2.60E-04 18.27 1 0.39829 0.40135 18.193

2 6.03E-05 1.14E-03 18.381 2 0.3981 0.40778 18.201
3 9.53E-05 4.56E-03 18.323 3 0.39834 0.41108 18.154
4 6.49E-04 8.16E-03 18.31 4 0.40054 0.42529 18.191
5 2.18E-04 9.41E-03 18.26 5 0.3982 0.43569 18.221

Step size F38 Step size F40

best ave time(s) best ave time(s)

1 0.12305 0.78028 18.587 1 3.0022 3.1521 17.98

2 0.17433 1.7612 18.63 2 3.0328 3.5912 18.067
3 0.4459 3.2396 18.62 3 3.0052 4.4567 18.021
4 0.36401 3.4993 18.633 4 3.1478 5.7398 17.981
(continued on next page)

18
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.3 (continued )

5 0.51501 5.0905 18.63 5 3.599 8.2665 18.039

Step size F42 Step size F43

best ave time(s) best ave time(s)

1 − 3.1344 − 3.0674 19.119 1 − 3.0423 − 3.0421 20.499

2 − 3.134 − 3.0282 19.163 2 − 3.0419 − 3.0411 20.439
3 − 3.1331 − 3.1072 19.16 3 − 3.0412 − 3.0371 20.522
4 − 3.1342 − 3.057 19.11 4 − 3.0405 − 3.0341 20.429
5 − 3.1303 − 3.0959 19.11 5 − 3.0392 − 3.0305 20.539

Step size F45 Step size F46

best ave time(s) best ave time(s)

1 0.000245 0.00425 19.419 1 − 10.3251 − 9.5631 19.322

2 0.004067 0.025156 19.42 2 − 9.9317 − 8.4032 19.24
3 0.00658 0.059513 19.413 3 − 9.5531 − 5.3931 19.26
4 0.034147 0.1461 19.43 4 − 8.3393 − 5.6754 19.263
5 0.02239 0.28347 19.469 5 − 6.6736 − 4.4552 19.22

Table B.4
step size sensitivity test data II

Step size F5 Step size F6

best ave time(s) best ave time(s)

1 − 894.529 − 648.42 18.366 1 − 0.5234 − 0.00789 14.452

2 − 956.893 − 882.42 18.429 2 − 0.8659 0.007196 14.438
3 − 956.754 − 906.78 18.553 3 − 0.8659 − 0.86594 14.415
4 − 956.34 − 934.19 18.545 4 − 0.8659 − 0.86594 14.422
5 − 956.34 − 900.43 18.43 5 − 0.8659 − 0.86594 14.462

Step size F8 Step size F9

best ave time(s) best ave time(s)

1 − 19.209 − 15.5262 18.321 1 − 4.1387 − 3.6095 21.648

2 − 19.208 − 17.4888 18.36 2 − 4.149 − 3.4423 21.658
3 − 19.207 − 18.5766 18.31 3 − 4.1524 − 3.1494 21.63
4 − 19.204 − 16.3631 18.357 4 − 4.1508 − 3.3424 21.618
5 − 19.206 − 16.1547 18.38 5 − 4.147 − 3.324 21.619

Step size F15 Step size F27

best ave time(s) best ave time(s)

1 118.6616 301.4638 21.079 1 116.3208 116.3208 21.132

2 0.2823 129.4756 21.102 2 116.3208 116.3208 21.088
3 0.70932 155.919 21.149 3 116.3208 116.3208 21.129
4 0.27675 142.9109 21.099 4 116.3208 116.3208 21.152
5 1.5691 201.0928 21.184 5 116.3208 116.3208 21.099

Step size F28 Step size F30

best ave time(s) best ave time(s)

1 0.000985 0.009114 21.099 1 − 1.0311 − 1.0271 18.255

2 0.000297 0.045736 21.082 2 − 1.0311 − 1.0157 18.14
3 0.004573 0.10183 21.128 3 − 1.0315 − 1.0169 18.11
4 0.008112 0.14053 21.069 4 − 1.0315 − 0.98223 18.132
5 0.015422 0.2694 21.061 5 − 1.0246 − 0.89889 18.04

Step size F34 Step size F39

best ave time(s) best ave time(s)

1 -0.99775 -0.97976 17.314 1 − 6.0179 − 1.0862 14.132

2 -0.99437 -0.9344 17.228 2 − 6.0179 − 1.0826 14.204
3 -0.99742 -0.83284 17.251 3 − 6.0179 − 6.0179 14.172
4 -0.97626 -0.71415 17.245 4 − 5.7948 − 1.1744 14.202
5 -0.99884 -0.65933 17.796 5 − 5.7948 − 5.7948 14.193

Step size F41 Step size F44

best ave time(s) best ave time(s)

1 − 3.8628 − 3.8318 18.85 1 1.12E-06 5.48E-05 21.049

2 − 3.8628 − 3.8624 18.813 2 1.05E-05 0.000208 21.071
3 − 3.8628 − 3.862 18.84 3 8.08E-06 0.000267 21.039
4 − 3.8627 − 3.8614 18.86 4 1.05E-05 0.000205 21.082
5 − 3.8627 − 3.8603 18.822 5 1.05E-05 0.000133 21.089
Step size F47

best ave time(s)

1 − 78.332 − 74.3458 21.242

2 − 78.3276 − 77.0811 21.199
3 − 78.3203 − 78.0852 21.188
(continued on next page)

19
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.4 (continued )

Step size F47

best ave time(s)

4 − 78.3116 − 75.3235 21.273

5 − 78.2781 − 77.3056 21.208

Table B.5
Cut number sensitivity analysis test data I

Cut number F2 Cut number F3

best Ave time(s) best Ave time(s)

5 0.044256 0.71038 18.135 5 − 2.0626 − 2.0612 18.447

10 0.044755 0.18078 19.696 10 − 2.0626 − 2.0626 20.146
15 0.01493 0.039875 21.226 15 − 2.0626 − 2.0626 21.776
20 0.0023962 0.013621 22.867 20 − 2.0626 − 2.0626 23.345

Cut number F4 Cut number F7

best Ave time(s) best Ave time(s)

5 − 0.93625 − 0.93624 19.272 5 0.00786 0.06862 22.331

10 − 0.93625 − 0.93625 20.096 10 0.0018744 0.011981 25.157
15 − 1 − 0.93752 21.176 15 4.02E-06 0.0091428 27.815
20 − 1 − 0.9388 22.447 20 1.45E-09 0.0085788 31.63

Cut number F10 Cut number F11

best ave time(s) best ave time(s)

5 5.94E-06 0.0015235 24.047 5 0.0018056 0.075656 18.14

10 9.14E-08 1.40E-06 27.343 10 1.05E-06 2.69E-05 19.613
15 1.58E-11 1.48E-09 30.572 15 1.75E-09 2.19E-08 21.11
20 5.05E-14 1.49E-12 33.768 20 8.00E-13 2.37E-11 22.598

Cut number F13 Cut number F14

best ave time(s) best ave time(s)

5 3.37E-06 0.0027052 17.981 5 0.29258 0.29314 18.003

10 9.97E-10 2.40E-06 19.439 10 0.29258 0.29258 19.483
15 1.58E-10 1.53E-09 21.022 15 0.29258 0.29258 21.009
20 3.49E-14 1.61E-12 22.379 20 0.29258 0.29258 22.538

Cut number F16 Cut number F17

best ave time(s) best ave time(s)

5 − 186.7304 − 155.3184 18.211 5 0.53143 2.6568 18.09

10 − 186.7309 − 173.0142 19.809 10 0.00014566 0.0345 19.57
15 − 186.7309 − 174.3674 21.319 15 1.60E-06 2.83E-05 21.079
20 − 186.7309 − 175.2025 22.751 20 4.11E-10 3.06E-08 22.548

Cut number F18 Cut number F19

best ave time(s) best ave time(s)

5 0.00043673 0.020279 21.049 5 0.061572 0.97442 20.789

10 6.14E-07 2.71E-05 23.526 10 2.29E-05 0.0009853 23.148
15 1.42E-09 2.34E-08 25.807 15 1.90E-08 9.91E-07 25.499
20 3.92E-14 2.17E-11 28.165 20 1.75E-11 1.02E-09 27.895

Cut number F20 Cut number F21

best ave time(s) best ave time(s)

5 2.54E-05 0.0046217 20.949 5 2.86E-07 6.37E-06 21.119

10 1.37E-07 4.68E-06 23.24 10 5.56E-13 2.02E-10 23.401
15 4.33E-10 4.27E-09 25.617 15 3.30E-16 3.93E-14 25.767
20 1.99E-13 4.03E-12 27.97 20 6.76E-19 5.38E-16 28.223

Cut number F22 Cut number F23

best ave time(s) best ave time(s)

5 0.0017871 0.024475 20.959 5 − 2 − 1.9987 21.021

10 1.04E-07 2.19E-05 23.248 10 − 2 − 2 23.428
15 6.30E-10 2.11E-08 25.48 15 − 2 − 2 25.666
20 2.13E-13 2.30E-11 27.715 20 − 2 − 2 27.928

Cut number F24 Cut number F25

best ave time(s) best ave time(s)

5 0.00042783 0.029729 17.973 5 0.00013012 0.0009181 17.951

10 2.74E-06 2.91E-05 19.45 10 8.85E-08 8.58E-07 19.382
15 2.54E-09 3.04E-08 20.959 15 6.78E-11 7.61E-10 20.859
20 3.50E-13 2.98E-11 22.475 20 1.24E-13 8.12E-13 22.328

Cut number F26 Cut number F27

best ave time(s) best ave time(s)

(continued on next page)

20
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.5 (continued )

5 − 1.9132 − 1.9125 18.043 5 116.3208 116.3208 21.089
10 − 1.9132 − 1.9132 19.529 10 116.3208 116.3208 23.558
15 − 1.9132 − 1.9132 20.989 15 116.3208 116.3208 25.986
20 − 1.9132 − 1.9132 22.508 20 116.3208 116.3208 28.415

Cut number F28 Cut number F30

best ave time(s) best ave time(s)

5 0.0004335 0.0094949 21.069 5 − 1.0314 − 1.0271 18.011

10 1.64E-06 1.08E-05 23.548 10 − 1.0316 − 1.0316 19.619
15 3.46E-11 1.04E-08 25.857 15 − 1.0316 − 1.0316 21.062
20 4.20E-13 9.67E-12 28.235 20 − 1.0316 − 1.0316 22.59

Cut number F31 Cut number F32

best ave time(s) best ave time(s)

5 0.0038901 0.021231 20.872 5 9.43E-06 0.0048588 20.988

10 1.52E-06 2.30E-05 23.227 10 1.20E-06 4.22E-05 23.438
15 6.34E-10 2.41E-08 25.517 15 4.54E-10 5.63E-06 25.676
20 1.47E-12 2.30E-11 27.79 20 5.31E-13 1.25E-06 27.546

Cut number F34 Cut number F35

best ave time(s) best ave time(s)

5 − 0.9994 − 0.23102 18.074 5 − 1.8013 − 1.7941 21.431

10 − 0.99986 − 0.99758 19.5 10 − 1.8013 − 1.8013 23.917
15 − 1 − 1 21.054 15 − 1.8013 − 1.8013 26.36
20 − 1 − 1 23.091 20 − 1.8013 − 1.8013 28.715

Cut number F36 Cut number F37

best ave time(s) best ave time(s)

5 1.11E-05 0.0002244 18.124 5 0.39798 0.40103 18.17

10 1.19E-08 2.66E-07 19.71 10 0.39789 0.39789 19.7
15 9.62E-12 3.03E-10 21.108 15 0.39789 0.39789 21.245
20 4.79E-15 3.59E-13 22.568 20 0.39789 0.39789 23.395

Cut number F39 Cut number F40

best ave time(s) best ave time(s)

5 − 0.98554 − 0.98554 14.135 5 3.002 3.1697 18.094

10 − 0.98632 − 0.98632 15.002 10 3 3.0001 19.399
15 − 0.98633 − 0.98633 15.791 15 3 3 20.999
20 − 6.0207 − 1.2884 16.591 20 3 3 22.394

Cut number F42 Cut number F43

best ave time(s) best ave time(s)

5 − 3.1344 − 3.0341 19.074 5 − 3.0423 − 3.0421 20.502

10 − 3.1345 − 3.0488 20.968 10 − 3.0425 − 3.0425 22.678
15 − 3.1345 − 3.0488 22.798 15 − 3.0425 − 3.0425 24.747
20 − 3.1345 − 3.0535 24.481 20 − 3.0425 − 3.0425 26.778

Cut number F44 Cut number F45

best ave time(s) best ave time(s)

5 3.54E-06 5.78E-05 21.053 5 0.00014667 0.0039973 19.55

10 1.90E-09 5.06E-06 23.487 10 3.19E-06 6.59E-05 21.428
15 8.07E-12 5.62E-06 25.711 15 1.56E-06 3.07E-05 23.237
20 3.36E-15 4.64E-06 27.865 20 5.45E-08 2.76E-05 24.457

Table B.6
Cut number sensitivity analysis test data II
Cut number F1 Cut number F5
best ave time(s) best ave time(s)

5 0.31725 3.0778 18.568 5 − 894.5538 − 650.3392 18.383

10 0.010722 0.43841 19.96 10 − 894.5788 − 648.9868 19.89
15 0.00048292 2.3867 21.516 15 − 894.5789 − 621.7105 21.418
20 1.48E-05 1.9887 23.044 20 − 894.5789 − 634.6642 22.911

Cut number F6 Cut number F8

best ave time(s) best ave time(s)

5 − 0.52336 0.014048 14.403 5 − 19.2085 − 16.003 18.421

10 − 0.5266 − 0.019327 15.284 10 − 19.2085 − 17.4688 19.892
15 − 0.5266 0.002703 16.152 15 − 19.2085 − 17.1788 21.348
20 − 0.5266 − 0.008324 16.991 20 − 19.2085 − 16.8889 22.887

Cut number F9 Cut number F12

best ave time(s) best ave time(s)

5 − 4.1542 − 3.6149 19.008 5 1.7353 11.5042 21.089

(continued on next page)

21
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.6 (continued )

10 − 4.1558 − 3.7033 20.423 10 0.99553 13.2731 23.482
15 − 4.1558 − 3.4508 22.124 15 0.99496 12.3374 25.706
20 − 4.1558 − 3.3497 23.784 20 0.99496 11.4221 28.105

Cut number F15 Cut number F29

best ave time(s) best ave time(s)

5 118.5725 294.7272 21.139 5 3.79E-05 0.1199 18.36

10 118.4386 274.7807 23.608 10 9.45E-07 0.09557 19.899
15 118.4384 286.228 25.956 15 7.85E-10 0.11348 21.501
20 118.4384 299.6515 28.374 20 2.88E-13 0.083619 23.138

Cut number F33 Cut number F38

best ave time(s) best ave time(s)

5 1.0402 2.0524 31.898 5 0.15487 0.81164 18.57

10 1.992 1.992 36.851 10 0.0004495 0.34399 20.279
15 0.998 1.9722 42.149 15 1.05E-05 0.21568 21.822
20 0.998 1.9324 47.464 20 0.00012523 0.4094 23.057

Cut number F41 Cut number F46

best ave time(s) best ave time(s)

5 − 3.8628 − 3.7545 18.86 5 − 10.4517 − 9.8149 19.19

10 − 3.8628 − 3.7082 20.659 10 − 10.5363 − 9.6396 21.168
15 − 3.8628 − 3.7855 21.978 15 − 10.5364 − 10.3831 23.091
20 − 3.8628 − 3.8009 23.149 20 − 10.5364 − 9.9805 24.897

Cut number F47

best ave time(s)

5 − 78.3312 − 73.496 21.239

10 − 78.3323 − 73.2431 23.68
15 − 78.3323 − 72.9604 26.116
20 − 78.3323 − 74.9395 28.625

Table B.6
Cut number sensitivity analysis test data II
Cut number F1 Cut number F5
best ave time(s) best ave time(s)

5 0.31725 3.0778 18.568 5 − 894.5538 − 650.3392 18.383

10 0.010722 0.43841 19.96 10 − 894.5788 − 648.9868 19.89
15 0.00048292 2.3867 21.516 15 − 894.5789 − 621.7105 21.418
20 1.48E-05 1.9887 23.044 20 − 894.5789 − 634.6642 22.911

Cut number F6 Cut number F8

best ave time(s) best ave time(s)

5 − 0.52336 0.014048 14.403 5 − 19.2085 − 16.003 18.421

10 − 0.5266 − 0.019327 15.284 10 − 19.2085 − 17.4688 19.892
15 − 0.5266 0.002703 16.152 15 − 19.2085 − 17.1788 21.348
20 − 0.5266 − 0.008324 16.991 20 − 19.2085 − 16.8889 22.887

Cut number F9 Cut number F12

best ave time(s) best ave time(s)

5 − 4.1542 − 3.6149 19.008 5 1.7353 11.5042 21.089

10 − 4.1558 − 3.7033 20.423 10 0.99553 13.2731 23.482
15 − 4.1558 − 3.4508 22.124 15 0.99496 12.3374 25.706
20 − 4.1558 − 3.3497 23.784 20 0.99496 11.4221 28.105

Cut number F15 Cut number F29

best ave time(s) best ave time(s)

5 118.5725 294.7272 21.139 5 3.79E-05 0.1199 18.36

10 118.4386 274.7807 23.608 10 9.45E-07 0.09557 19.899
15 118.4384 286.228 25.956 15 7.85E-10 0.11348 21.501
20 118.4384 299.6515 28.374 20 2.88E-13 0.083619 23.138

Cut number F33 Cut number F38

best ave time(s) best ave time(s)

5 1.0402 2.0524 31.898 5 0.15487 0.81164 18.57

10 1.992 1.992 36.851 10 0.0004495 0.34399 20.279
15 0.998 1.9722 42.149 15 1.05E-05 0.21568 21.822
20 0.998 1.9324 47.464 20 0.00012523 0.4094 23.057

Cut number F41 Cut number F46

(continued on next page)

22
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.6 (continued )

best ave time(s) best ave time(s)

5 − 3.8628 − 3.7545 18.86 5 − 10.4517 − 9.8149 19.19

10 − 3.8628 − 3.7082 20.659 10 − 10.5363 − 9.6396 21.168
15 − 3.8628 − 3.7855 21.978 15 − 10.5364 − 10.3831 23.091
20 − 3.8628 − 3.8009 23.149 20 − 10.5364 − 9.9805 24.897

Cut number F47

best ave time(s)

5 − 78.3312 − 73.496 21.239

10 − 78.3323 − 73.2431 23.68
15 − 78.3323 − 72.9604 26.116
20 − 78.3323 − 74.9395 28.625

Table B.7
Perturbation step sensitivity analysis test data

Cx F1 Cx F6
best ave time(s) best ave time(s)

0.01 0.20025 2.7473 20.218 0.01 0.025015 0.025015 14.885

0.1 0.42358 1.8433 19.558 0.1 − 0.52336 − 0.46852 14.979
0.5 0.1201 1.6905 19.305 0.5 − 0.52336 − 0.46852 14.956
1 0.44138 1.8136 19.299 1 − 0.52336 − 0.49045 15.389

Cx F10 Cx F13
best ave time(s) best ave time(s)

0.01 5.73E-06 0.001781 21.841 0.01 5.36E-06 0.003011 18.525

0.1 9.54E-05 0.001974 21.935 0.1 2.00E-06 0.003165 18.807
0.5 2.50E-05 0.001985 22.239 0.5 1.22E-05 0.002638 18.731
1 7.93E-05 0.001751 22.127 1 1.51E-06 0.00298 18.699

Cx F16 Cx F17
best ave time(s) best ave time(s)

0.01 − 186.685 − 166.074 19.021 0.01 0.22729 2.7857 18.64

0.1 − 186.671 − 171.983 19.048 0.1 0.41405 3.0703 18.556
0.5 − 186.712 − 168.56 19.065 0.5 0.68922 3.0741 18.665
1 − 186.587 − 173.248 19.022 1 0.87246 2.8891 18.713

Cx F22 Cx F25
best ave time(s) best ave time(s)

0.01 3.14E-05 0.020931 18.753 0.01 4.01E-06 0.00088 18.529

0.1 0.00354 0.022175 18.792 0.1 0.000108 0.000886 18.472
0.5 0.001229 0.024628 18.862 0.5 4.74E-05 0.000731 18.523
1 0.003026 0.024702 18.867 1 2.03E-05 0.000791 18.505

Cx F28 Cx F29
best ave time(s) best ave time(s)

0.01 0.000103 0.01173 18.8 0.01 0.000108 0.071607 18.736

0.1 9.44E-05 0.011194 18.835 0.1 0.000101 0.016966 18.725
0.5 0.002282 0.011364 18.952 0.5 0.000106 0.035504 18.837
1 0.001585 0.011204 18.976 1 0.000383 0.034957 18.963

Cx F32 Cx F34
best ave time(s) best ave time(s)

0.01 0.000174 0.006341 18.903 0.01 − 0.99732 − 0.27114 18.518

0.1 4.34E-05 0.008142 18.689 0.1 − 0.97874 − 0.27766 18.585
0.5 0.000391 0.006851 18.692 0.5 − 0.99355 − 0.22067 18.63
1 9.49E-05 0.078653 18.644 1 − 0.98628 − 0.22228 18.758

Cx F41 Cx F43
best ave time(s) best ave time(s)

0.01 − 3.8628 − 3.8163 19.792 0.01 − 3.0423 − 3.0421 21.173

0.1 − 3.8628 − 3.8318 19.546 0.1 − 3.0423 − 3.0421 21.121
0.5 − 3.8628 − 3.8627 19.616 0.5 − 3.0423 − 3.0384 21.294
1 − 3.8628 − 3.8627 19.604 1 − 3.0423 − 3.0372 21.167

23
J. Xu et al. Journal of Petroleum Science and Engineering 209 (2022) 109755

Table B.8
Initial position sensitivity analysis test data

Initial search location F5 Initial search location F6

best ave time(s) best ave time(s)

Fixed − 894.4586 − 657.4742 17.904 Fixed − 0.52336 0.014048 14.68

Random − 956.8444 − 712.2101 18.564 Random − 0.86901 − 0.4426 14.912
Initial search location F8 Initial search location F15
best ave time(s) best ave time(s)

Fixed − 19.2085 − 15.62 18.331 Fixed 118.8354 310.6765 21.099

Random − 19.2084 − 15.9639 18.391 Random 0.0052736 247.9856 21.335
Initial search location F16 Initial search location F39
best ave time(s) best ave time(s)

Fixed − 186.7079 − 165.0125 18.581 Fixed − 6.0179 − 1.2875 14.441

Random − 186.7146 − 132.8332 18.72 Random − 6.0207 − 3.3997 14.53
Initial search location F41 Initial search location F42
best ave time(s) best ave time(s)

Fixed − 3.8628 − 3.7854 18.99 Fixed − 3.1343 − 3.0627 19.626

Random − 3.8628 − 2.6606 19 Random − 3.1343 − 3.0436 19.736
Initial search location F43 Initial search location F46
best ave time(s) best ave time(s)

Fixed − 3.0423 − 3.0421 20.838 Fixed − 10.3864 − 9.471 19.497

Random − 3.0423 − 3.0421 20.739 Random − 10.2707 − 6.7845 19.483

Credit roles

Jianchun Xu: Conceptualization, Methodology, Shuyang Liu: Writing-Reviewing and Editing, Hangyu Li and Ling Fan: Supervision, Xiaopu Wang:
Writing- original draft, Wenxin Zhou: Data curation, Visualization, Investigation.

References Liu, Z., Forouzanfar, F., 2018. Ensemble clustering for efficient robust optimization of
naturally fractured reservoirs. Comput. Geosci. 22 (1), 283–296.
Liu, Z., Forouzanfar, F., Zhao, Y., 2018. Comparison of SQP and AL algorithms for
Al Dossary, M.A., Nasrabadi, H., 2016. Well placement optimization using imperialist
deterministic constrained production optimization of hydrocarbon reservoirs.
competitive algorithm. J. Petrol. Sci. Eng. 147, 237–248.
J. Petrol. Sci. Eng. 171, 542–557.
Ali, D.H., Al-Jawad, M.S., Van Kirk, C.W., 2015. Distribution of new horizontal wells by
Liu, Z., Reynolds, A.C., 2020. A sequential-quadratic-programming-filter algorithm with
the use of artificial neural network algorithm. In: SPE Middle East Oil & Gas Show
a modified stochastic gradient for robust life-cycle optimization problems with
and Conference. OnePetro.
nonlinear state constraints. SPE J. 25, 1938–1963, 04.
Babadagli, T., 2007. Development of mature oil fields—a review. J. Petrol. Sci. Eng. 57
Liu, Z., Reynolds, A.C., 2021. Gradient-enhanced support vector regression for robust
(3–4), 221–246.
life-cycle production optimization with nonlinear-state constraints. SPE J. 26,
Chen, B., Reynolds, A.C., 2016. Ensemble-based optimization of the water-alternating-
1590–1613, 04.
gas-injection process. SPE J. 21 (03): 0786-0798.
Lee, S., Stephen, K., 2019. Field application study on automatic history matching using
Chen, B., Reynolds, A.C., 2018. CO2 water-alternating-gas injection for enhanced oil
particle swarm optimization. In: SPE Reservoir Characterisation and Simulation
recovery: optimal well controls and half-cycle lengths. Comput. Chem. Eng. 113,
Conference and Exhibition. OnePetro.
44–56.
Leeuwenburgh, O., Egberts, P.J., Abbink, O.A., 2010. Ensemble methods for reservoir
Chen, B., Xu, J., 2019. Stochastic simplex approximate gradient for robust life-cycle
life-cycle optimization and well placement. In: SPE/DGS Saudi Arabia Section
production optimization: applied to brugge field. J. Energy Resour. Technol. 141 (9).
Technical Symposium and Exhibition. OnePetro.
Chen, B., Fonseca, R.M., Leeuwenburgh, O., et al., 2017a. Minimizing the risk in the
Lorentzen, R.J., Berg, A.M., Naevdal, G., Vefring, E.H., 2006. A new approach for
robust life-cycle production optimization using stochastic simplex approximate
dynamic optimization of waterflooding problems. Amsterdam, Netherlands. In:
gradient. J. Petrol. Sci. Eng. 153, 331–344.
Proceedings of the SPE Intelligent Energy Conference and Exhibition. Apr. 11–13,
Chen, H., Feng, Q., Zhang, X., et al., 2017b. Well placement optimization using an
No. SPE 99690.
analytical formula-based objective function and cat swarm optimization algorithm.
Lu, R., Forouzanfar, F., Reynolds, A.C., 2017a. An efficient adaptive algorithm for robust
J. Petrol. Sci. Eng. 157, 1067–1083.
control optimization using StoSAG. J. Petrol. Sci. Eng. 159, 314–330.
Chen, Y., Oliver, D.S., Zhang, D., 2009. Efficient ensemble-based closed-loop production
Lu, R., Forouzanfar, F., Reynolds, A.C., 2017b. Bi-objective optimization of well
optimization. SPE J. 14, 634–645, 04.
placement and controls using stosag. In: SPE Reservoir Simulation Conference.
Forouzanfar, F., Poquioma, W.E., Reynolds, A.C., 2016. Simultaneous and sequential
OnePetro.
estimation of optimal placement and controls of wells with a covariance matrix
Nasrabadi, H., Morales, A., Zhu, D., 2012. Well placement optimization: a survey with
adaptation algorithm. SPE J. 21, 501–521, 02.
special focus on application for gas/gas-condensate reservoirs. J. Nat. Gas Sci. Eng.
Fonseca, R.M., Leeuwenburgh, O., Van den Hof, P.M.J., et al., 2015a. Improving the
5, 6–16.
ensemble-optimization method through covariance-matrix adaptation. SPE J. 20,
Nwaozo, J.E., 2006. Dynamic Optimization of a Water Flood Reservoir. University of
155–168, 01.
Oklahoma.
Fonseca, R.R.M., Chen, B., Jansen, J.D., et al., 2017. A stochastic simplex approximate
Tabatabaei Nejad, S.A., Aleagha, A.A.V., Salari, S., 2007. Estimating optimum well
gradient (StoSAG) for optimization under uncertainty. Int. J. Numer. Methods Eng.
spacing in a Middle East onshore oil field using a genetic algorithm optimization
109 (13), 1756–1776.
approach. In: SPE Middle East Oil and Gas Show and Conference. OnePetro.
Fonseca, R.M., Kahrobaei, S.S., Van Gastel, L.J.T., et al., 2015b. Quantification of the
Tukur, A.D., Nzerem, P., Nsan, N., et al., 2019. Well placement optimization using
impact of ensemble size on the quality of an ensemble gradient using principles of
simulated annealing and genetic algorithm. In: SPE Nigeria Annual International
hypothesis testing. In: SPE Reservoir Simulation Symposium. OnePetro.
Conference and Exhibition. OnePetro.
Hosseini-Moghari, S.M., Morovati, R., Moghadas, M., et al., 2015. Optimum operation of
Van Essen, G.M., Van den Hof, P.M.J., Jansen, J.D., 2011. Hierarchical long-term and
reservoir using two evolutionary algorithms: imperialist competitive algorithm (ICA)
short-term production optimization. SPE J. 16, 191–199, 01.
and cuckoo optimization algorithm (COA). Water Resour. Manag. 29 (10),
Wang, X., Haynes, R.D., Feng, Q., 2016. A multilevel coordinate search algorithm for
3749–3769.
well placement, control and joint optimization. Comput. Chem. Eng. 95, 75–96.
Hanea, R.G., Casanova, P., Wilschut, F.H., et al., 2017. Well trajectory optimization
Yang, H., Kim, J., Choe, J., 2017. Field development optimization in mature oil
constrained to structural uncertainties. In: SPE Reservoir Simulation Conference.
reservoirs using a hybrid algorithm. J. Petrol. Sci. Eng. 156, 41–50.
OnePetro.
Zhang, K., Li, G., Reynolds, A.C., et al., 2010. Optimal well placement using an adjoint
Lerlertpakdee, P., Jafarpour, B., Gildin, E., 2014. Efficient production optimization with
gradient. J. Petrol. Sci. Eng. 73 (3–4), 220–226.
flow-network models. SPE J. 19, 1083–1095, 06.
Liu, X., Reynolds, A.C., 2016. A multiobjective steepest descent method with applications
to optimal well control. Comput. Geosci. 20 (2), 355–374.

24