You are on page 1of 12

Computers and Chemical Engineering 117 (2018) 209–220

Contents lists available at ScienceDirect

Computers and Chemical Engineering


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

A meta-optimized hybrid global and local algorithm for well


placement optimization
Hongwei Chen a, Qihong Feng a,∗, Xianmin Zhang a, Sen Wang a, Zhiyu Ma a,
Wensheng Zhou b, Chen Liu b
a
School of Petroleum Engineering, China University of Petroleum (East China), Qingdao, Shandong 266580, China
b
CNOOC Research Institute, Beijing 100027, China

a r t i c l e i n f o a b s t r a c t

Article history: Well placement optimization is a complex and time-consuming task. An efficient and robust algorithm
Received 16 January 2018 can improve the optimization efficiency. In this work, we propose a meta-optimized hybrid cat swarm
Revised 17 April 2018
mesh adaptive direct search (O-CSMADS) algorithm for well placement optimization. By coupling Cat
Accepted 17 June 2018
Swarm Optimization (CSO) algorithm, Mesh Adaptive Direct Search (MADS) algorithm, and Particle Swarm
Available online 24 June 2018
Optimization (PSO) meta-optimization approach, O-CSMADS has global search ability and local search
Keywords: ability. We perform detailed comparisons of optimization performances between O-CSMADS, hybrid cat
Well placement swarm mesh adaptive direct search (CSMADS) algorithm, CSO, and MADS in three different examples. Re-
Optimization efficiency sults show that O-CSMADS algorithm outperforms stand-alone CSO, MADS, and CSMADS. Besides, optimal
Hybrid optimization algorithm controlling parameters are not same for different problems, which indicates that the optimization of al-
Meta-optimization approach gorithmic parameters is necessary. The proposed method also shows great potential for other petroleum
Cat swarm optimization algorithm
engineering optimization problems, such as well type optimization and joint optimization of well place-
Mesh Adaptive Direct Search
ment and control.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction ical simulator (Siavashi et al., 2016; Isebor et al., 2014; Bouzark-
ouna et al., 2012; Yeten et al., 2002). Reservoir numerical sim-
In order to maximize the oil recovery, many methods involving ulator, which serves as a black box, calculates development per-
water flooding, steam flooding and fracturing have been applied formance under specific well locations by finite difference and
(Wang et al., 2016; Chen and Reynolds, 2018; Feng et al., 2013; Guo finite element solution technology (Singh and Srinivasan, 2013,
et al., 2017; Rui et al., 2018). For oilfield development, determin- 2014). The derivative-free algorithm is used to find the optimal
ing optimal well placement is the essential approach to realize the solution for well placement optimization problem. For derivative-
maximum oil production. Optimal well placement is influenced by free algorithms, the gradient computation is not involved
reservoir heterogeneity, fluid property, economic parameters, and during the optimization process. Derivative-free optimization
other factors (Gildin et al., 2006; Bouzarkouna et al., 2012; Tavallali algorithms show good performance when gradients of the opti-
et al., 2013; Zhang et al., 2017; Feng et al., 2017; Chen et al., 2018). mization problem are not available, difficult to obtain or ill-defined
Optimization algorithms provide a tool to find the optimal well (Isebor et al., 2014). Search strategies of derivative-free algorithms
placement effectively and automatically. Due to the nature of non- contain global search strategy and local search procedure. Ge-
linear, discrete, constrained, and multi-modal, well placement opti- netic Algorithm (GA) (Güyagüler and Horne, 2004; Emerick et al.,
mization problem is a complex and challenging task (Dossary and 2009; Lyons and Nasrabadi, 2013; Sampaio et al., 2015a,b), Parti-
Nasrabai, 2016). It is necessary to find an efficient, reliable cle Swarm Optimization (PSO) (Humphries and Haynes, 2015; On-
and robust algorithm to deal with well placement optimization wunalu and Durlofsky, 2010), Covariance Matrix Adaptation Evolu-
problem. tion Strategy (CMA-ES) (Bouzarkouna et al., 2012; Feng et al., 2016)
Because the optimization surface is rough with multiple lo- and Differential Evolution (DE) (Nwankwor et al., 2013; Awotunde,
cal optimum, well placement optimization problem is usually 2014; Atashnezhad et al., 2017) conduct global optimization strat-
solved by coupling derivative-free algorithm with reservoir numer- egy. Algorithms with local search strategy include Generalized Pat-
tern Search (GPS) (Humphries et al., 2014) and Hooke-Jeeves Di-
rected Search (HJDS) (Aliyev, 2011). By using stochastic parameters

Corresponding author.
in the optimization process to escape from local optimum, global
E-mail addresses: fengqihong.upc@gmail.com, wangsena1@126.com (Q. Feng).

https://doi.org/10.1016/j.compchemeng.2018.06.013
0098-1354/© 2018 Elsevier Ltd. All rights reserved.
210 H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220

2017). MADS algorithm is a local optimization algorithm based


Nomenclature on polling. Due to its excellent local search ability, it has been
combined with Particle Swarm Optimization (PSO) algorithm to
b annual discount rate construct hybrid PSO-MADS procedure (Isebor et al., 2014; Aliyev
c acceleration constant and Durlofsky, 2015; Humphries and Raynes, 2015) and Dif-
c1 controlling parameter in PSO ferential Evolution (DE) algorithm to build DE-MADS algorithm
c2 controlling parameter in PSO (Yang et al., 2017). By applying MADS algorithm to conduct
Cdyn constant in dynamic penalization method the local search, these hybrid algorithms outperform the stand-
Cinj water production costs, USD/m3 alone algorithms in average objective function and standard
Coil oil price, USD/m3 deviation. Controlling parameters have a significant influence on
Cwater water production costs, USD/m3 optimization efficiency. Meta-optimization approach is a powerful
Cwell cost to drill well, USD/well and effective method to find the reasonable controlling parame-
CDC count of dimension to change ters for optimization algorithm (Parsopoulos and Vrahatis, 2002;
eval objective function with penalization method Meissner et al., 2006). Meta-optimization approach applies an op-
fcon constraint violation measures timization algorithm to find the good behavior parameters for an-
i ith well other optimization algorithm. It has been applied to find the opti-
iter iteration step mal controlling parameters for GA, PSO, DE, Ant Colony Optimiza-
j jth well tion (ACO), and hybrid PSO–HJDS (Birattari et al., 2002; Meissner
l well length et al., 2006; Pedersen and Chipperfield, 2010; Aliyev, 2011). The
max maximum controlling parameters yielded by meta-optimization approach can
min minimum effectively improve the optimization performance. In order to avoid
MR mixture ratio trapping into local optimum and find the global optimum effec-
Ncon total number of constraints tively, the algorithm should have both global search ability and lo-
Ninj total number of injectors cal search ability. Therefore, we apply CSO and MADS algorithm to
Nprod total number of producers construct the hybrid algorithm. Because the controlling parameters
NPV net present value have a significant influence on optimization efficiency, PSO meta-
Qinj water injection rate, m3 /d optimized algorithm is coupled in optimization process of the
Qoil oil production rate, m3 /d hybrid algorithm to find the optimal controlling parameters, in-
Qwater water production rate, m3 /d cluding parameters in CSO algorithm and the conversion time from
SMP seeking memory pool CSO to MADS. It is guaranteed to use the optimal controlling pa-
SPC self-position considering rameters for each specific optimization problem. Our established
SRD seeking range of the selected dimension algorithm is a meta-optimized hybrid cat swarm mesh adaptive di-
t production time, s rect search algorithm, denoted by “O-CSMADS”.
T total production time, s This paper is organized as follows. In Section 2, the formula-
w inertia weight tion of well placement optimization problem is presented and con-
x coordinate in x direction for vertical well straint handling method is investigated. In Section 3, we present
xh heel coordinate in x direction. the O-CSMADS algorithm approach. In Section 4, we present op-
X optimization variables timization results for three different well placement optimization
y coordinate in y direction for vertical well cases. In Section 5, we provide some concluding remarks.
yh heel coordinate in y direction.
2. Well placement optimization problem formulation
zh heel coordinates in z direction
α dyn constant in dynamic penalization method
For general oil field development, well placement optimization
β dyn constant in dynamic penalization method
is to maximize the economic value or oil production (Yeten, 2003;
ϕ azimuth angle
Isebor et al., 2014; Humphries et al., 2014). In this work, we opti-
mize the well placement to seek the maximum Net Present Value
algorithms can converge to global optimum with high probability. (NPV). NPV reflects the economic effect caused by well placements
However, the local search ability is poor. When the optimization and it is computed based on the fluid production data generated
problem is nonlinear, nonsmooth and noncontinuous, the global by reservoir numerical simulator. Taking the oil sale prices, water
optimization algorithm has the probability of premature conver- production and injection costs, and drilling cost into account, NPV
gence leading to potential solutions being trapped in local opti- for a two-phase (oil and water) flow reservoir model can be for-
mum (Nwankwor et al., 2013). Local algorithms show excellent lo- mulated as:
 N
cal search performance. However, qualities of solutions found by 
T 
prod N prod

1 (time ) (time )
local algorithms are dependent on the initial guess (Aliyev, 2011). NPV = · Coil Qoil p
− Cwater Qwater p
In order to solve well placement optimization problem ef- time=1 (1 + b ) time
p=1 p=1
ficiently, we propose a meta-optimized hybrid global and lo- Nin j
 N prod +Nin j
 
cal algorithm. CSO algorithm is a global derivative-free (stochas- − Cin j Qin(t jin
)
− Cwell (1)
tic) algorithm that mimics the behavior of cats (Chu et al., in=1 wel l =1
2006). It has been used to solve many engineering problems
shown flexibility, fast convergence and producing highly con- where Qoil means oil production rate, m3 /d; Qwater and Qinj respec-
sistent results abilities. CSO algorithm shows better global op- tively represent water production and injection rate, m3 /d; Coil is
timization performance than GA, PSO, SA and DE in conver- oil sale price, USD/m3 ; Cwater and Cinj respectively denote water
gence speed and residual mean square error (Guo et al., 2016; production and injection costs, USD/m3 ; t is the production time,
Panda et al., 2011; Pradhan and Panda, 2012). In our previous s; T represents the total production time, s; b is the annual dis-
work, we have validated that CSO algorithm outperforms DE al- count rate; Ninj is the total number of injectors; Nprod denotes the
gorithm for well placement optimization problem (Chen et al., total number of producers; Cwell is the drilling cost, USD per well.
H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220 211

In order to guarantee that optimized well placements are Table 1


Controlling parameters for PSO meta-
applicable, reservoir boundary constraints, minimum inter-well
optimization algorithm.
distance constraints and well length constraints are taken into
account. If the optimized well location is out of the range of Dimension w c1 c2
reservoir, the simulation is not performed (Yeten, 2003). Inter-well 5 −0.2 −0.5 2.5
distance should be longer than the minimum inter-well distance to
avoid well interference and the detailed calculation method can be
found in many researches (Jesmani et al., 2016; Zhang et al., 2017).
Well length also should be in a reasonable range to satisfy the fea- ber of function evaluations reaches the maximum, MADS algorithm
sible drilling technique. stops and the optimal solutions are outputted.
Many methods are used to handle constraints for well The meta-optimized global and local algorithm for well place-
placement optimization problem, such as rejection method ment optimization combines the strengths of CSO and MADS. The
(Michalewicz, 1995), repair method (Jesmani et al., 2016), penaliza- optimization efficiency depends upon the controlling parameters.
tion method (Bouzarkouna et al., 2012), filter method (Isebor et al., In order to improve the algorithm efficiency, we also couple PSO
2014) and decoder method (Jesmani et al., 2016). In this work, we meta-optimization approach in the optimization process to estab-
apply dynamic penalization method to deal with constraints. For lish meta-optimized hybrid cat swarm mesh adaptive direct search
dynamic penalization method, dynamic penalization factor is used (O-CSMADS) algorithm. Superswarm and subswarm are included
to construct penalization function. It is related to the generation in PSO meta-optimized approach. Firstly, we initialize the super-
number and increases with the optimization iteration. This method swarm, which contains the controlling parameters of CSO and
can provide higher pressure on unfeasible solutions by penalizing switch time of CSO algorithm to MADS algorithm. Then these con-
the objective function when the optimization process towards the trolling parameters are transmitted to subswarm. With these con-
end (Michalewicz, 1995). The objective function combined dynamic trolling parameters, the subswarm conducts the optimization by
penalization method can be written as: hybrid cat swarm mesh adaptive direct search (CSMADS) algo-
rithm. For subswarm, the initial positions and velocities of cats
 αdyn 
Ncon
β are set randomly. Based on the inputted controlling parameter MR,
eval (X ) = NPV(X ) − Cdyn × iter fcondyn (X ) (2)
con=1
these cats are randomly distributed into seeking mode and tracing
mode. Then the position and velocity of each cat are updated ac-
where eval is the objective function with penalization method; X cording to the controlling parameters SMP, SRD, CDC, and c, which
represents the optimization variables; Cdyn , α dyn and β dyn are the are generated in superswarm. Optimal solutions found by CSO are
constants; Ncon is the total number of constraints; fcon is constraint outputted if the end condition is reached, which is controlled by
violation measure; iter is the iteration step. the switch time. Beginning with these solutions, MADS is applied
to conduct the local search. MADS updates the poll stencil center
3. Optimization approach and step size according to the objective function. When the max-
imum simulation number is reached, optimal solutions and opti-
In this section, we establish O-CSMADS algorithm by combining mal objective function value are sent back to superswarm. Dur-
CSO, MADS, and PSO meta-optimization method. ing the meta-optimization procedure, all subswarms complete one
CSO algorithm is a heuristic and population-based stochastic al- entire optimization procedure in each iteration of superswarm.
gorithm which describes and models the behavior of cats. It is in- The controlling parameters are updated in superswarm and then
troduced by Chu et al. (2006). Cats are divided into two modes by re-inputted into subswarm. When the stop condition is satisfied,
mixture ratio (MR) in the hunting process. One is seeking mode final optimal solutions are found with the optimal controlling pa-
and the other is tracing mode (Pradhan and Panda, 2012). In seek- rameters. Based on the whole optimization process, optimal solu-
ing mode, cats stay in alert and move their positions slowly. In tions can be found after both global searching and local searching
the tracing mode, cats chase the objective very quickly based on with the optimal controlling parameters. Fig. 1 illustrates the op-
their own velocities. Five parameters control the optimization pro- timization flowchart of O-CSMADS algorithm. The differences be-
cess of CSO. Seeking memory pool (SMP) represents the number tween the methodology for well placement optimization proposed
of candidate positions, which is also the size of seeking memory in this work and previous studies are as follows. (1) The compo-
of each cat. Seeking range of the selected dimension (SRD) reflects nent global algorithm used in this work is CSO. In our previous
the change ratio for the selected dimension. Count of dimension work, we have validated that CSO algorithm outperforms DE algo-
to change (CDC) denotes the number of dimensions to be varied. rithm for well placement optimization problem. Due to the great
Self-position considering (SPC) is a Boolean variable and it deter- potential of CSO for well placement optimization, we apply CSO al-
mines whether the present position will be the candidate posi- gorithm to construct the hybrid algorithm. (2) The controlling pa-
tion. c denotes the acceleration constant which controls the trac- rameters for the proposed hybrid algorithm are optimized for each
ing mode. When these controlling parameters are set properly, CSO specific well placement optimization problem.
shows good global search ability in convergence speed and robust- In CSO algorithm, SMP is the number of candidate positions.
ness (Guo et al., 2016; Panda and Pradhan, 2012). SMP can reflect the objective function evaluation number when
MADS is a local derivative-free optimization algorithm (Audet the total iteration step is known. That is, SMP controls the switch
and Dennis, 2006; Isebor et al., 2014). It explores the local opti- time of CSO algorithm to MADS algorithm. Therefore, optimiza-
mum which provides a better objective function value than other tion parameters in superswarm of O-CSMADS algorithm are SMP,
solutions in its neighborhood. The local optimum is searched using SRD, CDC, MR, and c. The dimension of PSO meta-optimization ap-
polling centered at the current polling stencil (Isebor et al., 2014). proach is 5. For this scale optimization problem, parameters in-
At each iteration, the objective function is calculated at stencil end cluding controlling parameter c1 , c2 and inertia weight w for PSO
point. If the objective function is improved at one trail poll point, meta-optimization approach are set based on previous researches
the poll stencil center is shifted to this better point and the polling (Pedersen and Chipperfield, 2010; Atashnezhad et al., 2014). And
step size is increased. If the objective function can not be improved they are shown in Table 1. Optimization dimension of the sub-
at any stencil end point, the polling step size is decreased. When swarm depends on the number of variables in the well placement
the poll stencil is reduced to the minimum threshold or the num- optimization problem.
212 H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220

Fig. 1. Flowchart of the proposed O-CSMADS algorithm.

4. Case studies case, the joint optimization of locations and control of 12 wells in
a channelized reservoir are solved. The total number of optimiza-
CSO, MADS, CSMADS and O-CSMADS algorithms are applied to tion variables for three cases is respectively 16, 21 and 60. A reser-
solve three well placement optimization problems in this section. voir numerical simulator, ECLIPSE (GeoQuest, 2010), is applied to
The optimization problem in the first case is to find the optimal calculate the production data with specific well placements. Opti-
well locations of six vertical wells and well types (producers or in- mized well placements are generated by optimization algorithm in
jectors) of four wells in a two-dimensional reservoir. In the second MATLAB software (MATLAB, 2013). Through coupling ECLIPSE with
case, well locations for three horizontal wells and three vertical MATLAB in the optimization process, the optimal well placement
wells in a three-dimensional reservoir are optimized. In the third can be obtained.
H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220 213

Fig. 2. The log-permeability and porosity distribution for Case 1.

Table 2 Table 3
Economic parameters for NPV used in Case 1. Controlling parameters of stand-
alone CSO and CSMADS algorithms
Parameters Value for Case 1.
3
Oil price Coil , USD/m 400
MR SMP SRD CDC c
Water production cost Cwater , USD/m3 20
Water injection cost Cinj , USD/m3 40 0.5 9 0.1 0.2 2
Annual discount rate b 0
Cost of drilling well Cwell , USD/well 20 million
Table 4
Upper and lower value constraints of control-
ling parameters in O-CSMADS algorithm for
Case 1.

Range MR SMP SRD CDC c

Lower 0.1 2 0.1 0.1 0.2


4.1. Case 1: Placement and type of six vertical wells Upper 0.9 9 0.8 0.8 2

4.1.1. Model description


For the first case, we use a two-dimensional reservoir 4.1.2. Results and discussions
model with 60 × 60 × 1 grid blocks and the block size is CSO, MADS, CSMADS and O-CSMADS algorithms are applied to
20 m × 20 m × 20 m. Permeability and porosity fields (Fig. 2) are find the optimal types and locations for these six wells. The num-
taken from SPE 10 benchmark model (Christie and Blunt, 2001). ber of objective function calculations is 10 0 0 for CSO, MADS, CS-
The SPE 10 model is a comparative solution project proposed by MADS. The number of simulations in one iteration of PSO meta-
the Society of Petroleum Engineers. This reservoir model or the optimization approach is 10 0 0. We assume SPC for CSO algorithm
part of this model has been used as the benchmark model to study is equal to 1.
the waterflooding project in order to compare the performance of In stand-alone CSO optimization process, there are 5 control-
different simulators (Siavashi et al., 2014; Siavashi et al., 2016), dif- ling parameters: SMP, SRD, CDC, MR, and c. Values of these pa-
ferent well placement and control strategies (Jesmani et al., 2016; rameters are set according to related researches (Guo et al., 2016)
Wang et al., 2016). We assume there is a two-phase oil-water with- and they are listed in Table 3. For CSMADS algorithm, the behav-
out capillary pressure. ioral parameters of CSO are same with stand-alone CSO algorithm
There are six vertical wells. Optimal locations for all six wells (listed in Table 3). The switch objective function evaluation num-
are optimized. For this case, the well location of each well is rep- ber of CSO algorithm to MADS algorithm is 500. For O-CSMADS
resented by (x, y). The number of location variables is equal to algorithm, PSO meta-optimization approach is applied to optimize
6 × 2 = 12. Besides, the well type, injector or producer, is also opti- the controlling parameters. Parameters for PSO meta-optimization
mized. There is one injector and one producer at least. Well types approach are listed in Table 1. In order to reduce the optimiza-
for other four wells need to be optimized. Therefore, the optimiza- tion time, the number of subswarm calculations is equal to 15. The
tion dimension is equal to 6 × 2 + 4 = 16. The liquid production rate ranges of controlling parameters for O-CSMADS algorithm are se-
of each producer is equal to 100 m3 /d and the water injection rate lected according to conclusions of Guo et al. (2016). The upper and
of each injector is 400 m3 /d. Total production time is 10 years. Dur- lower value constraints for SMP, SRD, CDC, MR, and c are shown in
ing the optimization process, boundary constraint and minimum Table 4.
inter-well distance constraint are considered. Locations of these six Dynamic penalization method is applied to handle the mini-
vertical wells should be in the area of this model. The inter-well mum inter-well distance constraint in the optimization process. As
distance between any two wells should be larger than 40 m. The shown in Eq. (2), parameters used in dynamic penalization method
objective function is to seek the maximum NPV and related eco- are set to be Cdyn = 5, α dyn = β dyn = 2. Because there are random
nomic parameters are listed in Table 2. parameters involved in the optimization processes of CSO, MADS,
214 H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220

Fig. 3. Average convergence curves of stand-alone CSO, MADS, CSMADS and O-CSMADS algorithms for Case 1.

Table 5 NPV of CSMADS. Therefore, the controlling parameters can influ-


Optimal controlling parameters of O-
ence the optimization performance. In comparison with CSMADS,
CSMADS algorithm for Case 1.
O-CSMADS algorithm performs better optimization ability with
MR SMP SRD CDC c higher objective function and lower standard deviation.
0.47 7.0 0.21 0.73 0.55 O-CSMADS has the advantages of global search ability, local
search ability, and optimal controlling parameters. Compared with
stand-alone CSO and MADS, the standard deviation of O-CSMADS
CSMADS, and O-CSMADS algorithms (Guo et al., 2016; Isebor et al., obtains a 70.89% and 86.29% decrease respectively. The worst NPV
2014; Humphries and Haynes, 2015), the optimization performance found by O-CSMADS is close to the best NPV of CSO and MADS.
is different for each trial. In order to overcome the stochastic prop- The average NPV of O-CSMADS is 33.03% higher than the average
erty of the algorithm, each algorithm is performed 10 times to ob- NPV of CSO and 65.73% higher than the average NPV of MADS. It
tain average performance. demonstrates the validity of our proposed O-CSMADS algorithm.
Fig. 3 presents the average NPV versus the number of simula- The remaining oil saturation distribution is shown Fig. 4. For
tions and boxplots of the final NPVs for 10 runs. The best, worst this optimization problem, the highest NPV is found by O-CSMADS
and average NPV over 10 runs for three algorithms, together with algorithm with NPV of 2.278 × 108 USD. This reservoir can achieve
the standard deviation are provided in the Supplementary mate- good development performance with one injector and five produc-
rial. According to the convergence curves plotted in Fig. 3, we can ers, which is found by these four algorithms. According to the opti-
see that MADS shows gradual convergence at the early stage of the mal well configurations shown in Fig. 4, injector placed in low per-
optimization process and CSO converges faster than MADS. CSO is meability region can achieve a larger sweep area for this reservoir
a global algorithm and it can avoid poor local optimum with high model. It also can be seen that placing a producer in high perme-
robustness. In contrast, MADS is a local algorithm. Due to the local ability region can produce more oil from the reservoir.
search nature, MADS may find a relatively high NPV or a poor lo-
cal optimum which highly depends on the initial solution. As a re- 4.2. Case 2: Placement of six wells in 3-D heterogeneous model
sult, the standard deviation of MADS (standard deviation is 0.620)
is higher than that of CSO (standard deviation is 0.292). The best 4.2.1. Model description
NPV of MADS algorithm (2.074 × 108 USD) is higher than the best In this case, we optimize the locations of six wells in a three-
NPV of CSO algorithm (2.059 × 108 USD). However, according to the dimensional reservoir model with 60 × 60 × 3 grid. The block size
average NPV and the standard deviation, the search performance of we set is 20 m × 20 m × 10 m. The permeability and porosity field,
CSO is better than that of MADS. shown in Fig. 5, is a cut-off of SPE 10 model (Christie and
CSMADS algorithm, which combines the advantages of CSO Blunt, 2001) from the third to the fifth layers. The simulation
and MADS, shows better search ability than stand-alone CSO and model involves a two-phase oil-water flow. We assume there is no
MADS. Fig. 3 also show that although the search ability of CSMADS capillary pressure.
is improved by combining CSO and MADS algorithms, the improve- There are two horizontal producers and one horizontal injector.
ment degree is not significant. The roughness of solution surface We assume that the horizontal well is placed in one layer. Five pa-
and the values of controlling parameters have significant influ- rameters are used to describe the location of horizontal well: heel
ences on the optimization performance. For O-CSMADS algorithm, coordinates xh , yh , zh , well length l, and azimuth angle ϕ . The
the controlling parameters are optimized by PSO meta-optimized variable number for horizontal well is equal to 15. Besides, one
approach. The optimal controlling parameters for O-CSMADS are vertical injector and two producers are drilled in this reservoir. Lo-
listed in Table 5. cation of the vertical well is determined by two parameters x, y.
With the optimized controlling parameters, the robustness of Therefore, the total number of optimization variables is 21. The in-
O-CSMADS is significantly improved. The standard diversion of jector is controlled with constant water injection rate 600 m3 /d
O-CSMADS is 0.085, while the standard deviation of CSMADS is and the producer is controlled with constant liquid production rate
0.266. Besides, with these optimal controlling parameters, the ex- 300 m3 /d. The system is simulated for 10 years.
ploration ability of O-CSMADS at the early stage is improved. Boundary constraint, well length constraint, and minimum
The worst NPV found by O-CSMADS is higher than the average inter-well distance constraint are considered during the optimiza-
H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220 215

Fig. 4. Oil saturation distribution under the optimal well locations for Case 1.

Table 6 Table 7
Controlling parameters of CSO algo- Upper and lower value constraints of control-
rithms for Case 2. ling parameters in O-CSMADS algorithm for
Case 2.
MR SMP SRD CDC c
Range MR SMP SRD CDC c
0.5 10 0.1 0.2 0.8
Lower 0.1 2 0.1 0.1 0.2
Upper 0.9 15 0.8 0.8 2

tion process. For this case, the length of a horizontal well ranges
from 200 m to 10 0 0 m. The inter-well distance should be larger Table 8
Optimal controlling parameters of O-
than 40 m. Case 1 deal with the cost of drilling well as a constant. CSMADS algorithm for Case 2.
We now treat the cost of drilling well as a function of well length.
MR SMP SRD CDC c
The basic drilling cost is equal to 18 million USD per well and the
drilling cost per meter is 5 × 104 USD. Other related economic pa- 0.76 12.0 0.14 0.71 0.64
rameters are listed in Table 2.

4.2.2. Results and discussions it demonstrates that the optimal control parameters for different
CSO, MADS, and O-CSMADS algorithms are respectively applied optimization problems are different. They should be set appropri-
to deal with this optimization problem. The maximum number of ately based on the specific optimization problem to improve the
objective function evaluation is equal to 20 0 0. Parameters for PSO optimization efficiency. PSO meta-optimization approach can find
meta-optimization approach are listed in Table 1. Parameters for the optimal controlling parameters for algorithms. For O-CSMADS
CSO are presented in Table 6. For O-CSMADS algorithm, the con- algorithm, CSO algorithm can conduct 60%–70% of the total num-
straints for SMP, SRD, CDC, MR, and c are shown in Table 7. In or- ber of objective function evaluations and then the solutions found
der to handle constraints, dynamic penalization method is applied by CSO are input as the initial populations of MADS.
and values of parameters are shown in Case 1. We analyze the op- Fig. 6 displays the average convergence curves and boxplot
timization performance based on 10 runs of each algorithm. comparisons between CSO, MADS, and O-CSMADS. The data set
The optimization results of controlling parameters for O- is presented in the Supplementary material. It again shows that
CSMADS algorithm are shown in Table 8. According to the opti- the O-CSMADS algorithm outperforms its component algorithms in
mal controlling parameters of O-CSMADS algorithm in two cases, term of average NPV, best NPV, and robustness. The average NPV
216 H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220

Fig. 5. The log-permeability and porosity distribution for Case 2.

Fig. 6. Average convergence curves of stand-alone CSO, MADS and O-CSMADS algorithms for Case 2.

of O-CSMADS is 17.12% higher than the average NPV of CSO and initial solutions of MADS. Therefore, O-CSMADS takes use of the
24.27% higher than the average NPV of MADS. The worst NPV ob- positive features of CSO, MADS, and PSO meta-optimization ap-
tained by O-CSMADS is higher than the average NPV found by CSO proach. In comparison with stand-alone CSO and MADS algorithms,
and MADS. Of the 10 runs for each algorithm, the standard devia- O-CSMADS shows the best search ability and robustness.
tion of O-CSMADS is only 0.143, while the standard deviations of The best well configurations for CSO, MADS, and O-CSMADS
CSO and MADS are 0.309 and 0.501, respectively. algorithms are presented in Fig. 7(a)–(c). The distributions of re-
From Fig. 6(a) we can see that the convergence curve of O- maining oil saturation are shown in Fig. 7(d)–(f). It can be seen
CSMADS with the optimized controlling parameters is higher than that the horizontal injector is placed in the middle layer which
that of CSO algorithm at early iteration stage of the optimization is found by this three algorithms. Besides, a horizontal producer
process. It illustrates that controlling parameters affect the opti- should be placed in the middle layer to obtain the highest NPV,
mization process. With the optimized controlling parameters, the which is found by O-CSMADS.
search ability is improved. According to the boxplots in Fig. 6(b),
it can be seen that the standard deviation of stand-alone MADS 4.3. Case 3: Joint optimization of well placement and control in
algorithm is the highest and the worst NPV of the MADS is the channelized model
lowest. It reflects that there are many local optimum solutions in
this case and stand-alone MADS algorithm easily gets stuck in lo- 4.3.1. Model description
cal optimum. MADS also shows an excellent local search ability to A three-dimensional channelized reservoir model is used in this
obtain a relatively high NPV, which is higher than the best NPV case. The model is the Egg model, which has high-permeability
found by stand-alone CSO. The optimization performance of MADS channels in the low-permeable background to represent the typical
highly depends on the initial solution. Therefore, in our proposed meandering river patterns (Jansen et al., 2015). There are 101 per-
method, the optimal solutions found by CSO are inputted as the meability realizations in standard Egg model. Egg model has been
H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220 217

Fig. 7. Optimized well locations and remaining oil saturation distribution for Case 2.

Table 9
Economic parameters for NPV used in Case 3.

Parameters Value

Oil price Coil , USD/m3 450


Water production cost Cwater , USD/m3 60
Water injection cost Cinj , USD/m3 50
Annual discount rate b 0

Table 10
Controlling parameters of CSO algo-
rithm for Case 3.

MR SMP SRD CDC c

0.5 12 0.1 0.2 0.8

neously. Therefore, the total number of optimization variables is


Fig. 8. The permeability distribution for Case 3. 12 × 2 + 12 × 3 = 60.
During the optimization process, all wells should be located in
the area of the model and the inter-well distance should be larger
than 32 m. The water injection rate or liquid production rate of ev-
applied in many researches related to optimization under geologi- ery well is from 40 m3 /d to 120 m3 /d. The total injection rate or
cal uncertainty (Shirangi and Durlofsky, 2016; Fonseca et al., 2017), production rate is from 320 m3 /d to 480 m3 /d. Besides, the to-
well placement optimization (Wang et al., 2016; Zandvliet et al., tal water injection rate is equal to the whole liquid production
2008), and well control optimization (Zandvliet et al., 2007; Feng rate. The related economic parameters to calculate NPV are listed
et al., 2016). There are 60 × 60 × 7 = 25,200 grid cells. For initial in Table 9.
model, there are 18,533 active cells. In our work, we modify the
model to make all cells active. The grid size is 8 m × 8 m × 4 m and 4.3.2. Results and discussion
the porosity is 0.2. The permeability distribution we used in this This optimization problem is respectively optimized by CSO,
case is shown in Fig. 8. The total production time is 3240 days. MADS and O-CSMADS algorithms. Given the complexity of this op-
There are 12 vertical wells containing 8 injectors and 4 pro- timization problem, the total number of simulation runs is 50 0 0.
ducers. Each well is controlled by constant water injection rate Behavior parameters of stand-alone CSO are listed in Table 10. For
or liquid production rate. The control interval is 1080 days and O-CSMADS algorithm, the controlling parameters are optimized by
each well has 3 control parameters. Location of the vertical well PSO meta-optimized approach. Parameters for PSO meta-optimized
is represented by two variables x, y. For this case, the optimiza- approach are shown in Table 1 and constraints for controlling pa-
tion problem is to find the optimal well location. Besides, the rameter are given in Table 11. Dynamic penalization method is
well control strategies of these well are also optimized simulta- used to deal with the optimization constraints which has been
218 H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220

Fig. 9. Average convergence curves of stand-alone CSO, MADS and O-CSMADS algorithms for Case 3.

Fig. 10. Optimized well locations and control strategies for Case 3.

Table 11 Table 12
Upper and lower value constraints of control- Optimal controlling parameters of O-
ling parameters in O-CSMADS algorithm for CSMADS algorithm for Case 3.
Case 3.
MR SMP SRD CDC c
Range MR SMP SRD CDC c
0.66 10.0 0.31 0.53 0.84
Lower 0.1 2 0.1 0.1 0.2
Upper 0.9 20 0.8 0.8 2

case are not same with that of Case 1 and Case 2. Therefore, the
controlling parameters should be set differently for the different
used in Case 1 and Case 2. We analysis the optimization perfor- optimization problem, which is also concluded in Case 2.
mance of each algorithm after 10 runs. Plots of the average NPV of CSO, MADS, and O-CSMADS al-
Optimal controlling parameters found by PSO meta- gorithms versus the number of simulations and the boxplots are
optimization approach for O-CSMADS are shown in Table 12. shown in Fig. 9. Detailed results about the best, worst, average and
It can be seen that the optimal controlling parameters of this standard deviation of NPV for each algorithm is collected in Sup-
H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220 219

plementary material. Even though the best NPV found by MADS is cific optimization problem. Our proposed algorithm O-CSMADS can
higher than the best NPV found by CSO, the average performance effectively solve the well placement optimization problem. Besides,
of CSO outperforms the average performance of MADS, which is this algorithm also shows great potential in dealing with well type
consistent with the observation in Case 1 and Case 2. The reason optimization problem and joint optimization of well placement
for this phenomenon is that MADS easily gets stuck in local optima and control problem.
with the randomly initialized solutions and the local search ability
of MADS is excellent. CSO is a global optimization algorithm and it
Acknowledgments
easily gets out of local optimum.
The performance of O-CSMADS algorithm is the best. It obtains
This research is supported by the National Science and Tech-
the average NPV of 1.746 × 108 USD, which is 9.75% higher than
nology Major Project of China (Grant No. 2016ZX05025001-
the average NPV of CSO algorithm and 12.35% higher than the av- 006), National Natural Science Foundation of China (51704312,
erage NPV of MADS algorithm. The worst NPV found by O-CSMADS
51474233, U1762213), National Postdoctoral Program for Innova-
is 1.677 × 108 USD, which is higher than the best NPV of MADS
tive Talents (BX201600153), China Postdoctoral Science Foundation
(1.657 × 108 USD) and close to the best NPV of CSO (1.680 × 108 (2016M600571), and Fundamental Research Funds for the Central
USD).
Universities (18CX07006A).
For O-CSMADS algorithm, the optimization process is conducted
by CSO algorithm with the optimized controlling parameters at
the early optimization stage. In comparison with the convergence Supplementary materials
curve of stand-alone CSO algorithm, the convergence curve of O-
CSMADS is higher. By using the initial solution provided by CSO Supplementary material associated with this article can be
algorithm, MADS conducts optimization process at the later stage. found, in the online version, at doi:10.1016/j.compchemeng.2018.
According to the optimization results, we can see that the ro- 06.013.
bustness of MADS is significantly improved in O-CSMADS algo-
rithm. With the standard deviation of 0.046, the standard devia- References
tion of O-CSMADS represents a 33.33% decrease in the standard
Aliyev, E., 2011. Use of Hybrid Approaches and Metaoptimization for Well Placement
deviation of CSO and a 48.89% decrease in the standard deviation
Problems. Thesis. Stanford University.
of MADS. Aliyev, E., Durlofsky, L., 2015. Multilevel field-development optimization using a se-
The best well configuration and well control strategy found quence of upscaled models. In: SPE Reservoir Simulation Symposium. 23–25
by each algorithm are shown in Fig. 10. Fig. 10(a)–(c) shows the February 2015, USA.
Atashnezhad, A., Cedola, A., Hareland, G., 2017. An empirical model to estimate a
pore-volume-weighted remaining oil saturation distribution for the critical stimulation design parameter using drilling data. In: SPE Western Re-
7 layers of the channelize reservoir and the optimal well loca- gional Meeting, 23–27 April 2017. California, USA.
tions. Fig. 10(d)–(f) shows the best well control strategy. Optimal Atashnezhad, A., Wood, D., Fereidounpour, A., Khosravanian, R., 2014. Designing
and optimizing deviated wellbore trajectories using novel particle swarm algo-
well configuration and control strategy obtained by O-CSMADS can rithms. J. Nat. Gas Sci. Eng. 21, 1184–1204.
achieve the highest NPV. It can be seen that the optimal well loca- Audet, C., Dennis, J., 2006. Mesh adaptive direct search algorithms for constrained
tions found by CSO are similar to the optimal well locations found optimization. SIAM J. Optim. 17 (1), 188–217.
Awotunde, A., 2014. On the joint optimization of well placement and control. In:
by MADS. While there is a big difference between the optimal SPE Saudi Arabia Section Technical Symposium and Exhibition. 21–24 April
well locations found by O-CSMADS and that of CSO and MADS. 2014, Saudi Arabia.
The reason for this phenomenon is possible that CSO and MADS Birattari, M., Stützle, T., Paquete, L., Varrentrapp, K., 2002. A racing algorithm for
configuring metaheuristics. In: Proceedings of the Genetic and Evolutionary
algorithms trap into the local optimum. By using O-CSMADS, the
Computation Conference (GECCO), pp. 11–18.
solutions get out of local optimum to convergence the global op- Bouzarkouna, Z., Ding, D., Auger, A., 2012. Well placement optimization with the
timum. According to the remaining oil saturation distribution, it covariance matrix adaptation evolution strategy and meta-models. Comput.
Geosci. 16, 75–92.
can be seen that the well locations and control strategies found by
Chen, B., Reynolds, A., 2018. CO2 water-alternating-gas injection for enhanced oil
O-CSMADS obtains the least remaining oil saturation and achieves recovery: optimal well controls and half-cycle lengths. Comput. Chem. Eng. 113,
the best production performance. 44–56.
Chen, H., Feng, Q., Zhang, X., Wang, S., Zhou, W., Geng, Y., 2017. Well placement op-
timization using an analytical formula-based objective function and cat swarm
5. Conclusions optimization algorithm. J. Pet. Sci. Eng. 157, 1067–1083.
Chen, H., Feng, Q., Zhang, X., Zhou, W., Geng, Y., 2018. A prediction formula for ratio
In this work, we propose an efficient algorithm for well place- of injection–production control area in triangle well pattern. J. Pet. Explor. Prod.
Technol. 8 (1), 195–203.
ment optimization problem. By coupling CSO algorithm, MADS Christie, M., Blunt, M., 2001. Tenth SPE comparative solution project: a comparison
algorithm, and PSO meta-optimization approach, we construct a of upscaling techniques. SPE Reservoir Eval. Eng. 4 (4), 308–317.
meta-optimized hybrid cat swarm mesh adaptive direct search Chu, S., Tsai, P., Pan, J., 2006. Cat Swarm Optimization. In: Lecture Notes Computer
Science, vol. 6, pp. 854–858.
(O-CSMADS) algorithm. We apply CSO, MADS, CSMADS, and O- Dossary, M., Nasrabadi, H., 2016. Well placement optimization using imperialist
CSMADS algorithms in three different optimization problems. We competitive algorithm. J. Pet. Sci. Eng. 147, 237–248.
validate the optimization efficiency of the proposed algorithm in Emerick, A., Silva, E., Messer, B., Almeida, L., Szwarcman, D., Pacheco, M., Vel-
lasco, M., 2009. Well placement optimization using a genetic algorithm with
vertical well placement optimization, horizontal well placement nonlinear constraints. In: SPE Reservoir Simulation Symposium. 2–4 February
optimization, well type optimization, and joint optimization of well 2009, Texas, USA.
placement and control. Results demonstrate that O-CSMADS algo- Feng, Q., Chen, H., Wang, X., Wang, S., Wang, Z., Yang, Y., Bing, S., 2016. Well control
optimization considering formation damage caused by suspended particles in
rithm has excellent global search ability and local search ability. O-
injected water. J. Nat. Gas. Sci. Eng. 35, 21–32.
CSMADS shows better performance than CSO, MADS, and CSMADS Feng, Q., Wang, S., Zhang, W., Song, Y., Song, S., 2013. Characterization of high-
algorithms in terms of optimization efficiency and robustness. PSO -permeability streak in mature waterflooding reservoirs using pressure transient
analysis. J. Pet. Sci. Eng. 110, 55–65.
meta-optimization approach can effectively determine the opti-
Feng, Q., Zhang, J., Wang, S., Wang, X., Cui, R., Wang, D., Bing, S., Rui, Z., 2017. Uni-
mal controlling parameters for O-CSMADS algorithm. With opti- fied relative permeability model and waterflooding type curves under different
mized controlling parameters, O-CSMADS algorithm outperforms levels of water cut. J. Pet. Sci. Eng. 154, 204–216.
CSMADS. Besides, the optimal controlling parameters of O-CSMADS Fonseca, R., Chen, B., Jansen, J., Reynolds, A., 2017. A stochastic simplex approximate
gradient (StoSAG) for optimization under uncertainty. Int. J. Numer. Meth. Eng.
algorithm are different for three examples. Therefore, it is neces- 109 (13), 1756–1776.
sary to apply the appropriate controlling parameters for the spe- GeoQuest, S., 2010. ECLIPSE reference manual 2010.1. Schlumberger.
220 H. Chen et al. / Computers and Chemical Engineering 117 (2018) 209–220

Gildin, A., Klie, H., Rodriguez, A., Wheeler, M., Bishop, R., 2006. Development of Rui, Z., Wang, X., Zhang, Z., Lu, J., Chen, G., Zhou, X., Patil, S., 2018. A realistic and in-
low-order controllers for high-order reservoir models and smart wells. In: SPE tegrated model for evaluating oil sands development with steam assisted grav-
Annual Technical Conference and Exhibition, 24–27 September 2006, Texas, ity drainage technology in Canada. Appl. Energ. 213, 76–91.
USA. Sampaio, M., Barreto, C., Schiozer, D., 2015a. Assisted optimization method for com-
Guo, L., Meng, Z., Sun, Y., Wang, L., 2016. Parameter identification and sensitivity parison between conventional and intelligent producers considering uncertain-
analysis of solar cell models with cat swarm optimization algorithm. Energy ties. J. Pet. Sci. Eng. 133, 268–279.
Convers. Manage. 108, 520–528. Sampaio, M., Gildin, E., Schiozer, D., 2015b. Short-term and long-term optimizations
Guo, T., Li, Y., Ding, Y., Qu, Z., Gai, N., Rui, Z., 2017. Evaluation of acid fracturing for reservoir management with intelligent wells. In: SPE Latin American and
treatments in shale formation. Energ. Fuel 31 (10), 10479–10489. Caribbean Petroleum Engineering Conference. 18-20 November 2015, Buenos
Güyagüler, B., Horne, R., 2004. Uncertainty assessment of well placement optimiza- Aires, Argentina.
tion. SPE J. 7 (1), 24–32. Shirangi, M., Durlofsky, L., 2016. A general method to select representative models
Humphries, T., Haynes, R., 2015. Joint optimization of well placement and control for decision making and optimization under uncertainty. Comput. Geosci. 96,
for nonconventional well types. J. Pet. Sci. Eng. 126, 242–253. 109–123.
Humphries, T., Haynes, R., James, L., 2014. Simultaneous and sequential approaches Siavashi, M., Blunt, M., Raisee, M., Pourafshary, P., 2014. Three-dimensional stream-
to joint optimization of well placement and control. Comput. Geosci. 18, line-based simulation of non-isothermal two-phase flow in heterogeneous
433–448. porous media. Comput. Fluids 103, 116–131.
Isebor, O., Durlofsky, L., Ciaurri, D., 2014. A derivative-free methodology with local Siavashi, M., Tehrani, M., Nakhaee, A., 2016. Efficient particle swarm optimization of
and global search for the constrained joint optimization of well locations and well placement to enhance oil recovery using a novel streamline-based objec-
controls. Comput. Geosci. 18 (3-4), 463–482. tive function. J. Energy Resour. 138 (5), 77–78.
Jansen, J., Fonseca, R., Kahrobaei, S., Siraj, M., Van Essen, G., Van den Holf, P., 2015. Singh, H., Srinivasan, S., 2013. Uncertainty analysis by model selection technique
The egg model-a geological ensemble for reservoir simulation. Geosci. Data J. 1 and its application in economic valuation of a large field. In: SEP North Africa
(2), 192–195. Technical Conference and Exhibition, 15–17 April 2013, Cairo, Egypt.
Jesmani, M., Bellout, M., Hanea, R., Foss, B., 2016. Well placement optimization Singh, H., Srinivasan, S., 2014. Scale up of reactive processes in heterogeneous me-
subject to realistic field development constraints. Comput. Geosci. 20 (6), dia-numerical experiments and semi-analytical modeling. In: SPE SPE Improved
1185–1209. Oil Recovery Symposium, 12–16 April 2014, Oklahoma, USA.
Lyons, J., Nasrabadi, H., 2013. Well placement optimization under time-dependent Tavallali, M., Karimi, I., Teo, K., Baxendale, D., Ayatollahi, S., 2013. Optimal producer
uncertainty using an ensemble Kalman filter and a genetic algorithm. J. Pet. Sci. well placement and production planning in an oil reservoir. Comput. Chem. Eng.
Eng. 109, 70–79. 55, 109–125.
MATLAB R, 2013. Version 8.1. 0.604 (R2013a). The MathWorks Inc, Natick. Wang, X., Haynes, R., Feng, Q., 2016. A multilevel coordinate search algorithm for
Meissner, M., Schmuker, M., Schneider, G., 2006. Optimized Particle Swarm Opti- well placement, control and joint optimization. Comput. Chem. Eng. 95, 75–96.
mization (OPSO) and its application to artificial neural network training. BMC Yang, H., Kim, J., Choe, J., 2017. Field development optimization in mature oil reser-
Bioinf. 7 (1), 125. voirs using a hybrid algorithm. J. Pet. Sci. Eng. 156, 41–50.
Michalewicz, Z., 1995. A survey of constraint handling techniques in evolutionary Yeten, B., 2003. Optimum Deployment of Unconventional Wells. Ph.D. thesis. Stan-
computation methods. Evol. Program. 4, 135–155. ford University.
Nwankwor, E., Nagar, A., Reid, D., 2013. Hybrid differential evolution and particle Yeten, B., Durlofsky, L., Aziz, K., 2002. Optimization of nonconventional well type,
swarm optimization for optimal well placement. Comput. Geosci. 17, 1–20. location and trajectory. SPE 77565.
Onwunalu, J., Durlofsky, L., 2010. Application of a particle swarm optimization al- Zandvliet, M., Bosgra, O., Jansen, J., Van den Hof, P., Kraaijevanger, J., 2007.
gorithm for determining optimum well location and type. Comput. Geosci. 14, Bang-bang control and singular arcs in reservoir flooding. J. Pet. Sci. Eng. 58,
183–198. 186–200.
Panda, G., Pradhan, P., Majhi, B., 2011. IIR system identification using cat swarm op- Zandvliet, M., Handels, M., Van Essen, G., Brouwer, R., Jansen, J., 2008. Adjoin-
timization. Expert Syst. Appl. 38 (10), 12671–12683. t-based well-placement optimization under production constraints. SPE J. 13 (4),
Parsopoulos, K., Vrahatis, M., 2002. Recent approaches to global optimization prob- 392–399.
lems through particle swarm optimization. Nat. Comput. 1 (2-3), 235–306. Zhang, Y., Lu, R., Forouzanfar, F., Reynolds, A., 2017. Well placement and control opti-
Pedersen, M., Chipperfield, A., 2010. Simplifying particle swarm optimization. Appl. mization for WAG/SAG processes using ensemble-based method. Comput. Chem.
Soft Comput. 10 (2), 618–628. Eng. 101, 193–209.
Pradhan, P., Panda, G., 2012. Solving multiobjective problems using cat swarm opti-
mization. Expert. Syst. Appl. 39 (3), 2956–2964.

You might also like