Fuel 308 (2022) 121872

Contents lists available at ScienceDirect

Fuel
journal homepage: www.elsevier.com/locate/fuel

Full Length Article

Robust hybrid machine learning algorithms for gas flow rates prediction
through wellhead chokes in gas condensate fields
Abouzar Rajabi Behesht Abad a, Hamzeh Ghorbani b, *, Nima Mohamadian b, Shadfar Davoodi c,
Mohammad Mehrad d, Saeed Khezerloo-ye Aghdam e, Hamid Reza Nasriani f
a
Department of Petroleum Engineering, Omidiyeh Branch, Islamic Azad University, Omidiyeh, Iran
b
Young Researchers and Elite Club, Ahvaz Branch, Islamic Azad University, Ahvaz, Iran
c
School of Earth Sciences & Engineering, Tomsk Polytechnic University, Lenin Avenue, Tomsk, Russia
d
Faculty of Mining, Petroleum and Geophysics Engineering, Shahrood University of Technology, Shahrood, Iran
e
Department of Petroleum Engineering, Amirkabir University of Technology, Tehran, Iran
f
School of Engineering, Faculty of Science and Technology, University of Central Lancashire, Preston, United Kingdom

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

Keywords: Condensate reservoirs are the most challenging hydrocarbon reservoirs in the world. The behavior of condensate
Gas flow rate gas reservoirs regarding pressure and temperature variation is unique. Adjusting fluid flow rate through wellhead
Multi-hidden layer extreme learning machine chokes of condensate gas wells is critical and challenging for reservoir management. Predicting this vital
Hybrid machine learning algorithms
parameter is a big step for the development of condensate gas fields. In this study, a novel machine learning
Least squares support vector machine
approach is developed to predict gas flow rate (Qg) from six input variables: temperature (T); upstream pressure
Wellhead choke
(Pu); downstream pressure (Pd); gas gravity (γg); choke diameter (D64) and gas–liquid ratio (GLR). Due to the
absence of accurate recombination methods for determining Qg, machine learning methods offer a functional
alternative approach. Four hybrid machine learning (HML) algorithms are developed by integrating multiple
extreme learning machine (MELM) and least squares support vector machine (LSSVM) with two optimization
algorithms, the genetic algorithm (GA) and the particle swarm optimizer (PSO). The evaluation conducted on
prediction performance and accuracy of the four HML models developed indicates that the MELM-PSO model has
the highest Qg prediction accuracy achieving a root mean squared error (RMSE) of 2.8639 Mscf/d and a coef­
ficient of determination (R2) of 0.9778 for a dataset of 1009 data records compiled from gas-condensate fields
around Iran. Comparison of the prediction performance of the HML models developed with those of the previous
empirical equations and artificial intelligence models reveals that the novel MELM-PSO model presents superior
prediction efficiency and higher computational accuracy. Moreover, the Spearman correlation coefficient anal­
ysis performed demonstrates that D64 and GLR are the most influential variables in the gas flow rate for the large
dataset evaluated in this study.

engineering concerns adversely affects wells’ productivity and shortens

1. Introduction their production life [5]. Such problems will be exacerbated, especially
in unconventional gas reservoirs with tight carbonated structures and
Hydrocarbon fuels are still recognized worldwide as the driving force very low permeability [6]. The unique phase behavior of condensate gas
and strategic energy to develop leading economic and industrial goals makes the production rate control techniques even more challenging
[1–3]. A sustainable production approach from hydrocarbon reservoirs and vital in such reservoirs. [7]. In condensate reservoirs, the production
is an essential production management policy that enables upstream rate declines significantly due to the accumulation of unproducible
companies to exploit hydrocarbon reservoirs efficiently [4]. Regardless liquid in the near-wellbore region [8]. The reservoir fluid in the regions
of the economic perspective, controlling the production rate by well­ far from the wellbore is a combination of rich gas and non-moveable
head chokes is the most important management lever for optimizing the connate water. At the early production stage, the pressure drops
production process. Increasing the production rate without involving below the dew point near the wellbore region, and the rich gas is

* Corresponding author.
E-mail addresses: [email protected] (H. Ghorbani), [email protected] (N. Mohamadian), [email protected] (S. Davoodi), hrnasriani@
uclan.ac.uk (H.R. Nasriani).

https://doi.org/10.1016/j.fuel.2021.121872
Received 19 March 2021; Received in revised form 8 July 2021; Accepted 28 August 2021
Available online 16 September 2021
0016-2361/© 2021 Elsevier Ltd. All rights reserved.
A.R.B. Abad et al. Fuel 308 (2022) 121872

Nomenclature M, l, O Experimental coefficients

MELM Multiple Extreme Learning Machine
ANN Artificial Neural Network MLP Multi-Layer Perceptron
ANFIS Adaptive Neuro-Fuzzy Inference System N Number of samples in dataset
b Bias vector n Number of inputs parameters
BP Backpropagation PSO Particle swarm optimization
CFD Cumulative distribution functions Pb The cognitive best value of particle
c1 Positive cognitive coefficient (individual learning factors Pwh Wellhead pressure
PSO) Pd Downstream pressure
c2 Positive social coefficient (global learning factor for PSO) Pu Upstream pressure
D64 Choke size Qg Gas flow rate
E Mean value for variable E Qliq Rate of liquids production
Ei Input variable value of data record i RBF Radial basis function
ELM Extreme Learning Machine RMSE Root mean square error
F Mean value for dependent variable F SVM Support Vector Machines
Fi Input variable value of data record i T Transpose matrix
FN Functional Network Vi Particle ith velocity in PSO swarm
GA Genetic algorithm W Inertial weight (PSO)
Gb The global best value found in the swarm w Weight vector
GEP Gene expression programming Xi Particle ith position in PSO swarm
GLR Gas to liquid ratio △p Differential pressure
LSSVM Least Squares Support Vector Machine ϕ(xi ) Kernel function

Table 1
Empirical equations proposed by some researchers to determine the flow rate of condensate gas through wellhead chokes.
Year Authors/ Formula Dataset Units Coefficient R2 Accuracy
Reference metric

1990 Osman & Pu b *D64 c 87 data Qg: Mscf/d, Pu: Psia, a = 0.00130, b = 1, c = 1.8298, – Best result:
Dokla [27] Qg = a* points D64: inch, LGR: STB/ d = 0.5598 AAPD% =
LGRd
MScf 10.64
2008 Al-Attar [29] Qg = a*Δpb *D64 c *GLRd 97 data Qg: MMscf/d, Δp: a = 3.37230e-5, b = 1, c = Best Best result:
points Psi, D64: inch, LGR: 1.15537, d = 0.84695 result: AAPD% =
STB/MScf 0.9521 7.144
2015 Seidi and Δpb *D64 c 106 data Qg: MMscf/d, Δp: a = 0.015, b = 0.65, c = 1.27, d Best Best result:
Qg = a*
Sayahi [31] LGRd points Psi, D64: inch, LGR: = 0.4 result: APD%= 23.93
STB/MMscf 0.9161
2017 Ghorbani 92 data Qg: Mscf/d, Pu and a = 0.0001, b = 2.3481935, c = 0.9677 APD%= 5.32
et al. [18] Qg = points Pd: Psig, D64: inch, 1, d = 0.0001, e = 1:0360972, f
γg: -, T: 0F = 1.498291
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
) √( )d [( )e ( )f ]
(
bPu c √
√ 1 Pd Pd
aD64 −
14.7 γg T Pu Pu
2019 Nasriani et al. Δpb *D64 c 234 data Qg: MMscf/d, Δp: a = 0.0437, b = 0.4836, c = Best Best result:
Qg = a*
[30] LGRd points Psi, D64: inch, LGR: 1.1136, d = 0.3129 result: AAPD%= 8.71
STB/MMscf 0.97

converted into condensate. This isothermal condensation is known as engineers [14]. The values determined in these measurements are the
retrograde condensation [9]. The accumulation of valuable condensate basic input parameters for calculation in many reservoir performance
droplets around the wellbore, also known as the condensate bank/ring, relationships. Determination of multiphase flow rate is crucial in plan­
has not yet reached critical saturation for portability, resulting in a ning and adopting correct measures and reforms in production policies
positive skin factor [10]. Production from gas condensate reservoirs commensurate with the reservoir’s performance during operation [15].
requires meticulous planning and management [11]. Scheduled pro­ The back pressure applying by wellhead chokes has several advantages,
duction plans for sale and export contracts of gas and gas condensate such as stabilizing the multiphase flow rate [16], preventing further
productive [12] require continuous production at the desired rate. Any pressure drop at the bottom hole section and condensate drop out,
disruption to the production process may damage economic obligations. avoiding to create the skin factor due to pressure drop, and preventing
Therefore, accurate control and management of production rates and water coning in gas condensate reservoirs [17,18]. Numerous experi­
pressure drop through production wells are essential to implement mental and theoretical relationships have been introduced to estimate
sustainable production programs from condensate reservoirs. Under­ the multiphase flow rate through wellhead chokes. In most of them, the
standing the importance of preserving and sustainable production from basis of flow calculations depends on the pressure difference between
gas condensate resources makes the position and credibility of efficient the upstream and downstream instruments [19–21]. One of the most
tools for controlling and handling this vital goal clearer. Wellhead popular computational models proposed belongs to Gilbert (1954),
chokes are a very cost-effective and efficient tool for measuring and which has been widely used to calculate the liquids flow rate through the
controlling multiphase flow rates at an optimum level [13]. Accurate wellhead choke and in recent years has been adapted for data from
measurement of multiphase flow is one of the concerns of production different regions (show in Eq. (1)) [22–26]:

2
A.R.B. Abad et al. Fuel 308 (2022) 121872

Table 2
Implementation of some machine learning algorithms to predict oil, gas, and gas condensate flow rates through wellhead chokes.
Fluid flow type Authors(year) Machine learning techniques Dataset Input parameters R2 Accuracy
metric

Oil flow rate Payaman & Salavati Artificial Neural Network (ANN) 196 data Pu - D64 - GOR 0.98 APD%=
(2012) [54] points − 0.33
Nejatian et al. (2014) Least-Squares Support Vector 171 data Reynolds number - d/D - 0.99 AAPD%=
[55] Machine (LSSVM) point Choke flow coefficient 0.256
Gholgheysari Gorjaei Particle swarm optimization (PSO)-Least 276 data Pu - D64 - GLR 0.965 APD%=
et al. (2015) [56] square support vector machine (LSSVM -PSO) points − 0.80
Rostami & Ebadi (2017) Gene expression programming (GEP) 119 data Pu - D64 - GOR - γg - API 0.96 AAPD%=
[57] points 14.808
Ghorbani et al. (2019) Genetic Algorithm and solver optimizers 127 data Pu - D64 - GLR - BS&W% 0.99 AAPD%=
[50] points 7.33
Ghorbani et al. (2020) Adaptive Neuro Fuzzy Inference System 182 data Pu - D64 - GLR - BS&W% 0.998 AAPD%=
[49] (ANFIS) points 6.62
Oil flow rate assisted with Khan et al. (2020) [51] ANN 1950 data Pu - D64 - Tup - Pd - Oil API 0.99 AAPD%=
gas lift points 2.56
Gas flow rate in gas & gas ZareNezhad & Aminian ANN 97 data ΔP - GOR - D64 0.99 APD%=
condensate reservoir (2011) [58] points 0.486
Elhaj et al. (2015) [59] ANN 162 data Pu - D64 - Pd - T- γg 0.99 AAPD%=
points 0.828
Fuzzy Logic (FL) 0.97 AAPD%=
0.681
Kalam et al. (2019) [59] ANN 17,097 data Pu - D64- T – Qg 0.953 AAPD%=
points 7.386
Functional Network (FN) 0.91 AAPD%= 12
ANFIS 0.95 AAPD%= 14

Fig. 1. Schematic of workflow proposed for the constrction and evaluation of four HML algorithms used for Qg prediction.

Pwh DO64 (replacing the upstream pressure with the pressure drop across the
Qliq = M (1)
GLRl choke) for the wells’ data in gas condensate reservoirs. Guo et al. 2002
evaluated data from 239 condensate gas wells with Sachdeva’s multi­
where Qliq is the rate of liquids production (STB/D), Pwh is the wellhead phase choke flow equation and compared the results with field mea­
pressure (psi), D64 is the choke size (1/64 in.), GLR is the gas to liquid surements. After receiving the under-estimated performance feedback
ratio (SCF/STB), and M, l, O are experimental coefficients calculated from this model, they could adapt it using different choke discharge
where sufficient data is available for specific reservoir systems. coefficients (CD) to obtain less computational error [28]. Al-Attar 2008
Osman and Dokla 1990 used a dataset from gas condensate wells in developed an empirical equation to describe a sub-critical flow model in
the Middle East region to develop an empirical relationship for calcu­ gas condensate wellhead chokes ranging from 24/64 to 128/64 in. for
lating the flow through wellhead chokes [27]. They adapted the Gilbert different choke sizes [29]. Nasriani and Kalantariasl (2019) also pre­
equation in three modified forms by changing the pressure parameters sented a tuned equation derived from the Gilbert basic equation to

3
A.R.B. Abad et al. Fuel 308 (2022) 121872

2. Methodology

2.1. Work flow

A systematic methodology involving ten steps (Fig. 1) is developed

for constructing and evaluating the four hybrid machine learning algo­
rithms employed for the prediction of gas flow rate through wellhead
chokes. The first step in the proposed workflow is data gathering from
gas condensate fields. Next, the maximum and minimum values of
variables need to be determined. Afterward, the variables are normal­
ized between − 1 and + 1 (Eq. (2)). Once the data are normalized, the set
of data is divided into two subsets, training and testing. Then, the ma­
chine learning optimizer’s accuracy is determined by statistical in­
Fig. 2. Schematic of GA cycle. dicators such as AAPD%, SD, MSE, RMSE, and R2. Results obtained from
accuracy evaluation are compared with empirical equations and hybrid
measure flow rate in sub-critical flow regime based on data collected machine learning techniques [47].
from 50 wells in some gas condensate reservoirs in southern Iran [30]. ( l
xi − xminl
)
Seidi and Sayahi (2015), by adapting Gilbert’s basic equation using the xli = l l
*2 − 1 (2)
xmax − xmin
genetic algorithm and nonlinear regression methods and applying them
to 67 datasets gathered from different gas condensate fields, proposed an
where xli is the value of attribute l for data record I; xminl is the minimum
optimized model for estimating the condensate gas flow rate [31]. The
value of the attribute l among all the data records in the dataset; and
equations presented by these researchers are summarized in Table 1.
xmaxl is the maximum value of the attribute l among all the data records
Recently, some researchers have strived to solve many problems in
in the dataset.
oil and gas industry [32–36]. However, data science has provided a new
way to move from conventional computing systems to faster, more ac­
curate, and cost-effective computing methods. Today, new machine 2.2. Least square support vector machine (LSSVM)
learning techniques are efficient tools for optimization and sophisticated
computing that reduce operating costs and improve system perfor­ The least-square support vector machine (LSSVM) is an expanded
mance. Extensive research has been conducted in recent years on the version of the support vector machine (SVM) that Suykens and Vande­
application of intelligent machine learning methods in various sectors of walle developed in 1998 [60,61]. LSSVM technique uses powerful fea­
the upstream oil and gas industry, such as desalting system analysis tures of SVM [62,63]. However, there are two major differences between
[37], hydrocarbon phase behavior prediction [38–42], determination of the LSSVM and SVM learning techniques. First, the LSSVM technique
oil and gas flow through orifice [43–46], and determination of flow rate uses square errors in the cost function instead of nonnegative errors, and
through wellhead choke [18,47–53]. Predicting multiphase flow rate second, the LSSVM technique applies equality constraints instead of
from wellhead chokes is the subject of other studies on machine learning inequality constraints. Consequently, in LSSVM, a linear system of
application in flow measurement concepts. Table 2 summarizes some of equations is solved instead of a quadratic programming problem, lead­
the recently published research on these smart models’ performance in ing to a considerable reduction in the model’s learning computational
this field. time [64,65].
As shown in Table 2, in recent years, intelligent machine learning In the LSSVM method, the following nonlinear cost function (Eq. (3))
models for accurately estimating the flow rate of hydrocarbon fluids is used for approximation [66,67]:
passing through wellhead chokes have been found to be inexpensive, f (x) = wT ϕ(xi ) + b (3)
fast, and accurate solutions for calculating the production flow of hy­
drocarbon fluids. Machine learning models require a large and extensive In which xi denotes the input variable to the function, the dimension
range of data set to create a comprehensive and more accurate model. of which is N × n,where N and n stand for the number of samples in the
There is still a shortage of model construction by vast data sets specif­ dataset and the number of inputs parameters, respectively. w and b
ically structured to predict gas flow rates. represent the weight and bias vector of output layer respectively, ϕ(xi )
Table 1 provides a comparison of previous empirical relationships, indicates kernel function, and T is transpose matrix. For the sake of
and Table 2 shows the results of intelligent methods proposed in pre­ brevity, the readers are advised to refer to the previously published
vious studies. It is worth noting that the methods proposed in this paper works, where a detailed theoretical description of the LSSVM model is
are compared with those empirical methods in previous studies that provided [61,62,68–73]. Since the LSSVM model parameters have a
presented better performance. In addition, as shown in Table 2, a limited considerable influence on the model accuracy and performance, GA and
number of studies have been performed on the gas flow rate prediction PSO optimization algorithms were applied for optimizing those pa­
in gas & gas condensate reservoirs. As a result, this research, based on a rameters in the present study. Besides these control parameters, the type
dataset of more than 1009 data records, has endeavored to develop of kernel applied in LSSVM model construction also has a pronouncing
novel models for gas flow rate prediction (MELM with PSO/GA opti­ effect on the performance and accuracy of the LSSVM model. Given that
mizer) with minimized RMSE. The model developed employs six input there is no standard way in kernel function selection, four of the most
variables, including temperature (T), upstream pressure (Pu), down­ commonly applied kernel functions, including the linear kernel, poly­
stream pressure (Pd), gas gravity (γg), choke diameter (D64), and nomial kernel, radial basis function (RBF) kernel, and multilayer per­
gas–liquid ratio (GLR) to accurately predict gas flow rate from wellhead ceptron kernel, have been tested out in the present study. Among those,
chock. Moreover, to create the best possible prediction performance and the RBF kernel is found to be the most efficient one.
accuracy as well as to avoid overfitting, several control measures are
applied in the present study. 2.3. Multilayer extreme learning machine (MELM)

The extreme learning machine (ELM), as a new quick single hidden

layer feedforward network, was first developed by Huang et al. in 2005
[74]. Since its emergence, ELM has been widely used in generating

4
A.R.B. Abad et al. Fuel 308 (2022) 121872

Fig. 3. Typical flow diagram for the LSSVR-PSO/GA hybrid models developed for Qg prediction.

solutions to various problems, namely regression, classification, and Complex variants of ELM with several hidden layers are recom­
clustering. The basic structure of ELM resembles a single hidden layer mended to solve problems with a nonlinear dataset of high complexity.
backpropagation (BP) neural network that is composed of three layers Therefore, a complex form of ELM that includes multiple hidden layers,
which are input, hidden, and output layers. However, the method used called MELM, was developed based on the deep learning concept [80].
in training ELM is soundly different from that of the conventional The construction procedure of the MELM learning model is elaborated in
network. Indeed, the ELM technique randomly assigns the hidden pa­ recently published works [38,63].
rameters, the hidden nodes biases, and the input weights to hidden
nodes and analytically calculates the output weights. As a result, the
time required for optimizing the hidden parameters of the model is 2.4. Optimization algorithm techniques
significantly decreased by avoiding iterative calculations during model
training [75,76]. Elaboration on structures and the theoretical principles 2.4.1. Genetic algorithm (GA)
of conventional artificial neural networks and ELM models can be Genetic algorithm is a class of evolutionary algorithms developed
discovered in previous publications [74,77–79]. based on natural selection and evaluation principles. This method is
commonly applied for solving search and optimization problems. This

5
A.R.B. Abad et al. Fuel 308 (2022) 121872

Table 3 2.4.2. Particle swarm optimization (PSO)

Optimal values of control parameters for the LSSVM-PSO/GA models established Particle Swarm Optimization (PSO), an optimization algorithm
for Qg prediction. inspired by natural swarming and flocking of birds and insects, was
LSSVM PSO GA proposed by Kennedy and Eberhart [87]. This optimization method
Control Value Control Value Control Value
initiates a population or “swarm” made of random solutions and, by
parameter parameter parameter updating generation, attempts to obtain the optimal solution. In the PSO
algorithm, solutions are named “particles” [38]. The population parti­
Variance of RBF 9.8507 Swarm size 80 Population 80
kernelσ2
cles go through the space of the problem by following the current best
Regularization 53.1392 Maximum 200 Maximum 200
particles in the population. Each population particle possesses a velocity
parameter iterations iterations and a position, and they seek positions with good fitness in the space.
Objective Social 2.05 Selection Roulette During the optimization process, two main pieces of information are
function constant method wheel memorized by each particle i) the best position heaving been so far
Cognitive 2.05 Crossover Uniform
visited by the particle (Pb) ii) the global best position attained by the
constant (p = 1)
Inertia 0.98 Mutation Uniform particles in the whole swarm (Gb) [29,38]. To obtain the best solution,
weight (p = 1) several iterations are performed by PSO. In each step, the solution
Mutation 0.08
rate
Selection 2 Table 4
pressure RMSE obtained for different MELM structures for pre-processing the MELM-
(Roulette PSO/GA models applied for Qg prediction.
wheel)
Number of hidden layers Number of neurons in the layers

3 5 7 9

5 6.3296 5.7488 6.0634 6.0985

method obtains the global optimum solution within a complex multi- 10 5.8175 5.2953 5.3296 5.3298
dimensional space. In the GA method, the poorer population of par­ 15 5.9542 5.0098 5.0108 5.0152
ents is replaced with the better offspring population by each generation 20 5.9533 5.0279 5.0295 5.1841
of the population using three operations: selection, crossover, and mu­
tation. This process is repeated by the GA until a high accuracy of pre­
diction is achieved. Hence, the population’s final output individual is the Table 5
best parameter group [81,82]. Fig. 2 illustrates the cycle of GA. To keep The best structure for pre-processing the MELM-PSO/GA models applied for Qg
the study concise, the readers are advised to read previously published prediction.
studies in which detailed theoretical descriptions on the GA technique Layer Layer Layer Layer Layer Layer Layer RMSE
are provided [83–86]. 1 2 3 4 5 6

Neurons 10 9 14 12 12 8 4.9637

Fig. 4. Typical Flow diagram of the MELM-PSO/GA developed for Qg prediction.

6
A.R.B. Abad et al. Fuel 308 (2022) 121872

Table 6 MELM learning algorithms are coupled with two optimization algo­
Optimal values of control parameters for the MELM-PSO/GA hybrid models rithms (GA and PSO) to develop these predictive models.
established for Qg prediction.
MELM PSO GA 2.5.1. LSSVM-PSO/GA hybrid models
Control Value Control Value Control Value
In this study, two hybrid models LSSVM-PSO and LSSVM-GA, are
parameter parameter parameter developed for predicting gas flow rate through the chocks. Fig. 3 dis­
plays the flow diagram for the LSSVM- PSO/GA models developed.
Number of 6 Swarm size 80 Population 80
input The optimal values of the LSSVM model hyperparameters were ob­
variables tained using the PSO and GA optimization algorithms. RBF kernel
Number of 20 Maximum 200 Maximum 200 function was employed in the LSSVM predictive model construction
hidden iterations iterations since it provides the best performance among all the kernel functions
layers
Number of 5 Social 2.05 Selection Roulette
tested (Table 1). The LSSVM hyperparameters for the hybrid models
neurons in constant method wheel developed, LSSVM-GA and LSSVM-PSO, and the control parameters for
each layer the GA and PSO optimization algorithms applied are listed in Table 3.
RMSE Cognitive 2.05 Crossover Uniform
constant (p = 1)
2.5.2. MELM-PSO/GA hybrid models
Inertia 0.98 Mutation Uniform
weight (p = 1) Coupling MELM algorithm with GA and PSO optimization, two other
Var Mutation rate 0.08 hybrid models, MELM-PSO and MELM-GA, were constructed for accu­
minimum rately and reliably predicting gas flow rate through wellhead chokes.
Minimum Selection 2 The genetic algorithm is inherently discrete, while the PSO algorithm is
velocity pressure
(Roulette
a continuous method. Both of these algorithms generate new responses
wheel) in the neighborhood of the two parents (in the genetic algorithm with
Minimum the crossover operator and the PSO by adsorption to the best position in
velocity the Pbest particle community). Generating answers in the neighborhood
of two parents can be one of the most obvious differences with point-
based methods such as simulated annealing and taboo search. Execu­
achieved is compared with both the global best and the self-local best of
tion time in GA is longer than in PSO, and it converges more slowly. The
the population. The new position of particles can be obtained by Eqs. (4)
PSO, on the other hand, converges faster due to fewer operators and
and (5).
fewer parameters. More details on the GA and PSO algorithms can be
Vi (t + 1) = wVi (t) + c1 r1 (Pbi (t) − xi (t) ) + c2 r2 (Gb (t) − xi (t)) (4) found in previous publications [90–93]. A flow diagram for the MELM-
PSO/GA hybrid models developed is illustrated in Fig. 4. As can be seen
xi (t + 1) = xi (t) + Vi (t + 1), i = 1, 2, ⋯, N (5) from Fig. 4, the developed hybrid models include a two-step procedure
of optimization, which is briefly described below:
where N indicates the number of swarm particles, xi and Vi represent the Step1: Determining the optimal number of hidden layers using the
position and velocity of the particles respectively, w stands for inertia optimizers applied by a tuning optimization procedure. The ranges of
weight, controlling the influence of the previous velocity on the new the numbers of hidden layers and the nodes in those layers are nar­
one, c1 and c2 denote the cognitive and social acceleration coefficient, rowing optimally down. The narrow ranges will then be employed as
respectively, and r1 and r2 are two random numbers ranging from 0 to 1. constraints in constructing hybrid models.
It should be noted that,w, c1 , and c2 can be obtained through performing Step 2: Calculating the MELM model’s control parameters (weights
a trial and error analysis on the dataset under evaluation [88,89]. and biases) for the constrained ranges of the hidden layers and the nodes
in those layers obtained at step 1.
2.5. Hybrid machine-learning models developed for Qg prediction Based on the first step optimization carried out for the MELM con­
struction, the number of hidden layers for MELM is constrained to a
In this study, four hybrid machine-learning models equipped with range from 5 to 20. The number of nodes in those hidden layers is
effective optimizers are proposed, which provide accurate and reliable constrained to a range from 3 to 9. Table 4 lists the results for the first
predictions of gas flow rate through wellhead chokes. The LSSVM and

Fig. 5. Marun-Khami, Aghajari-Khami, and Ahvaz-Khami gas condensate fields located onshore Iran in the Zagros Basin.

7
A.R.B. Abad et al. Fuel 308 (2022) 121872

Table 7
Statistical characterization of data variables in Iranian gas condensate fields for Qg prediction.
Field Variables Temperature Upstream Downstream Gas specific Choke Gas liquid Gas flow
pressure pressure gravity diameter ratio rate
Symbol T Pu Pd γg D64 GLR Qg
Units ◦
F psig psig - Inch Scf/STB Mscf/d

297 data records from Marun-Khami gas Mean 125.81 1791.70 795.82 0.68 41.94 8.60E+04 20.08
condensate field Std. 19.67 755.80 350.69 0.04 22.22 7.64E+04 5.99
deviation
Variance 385.67 5.69E+05 1.23E+05 0.00 492.05 5.82E+09 35.76
Minimum 74.00 217.00 100.00 0.61 16.00 7.46E+03 5.40
Maximum 187.00 6115.00 2615.00 0.82 160.00 3.22E+05 29.55
Skewness 0.0894 − 0.1461 − 0.0105 1.4158 3.0942 1.29E+00 − 0.2408
Kurtosis − 0.1333 3.03393 1.9411 3.1418 10.9981 1.20E+00 − 0.9674
Median 125.00 2043.00 891.00 067 40.00 5.67E+04 20.37
Mode 114.00 2350.00 969.00 068 40.00 1.61E+05 28.25

399 data records from Aghajari-Khami gas Mean 132.18 2045.11 912.01 0.68 128.29 5.98E+04 73.55
condensate field Std. 22.74 749.42 338.94 0.03 48.86 4.46E+04 17.16
deviation
Variance 515.71 5.60E+05 1.15E+05 0.00 2381.35 1.99E+09 293.70
Minimum 77.00 1036.00 223.86 0.61 42.00 6.36E+03 54.13
Maximum 189.00 4658.00 1.7252 0.82 194.00 2.69E+05 122.46
Skewness 0.0935 1.8326 2366.82 1.3309 − 0.1945 1.68E+00 0.8917
Kurtosis − 0.1346 3.3988 5.0902 3.8104 − 1.2896 3.99E+00 − 0.2748
Median 132.00 1880.00 887.00 0.67 130.00 5.36E+04 67.25
Mode 114.00 1653.00 1003.00 0.69 66.00 8.14E+04 67.41

371 data records from Ahvaz-Khami gas Mean 129.15 2083.33 896.10 0.67 82.99 7.46E+04 42.58
condensate field Std. 18.91 607.65 322.13 0.04 43.71 6.05E+04 7.16
deviation
Variance 356.56 3.68E+05 1.03E+05 0.00 1905.69 3.66E+09 51.06
Minimum 85.00 952.00 125.00 0.61 26.00 7.91E+03 29.57
Maximum 187.00 5910.00 2265.30 0.82 194.00 3.22E+05 53.99
Skewness 0.1189 2.2510 0.7916 1.4640 1.0655 1.75E+05 − 0.1667
Kurtosis − 0.0658 10.1780 3.6853 3.06545 − 0.0534 3.00E+00 − 1.1974
Median 129.00 2088.74 951.00 0.67 64.00 6.18E+04 40.00
Mode 146.00 1820.00 1070.00 0.68 56.00 8.19E+04 49.44

1099 data records from the gas condensate Mean 129.35 1987.86 864.12 0.68 88.50 7.23E+04 74.91
fields (Marun-Khami, Aghajari-Khami, and Std. 20.76 715.25 342.50 0.04 53.84 6.12E+04 24.68
Ahvaz-Khami) deviation
Variance 430.67 5.11E+05 1.17E+05 0.000.00 2896.15 3.75E+09 608.75
Minimum 74.00 217.00 100.00 0.61 16.00 6.36E+09 5.40
Maximum 189.00 6115.00 2615.00 0.82 194.00 3.22E+05 122.46
Skewness 0.1442 1.1531 0.8138 1.3982 0.6632 1.73E+00 0.6162
Kurtosis − 0.0303 5.0324 3.8315 3.2852 − 0.9592 3.21E+00 − 0.0940
Median 129.00 1990.00 908.00 0.67 64.00 5.57E+04 46.84
Mode 114.00 1986.00 611.50 0.67 64.00 8.14E+04 67.41

optimization step, and Table 5 shows the best structure for MELM-PSO/ variables have never been used simultaneously in previously published
GA models. The control parameters for the MELM-PSO/GA hybrid studies to predict the target parameter of this study. Therefore, the
models are presented in Table 6. models developed in the present study can be considered novel ap­
proaches in this field. Table 7 shows the statistical characteristics of the
3. Data collection and distribution data variables used to predict the Qg for each reference in this paper.
One of the descriptive diagrams to describe the input data is cumu­
In this paper, for predicting Qg from gas condensate reservoirs lative distribution functions (CDF) (see Fig. 6). In this figure (Fig. 6), the
through wellhead chokes, 1067 datasets were collected from three gas 1067 dataset distribution diagram is used, and the CFD formula is shown
condensate fields Marun-Khami, Aghhajari-Khami, and Ahvaz-Khami in Eq. (6) [47,49]:
that located in southwestern Iran (see Fig. 5). Khami group is a group
FX (x) = P(X ≤ x), forallx∊R (6)
of geological formations of Zagros, which includes Heath and Surmeh
formations from the Jurassic period and Fahlian, Gadvan, and Darian X is the data variable value range, X is the value of variable × in a
formations from the Cretaceous period. This group has crude oil reserves specific data record, and R is the dataset of data records.
in some oil fields plus gas and condensate gas in most fields. Khami CFD is used to describe the input variables in Fig. 6. The CFD for
reservoir rock is deeper than the Asmari and Bangestan reservoir rocks. temperature is T < 112 F0 for ~ 20.3% of the data records, 112 < T <
Ahvaz, Gachsaran, Maroon, Karanj, Bibi Hakimeh, and Aghajari fields 152 F0 for ~ 64.7% of the data records, and T > 152 F0 for the remaining
are among the fields that have crude hydrocarbon reserves (data used in 15% of the data. The CFD for initial gas specific gravity is γg < 0.6588
this study are confidential, and the authors have no permission to share for ~ 29.8% of the data records, 0.6588 < γg < 0.7188 for ~ 64.4% of
them in public). To predict Qg, six input variables were used in this the data records, and γg > 0.7188 for the remaining 5.8% of the data.
study, including temperature (T), the upstream pressure (Pu), down­ The CFD for gas to liquid ratio is GLR < 22243 Scf/STB for ~ 21.2% of
stream pressure (Pd), gas gravity (γg), choke diameter (D64), and the data records, 22243 < GLR < 140000 Scf/STB for ~ 60% of the data
gas–liquid ratio (GLR). To the authors’ best knowledge, these six input records, and GLR > 140000 Scf/STB for the remaining 18.8% of the

8
A.R.B. Abad et al. Fuel 308 (2022) 121872

Fig. 6. Cumulative distribution function (CDF) for the input variables and output values used for the Qg prediction (thinner blue line) compared to cumulative
distribution functions for normal distributions defined by variable means and standard deviations (thicker red line). (For interpretation of the references to colour in
this figure legend, the reader is referred to the web version of this article.)

9
A.R.B. Abad et al. Fuel 308 (2022) 121872

Fig. 7. Cross plot of input variables versus Qg, indicating the effect of boundaries on the performance of four HML models developed.

Table 8 Table 9
Gas flow rate prediction accuracy statistics for the training subset (712 available Gas flow rate prediction accuracy statistics for the testing subset (297 available
data records; ~70%) Marun-Khami, Aghajari-Khami, and Ahvaz-Khami gas data records; ~30%) Marun-Khami, Aghajari-Khami, and Ahvaz-Khami gas
condensate fields (Qg; Mscf/d). condensate fields (Qg; Mscf/d).
Models APD AAPD SD MSE RMSE R2 Models APD AAPD SD MSE RMSE R2
Units (%) (%) Mscf/ Mscf/d Mscf/d – Units (%) (%) Mscf/ Mscf/d Mscf/d –
d d

Empirical models Empirical models

Osman & − 93.977 97.986 54.156 5004.2328 70.7406 0.4017 Osman & − 79.112 85.145 47.094 6570.9979 59.7578 0.4392
Dolka Dolka
Al-Attar 74.364 83.219 59.382 3991.2958 63.1767 0.4271 Al-Attar 69.478 78.021 56.247 3484.8769 59.0328 0.4495
Seidi & 52.487 61.372 25.974 982.3102 31.3418 0.4952 Seidi & 55.115 62.059 21.376 924.76825 30.4100 0.4604
Sayahi Sayahi
Ghorbani 31.663 47.116 19.167 602.5881 24.5477 0.4954 Ghorbani 31.160 44.767 18.457 555.9255 23.5781 0.4895
et al. et al.
Nasriani 47.380 77.087 47.066 2228.3030 47.2049 0.4862 Nasriani 49.449 78.246 46.719 2182.6987 46.7194 0.4695
et al. et al.

Hybrid machine learning optimizer models Hybrid machine learning optimizer models
MELM-PSO − 2.237 5.471 2.592 6.7242 2.5931 0.9900 MELM-PSO − 3.150 7.220 3.426 11.7437 3.4269 0.9833
MELM-GA − 3.179 6.459 3.048 9.3848 3.0635 0.9862 MELM-GA − 6.576 12.638 5.840 34.2741 5.8544 0.9508
LSSVM- − .5.115 10.870 4.734 22.4867 4.7420 0.9655 LSSVM- − 8.134 16.594 6.939 48.2444 6.9458 0.9269
PSO PSO
LSSVM-GA − 5.194 10.961 5.025 25.3992 5.0398 0.9595 LSSVM-GA − 7.777 15.653 7.051 49.8530 7.0607 0.9241

10
A.R.B. Abad et al. Fuel 308 (2022) 121872

Table 10
Gas flow rate prediction accuracy statistics for the total dataset (1009 available data records; ~100%) Marun-Khami, Aghajari-Khami, and Ahvaz-Khami gas
condensate fields (Qg; Mscf/d).
Models APD AAPD SD MSE RMSE R2
Units (%) (%) Mscf/d Mscf/d Mscf/d –

Empirical models
Osman & Dolka − 89.602 94.1940 52.391 4582.3589 67.6931 0.4190
Al-Attar 65.448 75.645 53.425 2699.2699 51.9545 0.4239
Seidi & Sayahi 49.657 61.574 24.809 965.3727 31.0704 0.4810
Ghorbani et al. 31.515 46.424 18.964 588.8529 24.2663 0.4905
Nasriani et al. 48.831 77.428 47.004 2214.8793 47.0625 0.4744

Hybrid machine learning optimizer models

MELM-PSO − 2.506 5.986 2.863 8.2017 2.8639 0.9778
MELM-GA − 4.176 5.278 4.074 16.7110 4.0879 0.9693
LSSVM-PSO − 6.004 12.555 5.476 30.0685 5.4835 0.9534
LSSVM-GA − 5.655 12.342 5.679 32.5972 5.7094 0.9484

data. The CFD for gas flow rate is Qg < 18.3 Mscf/d for ~ 11.3% of the Standard Deviation (SD):
data records, 18.3 < Qg < 72.8 Mscf/d for ~ 76.3% of the data records, √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n
and Qg > 72.8 Mscf/d for the remaining 12.4% of the data. Based on the i=1 (Di − Dimean)
2
SD = (10)
CFD’s shown in Fig. 6, three variable parameters, including T, γg, and n− 1
GLR are normally distributed.
The CFD for upstream pressure is Pu < 2080 psig for ~ 56.4% of the 1∑ n
Dimean = (HMeasured i − HPredicted i )
data records, 2080 < Pu < 3220 psig for ~ 39% of the data records, and n i=1
Pu > 3220 psig remaining 4.6% of the data. The CFD for downstream
pressure is Pd < 498.1 psig for ~ 13.67% of the data records, 498.1 < Pd Mean Square Error (MSE):
< 966 psig for ~ 48.43% of the data records, 966 < Pd < 1421 psig for ~ 1∑ n

35% of the data records, and Pd > 1421 psig for the remaining 2.9% of MSE = (ZMeasured i − ZPredicted i )2 (11)
n i=1
the data. The CFD for choke size is D64 < 40 in. for ~ 20% of the data
records, 40 < D64 < 108 in. for ~ 46% of the data records, and D64 > Root Mean Square Error (RMSE):
108 in. for the remaining 34% of the data. Based on the CFDs shown in √̅̅̅̅̅̅̅̅̅̅
Fig. 6, three variable parameters, including Pu, Pd, and D64, are not RMSE = MSE (12)
normally distributed. Coefficient of Determination (R2):
∑N
4. Results & discussion
2
i=1 (HPredicted i − HMeasured i )
R2 = 1 − ∑n 2
(13)
∑N H
I=1 Measured i
i=1 (HPredicted i − )
Fig. 7 presents the relationship between the input variables (T, Pu, n

Pd, D64, γg, and GLR) and Qg for 1009 data records collected around These statistical indicators are among the most commonly used in­
Iran. Comparison of the input variables correlation with Qg indicates dicators to evaluate the prediction performance accuracy and compare
that D64 presents a strong correlation with Qg, which suggests this HML algorithms and empirical equations. Among these indicators,
parameter is more influential on Qg than other parameters. Besides, the RMSE is considered the most important one for evaluating HML models’
least influential parameter on the output variable (Qg) is found to be γg. prediction accuracy. Given these algorithms are configured to minimize
This evaluation of the inputs parameters’ correlation degree with Qg can the RMSE, this accuracy indicator is more important than other statis­
assist in the proper selection of features for the algorithms, leading to tical errors studied in this research.
enhanced prediction performance accuracy. Using statistical errors, the data are divided into two parts: test and
One way to compare HML and empirical equations’ efficiency in Qg train. Tables 8–10 show a comparison between the performance accu­
prediction is to use statistical errors. For this purpose, the equations racy of HML algorithms and empirical models for training (712 data
determining the magnitude of error, including percentage deviation records; 70%), testing (297 data records; 30%), and overall dataset
(PD) or relative error (RE), average percentage deviation (APD), abso­ (1009 data records; 100%) of Iran condensate field data, respectively.
lute average percentage deviation (AAPD), standard deviation (SD), Having a close look at the results presented in Tables 8–10 reveals
mean square error (MSE), root mean square error (RMSE; the objective that the prediction accuracy of the MELM-PSO algorithm, which is a
function of the HML models), and coefficient of determination (R2) are novel algorithm, is higher than other HML algorithms and empirical
selected for prediction accuracy evaluation, which are given in Eqs. (7)– equations. For instance, the MELM-PSO model has: RMSE = 2.5931
(13): Mscf/d; AAPD = 5.471٪; R2 = 0.9900 (for the training subset); RMSE =
Percentage deviation (PD) or relative error (RE): 3.4269 Mscf/d; AAPD = 7.220٪; R2 = 0.9833 (for the testing subset);
H(Measured) − H(Predicted) and RMSE = 2.8639 Mscf/d; AAPD = 5.986٪; R2 = 0.9778 (for the
PDi = x100 (7) overall dataset). Besides, HML models are found to be much more effi­
H(Measured)
cient than empirical models in terms of prediction accuracy. Comparing
Average percentage deviation (APD): the HML models’ prediction performance suggests that comparable
∑n prediction accuracy is reached by all four models. Still, the prediction
PDi
APD = i=1 (8) accuracy reached by the MELM-PSO model is slightly higher than those
n
of the MELM-GA and the LSSVM-PSO/GA models.
Absolute average percentage deviation (AAPD): Fig. 8 shows the measured versus predicted gas flow rate (Qg) for
∑n each data record in the training, testing, and total subset evaluated for
|PDi |
AAPD = i=1 (9) the Iranian condensate fields. Based on the performance accuracy shown
n

11
A.R.B. Abad et al. Fuel 308 (2022) 121872

Fig. 8. Measured versus predicted gas flow rate (Qg) for each data record in the training, testing, and overall dataset evaluated for HML models (MELM-PSO/GA and
LSSVM-PSO/GA) from the Iranian condensate fields (Marun-Khami, Aghajari-Khami, and Ahvaz-Khami).

in Fig. 8, it is clear that the performance accuracy of HML algorithms is LSSVM-PSO > LSSVM-GA.
close to each other. In other words, the results of the LSSVM algorithm Comparison of the results displayed in Figs. 8 and 9 demonstrate that
hybridized with GA / PSO are very close to MELM hybridized with GA / the prediction accuracy of the four HML models developed is much
PSO. As shown in Fig. 8, the coefficient of determination value for the higher than those of previous empirical equations. Based on the pre­
MELM-PSO algorithm is much better than other hybrid algorithms. diction accuracy (RMSE), they are as follows:
Comparison of the results presented in Tables 8–10 and Fig. 8 suggests MELM-PSO > MELM-GA > LSSVM-PSO > LSSVM-GA > Ghorbani
that the MELM-PSO can achieve higher performance accuracy compared et al. > Seeidi & Sayahi > Nasriani et al. > Al-Attar > Osman & Dokla.
to other models developed in this study. Based on the accuracy, the Figs. 10 and 11 display the histograms of gas flow rate prediction
models implemented can be sorted as MELM-PSO > MELM-GA > error with normal distributions (red line) for the HML algorithms and

12
A.R.B. Abad et al. Fuel 308 (2022) 121872

Fig. 9. Measured versus predicted gas flow rate (Qg) for each data record in the training, testing, and overall dataset evaluated for empirical equations (Osman &
Dokla, Al-Attar, Seeidi & Sayahi, Ghorbani et al., and Nasriani et al.) from the Iranian condensate fields (Marun-Khami, Aghajari-Khami, and Ahvaz-Khami).

the empirical equations based on 1009 data records from the Iranian accuracy. To identify and remove erroneous data in the dataset under
condensate fields. As shown in Fig. 10, the error rate for the HML is close study, K-means clustering method in a multidimensional space is used.
to zero, and the lowest error for these models is obtained by MELM-PSO. For this purpose, two to five clusters are considered, which are then
However, the error for all empirical equations is shifted to the right divided into smaller clusters [96]. Remote data sets are used as part of
(Fig. 11). According to the results of this figure (Fig. 11), it is clear that the data processing phase to input data into a single-layer ANN network
the error distribution for the empirical models, Osman & Dokla and Al- with five neurons to predict Qg.
Attar, is asymmetric. All the empirical models involve some individual The results of the K-means clustering performed are shown in Fig. 12.
predictions involving quite large errors, particularly in the positive di­ As it can be seen, 3 clusters demonstrate the lowest RMSE for Qg pre­
rection (i.e., overestimates of Qg). The lowest Qg prediction error range diction. Based on this modeling, 58 data sets are identified as outlier
is associated with is MELM-PSO model. data sets. The K-means clustering algorithm can retrieve remote data to
One of the most important and influential factors on the performance predict Qg. Fig. 13 displays that the k-means clustering presents a
accuracy of a prediction model is the use of high-quality data [94,95]. promising efficiency in outlier detection for the prediction of Qg.
However, due to the lack of calibration of measuring devices, field data Fig. 14 demonstrates how the HML models developed progress to­
always presents a degree of errors [38]. In other words, there can be data wards optimal and accurate prediction of Qg through two hundred it­
recodes among datasets that are far from the truth. These poor-quality erations. Comparing the results displayed in Fig. 14 indicates that all
data cause problems in the machine learning process and the training four HML algorithms present relatively similar convergence velocity in
model built on artificial intelligence. iteration #3. As seen in iteration #86, the prediction accuracy of
When dealing with such data, identifying and deleting unreliable LSSVM-PSO is better than that of LSSVM-GA. As for MELM-PSO/GA
data with distinct outlying values is the best way to increase the model’s models, PSO presents a quicker convergence to achieve its best

13
A.R.B. Abad et al. Fuel 308 (2022) 121872

Fig. 10. Gas flow rate prediction error (Qg) histograms displayed with normal distributions (red line) for HML models based on 1009 data records from the Iranian
condensate fields (Marun-Khami, Aghajari-Khami, and Ahvaz-Khami). (For interpretation of the references to colour in this figure legend, the reader is referred to the
web version of this article.)

solution than the GA optimizer. From iteration #120, the MELM-PSO diameter (D64) > downstream pressure (Pd) > gas–liquid ratio (GLR) >
performs better than the MELM-GA in terms of prediction accuracy. upstream pressure (Pu) > temperature (T) > gas gravity (γg).
All in all, the MELM-PSO/GA models are found to present higher fore­
1
cast accuracy than those of the LSSVM-PSO/GA. In addition, the PSO Qg ∝(D64 , Pd , Pu , T)andQg ∝( ) (15)
GLR, γg
optimizer is found to be more efficient in reaching the optimal solution
for both networks, the MELM and the LSSVM, when compared to the GA
optimizer. 5. Conclusion
To determine the degree of influence of each input variable on Qg,
Spearman’s non-parametric correlation coefficient (ρ) is used [97]. The In this research, 1009 input data from Iranian condensate fields
range of this parameter is between − 1 (complete negative correlation) (Marun-Khami, Aghajari-Khami, and Ahvaz-Khami) are used to
to 1 (complete positive correlation), which indicates a relatively low or construct four models to predict gas flow rate (Qg) through six input
high impact [98]. The Spearman parameter equation (Eq. (14)) is variables. The input variables to the developed models are temperature
defined as follows: (T), the upstream pressure (Pu), downstream pressure (Pd), gas gravity
∑n (γg), choke diameter (D64), and gas–liquid ratio (GLR). This is the first-
i=1 (Ei − E)(Fi − F) ever research work constructing a model based on these variables for Qg
ρ = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (14)
∑n 2 ∑n 2 prediction.
i=1 (Ei − E) i=1 (Fi − F)
Hybrid machine learning algorithms have several advantages over
simple machine learning algorithms. For instance, when the predictive
where Ei is E input variable value of data record I, E is mean value for
machine learning algorithms are combined with the PSO algorithm to
variable E, Fi is F dependent variable (Qg) value of data record I, F is
determine control parameters of the algorithms, the computational
mean value for dependent variable F, and n is the number of input
speed and accuracy enhance remarkably. In the case of the MELM
parameters.
model, they are optimized in two steps. The first step is to determine the
Fig. 15 shows the calculated ρ value for the total of 1009 processed
number of hidden layers and neurons in the network. The next is to
learning datasets. Based on the correlation coefficients determined, it is
identify the desired weight and biases applied to those layers and neu­
observed that D64, Pd, Pu, and T parameters positively influence Qg,
rons. In the case of LSSVM, the optimization setting is done in one step
whereas GLR and γg parameters present a negative influence on it. The
for the development of LSSVM with PSO/GA optimizer, which ulti­
greatest positive influence on Qg is observed for D64, while the greatest
mately leads to LSSVM-PSO and LSSVM-GA hybrid machine learning
negative influence is presented by GLR (see Eq. (15)). In general, the
optimizer algorithms.
order of input variables’ influence degree on Qg is as follows: choke

14
A.R.B. Abad et al. Fuel 308 (2022) 121872

Fig. 11. Gas flow rate prediction error (Qg) histograms displayed with normal distributions (red line) for empirical equations based on 1009 data records from the
Iranian condensate fields (Marun-Khami, Aghajari-Khami, and Ahvaz-Khami). (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)

Coupling the PSO to the GA algorithm is an effective approach in by the MELM-PSO is significantly higher than the prediction accuracy of
achieving high prediction accuracy in the HML algorithms. The multi- empirical equations. The best prediction performance accuracy obtained
hidden layer extreme learning machine (MELM) algorithm coupled from the empirical equations considered is observed for Ghorbani et al.’s
with the PSO optimizer presents the best performance among the four equation (RMSE = 24.2663 Mscf/d and R2 = 0.4905). Comparing the
HML models evaluated. This algorithm uses two hybrid stages with PSO developed MELM-PSO model with the previous empirical (Table 1) and
to improve its performance. This algorithm (MELM) first reduces the AI models (Table 2) suggest that the MELM-PSO model has superior
number of layers and nodes in each hidden layer by combining with prediction performance and higher accuracy.
PSO. In combination with the second PSO, determines the appropriate Sensitivity analysis obtained from the Spearman coefficient model
weight and bias for the nodes of the selected hidden layers. demonstrates that the input variables, including D64, Pd, Pu, and T, have
The most promising prediction performance accuracy is obtained by positive correlations with Qg. In contrast, GLR and γg parameters present
the MELM-PSO model applied to the overall dataset evaluated (RMSE = negative correlations with Qg. D64 displays the greatest positive corre­
2.8639 Mscf/d and R2 = 0.9778). The prediction performance achieved lation with Qg, whereas the poorest negative correlation with Qg is

15
A.R.B. Abad et al. Fuel 308 (2022) 121872

Fig. 12. Schematic of identifying and deleting the outlier data using the K-means clustering algorithm [38].

Fig. 14. RMSE values for the training subset based on HML algorithms (MELM-
PSO, MELM-GA, LSSVM-PSO, and LSSVM-GA) developed for the prediction of
Qg during supervised learning from the Iranian condensate fields (Marun-
Khami, Aghajari-Khami, and Ahvaz-Khami).

observed for GLR.

CRediT authorship contribution statement

Abouzar Rajabi Behesht Abad: Investigation, Visualization,

Writing - original draft, Data curation. Hamzeh Ghorbani: Supervision,
Writing - original draft, Validation, Methodology, Formal analysis,
Software. Nima Mohamadian: Conceptualization, Formal analysis,
Fig. 13. Results of outlier detection by the K-means clustering algorithm a) Writing - original draft. Shadfar Davoodi: Formal analysis, Conceptu­
Status of remote data detected for Qg prediction and b) Number of outlying data alization, Methodology, Writing - original draft, Writing - review &
detected per number of different clusters and ANN modeling error after the editing. Mohammad Mehrad: Software. Saeed Khezerloo-ye Agh­
remote data removal to predict Qg. dam: Writing - review & editing. Hamid Reza Nasriani: Supervision,
Validation.

Declaration of competing interest

The authors declare that they have no known competing financial

16
A.R.B. Abad et al. Fuel 308 (2022) 121872

[16] Ki S, Jang I, Cha B, Seo J, Kwon O. Restoration of missing pressures in a gas well
using recurrent neural networks with long short-term memory cells. Energies 2020;
13(18):4696. https://doi.org/10.3390/en13184696.
[17] Lak A, Azin R, Osfouri S, Gerami S, Chahshoori R. Choke modeling and flow
splitting in a gas-condensate offshore platform. J Nat Gas Sci Eng 2014;21:
1163–70. https://doi.org/10.1016/j.jngse.2014.07.020.
[18] Ghorbani H, Moghadasi J, Wood DA. Prediction of gas flow rates from gas
condensate reservoirs through wellhead chokes using a firefly optimization
algorithm. J Nat Gas Sci Eng 2017;45:256–71. https://doi.org/10.1016/j.
jngse.2017.04.034.
[19] Chong D, Yan J, Wu G, Liu J. Structural optimization and experimental
investigation of supersonic ejectors for boosting low pressure natural gas. Appl
Therm Eng 2009;29(14–15):2799–807. https://doi.org/10.1016/j.
applthermaleng.2009.01.014.
[20] Schüller R, Solbakken T, Selmer-Olsen S, facilities.. Evaluation of multiphase flow
rate models for chokes under subcritical oil/gas/water flow conditions. SPE
Product 2003;18(03):170–81. https://doi.org/10.2118/84961-PA.
[21] Coutinho RP, Waltrich PJ, Williams WC, Mehdizadeh P, Scott S, Xu J, et al.
Experimental characterization of two-phase flow through valves applied to liquid-
assisted gas-lift. J Energy Res Technol 2020;142(6). https://doi.org/10.1115/
1.4045921.
Fig. 15. Input variables assessed based on Spearman’s non-parametric corre­ [22] Gilbert W. Flowing and gas-lift well performance. Drilling and production practice.
lation coefficient values for Qg prediction calculated for 1009 data records of American Petroleum Institute; 1954. https://onepetro.org/APIDPP/proceedings-ab
stract/API54/All-API54/API-54-126/51072.
supervised learning dataset (from Iranian condensate fields (Marun-Khami,
[23] Safar Beiranvand M, Mohammadmoradi P, Aminshahidy B, Fazelabdolabadi B,
Aghajari-Khami, and Ahvaz-Khami)). Aghahoseini S. New multiphase choke correlations for a high flow rate Iranian oil
field. Mech Sci 2012;3(1):43–7. https://doi.org/10.5194/ms-3-43-2012.
[24] Aladwani F, Alatefi S. Toward the Development of a Universal Choke
interests or personal relationships that could have appeared to influence Correlation–Global Optimization and Rigorous Computational Techniques. Journal
the work reported in this paper. of Engineering Research 2020;8(3). 10.36909/jer.v8i3.7717.
[25] Bennis M, Gellert J, Nougués M, Crespo P. Decline Curve Analysis in Vaca Muerta
with Choke Size Normalization of Gas Rates. SPE/AAPG/SEG Latin America
Acknowledgment Unconventional Resources Technology Conference. Unconventional Resources
Technology Conference; 2020. 10.15530/urtec-2020-1403.
This research was supported by Tomsk Polytechnic University [26] Khamis M, Elhaj M, Abdulraheem A. Optimization of choke size for two-phase flow
using artificial intelligence. J Pet Explor Prod Technol 2020;10(2):487–500.
development program. https://link.springer.com/article/10.1007/s13202-019-0734-6.
[27] Osman ME, Dokla ME. Gas condensate flow through chokes. European Petroleum
References Conference. Society of Petroleum Engineers; 1990. 10.2118/20988-MS.
[28] Guo B, Al-Bemani AS, Ghalambor A. Applicability of Sachdeva’s choke flow model
in Southwest Louisiana gas condensate wells. SPE Gas Technology Symposium.
[1] Amirian E, Dejam M, Chen Z. Performance forecasting for polymer flooding in
Society of Petroleum Engineers; 2002. 10.2118/75507-MS.
heavy oil reservoirs. Fuel 2018;216:83–100. https://doi.org/10.1016/j.
[29] Al-Attar H. Performance of wellhead chokes during sub-critical flow of gas
fuel.2017.11.110.
condensates. J Petrol Sci Eng 2008;60(3–4):205–12. https://doi.org/10.1016/j.
[2] Litvinenko V. The role of hydrocarbons in the global energy agenda: The focus on
petrol.2007.08.001.
liquefied natural gas. Resources 2020;9(5):59. https://doi.org/10.3390/
[30] Nasriani HR, Khan K, Graham T, Ndlovu S, Nasriani M, Mai J, et al. An
resources9050059.
investigation into sub-critical choke flow performance in high rate gas condensate
[3] Zou C, Zhao Q, Zhang G, Xiong B. Energy revolution: from a fossil energy era to a
wells. Energies 2019;12(20):3992. https://doi.org/10.3390/en12203992.
new energy era. Nat Gas Ind B 2016;3(1):1–11. https://doi.org/10.1016/j.
[31] Seidi S, Sayahi T. A new correlation for prediction of sub-critical two-phase flow
ngib.2016.02.001.
pressure drop through large-sized wellhead chokes. J Nat Gas Sci Eng 2015;26:
[4] Song X, Liu Y, Xue L, Wang J, Zhang J, Wang J, et al. Time-series well performance
264–78. https://doi.org/10.1016/j.jngse.2015.06.025.
prediction based on long short-term memory (LSTM) neural network model.
[32] Hassanpouryouzband A, Joonaki E, Edlmann K, Haszeldine RS. Offshore geological
J Petrol Sci Eng 2020;186:106682. https://doi.org/10.1016/j.petrol.2019.106682.
storage of hydrogen: is this our best option to achieve net-zero? ACS Energy Lett
[5] Yadua AU, Lawal KA, Okoh OM, Ovuru MI, Eyitayo SI, Matemilola S, et al. Stability
2021:2181–6. https://doi.org/10.1021/acsenergylett.1c00845.
and stable production limit of an oil production well. J Pet Explor Prod Technol
[33] Hassanpouryouzband A, Yang J, Okwananke A, Burgass R, Tohidi B, Chuvilin E,
2020;10(8):3673–87. https://link.springer.com/article/10.1007/s13202-02
et al. An experimental investigation on the kinetics of integrated methane recovery
0-00985-3.
and CO 2 sequestration by injection of flue gas into permafrost methane hydrate
[6] Naik G. Tight gas reservoirs–an unconventional natural energy source for the
reservoirs. Sci Rep 2019;9(1):1–9. https://doi.org/10.1038/s41598-019-52745-x.
future. Accessado em 2003;1(07):2008. www.pinedaleonline.com/socioeconomic
[34] Hassanpouryouzband A, Yang J, Tohidi B, Chuvilin E, Istomin V, Bukhanov B.
/pdfs/tight_gas.pdf.
Geological CO2 capture and storage with flue gas hydrate formation in frozen and
[7] El-Banbi AH, McCain Jr W, Semmelbeck M. Investigation of well productivity in
unfrozen sediments: method development, real time-scale kinetic characteristics,
gas-condensate reservoirs. SPE/CERI Gas Technology Symposium. Society of
efficiency, and clathrate structural transition. ACS Sustain Chem Eng 2019;7(5):
Petroleum Engineers; 2000. 10.2118/59773-MS.
5338–45. https://pubs.acs.org/doi/abs/10.1021/acssuschemeng.8b06374.
[8] Hekmatzadeh M, Gerami S. A new fast approach for well production prediction in
[35] Hassanpouryouzband A, Yang J, Tohidi B, Chuvilin E, Istomin V, Bukhanov B, et al.
gas-condensate reservoirs. J Petrol Sci Eng 2018;160:47–59. https://doi.org/
CO2 capture by injection of flue gas or CO2–N2 mixtures into hydrate reservoirs:
10.1016/j.petrol.2017.10.032.
dependence of CO2 capture efficiency on gas hydrate reservoir conditions. Environ
[9] Alavi F, Mowla D, Esmaeilzadeh F. Production performance analysis of Sarkhoon
Sci Technol 2018;52(7):4324–30. https://doi.org/10.1021/acs.est.7b05784.
gas condensate reservoir. J Petrol Sci Eng 2010;75(1–2):44–53. https://doi.org/
[36] Hassanpouryouzband A, Yang J, Tohidi B, Chuvilin E, Istomin V, Bukhanov B, et al.
10.1016/j.petrol.2010.10.002.
Insights into CO2 capture by flue gas hydrate formation: gas composition evolution
[10] Mokhtari R, Varzandeh F, Rahimpour M. Well productivity in an Iranian gas-
in systems containing gas hydrates and gas mixtures at stable pressures. ACS
condensate reservoir: a case study. J Nat Gas Sci Eng 2013;14:66–76. https://doi.
Sustainable Chem Eng 2018;6(5):5732–6. https://pubs.acs.org/doi/abs/10.1021/
org/10.1016/j.jngse.2013.05.006.
acssuschemeng.8b00409.
[11] Janiga D, Czarnota R, Stopa J, Wojnarowski P, Kosowski P. Utilization of nature-
[37] Ranaee E, Ghorbani H, Keshavarzian S, Ghazaeipour Abarghoei P, Riva M, Inzoli F,
inspired algorithms for gas condensate reservoir optimization. Soft Comput 2019;
et al. Analysis of the performance of a crude-oil desalting system based on
23(14):5619–31. https://link.springer.com/article/10.1007/s00500-018-3218-6.
historical data. Fuel 2021;291:120046. https://doi.org/10.1016/j.
[12] Hassan AM, Mahmoud MA, Al-Majed AA, Al-Shehri D, Al-Nakhli AR, Bataweel MA.
fuel.2020.120046.
Gas production from gas condensate reservoirs using sustainable environmentally
[38] Rashidi S, Mehrad M, Ghorbani H, Wood DA, Mohamadian N, Moghadasi J, et al.
friendly chemicals. Sustainability 2019;11(10):2838. https://doi.org/10.3390/
Determination of bubble point pressure & oil formation volume factor of crude oils
su11102838.
applying multiple hidden layers extreme learning machine algorithms. Journal of
[13] Kaydani H, Najafzadeh M, Mohebbi A, Practice. Wellhead choke performance in oil
Petroleum Science and Engineering 2021:108425. 10.1016/j.petrol.2021.108425.
well pipeline systems based on genetic programming. Journal of Pipeline Systems
[39] Seyyedattar M, Ghiasi MM, Zendehboudi S, Butt S. Determination of bubble point
Engineering 2014;5(3):06014001. 10.1061/(ASCE)PS.1949-1204.0000165.
pressure and oil formation volume factor: Extra trees compared with LSSVM-CSA
[14] Hansen LS, Pedersen S, Durdevic P. Multi-phase flow metering in offshore oil and
hybrid and ANFIS models. Fuel 2020;269:116834. https://doi.org/10.1016/j.
gas transportation pipelines: Trends and perspectives. Sensors 2019;19(9):2184.
fuel.2019.116834.
https://doi.org/10.3390/s19092184.
[40] Daneshfar R, Keivanimehr F, Mohammadi-Khanaposhtani M, Baghban A. A neural
[15] Mokhtari K, Waltrich P. Performance evaluation of multiphase flow models applied
computing strategy to estimate dew-point pressure of gas condensate reservoirs.
to virtual flow metering. WIT Trans Eng Sci 2016;105:99–111.

17
A.R.B. Abad et al. Fuel 308 (2022) 121872

Pet Sci Technol 2020;38(10):706–12. https://doi.org/10.1080/ [65] Suykens JA, Van Gestel T, De Brabanter J. Least squares support vector machines.
10916466.2020.1780257. World scientific 2002.
[41] Wang K, Luo J, Wei Y, Wu K, Li J, Chen Z. Practical application of machine learning [66] Ahmadi MA, Ebadi M. Evolving smart approach for determination dew point
on fast phase equilibrium calculations in compositional reservoir simulations. pressure through condensate gas reservoirs. Fuel 2014;117:1074–84. https://doi.
J Comput Phys 2020;401:109013. https://doi.org/10.1016/j.jcp.2019.109013. org/10.1016/j.fuel.2013.10.010.
[42] Rashidi S, Khajehesfandeari M. Committee machine-ensemble as a general [67] Eslamimanesh A, Gharagheizi F, Mohammadi AH, Richon D. Phase equilibrium
paradigm for accurate prediction of bubble point pressure of crude oil. J Energy modeling of structure H clathrate hydrates of methane+ water “insoluble”
Res Technol 2021;143(2). https://doi.org/10.1115/1.4047977. hydrocarbon promoter using QSPR molecular approach. J Chem Eng Data 2011;56
[43] Farsi M, Barjouei HS, Wood DA, Ghorbani H, Mohamadian N, Davoodi S, et al. (10):3775–93. https://doi.org/10.1021/je200444f.
Prediction of oil flow rate through orifice flow meters: Optimized machine-learning [68] Ahmadi MA, Zahedzadeh M, Shadizadeh SR, Abbassi R. Connectionist model for
techniques. Measurement 2021;174:108943. https://doi.org/10.1016/j. predicting minimum gas miscibility pressure: application to gas injection process.
measurement.2020.108943. Fuel 2015;148:202–11. https://doi.org/10.1016/j.fuel.2015.01.044.
[44] Fadaei M, Ameli F, Hashemabadi SH. Investigation on different scenarios of two- [69] Ahmadi MA. Connectionist approach estimates gas–oil relative permeability in
phase flow measurement using Orifice and Coriolis flow meters: Experimental and petroleum reservoirs: application to reservoir simulation. Fuel 2015;140:429–39.
modeling approaches. Measurement 2021;175:108986. https://doi.org/10.1016/j. https://doi.org/10.1016/j.fuel.2014.09.058.
measurement.2021.108986. [70] Ahmadi MA, Mahmoudi B. Development of robust model to estimate gas–oil
[45] Dayev ZA. Application of artificial neural networks instead of the orifice plate interfacial tension using least square support vector machine: experimental and
discharge coefficient. Flow Meas Instrum 2020;71:101674. https://doi.org/ modeling study. J Supercrit Fluids 2016;107:122–8. https://doi.org/10.1016/j.
10.1016/j.flowmeasinst.2019.101674. supflu.2015.08.012.
[46] Jamei M, Ahmadianfar I, Chu X, Yaseen ZM. Estimation of triangular side orifice [71] Ahmadi MA. Toward reliable model for prediction drilling fluid density at wellbore
discharge coefficient under a free flow condition using data-driven models. Flow conditions: a LSSVM model. Neurocomputing 2016;211:143–9. https://doi.org/
Meas Instrum 2021;77:101878. https://doi.org/10.1016/j. 10.1016/j.neucom.2016.01.106.
flowmeasinst.2020.101878. [72] Ahmadi MA, Pournik M. A predictive model of chemical flooding for enhanced oil
[47] Shojaei Barjouei H, Ghorbani H, Mohamadian N, Wood DA, Davoodi S, recovery purposes: Application of least square support vector machine. Petroleum
Moghadasi J, et al. Prediction performance advantages of deep machine learning 2016;2(2):177–82. https://doi.org/10.1016/j.petlm.2015.10.002.
algorithms for two-phase flow rates through wellhead chokes. J Petrol Explor Prod [73] Ahmadi MA, Rozyn J, Lee M, Bahadori A. Estimation of the silica solubility in the
2021:1–29. https://link.springer.com/article/10.1007/s13202-021-01087-4. superheated steam using LSSVM modeling approach. Environ Prog Sustain Energy
[48] Choubineh A, Ghorbani H, Wood DA, Moosavi SR, Khalafi E, Sadatshojaei E. 2016;35(2):596–602. https://doi.org/10.1002/ep.12251.
Improved predictions of wellhead choke liquid critical-flow rates: modelling based [74] Huang G-B, Zhu Q-Y, Siew C-K. Extreme learning machine: theory and
on hybrid neural network training learning based optimization. Fuel 2017;207: applications. Neurocomputing 2006;70(1–3):489–501. https://doi.org/10.1016/j.
547–60. https://doi.org/10.1016/j.fuel.2017.06.131. neucom.2005.12.126.
[49] Ghorbani H, Wood DA, Mohamadian N, Rashidi S, Davoodi S, Soleimanian A, et al. [75] Huang G-B, Siew C-K. Extreme learning machine with randomly assigned RBF
Adaptive neuro-fuzzy algorithm applied to predict and control multi-phase flow kernels. Int J Inform Technol 2005;11(1):16–24. https://citeseerx.ist.psu.edu/view
rates through wellhead chokes. Flow Meas Instrum 2020;76:101849. https://doi. doc/download?doi=10.1.1.100.1555&rep=rep1&type=pdf.
org/10.1016/j.flowmeasinst.2020.101849. [76] Jiang X, Yan T, Zhu J, He B, Li W, Du H, et al. Densely connected deep extreme
[50] Ghorbani H, Wood DA, Moghadasi J, Choubineh A, Abdizadeh P, Mohamadian N. learning machine algorithm. Cognit Comput 2020;12(5):979–90. https://link.sprin
Predicting liquid flow-rate performance through wellhead chokes with genetic and ger.com/article/10.1007/s12559-020-09752-2.
solver optimizers: an oil field case study. J Pet Explor Prod Technol 2019;9(2): [77] Ahmadi MA, Ebadi M, Shokrollahi A, Majidi SMJ. Evolving artificial neural
1355–73. https://link.springer.com/article/10.1007/s13202-018-0532-6. network and imperialist competitive algorithm for prediction oil flow rate of the
[51] Khan MR, Tariq Z, Abdulraheem A. Application of artificial intelligence to estimate reservoir. Appl Soft Comput 2013;13(2):1085–98. https://doi.org/10.1016/j.
oil flow rate in gas-lift wells. Nat Resour Res 2020;29(6):4017–29. https://link.spr asoc.2012.10.009.
inger.com/article/10.1007/s11053-020-09675-7. [78] Ahmadi MA, Chen Z. Comparison of machine learning methods for estimating
[52] Alakeely AA, Horne RN. Application of deep learning methods to estimate permeability and porosity of oil reservoirs via petro-physical logs. Petroleum 2019;
multiphase flow rate in producing wells using surface measurements. J Petrol Sci 5(3):271–84. https://doi.org/10.1016/j.petlm.2018.06.002.
Eng 2021;108936. https://doi.org/10.1016/j.petrol.2021.108936. [79] Sajjadi S, Shamshirband S, Alizamir M, Yee L, Mansor Z, Manaf AA, et al. Extreme
[53] Gomaa I, Gowida A, Elkatatny S, Abdulraheem A. The prediction of wellhead learning machine for prediction of heat load in district heating systems. Energy
pressure for multiphase flow of vertical wells using artificial neural networks. Build 2016;122:222–7. https://doi.org/10.1016/j.enbuild.2016.04.021.
Arabian J Geosci 2021;14(9):1–10. https://link.springer.com/article/10.1007/s12 [80] Tang J, Deng C, Huang G-B, Hou J. A fast learning algorithm for multi-layer
517-021-07099-y. extreme learning machine. 2014 IEEE International Conference on Image
[54] Mirzaei-Paiaman A, Salavati S. The application of artificial neural networks for the Processing (ICIP). IEEE; 2014:175-8. 10.1109/ICIP.2014.7025034.
prediction of oil production flow rate. Energy Sour Part A 2012;34(19):1834–43. [81] Mokarizadeh H, Atashrouz S, Mirshekar H, Hemmati-Sarapardeh A, Pour AM.
https://doi.org/10.1080/15567036.2010.492386. Comparison of LSSVM model results with artificial neural network model for
[55] Nejatian I, Kanani M, Arabloo M, Bahadori A, Zendehboudi S. Prediction of natural determination of the solubility of SO2 in ionic liquids. J Mol Liq 2020;304:112771.
gas flow through chokes using support vector machine algorithm. J Nat Gas Sci Eng https://doi.org/10.1016/j.molliq.2020.112771.
2014;18:155–63. https://doi.org/10.1016/j.jngse.2014.02.008. [82] Mo L, Chen H, Chen W, Feng Q, Xu L. Study on evolution methods for the
[56] Gorjaei RG, Songolzadeh R, Torkaman M, Safari M, Zargar G. A novel PSO-LSSVM optimization of machine learning models based on FT-NIR spectroscopy. Infrared
model for predicting liquid rate of two phase flow through wellhead chokes. J Nat Phys Technol 2020;108:103366. https://doi.org/10.1016/j.infrared.2020.103366.
Gas Sci Eng 2015;24:228–37. https://doi.org/10.1016/j.jngse.2015.03.013. [83] Shavinina LV. The international handbook on innovation. Elsevier; 2003.
[57] Rostami A, Ebadi H. Toward gene expression programming for accurate [84] Goldberg DE, Holland JH. Genetic algorithms and machine learning. 1988. https://
prognostication of the critical oil flow rate through the choke: Correlation deepblue.lib.umich.edu/bitstream/handle/2027.42/46947/10994_2005_Article_
development. Asia-Pac J Chem Eng 2017;12(6):884–93. https://doi.org/10.1002/ 422926.pdf.
apj.2126. [85] Sivanandam S, Deepa S. Genetic algorithms. Introduction to genetic algorithms.
[58] ZareNezhad B, Aminian A. Accurate prediction of the dew points of acidic Springer; 2008, p. 15-37.
combustion gases by using an artificial neural network model. Energy Conver [86] Ahmadi M, Chen Z. Machine learning-based models for predicting permeability
Manage 2011;52(2):911–6. https://doi.org/10.1016/j.enconman.2010.08.018. impairment due to scale deposition. J Pet Explor Prod Technol 2020;10(7):
[59] Elhaj MA, Anifowose F, Abdulraheem A. Single gas flow prediction through chokes 2873–84. https://link.springer.com/article/10.1007/s13202-020-00941-1.
using artificial intelligence techniques. SPE Saudi Arabia Section Annual Technical [87] Eberhart R, Kennedy J. A new optimizer using particle swarm theory. MHS’95.
Symposium and Exhibition. Society of Petroleum Engineers; 2015. 10.2118/ Proceedings of the Sixth International Symposium on Micro Machine and Human
177991-MS. Science. Ieee; 1995:39-43. 10.1109/MHS.1995.494215.
[60] Yuan X, Chen C, Yuan Y, Huang Y, Tan Q. Short-term wind power prediction based [88] Kuo R, Hong S, Huang Y. Integration of particle swarm optimization-based fuzzy
on LSSVM–GSA model. Energy Convers Manage 2015;101:393–401. https://doi. neural network and artificial neural network for supplier selection. Appl Math
org/10.1016/j.enconman.2015.05.065. Model 2010;34(12):3976–90. https://doi.org/10.1016/j.apm.2010.03.033.
[61] Ahmadi MA, Ebadi M, Hosseini SM. Prediction breakthrough time of water coning [89] Kıran MS, Özceylan E, Gündüz M, Paksoy T. A novel hybrid approach based on
in the fractured reservoirs by implementing low parameter support vector machine particle swarm optimization and ant colony algorithm to forecast energy demand
approach. Fuel 2014;117:579–89. https://doi.org/10.1016/j.fuel.2013.09.071. of Turkey. Energy Convers Manage 2012;53(1):75–83. https://doi.org/10.1016/j.
[62] Ahmadi M-A, Bahadori A. A LSSVM approach for determining well placement and enconman.2011.08.004.
conning phenomena in horizontal wells. Fuel 2015;153:276–83. https://doi.org/ [90] Katiyar S. A comparative study of genetic algorithm and the particle swarm
10.1016/j.fuel.2015.02.094. optimization. Int J Technol 2010;2(2):21–4.
[63] Farsi M, Mohamadian N, Ghorbani H, Wood DA, Davoodi S, Moghadasi J, et al. [91] Hassan R, Cohanim B, De Weck O, Venter G. A comparison of particle swarm
Predicting formation pore-pressure from well-log data with hybrid machine- optimization and the genetic algorithm.1897. 10.2514/6.2005-1897.
learning optimization algorithms. Nat Resour Res 2021:1–27. https://link.springer. [92] Duan Y, Harley RG, Habetler TG. Comparison of particle swarm optimization and
com/article/10.1007/s11053-021-09852-2. genetic algorithm in the design of permanent magnet motors. IEEE:822-5.
[64] Suykens JA, Vandewalle J. Least squares support vector machine classifiers. Neural 10.1109/IPEMC.2009.5157497.
Process Lett 1999;9(3):293–300. https://link.springer.com/article/10.1023/A:1 [93] Panda S, Padhy NP. Comparison of particle swarm optimization and genetic
018628609742. algorithm for FACTS-based controller design. Appl Soft Comput 2008;8(4):
1418–27. https://doi.org/10.1016/j.asoc.2007.10.009.

18
A.R.B. Abad et al. Fuel 308 (2022) 121872

[94] Hodge V, Austin J. A survey of outlier detection methodologies. Artif Intell Rev [96] Mehrad M, Bajolvand M, Ramezanzadeh A, Neycharan JG. Developing a new
2004;22(2):85–126. https://link.springer.com/article/10.1023/B:AIRE.00000 rigorous drilling rate prediction model using a machine learning technique.
45502.10941.a9. J Petrol Sci Eng 2020;192:107338. https://doi.org/10.1016/j.petrol.2020.107338.
[95] Kovács L, Vass D, Vidács A. Improving quality of service parameter prediction with [97] Myers L, Sirois MJ. S pearman correlation coefficients, differences between. Encycl
preliminary outlier detection and elimination. Proceedings of the second Statist Sci 2004. https://doi.org/10.1002/0471667196.ess5050.
international workshop on inter-domain performance and simulation (IPS 2004), [98] Artusi R, Verderio P, Marubini E. Bravais-pearson and spearman correlation
Budapest. 2004. 2004:194-9. coefficients: meaning, test of hypothesis and confidence interval. Int J Biol Markers
2002;17(2):148–51. https://doi.org/10.1177/172460080201700213.

19