You are on page 1of 13

Fuel 268 (2020) 117389

Contents lists available at ScienceDirect

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

State-of-the-art modeling permeability of the heterogeneous carbonate oil T


reservoirs using robust computational approaches
Mehdi Mahdaviaraa, Alireza Rostamib,⁎, Khalil Shahbazib,⁎
a
Department of Petroleum Engineering, Omidieh Branch, Islamic Azad University, Omidieh, Iran
b
Department of Petroleum Engineering, Petroleum University of Technology (PUT), Ahwaz, Iran

GRAPHICAL ABSTRACT

The contour plot of the absolute relative deviation changes in various ranges of porosity and irreducible water saturation of different models including (a) GEP, (b)
GMDH, (c) Chilingarian (1990), (d) Wyllie and Rose (1950), (e) Armstrong (2003), (f) Archie (I), (g) Archie (II), and (h) Gholinezad and Masihi (2012)

ARTICLE INFO ABSTRACT

Keywords: The current study is aimed at developing straightforward and robust models for permeability prediction in
Permeability heterogeneous carbonate oil reservoirs. Two heuristic methods including Group Method of Data Handling
Heterogeneous carbonate reservoirs (GMDH) and Gene Expression Programming (GEP) were utilized to model the formation permeability with
Gene Expression Programming (GEP) respect to the static parameters including porosity, irreducible water saturation, and pore specific surface area.
Group Method of Data Handling (GMDH)
To do this, the required data points were assembled from the literature and were classified into the training and
Empirically derived models
Outliers detection
test data sets. After rigorous processing, two unique GMDH and GEP derived models were proposed for the first
time for fulfilling the scope of this work. The validity of the new suggested models was examined by utilizing
various statistical parameters integrated with visual analysis. Consequently, the results of the analysis reveal the
high exactness of the aforementioned GMDH and GEP models by Average Absolute Relative Deviations (AARD
%) of 37.15% and 16.55%, respectively. Implementing a comprehensive data and error assessment prove that
GEP based model provides the most accurate forecast of permeability than all the available and generally
published correlations for permeability. This is due to the fact that existing models are mainly developed for the
sandstone reservoirs over a restricted range of parameters; whereas, the proposed tools are created for het-
erogeneous carbonate reservoirs over a comprehensive databank. The results of sensitivity analysis demonstrate


Corresponding authors.
E-mail addresses: alireza.rostami@afp.put.ac.ir, alireza.rostami.put2014@gmail.com (A. Rostami), shahbazi@put.ac.ir (K. Shahbazi).

https://doi.org/10.1016/j.fuel.2020.117389
Received 6 February 2020; Accepted 11 February 2020
Available online 24 February 2020
0016-2361/ © 2020 Elsevier Ltd. All rights reserved.
M. Mahdaviara, et al. Fuel 268 (2020) 117389

the high impact value of irreducible water saturation and pore specific surface area on the modeling process.
Moreover, the validity of the database and GEP modeling is verified by Williams’ method, in which about 95% of
data are located in a valid region. To end with, reliable methods for permeability calculation as the most eminent
feature of the porous media are proposed here, which can be integrated with simulators or any modeling study
dealing with fluid flow in heterogeneous carbonate reservoirs.

1. Introduction regarding the abovementioned techniques [13,21,22,24–30].


Although the previously stated correlations provide precise in-
In the history of petroleum engineering development, evaluation of formation in the absence of experimental/operational data, they suffer
the inherent properties of the porous media has been accounted as a key from several limitations from the standpoints of accuracy as well as
factor in the management of the hydrocarbon-bearing reservoirs. The precision. The primary impediment is the universality of the established
key parameters in the discipline of reservoir description are the hy- empirical equations in literature. The abovementioned correlations
draulic properties of the porous media such as the formation perme- have been developed by utilizing the restricted range of parameters
ability. The absolute permeability, which is defined as the measure of with insufficient data points for a unique reservoir rock. Accordingly,
intrinsic flow capacity of a medium, plays a pivotal role in production
from the hydrocarbon reservoirs. Accordingly, measuring the afore-
mentioned parameter is a fundamental step for efficiently designing
and controlling the development of hydrocarbon fields.
Having eminent insight into the factors affecting the formation
permeability would yield useful information for measuring this para-
meter. The absolute permeability changes as a function of the complex
microstructure [1]. Hence, there have been several attempts to link the
permeability to the “straightforward-to-measure” quantities such as
porosity [2]. According to various experimental/petrophysical evalua-
tions, it is generally expected that the higher porosity would increase
the permeability [3–7]. However, this does not overwhelmingly support
the direct relationship between the porosity and permeability. In the
case of fractured carbonates, a high permeability with a low porosity
might be observed [8]. Conversely, some rocks such as pumice and
shales suffer from thoroughly poor permeability despite having entirely
high porosity [5,8]. Accordingly, an attempt was commissioned by
several researchers to connect this dynamic property (i.e., absolute
permeability) to various static properties of porous media such as
specific surface area, pore size distribution, pore throat radius as well as
irreducible water saturation [3,9,10].
In the case of absolute permeability measurement, there are three
plausible techniques including core measurement, well testing, and well
log analysis. Evaluation of the representative cores by carrying out
various flow experiments brings about valuable information regarding
the formation permeability [11]. Furthermore, the achievable data after
a cautiously designed well test operation can be employed for the cal-
culation of the average permeability of the reservoir [12]. Additionally,
by petrophysical evaluations of different logging tools such as Nuclear
Magnetic Resonance (NMR) and Dipole Shear Sonic Imager (DSI), the
engineers can calculate the absolute permeability of different forma-
tions [13]. However, the aforementioned strategies suffer from several
major drawbacks from the standpoints of time and expenses.
For alleviating the foregoing problems, several researchers have
developed theoretical and/or empirical models as well as artificial in-
telligence strategies for estimating the permeability in various reservoir
rocks [10,14–18]. Commencing with studies of Kozeny [14] and
Carman [15], many of the proposed correlations in literature linked the
permeability to the porosity for various formation rocks [13]. More-
over, some researchers have taken into account the effects of pore and
grain characteristics as well as the irreducible water saturation on the
permeability [2,19,20]. Zakirov and Galeeve [21] have utilized Lattice
Boltzmann Expressions (LBE) and Navier-Stokes Equations (NSE) to
calculate the absolute permeability of sandstones and achieved entirely
different results. Kamal et al. [22] have extended a new method to
calculate the in-situ formation permeability through the transient well
testing under the condition of three-phase flow in porous media. Qin
et al. [23] employed the petrophysical logs to establish a permeability
model for fractured tight sandstones. The interested readers are highly Fig. 1. 2D contour plot of permeability change with the input variables: (a) φ
suggested to refer to the available literature for further information and Swir versus k, (b) Sp and Swir versus k, and (c) Sp and φ versus k.

2
M. Mahdaviara, et al. Fuel 268 (2020) 117389

Table 1
A summary of the setting parameters of the developed GEP model.
Parameters Value

Number of chromosomes 30
Head size 7
Number of genes 3
Linking function Addition
Train/Test ratio 80/20
Fitness function RMSE
Mutation 0.00138
IS Transposition 0.00546
RIS Transposition 0.00546
Inversion 0.00546
Gene Recombination 0.00277
Random Chromosomes 0.0026
Random Cloning 0.00102
Operators +, -, *, /, Sqrt, Exp, Pow 10, Ln, Inv, X2, X3, X4, X5,
X1/3, X1/4, X1/5

they could not be employed for permeability assessment in another


reservoir [13]. It means that their applications to other reservoir rocks
may lead to considerable errors. Another major source of vulnerability
is the immense divergence between the lithology of reservoirs. Even
though the aforementioned correlations have been approximately ap-
plicable for estimation of permeability in sandstone rocks, they are
associated with limitations in the carbonate reservoir rocks. Two dis-
crete reasons emerge from the sophisticated procedure of permeability
Fig. 2. A schematic of the GEP model architecture. evaluation in carbonate rocks. First is the significantly more compli-
cated texture of this type of rock in comparison to the sandstones.
Second is the notable heterogeneous nature of the carbonate rocks, in
which the rock properties (particularly the rock permeability) alter
significantly [8].
To compensate for the aforesaid challenges, it is essential to employ
robust computational frameworks. Soft computing frameworks are
state-of-the-art approaches that have incrementally found their way
into the engineering applications to bridge the gap between the pre-
dictions/calculations and experimental measurements. In the last dec-
ades, a remarkable body of the literature has reported the success of
these smart approaches for putting an end on the petroleum industry
problems [31–36]. Ghiasi et al. [31] have used the Multilayer Percep-
tron (MLP) method to obtain close-to-reality predictions regarding the
water content of natural gas dried by calcium chloride dehydrator units.
Kamari et al. [33] have reported the prosper of the least square support
vector machine (LSSVM) technique for approximating the unloading
gradient pressure in continuous gas-lift systems. Chamkalani et al. [32]
have coupled the LSSVM and Particle Swarm Optimization (PSO)
methodology to estimate the deposition of asphaltene. A comprehensive
review of the application of hybrid models in chemical, petroleum, and
Fig. 3. The relative influence of the input variables on the target permeability energy systems can be found in relevant literature [36].
values. In the case of formation permeability, several artificial intelligence
models have been developed within the literature [12,18,37,38].
Input Layer Hidden Layers Output Layer
Huang et al. [37] have utilized the Back-Propagation Artificial Neural
Network (BP-ANN) for establishing the permeability model with respect
to the spatial position as well as logging data of six different wells.
Mohaghegh et al. [12] have also connected the permeability to the
petrophysical data of five wells by employing the ANN approaches.
Additionally, in the work represented by Basbug and Karpyn [18], an
ANN model was established to predict the permeability as a function of
some static properties of carbonate reservoirs. Although the afore-
mentioned models lead to fair accuracy of the assessments, they are
black boxes that do not explicitly represent the clear relationships be-
tween inputs and targets [39]. Accordingly, applying these models in
Fig. 4. A schematic of the developed GMDH model. the field applications is more strenuous than utilizing conventional
user-accessible correlations. An engineer requires a copy of trained
model software or at least the tuning parameters of the network to use
the model for solving a given field problem [39,40]. Hence, employing

3
M. Mahdaviara, et al. Fuel 268 (2020) 117389

Gene 1
K-
Expression Ln./.X2.X2.Exp./.c7.d1.c3
1

ln Ln: Natural logarithm


/: Division
X2: x to power of 2
/ Exp: Exponential
d1: Sp
Sub-ET 1 X2 X2 c7: 3.35245826593829
c3: -1.31390158187028
Exp /

c7 d1 c3

Gene 2
K-
Expression Ln./.X3.X2././.*.c5.c0.c7.d2.c7
2
ln
Ln: Natural logarithm
/: Division
/ X3: x to power of 3
X2: x to power of 2
X3 X2 *: Multiplication
d2:
Sub-ET 2
c0: -2.42027592635561
/ /
c5: -1.20650529338329
c7: -51.8854453475734
* c5 c0 c7

d2 c7

Gene 3
K-
Expression /.+.Exp././.Pow10.X3.c3.c4.d0.d2.c7
3
/: Division
/ +: Addition
Exp: Exponential
+ Exp Pow 10: 10^x
X3: x to power of 3
d0: Swir
Sub-ET 3 / / Pow 10 d2:
c3: -5.11587729318284
X3 c3 c4 d0 d2 c4: 1.06779628387339
c7: 1.31090979548069
c7

Fig. 5. A detailed description of the k-expressions and expression trees of the developed GEP model.

Table 2
The values of the various statistical parameters for the proposed GMDH and GEP models.
Group Method of Data Handling (GMDH) Gene Expression Programming(GEP)

Train Test All Train Test All

a
RE 1.19000 1.55860 1.55860 0.20260 0.14458 0.20268
SD 0.51930 0.51570 0.70320 0.64450 0.03130 0.64520
b
ARD% (min.) 0.05230 6.12000 0.05230 0.19180 0.40938 0.19183
AARD% 34.0962 61.8893 37.1469 19.1648 5.89050 16.5502
b
ARD% (max.) 228.101 375.218 375.218 412.038 13.4616 412.038
MSE 53832.9 2567.89 43735.3 297.666 2285.93 689.295
RMSE 232.019 50.6744 209.129 0.19870 47.8115 26.2544
R2 0.78640 0.96450 0.79880 0.99880 0.99630 0.99770

a
RE refers to residual error.
b
ARD refers to absolute relative deviation.

4
M. Mahdaviara, et al. Fuel 268 (2020) 117389

Fig. 6. The cross plots of the predicted permeability versus the measured per- Fig. 7. The relative deviation of the predicted training and test points from the
meability for the developed (a) GEP and (b) GMDH models. measured permeabilities for (a) GEP and (b) GMDH models.

robust computer analysis is unavoidable for developing universal and function of porosity, irreducible water saturation, and pore specific
symbolic models. surface area. The data points belong to four different carbonates in-
The current investigation was undertaken to develop computational cluding Vuktyl'skiy, central Asia, Kuybyhev, and Orenburg deposits.
intelligence models to satisfactorily estimate the absolute permeability Accordingly, this databank encompasses a wide scope of carbonate li-
in carbonate reservoirs. The authors seek to establish a novel approach thology types from limestone to dolomite, which can be used to initiate
which benefits from three key factors: (i) universality, (ii) user-acces- precious universal models. A snapshot of the permeability changes with
sibility, and (iii) high enough precision. To reach this goal, the heuristic the input variables is shown in the 2D contour plots of Fig. 1. As can be
self-organizing models including Group Method of Data Handling seen in three contour plots, the higher permeabilities correspond to the
(GMDH) and Gene Expression Programming (GEP) were utilized to higher porosities, lower irreducible water saturations, and also lower
create reasonable relationships between permeability and some com- pore specific surface areas.
monly used static variables. Three static parameters of porosity, irre-
ducible water saturation, and pore specific surface area were con-
sidered as the inputs. The training and test procedures of the modeling 3. Model architecture
were carried out by adopting the petrophysical properties of various
carbonates from the reliable literature [2]. In the end, the validity of the Due to the sophisticated procedure of absolute permeability mod-
established models was assessed using diverse statistical parameters/ eling in carbonate rocks, it is unavoidable to employ robust virtual
plots integrated with the comprehensive comparison with other lit- intelligence approaches. To address this issue, the Group Method of
erature published correlations. As an eminent step toward the perme- Data Handling (GMDH) and Gene Expression Programming (GEP) were
ability prediction in carbonate rocks, the contribution of this work is to utilized to establish permeability models with three various inputs in-
develop meticulous models that simultaneously embody both ad- cluding porosity, irreducible water saturation, and pore specific surface
vantages of the traditional correlations and artificial intelligent-based area. In order to preserve model normalization and prevent the over-
models, namely user-accessibility and higher reliability, respectively. fitting problem, the literature published data points employed in this
study were divided into training and test data sets.
Both GMDH and GEP approaches have been widely used for
2. Data presentation working with the problems of different research areas including pet-
roleum engineering. The GEP is a flexible nature-inspired paradigm
For the scope of the current study, the petrophysical properties of combining both advantages of genetic algorithm (GA) and genetic
various carbonate reservoirs have been supplemented from the reliable programming (GP) methods. Therefore, it benefits from the outstanding
literature [2]. In this dataset, the absolute permeability changes as a capability of data processing and universal searchability. The GMDH is

5
M. Mahdaviara, et al. Fuel 268 (2020) 117389

Fig. 8. Assessment of the agreement between the predicted and measured


permeability points for (a) GEP and (b) GMDH models.

a polynomial neural network benefiting from self-organizing nature,


avoiding overfitting pitfalls with high ability to solve complex non- Fig. 9. Evaluation of the AARD values pertinent to the developed GMDH and
GEP models in various ranges of the (a) effective porosity, (b) irreducible water
linear problems. The user-accessible mathematical expressions can be
saturation, and (c) pore specific surface area.
obtained from both GEP and GMDH approaches. Therefore, they in-
clude both features of smart models and traditional correlations si-
multaneously.
The GEP approach was established to overcome the prematurity
problems associated with the GP technique. However, Rao et al. [41]
have stated that simple GEP sometimes indicates instability. Besides,
the GEP tries to enhance its diversity by introducing complex numerical
constants. In some cases, however, it cannot improve the precision of
the generated model. Nevertheless, the superiority of the GEP over
GMDH from the standpoints of accuracy and flexibility has been re-
ported in the relevant literature [42]. Accordingly, appropriate
knowledge and experience regarding these approaches may have a di-
rect influence in success of the models.
A brief outline of the GMDH and GEP algorithms has been addressed
in two following subsections.

3.1. Group Method of data Handling (GMDH)

As a result of the executed investigation by Ivakhnenko in the 1960s


[43], the Group Method of Data Handling (GMDH) or Polynomial
Fig. 10. The William’s plot of the proposed GEP models.
Neural Network (PNN) has been constructed, which is a deep learning

6
M. Mahdaviara, et al. Fuel 268 (2020) 117389

Table 3
The literature published correlations utilized for the comparison with the established GMDH and GEP models.
Authors Correlation Constants All

Archie (I, II) k = ai (b )c . i = 1.2


a {2.55
9.35
Archie I
Archie II
b = 10
c = 5.65
Wyllie and Rose (1950) b c a = 79
k= a
Swi b=3
c=2
Modified Chilingarian (1990) k = 10[a + bSwi+ cSp+ d + eSwi Sp + fSwi + gSp ] a = 3.17597
b = 0.05748
c = 0.00363
d = 0.04083
e = 0.00002
f = 0.00351
g = 0.0001
Armstrong (2003)
( ) {
c
k=a b 1
1 10 k < 200
a
Swi 1 k 200
b = 1.5
c = 1.5
Modified Gholinezhad and Massihi (2012) b ×
1 Swi a = 138.408
k = a10 1 Swi
b = 0.163

Table 4
The values of the various statistical parameters for the new proposed models and the traditional correlations.
GEP GMDH Archie (I) Archie (II) Wyllie and Rose (1950) Modified Chilingarian (1990) Armstrong (2003) Modified Gholinezad and Massihi (2012)

SD 0.1992 0.5308 10.8361 40.7774 0.89480 0.78750 0.84370 70.1572


AARD% 16.5502 37.1469 470.590 1721.90 87.5490 69.5907 80.3311 2890.99
MSE 689.295 43735.3 1.49e07 1.68e07 1.59e07 2.32e05 1.74e07 1.08e07
RMSE 26.2544 209.130 3865.06 4103.89 3983.48 482.033 4170.06 3279.70
R2 0.99770 0.79880 0.05720 0.05720 0.38090 0.98980 0.00021 0.24400

Fig. 12. Comparing the relative deviation of the developed GMDH and GEP
models against the literature published correlations.
Fig. 11. A comparison amongst cross plots of the proposed GMDH and GEP
models and the literature published correlations. For the principle of the aforementioned algorithm, the intricate
discrete form of the Volterra series, namely the Kolmogorov-Gabor (KG)
machine. This strategy has recently grabbed the attention of scientists polynomial is manipulated to prognosticate the target using the mul-
in various fields such as petroleum engineering. This heuristic tech- tiple input variables as follows [44–47]:
nique provides means by which the non-linear and complex issues can
N N N N N N
be easily solved. Perhaps, the most serious advantage of the GMDH is its y = f (X ) = a 0 + wi xi + wij x i xj + wijk x i xj xk
self-organizing nature. Accordingly, the number of neurons/layers and i=1 i=1 j=1 i=1 j =1 k=1 (1)
the net/transfer functions are achieved autonomously. The second ad-
vantage of employing the GMDH method is the user accessibility of the The symbols y, x and w stand for the target, the input variable, and
final model. weight coefficient, respectively, and the entity N is the number of

7
M. Mahdaviara, et al. Fuel 268 (2020) 117389

head, which encompasses both operators (e.g. addition, subtraction,


multiplication, and division) and terminals (e.g. variables and con-
stants), and (2) the tail, which includes only terminal symbols. The
information is decoded from the chromosomes to the expression trees
(ETs). In this translation process, the genes in the chromosomes are
expressed in sub-trees, and they are connected to each other using a
pre-defined linking function (e.g., +, -, *) [54–56]. An example of the
chromosome and the expression tree of a typical GEP model is revealed
in Fig. 2. Following operations are executed during a GEP modeling:

1. Initialization: A set of chromosomes is generated randomly in order


to form the initial population.
2. Translation: All the produced chromosomes are converted to the
expression trees.
3. Execution: Each program encompasses chromosomes is im-
plemented.
4. Fitness Evaluation: A fitness evaluation function is employed to
survey the fitness value of each chromosome. If each chromosome
Fig. 13. The cumulative frequency of the ARD values of the established GMDH met the termination condition, the modeling process is accom-
and GEP models and the literature published correlations. plished. Otherwise, the best chromosomes are selected from the
population in order to generate the recent population. To this end,
inputs. To alleviate the sophisticated procedure of solving the aforesaid different selection approaches such as roulette-wheel selection
series, it can be streamlined to second-order transportation of the input strategy and tournament selection strategy can be employed [57].
variables as follows [43,48]: 5. Replication: In order to reproduce the modified generation of chro-
mosomes, the replication and genetic improvement procedures are
z iGMDH = w0 + w1 x i + w2 xj + w3 x i xj + w4 xi2 + w5 x 2j (2) carried out using various operations including transportation, mu-
tation, and recombination. At the end, the procedure is repeated
Accordingly, a network of linked neurons is generated to establish
until satisfying the termination criterion.
the input–output relationship in the Kolmogorov-Gabor (KG) poly-
nomial of Eq. (1). This network is a feed-forward transformation, in
The readers are referred to the literature for more enlightenments
which only the neurons that outperform others according to a “self-
regarding the theoretical and practical aspects of the GEP modeling
selection thresholds” are permitted to pass to the following layer [43].
approach [54,56,58].
To establish the superlative neurons, efforts are directed toward the
achievement of the optimum weights for the foregoing quadratic
4. Results and discussion
equation (i.e., Eq. (2)) by employing the least-squares technique as
follows:
The current section outlines and discusses the major findings of this
W T = YX T (XX T ) 1
(3) investigation.

where W, X, and Y stand for the vectors of unknown weights, inputs,


4.1. Establishment of the proposed models
and outputs, respectively, and the superscript T represents the trans-
pose operator.
The current study is particularly designed to speculate the absolute
permeability of the carbonate reservoirs by using the fact that it alters
3.2. Gene expression Programming (GEP) with various static parameters of the rock. As previously stated, many
published papers shed light on the factors affecting the absolute per-
Development of the Evolutionary Algorithms (EAs) in the last meability of reservoir rocks [2,3,6,9,10,12]. A recurrent theme in the
decade has been a snapshot of the state-of-the-art in the field of arti- literature published investigations was the robust relevance of the
ficial intelligence. Evolutionary algorithms are population-based heur- formation permeability to the porosity, irreducible water saturation,
istic algorithms that employ mechanisms aroused by genetics and nat- and pore specific surface area [2,3,6,9,10,12]. Due to this fact, we have
ural selection. The aforementioned algorithms utilize biological considered these parameters as the inputs of the models. Accordingly, a
evolutionary mechanisms including selection, mutation, and cross over wide variety of data points have been gathered from reliable open lit-
to deal with complicated problems. There are three significant gen- erature [2] for linking the aforementioned parameters.
erations of EAs, namely Genetic Algorithms (GA) [49,50], Genetic For evaluating the relevancy of the output to the independent input
Programming (GP) [51,52], and Gene Expression Programming (GEP) variables in the literature adopted data points, a sensitivity analysis was
[53]. GAs are the first generation of the EAs, which employ linear accomplished by applying the following mathematical expression [59]:
strings of fixed length (chromosomes) as the population of the in- n
dividuals [53]. GPs are the extended version of the GAs, which utilize i=1
(x k xk . i )(y yi )
r=
the nonlinear entities of different sizes and shapes (parse trees) as the n
(x k xk . i )
2 n
(y yi )
2
i=1 i=1 (4)
individuals [53]. GEP which has been employed in the current study is
a further supplement of the GAs and GPs. This generation of EAs utilizes where r stands for the relevancy factor. This entity alters between −1
linear strings of fixed length (linear representations called chromo- and +1 for evaluating the rate of permeability change with porosity,
somes) integrated with nonlinear entities of different sizes and shapes irreducible water saturation, and pore specific surface area. The r va-
(ramified representations called expression trees). Accordingly, the GEP lues of 0, −1, and +1 correspond to the lack of relevancy, thoroughly
implements based on two straightforward entities, namely chromo- inverse, and entirely direct relationships, respectively. The k and n
somes (Genotypes or Karva-expressions) and expression trees (Pheno- values represent the input type and the number of data points, re-
types). The chromosomes usually contain one or more parts of equal spectively. The entities x and y are the average input and target values,
length termed genes. The genes are comprised of two sections: (1) the respectively. Fig. 3 demonstrates the results of the aforementioned

8
M. Mahdaviara, et al. Fuel 268 (2020) 117389

Fig. 14. The contour plot of the absolute relative deviation change in various ranges of porosity and irreducible water saturation of different models including (a)
GEP, (b) GMDH, (c) Chilingarian (1990), (d) Wyllie and Rose (1950), (e) Armstrong (2003), (f) Archie (I), (g) Archie (II), and (h) Gholinezad and Masihi (2012).

9
M. Mahdaviara, et al. Fuel 268 (2020) 117389

assessment. As it is evident, the considered input variables entirely N


kiexp kipred
2
1
influence the target, which supports the aforesaid literature experi- SD =
N 1 kiexp (11)
mental findings. As shown, irreducible water saturation is the most i=1

affecting variable on permeability estimation. However, due to the N


stochastic nature of the carbonate rocks, there are data points that 100 kiexp kipred
ARD% =
contravene the abovementioned rule of thumb. Accordingly, the so- N i=1
kiexp (12)
phisticated procedure of permeability estimation becomes more pro-
N
blematic. 100 kiexp kipred
AARD% =
For the scope of the current study, two robust heuristics self-orga- N i=1
kiexp (13)
nizing strategies, namely Group Method of Data Handling (GMDH) and
Gene Expression Programming (GEP) were utilized to rigorously model 1
N

the permeability of carbonate reservoirs. For this, about 80% and 20% MSE = (kiexp kipred )2
N i=1 (14)
of data points were chosen randomly to design the training and test sub-
data, respectively. The three foregoing parameters namely porosity, N
1
irreducible water saturation, and pore specific surface area were con- RMSE = (kiexp kipred ) 2
N (15)
templated as the independent input variables: i=1

k = f ( , Swir , Sp) (5) N


i=1
(kipred kiexp )2
R2 = 1
N
Comprehensive modeling was accomplished according to the GMHD i=1
(kipred kiexp )2 (16)
methodology represented in the previous section. The model execution 2
where the entities RE, SD, ARD, AARD, RMSE, and R stand for the
was led to the following mathematical expressions:
Residual Error, Standard Deviation, Absolute Relative Deviation,
k = exp (a + bSwir + cSwir N1 + dSwir 2 + eN1) (6) Average Absolute Relative Deviation, Root Mean Square Error, and
Determination Coefficient, respectively. The pred and exp superscripts
in which denote the predicted and experimental values of the absolute perme-
N1 = f + g 2 + hN2 + iN2 2 (7) ability, respectively. Table 2 represents the values of the statistical
parameters for both GMDH and GEP based models. As can be seen from
N2 = j + mSp + lSP Swir (8) the table, despite the privations associated with modeling the perme-
ability in carbonate reservoirs, both proposed GEP (i.e., Eq. (9)) and
where N1 and N2 are neuron formulations that alter as a function of GMDH (i.e., Eq. (6)) based models benefit from fair enough accuracy/
porosity ( , fraction), irreducible water saturation (Swir, %), and pore precision for estimation of formation permeability. Moreover, the GEP
specific surface area (Sp, cm−1). The tuning parameters of the corre- model corresponds to more reliable statistical entities. For instance, the
lations are a = -0.0854968; b = 0.0317332; c = -0.00698438; d = - values of the AARD, RMSE, R2, and SD for the GEP model are
0.000571046; e = 1.05957; f = -0.809153; g = 55.6646; h = - 16.5502%, 26.2544, 0.9977, and 0.6452, and for the GMDH approach
0.0396009; i = 0.184891; j = 6.30324; m = -0.00137263; and are 37.1469%, 209.1299, 0.7988, and 0.7032, respectively.
l = 1.57506E-5. The root mean square of errors (RMSE) entity was
selected as the validation criterion of the developed GMDH model. The 4.2.2. Visual illustration
initial layer width and the maximum number of the layers are 300 and In addition to the pre-mentioned parameters, a visual illustration of
3, respectively. A detailed description of the established GMDH model the accuracy of the extended models was addressed by employing a
is schematically illustrated in Fig. 4. plethora of statistical plots including cross plot, relative divergence
In addition to the generated GMDH model, the succeeding simpli- distribution, group error assessment, and 2D surface and contour plots.
fied expression signifies another model extended by the GEP technique The cross plots of the developed GMDH and GEP models for the training
for permeability prediction in carbonate reservoirs: and test data are demonstrated in Fig. 6. As it is evident, all the training
3
and test data points are distributed in the vicinity of the unit slope line,
bSwir 1 + c
k= a exp which indicates the high consistency between the prognosticated points
Sp 2 exp (10 ) (9) using both GEP and GMDH heuristic algorithms and the experimental
data points. Furthermore, Fig. 7 represents the relative deviation of the
where a, b, and c which are tuning parameters are corresponding to the
calculated points from the targets, which provides more verification
values of 5.1516246092E10, 1.067796284, and -0.44035031, respec-
regarding the correctness of the developed GMDH and GEP models. The
tively. In Eq. (9), the units of porosity, irreducible water saturation and
points situated nearby the horizontal zero-line represent more fitness to
pore specific surface area are, in turn, fraction, percentage (%) and
the corresponding measured data. According to Fig. 7, the majority of
cm−1. Additionally, the setting values of the the suggested GEP model
the data points are distributed alongside the zero-line, which describes
are summarized in Table 1. As it is evident, the established GEP model
the fact that the suggested models are prosperous for the estimation of
encompasses 30 chromosomes and 3 genes, in which the genes linked to
permeability in carbonate and heterogeneous reservoir rocks. It should
each other using an addition (+) operator. The genes embody various
be mentioned that the GEP based model gives more precise estimates of
operators including +, *, /, Exp, Pow 10, Ln, X2, and X3. A detailed
permeability because the predicted values are highly accumulated
description of three k-expressions and the expression trees of the GEP-
around the 45° line in Fig. 6 and zero line in Fig. 7. Moreover, the range
derived model is demonstrated in Fig. 5.
of error for GEP and GMDH based models are mainly about −20 to
+20%, and −40 to +40%, respectively. Fig. 8 is another perceptible
4.2. Validity analysis of the proposed models benchmark for investigating the exactitude of the GMDH and GEP
based models. As can be seen, the training and test points pertinent to
4.2.1. Statistical parameters both established GMDH and GEP models have a satisfactory agreement
The genuineness of the proposed GMDH and GEP derived models with the measured data points. This is another attestation regarding the
has been investigated by utilizing copious statistical criteria as follows: potential of the proposed models here for estimating the permeability in
the heterogeneous carbonate reservoirs. As can be seen, the GEP model
RE = (kiexp kipred ) (10) gives more reliable estimates than the GMDH in comparison to the

10
M. Mahdaviara, et al. Fuel 268 (2020) 117389

experimental data points. literature correlations including Archie (I) and (II) [8], Wyllie and Rose
Furthermore, the faithfulness of the established GMDH and GEP [64], Chilingarian [2], Armstrong [19], and Gholinajad and Masihi [8].
methodologies was evaluated in various ranges of the employed input As it is evident, the aforementioned correlations embody the various
variables. The forgoing assessment is graphically depicted in Fig. 9(a–c) number of variables. The porosity is the only input variable of the Ar-
for porosity, irreducible water saturation, and pore specific surface chie (I, II) [8] models; the Wyllie and Rose [64], Armstrong [19], and
area, respectively. The following statements represent the major out- Gholinezhad and Masihi [8] correlations encompass the and Swi in-
lines of the analysis: puts; the correlation of the Chilingarian [2] is dependent to , Swi , and
Sp . Additionally, these literature correlations encompass diverse num-
• The proposed GMDH and GEP derived correlations indicate sa- bers of tuning parameters from two (e.g., Gholinezhad and Masihi [8]
tisfactory performance in all ranges of the input variables. correlation) to seven (e.g., Chilingarian [2] correlation). However,
• The GEP strategy claims preferable performance than the GMDH other literature correlations studied here utilize three constant para-
model in all ranges of the studied inputs. meters for fitting on the experimental data points. It should be worth
• The excellent correctness of the GEP model occurs at the porosity, pointing out that Chilingarian [2] and Gholinajad and Masihi [8] did
irreducible water saturation, and pore specific surface area ranges of not represent unique tuning parameters (constants) for their correla-
19–24%, 39–52%, and 8358–10164, respectively. This illustrates tions. Hence, a regression analysis was implemented using the Gauss-
the superior capability of the GEP method in high ranges of the Newton as well as Levenberg-Marquardt algorithms [65] in order to
input parameters. obtain the best constants for the aforementioned empirical equations.
Accordingly, the Chilingarian [2] and Gholinajad and Masihi [8] cor-
The benefits of the GEP model are not limited to the higher accuracy relations were modified in accordance with the data points employed in
and the lack of overfitting problems. This strategy has established a the current study.
more straightforward correlation with only three tuning parameters, The validity of the abovementioned literature published correla-
which are unique for all datasets. This ease of use integrated with tions for fitting on the data points employed in this study was in-
higher accuracy are vital components of the GEP derived correlation vestigated using various statistical parameters including SD, ARD%,
making it the obvious choice for estimation of permeability in carbo- AARD%, MSE, RMSE, and R2. Table 4 compares the values of these
nate rocks. parameters for the suggested GMDG and GEP models with the pre-
viously reported correlations. As can be seen in the table, both GMDH
4.2.3. Applicability domain of the established GEP model and GEP models easily outperform traditional equations in accordance
Another important issue is diagnosis the observation points, which with all the above-described statistical parameters. For instance, the
are allocated in an abnormal distance with respect to the majority of values of the AARD are 16.5502%, 37.1469%, 470.590%, 1721.90%,
data points. The existence of these outliers may be owing to various 87.5490%, 69.5907%, 80.3311%, and 2890.99% for the GEP, GMDH,
factors including the experimental errors. Accordingly, detection of the Archie (I), Archie (II), Wyllie and Rose [64], modified Chilingarian [2],
aforementioned data points is essential to avoid inaccuracy and un- Armstrong [19], and modified Gholinezad and Masihi [8] correlations,
reliability in the developed model [60–63]. There are several techni- respectively. These values reveal that the predicted permeabilities using
ques such as the Leverage approach for detecting and eliminating the the traditional correlations highly deviate from the corresponding ex-
outliers. The readers are advised to refer to the literature for more perimental values. According to the AARD parameter, the correctness of
description regarding the concepts of outliers detection [60,61,63]. In the studied models can be ranked in the following order:
the current study, the Leverage method integrated with Williams’ plot GEP > GMDH > modified Chilingarian [2] > Armstrong
was employed to visualize the outliers. Fig. 10 represents the Williams’ [19] > Wyllie and Rose [64] > Archie (I) > Archie (II) > modified
plot of the established GEP based model, in which the standardized Gholinezad and Masihi [8].
residual (R) of the model is sketched versus the Leverage values (Hat Hence, the GEP based correlation establishes itself as the most re-
indices). Allocating the points in various domains is associated with the liable model for estimation of permeability in carbonate rocks. The
following diagnoses: modified version of Chilingarian [2] correlation, on the other hand,
performs better in comparison to other literature published models.
• Bad High Leverage (Outliers): R > 3 or R < 3 The deviation of the permeability points calculated using various
• Good High Leverage: 3 R 3 and H > H models from the measured points are demonstrated in the cross plot of
• Applicability domain: 3 R 3 and H < H (the squared area) Fig. 11. The presence of the points related to the proposed GMDH and
GEP models in the vicinity of the unit slop line corresponds to the high
The entity H stands for the critical Leverage which can be defined superiority of these methodologies in comparison to other literature
as follows: correlations. However, there are discrepancies in the case of the tra-
3f + 1 ditional models, which introduce systematic errors in their process of
H = permeability assessment. Similar results are also achieved by Fig. 12
p (17)
which indicates the relative deviation of each model from the experi-
where f and p entities denote the number of model parameters and the mental data. The points pertinent to the GMDH and GEP approaches are
number of data points, respectively. accumulated nearby the horizontal line, which reveals the better per-
As it is demonstrated in Fig. 10, the predominant number of data formance of the proposed models in this study.
points (95.45%) are located within the squared area of 3 R 3 and The cumulative frequency plot of Fig. 13 depicts another visual tool
H <0.1818, namely the applicability domain of the proposed GEP for testing the suggested correlations in this study against the tradi-
model. There are only one and two points located at the Good High tional literature published models over the entire dataset utilized in this
Leverage and Bad High Leverage domains, respectively. This is another work. This statistical approach provides graphical plots through which
attestation regarding the validity and reliability of the established GEP the frequency of ARD values of different models can be precisely de-
model over the datasets employed in this study. tected. The higher the frequency of low ARDs, the lower uncertainty
can be assigned to the proposed model. A sharp ascending trend of the
4.2.4. Comparison with literature published models curve at low ARDs corresponds to an appropriate consistency between
The proposed GMDH and GEP derived models were assessed against the calculated and exact experimental values. As can be seen in Fig. 13,
various literature published correlations by utilizing different statistical about 95.45% and 54.54% of calculations conducted by GEP and
parameters as well as visual descriptions. Table 3 lists the foregoing GMDH models, respectively, correspond to the errors less than about

11
M. Mahdaviara, et al. Fuel 268 (2020) 117389

20%. This indicates that the recently proposed models perform well on permeability with the AARD and R2 values of 16.55% and 0.9977 in
the suit of permeability estimation. However, the GEP approach out- heterogeneous carbonate reservoirs.
performed the GMDH method, which supports the above-mentioned • The accuracy of the proposed GEP and GMDH derived models are
statistical results. According to Fig. 13, the calculations carried out by highly better than the widely used published correlations available
traditional correlations are fraught with uncertainties. The modified in the literature. The AARD value of the literature permeability
Chilingarian [2] correlation represents the best results among the lit- models varies from 69.59% to 2890.99%, which approves the su-
erature published models. However, only about 18% of Chilingarian [2] periority of the established methods in this study.
prognostications have ARDs less than 20% and the rest of the predic- • The irreducible water saturation is the most affecting variable in
tions are associated with remarkable mismatch with respect to the real estimating permeability based on the sensitivity analysis.
experimental values. • The outlier’s analysis shows that about 95.45% of the database and
In the end, the distribution of the ARD values over the various model predictions are valid and accommodate in the applicability
ranges of the porosity and irreducible water saturation is evaluated for domain.
different models. The results of the analysis are sketched in the 2D • Finally, it can be proposed that the created rigorous methods pre-
contour plot of Fig. 14. The dark blue and dark green colors represent sented here are crucially important in analyzing fluid flow in porous
the ARD values of less than 25% and more than 1200%, respectively. media through petroleum engineering studies, especially in sub-
Other ARD values allocate between the aforesaid limits. As can be seen, terranean reservoir simulations.
the major area of the GEP contour plot is uniformly covered by dark
blue shading [Fig. 14 (a)], whereas an entirely constrained plot area CRediT authorship contribution statement
with less accuracy is concentrated around = 14% and Swir = 25%
point. Similarly, the main portion of the GMDH contour map attribute Mehdi Mahdaviara: Writing - original draft, Visualization,
to the light and dark blue colors. However, the colors are not regularly Investigation. Alireza Rostami: Conceptualization, Formal analysis,
distributed over the map [Fig. 14 (b)]. In the contour maps pertinent to Methodology, Software, Validation, Visualization, Writing - original
the traditional literature correlations, there are a few or no regions draft, Writing - review & editing. Khalil Shahbazi: Supervision,
containing the aforementioned optimal color. The domain portion of Writing - review & editing.
the Chilingarian [2], Wyllie and Rose [64], and Armstrong [19] plots
corresponds to the ARD range of 50–100, namely light blue color Declaration of Competing Interest
[Figures (c)–(e)]. Additionally, the dominant part of the Archi (I, II) and
Gholinezad and Masihi [8] plots attribute to the light and dark green The authors declare that they have no known competing financial
shadings [Fig. 14 (f)-(h)]. Accordingly, this is another testimony re- interests or personal relationships that could have appeared to influ-
garding the more accuracy, precision, and applicability of the generated ence the work reported in this paper.
models in the current work in comparison to the other literature pub-
lished correlations. References
The proposed paradigms in this study utilize a wide range of data to
create two permeability models; so, the established models are uni- [1] Mostaghimi P, Blunt MJ, Bijeljic B. Computations of absolute permeability on
versal. As it has been known, the application of any model beyond the micro-CT images. Math Geosci 2013;45(1):103–25.
[2] Chilingarian GV, Chang J, Bagrintseva KI. Empirical expression of permeability in
applicability domain may lead to noticeable errors. However, if the terms of porosity, specific surface area, and residual water saturation of carbonate
established model uses an all-inclusive database, it may lead to less rocks. J. Petrol Sci. Eng. 1990;4(4):317–22.
prediction error when it will be applied to the out-range dataset be- [3] Chilingarian GV, Torabzadeh J, Rieke HH, Metghalchi M, Mazzullo S.
Interrelationships among surface area, permeability, porosity, pore size, and re-
cause the trends are highly understood by the established model. sidual water saturation. Developments in petroleum science. Elsevier 1992:379–97.
Consequently, the suggested models can give us good estimates of [4] Akam SA, Maher TF, Schell-Murdey C, Flow Arnott S. Quality Indicator (FQI): An
carbonate rock permeability when they are extrapolated beyond the Innovative Approach to Permeability Prediction. SPE Asia Pacific Oil and Gas
Conference and Exhibition. Soc Petrol Eng 2010.
range of applicability. [5] Dandekar A. Critical evaluation of empirical gas condensate correlations. J Nat Gas
The GEP and GMDH models can be used for reservoir character- Sci Eng 2015;27:298–305.
ization and simulation in any pertinent college studies or industrial [6] Chilingar GV. Relationship between porosity, permeability, and grain-size dis-
tribution of sands and sandstones. Dev Sedimentol 1964:71–5.
project. Moreover, the proposed techniques can be integrated into any
[7] Saar MO, Manga M. Permeability-porosity relationship in vesicular basalts. Geophys
theoretical modeling and commercial software. Thereby, the proposed Res Lett 1999;26(1):111–4.
tools can be of immense value for scientists and expert engineers [8] Gholinezhad S, Masihi M. A physical-based model of permeability/porosity re-
dealing with fluid flow in porous media studies. lationship for the rock data of Iran southern carbonate reservoirs. Iran J Oil Gas Sci
Technol 2012;1(1):25–36.
[9] Mortensen J., Engstrom F., Lind I. The relation among porosity, permeability, and
5. Conclusions specific surface of chalk from the Gorm field, Danish North Sea. SPE-83960-PA
1998;1(03):245-51.
[10] Timur A. An investigation of permeability, porosity, & residual water saturation
The present study was undertaken to model the permeability of relationships for sandstone reservoirs. Log Anal 1968;9(04).
carbonate and heterogeneous reservoirs by utilizing two heuristic self- [11] Sander R, Pan Z, Connell LD. Laboratory measurement of low permeability un-
organizing algorithms namely Group Method of Data Handling (GMDH) conventional gas reservoir rocks: A review of experimental methods. J Nat Gas Sci
Eng 2017;37:248–79.
and Gene Expression Programming (GEP) for the first time in literature. [12] Mohaghegh S, Arefi R, Ameri S, Rose D. Design and development of an artificial
At first, a set of permeability data points encompassing porosity, irre- neural network for estimation of formation permeability. SPE Computer
ducible water saturation, and pore specific surface area variables were Applications 1995;7(06):151–4.
[13] Babadagli T, Al-Salmi S. A review of permeability-prediction methods for carbonate
gathered from the reliable literature. Then, by a random computational reservoirs using well-log data. SPE Reservoir Eval. Eng. 2004;7(2):75–88.
approach, the database used here were separated into two groups of test [14] Kozeny J. Ueber kapillare Leitung des Wassers im Boden. Sitzungsber Akad
and train. As a result of GEP and GMDH modeling as robust strategies, 1927;136:271–306.
[15] Carman PC. Fluid flow through granular beds. Chem Eng Res Des 1997;75:S32–48.
two symbolic and unique equations were established. The accuracy and
[16] Coats GR, Denoo S. The producibility answer product. Tech Rev 1981;29:55–144.
applicability of the proposed models were assessed in multiple stages [17] Babadagli T, Al-Salmi S. Improvement of permeability prediction for carbonate
including visual methods and numerical parameters. The major results reservoirs using well log data. SPE Asia Pacific Oil and Gas Conference and Exhibition.
of the current investigation are as follows: Soc Petrol Eng 2002.
[18] Basbug B, Karpyn ZT. Estimation of permeability from porosity, specific surface
area, and irreducible water saturation using an artificial neural network. Latin
• The GEP based model gives the best results for forecasting

12
M. Mahdaviara, et al. Fuel 268 (2020) 117389

American & Caribbean Petroleum Engineering Conference. Society of Petroleum relative permeability of gas condensate reservoirs: Advanced computational fra-
Engineers; 2007. meworks. J Petrol Sci Eng 2020::106929.
[19] Armstrong OP. Permeability Correlations for Carbonate and Other Rocks. 2003 [43] Ivakhnenko AG. Polynomial theory of complex systems. IEEE Trans. Syst, Man,
Seattle Annual Meeting. 2003. Cybernetics 1971;4:364–78.
[20] Swanson B. A simple correlation between permeabilities and mercury capillary [44] Wu J, Wang Y, Zhang X, Chen Z. A novel state of health estimation method of Li-ion
pressures. J Petrol Technol 1981;33(12). 2,498–2,504. battery using group method of data handling. J Power Sources 2016;327:457–64.
[21] Zakirov T, Galeev A. Absolute permeability calculations in micro-computed tomo- [45] Rostami A, Hemmati-Sarapardeh A, Karkevandi-Talkhooncheh A, Husein MM,
graphy models of sandstones by Navier-Stokes and lattice Boltzmann equations. Int Shamshirband S, Rabczuk T, et al. Modeling heat capacity of ionic liquids using
J Heat Mass Transf 2019;129:415–26. group method of data handling: A hybrid and structure-based approach. Int J Heat
[22] Kamal MM, Morsy S, Suleen F, Pan Y, Dastan A, Stuart MR, et al. Determination of Mass Transf 2019;129:7–17.
in-situ reservoir absolute permeability under multiphase-flow conditions using [46] Sadi M, Shahrabadi A. Evolving robust intelligent model based on group method of
transient well testing. SPE-83960-PA 2019;22(01):336-50. data handling technique optimized by genetic algorithm to predict asphaltene
[23] Qin Z, Wu D, Luo S, Ma X, Huang K, Tian F, et al. A Novel Method to Obtain precipitation. J Petrol Sci Eng 2018;171:1211–22.
Permeability in a Dual-Pore System Using Geophysical Logs: A Case Study of an [47] Najafzadeh M, Barani G-A, Hessami-Kermani M-R. Group method of data handling
Upper Triassic Formation, Southwest Ordos Basin, China. Nat Resour Res to predict scour depth around vertical piles under regular waves. Scientia Iranica
2020:1–16. 2013;20(3):406–13.
[24] Ahmed U, Crary S, Coates G. Permeability estimation: the various sources and their [48] Hemmati-Sarapardeh A, Mohagheghian E. Modeling interfacial tension and
interrelationships. J Petrol Technol 1991;43(05):578–87. minimum miscibility pressure in paraffin-nitrogen systems: Application to gas in-
[25] Nelson PH. Permeability-porosity relationships in sedimentary rocks. Log Analyst jection processes. Fuel 2017;205:80–9.
1994;35(03). [49] Holland J. Adaptation in Natural and Artificial Systems“ Ann Arbor: University of
[26] Yao C, Holditch S. Estimating permeability profiles using core and log data SPE Michigan Press; 1975.
Eastern Regional Meeting. Society of Petroleum Engineers; 1993. [50] Goldberg DE, Holland JH. Genetic algorithms and machine learning. Machine
[27] Mohaghegh S, Balan B, Ameri S. Permeability determination from well log data. learning 1988;3(2):95–9.
SPE Form Eval 1997;12(03):170–4. [51] Cramer NL. A representation for the adaptive generation of simple sequential
[28] Saner S, Kissami M, Al Nufaili S. Estimation of permeability from well logs using programs. Proceedings of the first international conference on genetic algorithms.
resistivity and saturation data. SPE Form Eval 1997;12(1):27–31. 1985. p. 183–7.
[29] Wendt WA, Sakurai S, Nelson PH. Permeability prediction from well logs using [52] Koza JR, Koza JR. Genetic programming: on the programming of computers by
multiple regression Elsevier Reservoir Charact.1986:181–221. means of natural selection. MIT press; 1992.
[30] Singh N. Permeability prediction from wireline logging and core data: a case study [53] Ferreira C. Gene Expression Programming: A New Adaptive Algorithm for Solving
from Assam-Arakan basin. J Pet Explor Prod Technol 2019;9(1):297–305. Problems. 2001.
[31] Ghiasi MM, Bahadori A, Zendehboudi S. Estimation of the water content of natural [54] Hong T, Jeong K, Koo C. An optimized gene expression programming model for
gas dried by solid calcium chloride dehydrator units. Fuel 2014;117:33–42. forecasting the national CO2 emissions in 2030 using the metaheuristic algorithms.
[32] Chamkalani A., Zendehboudi S., Bahadori A., Kharrat R., Chamkalani R., James L., Appl Energy 2018;228:808–20.
et al. Integration of LSSVM technique with PSO to determine asphaltene deposition. [55] Alkroosh I, Nikraz H. Correlation of Pile Axial Capacity and CPT Data Using Gene
2014;124:243-53. Expression Programming. Geotech Geol Eng 2011;29(5):725–48.
[33] Kamari A., Bahadori A., Mohammadi A.H., Zendehboudi S.J.P.S., Technology. [56] Ferreira C. Gene expression programming: mathematical modeling by an artificial
Evaluating the unloading gradient pressure in continuous gas-lift systems during intelligence. Springer; 2006.
petroleum production operations. 2014;32(24):2961-8. [57] Zhong J, Hu X, Zhang J, Gu M. Comparison of performance between different se-
[34] Arabloo M, Bahadori A, Ghiasi MM, Lee M, Abbas A, Zendehboudi S. A novel lection strategies on simple genetic algorithms. International Conference on
modeling approach to optimize oxygen–steam ratios in coal gasification process. Computational Intelligence for Modelling, Control and Automation and
Fuel 2015;153:1–5. International Conference on Intelligent Agents, Web Technologies and Internet
[35] Ghiasi MM, Bahadori A, Zendehboudi S, Chatzis I. Rigorous models to optimise Commerce (CIMCA-IAWTIC'06). 2. IEEE; 2005:1115-21.
stripping gas rate in natural gas dehydration units. Fuel 2015;140:421–8. [58] Zhong J, Feng L, Ong Y. Gene Expression Programming: A Survey [Review Article].
[36] Zendehboudi S., Rezaei N., Lohi A.J.A.E. Applications of hybrid models in chemical, IEEE Comput Intell Mag 2017;12(3):54–72.
petroleum, and energy systems: A systematic review. 2018;228:2539-66. [59] Chok NS. Pearson's versus Spearman's and Kendall's correlation coefficients for
[37] Huang Z, Shimeld J, Williamson M, Katsube J. Permeability prediction with arti- continuous data. University of Pittsburgh; 2010.
ficial neural network modeling in the Venture gas field, offshore eastern Canada. [60] Rousseeuw PJ, Leroy AM. Robust regression and outlier detection. Wiley Online
Geophysics 1996;61(2):422–36. Library; 1987.
[38] Mohaghegh S. Virtual-intelligence applications in petroleum engineering: Part [61] Gramatica P. Principles of QSAR models validation: internal and external.
1—Artificial neural networks. J Petrol Technol 2000;52(09):64–73. 2007;26(5):694-701.
[39] Tu JV. Advantages and disadvantages of using artificial neural networks versus [62] Hemmati-Sarapardeh A, Ameli F, Dabir B, Ahmadi M, Mohammadi AH. On the
logistic regression for predicting medical outcomes. J Clin Epidemiol evaluation of asphaltene precipitation titration data: modeling and data assessment.
1996;49(11):1225–31. Fluid Phase Equilib 2016;415:88–100.
[40] Astion ML, Wener MH, Thomas RG, Hunder GG, Bloch DA. Application of neural [63] Goodall CR. 13 Computation using the QR decomposition. Handbook of Statistics.
networks to the classification of giant cell arteritis. Arthritis Rheumatism: Off J Am Elsevier 1993:467–508.
Coll Rheumatol 1994;37(5):760–70. [64] Wyllie M, Rose WD. Some theoretical considerations related to the quantitative
[41] Rao Y, Wang R-c, Yuan C-a. Improving gene expression programming using parallel evaluation of the physical characteristics of reservoir rock from electrical log data. J
taboo search. 2009 Fifth International Conference on Natural Computation. 3. IEEE; Petrol Technol 1950;2(04):105–18.
2009:144-8. [65] Marquardt DW. An algorithm for least-squares estimation of nonlinear parameters.
[42] Mahdaviara M, Menad NA, Ghazanfari MH, Hemmati-Sarapardeh A. Modeling J Soc Ind Appl Math 1963;11(2):431–41.

13

You might also like