Petroleum xxx (2017) 1e7

Contents lists available at ScienceDirect

Petroleum
journal homepage: www.keaipublishing.com/en/journals/petlm

Original Article

Decline curve based models for predicting natural gas well

performance
Arash Kamari a, Amir H. Mohammadi a, b, c, *, Moonyong Lee d, Alireza Bahadori e, f, **
a
Disciples of Chemical Engineering, School of Engineering, University of KwaZulu-Natal, Howard College Campus, King George V Avenue, Durban 4041,
South Africa
b
Institut de Recherche en Genie Chimique et Petrolier (IRGCP), Paris Cedex, France
c 
Departement de G enie des Mines, de la Metallurgie et des Materiaux, Facult
e des Sciences et de G
enie, Universit
e Laval, Qu
ebec, QC G1V 0A6, Canada
d
School of Chemical Engineering, Yeungnam University, Gyeungsan, Republic of Korea
e
School of Environment, Science & Engineering, Southern Cross University, Lismore, NSW 2480, Australia
f
Australian Oil and Gas Services Pty Ltd, Lismore, NSW 2480, Australia

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

Article history: The productivity of a gas well declines over its production life as cannot cover economic policies. To
Received 31 December 2015 overcome such problems, the production performance of gas wells should be predicted by applying
Received in revised form reliable methods to analyse the decline trend. Therefore, reliable models are developed in this study on
22 June 2016
the basis of powerful articial intelligence techniques viz. the articial neural network (ANN) modelling
Accepted 27 June 2016
strategy, least square support vector machine (LSSVM) approach, adaptive neuro-fuzzy inference system
(ANFIS), and decision tree (DT) method for the prediction of cumulative gas production as well as initial
Keywords:
decline rate multiplied by time as a function of the Arps' decline curve exponent and ratio of initial gas
Decline curve
Production
ow rate over total gas ow rate. It was concluded that the results obtained based on the models
Rate developed in current study are in satisfactory agreement with the actual gas well production data.
Model Furthermore, the results of comparative study performed demonstrates that the LSSVM strategy is su-
Gas well perior to the other models investigated for the prediction of both cumulative gas production, and initial
decline rate multiplied by time.
2016 Southwest Petroleum University. Production and hosting by Elsevier B.V. on behalf of KeAi
Communications Co., Ltd. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction forecast the well or reservoir production performance, many

Hydrocarbon reserves estimation and/or production perfor-
mance forecasting is regarded as one of the most challenging tasks
in petroleum engineering disciplines. To this end, a reliable method
called decline curve is used to fit recorded production rate related
to individual and group of wells or reservoirs by a mathematical/
computational function for forecasting the production performance
by conducting the extrapolation of the decline function fitted [1]. To
* Corresponding author. Thermodynamics Research Unit, School of Engineering,
University of KwaZulu-Natal, Howard College Campus, King George V Avenue,
Durban 4041, South Africa.
** Corresponding author. School of Environment, Science & Engineering, Southern
Cross University, Lismore, NSW 2480, Australia.
E-mail addresses: a.h.m@irgcp.fr, amir_h_mohammadi@yahoo.com
(A.H. Mohammadi), alireza.bahadori@scu.edu.au (A. Bahadori).
Peer review under responsibility of Southwest Petroleum University.
models have been proposed by using different approaches. Xu et al.
[2] proposed an analytical model to analyse the production data of
shale gas and tight gas reservoirs. The solutions made by them are
applied for analysing the type curve of production data for both
homogeneous and naturally fractured reservoirs. Furthermore,
they reported a new type curve to evaluate the production per-
formance of shale and tight gas reservoirs. They reported that there
is satisfactory agreement between the results obtained based on
the developed analytical model and the numerical methods to
which it was compared.
Zhang et al. [3] developed a composite method for analysing the
decline performance of a multiple fractured horizontal well in shale
gas reservoirs. They concluded that their developed composite
model can provide some visions into the mechanism of uid ow in
shale gas reservoirs, and help to analyse the decline production
performance of such wells and reservoirs. Fanchi et al. [4] analysed
the probabilistic decline curve workow for modelling the

http://dx.doi.org/10.1016/j.petlm.2016.06.006
2405-6561/ 2016 Southwest Petroleum University. Production and hosting by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Please cite this article in press as: A. Kamari, et al., Decline curve based models for predicting natural gas well performance, Petroleum (2017),
http://dx.doi.org/10.1016/j.petlm.2016.06.006
2 A. Kamari et al. / Petroleum xxx (2017) 1e7
production data of the some shale gas reservoirs. Khanamiri [1]
presented two semi theoretical methods for both oil and gas res-
ervoirs on the basis of time and rate variables. The results indicated
that the oil reservoir model outputs are same with those obtained
by hyperbolic equation, and moreover the gas reservoirs model
outputs are reliable in comparison with the Arp's method [5].
Regarding the importance of predicting the production perfor-
mance of hydrocarbon reservoirs, development of easy-to-use,
reliable, and accurate models is of much important. To this end,
intelligent based techniques show better capability for modelling
nonlinear problems than the traditional methods such as empiri-
cally derived correlations as well as production charts so that this
makes it more attractive for prediction of production well perfor-
mance. In comparison with the intelligent based techniques, the
empirically derived correlations as well as production charts have a
poor performance for prediction targets in the presence of a small
size of dataset because of their high number of adjustable param-
eters. Therefore, the empirically derived correlations as well as
production charts may have the over-tting problem so that they
may lead to error in the outside ranges of data which have been
developed based on [6]. In other words, the empirically derived
correlations and production charts may be applicable for analysing
the performance of just some production wells with specic data
while intelligent based models are powerful tools which can pre-
dict production well performance in any range of data associated
with various geographical locations.
Therefore, application of articial intelligence approaches viz.
articial neural networks, fuzzy logic, decision tree (DT), and least
square support vector machine, etc., is of helpful to propose such
models. Articial intelligence methods have fruitfully been imple-
mented in predicting the thermodynamics of petroleum reservoir
uids, oil production and optimization, enhanced oil recovery, pe-
troleum ow assurance, and reservoir engineering [7e9]. As a
result, the articial neural network (ANN) modelling strategy
optimized with particle swarm optimization (PSO), least square
support vector machine (LSSVM) approach coupled with the
coupled simulated annealing (CSA) optimization method, adaptive
neuro-fuzzy inference system (ANFIS), and nally DT method were
implemented in current study to accurately predict the cumulative
gas production and initial decline rate multiplied by time as a
function of the Arps' decline curve exponent and ratio of initial gas
ow rate over total gas ow rate.
2. Database
As a result, the decline curve model proposed by Arps [5] is still
considered as one of most reliable methods for predicting hydro-
carbon reserves estimation and production performance related to
oil and gas wells as well as reservoirs so that almost all conven-
tional decline curve methods/models are developed on the basis of
Arps' model as follows [10]:
qi
qt
1 bDi t1=b
where qt (MMscf/day) stands for the total gas ow rate at time t, qi
(MMscf/day) denotes the initial gas ow rate, Di (1/day) expresses
the initial decline rate, and b indicates the Arps' decline curve
exponent.
Regarding the matters discussed above, the cumulative gas
production ((GP2-GP1)/(tqi)) and initial decline rate multiplied by
time (Dit) as a function of the Arps' decline curve exponent and
ratio of initial gas ow rate over total gas ow rate have been
Please cite this article in press as: A. Kamari, et al., Decline curve based models for predicting natural gas well performance, Petroleum (2017),
http://dx.doi.org/10.1016/j.petlm.2016.06.006
models based on the data taken from the literature [11,12]. The data
used in this study are taken from decline curve analysis charts
relating production rate to time as well as cumulative production
reported by Gentry [11]. Gentry [11] tested the charts by the results
of production history of a Mississippi limestone well in North-
western Oklahoma with initial production rate (qi) 6292 bbl/
month; production rate at time t (qt) 730 bbl/month; cumulative
production at time t (Qt) 55,960 bbl; ultimate production (Qultimate)
104,000 bbl; and a production time 27 months (time between qi
and qt). Ranges (i.e. maximum, minimum and average) of the cu-
mulative gas production and initial decline rate multiplied by time
as outputs, as well as the Arps' decline curve exponent and ratio of
initial gas ow rate over total gas ow rate as input variables are
listed in Table 1.
3. Methodologies for model development
3.1. Articial neural network e multilayer perceptron (MLP) model
Articial neural network is an advanced model to process and
classify information which simulates the biological neural network
in human brain and is based on the mathematical system of pro-
cesses which happens in brain [13e15]. Multilayer perceptron
(MLP) denotes the advanced type and well researched type of ANNs
in regression and classication, applying a feed forward, supervised
and hetero-associative model [16]. The MLP and many other neural
networks use an algorithm named back-propagation. By using the
back-propagation, the input data is repetitively imported to the
neural network. With each steps that outputs are generated, the
values will be compared to measured outputs and consequently
error is calculated. This error is then provided back (back-propa-
gated) to the ANN and used to regulate the weights such that the
error declines with each iteration and the neural model produces
accurate outputs which are close to desired targets. This procedure
is recognized as training [17].
In artificial neural networks, information handling is executed in
numerous simple separate processors which are entitled neurons.
When inputs are imported to the neuron, they are multiplied by
related weight of each link. Then bias (threshold) of the neuron is
supplied to summation of weighted inputs. In articial neural
network approach, the data from the input neurons are propagated
through the network via weighted interconnections [18]. Every i
neuron in a k layer is connected to every neuron in adjacent layers.
The activation function of exponential sigmoid function which has
generally and traditionally been utilized to develop ANNs [19] as
following:
1
f x (2)
1 ex
where x stands for parameter of activation function. A bias term, b0 ,
is associated with every interconnection in order to introduce a
(1) supplementary degree of freedom. The expression of the weighted
sum, S, to the ith neuron in the kth layer (k  2) is [18].
X
Nk1 h  i
Sk;i wk1;j;i Ik1;j b0 k;i (3)
j1
where w is the weight parameter between each neuroneneuron
interconnection. Using this feed-forward network with activation
function, the output, O, of the i neuron within the hidden k layer is
[18].
 A. Kamari et al. / Petroleum xxx (2017) 1e7 3

Table 1
Ranges of data employed for the prediction of gas well performance using decline curve models developed in this study.

Variable Min. Avg. Max. Type

Arps' decline curve exponent 0 0.5 1 Input

Ratio of initial gas ow rate over total gas ow rate 0.2 3.0 10 Input
Cumulative gas production 0.7 7.8 40 Output
Initial decline rate multiplied by time 0.011 0.3 0.726 Output

1 1 X
n
Ok;i P (4) f x ak Kx; xk b (11)
1 eSk;i
wk1;j;i Ik1;j b0 k;i 
Nk1
 j1 k1
1e
In which K(x,xk) represent the Kernel function, made by inner
product of the vectors x and xk [22,25].

Kx; xk xT $xk (12)

3.2. Least squares support vector machine model
Radial basis function (RBF) is the utmost used relation for
Least squares support vector machines (LSSVM) are least calculating the Kernel function [23,26].
squares forms of support vector machines (SVM), which are a set of  
associated supervised learning methods that investigate data and  k xk  xk2
kx; xk exp (13)
identify patterns, and which are used for sorting and regression 2 s
analysis offered by Suykens et al. [20e22]. In LSSVM a linear
Here s2 is a decision variable. Its optimization controlled by an
approximation is prepared in kernel induced feature space. By
external procedure during model's internal computations. The
considering a data set (xi, yi), i 1, 2, N, with input data xi R and
mean square error (MSE) definition for the LSSVM can be dened as
output data yi R. The regression model can be established as fol-
follows [25,26].
lows [22e24]:
Pn  
T
yk w xk b ek ; K 1; 2; ; N (5) i1 T rep  Texp
MSE pred
(14)
ns
In these equations, w characterizes the linear regression
(regression weight), T is symbolic of the transpose matrix, e is Which T is the wax disappearance temperature, rep/pred and
training items regression error, b represents the model linear exp specify the represented/predicted and experimental data
regression intercept, and shows the feature map. The cost func- respectively and ns is the initial population number [25,26].
tion of LSSVM algorithm, QLSSVM is calculated as below [21e23].
3.3. Regression decision tree (DT) model
1 X
n
QLSSVM wT w g e2k (6)
2 Decision tree constructs models in purpose of regression or
k1
classication with the shape of a tree structure. It divides a dataset
g is the relative weight of the regression errors summation into smaller subsections while a related decision tree is incre-
compared to the regression weight. By assistance of Lagrange mentally developed simultaneously. The outcome is a tree with
function, the regression weight is normally shown as follows decision nodes and leaf nodes. A decision node has two or more
[25,26]. divisions, each signifying values for the feature tested. Leaf node
represents a decision on the target. The top decision node in a tree
X
n
which relates to the best predictor is called root node. Decision
w a k xk (7)
trees can process both categorical and numerical data [27e29].
k1
There are three kinds of decision tree: CRT, CHAID and Exhaustive
In which ak is dened as. CHAID, and Quest. The procedures of the three sorts follow the
following phases. Start tree building by allocating the node to
ak 2gek (8) classes, stopping tree building. Reach the optimum tree selection
With the assumption of linear regression between independent and perform cross-validation. CRT makes tree Pruning before
and dependent LSSVM variables, Eq. (5) can be re-written as creating the optimum tree selection, while CHAID method imple-
[23,25]. ments statistical tests at each step of splitting [30e32].
The CRT (Classication and Regression Tree) is a recursive sub-
X
n dividing technique to be used both for regression and classication.
y ak xTk x b (9) The CRT is built by division of data set into the subgroups using all
k1 predictor variables to create two child nodes repeatedly, beginning
With the subsequent equation the Lagrange multipliers, ak can with the whole data set. The best predictor is chosen using a di-
be considered as [21,22,26]. versity measures. The objective is to yield subdivisions of the data
which are as similar as possible regarding to the target variable
yk  b [33e35]. The CHAID (Chi-Square Automatic Interaction Detector)
ak (10)
technique is established based on the X2-test of association. A
xTk x 2g1
CHAID algorithm is a decision tree that is created by repetitively
By means of Kernel function the rst linear regression equation division subgroups of the space into two or more child nodes. To
will changed into a nonlinear form [24e26]. control the best splitting at any node, any acceptable couple of

Please cite this article in press as: A. Kamari, et al., Decline curve based models for predicting natural gas well performance, Petroleum (2017),
http://dx.doi.org/10.1016/j.petlm.2016.06.006
4 A. Kamari et al. / Petroleum xxx (2017) 1e7
groups of the predictor variables is combined till there is no sta-
tistically important alteration in the couple with respect to the
target variable. This CHAID method logically works with commu-
nications between the independent variables that are directly
accessible from an analysis of the tree. The nal nodes recognize
subsections dened by diverse groups of independent variables
[35,36].
The original CHAID procedure is not assured to discover the best
split of all of those inspected because it uses the last split tested.
The Exhaustive CHAID procedure tries to solve this issue by
continuing to combine groups, regardless of signicance level, until
only two groups stay for each predictor. It then uses the split with
the major importance value rather than the last one tried. The
Exhaustive CHAID needs additional computer time [30,35]. The
QUEST (Quick-Unbiased-Efcient Statistical Tree) is a binary split
decision tree process for classication and data analysis. The QUEST
can be used with univariant or linear grouping splits. An excep-
tional characteristic is that its attribute selection technique has
insignicant bias. If all the attributes are uninformative with
respect to the class attribute, then each has approximately the
identical variation of being selected to split a node time [30,35].
Finally, here it is worthwhile to note that a regression tree is similar
to a classification tree, except that the Y variable takes ordered
values and a regression model is fitted to each node to give the
predicted values of Y [37].
3.4. Adaptive neuro-fuzzy inference system model
The adaptive neuro-fuzzy inference system proposed by Jang in
1993 [38] is viewed as a smart hybrid methodology composed of,
and/or combining both fuzzy logic and articial neural networks.
The production of reasonable results by means of the imple-
mentation of the simple fuzzy inference system (FIS) is directly
associated with expert-knowledge rules. The absence of such rules
is recognized as a major disadvantage of this intelligent system.
The denition of intelligent rules and appropriate membership
functions can help to avoid imperfect results. The ANFIS combines
the strengths of fuzzy logic and articial neural networks in order
to develop models on the basis of an FIS with optimized rules and
Fig. 1. Crossplots of the initial decline rate multiplied by time data versus the predicted values by the developed model based on the LSSVM, DT, ANFIS, and ANN-PSO approaches.
Y axis represents the model results and X axis stands for actual data.
Please cite this article in press as: A. Kamari, et al., Decline curve based models for predicting natural gas well performance, Petroleum (2017),
http://dx.doi.org/10.1016/j.petlm.2016.06.006
membership functions. The ANFIS methodology relies on the
assumption that ANN trains the whole network and FIS structure to
pursue the fuzzy logic representation. It is worth mentioning that
the structures of rules and functions in models developed based on
the ANFIS method are similar to those which have been proposed
by the concept of fuzzy logic [39].
4. Results and discussion
To forecast the cumulative gas production and initial decline
rate multiplied by time, eight reliable models have been developed
on the basis of the ANN, LSSVM, ANFIS, and DT strategies. To this
end, the same variables viz. the Arps' decline curve exponent and
ratio of initial gas ow rate over total gas ow rate have been
considered as inputs to predict the gas well performance. For
visualizing the precision performance and capability of the models
developed based on the methods mentioned above, a reliable error
parameter called average absolute relative deviation (AARD) is
applied as follows:
n X 
1X  exp
 Xrep:=pred 
  100
AARD% (15)
n i1  Xexp 
As e result, the particle swarm optimization [40] strategy was
utilized in this study to tune the adjustable parameters (bias and
weight) related to the feed-forward ANN algorithm. Afterward, the
number of hidden neurons in hidden layer has been optimized by
conducting the try and error method. The most precise models
developed on the basis of ANN strategy were selected with respect
to the AARD percent. Furthermore, the coupled simulated anneal-
ing (CSA) [41] optimization strategy was used to tune the adjust-
able parameters of the LSSVM viz. g and s2. As a consequence, the
values optimized by the CSA method are s2 0.41 and g 634,776
for the model developed for prediction of the cumulative gas pro-
duction, and s2 5.4 and g 1.8E11 for the model proposed for
estimating the initial decline rate multiplied by time. Additionally,
the regression DT algorithm available in the MATLAB was used to
develop a predictive model for comparing the estimated data
values with the other methods investigated in this study. Finally,
 A. Kamari et al. / Petroleum xxx (2017) 1e7 5

Fig. 2. Crossplots of the actual cumulative gas production data versus the predicted values by the developed model based on the LSSVM, DT, ANFIS, and ANN-PSO approaches. Y axis
represents the model results and X axis stands for actual data.

having tried a number of membership functions to develop a reli-
able method on the basis of ANFIS method, the Gaussian mem-
bership provided reasonable results.
To evaluate the accuracy performance of the models developed
in this study, the AARD % was calculated for each model. The cu-
mulative gas production data was predicted with an AARD % of
6.95, 8.95, 14.66, and 30.5% for the LSSVM, DT, ANFIS, and ANN-PSO
models, respectively. Moreover, the models proposed on the basis
of the LSSVM, DT, ANFIS, and PSO-ANN could estimate the initial
decline rate multiplied by time with an AARD % of 2.26, 10.5, 23.62,
and 47.43%, respectively. As clear from the results obtained, the
LSSVM models proposed in current study have highest accuracy in
predicting both cumulative gas production and initial decline rate
multiplied by time. Moreover, the predicted values by the regres-
sion DT model developed in this study are in more agreement with
the actual production data in comparison with the ANFIS and ANN-
PSO models. Figs. 1 and 2 illustrate the parity diagrams of the
predicted values by all methods investigated in this study for the

Fig. 3. Absolute relative deviation contour of the predicted initial decline rate multiplied by time in the different ranges of Arps' decline curve exponent (b) and ratio of initial gas
ow rate over total gas ow rate (qi/qt).

Please cite this article in press as: A. Kamari, et al., Decline curve based models for predicting natural gas well performance, Petroleum (2017),
http://dx.doi.org/10.1016/j.petlm.2016.06.006
6 A. Kamari et al. / Petroleum xxx (2017) 1e7
Fig. 4. Absolute relative deviation contour of the predicted cumulative gas production in the different ranges of Arps' decline curve exponent (b) and ratio of initial gas ow rate
over total gas ow rate (qi/qt).
initial decline rate multiplied by time and cumulative gas produc-
tion data, respectively. As clear from the gures, distribution of the
predicted values by the developed LSSVM models around the unit
slope line is lowest compared to the other models studied. Addi-
tionally, Figs. 3 and 4 displays the absolute relative deviation (ARD)
contour of the predicted gas well production data in the different
ranges of Arps' decline curve exponent and ratio of initial gas ow
rate over total gas ow rate for the initial decline rate multiplied by
time and cumulative gas production data, respectively. Fig. 3
demonstrates that all methods investigated in this study have
better performance in the higher range of b and qi/qt for estimating
the initial decline rate multiplied by time. Additionally, the models
developed in current study (LSSVM, DT, ANFIS, and ANN-PSO) are
not recommended for the prediction of cumulative gas production
in the ranges of 0.7e0.9 for b and 3e5 for qi/qt, as shown in Fig. 4.
5. Conclusions
The challenges faced with the petroleum engineers to estimate
the hydrocarbon reserves, and predict the production performance
of oil and gas wells were considered in this study with develop-
ment of reliable models for prediction of the cumulative gas pro-
duction as well as the initial decline rate multiplied by time. To this
end, the Arps' decline curve exponent and ratio of initial gas ow
rate over total gas ow rate have been considered as inputs of the
models. Reliable articial intelligence techniques viz. the regression
decision tree, least squares support vector machine, adaptive
neuro-fuzzy inference system, and neural network were employed
to analyse the gas wells performance. It was concluded that the
results obtained based on the models developed in current study
are in satisfactory agreement with the actual gas well production
data. Finally, a comparative study was performed between the all
models developed in current study and the actual gas well pro-
duction data. It was found that the LSSVM strategy is superior to the
Please cite this article in press as: A. Kamari, et al., Decline curve based models for predicting natural gas well performance, Petroleum (2017),
http://dx.doi.org/10.1016/j.petlm.2016.06.006
other models investigated for the prediction of both cumulative gas
production, and initial decline rate multiplied by time.
References
[1] H.H. Khanamiri, A non-iterative method of decline curve analysis, J. Pet. Sci.
Eng. 73 (2010) 59e66.
[2] B. Xu, M. Haghighi, X. Li, D. Cooke, Development of new type curves for
production analysis in naturally fractured shale gas/tight gas reservoirs, J. Pet.
Sci. Eng. 105 (2013) 107e115.
[3] D. Zhang, L. Zhang, Y. Zhao, J. Guo, A composite model to analyze the decline
performance of a multiple fractured horizontal well in shale reservoirs, J. Nat.
Gas Sci. Eng. 26 (2015) 999e1010.
[4] J. Fanchi, M. Cooksey, K. Lehman, A. Smith, A. Fanchi, C. Fanchi, Probabilistic
decline curve analysis of Barnett, Fayetteville, Haynesville, and Woodford gas
shales, J. Pet. Sci. Eng. 109 (2013) 308e311.
[5] J. Arps, Analysis of decline curves, Trans. AIME 160 (1945) 228e247.
[6] A. Kamari, A. Bahadori, A.H. Mohammadi, On the determination of crude oil
salt content: application of robust modeling approaches, J. Taiwan Inst. Chem.
Eng. 55 (2015) 27e35.
[7] A. Kamari, M. Arabloo, A. Shokrollahi, F. Gharagheizi, A.H. Mohammadi, Rapid
method to estimate the minimum miscibility pressure (MMP) in live reservoir
oil systems during CO2 ooding, Fuel 153 (2015) 310e319.
[8] A. Kamari, A. Bahadori, A.H. Mohammadi, S. Zendehboudi, Evaluating the
unloading gradient pressure in continuous gas-lift systems during petroleum
production operations, Pet. Sci. Technol. 32 (2014) 2961e2968.
[9] A. Kamari, A.H. Mohammadi, A. Bahadori, S. Zendehboudi, Prediction of air
specic heat ratios at elevated pressures using a novel modeling approach,
Chem. Eng. Technol. 37 (2014) 2047e2055.
[10] A. Bahadori, Analysing gas well production data using a simplied decline
curve analysis method, Chem. Eng. Res. Des. 90 (2012) 541e547.
[11] R.W. Gentry, Decline-curve analysis, J. Pet. Technol. 24 (1972) 38e41.
[12] T. Ahmed, Reservoir Engineering Handbook, Access Online via Elsevier, 2006.
[13] R. Zabihi, M. Schafe, H. Nezamabadi-Pour, M. Ranjbar, Articial neural
network for permeability damage prediction due to sulfate scaling, J. Pet. Sci.
Eng. 78 (2011) 575e581.
[14] P.S. Hegeman, C. Dong, N. Varotsis, V. Gaganis, Application of articial neural
networks to downhole uid analysis, in: International Petroleum Technology
Conference, 2007.
[15] N. Al-Bulushi, M. Araujo, M. Kraaijveld, Predicting water saturation using
articial neural networks (ANNS), Neural Netw. 549 (2007) 57.
[16] M. Raq, G. Bugmann, D. Easterbrook, Neural network design for engineering
applications, Comput. Struct. 79 (2001) 1541e1552.
 A. Kamari et al. / Petroleum xxx (2017) 1e7
[17] G. Arulampalam, A. Bouzerdoum, A generalized feedforward neural network
architecture for classication and regression, Neural Netw. 16 (2003)
561e568.
[18] A.H. Mohammadi, D. Richon, Hydrate phase equilibria for hydrogen water
and hydrogen tetrahydrofuran water systems: predictions of dissociation
conditions using an articial neural network algorithm, Chem. Eng. Sci. 65
(2010) 3352e3355.
[19] A. Eslamimanesh, F. Gharagheizi, A.H. Mohammadi, D. Richon, Articial neural
network modeling of solubility of supercritical carbon dioxide in 24
commonly used ionic liquids, Chem. Eng. Sci. 66 (2011) 3039e3044.
[20] H. Yarveicy, A.K. Moghaddam, M.M. Ghiasi, Practical use of statistical learning
theory for modeling freezing point depression of electrolyte solutions: LSSVM
model, J. Nat. Gas Sci. Eng. 20 (2014) 414e421.
[21] P. Samui, D. Kothari, Utilization of a least square support vector machine
(LSSVM) for slope stability analysis, Sci. Iran. 18 (2011) 53e58.
[22] J.A. Suykens, J. Vandewalle, Least squares support vector machine classiers,
Neural Process. Lett. 9 (1999) 293e300.
[23] B.B. Ekici, A least squares support vector machine model for prediction of the
next day solar insolation for effective use of PV systems, Measurement 50
(2014) 255e262.
[24] L.-Z. Ding, S. Liao, Approximate model selection for large scale LSSVM, in:
ACML, 2011, pp. 165e180.
[25] H. Safari, A. Shokrollahi, M. Jamialahmadi, M.H. Ghazanfari, A. Bahadori,
S. Zendehboudi, Prediction of the aqueous solubility of BaSO4 using pitzer ion
interaction model and LSSVM algorithm, Fluid Phase Equilib 374 (2014)
48e62.
[26] M. Mesbah, E. Soroush, V. Azari, M. Lee, A. Bahadori, S. Habibnia, Vapor liquid
equilibrium prediction of carbon dioxide and hydrocarbon systems using
LSSVM algorithm, J. Supercrit. Fluids 97 (2015) 256e267.
[27] M. Erdogan, B. Mudford, G. Chenoweth, R. Holeywell, J. Jakubson, Optimiza-
tion of decision tree and simulation portfolios: a comparison, in: SPE Hy-
drocarbon Economics and Evaluation Symposium, Society of Petroleum
Engineers, 2001.
Please cite this article in press as: A. Kamari, et al., Decline curve based models for predicting natural gas well performance, Petroleum (2017),
http://dx.doi.org/10.1016/j.petlm.2016.06.006
7
[28] I.K. Sethi, B. Chatterjee, Efcient decision tree design for discrete variable
pattern recognition problems, Pattern Recognit. 9 (1977) 197e206.
[29] L.R. Heinze, H.W. Winkler, J.F. Lea, Decision tree for selection of articial lift
method, in: SPE Production Operations Symposium, Society of Petroleum
Engineers, 1995.
[30] M.S. Alkhasawneh, U.K. Ngah, L.T. Tay, N.A. Mat Isa, M.S. Al-Batah, Modeling
and testing landslide hazard using decision tree, J. Appl. Math. 2014 (2014).
[31] B. Chandra, R. Kothari, P. Paul, A new node splitting measure for decision tree
construction, Pattern Recognit. 43 (2010) 2725e2731.
[32] K.-M. Osei-Bryson, Evaluation of decision trees: a multi-criteria approach,
Comput. Op. Res. 31 (2004) 1933e1945.
[33] X. Wang, X. Liu, W. Pedrycz, L. Zhang, Fuzzy rule based decision trees, Pattern
Recognit. 48 (2015) 50e59.
[34] D.G. Laughton, G. Joe, M. Paduada, M. Samis, Complete decision-tree analysis
using simulation methods: illustrated with an example of Bitumen production
in alberta using steam injection, in: SPE Annual Technical Conference and
Exhibition, Society of Petroleum Engineers, 2006.
[35] P. Tan, M. Steinbach, V. Kumar, Introduction to Data Mining. Classication:
Basic Concepts, Decision Trees and Model Evaluation, AddisoneWesley, Bos-
ton, 2006.
[36] A.H. Gandomi, M.M. Fridline, D.A. Roke, Decision tree approach for soil
liquefaction assessment, Sci. World J. 2013 (2013).
[37] W.Y. Loh, Classication and regression trees, Wiley Interdiscip. Rev. Data Min.
Knowl. Discov. 1 (2011) 14e23.
[38] J.-S.R. Jang, ANFIS: adaptive-network-based fuzzy inference system, systems,
man and cybernetics, IEEE Trans. 23 (1993) 665e685.
[39] I. Rahimzadeh Kivi, M. Ameri Shahrabi, M. Akbari, The development of a
robust ANFIS model for predicting minimum miscibility pressure, Pet. Sci.
Technol. 31 (2013) 2039e2046.
[40] J. Kennedy, Particle swarm optimization, in: Encyclopedia of Machine
Learning, Springer, 2010, pp. 760e766.
[41] V. Fabian, Simulated annealing simulated, Comput. Math. Appl. 33 (1997)
81e94.