Fuel 215 (2018) 904–914

Contents lists available at ScienceDirect

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

Full Length Article

Developing a robust proxy model of CO2 injection: Coupling Box–Behnken T

design and a connectionist method
⁎ ⁎
Mohammad Ali Ahmadi , Sohrab Zendehboudi , Lesley A. James
Faculty of Engineering and Applied Science, Memorial University, St. John’s, NL, Canada

G RA P H I C A L AB S T R A C T

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

Keywords: The CO2 based enhanced oil recovery methods (EORs) in the petroleum industry are considered as one of the
CO2 based EOR efficient technologies for further production where the natural driving forces become weak. To determine which
Simulation strategy EOR method is more appropriate for a targeted reservoir, there is a need to develop a reliable and fast tool to
Genetic algorithm predict the performance of the EOR methods due to assumptions and central processing time (CPU) time of
Box–Behnken design
reservoir simulations. We develop a promising approach for predicting the ultimate oil recovery factor of the
Least square support vector machine (LSSVM)
miscible CO2 injection processes. To attain this goal, the least square support vector machine is used to build the
proxy model. The Box-Behnken design as a branch of response surface methods is employed to design simulation
runs for miscible CO2 injection processes, and the leverage method is applied to validate the proxy model in
terms of statistical perspective. An artificial heterogeneous reservoir is used to perform compositional reservoir
simulations. Five operational parameters of the miscible CO2 injection process are considered, including bottom-
hole flowing pressure (BHP) of injection well (psi), CO2 injection rate (MMSCF/D), injected CO2 concentration
(mole fraction), bottom-hole flowing pressure (BHP) of production well (psi), and oil production rate (STB/D).
The developed proxy model can be employed to forecast the ultimate oil recovery factor of the miscible CO2
injection operations at the different rock, fluids, and process conditions. The proposed method appears to be an
efficient simulation strategy that offers guidelines and screening criteria for the application of the miscible CO2
injection.

⁎
Corresponding authors.
E-mail addresses: [email protected] (M.A. Ahmadi), [email protected] (S. Zendehboudi).

https://doi.org/10.1016/j.fuel.2017.11.030
Received 20 July 2017; Received in revised form 6 November 2017; Accepted 9 November 2017
Available online 22 December 2017
0016-2361/ © 2017 Elsevier Ltd. All rights reserved.
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

Nomenclature RSM response surface method

SCF standard cubic feet
Abbreviations SLT statistical learning theory
SR standardized residual
AARD average absolute relative deviation STB standard barrel
ARD average relative deviation SVM support vector machine
BBD Box-Behnken design SVR support vector regression
BHP bottom hole pressure WAG water alternative gas
CCD central composite design WOC water oil contact
CF cost function
CMG computer modeling group Variables
CPU central processing unit
EOR enhanced oil recovery αk lagrange multipliers
EOS equation of state b a term of bias
GA genetic algorithm ek the regression error
IFT interfacial tension γ the regularization parameter
LM levenberg–Marquardt K kernel function
LSSVM least squares support vector machine n space’s dimension
MM million ns number of data samples
MSE mean squared error σ2 kernel sample variance
OOIP original oil in place w the vector of weight
QP quadratic programming xk kth input data in the input space
RBF radial basis function yk output value for a specified input variable (e.g., xk)
RF recovery factor

1. Introduction performance of miscible surfactant-CO2 flooding. Their proposed model

was built on hybridization of ant colony and support vector regression
Nowadays, the main source of energy is fossil fuels which are de- (SVR) method. They concluded that the SVR model is able to forecast
posited in the oil and gas reservoirs. Most of the oil reservoirs are ap- the performance of surfactant- CO2 flooding with a high degree of re-
proaching the end of their primary production lives. However, around liability and precision.
70% of the original oil in place (OOIP) remain in the geological for- Jaber et al. [27] proposed a proxy model to determine the perfor-
mations after primary production stage. To produce the remaining oil mance of the CO2-WAG (water alternative gas) injection for a hetero-
from the depleted reservoirs, enhanced oil recovery (EOR) techniques geneous clastic reservoir. They employed a Box-Behnken design method
should be implemented through effective engineering strategies in to build their proxy model. They considered four parameters (e.g.,
terms of theoretical and practical aspects [1–5]. controllable variables) including the ratio of CO2 slug size to water slug
To dynamically evaluate the performance of any EOR scenario (e.g., size, CO2 slug size, bottom hole pressure, and cyclic length. According
water flooding, CO2 injection, and chemical flooding) and to under- to their results, the developed model can be used at different levels of
stand the contributions of oil production mechanisms (e.g., interfacial operational parameters to reasonably estimate the incremental oil re-
tension (IFT) reduction, oil swelling, and oil viscosity reduction) to covery over the miscible CO2-WAG flooding processes.
fluids displacement, the reservoir simulation studies seem paramount to The main objective of this research work is to avoid performing
be conducted [5–12]. One of the promising EOR methods is gas injec- reservoir simulation runs, which are costly and time-consuming, by
tion where the injection fluid is a gas, such as N2, CO2, associated gas, introducing a simpler and valid approach. This paper is planned to
flue gas, and air. Among the gas injection methods, CO2 injection is not develop a promising proxy model for prediction of the ultimate oil re-
only an efficient EOR method, but it also provides a solution for re- covery achievable through miscible CO2 injection. To design reservoir
ducing emissions of greenhouse gases through injecting CO2 into de- simulations versus the operational parameters, the response surface
pleted oil reservoirs and aquifers [9,13,14]. method (RSM) is employed. Least square support vector machine
Various parameters are contributing to the oil production and oil (LSSVM) as a subset of connectionist models is used to develop the
sweep efficiency during CO2 injection [13,15,16]. There is no robust, proxy model for obtaining the target function. To statistically evaluate
fast, and easy-to-use method to determine the performance of miscible the applicability of the proxy model, the leverage method is im-
CO2 injection into a given oil reservoir. In addition, it is challenging to plemented.
categorize/screen the candidate reservoirs for miscible CO2 injection
operations [17–24].
2. Methodology
Several scholars made attempts to introduce dimensionless numbers
to consider different oil production mechanisms. For instance, Wood
2.1. Characterization of the reservoir model
et al. [25] proposed some dimensionless numbers such as the dip angle
group, effective aspect ratio, buoyancy number, and CO2-oil mobility
A synthetic oil reservoir was used to simulate the miscible CO2 in-
ratio to select appropriate candidates for CO2 injection. They neglected
jection process. The reservoir properties including initial oil saturation,
the impact of reservoir heterogeneity in their research work, while the
porosity, permeability, and pressure are depicted in Fig. 1. The initial
performance of CO2 injection (both miscible and immiscible) is con-
oil saturation distribution of the reservoir is shown in Fig. 1(a).
siderably affected by reservoir heterogeneity. Hence, developing a
Fig. 1(b) and (c) illustrate the distribution of porosity and permeability
straightforward and robust strategy such as a proxy model for pre-
of the reservoir, respectively. Permeability is assumed to be the same in
dicting and evaluating the performance of the miscible CO2 injection is
x, y and z directions. Fig. 1(d) shows the reservoir pressure variation
of great interest to the petroleum industry.
with depth and reservoir layers. The grid depth ranges from 6072.83 to
Helaleh and Alizadeh [26] developed a proxy model to forecast the
6258.87 ft and the reference pressure is considered to be 3932.3 psi at a

905
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

Fig. 1. 3D view of distribution for the a) oil saturation

b) porosity c) permeability d) reservoir pressure of the
synthetic reservoir used in this study.

depth of 6165 ft. Water-oil-contact (WOC) is set at 6200.87 ft. Initially, aquifer supports the reservoir over the production [4]. This aquifer has
the reservoir pressure is above the bubble point pressure since the in- a thickness of 60 ft, a porosity of 0.25, a permeability of 1.65 md and
itial gas saturation in the reservoir is zero. An infinite acting bottom radius of 518.22 ft. Two wells are drilled for the production and

906
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

injection. Both wells start operation from January 1st, 1901. Our con- where ɑ and b are defined below [37–45]:
trolling parameters in simulating miscible CO2 injection are bottom- −1
hole flowing pressure (BHP) of the injection well (psi), CO2 injection
b=
( )
1Tn Ω + γ In
1
y
rate (MMSCF/D), injected CO2 concentration (mole fraction), bottom- −1
1 (Ω + I )
T 1
hole flowing pressure (BHP) of the production well (psi), and the oil n γ n
1n
(8)
production rate (STB/D). The oil reservoir under CO2 injection was
−1
simulated for 35 years (1901–1935). For this goal, GEM package (as 1
α = ⎜⎛Ω + In⎟⎞ (y−1n b)
compositional reservoir simulator engine) of the computer modeling ⎝ γ ⎠ (9)
group (CMG) reservoir simulator 2016.1© was used. It is worth to
Eq. (10) uses a nonlinear regression through incorporating the
highlight that Peng-Robinson equation of state (EOS) was used in CMG
Kernel function K [37–45]:
package for reservoir fluid behaviour modeling purposes.
N

2.2. Least square support vector machine (LSSVM) f (x ) = ∑ αk K (x ,xk ) + b

k=1 (10)

The least square SVM theorem was proposed and developed by while K(x,xk) is the Kernel function which is related to the transfer
Suykens and Vandewalle (1999), based on the idea that the data sam- functions (to feature space) Ф(x) and Ф(xk) shown below [37–45]:
ples S = {(x1,y1),…,(xn,yn)} with a nonlinear trend can be formulated K (x ,xk ) = Φ(x )T Φ(xk ) (11)
as shown in Eq. (1). In Eq. (1), w stands for the weight factor, φ denotes
the nonlinear function which correlates the input space to a high-di- We employ the radial basis function (RBF) Kernel in this research
mension characterization area and conducts linear regression, and b work [37–41]:
represents the bias term [28–32]. The following expression was im- K (x ,xk ) = exp(−‖xk −x‖2 / σ 2) (12)
plemented as a cost function of the LSSVM in calculation steps [33–41].
where σ is the variance of the distribution and it is the only parameter
2
y = wT . φ (x ) + b with w ∈ ,  ∈ , φ (·) ∈  → ,  → ∞ to be tuned by GA. To obtain the optimal parameters of LSSVM, we use
(1) mean squared error (MSE) as the objective function to be minimized
Which is constrained as [37–45]: [32,45]:
n
yk = wT ϕ (xk ) + b + ek k = 1,2,…,N (2)
∑ (RFest .i−RFexp .i)2
i=1
For the function estimation, the structural risk minimization (SRM) MSE =
ns (13)
is suggested; the optimization objective function is shown with J below
in which γ is the regularization constant, and ek refers to the regression where, RF represents the recovery factor, subscripts est. and exp. re-
error [33–42]. present the predicted and actual recovery factor, respectively, and ns
N stands for the number of data from the initially assigned population.
1 T 1
J (w,e ) = w w + γ ∑ ek2
2 2 k=1 (3) 2.3. Genetic Algorithm (GA)
To obtain w and e, the Lagrange multiplier optimum programming
approach is performed to solve Eq. (3); the employed approach con- Genetic Algorithm (GA) as one of the best optimization methods
siders impartial and restriction parameters simultaneously. The men- which is attributed to its unique features which are searching quickly
tioned Lagrange function L is formulated by the following relationship and optimizing efficiently; the two essential characteristics which have
[33–45]: been derived from the principle of “survival of the fittest” element of
natural evolution with the genetic propagation of properties. In more
L (w,b,e,α ) = J (w,e )−Σm T
k = 1αi {w Ø(xk ) + b + ek−Yk } (4) details, GA operates through clarifying a variety of zones in the target
In Eq. (4), αi denotes the Lagrange multipliers that might be positive area determined by experts and defining simultaneously and randomly
or negative as the LSSVM has equality restrictions. Using Karush a large number of possible paths [32,35]. The GA has this capability of
Kuhn–Tucher’s (KKT) conditions, optimum solution for Eq. (4) is given being replaced with classic optimization techniques. This approach is
below [37–45]. based on the idea of Darwinian natural selection and genetics in bio-
logical systems. According to the concept of ‘survival of the fittest’, the
n
⎧ ∂ L = w− ⎫ GA can converge towards the best point in the prepared space soon
⎪ ω ∑ αi φ (xi) = 0 ⎪ after a series of repetitive calculations. Foundations of this searching
⎪ i = 1 ⎪
⎪ n
⎪ process are based on technical operations such as artificial mutation,
∂b L = ∑ αi = 0 crossover, and selection [32,35,38,45]. To run the above algorithm, it is
⎨ ⎬
⎪ i=1
⎪ preliminarily required to prepare an initial population containing a
⎪ ∂ei L = Cei−αi = 0 ⎪ particular number of so-called individuals which are representing the
⎪ ∂ L = (wT Ø(x ) + b + e −y = 0 ⎪
⎩ iα k k k ⎭ (5) possible paths toward the ideal goal. The next step which is vital to be
taken is turning each chromosome, already introduced under the title of
The linear set of equations can be demonstrated as [37–45]:
an individual, into an encoded string. After that, each string must show
T its suitability with the nature of the problem through becoming in-
⎡0 − 1 ⎤ b 0
⎢1 Ω + 1 IN ⎥ ⎡ α ⎤ = ⎡ y ⎤ troduced into the fitness function. Subsequently, the output of fitness
⎢ γ ⎥⎣ ⎦ ⎣ ⎥
⎢ ⎥ ⎢ ⎦ (6) function related to each chromosome is taken as a criterion to make a
⎣ ⎦
decision if the related string can provide a satisfactory performance.
where, y = (y1,…,yn) , In = (1,…,1) , α = (α1,…,αn) and Ωil = φ (xi)T
T T T
After removing a number of the weakest individuals which are de-
φ (xl) for i, l = 1,…,n. Using Mercer’s theorem, the resultant LSSVM termined by the designer, it is the turn to operate crossover and mu-
model for function approximation is as follows [37–45]: tation rates to produce new individuals with higher performance. Then,
N implementation of the crossover operation on the couple of chosen
f (x ) = ∑ αk K (x ,xk ) + b strings (chromosomes) to recombine them has to be followed. It has
k=1 (7)
been suggested by the previous studies that the best performance of the

907
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

GA is attainable when the crossover point of any two chromosomes is Fig. 5 illustrates the recovery factor determined by CMG software
randomly set. The process is followed by switching some randomly for miscible CO2 injection at various BHP of the production well. As it is
selected positions to 1 if they are 0, and vice versa. The last described clear from Fig. 5, the recovery factor drops from around 65% in case of
step is named mutation which is run to prevent the procedure to be BHP = 500 psi to almost 45% when BHP is equal to 2000 psi. Fig. 5
trapped in any local maxima. The final step is to return the generated reveals that the recovery factor is a strong function of the flowing
off-springs to the first step during the next population to be evaluated bottom-hole pressure of production well.
again [32,35,38,45]. Fig. 2 depicts the schematic of the hyperpara- Fig. 6 demonstrates the CMG simulation results for different mag-
meters optimization using the genetic algorithm. nitudes of production rate under miscible CO2 injection operation. As
illustrated in Fig. 6, no certain trend is observed between recovery
2.4. Proxy model development factor and the related oil production rate. This is mainly due to various
parameters contributed in oil production. For example, when oil pro-
Proxy methods are popular techniques for CPU-time reduction in duction rate is equal to 10,000 STB/D (standard barrel per day), the
population-based optimizers, in which the cost function (CF) is replaced ultimate recovery factor might be close to 42% or 62%. It means that
by a function, called proxy [46], and the proxy is employed to assess all for developing a proxy model to predict ultimate oil recovery factor of
or some of the individuals in the optimization process [47]. In other miscible CO2 injection, the linear regression methods do not work
words, proxy model is a representative behavior function of the whole properly.
reservoir model; such a model is used instead of full field reservoir si- Fig. 7 shows the variation of the ultimate oil recovery factor of
mulation to save time and money. However, the precision and suit- miscible CO2 injection with CO2 injection rate. As depicted in Fig. 7, the
ability of the proxy models highly depend on the reliability of the re- ultimate oil recovery factor varies with CO2 injection rate; however, it
servoir simulations. A proxy is trained by a set of samples taken from is not easy to obtain a specific mathematical relationship between the
the original fitness function landscape. Usually, a large number of recovery factor and injection rate through common statistical strate-
training data is needed to build an accurate proxy model that gives an gies.
acceptable approximation of the global optimum point of the CF The values of the global optimum for hyperparameters of the proxy
[48–52]. Although imperfect proxy models might not have the cap- model including σ2 and γ were determined for predicting the recovery
abilities to approximate the global optimum, they can provide an factor (RF) of miscible CO2 injection as 1.687654 and 27.578421,
overview of the entire fitness function landscape and a good estimation correspondingly.
of the sample regions. Referring to this fact, a new-generation of proxy Fig. 8 depicts the scatter plot of the outputs of the proxy model
models is proposed in which the CF is carried out in conjunction with versus the corresponding recovery factor obtained from CMG software.
the proxy for the fitness evaluation of the individuals. This method is As shown in Fig. 8(a), the recovery factor values predicted by the
effectively implemented in different disciplines [48,52,53]. Several proposed proxy model versus the CMG simulation results fall on the
techniques were proposed to enhance proxy-modeling, by applying straight line which is very close to Y = X line. One of the statistical
different sampling strategies and various types of proxy [52–54]. For indexes for evaluating the performance of the proposed proxy model in
instance, Silva et al. [55], Cullick et al. [46] and Sampaio et al. [56] this study is the correlation coefficient of the regression plot. As
employed an artificial neural network, as the proxy model, and attained
acceptable outcomes. In this study, LSSVM method as a promising
connectionist approach is utilized to develop a new generation of the
proxy model. Fig. 3 illustrates the schematic of the proxy model de-
velopment strategy. As noted previously, our influential variables for
simulating CO2 injection into the reservoir are BHP of the injection well
(psi), CO2 injection rate (MMSCF/D), injected CO2 concentration (mole
fraction), BHP of production well (psi), and oil production rate (STB/
D). Hence, it is required to define an acceptable and reasonable range
for the parameters listed above. Table 1 reports the ranges of the proxy
model input parameters. There are various methods for designing the
simulation runs such as 2-level full factorial, 2-level partial factorial,
and response surface methodology (RSM). RSM includes two main ca-
tegories; namely Box–Behnken design (BBD) and central composite
design (CCD). In this paper, Box–Behnken method is employed to de-
sign our CO2 injection scenarios. Table 2 reports the different scenarios
designed for CO2 injection using Box–Behnken method and their proxy
roles. As tabulated in Table 2, 37 simulation runs were conducted to
construct the proxy model. 9 simulation runs were also employed to
validate the proposed proxy model.

3. Results and discussion

3.1. Proxy model

Fig. 4 depicts the recovery factor determined by CMG software for

miscible CO2 injection versus the corresponding BHP of the injection
well. As illustrated in Fig. 4, the recovery factor significantly changes
from about 66% as the maximum recovery factor to 38% as the
minimum recovery factor. Fig. 4 shows that the recovery factor con-
siderably varies with the flowing bottom-hole pressure of injection well;
however, there is no linear relationship between the recovery factor
Fig. 2. Schematic of the hyper-parameters optimization using the genetic algorithm.
and flowing bottom-hole pressure of the injection well.

908
 M.A. Ahmadi et al. Fuel 215 (2018) 904–914

Fig. 3. Schematic of the proxy model development

Define Input Variables strategy.

Sensitivity Analysis

Input Dataset Sampling

Proxy-Model Estimation
Input Dataset Improvement

Proxy-Model Verification

Yes Proxy-Model
Proxy-Model Quality is
Sufficient Employment

NO

Table 1 model and CMG simulation runs is demonstrated in Fig. 10, where
Ranges of the proxy model input parameters. various oil production rates are examined in the CO2 injection process.
As it is clear, the maximum relative error for the testing phase corre-
Parameter Unit Min Max
sponds to the oil production rate of 7745 STB/D (standard barrel per
CO2 injection rate MMSCF/D 1000000 10000000 day), and the maximum relative error for the training phase occurs
Maximum bottom-hole pressure of psi 1500 7500 when the oil production rate is 9100 STB/D.
injection well
Fig. 11 depicts the relative deviation of the outputs of the proxy
Minimum bottom-hole pressure of psi 200 2000
injection well
model from the recovery factor of miscible CO2 injection achieved from
Oil production rate STB/D 1000 10000 CMG software versus corresponding values of the BHP of the injection
CO2 concentration Mole fraction 0.8 1 well. As shown in Fig. 11, the maximum relative deviation for the
testing data points is experienced when BHP of the injection well is
almost 3700 psi. Also, the maximum relative error for the training data
illustrated in Fig. 8, the best fitted straight line has a high correlation samples is noticed when BHP of the injection well is equal to 3900 psi.
coefficient (R2) which is equal to 0.9816. It confirms that the proxy Fig. 12 demonstrates the relative deviation of the proxy RF values from
model was trained adequately to forecast the ultimate oil recovery of the recovery factor of miscible CO2 injection obtained from CMG soft-
the miscible CO2 injection process. Fig. 8(b) represents the regression ware versus corresponding values of the BHP of production well. As
plot between the simulation results and the RF values estimated by the demonstrated in Fig. 12, the maximum relative error for the training
developed proxy model. As clear from this comparison, there is a better and testing stages occurs at the BHP of the production well equal to
agreement between the proxy model and CMG simulator outputs due to 1100 psi and 503 psi, respectively.
a very low number of data used in the testing stage, compared to the Table 3 reports the simulation results obtained from GEM package
training phase. Fig. 8(c) demonstrates the scatter plot of the recovery of CMG software, recovery factor predicted by the developed proxy
factor values determined using the proxy model and commercial si- model, residual error, mean squared error, and average relative de-
mulator based on the entire data samples. The high correlation coeffi- viation (ARD). According to Table 3, the minimum residual value is
cient of the fitted regression line reveals the promising effectiveness of −2.31396 and the maximum residual value is +2.2643. Also, the
the developed proxy model. maximum MSE value is equal to 5.3544, and the minimum MSE is
Fig. 9 shows the relative deviation of the outputs of the proxy model 0.001373. It implies that the developed proxy model offers a precise
in terms of recovery factor of miscible CO2 injection with respect to the and reliable tool for determining ultimate recovery factor of the mis-
CMG software results at different CO2 injection rates for both testing cible CO2 injection processes.
and training data samples. As depicted in Fig. 9, the maximum relative Table 4 reports the indexes for performance evaluation of the proxy
error for the training data samples belongs to the CO2 injection rate model proposed in this study. These indexes are the correlation coef-
interval of 4 × 106 to 6 × 106 MMSCF/D (million standard cubic feet ficient (R2), mean squared error (MSE), and average absolute relative
per day). The maximum relative error for the testing data points is deviation (AARD). As listed in Table 4, the proxy model introduced in
+3.54% which occurs at an injection rate of 4 × 106 MMSCF/D. Ac- this paper provides promising results from a statistical viewpoint. High
cording to Fig. 9, the relative error for both training and testing data correlation coefficient values and very low magnitudes of MSE and
samples falls between ± 5% lines. AARD confirm the outstanding efficiency of the developed proxy model
A relative comparison of oil recovery factor obtained from the proxy in forecasting the performance of the miscible CO2 injection processes.

909
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

Table 2
Box–Behnken designs for CO2 injection.

ID Proxy role Injector BHP (psi) CO2 mole fraction CO2 injection rate (MMSCF/D) Producer BHP (psi) Oil production rate (STB/D) RF (%OOIP)

1 Training 3900 0.895 3700000 1640 7300 41.787663

2 Training 5700 0.88 2800000 740 6400 55.878799
3 Training 2100 0.97 7300000 2000 2800 40.404835
4 Training 1500 1 1000000 1820 7300 38.104408
5 Training 5700 0.925 7300000 1640 8200 43.034195
6 Training 2700 0.925 3700000 920 1000 53.062904
7 Training 7500 0.91 1900000 920 6400 54.576023
8 Training 2700 1 3700000 740 1900 55.891567
9 Training 7500 0.85 8200000 1100 2800 47.592541
10 Training 5100 0.97 5500000 1820 4600 42.57518
11 Training 3900 0.85 1000000 1820 3700 40.414654
12 Training 5100 0.955 9100000 1280 8200 45.362228
13 Training 1500 0.85 9100000 1100 6400 47.569981
14 Training 6300 0.88 6400000 1460 5500 43.288555
15 Training 5700 0.925 9100000 1460 1000 43.350037
16 Training 4500 0.985 10000000 200 7300 61.073956
17 Training 2100 0.865 4600000 560 8200 63.530128
18 Training 6900 0.91 1000000 740 10000 53.268475
19 Training 5100 0.91 7300000 2000 3700 42.159241
20 Training 3900 0.88 10000000 1100 9100 47.142193
21 Training 6900 0.985 8200000 2000 4600 43.338001
22 Training 2700 0.895 2800000 1460 1900 42.260117
23 Training 3300 0.94 6400000 1280 10000 45.047085
24 Training 1500 0.895 10000000 560 5500 59.897743
25 Training 2100 1 1900000 920 9100 50.524914
26 Training 1500 0.94 6400000 1820 3700 38.096645
27 Training 2700 0.91 3700000 1100 7300 47.234138
28 Training 3300 0.91 6400000 740 3700 56.332222
29 Training 3900 0.88 8200000 380 10000 60.53804
30 Training 6900 0.955 7300000 920 7300 52.912228
31 Training 5100 0.925 9100000 1280 2800 45.21104
32 Training 6900 0.85 4600000 560 9100 64.277306
33 Training 6300 0.895 5500000 1820 5500 42.303593
34 Training 3900 0.865 7300000 560 6400 63.738056
35 Training 2100 0.88 3700000 1100 3700 47.920063
36 Training 3900 0.925 6400000 1100 3700 46.46006
37 Training 3300 1 2800000 560 7300 60.874733
38 Verification 6300 0.925 7300000 560 8200 65.921333
39 Verification 3300 0.85 4600000 920 5500 53.176949
40 Verification 2100 0.97 9100000 1280 5500 46.097
41 Verification 6300 1 4600000 1640 5500 42.837906
42 Verification 3300 0.88 3700000 740 9100 60.084507
43 Verification 3678.7058 0.98058705 3944118.5 503.8387 7745.4823 63.805046
44 Verification 6241.9439 0.91739496 2720340 1950.3223 9582.3154 41.126812
45 Verification 4902.9465 0.97855853 6192838.4 1617.1389 7192.1755 42.910915
46 Verification 5408.7404 0.89545431 6916595.7 943.61662 2932.1513 51.572277

70 70

60 60
Recovery Factor (%OOIP)

Recovery Factor (%OOIP)

50 50

40 40

30
30

20
20

10
10

0
0 1000 2000 3000 4000 5000 6000 7000 8000 0
0 500 1000 1500 2000 2500
Injection Well Bottom-Hole Pressure (psi)
Production Well Bottom-Hole Pressure (psi)
Fig. 4. Recovery factor of miscible CO2 injection versus the corresponding BHP of in-
jection well. Fig. 5. Recovery factor of miscible CO2 injection versus the corresponding BHP of pro-
duction well.

3.2. Validity of the proxy model

and to assess the quality of the simulation results for miscible CO2 in-
The Leverage method was employed to verify the applicability jection [57–59]. For this purpose, hat value and standardized residuals
range of the proposed proxy model for miscible CO2 injection process were determined for both data samples including training and testing.

910
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

70 80

75
60 y = 0.9758x + 1.294
70 R² = 0.9811
Recovery Factor (%OOIP)

Proxy Model Output

50 65

60
40
55

30 50

45
20 40

35
10
30
30 35 40 45 50 55 60 65 70
0
0 2000 4000 6000 8000 10000 12000 Simulated Recovery Factor

Oil Production Rate (STB/D) Training Data Best Fit Training Data : R² = 0.9811

Fig. 6. Recovery factor of miscible CO2 injection versus the corresponding oil production (a)
rate.
80

75
70
70

Proxy Model Output

60 65
y = 0.9301x + 2.7581
60 R² = 0.9934
Recovery Factor (%OOIP)

50
55

50
40
45
30 40

35
20
30
30 35 40 45 50 55 60 65 70
10 Simulated Recovery Factor

Testing Data Best Fit Testing Data : R² = 0.9934

0
0.E+00 2.E+06 4.E+06 6.E+06 8.E+06 1.E+07 1.E+07
(b)
CO2 Injection Rate (MMSCF/D)
80
Fig. 7. Recovery factor of miscible CO2 injection as a function of CO2 injection rate.
75
y = 0.9592x + 1.9446
70 R² = 0.9816
Fig. 13 demonstrates William plot containing hat value and standar-
Proxy Model Output

dized residuals for the whole data samples. As graphically shown in 65

Fig. 13, all the data samples fall in the range ± 3 standardized re- 60
siduals. The red horizontal lines indicate the doubtful index. For in- 55
stance, the data points have greater standardized residual (SR) value 50
than +3 or lower than −3 are doubtful. The blue vertical line re- 45
presents the value of the warning Leverage for the data samples
40
[57–59]. As depicted in Fig. 13, all the outputs of the proposed proxy
35
model are located within the intervals mentioned above. As a result, it
can be concluded that the presented model based on the LSSVM method 30
30 35 40 45 50 55 60 65 70
and BBD approach for estimation of the recovery factor of miscible CO2 Simulated Recovery Factor
injection is statistically correct.
Data Best Fit Overal Data : R² = 0.9816

3.3. Limitations of the proxy model (c)

Fig. 8. Scatter plot of the outputs of the proxy model versus the corresponding recovery
The proxy model developed in this study has the following draw-
factor attained from CMG software for a) training data points b) testing data points c)
backs: entire data points.

• The proxy model can be only applicable in the oil reservoir and 4. Conclusions
geological circumstances similar to the model/system considered in
this study.
• The model is valid within the ranges of the operating parameters
This study presents a new simulation tool which is employed to
model CO2 miscible injection processes through a reliable and accurate
mentioned in this study.
• It can be employed only to predict the performance of the CO2
manner. The main results obtained from the present research work are
as follows:
miscible injection operations.

1. The proposed proxy model to determine the ultimate recovery factor

(RF) of miscible CO2 injection method is simple, precise, and robust
for the purposes of design and optimal operating of EOR plants.

911
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

20 20
Training Data Training Data
15 15
Testing Data Testing Data
10 10
Relative Error %

Relative Error %
5 5

0 0
0 500 1000 1500 2000 2500
-5 -5

-10
-10

-15
-15

-20
0.00E+00 2.00E+06 4.00E+06 6.00E+06 8.00E+06 1.00E+07 -20
BHP of Production Well (psi)
CO2 Injection Rate (MMSCF/D)
Fig. 12. Relative deviation of the proxy model and CMG simulation outputs against
Fig. 9. Relative error of proxy RF with respect to CMG outputs versus injection rate. corresponding values of the BHP of production well.

20 Table 3
Training Data Simulation results, proxy model outputs, and error analysis in RF prediction.
15
Testing Data
ID RF-CMG RF-LSSVM MSE ARD Residual
10
Relative Error %

1 41.787663 40.6199 1.36367 2.794516 1.167763

5 2 55.878799 57.6858 3.265253 −3.23379 −1.807
3 40.404835 40.5007 0.00919 −0.23726 −0.09587
0
4 38.104408 38.6277 0.273835 −1.37331 −0.52329
0 2000 4000 6000 8000 10000 12000
5 43.034195 42.3646 0.448357 1.55596 0.669595
-5
6 53.062904 51.804 1.584839 2.372475 1.258904
7 54.576023 53.0936 2.197578 2.716253 1.482423
-10
8 55.891567 55.9757 0.007078 −0.15053 −0.08413
9 47.592541 48.0771 0.234797 −1.01814 −0.48456
-15
10 42.57518 41.2198 1.837055 3.183498 1.35538
11 40.414654 40.6365 0.049216 −0.54892 −0.22185
-20
Oil Production Rate (STB/D) 12 45.362228 45.7412 0.14362 −0.83544 −0.37897
13 47.569981 48.5135 0.890228 −1.98343 −0.94352
Fig. 10. Comparing proxy model and CMG results in terms of relative deviation at dif- 14 43.288555 44.7213 2.052758 −3.30975 −1.43275
ferent oil production rates. 15 43.350037 43.6891 0.114964 −0.78215 −0.33906
16 61.073956 61.7984 0.524819 −1.18618 −0.72444
17 63.530128 62.3648 1.357989 1.834292 1.165328
20
18 53.268475 54.5412 1.619829 −2.38926 −1.27273
Training Data 19 42.159241 40.7305 2.041301 3.388915 1.428741
15
Testing Data 20 47.142193 47.8575 0.511664 −1.51734 −0.71531
21 43.338001 43.6543 0.100045 −0.72984 −0.3163
10
22 42.260117 42.464 0.041568 −0.48245 −0.20388
Relative Error %

5 23 45.047085 44.9271 0.014396 0.266355 0.119985

24 59.897743 58.7616 1.290821 1.896804 1.136143
0 25 50.524914 50.8064 0.079234 −0.55712 −0.28149
0 1000 2000 3000 4000 5000 6000 7000 8000 26 38.096645 38.9281 0.691317 −2.18249 −0.83145
-5 27 47.234138 48.9685 3.008012 −3.67184 −1.73436
28 56.332222 56.6274 0.08713 −0.52399 −0.29518
-10 29 60.53804 62.852 5.354411 −3.82232 −2.31396
30 52.912228 53.6868 0.599962 −1.46388 −0.77457
-15 31 45.21104 45.2481 0.001373 −0.08197 −0.03706
32 64.277306 63.2059 1.147911 1.66685 1.071406
-20 33 42.303593 41.876 0.182836 1.010772 0.427593
BHP of Injection Well (psi) 34 63.738056 61.9744 3.110482 2.767038 1.763656
35 47.920063 48.261 0.116238 −0.71147 −0.34094
Fig. 11. Difference between the predicted and simulated RF for the miscible CO2 injection
36 46.46006 48.5317 4.291692 −4.45897 −2.07164
operation at various values of injection well BHP. 37 60.874733 59.4471 2.038136 2.345198 1.427633
38 65.921333 64.0203 3.613926 2.88379 1.901033
39 53.176949 53.5167 0.115431 −0.63891 −0.33975
2. Based on the magnitudes of the statistical indexes including MSE,
40 46.097 45.5089 0.345862 1.275788 0.5881
ARD, AARD, and residual values, the proxy model developed in this 41 42.837906 43.3724 0.285684 −1.24771 −0.53449
study provides reliable results, implying the model is statistically 42 60.084507 58.4017 2.831839 2.800734 1.682807
acceptable. 43 63.805046 61.5407 5.127263 3.548851 2.264346
44 41.126812 40.5923 0.285703 1.299668 0.534512
3. The Leverage method was employed to validate the applicability
45 42.910915 41.6025 1.71195 3.049143 1.308415
range of the proxy model for miscible CO2 injection processes and to 46 51.572277 51.1041 0.21919 0.907808 0.468177
evaluate the quality of the simulation outputs. According to the
William plot, the hybridization of the LSSVM method and
Box–Behnken design (BBD) approach for RF estimation of miscible 4. Employing a proxy model, a parametric sensitivity analysis was
CO2 injection operations is statistically correct so that the hybrid performed to study the impacts of important parameters (e.g.,
model can forecast the production behavior/trend of the recovery bottom-hole pressure, oil production rate, and CO2 injection rate) on
technique. the target variable. It is concluded that CO2 injection rate is the most

912
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

Table 4 CO2 EOR via thickeners, foams, and gels – a literature review of 40 years of research
Statistical parameters of the outputs achieved from the proxy model developed in this and pilot tests. In: SPE improved oil recovery symposium. Society of Petroleum
study for miscible CO2 injection. Engineers; 2012.
[2] Ding M, Yuan F, Wang Y, Xia X, Chen W, Liu D. Oil recovery from a CO2 injection in
Training Testing Overall heterogeneous reservoirs: the influence of permeability heterogeneity, CO2-oil
miscibility and injection pattern. J Nat Gas Sci Eng 2017;44:140–9.
Correlation coefficient (R2) 0.9811 0.9934 0.9811 [3] Ahmadi MA, Zeinali Hasanvand M, Behbahani SS, Nourmohammad A, Vahidi A,
Amiri M, et al. Effect of operational parameters on the performance of carbonated
Mean squared error (MSE) 1.153 1.615 1.243
water injection: experimental and numerical modeling study. J Supercrit Fluids
Average absolute relative deviation (AARD) 1.758 1.9613 1.797
2016;107:542–8.
[4] Ahmadi MA, Pouladi B, Barghi T. Numerical modeling of CO2 injection scenarios in
petroleum reservoirs: application to CO2 sequestration and EOR. J Nat Gas Sci Eng
4 2016;30:38–49.
[5] Zhang R, Yin X, Wu Y-S, Winterfeld P. A fully coupled model of non-isothermal
3
Training Data multiphase flow, solute transport and reactive chemistry in porous media. In: SPE
Standardized Residual

2 Testing Data annual technical conference and exhibition. Society of Petroleum Engineers; 2012.
[6] Wu Y-S, Chen Z, Kazemi H, Yin X, Pruess K, Oldenburg C, et al. Simulation of
1 coupled processes of flow, transport, and storage of CO2 in saline aquifers. Golden,
CO (United States): Colorado School of Mines; 2014.
0 [7] Zhao X, Rui Z, Liao X, Zhang R. The qualitative and quantitative fracture evaluation
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 methodology in shale gas reservoir. J Nat Gas Sci Eng 2015;27:486–95.
-1 [8] Zhao X, Rui Z, Liao X, Zhang R. A simulation method for modified isochronal well
testing to determine shale gas well productivity. J Nat Gas Sci Eng 2015;27:479–85.
-2
[9] Yao Y, Wang Z, Li G, Wu H, Wang J. Potential of carbon dioxide miscible injections
-3 into the H-26 reservoir. J Nat Gas Sci Eng 2016;34:1085–95.
[10] Zhang R, Winterfeld PH, Yin X, Xiong Y, Wu Y-S. Sequentially coupled THMC model
-4 for CO2 geological sequestration into a 2D heterogeneous saline aquifer. J Nat Gas
Hat Value Sci Eng 2015;27:579–615.
[11] Wu Y-S, Fakcharoenphol P, Zhang R. Non-Darcy displacement in linear composite
Fig. 13. William plot for the results obtained from the proposed proxy model for CO2
and radial flow porous media. In: SPE EUROPEC/EAGE annual conference and
miscible injection operations. exhibition. Society of Petroleum Engineers; 2010.
[12] Xiong Y, Fakcharoenphol P, Winterfeld P, Zhang R, Wu Y-S. Coupled geomechanical
and reactive geochemical model for fluid and heat flow: application for enhanced
important factor, affecting production performance. The outcomes geothermal reservoir. In: SPE reservoir characterization and simulation conference
are satisfactory, as well. This phase of study again confirms the re- and exhibition. Society of Petroleum Engineers; 2013.
liability and appropriateness of the developed model. [13] Yao Y, Ji Z. A quick evaluation model for CO2 flooding and sequestration. Pet Sci
2010;7:515–23.
5. The model developed in this study can be linked to the commercial [14] Zhang R, Wu Y-S, Fakcharoenphol P. Non-Darcy displacement in linear composite
reservoir simulation packages such as computer modeling group and radial aquifer during CO2 sequestration. Int J Oil Gas Coal Technol
(CMG) software to improve their performance and accuracy while 2014;7:244–62.
[15] Kovscek A. Screening criteria for CO2 storage in oil reservoirs. Pet Sci Technol
forecasting the recovery factor for the miscible CO2 injection pro- 2002;20:841–66.
cesses. [16] Mo S, Akervoll I. Modeling long-term CO2 storage in aquifer with a black-oil re-
servoir simulator. In: SPE/EPA/DOE exploration and production environmental
conference. Society of Petroleum Engineers; 2005.
Acknowledgments
[17] Ampomah W, Balch R, Cather M, Rose-Coss D, Dai Z, Heath J, et al. Evaluation of
CO2 storage mechanisms in CO2 enhanced oil recovery sites: application to morrow
The financial support from Natural Sciences and Engineering sandstone reservoir. Energy Fuels 2016;30:8545–55.
[18] Ampomah W, Balch R, Cathar M, Will R, Lee S, Dai Z. Performance of CO2-EOR and
Research Council of Canada (NSERC), InnovateNL, and Statoil Canada
storage processes under uncertainty. In: SPE Europec featured at 78th EAGE con-
is greatly appreciated. ference and exhibition. Society of Petroleum Engineers; 2016.
[19] Dai Z, Viswanathan H, Middleton R, Pan F, Ampomah W, Yang C, et al. CO2 ac-
Appendix A counting and risk analysis for CO2 sequestration at enhanced oil recovery sites.
Environ Sci Technol 2016;50:7546–54.
[20] Dai Z, Middleton R, Viswanathan H, Fessenden-Rahn J, Bauman J, Pawar R, et al.
For using the proxy model, we need to install the LSSVM toolbox in An integrated framework for optimizing CO2 sequestration and enhanced oil re-
MATLAB software firstly. The next step is dragging and dropping the covery. Environ Sci Technol Lett 2013;1:49–54.
[21] Pan F, McPherson BJ, Dai Z, Jia W, Lee S-Y, Ampomah W, et al. Uncertainty analysis
RF.mat file in the MATLAB workspace. The following command should of carbon sequestration in an active CO2-EOR field. Int J Greenhouse Gas Control
be then written in the MATLAB workspace to calculate the recovery 2016;51:18–28.
factor. [22] Dai Z, Viswanathan H, Fessenden-Rahn J, Middleton R, Pan F, Jia W, et al.
Uncertainty quantification for CO2 sequestration and enhanced oil recovery. Energy
RF = simlssvm ({trainX’, trainY’, type, gam, sig2, ‘RBF_kernel’, Procedia 2014;63:7685–93.
‘preprocess’}, {alpha, b}, [Injection BHP; CO2 mole fraction; Injection [23] Ampomah W, Balch RS, Grigg RB, McPherson B, Will RA, Lee SY, et al. Co-opti-
Rate; Producer BHP; Production Rate]) mization of CO2-EOR and storage processes in mature oil reservoirs. Greenh Gases
Sci Technol 2017;7:128–42.
For example, consider the following data:
[24] Ampomah W, Balch R, Cather M, Will R, Gunda D, Dai Z, et al. Optimum design of
CO2 Injection Rate = 4600000 MMSCF/D CO2 storage and oil recovery under geological uncertainty. Appl Energy
Maximum Bottom-Hole Pressure of Injection Well = 6300 psi 2017;195:80–92.
[25] Wood DJ, Lake LW, Johns RT, Nunez V. A screening model for CO2 flooding and
Minimum Bottom-Hole Pressure of Injection Well = 1640 psi
storage in Gulf Coast reservoirs based on dimensionless groups. In: SPE/DOE
Oil Production Rate = 5500 STB/D symposium on improved oil recovery. Society of Petroleum Engineers; 2006.
CO2 Concentration (mole fraction) = 1 [26] Helaleh AH, Alizadeh M. Performance prediction model of Miscible Surfactant-CO2
RF = simlssvm ({trainX’, trainY’, type, gam, sig2, ‘RBF_kernel’, displacement in porous media using support vector machine regression with
parameters selected by Ant colony optimization. J Nat Gas Sci Eng
‘preprocess’}, {alpha, b}, [6300 1 4600000 1640 6300]) 2016;30:388–404.
The recovery factor predicted by the proxy model is equal to [27] Jaber AK, Awang MB, Lenn CP. Box-Behnken design for assessment proxy model of
43.3724, while the corresponding simulated recovery factor using CMG miscible CO2-WAG in heterogeneous clastic reservoir. J Nat Gas Sci Eng
2017;40:236–48.
software is 42.837906, implying a very good agreement between the [28] Vapnik VN, Vapnik V. Statistical learning theory. New York: Wiley; 1998.
estimated and simulated recovery factor. [29] Cortes C, Vapnik V. Support-vector networks. Mach Learn 1995;20:273–97.
[30] Burges CJ. A tutorial on support vector machines for pattern recognition. Data Min
Knowl Disc 1998;2:121–67.
References [31] Suykens J, Gestel T, Brabanter J, Moor B, Vandewalle J. Least squares support
vector machines. KU Leuven, Belgium: World Scientific Publishing Co. Pvt. Ltd.;
[1] Enick RM, Olsen DK, Ammer JR, Schuller W. Mobility and conformance control for 2002.

913
M.A. Ahmadi et al. Fuel 215 (2018) 904–914

[32] Ahmadi MA. Connectionist approach estimates gas–oil relative permeability in [46] Cullick AS, Johnson WD, Shi G. Improved and more rapid history matching with a
petroleum reservoirs: application to reservoir simulation. Fuel 2015;140:429–39. nonlinear proxy and global optimization. In: SPE annual technical conference and
[33] Keerthi SS, Lin C-J. Asymptotic behaviors of support vector machines with Gaussian exhibition. Society of Petroleum Engineers; 2006.
kernel. Neural Comput 2003;15:1667–89. [47] Keating EH, Harp DH, Dai Z, Pawar RJ. Reduced order models for assessing CO2
[34] Ahmadi MA. Toward reliable model for prediction drilling fluid density at wellbore impacts in shallow unconfined aquifers. Int J Greenhouse Gas Control
conditions: a LSSVM model. Neurocomputing 2016;211:143–9. http://dx.doi.org/ 2016;46:187–96.
10.1016/j.neucom.2016.01.106i. [48] Jin Y. A comprehensive survey of fitness approximation in evolutionary computa-
[35] Ahmadi MA, Ebadi M, Marghmaleki PS, Fouladi MM. Evolving predictive model to tion. Soft Comput 2005;9:3–12.
determine condensate-to-gas ratio in retrograded condensate gas reservoirs. Fuel [49] Jin Y. Surrogate-assisted evolutionary computation: Recent advances and future
2014;124:241–57. challenges. Swarm Evol Comput 2011;1:61–70.
[36] Ahmadi MA, Pournik M. A predictive model of chemical flooding for enhanced oil [50] Jin Y, Olhofer M, Sendhoff B. Managing approximate models in evolutionary
recovery purposes: application of least square support vector machine. Petroleum aerodynamic design optimization. In: Proceedings of the 2001 congress on evolu-
2016;2(2):177–82. tionary computation, 2001. IEEE; 2001. p. 592–99.
[37] Ahmadi MA, Mahmoudi B, Yazdanpanah A. Development of robust model to esti- [51] Razavi S, Tolson BA, Burn DH. Review of surrogate modeling in water resources.
mate gas–oil interfacial tension using least square support vector machine: ex- Water Resour Res 2012;48.
perimental and modeling study. J Supercrit Fluid 2016;107:122–8. [52] Sayyafzadeh M. History matching by online metamodeling. In: SPE reservoir
[38] Ahmadi MA, Ebadi M. Evolving smart approach for determination dew point characterisation and simulation conference and exhibition. Society of Petroleum
pressure through condensate gas reservoirs. Fuel 2014;117:1074–84. Engineers; 2015.
[39] Van Gestel T, Suykens JA, Baesens B, Viaene S, Vanthienen J, Dedene G, et al. [53] Sayyafzadeh M, Haghighi M. Regularization in history matching using multi-ob-
Benchmarking least squares support vector machine classifiers. Mach Learn jective genetic algorithm and Bayesian framework (SPE 154544). In: 74th EAGE
2004;54:5–32. conference and exhibition incorporating EUROPEC 2012; 2012.
[40] Suykens JA, Vandewalle J. Training multilayer perceptron classifiers based on a [54] Zubarev DI. Pros and cons of applying proxy-models as a substitute for full reservoir
modified support vector method. IEEE Trans Neural Networks 1999;10:907–11. simulations. In: SPE annual technical conference and exhibition. Society of
[41] Suykens J, Van Gestel T, De Brabanter J, De Moor B, Vandewalle J. Least squares Petroleum Engineers; 2009.
support vector machines. Singapore: World Scientific Publishing; 2002. [55] Silva PC, Maschio C, Schiozer DJ. Application of neural network and global opti-
[42] Suykens JA, Vandewalle J. Recurrent least squares support vector machines. IEEE mization in history matching. J Can Pet Technol 2008;47.
Trans Circuits Syst I Fundam Theory Appl 2000;47:1109–14. [56] Sampaio TP, Ferreira Filho VJM, Neto ADS. An application of feed forward neural
[43] Suykens JA, De Brabanter J, Lukas L, Vandewalle J. Weighted least squares support network as nonlinear proxies for use during the history matching phase. In: Latin
vector machines: robustness and sparse approximation. Neurocomputing American and Caribbean petroleum engineering conference. Society of Petroleum
2002;48:85–105. Engineers; 2009.
[44] Ahmadi MA, Zahedzadeh M, Shadizadeh SR, Abassi R. Connectionist model for [57] Rousseeuw PJ, Leroy AM. Robust regression and outlier detection. John Wiley &
predicting minimum gas miscibility pressure: application to gas injection process. Sons; 2005.
Fuel 2015;148:202–11. [58] Gramatica P. Principles of QSAR models validation: internal and external. Mol
[45] Ahmadi MA, Ebadi M, Hosseini SM. Prediction breakthrough time of water coning Inform 2007;26:694–701.
in the fractured reservoirs by implementing low parameter support vector machine [59] Goodall CR. 13 Computation using the QR decomposition. Handb Stat
approach. Fuel 2014;117:579–89. 1993;9:467–508.

914