See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220606228

Support Vector Regression and Functional Networks for viscosity and Gas/Oil
Ratio Curves Estimation.

Article  in  International Journal of Computational Intelligence and Applications · September 2011

DOI: 10.1142/S1469026811003100 · Source: DBLP

CITATIONS READS

17 243

5 authors, including:

Amar Khoukhi Munirudeen Oloso

New York City College of Technology Intenttech Ltd
77 PUBLICATIONS   588 CITATIONS    13 PUBLICATIONS   96 CITATIONS   

SEE PROFILE SEE PROFILE

Moustafa Elshafei Abdulaziz Al-Majed

Zewail University of Science and Technology King Fahd University of Petroleum and Minerals
165 PUBLICATIONS   1,899 CITATIONS    103 PUBLICATIONS   1,435 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Wax Removal Using Thermochemical fluids View project

Improving the Productivity of Tight Reservoirs View project

All content following this page was uploaded by Amar Khoukhi on 11 February 2015.

The user has requested enhancement of the downloaded file.

 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

International Journal of Computational Intelligence and Applications

Vol. 10, No. 3 (2011) 269–293

c Imperial College Press
DOI: 10.1142/S1469026811003100

SUPPORT VECTOR REGRESSION AND FUNCTIONAL

NETWORKS FOR VISCOSITY AND GAS/OIL
RATIO CURVES ESTIMATION

AMAR KHOUKHI∗,‡ , MUNIRUDEEN OLOSO∗,§ , MOSTAFA ELSHAFEI∗,¶ ,

ABDULAZEEZ ABDULRAHEEM†, and ABDULAZIZ AL-MAJED†,∗∗
∗Systems Engineering Department

King Fahd University of Petroleum and Minerals, Dhahran, KSA

†Petroleum Engineering Department

King Fahd University of Petroleum and Minerals, Dhahran, KSA

‡[email protected]
§[email protected]
¶[email protected]
[email protected]
∗∗[email protected]

Received 10 October 2010

Revised 22 December 2010

In oil and gas industry, prior prediction of certain properties is needed ahead of explo-
ration and facility design. Viscosity and gas/oil ratio (GOR) are among those properties
described through curves with their values varying over a specific range of reservoir pres-
sures. However, the usual single point prediction approach could result into curves that
are inconsistent, exhibiting scattered behavior as compared to the real curves. Support
Vector Regressors and Functional Networks are explored in this paper to solve this prob-
lem. Inputs into the developed models include hydrocarbon and non-hydrocarbon crude
oil compositions and other strongly correlating reservoir parameters. Graphical plots
and statistical error measures, including root mean square error and average absolute
percent relative error, have been used to evaluate the performance of the models. A com-
parative study is performed between the two techniques and with the conventional feed
forward artificial neural networks. Most importantly, the predicted curves are consistent
with the shapes of the physical curves of the mentioned oil properties, preserving the
need of such curves for interpolation and ensuring conformity of the predicted curves
with the conventional properties.

Keywords: Reservoir characterization; viscosity; gas/oil ratio (GOR); Artificial Neural

Networks; Support Vector Regressors; Functional Networks.

1. Introduction
1.1. Motivation
Reservoir fluid properties are very important in petroleum engineering compu-
tations such as material balance calculations, well test analysis, reserve estima-
tion, inflow performance calculations, fluid flow in porous media, evaluation of
new formation for potential development, numerical reservoir simulations, design

269
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

270 A. Khoukhi et al.

of production equipment, and planning future enhanced oil recovery projects. At

every stage of the petroleum exploration and production business, a priori knowl-
edge of how the fluids will behave under a wide range of pressure and temperature
conditions, particularly in terms of their volumetric and thermo physical properties,
is required. The relationships of Pressure-Volume-Temperature (PVT) properties
for oil and gas are traditionally estimated using empirical studies. Ideally, those
properties could be measured in laboratories. The problem with those measure-
ments is the availability of those laboratory tests and the right samples collected
from the well-bore or well surface. The need for the prediction of PVT properties
was then accomplished through the use of equations of states (EOS), which are
derived from the basic mass, energy, and chemical balance equations. Because the
EOS are derived for pure substances, correction factor(s) are always added when
used on practical data. To overcome these problems, empirically derived correla-
tions between those properties and well data had been developed based on available
data for different regions in the world. During the last several years, neural net-
works have been used to obtain better prediction models than the empirical ones
and they have shown a significant prediction improvement.
Unfortunately, the developed neural networks correlations are often limited and
global correlations are usually less accurate compared to local correlations. Nev-
ertheless, the achievements of neural networks opened the door to computational
intelligence techniques to play a major role in oil and gas industry. To improve
prediction accuracy, computational intelligence techniques, such as Artificial Neu-
ral Networks (ANN), Support Vector Machines (SVM), Adductive Networks and
Genetic Algorithms (GA) among others, have been applied.
Some of these properties described as curves are estimated through single or
multi-data point prediction. However, the usual single or multi-data point predic-
tions could compromise the original shape of the curves. Henceforth, prediction
techniques for entire curves are needed to elaborate for some of these properties.
Two of such PVT properties that need to be presented as curves are oil vis-
cosity and gas oil ratio (GOR). These two properties vary with pressure and their
prediction is defined over a specified range of pressures. Fluid viscosity is a measure
of its internal resistance to flow. This property is highly needed for many calcula-
tions and applications in the petroleum industry such as oil recovery estimation,
multi-phase flow calculation, gas-lifting and pipeline design. The GOR is basically
determined through separator calculations. In this paper, two main advanced neu-
ral networks techniques are implemented for oil viscosity and gas oil ratio (GOR)
curves prediction.

2. Related Works
2.1. Viscosity correlations
A good number of empirical correlations have been developed in the literature to
estimate crude oil viscosity at, below and above bubble point pressure (Pb ). Pb is
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 271

the pressure at which the light hydrocarbon components in the oil starts to change
to the gas phase, and hence appear as bubbles in the oil sample. Correlations
based on soft computing techniques have recently been developed to predict this
important PVT property. Graphical correlations1 were developed for predicting vis-
cosity at different reservoir pressures using a data set from U.S.A. oil samples. The
authors correlated the under-saturated oil viscosity with the viscosity at Pb (µob ),
and pressure above bubble point Pb , while using oil gravity and temperature of
range 100–220◦ F to develop the dead oil (at the standard temperature and pres-
sure) viscosity (µod ) correlation. In the same vein, a graphical correlation2 was
presented for predicting oil viscosity at Pb . The correlation was developed as a
function of GOR using viscosity data of 457 crude oil samples from Canada and
U.S.A.
Correlations for µob and µod were developed, with a data set of 2073 oil viscosity
measurements used to develop the µob correlation while 460 dead oil observations
was used to develop µod correlation.3 Another correlation was developed in Ref. 4,
for µod in the temperature range of 50–300◦ F. The variable µod was correlated as
a function of American Petroleum Institute (API) oil gravity scale and tempera-
ture. A large set of PVT measurements was used5 to develop an under-saturated
oil viscosity correlation as a function of µob , Pb and reservoir pressure. Also, a µod
correlation6 based on a modification of Beggs and Robinson’s correlation was pre-
sented. Correlations for µob was used to predict oil viscosity below and above bubble
point were developed in Ref. 7. Total data points of 150, 1503 and 1691 were used
to develop the three correlations. The viscosity data used in the study were from
Saudi crude oil. For µob correlation, gas relative density, solution GOR, relative
temperature and oil relative density, independent variables were used. For viscosity
above and below Pb correlations, the correlating variables used were µob , reservoir
pressure and Pb . Correlations were also developed8 for µod , µob and under-saturated
oil viscosity using light crude oil data of Libya. µod was correlated as a function of
stock tank oil gravity and temperature. The µob was correlated with API oil gravity,
µod and Pb , while the under-saturated oil viscosity was correlated with pressure,
Pb , µob , µod and API oil gravity.
In Ref. 9, new empirical correlations for µod , µob and under-saturated oil vis-
cosity were suggested. The µod was correlated as a function of API oil gravity and
reservoir temperature, µob as a function of dead oil viscosity and solution GOR,
and under-saturated oil viscosity as a function of µob , Pb and reservoir pressure. A
total of 126 laboratory PVT analyses from Texas and Louisiana in U.S. were used
to develop oil viscosity correlations. Other correlations for µob and under-saturated
oil viscosity based on UAE crude oil were presented in Ref. 10. The correlated µob
with solution GOR, reservoir temperature, gas specific gravity and API oil grav-
ity using 57 data points. An under-saturated oil viscosity correlation was devel-
oped as a function of Pb , µob , reservoir pressure and solution GOR using 328 data
points.
Other correlation functions were introduced in Ref. 11 for µod , µob and under-
saturated oil viscosity for Gulf of Mexico crude oils based on 100 PVT laboratory
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

272 A. Khoukhi et al.

reports. The authors correlated µod with temperature, pressure and solution GOR
at Pb , and API oil gravity. The µob was correlated with µod and solution GOR, while
the under saturated oil viscosity was correlated with µob , Pb , reservoir pressure and
solution GOR.
In the same vein, a number of researchers have developed viscosity correlations
using soft computing techniques. A Radial Basis Function Network model12 was
used for predicting oil viscosity using reservoir pressure and temperature, API
oil gravity and gas gravity as the network inputs. A universal neural-network-
based model for estimating PVT properties of crude oil systems was introduced
in Ref. 13. Another ANN model14 was developed using all the data points of a
reservoir to obtain the viscosity curve. A PVT data of 650 reservoir fluids from
around the world was used to develop the viscosity correlation model and also
for some other PVT properties. An ANN correlation model15 was introduced to
predict brine viscosity using temperature and salinity as the network inputs. A
total of 1040 data points were used to build the model. Another neural network
was constructed16 to predict viscosity below Pb for Pakistani crude oil. The cor-
relating parameters were: pressure, reservoir temperature Pb , oil formation volume
factor, solution GOR, gas specific gravity and API gravity. More recently, an oil
viscosity correlation was presented in Ref. 17 for Iranian crude oil using genetic
algorithms. The input parameters were the pressure, temperature, and reservoir
fluid GOR and oil density. An excellent study had been reported recently,18 imple-
menting a support vector regression technique on PVT Correlations for Indian
Crude oil.
In all abovementioned prediction techniques, including empirical correlations
and soft computing models, data points are being predicted and the shapes of the
resulting estimated curves may not be consistent with the experimental ones.

2.2. Gas/oil ratio correlations

Ordinarily, GOR correlations are derived from Pb correlation. However, a precedent
was set in Ref. 5 where a regression analysis was used to obtain an empirical corre-
lation for Rs . They used 5008 data points to perform the regression analysis. The
solution GOR was correlated as a function of pressure, gas relative density, oil API
gravity and temperature. Other researchers have developed general correlations for
GOR including Refs. 11, 19 and 20. They have all developed separate empirical
correlations for GOR.
Several Pb correlations had been published. Good reviews of these correlations
can be found in Refs. 21–24. Generally, Pb correlation is developed as a function
of solution GOR, gas specific gravity, oil API gravity and reservoir temperature.
Though solution GOR correlation can be obtained from any of the existing Pb
correlation, some possible complexities in solving such resulting solution GOR cor-
relation have been observed in Ref. 11, when there is the need to have separate
correlation for solution GOR.
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 273

3. PVT Data Acquisition and Pre-Processing

At the implementation phase, it is important to make sure that the input data
values fall in a natural domain. Such a quality control step is a must to have very
accurate and reliable results at the end.
The following are the most common domains for the input/output variables,
gas–oil ratio, bubble point pressure, API oil gravity, relative gas density, reservoir
temperature and oil formation volume factor that are used in both input and out-
put layers of modeling schemes for PVT analysis (a nomenclature is provided in
Appendix 1).

• Gas oil ratio which varies from 151 to 1332, scf/stb.

• Bubble point pressure, starting from 210, and ending with 2985, psia.
• Reservoir temperature with its range from 100◦ F to 250◦ F.
• API gravity which changes between 21.4 and 47.6.
• Gas relative density, changing from 0.744 to 1.367.
• Bubble point/gas-saturated oil viscosity 0.88 to 6.49 (cP ).
• Dead oil viscosity varies from 0.305 to 1.91 (cP ).

The implementation studies of the presented work were achieved based on three
databases say, data sets A, B and C. Data set A consists of the hydrocarbon and
non-hydrocarbon components, and some other properties of the crude oil. Data set
B consists of the viscosity-pressure measurements to generate viscosity curves for
the corresponding wells in data set A, while data set C consists of GOR pressure
measurements to generate gas/oil ratio curves for corresponding wells in A. These
data were taken from Middle East crude oil reservoirs.
Initially, there were 106 data points in the set A. In preprocessing the data, we
applied two different outlier-detection methods on the data set A before utilizing it.
These are Cook’s distance method and Chauvenet’s criterion.25 The former method
was implemented using STATISTICA software while details on the latter method
can be found in Ref. 25. Only data points that were detected to be outliers by the
two methods have been declared as outliers and removed.
Eventually, seven data points were declared as outliers. After the removal of
the outliers from data set A, it was reduced to 99 data points. As recommended
and is usually done, the predictors were normalized within the interval [0 1] using
formula (1) below. This makes the input data dimensionless and ensures that the
predictors are independent of the measurement units.

(xold − min(xi ))
xnew = i
, (i = 1, 2, . . . , n). (1)
i
(max(xi ) − min(xi ))
The data set A was then divided into training and testing sets. The training set
consists of 70% (approx. 70 data points) while the testing set consists of 30%
(approx. 29 points). The idea of curve prediction using artificial neural network
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

274 A. Khoukhi et al.

technique was first introduced in Ref. 26 using ANN, in Ref. 27 using RBF, and
Refs. 28 and 29 using a hybrid ANN with differential evolution.

4. Problem Statement
Typical viscosity and GOR curves are shown in Figs. 1 and 2. Each curve represents
the variation of viscosity or GOR for the corresponding oil well. The significance
of such a curve is compromised, if single point or multi-point based prediction is
utilized. Equations (2) and (3) can be used to represent any crude oil viscosity curve
and Eq. (4) is used to represent GOR curves.

β
P − Pd
µ = µod + (µob − µod ) for P < Pb , (2)
Pb − Pd
µ = µob + α(P − Pb ) for P ≥ Pb , (3)

τ
P − Pb
Rs = Rsb , (4)
Pb − Pd
where α and β are viscosity curve coefficients and τ is the fitting GOR curve
coefficient. The statistical distribution of the fitting coefficients is shown in Table 1.
From Eqs. (2) and (3), three parameters µob , α and β are needed to generate
the viscosity curve. The first parameter µob is determined from the laboratory
PVT analyses while α and β are to be generated from the curve fitting. Two

2
Experimental
Predicted
1.8

1.6

1.4
Viscosity

1.2

0.8

0 500 1000 1500 2000 2500 3000

Pressure

Fig. 1. Typical result from single or multi-data point prediction for viscosity curve.
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 275

Gas/Oil Ratio Curve

Experimental
600
Predicted

500

400
Gas/Oil Ratio

300

200

100

0
0 200 400 600 800 1000 1200 1400 1600 1800
Pressure

Fig. 2. Typical result from single or multi-data point prediction for GOR curve.

Table 1. Statistical distribution of the

fitting coefficients.

Parameter Max. Value Min. Value

α 1.91E-04 1.48E-05
β 0.894 0.1251
τ 0.941657 0.42174

parameters, Rsb and τ are needed to generate GOR curves, Rsb is determined
from PVT laboratory analyses while τ is obtained from curve fitting.

5. Approach
5.1. Support vector regression
Support vector machine modeling schemes and methods are among the most suc-
cessful and effective algorithms in both machine learning and data mining commu-
nities. It has been widely used as a robust tool for classification and regression. An
overview can be found in Refs. 30 and 31. Support Vector Regression (SVR) is a
regression version of Support Vector Machines (SVMs) (Fig. 3). Unlike classifica-
tion problems where the outputs are either 1 and 0 or 1 and −1, the outputs in the
regression problems are real numbers. This makes it a bit difficult to model this type
of information which has infinite possibilities. With the introduction of Vapnik’s
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

276 A. Khoukhi et al.

ε-insensitive loss function, SVM has been extended to solve nonlinear regression
estimation problems, leading to techniques known as SVR. These have been shown
to exhibit excellent performance.32 SVR has been found to be very robust to predict
complex nonlinear relationship problems in many applications, including problems
such as optical character recognition, text categorization, and face detection in
images.33 In the case of regression, a margin of tolerance ∈ is set in approximation
to the SVM which would have already being inferred from the problem. As shown
in Figs. 3 and 4, SVMs map input vectors to a higher dimensional space, where a
maximal separating hyperplane is constructed.34–36
The kernel function is responsible for transforming the data set into hyperplane.
The variables of the kernel must be computed accurately since they determine the
structure of high-dimensional feature space which governs the complexity of the
final solution.

Fig. 3. The original input space mapped to a higher feature space with a separable training set.

Fig. 4. Soft margin loss setting for a linear dimensional SVR (Schölkopf and Smola, 2002).
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 277

After the selection of a kernel, the other highly influential parameters in any
SVR model based on the observation are “C” and “kernel option”. For polynomial
kernel, kerneloption denotes the degree of the kernel polynomial while it denotes
kernel bandwidth for “Gaussian”. “C” is the trade-off between achieving minimal
training error and complexity of the model. The kernel functions and SVR options
used in this study appear in Appendix 2.

5.2. Functional networks

Functional networks were introduced as a powerful alternative to neural
networks.37,38 Unlike neural networks, functional networks have the advantage of
using domain knowledge in addition to data knowledge. The network initial topol-
ogy can be derived based on the modeling of the properties of the real world. Once
this topology is available, functional equations allow one to obtain a much simpler
equivalent topology. Although functional networks can also deal with data only, the
class of problems where functional networks are most convenient is the class where
the two sources of knowledge about domain and data are available. In functional
networks, neural functions are to be learned instead of weights (Figs. 5 and 6). To
learn these neural functions, a set of linearly independent functions are to be used.
These are called basis functions. Possible basis functions are: polynomial, expo-
nential, Fourier and logarithm functions or their combinations. The selection of the
basis function along with the possible learning method is essential in developing the
FN model. To learn (parametric) functional networks, one can choose different sets
of linearly independent functions for the approximation of the neuron functions.
At the same time, there is a need to select the best model according to some
criterion of optimality. The Minimum Description Length Principle (MDLP) is one
of the model selection principles we can use as discussed in Ref. 37. This consists
of finding the minimum information required to store the given training set using
a functional network model. Therefore, it was demonstrated that the best func-
tional network model for a given problem corresponds to that with the minimum
description length value.38,39 The code length L(x) of x is defined as the amount
of memory needed to store the information x.

W41
X1 X4

f (ΣW4i Xi)
W42 W64

X2 W52 X6
f (ΣW6i Xi)
f (ΣW5i Xi) X5
W65
X3
W53

Fig. 5. A standard neural network.

 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

278 A. Khoukhi et al.

X4
X1
f1 (X1, X2)
X6
X2 f3 (X4, X5)

X3
f2 (X2, X3) X5

Fig. 6. A standard functional network.

The output is given by:

y = f1 (x1 ) + f2 (x2 ) + f3 (x3 ) + · · · + f12 (x12 ). (5)

For the case at hand, the different families of functions used for each parameter
appear in Appendix 3. The MDLP was used to optimize the network and select the
best model. It can be noted that in some cases, some functions are zero, this means
that the corresponding input to that node does not really affect the predicting
output at that instance.

5.3. Statistical quality measures

The performance and accuracy of SVR and FN as well as a feed-forward neural
network (FFNN) are compared. In doing so, two commonly used statistical tech-
niques have been adopted along with the graphical plots of the predicted curves
(only sample plots are shown here for comparison). These are the root mean square
error (RMSE) of the training and testing wells Eq. (6) and average absolute percent
relative error (AAPRE) of the training and testing wells Eq. (8).40,41
The formulas for the two statistical measures RMSE and AAPRE are given as
follows:

(1) Root mean square error


(x1 − y1 )2 + (x2 − y2 )2 + · · · + (xn − yn )2
RMSE = . (6)
n
(2) Average absolute percent relative error


xi − yi
Ei = × 100; (i = 1, 2, 3, . . . , n), (7)
yi

1
n
AAPRE = |Ei |. (8)
n i

A good model should have low RMSE and AAPRE values.

 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 279

In Eqs. (6)–(8), x s are the predicted values, y  s are the actual/experimental

values and n is the total number of data points in all training wells (70 data points)
or testing wells (29 data points).

6. Results and Discussion

6.1. Viscosity curve prediction
In the performed study, a comparison is done with the artificial neural networks to
assess the performance of SVR, as compared to FN and ANN. A feedforward neural
network (FFNN) model is developed for five aforementioned prediction variables.
A number of trials were made viz: selecting the number of hidden layers, number
of neurons in each hidden layer and the training algorithm. For µob and Rsb , we
eventually used two hidden layers with thirteen and six nodes respectively. Hence,
we have 12-13-6-1 FFNN structure, (12 input neurons, 13 nodes in the first hidden
layer, 6 nodes in the second hidden layer and 1 output neuron), for each case
parameter. For the three fitting variables, we used 12-12-5-1 FFNN architecture.
In all cases, tangent sigmoid transfer function and Levenberg-Marquardt training
optimization were eventually used, and the best network out of 1000 runs in each
case was taken.
Also, a sample of prediction plots of training and testing wells from the two
frameworks are shown in Figs. 7 through 12. The statistical performance measures

Viscosity Curve - Experimental and Predicted (Training)

Experimental
SVR
FN
2

1.5
Viscosity (cP)

0.5

0
0 500 1000 1500 2000 2500
Pressure (psi)

Fig. 7. Viscosity vs. pressure plot for sample well TR1.

 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

280 A. Khoukhi et al.

Viscosity Curve - Experimental and Predicted (Testing)

2 Experimental
SVR
1.8 FN

1.6

1.4
Viscosity (cP)

1.2

0.8

0.6

0.4

0.2

0
0 500 1000 1500 2000 2500 3000 3500 4000
Pressure (psi)

Fig. 8. Viscosity vs. pressure plot for sample well TS1.

Viscosity Curve - Experimental and Predicted (Training)

2 Experimental
SVR
1.8
FN
1.6

1.4
Viscosity (cP)

1.2

0.8

0.6

0.4

0.2

0
0 500 1000 1500 2000 2500 3000
Pressure (psi)

Fig. 9. Viscosity vs. pressure plot for sample well TR2.

 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 281

Viscosity Curve - Experimental and Predicted (Testing)

1.5 Experimental
SVR
FN

1
Viscosity (cP)

0.5

0
0 500 1000 1500 2000 2500 3000 3500 4000
Pressure (psi)

Fig. 10. Viscosity vs. pressure plot for sample well TS2.

Viscosity Curve - Experimental and Predicted (Training)

Experimental
SVR
2 FN

1.5
Viscosity (cP)

0.5

0
0 500 1000 1500 2000 2500
Pressure (psi)

Fig. 11. Viscosity vs. pressure plot for sample well TR3.
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

282 A. Khoukhi et al.

Viscosity Curve - Experimental and Predicted (Testing)

3.5
Experimental
SVR
3
FN

2.5
Viscosity (cP)

1.5

0.5

0
0 500 1000 1500 2000 2500 3000 3500 4000
Pressure (psi)

Fig. 12. Viscosity vs. pressure plot for sample well TS3.

Table 2. Statistical performance measures of SVR, FN and FFNN models for

viscosity curves prediction.

Model SVR FN FFNN

RMSE AAPRE% RMSE AAPRE% RMSE AAPRE%
Training 0.07495 6.3953 0.067648 5.400661 0.08664 8.33945
Testing 0.07659 8.5969 0.079412 8.551437 0.08712 10.24569

for the two frameworks are shown in Table 2. For this pair of techniques, the
performance of both frameworks, SVR and FN, are very competitive. While FN
performance is better than that of SVR in the training phase with lower RMSE
and AAPRE, which are 0.06765 and 5.4% respectively, against those of SVR which
are 0.07495 and 6.3953% respectively, SVR performance is very competitive with
that of FN for the testing wells. For the testing phase, FN has lower AAPRE
of 8.5514%, against that of SVR which is 8.5969%, while SVR has lower RMSE,
0.0765, against that of FN which is 0.07941. Also, from Table 2, the RMSE and
APPRE for the FFNN predictions are the highest for both training and testing.
In essence, the results of SVR and FN are very competitive for viscosity curve
prediction, while both clearly outperform FFNN. The predicted curves from the
two SC techniques show good matching with the experimental curves for both
training and testing wells with little deviation in some testing wells. Table 3 shows
a sample of predicted viscosity curve parameters by SVR FN and FFNN models.
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 283

Table 3. Sample predicted viscosity curve parameters by SVR, FN and

FFNN models.
Actual SVR FN FFNN
Training: α 7.19E-05 5.79E-05 5.7914E-05 5.89E-05
β 0.6688 0.626296 0.625281 0.50256
µob 0.69 0.686544 0.71764 0.6628
Testing: α 3.87E-05 4.88E-05 4.32E-05 5.38E-05
β 0.3388 0.3386 0.3190 0.4889
µob 0.58 0.5736 0.5549 0.5228

Table 4. Time complexity of all models

for viscosity curve prediction.

CPU Time (seconds)

Model Training Testing
SVR 2.6208 0.00103
FN 2.318 0.0936
FFNN 11.466 0.0468

Comparison based on time to complete development of each model is shown in

Table 4. The training time may not be necessary or could be traded off, since after
development of the model, only the testing phase will be utilized. And from this
viewpoint, it is clear that SVR is the best of the three techniques.

6.2. Gas/oil ratio curve prediction

The SVR and FN frameworks were implemented to predict the required variables,
τ and Rsb for gas/oil ratio curve prediction. Similar to the previous cases, only
sample training and testing plots of the predicted gas/oil ratio curves are shown
in Figs. 13 through 18. The predicted curves from these two techniques show good
matching with the experimental curves for training and testing wells. Table 5 shows
the statistical measures for evaluating the performance of SVR, FN and FFNN
techniques in predicting gas/oil ratio curves. In this case, unlike the viscosity curve
prediction where performances of both SVR and FN are very competitive, SVR
has better average performance than FN in both training and testing phases, based
on the statistical measures used for evaluation. For the training, SVR has RMSE
19.0043 and AAPRE of 7.5279%, while FN has RMSE of 21.6942 and AAPRE of
8.4167%. For testing, SVR has RMSE of 30.0170 and AAPRE of 9.0757%, while
FN has RMSE of 32.8196 and AAPRE of 10.2012%. Table 6 shows that FFNN
predictions are the worst with the highest RMSE and AAPRE.
Based on the preceding analysis, though performance of FN is also good, SVR
framework gives better performance in predicting gas/oil ratio than FN. This is
also evident from the sample predicted curves. Table 6 shows a sample of predicted
GOR curve parameters by SVR, FN and FFNN models.
Table 7 shows the computational time for both training and testing of SVR, FN
and FFNN models for gas/oil ratio curve prediction. It is noteworthy that SVR is
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

284 A. Khoukhi et al.

Solution Gas/Oil Ratio Curve - Experimental and Predicted (Training)

800
Experimental
SVR
700
FN
SolutionGas/Oil Ratio (SCF/STB)

600

500

400

300

200

100

0
0 200 400 600 800 1000 1200 1400 1600 1800
Pressure (psi)

Fig. 13. Gas/oil ratio vs. pressure plot for sample well TR1.

Solution Gas/Oil Ratio Curve - Experimental and Predicted (Testing)

Experimental
800 SVR
FN
700
SolutionGas/Oil Ratio (SCF/STB)

600

500

400

300

200

100

0
0 200 400 600 800 1000 1200 1400 1600 1800
Pressure (psi)

Fig. 14. Gas/oil ratio vs. pressure plot for sample well TS1.
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 285

Solution Gas/Oil Ratio Curve - Experimental and Predicted (Training)

600
Experimental
SVR
FN
500
Solution Gas/Oil Ratio (SCF/STB)

400

300

200

100

0
0 200 400 600 800 1000
Pressure (psi)

Fig. 15. Gas/oil ratio vs. pressure plot for sample well TR2.

SolutionGas/Oil Ratio Curve - Experimental and Predicted (Testing)

1000 Experimental
SVR
900
FN
Solution Gas/Oil Ratio (SCF/STB)

800

700

600

500

400

300

200

100

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Pressure (psi)

Fig. 16. Gas/oil ratio vs. pressure plot for sample well TS2.
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

286 A. Khoukhi et al.

Solution Gas/Oil Ratio Curve - Experimental and Predicted (Training)

400
Experimental
SVR
350
FN
Solution Gas/Oil Ratio (SCF/STB)

300

250

200

150

100

50

0
0 100 200 300 400 500 600 700 800
Pressure (psi)

Fig. 17. Gas/oil ratio vs. pressure plot for sample well TR3.

Solution Gas/Oil Ratio Curve - Experimental and Predicted (Testing)

Experimental
900 SVR
FN
800
Solution Gas/Oil Ratio (SCF/STB)

700

600

500

400

300

200

100

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Pressure (psi)

Fig. 18. Gas/oil ratio vs. pressure plot for sample well TS3.
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 287

Table 5. Statistical performance measures of SVR, FN and ANN models for

gas/oil curve prediction.

Model SVR FN FFNN

RMSE AAPRE% RMSE AAPRE% RMSE AAPRE%
Training 19.0043 7.5279 21.6942 8.4167 23.5217 8.804342
Testing 30.0170 9.0757 32.8196 10.2012 39.0161 12.701

Table 6. Sample predicted gas/oil curve parameters by SVR, FN

and ANN models.
Actual SVR FN FFNN
Training: τ 0.7762 0.6655 0.6813 0.6558
Rsb 558 584.0497 593.5556 599.3255
Testing: τ 0.61900 0.6397 0.5983 0.6468
Rsb 689 689.0279 669.7892 692.3316

Table 7. Time complexity of all models

for gas/oil ratio curve prediction.

CPU Time (seconds)

Model Training Testing
SVR 3.0264 0.0001
FN 2.4804 0.0624
FFNN 7.566 0.078

Table 8. Gas/oil ratio correlations.

Model Training Testing

SVR 0.997233 0.987923
FN 0.994975 0.964899

Table 9. Viscosity correlations.

Model Training Testing

SVR 0.986636 0.921167
FN 0.991126 0.920655

more demanding than FN at the training and testing with a significant difference,
whereas the FFNN is the most demanding at training. Furthermore, Tables 8 and 9
show the correlation coefficient as to relate the statistical significance of the results
which also illustrate that SVR performs better than FN for predicting gas/oil Curve
and that SVR and FN are competitive regarding viscosity curve prediction.

7. Conclusion
In this paper, we have presented two advanced computational intelligence tech-
niques to predict crude oil Pressure-Volume-Temperature (PVT) properties that
need to be represented as curves over a specified range of reservoir pressures.
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

288 A. Khoukhi et al.

Instead of the usual single or multi-data points prediction, which could distort
the consistency of the curve’s shape, an efficient approach for predicting such PVT
properties curves has been introduced and implemented. In all predictions, we have
implemented different independent neural network techniques, viz: Support Vector
Regression, Functional Networks and Feedforward Neural Network. The viscosity
and gas/oil ratio curves prediction problems were formulated and implemented
using these approaches. Simulation of these results had been reported and compar-
isons between the three techniques were discussed on both viscosity and solution
GOR curve predictions. Interestingly, the shapes of the predicted viscosity and
solution GOR curves are consistent with the physical law and the experimental
curves. This makes the use of such predicted curve practicable for use, contributing
thereby to enhanced exploration and production processes through cost reduction
and improving human operator conditions.

Acknowledgment
This work was supported by King Fahd University of Petroleum and Minerals
(Grant No. SB100014), and partially by King Abdulaziz City for Science and Tech-
nology under Grant MSTP-KACST-08-OIL82-4, Saudi Arabia.

Appendix 1: Nomenclature and Abbreviations

P: Pressure (psi)
Pb : Bubble point pressure (psi)
Pod : Pressure at dead oil viscosity (psi)
Rs : Solution gas/oil ratio, SCF/STB (m3 /m3 )
Rsb : Bubble point solution gas/oil ratio, SCF/STB (m3 /m3 )
T: Temperature (◦ F)
V: Volume (m3 )
µa : Viscosity above bubble point (cP )
µb : Viscosity below bubble point (cP )
µo : Oil viscosity (cP )
µob : Bubble point/gas-saturated oil viscosity (cP )
µod : Dead oil viscosity (cP )
Res Temp: Reservoir temperature (◦ F)
Mol N2 : Mole fraction of N2 (mol%)
Mol CO2 : Mole fraction of CO2 (mol%)
Mol H2 S: The mole fraction H2 S (mol%)
PVT: Pressure-Volume-Temperature
EOS: Equations of States
GOR: Gas/Oil Ratio,
RMSE: Root Mean Square Error
AAPRE: Average Absolute Percent Relative Error
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 289

Appendix 2: Kernel of SVR

For the five predicting variables, using Matlab, the selected option for optimal
relevant variables are stated as follows.
(1) α : C = 10000; lambda = 1e-7; epsilon = 0.09; kernel option = 0.9; kernel =
‘poly’; verbose = 1.
(2) β : C = 60; lambda = 1e-7; epsilon = 0.08; kernel option = 0.8; kernel = ‘poly’;
verbose = 1.
(3) µob : C = 40000; lambda = 1e-7; epsilon = 0.001; kernel option = 0.994; kernel =
‘Gaussian’; verbose = 1.
(4) τ : C = 100000; lambda = 1e-7; epsilon = 0.001; kernel option = 2.8; kernel =
‘poly’.
(5) Rsb : C = 500000; lambda = 1e-7; epsilon = 0.001; kernel option = 0.12;
kernel = ‘Gaussian’; verbose = 1.

Appendix 3: Families of Functions Used for Each Parameter for

FN Model
For α, polynomial family of degree 3 was used and f1 · · · f12 are
f1 (x1 ) = −0.78629 − 0.00026x1 + 0.00012x21 − 4.5 × 10−5 x31 ;
f2 (x2 ) = −0.0002x2; f3 (x3 ) = −0.00019x3 + 7.6 × 10−7 x23 ;
f4 (x4 ) = −0.00021x4 + 2.29 × 10−7 x24 ;
f5 (x5 ) = −1.7 × 10−5 x25 + 4.93 × 10−7 x35 ;
f6 (x6 ) = −4.7 × 10−6 x26 + 3.72 × 10−8 x36 ;
f7 (x7 ) = 2.58 × 10−4 x7 − 5.3 × 10−6 x27 ;
f8 (x8 ) = 0.03466x8 − 0.0005x28 + 2.4 × 10−6 x28 ;
f9 (x9 ) = −1.7 × 10−7 x9 + 4.38 × 10−11 x29 ;
f10 (x10 ) = 5.44 × 10−5 x10 − 8.7 × 10−7 x210 ;
f11 (x11 ) = −4.2 × 10−5 x11 + 2.25 × 10−7 x211 − 4 × 10−10 x311 ;
f12 (x12 ) = −0.00011x12 + 3.3 × 10−5 x212 − 2.3 × 10−6 x312 .
For β, polynomial family of degree 3 gave the best result and f1 · · · f12 are
f1 (x1 ) = −119.266 + 1.37266x21 − 0.4728x31; f2 (x2 ) = 0.80142x2;
f3 (x3 ) = 0.752408x3 + 0.02248x3 − 0.00108x23;
f4 (x4 ) = = 0.80746x4 + 0.000544x24;
f5 (x5 ) = 1.83579x5 − 0.09053x25 + 0.002752x35;
f6 (x6 ) = 1.72097x6 − 0.02352x26 + 0.000207x36;
f7 (x7 ) = 0.01923x27 − 0.00052x37; f8 (x8 ) = 0;
f9 (x9 ) = 0.000295x9 − 1.6 × 10−6 x29 + 2.67 × 10−10 x39 ;
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

290 A. Khoukhi et al.

f10 (x10 ) = 3.2953x10 − 0.09306x210 + 0.000853x310;

f11 (x11 ) = −0.33375x11 + 0.00175x211 − 3 × 10−6 x311 ;
f12 (x12 ) = −0.5895x12 + 0.070413x212.
For µob , polynomial family of degree 3 gave the best result and f1 · · · f12 are
f1 (x1 ) = −25.5869x1;
f2 (x2 ) = 0.162028x2 − 0.105347x22 + 0.004261x32;
f3 (x3 ) = 0.032553x3 − 0.00362x23; f4 (x4 ) = 0;
f5 (x5 ) = 0; f6 (x6 ) = 0;
f7 (x7 ) = 1.9566x7 − 0.07975x27 + 0.001032x37;
f8 (x8 ) = 0; f9 (x9 ) = −3.9 × 10−7 x29 + 8.62x39 ;
f10 (x10 ) = 1.70833x10 − 0.04799x210 + 0.000443x310;
f11 (x11 ) = −0.13176x11 + 0.000663x211 − 1.1 × 10−6 x311 ;
f12 (x12 ) = 0.152996x12 + 0.01277x212.
For τ , logarithm family gave the best result and f1 · · · f12 are
f1 (x1 ) = −14589.5 − 9.80634 log(x + 2) + 15.5367 log(x + 3);
f2 (x2 ) = 107.4192 log(x2 + 3) − 331.197 log(x2 + 4) + 239.503 log(x2 + 5);
f3 (x3 ) = 77.258 log(x3 + 3) − 251.037 log(x3 + 4) + 188.3491 log(x3 + 5);
f4 (x4 ) = −903.557 log(x4 + 3) + 949.7861 log(x4 + 4);
f5 (x5 ) = 1.94 × 105 log(x5 + 2) − 7.1 × 105 log(x5 + 3)
+ 8.7 × 105 log(x5 + 4) − 349768 log(x5 + 5);
f6 (x6 ) = 3105700 log(x6 + 2) − 10000000 log(x6 + 3)
+ 10858916 log(x6 + 4) − 3895946 log(x6 + 5);
f7 (x7 ) = −2.3 × 107 log(x7 + 2) + 7.59 × 107 log(x7 + 3)
− 8.5 × 107 log(x7 + 4) + 3.13 × 107 log(x7 + 5);
f8 (x8 ) = 1.2 × 104 log(x8 + 2) − 12235.2 log(x8 + 4);
f9 (x9 ) = 1.05 × 109 log(x9 + 2) − 3.2 × 109 log(x9 + 3)
+ 3.16 × 109 log(x9 + 4) − 1.1 × 109 log(x9 + 5);
f10 (x10 ) = − 1.2502 log(x10 + 4); f11 (x) = 0;
f12 (x12 ) = − 12.0186 log(x12 + 2) + 14.2783 log(x12 + 3).
For Rsb , logarithm family gave the best result and f1 · · · f12 are
f1 (x1 ) = 251353 + 3409.847 log(x1 + 2) − 6184.59 log(x1 + 3);
f2 (x2 ) = −17732.5 log(x2 + 2) + 66166.75 log(x2 + 3) − 54857.3 log(x2 + 4);
f3 (x3 ) = 10315.9 log(x3 + 3) − 14498.5 log(x3 + 4);
f4 (x4 ) = −1.3 × 107 log(x4 + 2) + 2.68 × 107 log(x4 + 3) − 1.4 × 103 log(x4 + 4);
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 291

f5 (x5 ) = 2.63 ×106 log(x5 + 2) − 5.9 ×106 log(x5 + 3) + 3.3144 × 106 log(x5 + 4);
f6 (x6 ) = −2.2 × 107 log(x6 + 2) + 4.7 × 107 log(x6 + 3) − 2.5 × 107 log(x6 + 4);
f7 (x7 ) = 1.65 × 106 log(x7 + 3) − 1.71 × 106 log(x7 + 4);
f8 (x8 ) = 2056.883 log(x8 + 4);
f9 (x9 ) = 1.47 × 109 log(x9 + 2) − 2.9 × 109 log(x9 + 3) + 1.48 × 109 log(x9 + 4);
f10 (x10 ) = −5 × 105 log(x10 + 3) + 5.119 × 105 log(x10 + 4); f11 (x11 ) = 0;
f12 (x12 ) = 84618.57 log(x12 + 2) − 246973 log(x12 + 3) + 169459 log(x12 + 4).

References
1. C. Beal, The viscosity of air, water, natural gas, crude oil and its associated gases at
oil field temperature and pressures, Trans. AIME 165 (1946) 94–115.
2. J. Chew and C. A. Jr. Connally, A viscosity correlation for gas–saturated crude oils,
Trans. AIME 216 (1959) 23–25.
3. H. D. Beggs and J. R. Robinson, Estimating the viscosity of crude oil system, J. Pet.
Tech. 9 (1975) 1140–1149.
4. O. Glaso, Generalized pressure-volume temperature correlations, J. Pet. Tech. 32(5)
(1980) 785–795.
5. M. Vazuquez and H. D. Beggs, Correlation for fluid physical property prediction,
J. Pet. Tech. 32(6) (1980) 968–970.
6. E. O. Egbogah and T. Ng. Jack, An improved temperature-viscosity correlation for
crude oil systems, J. Pet. Sc. & Eng. 5 (1990) 197–200.
7. S. A. Khan, M. A. Al-Marhoun, S. O. Duffuaa and S. A. Abu-Khamsin, Development
of viscosity correlations for crude oils, Fifth SPE Middle East Oil Show, Bahrain,
March (1987), pp. 7–10.
8. R. Labedi, Improved correlations for predicting the viscosity of light crudes, J. Pet.
Sc. & Eng. 8 (1992) 221–234.
9. G. E. Jr. Petrosky and F. F. Farshad, Viscosity, correlations for Gulf of mexico crude
oils, Paper SPE 29468 Presented at Production Operations Symposium (Oklahoma
City, OK, U.S.A.), 2–4 April, 1995.
10. R. A. Almehaideb, Improved PVT correlations for UAE crude oils, SPE Middle East
Oil & Gas Show and Conference (Bahrain, March 1997), pp. 15–18.
11. B. Dindoruk and P. G. Christman, PVT properties and viscosity correlations for Gulf
of Mexico oils, SPE Res. Eng. 7(6) (2004) 427–437.
12. A. M. El-Sharkawy, Modeling the properties of crude oil and gas systems using RBF
network, SPE Asia Pacific Oil & Gas Conference (Perth, Australia, October 1998),
pp. 12–14.
13. Ridha B. Gharbi, Adel M. Elsharkawy and Mansour Karkoub, Universal Neural-
Network-Based Model for estimating the PVT properties of crude oil systems, Energy
Fuels 13(2) (1999) 454–458, DOI: 10.1021/ef980143v.
14. N. Varotsis, V. Gaganis, J. Nighswander and P. Guieze, A novel non-iterative method
for the prediction of the PVT behavior of reservoir fluids, SPE Annual Technical
Conference and Exhibition (Houston, Texas, October 1999), pp. 3–6.
15. E. Osman and M. Al-Marhoun, Artificial neural networks models for predicting PVT
properties of oil field brines, 14th SPE Middle East Oil & Gas Show and Conf.
(Bahrain, 12–15 March, 2005).
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

292 A. Khoukhi et al.

16. M. A. Ayoub, D. M. Raja and M. A. Al-Marhoun, Evaluation of below bubble point

viscosity correlations and construction of a new neural network model, SPE Asia
Pacific Oil & Gas Conference and Exhibition (Indonesia, 30 October–1 November,
2007).
17. Y. Hajizadeh, Viscosity prediction of crude oils with genetic algorithms, SPE Latin
American and Caribbean Petroleum Engineering Conference (Argentina, April 2007),
pp. 15–18.
18. S. Dutta and J. P. Gupta, PVT correlations of Indian crude using support vector
regression, Energy Fuels 23(11) (2009) 5483–5490, DOI: 10.1021/ef900518f.
19. F. Farshad, J. L. LeBlanc and J. D. Garber, Empirical PVT correlations for Colombian
crude oils, SPE Fourth Latin American and Caribbean Petroleum Engineering Conf.
(Tobago, April 1996), pp. 23–26.
20. G. E. Petrosky and F. F. Farshad, Pressure volume temperature correlations for gulf
of Mexico crude oils, SPE Res. Eval. & Eng. 1(5) (1998) 416–420.
21. A. A. Al-Shammasi, A review of bubble point pressure and oil formation volume factor
correlations, SPE Res. Eval. & Eng. 4(2) (2001) 146–160.
22. M. A. Al-Marhoun, Evaluation of empirically derived PVT properties for Middle East
crude oils, J. Pet. Sc. & Eng. 42 (2004) 209–221.
23. S. S. Ikiensikimama, J. Madu and L. Dipeolu, Black oil empirical PVT correlations
screening for the Niger delta crude, 30th Annual SPE International Technical Con-
ference and Exhibition (Nigeria, 31 July–2 August, 2006).
24. M. N. Hemmati and R. Kharrat, Evaluation of empirically derived PVT properties
for Middle East crude oils, Scientia Iranica 14(4) (2007) 358–368.
25. J. P. Holman, Experimental Methods for Engineers (Seventh Edition, McGraw-Hill,
2001).
26. M. A. Al-Marhoun, A. A. Abdul Raheem, S. Nizamuddin and S. Shujath Ali, Pre-
diction of crude oil viscosity curve using empirical derived correlations and artificial
intelligence techniques, submitted, J. Pet. Sc. & Eng., 2010.
27. M. A. Oloso, A. Khoukhi, A. Abdulraheem and M. Elshafei, Prediction of crude oil
viscosity and gas/oil ratio curves using advances to neural networks. 2009 SPE/EAGE
Reservoir Characterization and Simulation Conference (Abu Dhabi, UAE, October
2009), pp. 19–21.
28. A. Khoukhi, M. Oloso, A. Abdulraheem and M. Elshafei, Evolutionary neural net-
works for estimating viscosity and gas/oil ratio curves, Proc. of the 21st IASTED Int’l
Conf. Modelling & Simulation MS2010 (Banff, Alberta, Canada, July 15–17, 2010).
29. A. Khoukhi, M. Oloso, A. Abdulraheem and M. Elshafei, Viscosity and gas/oil ratio
curves estimation using advances to neural networks, 2011, 7th International Work-
shop on Systems, Signal Processing and their Applications (WOSSPA) (Tipaza, Alge-
ria, May 9–11, 2011).
30. C. Cortes and V. Vapnik, Support-vector networks, Mach. Learn. 20 (1995) 273–297.
31. P. Schölkopf and A. J. Smola, Learning with Kernels: Support Vector Machines, Reg-
ularization, Optimization, and Beyond (MIT Press, Cambridge, MA, 2002).
32. A. E. El-Sebakhy, A fast and efficient algorithm for multi-class support vector
machines classifier. ICICS2004: 28–30 November, IEEE Computer Society (2004),
pp. 397–412.
33. W. T. A. Littman, Introduction to Support Vector Machines, Machine Learning
Course 536. In Tom Mitchell, Text book: Machine learning, McGraw Hill, 1997.
Department of Computer Science, Rutgers University (The State University of New
Jersey, USA, 2003).
 October 7, 2011 10:10 WSPC/S1469-0268 157-IJCIA
S1469026811003100

Support Vector Regression and Functional Networks 293

34. C. Cristianini and J. Shawe-Taylor, An Introduction to Support Vector Machines

(Cambridge Univ. Press, Cambridge, U.K., 2000).
35. M. S. Mohamad, S. Deris, R. M. Illias, A hybrid of genetic algorithm and support
vector machine for features selection and classification of gene expression microar-
ray, Intl. Journal of Computational Intelligence and Applications IJCIA 5(1) (2005)
91–107.
36. L. Zhang, W. Zhou and L. Jiao, Support vector machines based on the orthogonal
projection kernel of father wavelet, Intl. Journal of Computational Intelligence and
Applications IJCIA 5(3) (2005) 283–303.
37. E. Castillo, Functional networks, Neural Processing Letters 7 (1998) 151–159.
38. E. Castillo, A. Cobo, J. M. Guteirrez and R. E. Pruneda, Functional networks: A new
network-based methodology, Computer-Aided Civil and Infrastructure Engineering
15(2) (2000) 90–106.
39. Available: <http://www.cs.rutgers.edu/∼mlittman/courses/ml03/>.
40. A. Khoukhi and S. Al-Bukhitan, Crude oil PVT properties prediction using hybrid
Genetic-Neuro-Fuzzy techniques, International Journal of Oil Gas and Coal Technol-
ogy 4(1) (2011) 47–63.
41. A. Khoukhi and S. Al-Bukhitan, Data-Driven Genetic-Neuro-Fuzzy systems to crude
oil PVT properties prediction, NAFIPS10 (Toronto, Canada, July 12–14, 2010),
pp. 1–7.

View publication stats