Journal of Petroleum Science and Engineering 147 (2016) 4755

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

journal homepage: www.elsevier.com/locate/petrol

Forecasting density, oil formation volume factor and bubble point

pressure of crude oil systems based on nonlinear system identication
approach
Saeed Salehinia a,n, Yaser Salehinia b, Fatemeh Alimadadi c, Seyed Hossein Sadati b
a
Mechanical Engineering Department, Shahid Rajaee Teacher Training University, Tehran, Iran
b
Mechanical Engineering Department, K. N. Toosi University of Technology, Tehran, Iran
c
Central Training Department of National Iranian Oil Company (NIOC), Tehran, Iran

art ic l e i nf o a b s t r a c t

Article history: Accurate predictions of uid properties, such as density, oil formation volume factor and bubble point
Received 19 September 2015 pressure, are essentials for all reservoir engineering calculations. In this paper, an approach based on
Received in revised form nonlinear system identication modeling; Nonlinear ARX (NARX) and Hammerstein-Wiener (HW) pre-
4 April 2016
dictive model, is proposed for forecasting the pressure/volume/temperature (PVT) properties of crude oil
Accepted 6 May 2016
systems. To this end, two datasets; one containing 168 PVT samples from different Iranian oil reservoirs
Available online 7 May 2016
and other a databank containing 755 data from various geographical locations, were employed to con-
Keywords: struct (i.e. train) and evaluate (i.e. test) the models. Simulation results demonstrate that the proposed
PVT NARX and HW models outperform previously employed methods including three types of articial
Oil formation volume factor
neural networks models (committee machine, multilayer perceptron and radial basis function), two
Density
types of ANFIS models (grid partition and fuzzy c-mean) and several empirical correlations with the
Bubble point pressure
Nonlinear autoregressive exogenous (NARX) smallest prediction error, and that they are reliable models for predicting the oil properties in reservoirs
Hammerstein-Wiener engineering among other soft computing approaches.
& 2016 Elsevier B.V. All rights reserved.

1. Introduction
Obtaining properties of reservoir uid has a very important
application in reservoir engineering computation such as inow
performance calculations, well test analysis, numerical reservoir
simulation, reserve estimates and material balance calculations
(Osman et al., 2001). One essential necessity for all types of pet-
roleum calculations is the PVT properties such as design of uid
handling equipment, reservoir volumetric estimates and de-
termination of hydrocarbon owing properties. Idyllically, in order
to extract desired PVT properties, these properties must be ac-
quired through laboratorial analysis of real measurements/sam-
ples collected from the oil wells location. Most often, though, these
measurements are not fully available, because of one or more of
these reasons: a) Samples collected are not reliable. b) Samples
have not been taken because of cost saving. c) PVT analyses are not
available when needed (Hemmati and Kharrat, 2007). As a result,
many correlations models have been developed to relate these
desired properties (e.g. density and bubble point pressure) to
other measures which are relatively abundant and/or easier to
n
Corresponding author.
E-mail address: saeed.salehinia@gmail.com (S. Salehinia).
measure (e.g. temperature, pressure and API gravity). The geolo-
gical condition is considered important for the development of a
correlation, since the chemical composition of crude oil varies
from region to region (Al-Marhoun, 2004). Therefore, all calcula-
tions will depend on the precision of the correlations used for
predicting the uid properties (Osman et al., 2001).
During the past sixty years, the importance of developing and
utilizing empirical correlations for PVT properties has been re-
cognized by engineers. Consequently, much research has been
conducted in this eld and lead to the development of new em-
pirical correlations. Katz (1942) published a graphical correlation
for predicting oil formation volume factor (Bo) factor. Katz used U.
S. mid-continent crude to develop his correlations. Al-Marhoun
(1985, 2004) correlations consist of equations for estimating
bubble point pressure (Pb), solution gas-oil ratio and Bo for Saudi
Arabia oils. These correlations were developed using 75 bottom-
hole samples from 62 reservoirs in Saudi Arabia. Petrosky and
Farshad (1993) presented correlations for estimating Pb, solution
gas oil ratio, Bo factor and oil compressibility for Gulf of Mexico
oils. The correlations were developed with uid samples taken
from offshore Texas and Louisiana. Khan et al. (1987) used samples
from 75 bottom-hole from 65 Saudi Arabia reservoirs and devel-
oped equations for estimating oil viscosity for Saudi Arabia oils at
above and below the bubble point. Glaso (1980) presented

http://dx.doi.org/10.1016/j.petrol.2016.05.008
0920-4105/& 2016 Elsevier B.V. All rights reserved.
48 S. Salehinia et al. / Journal of Petroleum Science and Engineering 147 (2016) 4755

Nomenclature HW Hammerstein-wiener
M membership function
ANFIS adaptive neural fuzzy interface system NARX nonlinear autoregressive exogenous
ANN articial neural network P pressure (psia)
API oil API gravity Pb bubble point pressure
Bo oil formation volume factor R2 coefcient of determination
Ea average absolute percent relative error Rs solution gas/oil (scf/stb)
Emax maximum absolute percent relative error T temperature (F)
Emin minimum absolute percent relative error o oil specic gravity
Er average percent relative error g gas specic gravity
FCM fuzzy c-mean o density (g/cm3)
GP grid partition

Fig. 1. ANFIS structure.

correlations for estimating Pb, solution gas-oil ratio and Bo for

Fig. 2. Structure of NARX model.
North Sea oils. In other studies such as Hanafy et al. (1997),
Egyptian crude oil were used to estimate correlations for Pb, so-
lution gas-oil ratio, Bo, oil compressibility, oil density and oil
viscosity. Also Standing (1962), using 105 experimental data points
on 22 different oil-gas mixtures from California reservoirs, pro-
posed correlations for estimating Pb, solution gas-oil ratio and Bo
for California oils. In another study, Vazquez and Beggs (1980),
using analyses of 600 laboratory PVT from elds all over the world,
correlations for solution gas-oil ratio, Bo and oil compressibility
were presented.
The main problem of these correlations is that they are not
adequately accurate, since they are usually based on specic re-
gions of oil elds. Therefore, researchers turned to soft computing
methods for a greater accuracy, faster, and easier way of predicting
important properties from available PVT parameters. Another
popular means of obtaining reservoir properties is the method of
Articial Neural Networks (ANNs). Many researchers have in-
vestigated that ANNs are able to solve the PVT correlation pro-
blems in the petroleum industry (Shokir et al., 2004). Therefore, in
recent years researchers have utilized ANNs to predict more ac-
curate PVT correlations (Osman and Al-Marhoun, 2005; Asadisa-
ghandi and Tahmasebi, 2011; Alimadadi et al., 2011; Abedini et al.,
2012; Talebi et al., 2014). Obviously recurrent neural networks like
ANN, because of their intrinsic nonlinearity and computational
simplicity, are natural candidates to approximate a given model for
a predictive purpose such as oil system properties. However, sys-
tem identication approach as one of the widely employed
method for prediction purposes; including such convenient model
forms for prediction purposes like the Nonlinear AutoRegressive
with exogenous inputs (NARX) model and Hammerstein-Wiener
(HW) model has not been employed before, to the authors
knowledge, for crude oil properties forecasting. This study propose
an approach, i. e. NARX and HW modeling for predicting reservoir-
uid PVT properties such as density, Bo and Pb with high accuracy.
 S. Salehinia et al. / Journal of Petroleum Science and Engineering 147 (2016) 4755 49

Fig. 3. Structure of Hammerstein-Wiener model.

Table 1
Ranges of inputs and outputs data used in training of the models.

Inputs Outputs

Pb (psia) Temperature (F) Pressure (psia) Rs (Scf/STB) API o g Bo o (g/cm3)

Maximum 4416 277 6548 1998.4 34.77 0.8807 1.3103 2.3947 0.8798
Minimum 1090 140 1413 255.16 18.91 0.5351 0.7883 1.1680 0.5345
Average 2805.4 222.3 3778.1 899.92 26.87 0.7086 1.0405 1.5908 0.7079

Table 2 brief introduction of ANFIS and then the proposed NARX and HW
Ranges of output data used in validation (calibration) and test of the models. models and their structure in presented. Section 3 provides com-
parative simulation results of the proposed NARX and HW models
Validation (calibration) Test
with ANFIS, ANN and empirical correlations methods. Finally, the
Bo o Bo o conclusions and future work recommendations are given in Sec-
tion 4.
Maximum 2.3532 0.5839 2.1782 0.6165
Minimum 2.0612 0.5345 1.9722 0.5839
Average 2.2101 0.5577 2.0671 0.6022
2. Methodology

Table 3 2.1. Adaptive Neuro-Fuzzy Inference System (ANFIS)

Inputs specication used for each model.
Adaptive Neuro Fuzzy Inference System (ANFIS) was developed
Type of input Type of model PVT
by Jang (1993) to minimize an output error measure and maximize
Input number Properties NARX and ANFIS-FCM HW and ANFIS-GP cost index. ANFIS is a multilayered feed-forward neural network
density Bo based on sugeno fuzzy inference system which has been widely
u1 Pb applied for modeling, predicting and control of stochastic and
u2 T
nonlinear processes. ANFIS concurrently uses fuzzy reasoning and
u3 P
u4 Rs neural network calculation to establish mapping relation from
u5 API input variables to output parameter. Faster convergence, smooth-
u6 o ness, adaptability, smaller size of the training set and optimization
u7 g search space are the main advantages of using ANFIS rather than
neural networks and fuzzy systems. Fig. 1 describe the structure of
an ANFIS model. The rst layer in the ANFIS structure is the fuz-
Table 4
zication layer; the second layer performs the rule base layer; the
Conguration of the HW model.
third layer carry out the normalization of membership functions
Output Inputs nonlinearity Linear block order Output nonlinearity (MFs); the fourth and fth layers are the defuzzication and
summation layers, respectively.
Bo One-dimensional poly- nb = [1 1 1 1 1 1 1] One-dimensional poly-
Design process of ANFIS model consist of two stages: building
nomial of degree 2 nf = [3 3 3 3 3 3 3] nomial of degree 1
and training. The building stage is to dene number of inputs,
nk = [0 0 0 0 0 0 0] number and type of MFs. In this stage the input/output relation is
made into rule patches. This can be achieved by using a number of
Table 5
methods such as grid partitioning (GP), subtractive clustering and
Conguration of the NARX model. fuzzy c-means (FCM) methods (Jang, 1993). In the training stage,
the ANFIS is given the ability to learn from training data by em-
Output Class of estimator Regressors orders ploying the ANN technique to update the parameters of the Ta-
kagiSugeno type inference model. Then, the solutions mapped
Density Wavelet network with 1 unit na = [1]
out onto a Fuzzy Inference System (FIS) can therefore be described
nb = [1 1 1 1 1 1]
in linguistic terms.
nk = [0 0 0 0 0 0]

2.2. Nonlinear ARX model

Moreover, to illustrate the performance of the proposed approach,
its performance will be compared with empirical correlations,
ANFIS and ANN approaches. ANFIS as another commonly used
method for prediction of reservoir oil properties (Olatunji et al.,
2011; Khoukhi, 2012; Zamani et al., 2015) is also employed in this
study for comparison purpose.
The rest of this paper organized as follow. In Section 2, at rst a
NARX models, as a nonlinear black-box model, can be used as
indicative of nonlinear dynamic from a wide variety of behaviors
and have been extensively used in various applications. (Ljung,
2012; Andalib and Atry, 2009; Salehinia et al., 2013). Typically, the
NARX models will be used as black-box structures and have
structure as Eq. (1):
50 S. Salehinia et al. / Journal of Petroleum Science and Engineering 147 (2016) 4755
Fig. 4. Structure of the HW model with highest performance.
y(t 1), ... , y(t na), u(t nk ), ...
y(t ) = f ( )
,u(t nk nb + 1)
where the function f depends on a nite number of previous
inputs u and outputs y. na is the number of past output terms and
nb is the number of past input terms used to predict the current
output. nk is the delay from the input to the output, specied as
the number of samples. As it can be seen from the NARX structure
in Fig. 2, the NARX model uses a parallel combination of nonlinear
and linear blocks. The NARX model utilizes regressors as variables
for nonlinear and linear functions. Regressors are functions of
measured input-output data (Ljung, 2012). The predicted output
y^ (t ) of a nonlinear model at time t is given by the general Eq. (2):
y^ (t ) = F (x(t ))
where x(t ) represents the regresses, F is a nonlinear regresses
command, which is estimated by nonlinearity estimators/classes
(Kumar et al., 2010). As shown in Fig. 2, the command F can in-
clude both linear and nonlinear functions of x(t ). Eq. (3) gives the
description of F.
F (x) = LT (x r ) + d + g (Q (x r ))
The exact form of F(x) depends on type of the nonlinearity
estimator. However, in general, the rst part is the output from
linear function block, d is the output offset, and the third part is
the output from nonlinear function block; where L is coefcient of
the linear function, g is the unit nonlinear command, r is the mean
of regressors x, and Q is the parameter of the nonlinearity esti-
mators (Ljung, 2012). The NARX model computes the output in
two stages: 1) Using current and past input values and past output
data, computes regressors. 2.) Using a combination of nonlinear
and linear function, the nonlinearity estimator block maps the
regressors to the model output.
B1(z) / F1(z) =1 / 1 + 0.008944 z-1 + 0.01702 z-2 + 0.009564 z-3
B2(z) / F2(z) =1 / 1 - 0.02326 z-1 - 0.018 z-2 + 0.05053 z-3
B3(z) / F3(z) =1 / 1 - 0.02494 z-1 - 0.001768 z-2 + 0.05377 z-3
B4(z) / F4(z) =1 / 1 - 0.0008197 z-1 + 0.0006179 z-2 - 0.009482 z-3
B5(z) / F5(z) =1 / 1 - 0.8269 z-1 - 0.5202 z-2 + 0.7 z-3
B6(z) / F6(z) =1 / 1 - 0.9308 z-1 + 0.01357 z-2 + 0.07072 z-3
B7(z) / F7(z) =1 / 1 - 0.02347 z-1 - 0.2675 z-2 - 0.05565 z-3
2.3. Hammerstein-Wiener model
(1)
When the output of a system depends nonlinearly on its inputs,
sometimes it is possible to decompose the input-output relation-
ship into two or more interconnected elements. In this case, the
dynamics can be represented by a linear transfer function and
capture the nonlinearities using nonlinear functions of inputs and
outputs of the linear system. The HW model achieves this con-
guration as a series connection of static nonlinear blocks with a
dynamic linear block as shown in Fig. 3. The HW model can be
used as a black-box model structure because it provides a exible
parameterization for nonlinear models (Ljung 2012).
where B and F are backward shifting operator, and w(t) f(u(t))
is a nonlinear function transforming input data u(t). For ny outputs
(2) and nu inputs, the linear block is a transfer function matrix con-
taining entries:
Bj, i(q)
Fj, i(q) (4)
where j 1,2,,ny and i1,2,,nu.
y(t) h(x(t)) is a nonlinear function that maps the output of the
linear block to the system output. w(t) and x(t) are internal vari-
(3) ables that dene the input and output of the linear block, re-
spectively. Function f is called the input nonlinearity and function
h is called the output nonlinearity. The nonlinearities f and h are
scalar functions, one nonlinear function for each input and output
channel. The linear block setting [nb, nf, nk] sets the order of the
linear transfer function, where nb is the number of zeros plus 1, nf
is the number of poles, and nk is the input delay. The HW model
computes the output y in three stages: 1) Computes w(t) f(u(t))
from the input data. w(t) is an input to the linear transfer function
B/F. 2) Computes the output of the linear block using w(t) and
initial conditions: x(t)(B/F)w(t). 3) Compute the model output by
transforming the output of the linear block x(t) using the nonlinear
function h: y(t)h(x(t)).
 S. Salehinia et al. / Journal of Petroleum Science and Engineering 147 (2016) 4755 51

volume at a specic pressure and temperature. This property is one of the most
important parameters necessary for reservoir calculations and uid mobility. Pb is
one of the key PVT property of reservoir uid, which plays a critical role in almost
all tasks related to reservoir and production engineering. On one hand, obtaining
uid properties of reservoirs is very important for petroleum engineering. On the
other hand, a complete data information of PVT properties is not available for some
reservoirs, and moreover, calculating uid properties in laboratories is time con-
suming and expensive-not to mention human errors. Hence, a soft computing
technique to predict necessary uid properties is the solution which engineers have
been using recently.

3.1. Dataset 1: Bo and density

To design the predictive models for Bo and density, a dataset comprising of 168
data points from 21 different petroleum elds from the south and west reservoirs
of the Iran have been used. In this dataset, there are 8 data points from each eld
with limited range of API gravity. To develop the models, 136 out of 168 PVT data
points was used to train the models. In order to prevent the over tting of the
models during training process, 16 data points was used for validation. The re-
maining 16 data points that were not shown to the models (blind data) were used
to investigate and test the accuracy of the models. These data has not been in-
troduced to the system in the training process. All the PVT data used in this study
were selected from undersaturated petroleum reservoirs. Choosing appropriate
ranges of input and output is a really important step to have well-trained models
and leads to very accurate and reliable results at the end. The ranges of input and
output data used for training of the models are shown in Table 1. Also, the ranges of
output data used for validation (calibration) and test are shown in Table 2.
Finding the best conguration for an ANFIS model must be done through
testing different setting; there is yet no basis to nd most effective conguration
for ANFIS at any dened problem. However, monitoring root mean squared errors
at the different number of epochs for training and checking data is the simplest
way to steer at desired performance. To design an ANFIS model, number and type
of MF have a key role. Extensive search carried out to nd the proper structure and
optimal number of MFs for the best ANFIS model(s). To this end, best model for
density obtained through generating a FIS using FCM with 11 Gaussian MFs for each
input and constant MF for output. Then the obtained FIS used to train an ANFIS
model. Furthermore, best model for Bo obtained through generating a FIS using GP
with 2 MFs for each input and linear MF for the output. By applying various linear
and nonlinear MF, triangular MF for inputs 1, 4 and 7; and generalized-bell MF for
inputs 2, 3, 5 and 6 have displayed better performance. Then the obtained FIS used
for training of an ANFIS model. In order to tune adaptive parameters of ANFIS
models, training of network and optimal distribution of MFs are executed by hybrid
learning algorithm composed of least squares estimates and back propagation error
to perform forward pass and backward pass respectively.
Fig. 5. Structure of the NARX model with highest performance.
In the case of proposed blocks-structured models, also several NARX and HW
models were obtained. Apart from choosing the type of inputs for models, in ad-
Table 6 dition, the combination of various estimator and order were tested too. To speed up
Testing results for prediction of density (o). the process of nding the best models and also decreasing the complexity choosing
the proper ones; best-t as shown in Eq. (5) between model output and real data as
Model R2 Er (%) Ea (%) |Er|min (%) |Er|max (%) a percentage value was monitored. In this equation, y is the measured output, y^ is
the predicted model output, and y is the mean of y. The higher the value is closer to
Hanafy et al. (1997) 0.994  0.2471 0.2580 0.0423 1.1085 100%, the better the model can perform and predict the real data. After in-
Vazquez and Beggs (1980) 0.996 0.1555 0.1555 0.0343 0.2549 vestigating several possibilities, the best result achieved through using a NARX
Ahmed (1985) 0.996 1.0139 1.0139 0.2820 2.3367 model for predicting density and a HW model for Bo.
ANN (Alimadadi et al., 2011) 0.991  0.470 0.479 0.252 0.643
ANFIS-FCM 0.996  0.0646 0.0718 3.48E  3 0.3140 y y^
NARX 0.999 6.60E  4 0.0028 8.36E  7 0.0163 Best fit=(1 )*100
y y (5)

The inputs used for estimating the NARX, HW and ANFIS models are shown in
Table 7 Table 3. The real dataset contains 26 different properties of oil. As it can be seen in
Testing results for prediction of Bo. Table 3, after thorough study of the inputs and their effect on the behavior of the
outputs, the appropriate inputs were selected. The selected inputs for the best
Model R2 Er (%) Ea (%) |Er|min (%) |Er|max (%) NARX, HW and ANFIS models are the same. Tables 4 and 5 summarize the con-
guration, and Figs. 4 and 5 illustrate the structure and properties of linear and
Vazquez and Beggs (1980) 0.996 9.2714 9.2714 9.1052 9.5440 nonlinear blocks of the best attained HW and NARX models, respectively.
Glaso (1980) 0.997  10.8012 10.8012 9.3991 12.2908 To evaluate the performance and accuracy of the selected NARX, HW and ANFIS
Ahmed (1985) 0.996 8.4578 8.4578 7.0791 9.3193 models, the 16 blind data points were fed to the models, and the predicted outputs
ANN (Alimadadi et al., 0.973  1.743 1.743 0.466 2.71 of the models were compared with real data of density and Bo using important
2011) statistical criteria including coefcient of determination (R2), average percent re-
ANFIS-GP 0.999  0.2004 1.2723 0.8766 1.9974 lative error (Er) and average absolute percent relative error (Ea). These statistical
HW 0.999 7.14E  03 0.0137 5.01E  07 0.0989 quality measures are calculated as follows:

n
i = 1 (yireal yimodel )2
R2 = 1
n
3. Results and discussion i = 1 (yireal average(yireal ))2 (6)

Bo is a key factor in terms of economic issues, which is dened as the volume of n

yireal yimodel
100
reservoir oil that would take up one stock-tank barrel plus any dissolved gas at the Er = ( )
reservoir pressure and temperature. Density is dened as the mass of oil per unit of
n i =1 yireal (7)
52 S. Salehinia et al. / Journal of Petroleum Science and Engineering 147 (2016) 4755
Fig. 6. Comparison of the predicted oil properties versus Ea.
100
n
yireal yimodel
Ea = ( )
n i =1 yireal (8)
where yireal is the real/measured value of ith sample, yimodel is the predicted
value of the ith sample by the model, and n represent the total number of data
points.
The empirical correlations of Ahmed (1985), Hanafy et al. (1997) and Vazquez
and Beggs (1980) have been selected for estimating and comparing the results of
density; and correlations of Ahmed (1985), Glaso (1980) and Vazquez and Beggs
(1980) were used for estimating and comparing the results of Bo. Furthermore, the
results also were compared with a committee machine type ANN model that was
trained and evaluated with the same dataset used in this study. The comparison
results are summarized in Tables 6 and 7.
As can be seen from the Tables 6 and 7, performance of NARX and HW models
are outstanding. The results show that the NARX and HW models have predicted
the density and Bo values to the satisfactory extent. It can be seen that the NARX
and HW model results outperforms all ANFIS, ANN models and three most com-
mon published empirical correlations. The proposed NARX and HW models,
achieved the lowest Er, lowest Ea and a high R2 in the prediction of density and Bo
among the empirical correlations, ANN and ANFIS models. The Ea is an important
indicator of the accuracy of the models, which NARX and HW models achieved the
highest accuracy. It can be seen that NARX model shows a 96.6% improvement in
term of Ea over the ANFIS model, 99.4% over the ANN model, and also 98.2% over
best empirical correlation at predicting density. Also, the HW model shows a 98.9%
improvement in term of Ea over the ANFIS model, 99.2% over the ANN model, and
also 99.8% over best empirical correlation at predicting Bo.
Fig. 6 shows a comparison of the Ea and further illustrate the superiority of the
performance of the proposed NARX and HW models for predicting density and Bo
against the results from the ANFIS and ANN models. It is evident from the Fig. 6
that the performance of NARX and HW models is excellent. The results for em-
pirical correlations had to be plotted separately, since their numerical results are
much greater than the NARX, HW and ANFIS models; would have made the results
in Fig. 6 incomparable.
Next, a sensitivity study carried out to illustrate the effect of the number of
training data on the accuracy of the results obtained from NARX, HW and ANFIS
methods. To this end, the training data have been reduced from 136 data in a
multistep test by the amount of 16 data in each step, while the 16 validation data
and 16 test data were remained the same (same as the previous test) at each step.
Fig. 7 demonstrate the results of sensitivity analysis for prediction of density and Bo.
According to the results from Fig. 7, the ANFIS models showed the highest
correlation with the number of training data; as the number of the training data
decreased so was the performance of ANFIS models. On the other hand, the ac-
curacy of NARX model remained in a relatively constant range (less than 4e  3) up
until step ve. Even though the accuracy of NARX model decreased when the
number of training data became less than 72 data points, but even then the pre-
diction accuracy of NARX model was to the satisfactory extend. It can be seen that
the NARX model was able to accurately predict the density using only 24 training
data. The sensitivity analysis of HW model showed that more training data is not
always equal to the greater accuracy, and that the HW model also can achieve high
prediction accuracy using less training data. The NARX and HW models showed
that they can provide satisfactory accuracy even when a few numbers of experi-
mental data available to train with.
Finally, to achieve an unbiased estimation of the model performance for each
oil elds, the dataset can be divided randomly into k subsets of equal size and k-
fold cross-validation performed. In this method, models are built k times, each time
leaving out a single subset as the validation data for testing the model and the
remaining k  1 subsets are used as training data. The k results from the folds, then
averaged to yield a single result. Since our dataset contains 168 data points from 21
different oil elds and each oil eld has 8 data points. Thus, the dataset can be
divided into 21 folds in which each fold contains 8 data (i.e., each fold contains the
data of one oil eld). This way, the prediction performances of the models can be
evaluated for 21 different elds. As stated, the result of a cross-validation test is
presented as a single averaged number presenting the performance of the pro-
posed method over k subsets. However, here the prediction results and CPU time
from each subsets also provided so that the results in more details are available for
comparison. The results of 21-fold cross-validation for density and Bo are shown in
Fig. 8.
It can be seen from cross-validation results in Fig. 8 that the blocks-structured
NARX and HW models with the lowest average of Ea have an excellent predictive
performance of PVT properties. The results demonstrate that the NARX and HW
models estimated 8 data points of the density and Bo in each 21 fold to the sa-
tisfactory extend. Hence, it could be concluded at this point that the NARX and HW
modeling approaches have better and reliable performance compared to the
 S. Salehinia et al. / Journal of Petroleum Science and Engineering 147 (2016) 4755 53

Fig. 7. Sensitivity results for number of training data versus Ea.

Fig. 8. Error results of cross-validation: (a) NARX (b) ANFIS-FCM (c) HW (d) ANFIS-GP.
54 S. Salehinia et al. / Journal of Petroleum Science and Engineering 147 (2016) 4755

Table 8
Ranges of PVT properties of crude oil systems in the databank.

Rs (Scf/STB) g API T (F) Pb (psi)

Maximum 3298.66 3.4445 56.8 360.93 6613.82

Minimum 1.395 0.521 6.0 74.00 58.0152
Average 592.387 1.1156 34.36 207.167 1846.0504

Table 9
Conguration of the developed NARX, HW and ANFIS models for Pb.

Model Class of nonlinear estimator Orders of

NARX Sigmoid network with 40 units Regressors:

na [0]
nb [1 1 1 1]
nk [0 0 0 2]

HW input1 (T): sigmoid network with 20 units Linear block:

input2 (Rs): piecewise linear with 30 units nb [3 3 3 3]
input3 (API): piecewise linear with 30 units nf [4 4 4 4]
input4 ( g ): piecewise linear with 30 units nk [0 0 0 0]
output (Pb): wavelet network with 5 units

ANFIS-GP input1 (T): Difference between two sigmoidal MF MFs: [2 2 2 2]

input2 (Rs): Difference between two sigmoidal
MF
input3 (API): Gaussian MF
input4 ( g ): Difference between two sigmoidal
MF
output (Pb): Linear MF

developed empirical correlations and the two types of ANFIS models presented
here.

3.2. Dataset 2: bubble point pressure

In this subsection, in order to investigate the effectiveness of the proposed

NARX and HW modeling method on other oil elds and as well as for estimating
other PVT properties, the predictive performance of the NARX and HW models
were assessed for prediction of reservoir oil saturation pressure at various condi-
tions. For this purpose, a large databank attained from Talebi et al. (2014) which in
turn they have gathered from nine other studies. The gathered databank includes
755 experimental data of Rs, , T, API, and bubble point pressure (Pb) from a variety
of crude oils (heavy oils to volatile oil) ranges and from various geographical lo-
cations. The PVT data in the databank cover a wide range of conditions, as shown
Table 8.
Talebi et al. (2014) developed two models based on Multilayer Perceptron Fig. 9. Cross plots of developed NARX and HW models.
(MLP) and Radial Basis Function (RBF) neural networks for prediction of Pb, and
considered the Pb to be a function of Rs, g , T and API. At the end, the MLP-ANN and
RBF-ANN models prediction results have been compared with sixteen published Table 10
correlations. Similarly, in this study two models based on NARX and HW methods Prediction results of Pb by NARX, HW and ANFIS models in comparison with the
have been built using Rs, g , T and API as inputs and Pb as output. Similar to the published results in Talebi et al. (2014).
previous section, an ANFIS model also has been built for comparison purposes. The
outlines of the developed NARX, HW and ANFIS models are presented in Table 9. Method Er (%) Ea (%) R2
Moreover, the amount of data for training set and testing set has been selected
equal to the chosen numbers of the RBF-ANN model in Talebi et al. (2014). Ikiensikimama and Ogboja (2009)  555.5 555.5 0.4
Fig. 9 shows cross plots of the developed NARX and HW models predicting Pb. Yi (2000) 94.0 94.0 0.77
In cross plots, a close-tting cloud of points about 45 line for all of the data in- Valko and McCain (2003) 0.1 25.7 0.82
dicate the robustness of the developed models. Table 10 shows the result of sta- Dindoruk and Christman (2004)  2.8 25.6 0.86
tistical quality measures ((Eqs. (6)8)) of NARX and HW models compared with Arabloo et al. (2014) 3.2 18.9 0.86
different approaches for estimation of Pb in the databank. Talebi et al. (2014) (MLP-ANN)  2.73 16.94 0.94
As can be seen from Table 10, the HW model and especially the NARX model Talebi et al. (2014) (RBF-ANN)  3.28 15.53 0.95
shows a very good performance estimating Pb from various geographical locations ANFIS-GP  1.29 16.18 0.92
at various conditions. The superiority of NARX model is also evident from Fig. 9 in NARX  0.79 15.06 0.94
which the cloud points of NARX model is tighter around 45 line than that of HW HW  1.37 32.54 0.86
model. Talebi et al. (2014) reported in their study that the majority of the presented
correlations have high error and could not accurately estimate the experimental
saturation pressures; among published empirical correlations the Ikiensikimama
for prediction of Pb is a promising approach to achieve high prediction accuracy.
and Ogboja (2009) correlation had the worst prediction performance and Arabloo
et al. (2014) correlation had the best result. Table 10 demonstrates that the de-
veloped NARX model has the smallest Er, Ea and also it has a high value of R2 close
4. Conclusions and recommendations
to 1 in comparison to the two types of ANN models and sixteen empirical corre-
lations presented by Talebi et al. (2014). Therefore, using NARX modeling method Obtaining reservoir-uid characteristics is very important and
 S. Salehinia et al. / Journal of Petroleum Science and Engineering 147 (2016) 4755
necessary for all reservoir calculations. Among such reliable
techniques, soft computing is accurate, quick, and cost effective
method for extracting precise values of important properties from
PVT data, as applied in this research. In this study, two blocks-
structured NARX and HW models based on nonlinear system
identication approach was proposed for the prediction of density
(o), Oil FVF (Bo) and bubble point pressure (Pb) of crude oil
systems.
At rst, in order to predict density and Bo, a dataset consist of
168 PVT data points from 21 reservoirs in west and south of Iran
were used for training of the NARX and HW models. A compara-
tive study carried out between the trained NARX and HW models
with three common empirical correlations, two types of ANFIS
models and a committee machine type ANN model. Simulation
results revealed that the NARX and HW models predicted the
density and Bo with the highest accuracy and outperformed the
empirical correlations, ANFIS and ANN models. The numerical
results demonstrate that both the NARX and HW models has the
smallest Ea and the smallest Er compared to other methods.
Moreover, two sensitivity study was conducted to assess the im-
pact of the number of training data on prediction performance and
CPU time. The results of the sensitivity studies showed that the
NARX and HW models can provide satisfactory prediction accuracy
when fewer PVT data are available for training. In addition, a
cross-validation test was performed on the NARX, HW and ANFIS
models to further evaluate their predictive performances over 21
different oil elds. The cross-validation results revealed that the
NARX and HW models predicted the density and Bo from 21 dif-
ferent Iranian reservoirs to the satisfactory extend.
Finally, to assess the efciency of the NARX and HW models for
other oil elds and also estimating other PVT properties, a data-
bank consist of 755 PVT data covering wide ranges of crude oils
was used to evaluate the accuracy of the NARX and HW models
predicting Pb from various geographical locations. The simulation
results demonstrate the superior performance of the NARX model
when compared with two MLP and RBF neural network models, an
ANFIS model and several empirical correlations.
For future work recommendations, we suggest employing the
proposed approach for forecasting other oil PVT properties such as
compressibility and viscosity since NARX and HW showed out-
standing predictive performances. Moreover, we propose that
these two modeling approaches be used in another area like
porosity prediction, history matching, predicting the rock me-
chanics properties, ow regimes and liquid-holdup multiphase
follow since the proposed approach is exible and reliable and has
been used for modeling and prediction purposes in various area
(Andalib and Atry, 2009; Kumar et al., 2010; Salehinia et al., 2013;
Fathi and Ahmad, 2013).
Acknowledgment
The authors extend their thanks to Mr. Mohammad Mahdi
Ghiasi for sharing the PVT databank with us that made the im-
provement of the manuscript possible.
References
Abedini, R., Esfandyari, M., Nezhadmoghadam, A., Rahmanian, B., 2012. The pre-
diction of undersaturated crude oil viscosity: an articial neural network and
55
fuzzy model approach. Pet. Sci. Technol. 30 (19), 20082021.
Ahmed, T., 1985. Compositional Modeling of Tyler and Mission Canyon Formation
Oils with Co2 and Lean Gases. Final report submitted to Montanans on a New
Track Science (MONTS) (Montana National Science Foundation Grant Program).
Alimadadi, F., Fakhri, A., Farooghi, D., Sadati, S.H., 2011. Sadati Using a Committee
Machine with Articial Neural Networks to Predict PVT of Iran Crude Oil. Paper
SPE 141165.
Al-Marhoun, M.A., 2004. Evaluation of empirically derived PVT properties for
Middle East crude oils. J. Pet. Sci. Eng. 42, 209221.
Al-Marhoun, M.A., 1985. Pressure-Volume-Temperature Correlations for Saudi
Crude Oils. Paper SPE 13718.
Andalib, A., Atry, F., 2009. Multi-step ahead forecasts for electricity prices using
NARX: a new approach, a critical analysis of one-step ahead forecasts. Energy
Convers. Manag. 50, 739747.
Arabloo, M., Amooie, M.A., Hemmati-Sarapardeh, A., Ghazanfari, M.H., Mo-
hammadi, A.H., 2014. Application of constrained multi-variable search methods
for prediction of PVT properties of crude oil systems. Fluid Phase Equilib. 363,
121130.
Asadisaghandi, J., Tahmasebi, P., 2011. Comparative evaluation of back-propagation
neural network learning algorithms and empirical correlations for prediction of
oil PVT properties in Iran oilelds. J. Pet. Sci. Eng. 78 (2), 464475.
Dindoruk, B., Christman, P.G., 2004. PVT Properties and Viscosity Correlations for
Gulf of Mexico Oils. SPE Reservoir Evaluation & Engineering 7 (06), 427437.
Fathi, A., Ahmad, M., 2013. Identication of a dynamic model for shape memory
alloy actuator using Hammerstein-Wiener gray box and mutable smart bee
algorithm. Int. J. Intell. Comput. Cybern. 6 (4), 328357.
Glaso, O., 1980. Generalized pressure-volume-temperature correlations. J. Pet.
Technol. 32 (5), 785795.
Hanafy, H.H., Macary, S.M., ElNady, Y.M., Bayomi, A.A., El Batanony, M.H., 1997. A
new approach for predicting the crude oil properties. Paper SPE 37439.
Hemmati, M.N., Kharrat, R., 2007. Evaluation of empirically derived PVT properties
for Middle East crude oils. Sci. Iran. 14 (4), 358368.
Ikiensikimama, S.S., Ogboja, O., 2009. New bubblepoint pressure empirical PVT
correlation. In: Proceedings of Nigeria Annual International Conference and
Exhibition. Society of Petroleum Engineers.
Jang, J.S., 1993. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans.
Syst. Man Cybern. 23 (3), 665685.
Katz, D.L., 1942. Prediction of shrinkage of crude oils. Drill. Prod. 42 (137), 137147.
Khan, S.A., Al-Marhoun, M.A., Duffuaa, S.O., Abu-Khamsin, S.A., 1987. Viscosity
correlations for Saudi Arabian crude oils. Paper SPE 15720.
Khoukhi, A., 2012. Hybrid soft computing systems for reservoir PVT properties
prediction. Comput. Geosci. 44, 109119.
Kumar, P., Potluri, C., Sebastian, A., Chiu, S., Urfer, A., Naidu, D.S., Schoen, M., 2010.
An adaptive multi sensor data fusion with hybrid nonlinear ARX and Wiener-
Hammerstein models for skeletal muscle force estimation. In: ICS'10 Int. Con-
ference on Systems: Part of the 14th WSEAS CSCC Multi Conference. (I), pp.
186191.
Ljung, L., 2012. System Identication Toolbox TM 8 Users Guide. The Math Works,
Inc.
Olatunji, S.O., Selamat, A., Abdul Raheem, A.A., 2011. Predicting correlations prop-
erties of crude oil systems using type-2 fuzzy logic systems. Expert Syst. Appl.
38, 1091110922.
Osman, E.A., Al-Marhoun, M.A., 2005. Articial neural networks models for pre-
dicting PVT properties of oil eld brines. Paper SPE 93765..
Osman, E.A., Abdel-Wahhab, O.A., Al-Marhoun, M.A., 2001. Prediction of Oil PVT
properties using neural networks. Paper SPE 68233.
Petrosky, G.E., Farshad, F.F., 1993. Pressure-volume-temperature correlations for
Gulf of Mexico crude oils. Paper SPE 26644.
Salehinia, Y., Salehinia, S., Naja, F., Sadati, S.H., Shiee, M., 2013. Solving forward
kinematics problem of stewart robot using soft computing. In: RSI/ISM Inter-
national Conference on Robotics and Mechatronics (ICRoM).
Shokir, E.M., El, M., Goda, H.M., Fattah, K.A., Sayyouh, M.H., 2004. Modeling ap-
proach for predicting PVT data. Eng. J. Univ. Qatar 17, 1128.
Standing, M.B., 1962. Oil-system Correlation: Petroleum Production Handbook.
Mcgraw-Hill Book Co., New York City.
Talebi, R., Ghiasi, M.M., Talebi, H., Mohammadyian, M., Zendehboudi, S., Arabloo, M.,
Bahadori, A., 2014. Application of soft computing approaches for modeling
saturation pressure of reservoir oils. J. Nat. Gas Sci. Eng. 20, 815.
Valko, P.P., McCain, W.D., 2003. Reservoir oil bubblepoint pressures revisited; so-
lution gasoil ratios and surface gas specic gravities. Journal of Petroleum
Science and Engineering 37 (3), 153169.
Vazquez, M., Beggs, H.D., 1980. Correlation for uid physical property prediction. J.
Pet. Technol. 32 (6), 968970, SPE-6719-(PA).
Yi, X., 2000. Using wellhead sampling data to predict reservoir saturation pressure.
In: SPE Permian Basin Oil and Gas Recovery Conference. Society of Petroleum
Engineers.
Zamani, H.A., Raee-Taghanaki, S., Karimi, M., Arabloo, M., Dadashi, A., 2015. Im-
plementing ANFIS for prediction of reservoir oil solution gas-oil ratio. J. Nat.
Gas. Sci. Eng. 25, 325334.