Petroleum 2 (2016) 40e48

Contents lists available at ScienceDirect

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

Original article

The most accurate heuristic-based algorithms for estimating

the oil formation volume factor
Mohammad Reza Mahdiani*, Ghazal Kooti
Faculty of Petroleum Engineering, Amirkabir University of Technology (Tehran Polytechnic), Hafez Avenue, Tehran, Iran

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

Article history: There are various types of oils in distinct situations, and it is essential to discover a model for
Received 6 November 2015 estimating their oil formation volume factors which are necessary for studying and simulating the
Received in revised form reservoirs. There are different correlations for estimating this, but most of them have large errors
30 December 2015
(at least in some points) and cannot be tuned for a specific oil. In this paper, using a wide range of
Accepted 30 December 2015
experimental data points, an artificial neural network model (ANN) has been created. In which its
internal parameters (number of hidden layers, number of neurons of each layer and forward or
Keywords:
backward propagation) are optimized by a genetic algorithm to improve the accuracy of the model.
Genetic programming
Neural network
In addition, four genetic programming (GP)-based models have been represented to predict the oil
Modeling formation volume factor In these models, the accuracy and the simplicity of each equation are
surveyed. As well as, the effect of modifying of the internal parameters of the genetic programming
(by using some other values for its nodes or changing the tree depth) on the created model. Finally,
the ANN and GP models are compared with fifteen other models of the most common previously
introduced ones. Results show that the optimized artificial neural network is the most accurate and
genetic programming is the most flexible model, which lets the user set its accuracy and simplicity.
Results also recommend not adding another operator to the basic operators of the genetic
programming.
Copyright © 2016, Southwest Petroleum University. Production and hosting by Elsevier B.V. on
behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction available. Thus, an accurate correlation for different ranges of

Oil formation volume factor (Bo) is one of the PVT properties
of oil, which is the ratio of a specific petroleum volume at the
reservoir condition to its corresponding volume at standard
condition. Bo's value is essential in calculating various parame-
ters such as the depletion rate, oil in place, predicting the future
of the reservoir, optimizing the rate of production and some
other simulation and optimization problems. Distinctive corre-
lations are used whenever the experimental value for Bo of a
specific oil at particular pressure and temperature is not
* Corresponding author.
E-mail address: mahdiani@aut.ac.ir (M.R. Mahdiani).
Peer review under responsibility of Southwest Petroleum University.
Production and Hosting by Elsevier on behalf of KeAi
parameters is necessary. There are numerous correlations for
estimating Bo in the literature, but most of them have poor
prediction for some points even in their applicable range. In
addition to that they can't be tuned for a special oil, and their
accuracy and speed can't be changed base on the need of the
problem. They are not flexible to be used in different problems
and also there is not a comprehensive study to help users to
select one for their studies. Here, one artificial neural network
(ANN), one hybrid of ANN and genetic algorithm, four genetic
programming (GP) methods, and fifteen previously introduced
correlations will be surveyed, and their strengths and weak-
nesses will be discussed. Considering the pressure, reservoirs can
be classified into three categories: below the bubble point
pressure, at bubble point pressure and above the bubble point
pressure. Usually, there is some free gas in the reservoir at
pressures below the bubble point. Above that, Bo changes
because of the oil and gas compressibility. Knowing the value of
Bo at bubble point condition, Bo at higher pressures can be easily

http://dx.doi.org/10.1016/j.petlm.2015.12.001
2405-6561/Copyright © 2016, Southwest Petroleum University. Production and hosting by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open
access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 M.R. Mahdiani, G. Kooti / Petroleum 2 (2016) 40e48
calculated using compressibility factor. Thus, most of the corre-
lations estimate Bo at bubble point pressure [1e3]. The models of
this paper also estimate the Bo at bubble point pressure.
Different correlations have been represented for predicting
the oil formation volume factor. In 1942, Katz [4] using 117
measurement data developed a correlation to predict the oil
formation volume factor using gas gravity, oil gravity, gas oil
ratio, and reservoir pressure and reservoir temperature. In 1947,
Standing [5e8] created a correlation which was the first corre-
lation that needs only four parameters (gas gravity, oil gravity,
and solution gas oil ratio and reservoir temperature) to estimate
the oil formation volume factor. Vasquez and Begg [9](1980)
developed a correlation which was highly dependent on gas
gravity. In 1980, Glaso [10] introduced a correlation for a
methane contained black oil based on gaseoil equilibrium. Al-
Marhoun [11,12] (1988) used the linear regression for Middle
East oil. After that a lot of researchers tried to introduce co-
efficients for previous correlations. Among them are Abdul-
Majeed, 1988 [13], Dokla (1992) [14], Al-Marhoun (1992) [15],
Petrosky (1993) [17], Omar (1993) [18], Kartoamodj 1994 [19],
Farshad 1996 [20]. In more recent years some researchers tried to
introduce completely new correlations. Maybe by adding pa-
rameters or different techniques. Sutton [24] developed a cor-
relation for density and using that estimated Bo. Sebakhy [25]
(2009) used support vector machine modeling Nassar (2013)
[26]considered separator pressure, temperature and its config-
uration (two or three stage) and introduced a correlation for Bo,
In 2014 Sulaimon [27] used a group method of data handling to
create a model.
As it can be seen, there are different correlations for the oil
formation volume factor in the literature. Now, the question is
which one of them is more accurate. There are a lot of studies
that compare these correlations. In 1994, using 195 data points,
Ghetto [28]compared different correlations and suggested the
Vasquez and Begg model for estimating the oil formation volume
factor. In 1997, Almehaideb [22] used the data of 13 oil fields in
UAE to show the performance of common correlations. He sug-
gested Al-Marhoun, Glaso, and Standing as the top three accurate
correlations. In 2012, Godefroy [29] compared various correla-
tions and suggested Glaso and Al-Marhoun to estimate Bo. Un-
fortunately none of the above studies have a comprehensive
investigation on distinctive models, they are old and are not
surveyed the modern models and the new methods for modeling
such as neural network, genetic programming and simulated
annealing programming [46,47] and the database of most of
them is not wide enough. Here, using an extensive database, 15
models of the most common literature correlations in addition to
modern models such as neural network and different genetic
programming models are compared and their strengths and
weaknesses are illustrated.
2. Methodology
As mentioned earlier, here using some artificial neural
network and genetic programming some new models are
created and compared with the most common previously
introduced ones in literature. For this purpose there is need to
have a dataset of the measured oil formation volume factors in
various conditions. This data set, gained from different literature
works and experiment, consists of 160 data points of the Middle
East oil of Al-Marhoun [12] study, 195 points of Ghetto [28] from
Persian gulf, middle east and Africa oils, 41 data point of Glaso
[10] from north sea, 177 points of Katz [4] study 93 points of
Malaysian oil of Omar's study as well as 200 other experimental
data points which were measured of Iranian south oil fields. The
41
Table 1
Range of parameters, used in building the model.
Gas gravity Oil gravity Temperature (
F) Solution gas
oil ratio (SCF/STB)
Minimum 0.612 0.766 100 104
Maximum 1.315 0.895 280 1439
number of all points was 866. From them 395 points that they
were not used in building of any surveyed correlations of this
paper put aside for comparing different correlations and 471
points used for building new models for oil formation volume
factor using artificial neural network and genetic programming.
Artificial neural network (ANN) is a modeling method that is
based on biological neurons. This network is consisted of units
called artificial neurons. These neurons receive inputs and sum
them to produce the output. Each neuron has a weight and thus
the summation is weighted. The sum is passed through a func-
tion called transfer function. And finally the model is created
[30]. ANN has been used widely in the different courses of sci-
ence [31e33].
To create the ANN model, the same data that were used in
creating the genetic programming models is used. Their range is
shown in Table 1 and the property of the neural network model
is shown In Table 2. In this Table, ‘LevenbergeMarquart’ is the
name of a method; for more information see Ref. [34]. As shown
in this table, type of neural network is a forward propagation and
all its details are listed in that. Here using artificial neural
network a model for estimating the oil formation volume factor
created. In next section the created ANN model statistical
properties will be discussed and compared with other Bo
correlations.
Artificial neural network is a good method but it is very
random base and different runs of that, results to different
models. As well as changing its internal parameters such as the
number of hidden layers, the number of neurons of each layer
and having forward or backward propagation can change the
final model and its efficiency. Thus, here ANN is coupled with GA.
GA optimizes the number of hidden layers and the neurons of
each layer to improve the quality of the models in addition to
that it selects the type of propagation (forward or backward). The
basic properties of the artificial neural network are mentioned in
Table 2. The properties of GA are listed in Table 3. For knowing
more about the parameters of this table the interested reader can
refer to [35]. As mentioned before. For that reason in each iter-
ation for every set of inputs (number of hidden layers and
number of neurons of each layer and the type of propagation)
five times the ANN ran and the best model selected. The output
of fitness function of genetic algorithm is average relative error
that should be minimized. Five was selected because it is not so
small that random structure of algorithm affects that and is not
so large that causes the low speed of the algorithm. In this study,
Table 2
The parameters of the artificial neural network model used in this study.
Parameter Value
Type of ANN MLPs, forward
Training LevenbergeMarquart
Performance Mean Square Error
Stop tolerance 1.00E-05
Number of nodes of each layer 8
Number of input layers 4
Number of hidden layers 10
Number of output layers 1
42 M.R. Mahdiani, G. Kooti / Petroleum 2 (2016) 40e48

Table 3 developed for symbolic optimization [36,37]. Despite the fact

The genetic algorithm parameters of the model of this study. that the genetic algorithm uses an array of numbers for opti-
Property name Property value mization, genetic programming uses tree (Fig. 2) representation
Population type Integer vector
[38]. As Fig. 2 shows that a tree consists of nodes that each
Population size 20 contains an operator and terminals which each contains a
Fitness scaling Rank parameter and represents an equation. Using trees is a flexible
Selection function Stochastic uniform method that is used in different courses of science and engi-
Elite count 4
neering such as differential equations [39], chemical reactor
Crossover function Scattered
Crossover fraction 0.8 modeling [40e42], circuit designing, etc. [43]. In the genetic
Mutation function Uniform programming technique, at first an initial population of trees is
Mutation probability 0.05 created. Then in each iteration, the individuals are evaluated and
Migration Forward
some individuals are selected for reproduction. The population
Max generation 100
Max stall generation 50
of selected individuals is increased by cross over and mutation.
Function tolerance 1.00E-06 When an individual with acceptable fitness is found or the
number of iteration passes a specific number, the algorithm stops
[44,45].
GA is changed, not just to save the inputs but also to save the Table 1 Shows the range of parameters used for creating (or
created network. The flowchart of the optimized neural network training) the model. In this range, we are sure that the model is
can be seen in Fig. 1. After running the model it concluded that 10 working well but extrapolating the model can lead to unrealistic
hidden layers, 8 neurons for each layer and forward propagation results. These data are used for creating all the models of this
lead to best results. Finally the created ANN model was study and thus the range of applicability of all the models of this
compared with usual ANN, GP and other correlations. study are as Table 1. In this study, the type of nodes are basic
Now genetic programming models will be created. Genetic mathematical operations {þ, , , /} and the parameters of the
programming is an extension of the genetic algorithm, terminals are {T, gg, go, RS}. Afterward, GP changes the tree
structures to minimize the average relative error of its

Fig. 1. Flowchart of optimized artificial neural network.

 M.R. Mahdiani, G. Kooti / Petroleum 2 (2016) 40e48 43

Fig. 2. Three examples of tree structures.

corresponding equation over the data, the GP stops whenever
the values of the average relative error falls below a specific value
or its changes fall below a defined tolerance. GP has different
parameters and changing them my cause different results. Thus
to be able to repeat this modeling these parameters should be
known. They are listed in Table 4.
In this study GP was run multiple times. In first time,
maximum tree depth was set to 4 for having a simpler equation
(in Fig. 1 the left tree depth is two and the others is three). In fact
if the number of levels exceeded from four, algorithm considered
a penalty for that. Running this led to the equation (1).
purpose this time the value of maximum tree depth was set to 4
and GP was run which resulted in the equation (3).
Bo ¼ 0:83161 þ 0:0012246  T þ 0:00052734  Rs
.
þ 115188:5563  Rs T ðRsþggþ1Þ (3)
Again, similar to pervious part, the value of maximum tree
depth was set to 8 and algorithm was run and the equation (4)
was the result:
.
Bo ¼ 1:1292  0:67343  Rs ðTggÞ2 þ 8:3272e  06  T 2

Bo ¼ 0:60641 þ 22:1992=T þ 5:0918e  07  Rs  T þ 0:00053679  Rs  0:0019306  T

þ 0:0018287  T þ 0:00040493  Rs (1) (4)

The accuracy of these genetic programming models and other
Its accuracy will be discussed later. Then for having a better
different correlations of this part will be discussed in next
accuracy the level of tree branches was set to 8 when the algo-
section.
rithm was run again it resulted in the equation (2).

Bo ¼ 0:9135 þ 0:052341  1=Rs þ 5:3578e  06  Rs  go 3. Results and discussion

þ 8:9699e  07  Rs  T=go þ 0:00088254  T In this part one usual and one optimized artificial neural
þ 0:00027237  Rs network, and four genetic programming models as well as 15
other literature correlations, for estimating oil formation volume
(2)
factor, will be compared and their different statistical properties
Considering the different correlations that were previously will be discussed to help the user to select a correlation for
introduced reveal that most of them are exponential. Thus, for evaluating Bo which is best suited for his studies. The correlations
third run the genetic programming is let to use exponential in are selected from the most common ones and their equations are
addition to its basic operators for its nodes' values. For this listed in Table 5. Each correlation is applicable just in a range of
data in which it has been created. The range of applicability of
each correlation is shown in Table 6. It is clear that for comparing
Table 4
the models, the test data should be in the range of applicability of
The parameters of the GP model used in this study.
all correlations. These data points are selected consciously from
Parameter Value data base and their ranges are showed in Table 7. All these points
Population type Double vector are in the range of all used correlations. As this table shows,
Population size 15 Almehaideb has the simplest form of equation, after that there
Initial population generation Random
are Macary and Batanoney and Standing, El-Banbi, Sulaimon and
Initial score Calculated by fitness function
Fitness scaling Rank Kartoamodjo and Schmidt have the simplest form. Of course,
Selection function Tournament now, using all these equations are easy, by computers, but still in
Mutation function Uniform some cases having a simple form of equation is desired.
Crossover function Scattered (Two parents) To compare the applicability of different algorithms the cross
Type of replacement elitist
Elite count 3
plot of the estimation oil formation volume factor of 395 sample
Crossover probability 0.7 data points for all correlations versus their measured one plotted
Mutation probability 0.3 (Fig. 3). In this figure exact Bois the Bo that is measured in lab-
Probability of changing 0.3 oratory with PVT tests. And estimated Bo is the Bo that is pre-
terminal - nonterminal
dicted by correlations. In Fig. 3, standing (a), Vazquez and Begg
nodes (vice versa) during
mutation (b), Al-Marhoun (d) and (g), Macary and Batanoney (h), Omar
Hybrid No and Todd (i), Petrosky (j), Kartoamodjo and Schmidt (k),
max generation 500 Almehdajeb (m) and sulaimin (o) all have similar prediction, they
Max stall generation 100 had a good estimation for most of the points but for a small range
Function tolerance 1.00E-15
of points they had a poor result. In these correlations, Standing
44 M.R. Mahdiani, G. Kooti / Petroleum 2 (2016) 40e48

Table 5
The equation of different correlations for estimating oil formation volume factor.

Authors Equation

Standing (1947) [6] Bo ¼ 0.972 þ 0.0001472  (Rs  (gg/go)0.5 þ 1.125  T)1.175

Vazquez and
Begg (1980) [9]
Glaso (1980) [10]
Al-Marhoun (1988) [12]
Abdul-Majeed and
Salmon (1988) [13]
Dokla and
Osman (1992) [14]
Al-Marhoun (1992) [15]
Macary and Batanoney
(1992) [16]
Omar and Todd (1993) [18]
Petrosky and
Farshad (1993) [17]
Kartoamodjo and
Schmidt (1994) [19]
Farshad et al. (1996) [20]
Almehaideb (1997) [22]
El-Banbi (2006) [23]
Sulaimon (2014) [27]
Bo ¼ 1 þ 0.0004677  Rs þ 1.751  10 5  ((go/gg)  (T e 60))1.8106  108  (Rs  (go/gg)  (T e 60))
Bo ¼ 1 þ (6.58511 þ 2.91329  Log(Rs  (gg/go)0.526 þ 0.968  T) e 0.27683  (Log(Rs  (gg/go)0.526 þ 0.968  T))2)
Bo ¼ 0.497069 þ 0.000862963  (T þ 460) þ 0.00182594  (Rs0.74239  gg0.323294  go1.20204) þ 0.00000318099 
(Rs0.74239  gg0.323294  go 1.20204)2
Bo ¼ 0.9657876 þ 0.000773  (T þ 460) þ 0.000048141  (Rs1.2  gg0.147  go5.222) þ  0.00000000068987 
(Rs1.2  gg0.147  go5.222)2
Bo ¼ 0.0431935 þ 0.00156667  (T þ 460) þ 0.00139775  (Rs0.77357  gg0.404  go0.882505) - 0.00000380525  (Rs0.773572
 gg0.404020  go0.882605)2
Bo ¼ 1 þ 0.000177342  Rs þ 0.000220163  Rs  (gg/go) þ 0.000004292580  Rs  (1  go)  (T e 60) þ 0.000528707
 (T e 60)
Bo ¼ (1.0031 þ 0.0008  T)  (EXP(0.0004  Rs þ 0.0006  (go/gg)))
Bo ¼ 0.972 þ 0.0001472  (Rs  (gg/go)0.5 þ 1.125  T(1.166 þ0.000762  go/gg 0.0399  gg)
Bo ¼ 1.0113 þ 0.0000702046  (Rs0.3738  (gg0.2914/go0.6265)0.5 þ 0.24626  T0.5371)3.0936
Bo ¼ 0.98496 þ 0.0001  (Rs0.755  (gg0.25/go1.5) þ 0.45  T)1.5
Bo ¼  1 e (>2.65 þ 0.5576  Log(Rs0.5956  (gg0.2369  go1.3282) þ 0.0976  T) - 0.333  (Log(Rs0.5956  (gg0.2369  go1.3282)
þ 0.0976  T))2
Bo ¼ 1.122 þ 0.00000141  Rs  T/go2
Bo ¼ 47.23e8.833  (Rs  ((gg/go)0.5) þ 1.325  (T))0.0092  36.6
Bo ¼ 0.0000009  (Rs  (go/gg) þ Rs  (T þ 460)) þ 1.0367

Table 6
Range of applicability of different correlations for estimating oil formation volume factor.

Correlation Temperature (F) Solution gas oil ratio (SCF/STB) Oil gravity Gas gravity

min max min max min max min max

Standing (1947) [6] 100 258 20 1425 0.72 0.96 0.59 0.95
Vazquez and Begg (1980) [9] 75 294 0 2199 0.74 0.96 0.511 1.35
Glaso (1980) [10] 80 280 90 2637 0.79 0.92 0.65 1.28
Al-Marhoun (1988) [12] 74 240 26 1602 0.80 0.94 0.75 1.37
Abdul-Majeed and Salmon (1988) [13] 190 275 181 2266 0.82 0.89 0.8 1.29
Dokla and Osman (1992) [14] 190 275 181 2266 0.82 0.89 0.8 1.29
Al-Marhoun (1992) [15] 75 300 0 3265 0.75 1.00 0.575 2.252
Macary and Batanoney (1992) 130 290 200 1200 0.82 0.90 0.7 1
Omar and Todd (1993) [18] 125 280 142 1440 0.77 0.89 0.612 1.32
Petrosky and Farshad (1993) [17] 114 288 217 1406 0.80 0.96 0.58 0.85
Kartoamodjo and Schmidt (1994) [19] 75 320 0 2890 0.74 0.97 0.38 1.171
Farshad et al. (1996) [20] 95 260 6 1645 0.80 0.95 0.66 1.7
Almehaideb (1997) [22] 190 306 128 3871 0.79 0.87 0.75 1.12
El-Banbi (2006) [23] 186 312 0 8280 0.74 0.85 0.67 1.14
Sulaimon (2014) [27] 125 280 142 1440 0.77 0.89 0.038 1.315

has the simplest equation and thus is preferred. Dokla and more complicated. This is the advantage of GP that the level of
Osman (f) prediction is not good at all, its r-square is about 0.34 accuracy and complexity of its models can be tuned. In usual GP
which is not acceptable however in small Botheir prediction is just the basic operators are applied. But most oil formation vol-
acceptable. Glaso (c) estimated better than Dokla and Osman in ume factor correlations are exponential. In GP exponential model
general case; but despite the good prediction of Dokla and 1 (r) and 2 (s) it's allowed to use exponential as well as four basic
Osman in small Bo, Glaso did not have a good estimation in any operators. The level of complexity of GP (exponential) 1 is similar
range of Bo. Farshd (l) and E-Banbi (n) just have a good prediction to genetic programming 1 but as shown earlier for both of them
for a small range of data. These correlations predictions are not GP found the same equation. Which means adding exponential
accurate for very small or very large Bo. Abdul-Majeed (e) pre- to GP operators only complicates its search process and does not
diction for each point is more than measured one and the increase its accuracy. Again GP exponential 2 is similar to GP 1
amount of his overestimation in all cases is similar. Thus but in this case adding exponential component to operators have
assigning a coefficient to that could have increased its
performance.
Genetic programming model 1(p) and model 2 (q) are similar Table 7
Range of parameters, for testing different correlations.
and both of them are good but model 2 has a slightly better
estimation. The difference between them is the level of their Gas gravity Oil gravity Temperature (
F) Solution gas
equation complexity. Model 2 is more accurate but its equation is oil ratio (SCF/STB)
also more complicated. Using GP, an equation even more accu- Minimum 0.8 0.8 190 217
rate than GP model 2 can be gained but its equation will be much Maximum 0.85 0.85 240 1200
 M.R. Mahdiani, G. Kooti / Petroleum 2 (2016) 40e48 45

Fig. 3. Regression fitting of different models: (a) standing 1947 [6], (b) Vazequez and Begg 1980 [9], (c) Galso (1980) [10], (d) Al-Marhoun 1988 [12], (e) Abdul-Majeed and
Salmon (1988) [13], (f) Dokal and Osman 1992 [14], (g) Al-Marhoun 1992 [15], (h) Macary and Batanoney (1992) [16], (i) Omar and Todd (1993) [18], (j) Petrosky and Farshd
(1993) [17], (k) Karoamodjo and schmidt (1994) [19], (l) Farshad et al. (1996) [20], (m) Almehaideb (1997) [22], (n) El-Bani (2006) [23], (o) Sulaimon (2014) [27], (p) Genetic
Programming Model 1, (q) Genetic Programming Model 2, (r) Genetic Programming (Exponential) Model 1, (s) Genetic Programming Model 2, (t) Artificial Neural Network, (u)
optimized neural network.
46 M.R. Mahdiani, G. Kooti / Petroleum 2 (2016) 40e48

Fig. 3. (continued).

decreased the accuracy of the model in fact adding exponential
component had made the model 2 too hard for GP to find even
the model that it found without exponential component. In
artificial neural network (t) just two points out of 395 points
have a poor estimation and the R-square of ANN is good. But
some of genetic programming models were more accurate
(comparing r square). Optimized neural network (u) has the best
accuracy. This model prediction is excellent and no point is
highly overestimated or underestimated. Thus its prediction is
reliable. For further investigation of different correlations Fig. 4
compares average, median and maximum relative error of
different correlations. These parameters are of great importance
in comparing the applicability of two different methods [48]. In
addition to r-square in Fig. 4(a) average relative errors of
different correlations are compared. In this figure optimized ANN
has the lowest average relative error, usual ANN and GP models
have a small error. After them, Standing, Al-Marhoun, Kartoa-
modjeb, Almehaideb and Sulaimon have the smallest errors. In
other side, Glasso, Abdulmajeed and Farshad et al. have the
highest average relative errors. Fig. 4(b) compares the median
relative errors of different correlations. As illustrated in this
figure, the correlations that had a higher average relative error
have a higher average median error too, and the correlations that
have a lower average relative error, have a lower median relative
error. The point of this figure is that in all correlations, median
relative error is smaller than average relative error. Fig. 4(c)
 M.R. Mahdiani, G. Kooti / Petroleum 2 (2016) 40e48 47

80 shows the maximum relative error of different correlations. As

Average RelaƟve Error (%)
70 this figure shows, the optimized ANN has the lowest maximum
60
relative error and again the good performance of the new model
50
40 is confirmed. In usual ANN, despite the good performance of its
30 maximum error is high and thus it is not recommended. Genetic
20 programming model 2 has the least maximum error and thus the
10
best accuracy. After that other GP models are good and after
0

Macary and Batanoney…

Petrosky and Farshad…

Kartoamodjo and…
Abdul-Majeed and…

GeneƟc Programming…
GeneƟc Programming…
GeneƟc Programming…
GeneƟc Programming…
OpƟmized ArƟficial…
them Al-Marhoun, standing, Vazquez and Begg, Kartaoamodjo
Glaso (1980)

El-Banbi (2006)
Sulaimon (2014)
Al-Marhoun (1988)

Farshad et al (1996)
Almehaideb (1997)
Dokla and Osman (1992)
Al-Marhoun (1992)
Standing (1947)
Vazquez and Begg (1980)

Omar and Todd (1993)

ArƟficial Neural Network

and sulaimon are the most accurate ones. Finally Fig. 4(d) com-
pares the r-squares of different correlations. As this figure shows
again the optimized artificial neural network has the best per-
formance, after that genetic programming have the best r and
Dokla and Osman and Glaso have the worst ones.
Now after investigating the different correlations it's the time
to recommend the best correlations for different usages. First of
(a)
all the optimized artificial neural network, has the best accuracy.
Usually neural network models just have simple production and
70
summation and thus they are fast. So this model have a good
Median RelaƟve Error (%)

60
efficiency and using that or in a better situation creating this kind
50
of models for each specific oil is highly recommended. Usual
40
30
artificial neural network has a good estimation for most points
20 but because of bad estimation for a small fraction of points its use
10 is not safe. After optimized the ANN, GP model 2 has the least
0 errors and best r. It's recommended for all cases (in its applicable
Macary and Batanoney…

Kartoamodjo and…

GeneƟc Programming…
GeneƟc Programming…
GeneƟc Programming…
GeneƟc Programming…
Petrosky and Farshad…
Abdul-Majeed and…

OpƟmized ArƟficial…
Glaso (1980)

El-Banbi (2006)
Dokla and Osman (1992)

Sulaimon (2014)
Standing (1947)
Vazquez and Begg (1980)

Al-Marhoun (1988)

Omar and Todd (1993)

Farshad et al (1996)
Almehaideb (1997)
Al-Marhoun (1992)

ArƟficial Neural Network

range). But it is a little more complicated. If a simpler equation is

desired GP model 1 is recommended. It should be mentioned
that using exponential in GP models is not recommended at all.
After that Al-Marhoun (1988, 1992) [12,15] models have low
average and median relative errors and they don't have points
with a huge over estimation or underestimation. Thus their
predictions are good. In next positions there are, Kartoamodjo
and Schmidt (1994) [19], Petrosky and Farshad (1993) [17],
(b)
Standing (1947) [6] which have an average relative error about
180 2% and median relative error less than 2% as well as their
Maximum RelaƟve Error (%)

160
maximum relative error is less than 30% (Fig. 4) thus they are
140
120 recommended. Among them, Standing (1947) [6] has the
100 simplest form of equation. Other equations are similar except
80
60
Glaso (1980) [10], Abdul-Majeed and Salmon (1988) [13], Dokla
40 and Osman (1992) [14], Farshad et al. (1996) [20] and El-Banbi
20 (2006) [23] which had a poor performance.
0
Macary and Batanoney…

Petrosky and Farshad…

Kartoamodjo and…

GeneƟc Programming…
GeneƟc Programming…
GeneƟc Programming…
GeneƟc Programming…
Abdul-Majeed and…

OpƟmized ArƟficial…
Glaso (1980)

El-Banbi (2006)
Sulaimon (2014)
Al-Marhoun (1988)

Dokla and Osman (1992)

Al-Marhoun (1992)

Farshad et al (1996)
Almehaideb (1997)
Standing (1947)
Vazquez and Begg (1980)

Omar and Todd (1993)

ArƟficial Neural Network

4. Conclusion

1. Usual neural network is more accurate than most of the other

previously introduced correlations for most data points, but
because of bad estimation for some points it is not safe to use.
2. Optimizing the parameters of artificial neural network with
(c)
optimization algorithms such as genetic algorithm can highly
1.2 improve the accuracy of the model. By using this method
1 there is no need to run the artificial neural network over and
0.8
over to find a good model. This method has the highest
R Square

accuracy.
0.6
3. Genetic programming is the most flexible model for pre-
0.4
dicting different properties such as oil formation volume
0.2
factor. It is very accurate and its level of accuracy and its
0
complexity can be tuned based on the problem. Of course
Macary and Batanoney…

Kartoamodjo and…
Petrosky and Farshad…

GeneƟc Programming…
GeneƟc Programming…
GeneƟc Programming…
Abdul-Majeed and…

GeneƟc Programming…
OpƟmized ArƟficial…
Glaso (1980)

Dokla and Osman (1992)

El-Banbi (2006)
Sulaimon (2014)
Standing (1947)
Vazquez and Begg (1980)

Al-Marhoun (1992)

Omar and Todd (1993)

Farshad et al (1996)
Almehaideb (1997)
Al-Marhoun (1988)

ArƟficial Neural Network

having more accurate model needs a more complicated one.

4. Adding an exponential operator to the genetic programming
node's values, in some cases does not have any effect on the
accuracy of the model and in some cases decreases its

Fig. 4. Statistical properties of different models: (a) Average relative error (%), (b)
(d)
Median relative error (%), (c) Maximum relative error, (d) R square.
48 M.R. Mahdiani, G. Kooti / Petroleum 2 (2016) 40e48
efficiency. In fact adding exponential operator just compli-
cates the searching process of the genetic programming and
is not recommended at all.
5. As a summary, in investigated equations, optimized artificial
neural network and after that GP models (whiteout expo-
nential operator) have the best performance. After them Al-
Marhoun (1988, 1992) [12,15], Kartoamodjo and Schmidt
(1994) [19], Petrosky and Farshad (1993) [17], Standing (1947)
[6] are recommended. Among them Standing (1947) [6] has
the simplest form of equation. Other equations are similar
except Glaso (1980) [10], Abdul-Majeed and Salmon (1988)
[13], Dokla and Osman (1992) [14], Farshad et al. (1996) [20],
El-Banbi (2006) [23] which have a poor performance.
Nomenclature
gO Oil Gravity (fraction)
gg Gas Gravity (fraction)
RS Solution Gas Oil Ratio (SCF/STB)
BO Formation Volume Factor (Reservoir barrel/STB)
T Temperature (oF)
P Pressure (psi)
SI metric conversion factor
SCF * 0.02832s sm
STB * 0.15899 sm
psi * 6870.75 pa
0.56F þ 255.4 K
References
[1] James P. Brill, H. Dale Beggs, Two Phase Flow in Pipes, Gulf Publishing,
1991.
[2] William D. McCain Jr., The Properties of Petroleum Fluid, PennWell, United
States of America, 1990.
[3] W.D. McCain Jr., The Properties of Petroleum Fluids, Penn Wel Booksl,
Tulsa, 1990.
[4] D.L. Katz, Prediction of Shrinkage of crude oils, in: Drill. Prod. & Prac, API,
Dallas, 1942, p. 137.
[5] A.A. Al-Shammasi, A review of bubblepoint pressure and oil formation
volume factor correlations, SPE Reserv. Eval. Eng. 4 (2) (2001) 146e160.
[6] M.B. Standing, A pressure-volume-temperature correlation for mixtures of
California oils and gases, in: Drilling and Production Practice, vol. 275,
1947. Dallas.
[7] M.B. Standing, Oil-System Correlation, McGraw-Hill Book Co, New York
City, 1962.
[8] M.B. Standing, Volumetric and Phase Behavior of Oil Field Hydrocarbon,
Reinhold Publishing Corp., New York City, 1952.
[9] Milton Vazquez, H. Dale Beggs, Correlations for fluid physical property
prediction, J. Petroleum Technol. 32 (06) (1980) 968e970.
[10] O. Glaso, Generalized pressureevolumeetemperature correlations, JPT 32
(5) (1980) 785e795.
[11] M.A. Al-Marhoun, PVT correlations for Saudi crude oils, in: Middle East Oil
Technical Conference & Exhibition, Manama, Bahrein, 1985.
[12] M.A. Al-Marhoun, PVT correlations for middle east crude oils, JPT 40 (5)
(1988) 650e666.
[13] G.H.A. Abdul-Majeed, N.H. Salman, s sdasda, An empirical correlation for
FVF prediction, J. Cdn. Pet. Tech. 27 (6) (1988) 118.
[14] M. Dokla, M. Osman, Correlation of PVT properties for UAE crudes, in:
SPEFE, 1992, p. 41.
[15] M.A. Al-Marhoun, New correlation for formation volume factor of oil and
gas mixtures, J. Cdn. Pet. Tech. (1992) 3e22.
[16] S.M. Macary, M.H. El-Batanoney, Derivation of PVT correlations for the Gulf
of Suez crude oils, in: EGPC 11th Pet. Exp. & Prod. Conf., 1992.
[17] G. Petrosky, F. Farshad, Pressure-volume-temperature correlation for the
Gulf of Mexico, in: SPE Annual Technical Conference and Exhibition,
Houston, 1993.
[18] Development of new modified black oil correlation for Malaysian crudes,
in: SPE Asia Pacific Oil and Gas Conference, Singapore, 1993.
[19] Trijana Kartoatmodjo, Zelimir Schmidt, Large data bank improves crude
physical property correlations, Oil Gas J. 92 (27) (1994) (United States).
[20] Empirical PVT correlations for colombian crude oils, in: SPE Latin American
and Caribbean Petroleum Enginerring Conference, Port of Spain, Trinidad
and Tobago, 1996.
[22] R.A. Almehaideb, Improved PVT correlations for UAE crude oils, in: SPE
Middle East Oil Show, Manama, Bahrain, 1997.
[23] A.H. El-Banbi, K.A. Fattah, M.H. Sayyouh, New modified black-oil PVT
correlations for Gas condensate and volatile oil fluids, in: SPE Annual
Technical Conference and Exhibition, Texas, USA, 2006.
[24] R.P. Sutton, An accurate method for determining oil PVT properties using
the Standing-Katz Gas Z-factor chart, SPE Reserv. Eval. Eng. 11 (2) (2008)
246e266.
[25] E.A. El-Sebakhy, Forecasting PVT properties of crude oil systems based on
support vector machines modeling scheme, J. Pet. Sci. Eng. 34 (1e4) (2009)
25e34.
[26] I.S. Nassar, A.H. El-Banbi, M.H. Sayyouh, Modified black oil PVT properties
correlations for volatile oil and gas condensate reservoirs, in: North Africa
Technical Conference & Exhibition, Cairo, Egypt, 2013.
[27] A.A. Sulaimon, N. Ramli, B.J. Adeyemi, Modified black oil PVT properties
correlations for volatile oil and gas condensate reservoirs, in: SPE Nigerian
Annual International Conferance and Exhibition, Lagos, Nigeria, 2014.
[28] D. Ghetto, G. Marco, V. Marco, Reliability analysis on PVT correlations, in:
SPE European Petroleum Conference, London, 1994.
[29] S. Godefroy, S.H. Khor, D. Emms, Comparison and validation of theoretical
and empirical correlations for black oil reservoir fluid properties, in:
Offshore Technology Conference, Texas, USA, 2012.
[30] J.A. Anderson, An Introduction to Neural Networks, A Bradford Book, 1995.
[31] R. Gosukonda, A.K. Mahapatra, X. Liu, G. Kannan, Application of artificial
neural network to predict Escherichia coli O157:H7 inactivation on beef
surfaces, Food Control (2015) 606e614.
[32] V. Chandwani, V. Agrawal, R. Nagar, Modeling slump of ready mix concrete
using genetic algorithms assisted training of artificial neural networks,
Expert Syst. Appl. (2015) 885e893.
[33] H. Cao, F. Cao, D. Wang, Quantum artificial neural networks with appli-
cations, Inf. Sci. (2015) 1e6.
[34] A. Adly, S.K. Abd-El-Hafiz, Utilizing neural networks in magnetic media
modeling and field computation: a review, J. Adv. Res. 5 (2014) 615e627.
[35] M. Mitchell, An Introduction to Genetic Algorithms (Complex Adaptive
Systems), A Bradford Book, 1998.
[36] E. Shokir, M. El, Dew point pressure model for gas condensate reservoirs
based on genetic programming, Energy Fuels 22 (2008).
[37] J.R. Koza, Genetic Programming: on the Programming of Computers by
Means of Natural Evolution, MIT Press, Cambridge, MA, 1992.
[38] C.R. Reeves, Genetic algorithm for the operations research, J. Comput.
Chem. Eng. 9 (1997) 231e250.
[39] E. Sakamoto, H. Iba, Inferring a system of differential equations for a gene
regulatory network by using genetic programming, in: 14th European
Conference, EuroGP, Seoul, Korea, 2001.
[40] H. Cao, J. Yu, L. Kang, Y. Chen, The Kinetic evolutionary modelling of
complex systems of chemical reactions, Comput. Chem. Eng 23 (2) (1999)
143e151.
[41] B. McKay, M. Willis, G. Barton, Steady-state modelling of chemical process
systems using genetic programming, Comput. Chem. Eng. 21 (4) (1997).
[42] B.D. Al-Anazi, A.A. AlQuraishi, New correlation for Z-factor using genetic
programming technique, in: SPE Oil and Gas India Conference and Exhi-
bition, Mumbai, 2010, p. 9.
[43] E.M. Shokir, M.K. Emera, S.M. Eid, A.A. Abd Wally, A new optimiztion
model for 3D well design, Oil Gas Sci. Technol. 59 (2004) 255e256.
[44] J. Madar, J. Abonyi, F. Szeifert, Genetic programming for system identifi-
cation, in: Intelligent Systems Design and Applications (ISDA 2004) Con-
ference, Budapest, Hungary, 2004.
[45] R. Pearson, Selecting nonlinear model structures for computer control,
J. Process Control 13 (2003).
[46] M.R. Mahdiani, E. Khamehchi, A new method for building proxy models
using simulated annealing, Middle-East J. Scientific Res. (2014) 324e328.
[47] M.R. Mahdiani, E. Khamehchi, R. Soltanmohammadi, B. Azkayi, A new
proxy model, based on meta heuristic algorithms for estimating gas
compressor torque, 11th international industrial conference (2015).
[48] M.R. Mahdiani, E. Khamehchi, Preventing instability phenomenon in gas-
lift optimization, Iran. J. Oil Gas Sci. Technol. (2015) 49e65.