Sie sind auf Seite 1von 10

Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10

Contents lists available at ScienceDirect

Journal of the Taiwan Institute of Chemical Engineers


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

A soft computing approach for the determination of crude oil viscosity:


Light and intermediate crude oil systems
Abdolhossein Hemmati-Sarapardeh a,∗, Babak Aminshahidy a, Amin Pajouhandeh b,
Seyed Hamidreza Yousefi a, Seyed Arman Hosseini-Kaldozakh a
a
Department of Petroleum Engineering, Amirkabir University of Technology, Tehran, Iran
b
Department of Petroleum Engineering, Shahid Bahonar University of Kerman, Kerman, Iran

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

Article history: Crude oil viscosity is a key property needed for petroleum engineering analysis such as evaluation of fluid
Received 8 May 2015 flow in porous media, reservoir performance, reservoir simulation, etc. This property is traditionally mea-
Revised 8 July 2015
sured through expensive and time consuming laboratory measurements. In this communication, about 1500
Accepted 10 July 2015
dead oil viscosity data points of light and intermediate crude oil systems from various geological locations
Available online 28 August 2015
have been collected. Afterward, a soft computing approach, namely least square support vector machine
Keywords: (LSSVM), has been utilized to develop two distinct viscosity models for temperatures below and above
Crude oil 313.15 K. The parameters of these models have been optimized using coupled simulated annealing (CSA)
Viscosity optimization tool. The results of this study indicated that the developed models can predict dead oil viscos-
Least square support vector machine ity at all temperatures and oil API gravities with enough accuracy. In addition, statistical and graphical error
Temperature, Oil API gravity analyses illustrated that the proposed CSA-LSSCM models outperform all of pre-existing models. Besides, the
relevancy factor showed that oil API gravity has the greatest effect on dead oil viscosity. Finally, the Leverage
approach demonstrated that the proposed models are statistically valid and acceptable, and only 2% of the
data points may be regarded as the probable outliers.
© 2015 Taiwan Institute of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

1. Introduction Moreover, viscosity data at other temperatures for production equip-


ment and pipelines design, and for planning thermal enhanced oil re-
Crude oil consumption is projected to increase year after year. covery methods is required. Due to these problems, developing pre-
Problems of crude oil consumption in the next decades will revolve dictive models for estimation of crude oil viscosity have been much
around two main factors consisting of population and the increas- attended.
ing use in developing countries. For this reason, accurate and reliable Several correlations have been proposed to predict crude oil vis-
oil properties are required. One of these properties is viscosity which cosity. These correlations are generally categorized into two classes,
is defined as the internal friction of the fluid to flow [1–4]. Reser- depending on the input variables. The first type refers to black oil type
voir oil viscosity is a key property needed for petroleum engineering correlations which use oilfield data, such as oil API gravity, reservoir
analysis such as evaluation of fluid flow in porous media, reservoir temperature, solution gas–oil ratio, saturation pressure, and pressure
performance, reservoir simulation, well testing, and design of pro- to predict oil viscosity. The second type is empirical and/or semi-
duction facilities and transport equipment [5–14]. Crude oil viscosity empirical correlations which use reservoir fluid composition, nor-
is a function of several thermodynamic and physical properties such mal boiling point, pour point temperature, critical temperature, mo-
as type and the nature of its chemical composition, pressure, bub- lar mass, and acentric factor of components [1,13,14,19,21,22].
ble point pressure, temperature, oil specific gravity, gas gravity, and In the past decades, several empirical and semi-empirical correla-
gas solubility [1,2,15,16]. Traditionally, this property is measured ex- tions have been derived mostly from corresponding state equations
perimentally on subsurface or surface (recombined) samples at the to predict crude oil viscosity. Most of presented correlations were de-
reservoir temperature and pressure. Laboratory measurements of this veloped for a given region. Application of these correlations to crude
property are always money and time consuming [2,10,11,15,17–20]. oils of different sources leads to large errors. This difference is at-
tributed to the difference in the nature of crude oil chemical compo-
∗ sition. Above the bubble point pressure, increasing pressure causes
Corresponding author. Tel.: +989132437785.
E-mail address: aut.hemmati@gmail.com, hossein.hemmati1368@yahoo.com, an increase in crude oil viscosity. However, below the bubble point,
aut.hemmati@aut.ac.ir (A. Hemmati-Sarapardeh). increasing pressure causes a decrease in oil viscosity. Crude oil has its

http://dx.doi.org/10.1016/j.jtice.2015.07.017
1876-1070/© 2015 Taiwan Institute of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
2 A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10

Table 1
The origin and PVT data ranges used in dead oil viscosity models.

Author Source of data T, K API μod , cP

Beggs and Robinson (1975) [27] – 294.2–419.2 16–58 –


Glasø (1980) [28] North Sea 283.1–422.0 20–48 0.60–39
Labedi (1982) [31] Nigeria and Angola 313.1–378.1 25–45.5 0.72–21.15
Kaye (1985) [29] Offshore California 334.8–412.0 7–41 –
Al-Khafaji et al. (1987) [30] – 288.7–422.0 15–51 –
Egbogah and Ng (1990) [9] – 288.1–353.1 5–58 –
Labedi (1992) [32] Libya 310.9–425.4 32–48 0.66–4.79
Kartoatmodjo and Schmidt (1994) [33] Worldwide 299.8–433.1 14–59 0.50–586
Petrosky (1995) [34] Gulf of Mexico 318.7–415.3 25–46 0.72–10.25
Bennison (1998) [35] North Sea 277.0–422.0 11–20 6.40–8396
Elsharkawy and Alikhan (1999) [36] Middle East 310.9–422.0 20–48 0.60–33.7
Hossain et al. (2005) [38] Worldwide 273.1–374.8 7–22 12–451
Naseri et al. (2005) [13] Iran 313.7–420.9 17–44 0.75–54
Hemmati-Sarapardeh et al. (2013) [5] Iran 283.1–416.5 17–44 0.39–70
This study Worldwide 273.1–533.1 20–50 0.34–1393

highest viscosity value at atmospheric pressure. Thus, based on pres- reservoir temperature [13]. Very recently, Hemmati-Sarapardeh et al.
sure, these correlations can be classified into three categories: dead presented a simple correlation for dead oil viscosity using 120 dead
oil (stock tank), saturated (below and at bubble point) and under- oil viscosity data points [5]. Afterward, Hemmati-Sarapardeh et al.
saturated (above bubble point) [1,21,23]. used the same data bank and developed an intelligent model for pre-
The evaluation of dead oil viscosity is an important step in the de- diction of crude oil viscosity [7]. Some of the models including Kaye
sign of various operations. Previous studies have demonstrated that [29], Al-Khafaji et al. [30], Egbogah and Ng [9], Kartoatmodjo and
correlations developed for dead oil viscosity cannot predict dead oil Schmidt [33] can be used to estimate all API range of crude oil (heavy
viscosity with enough accuracy. Moreover, dead oil viscosity is the to light crude oils). In 1985, Kaye derived a correlation for the viscos-
input parameter for saturated oil viscosity correlations; therefore, in- ity of California dead oil [29]. In 1987, Al-Khafaji et al. presented a vis-
accurate estimation of dead oil viscosity leads to inaccurate predic- cosity correlation for dead oil by modifying Beal correlation [30]. In
tion of saturated oil viscosity as well as under-saturated oil viscos- 1988, Egbogah and Ng developed two different correlations to com-
ity. Hence, seeking for a more accurate dead oil viscosity model is pute dead oil viscosity. In 1994, Kartoatmodjo and Schmidt modified
important. Glasø correlation [33].
Literature search and review have shown that most of dead oil Most of previously published models have been developed based
viscosity correlations use oil API gravity and temperature for viscos- on a specific region and cannot be used globally for prediction of dead
ity prediction, while some of authors correlated viscosity to molar oil viscosity, because characteristics of fluids are different in each re-
mass, normal boiling point, critical temperature, and acentric fac- gion. In addition, most of these correlations have been derived based
tor [24,25]. The most popular empirical models which are used in on limited data points and limited range of parameters. The main pur-
petroleum engineering calculations for predicting dead oil viscos- pose of this study is to propose a universal model for accurate pre-
ity are those ones developed by Beal [26], Beggs and Robinson [27], diction of dead oil viscosity of light and intermediate crude oil sys-
Glasø [28], Kaye [29], Al-Khafaji et al. [30], Egbogah and Ng [9], Labedi tems (API ≥ 20) as a function of temperature and oil API gravity. It
[31,32], Kartoatmodjo and Schmidt [33], Petrosky [34], Bennison [35], should be pointed out that heavy and extra-heavy crude oil systems
Elsharkawy and Alikhan [36], Bergman [37], Hossain et al. [38], Naseri (API < 20) cannot be modeled using only temperature and API, as
et al. [13] and Hemmati-Sarapardeh et al. [5]. The ranges and ori- these crude oils contain a large amount of heavy components such as
gin of data used by these authors to develop their correlations are asphaltenes and resins, which control their rheological behaviors. In
listed in Table 1. These models can be categorized into two sections: this study, a novel soft computing modeling approach, namely least
1—models developed for heavy and extra heavy crude oil systems square support vector machine (LSSVM), is employed to model dead
(API < 20) 2—models developed for intermediate and light crude oil oil viscosity of light and intermediate crude oil systems. The param-
systems (API ≥ 20). Bennison [35] and Hossain et al. [38] correla- eters of the model are optimized using a global optimizer, namely
tions are identified for the first group. In 1998, Bennison used viscos- coupled simulated annealing (CSA). In summary, the following are
ity data to derive a new dead oil viscosity correlation [35]. In 2005, the main goals of this study:
Hossain proposed an empirical correlation for dead oil, which is ap-
plicable for the heavy oils with API gravity ranging from 10 to 22.3 1. Providing a large data bank of dead oil viscosity from various geo-
[38]. Beggs and Robinson [27], Glasø [28], Labedi [31,32], Petrosky logical locations, covering a wide range of temperature and oil API
[34], Elsharkawy and Alikhan [36], Naseri et al. [13], and Hemmati- gravity.
Sarapardeh et al. [5] correlations have been proposed to predict the 2. Developing a robust and accurate model for prediction of dead oil
viscosity of intermediate and light crude oils. In 1975, Beggs and viscosity as a function of temperature and API, using CSA-LSSVM
Robinson published a correlation for prediction of dead oil viscosity. modeling procedure.
The correlation developed based on analyzing 460 dead oil viscosity 3. Comparing the results obtained by the proposed CSA-LSSVM
measurements [27]. In 1980, Glasø presented another correlation for model with those obtained by pre-existing correlations through
predicting dead oil viscosity. Glasø used the data from six North Sea statistical and graphical error analyses.
oil samples [28]. In 1982, Labedi derived a correlation for the viscos- 4. Checking the validity of the proposed model to see whether the
ity of Nigeria and Angola [31]. In 1992, Labedi published a new cor- model can capture the physically expected trends with variation
relation for dead oil viscosity of Libya [32]. In 1999, Elsharkawy and of input parameters.
Alikhan developed empirical correlations for estimating dead oil vis- 5. Investigating the relative effect of each input parameter on dead
cosity of Middle East crude oils [36]. In 2005, Naseri et al. proposed oil viscosity by means of relevancy factor.
a correlation for prediction of Iranian dead oil viscosity. Naseri’s 6. Identifying the probable suspected data and applicability domain
correlation is considered to be a function of oil API gravity and of the applied CSA-LSSVM model.
A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10 3

2. Database set of equations [55,56]. The cost function of the least square support
vector machine (LSSVM) is determined as:
Dead oil viscosity is assumed to be a function of two variables:
1 T 1 
N
reservoir oil gravity (API) and reservoir temperature (T). Cost function = w w+ γ e2k (7)
2 2
μod = f (API, T ) (1) k=1

Subjected to the following constraint:


In this study, almost all experimental dead oil viscosity data,
which are available in open literature sources, have been used yk = wT ϕ( xk ) + b + ek (8)
[5,6,13,14,21,23,28,38–45]. As can be seen, these data are from vari- where γ and ek are tuning parameter in LSSVM method and the vari-
ous geological locations and cover a wide range of temperature and able error, respectively. The Lagrangian for this problem is as follows:
API. These data points include oil API gravity, reservoir temperature,
and dead oil viscosity. The PVT data ranges used in this study are sum-
1  2 
N N
marized in Table 1. 1
L(w, b, e, a) = wT w + γ ek − ak (wT ϕ(xk ) + b + ek − yk )
2 2
k=1 k=1
3. Model development
(9)
Support vector machine (SVM) methodology is a tool for both clas- where ak is Lagrangian multipliers. In order to solve the problem, the
sification and regression analysis [46,47] and recently has been ap- derivatives of Eq. (9) should be equated to zero. Thus, the following
plied in several fields [48–53]. According to SVM primary formula- equations are obtained:
tions, any function f(x) can be written as follows [54]: ⎧
⎪ ∂L N
f (x) = w T
ϕ(x) + b ⎪
⎪ = 0 ⇒ w = a ϕ( x )
(2) ⎪
⎪ ∂w


k k
where w and ϕ(x) show the transposed output layer vector and the
T ⎪
⎪ 
k=1
⎪ ∂ L
⎨ =0⇒
N
kernel function, respectively, and b is the bias. x is the input of the ak = 0
model which has a dimension of N × n, where N and n represent the
∂b (10)

⎪ ∂L
k=1
number of data points and number of input variables, respectively. ⎪
⎪ = 0 ⇒ a = γ ek , k = 1, 2, . . . , N

⎪ ∂ ek
Vapnik used minimization of the following cost function to calculate ⎪

k
w and b [54]: ⎪
⎪ ∂L
⎩ = 0 ⇒ wT ϕ(xk ) + b + ek − yk = 0 k = 1, 2, . . . , N
 N ∂ ak
1 T
Cost function = w +c ( ξk − ξk∗ ) (3) where γ is a tuning parameter of LSSVM. As can be found in
2
k=1 Eq. (10), there are 2N + 2 equations and 2N + 2 unknown parameters
Subjected to the below constraints: ( ak , ek , w and b). Therefore, the parameters of LSSVM are achieved
⎧ by solving the system of equations depicted in Eq. (10). In this work,
⎨yk − w ϕ( xk ) − b ≤ ε + ξk , k = 1, 2, 3, . . . , N
T
the radial basis function (RBF) Kernel was employed, which is pre-

w ϕ( xk ) + b − yk ≤ ε + ξk ,
T
k = 1, 2, 3, . . . , N (4) sented as below:
⎩ ∗
ξk , ξk ≥ 0, k = 1, 2, 3, . . . , N K ( x, xk ) = exp (−|| xk − x || / σ 2 )
2
(11)
where xk , yk and ε are the kth data input, kth data output, and fixed where σ 2 is the other tuning parameter. Hence, in LSSVM with RBF
precision of the function approximation, respectively. ξk and ξk∗ stand kernel function two tuning parameters exist, which can be obtained
for slack variables, which should be used to determine the allowed by minimizing deviation of the experimental data from the predicted
margin of error. c is considered as the tuning parameter of the SVM. values. Normally, minimization of mean square error is considered as
In order to minimize the cost function, the Lagrangian of this problem the objective function to find the tuning parameters of LSSVM model.
should be used as follows [54]: This error is defined as follows:
n
1 
N
( Orep./ pred.i − Oexp .i )2
L ( a, a ) = − ( ak − a∗k )( al − a∗l )K ( xk − xl )
∗ i=1
MSE = (12)
2 n
k,l=1
where O is the output, subscripts rep./pred. and exp. denote the rep-

N 
N
resented/predicted and experimental values, respectively, and n ex-
− ε ( ak − a∗k )+ (
yk ak − a∗k ) (5)
presses the number of data points. In this study, coupled simulated
k=1 k=1
annealing has been used to optimize the parameters of the devel-

N oped LSSVM model. More details about this optimization technique
( ak − a∗k ) = 0, ak , a∗k ∈ [0, c] (5a) can be found elsewhere [57]. A schematic flowchart of the applied
k=1 CSA-LSSVM strategy is demonstrated in Fig. 1.

K ( x k − xl ) = ϕ ( x k ) ϕ ( xl ) , k = 1, 2, . . . , N
T
(5b) 4. Performance evaluation
where αk and αk∗
are Lagrangian multipliers. Consequently, the final
form of the SVM is obtained as follows: Statistical and graphical error analyses can be used to evaluate the
performance of the proposed models.

N
f (x) = ( ak − a∗k )K ( x − xk ) + b (6)
4.1. Statistical error analysis
k,l=1

SVM are usually solved by finding solutions to quadratic program- These statistical parameters are normally used to quantitatively
ming issues with linear inequality constraints. For this reason, in evaluate the performance of a model: mean percentage error (MPE),
1999, Suykens and Vandewalle [55,56] proposed a least square mod- mean absolute percentage error (MAPE), root mean square error
ification to the original SVM (LSSVM) in order to improve the SVM (RMSE), and standard deviation (SD). Formulation of these parame-
method. In LSSVM method, solution is obtained by solving a linear ters is given below.
4 A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10

Input
data

Random division of data


into training and testing

Implement Coupled Simulated Training Testing


Annealing (CSA) data data

Select model features Employ feature subset


(σ2,γ) (σ 2,γ)

Construct dead oil


No viscosity model

Meet stopping Model evaluation using both


criterion? training and testing data

Yes
Optimum model features Retrain the LSSVM model
(σ 2,γ) obtained using the optimum features

Final LSSVM
model

Fig. 1. A typical flowchart for the applied CSA-LSSVM algorithm in this study.

1. Mean percentage error (MPE). 4.2. Graphical analysis

1
n
Graphical tools can be used to check the accuracy and validity of
Er = Ei (13)
n a model. In this study, error distribution curve and cumulative fre-
i=1
quency plot are used.
where n and Ei are the total number of experimental data and
the relative deviation of an estimated value from the correspond- 1. Error distribution curve: it is a tool to quantify error distribution
ing experimental value (Ei is expressed as percent relative error), around the zero error line to designate if the model has an error
respectively: trend or not.
 2. Cumulative frequency plot: in this technique, the proportion of
(μ)exp . − (μ)est. the data points having absolute percent relative errors below dif-
Ei = × 100 ⇒ i = 1, 2, 3, . . . , n (14)
(μ)exp . ferent ascending values is plotted versus the absolute percent rel-
ative error in a cumulative manner for each model on the same
2. Mean absolute percentage error (MAPE). figure to facilitate the visual comparison.

1
n
Ea = |Ei | (15) 5. Results and discussion
n
i=1
In this study, viscosity of dead oil is predicted by a robust soft com-
3. Root mean square error (RMSE): puting approach. Selecting appropriate input parameters for model

development is a key factor. Similar to most of previously published
1  2
n
μiexp . − μiest. correlations, temperature and oil API gravity were selected as the
RMSE = (16)
n input parameters of the proposed model. Many attempts were un-
i=1
dertaken to develop a single model for the prediction of all viscos-
4. Standard deviation (SD). ity data; however, the proposed models were not able to satisfac-

2
torily predict dead oil viscosity of all data set. Afterward, the data
1 
n
μiexp . − μiest. bank were divided into two subsections; viscosity data at tempera-
SD = (17) tures below 313.15 K (40 °C) and viscosity data at temperatures above
n−1 μiexp .
i=1 313.15 K (40 °C). Therefore, two distinct models were developed for
A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10 5

Table 2 110
Statistical parameters of the proposed models for the determination of viscosity.

Mean Absolute Percentage Error, %


Statistical parameter 90

Below 313.15 K Above 313.15 K All data


70
Training set
MAPE% 17.92 16.56
50
MPE% –3.57 1.83
RMSE, cP 29.26 8.80
SD 0.24 0.21 30
Number of data 746 451
Test set
10
MAPE% 16.61 16.56
MPE% –1.27 –2.08
RMSE, cP 35.17 8.22
SD 0.23 0.22
Number of data 187 113
Total
MAPE% 17.66 16.38 17.17
MPE% –3.11 –1.30 –2.43
RMSE, cP 30.54 8.35 24.65 Fig. 2. Mean absolute percentage error of the models.
SD 0.24 0.22 0.23
Number of data 933 564 1497
70

temperatures below 313.15 K (Model 1) and above 313.15 K (Model


2). To achieve the most efficient and accurate models, the data points 50

RMSE, cP
in each region were randomly distributed into two subsets as train-
ing and testing sets. In each region, the training set was used to con-
struct the model, to achieve the optimal model structure, and to ed-
30
ucate existing physical rules in the system. The testing set was used
to evaluate the performance of the developed model in terms of ac-
curacy and reliability. 80% of the data points were used for model
development and the remaining 20% were used as testing set. In dis- 10
tribution of data points into training and testing sets, several distribu-
tions were done to escape from local accumulation of the data points
in the feasible region of the problem. It should be noted here that,
933 of the viscosity data points have a temperature below 313.15 K,
while 565 of the data points are for crude oil systems with temper-
atures above 313.15 K. The values of σ 2 and γ for Model 1 (tem-
perature below 313.15 K) were obtained 0.6807 and 4469.1858, re-
Fig. 3. Root mean square error of the models.
spectively. As pointed out earlier, coupled simulated annealing was
used to find the optimal model parameters. For Model 2 (temperature
above 313.15 K), the values of σ 2 and γ were obtained 22.8155 and Table 3
Statistical parameters of the proposed model and previously published correlations
950.1749, respectively. It should be noted that it is not necessary to for the determination of viscosity.
develop a distinct model for higher temperatures (for example, tem-
Author MPE (%) MAPE (%) RMSE, cP SD
peratures higher than 500 K) as viscosity is not strongly affected by
temperature at very high temperatures. Model 1 predicts dead oil vis- Beggs and Robinson –1643.47 1651.05 595573.92 444.13
cosity data with an MAPE of 17.66%, while Models 2 can predict vis- Glasø 99.97 99.97 69.29 1.00
cosity data of dead oils with an MAPE of 16.38 %, which shows that Labedi—Libya 16.08 36.86 61.52 0.61
Labedi—Nigeria and Angola –17.96 41.49 57.82 0.53
Model 2 is slightly more accurate than Model 1. On the other hand,
Egbogah and Ng 17.64 26.42 42.58 0.36
RMSE (cP) obtained by Model 1 (30.54 cP) is much higher than that Kaye 21.08 32.78 543.62 0.55
of Model 2 (8.35 cP). For a given crude oil system, viscosity data at Al-Khafaji 5.18 21.55 54.96 0.29
lower temperatures have very large values compared to those of high Petrosky 25.60 30.38 59.93 0.37
temperatures. As it is known, RMSE (Eq. 16) is strongly affected by Kartoatmodjo and Schmidt 15.12 22.81 48.69 0.29
Bennison 84.80 88.84 40.32 0.91
the values of experimental data and predicted values; therefore, as Elsharkawy –6.93 23.08 321.41 0.40
expected, the value of RMSE (cP) for Model 1 is higher than that of Hossain 85.65 87.68 39.73 0.90
Model 2. The other statistical quality measures of the proposed mod- Naseri 29.13 32.24 55.68 0.38
els are summarized in Table 2. Hemmati-Sarapardeh 31.09 32.95 56.68 0.38
This study –2.43 17.17 24.65 0.23

5.1. Validation of the developed models compared


to pre-existing models
it is found that Glasø [28], Hossain [38], and Bennison [35] correla-
To pursue the objective of this study, performance of the devel- tions predict dead oil viscosity with an MAPE more than 80%. Also,
oped models were compared to pre-existing correlations by means Beggs and Robinson [27] correlation predicts dead oil viscosity with
of statistical and graphical error analyses. Generally speaking, MAPE an MAPE of 1651%, which was not shown in this figure. Fig. 3 illus-
and RMSE are the most relevant statistical quality measures. MAPE trates RMSE (cP) of different models in predicting viscosity of dead
of different models have been sketched in Fig. 2, demonstrating that oils. As can be seen, the CSA-LSSVM models predict dead oil viscosity
the proposed models outperform all of pre-existing models. Besides, with the smallest RMSE (cP), showing their excellent performance.
6 A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10

300 300

200 200

Relative Error, %

Relative Error, %
100 100

0 0

-100 -100

-200 Al-Khafaji -200 Kartoatmodjo and Schmidt

-300 -300
20 30 API 40 50 20 30 API 40 50

300 300

200 200
Relative Error, %

Relative Error, %
100 100

0 0

-100 -100

-200 Elsharkawy -200 This study

-300 -300
20 30 API 40 50 20 30 API 40 50

Fig. 4. Error distribution curves of the models.

0.9

0.8

0.7
Cumulative Frequency

0.6

0.5

0.4

0.3 Hemmati-Sarapardeh
Naseri
0.2 Elsharkawy
Kartoatmodjo and Schmidt
0.1 Al-Khafaji
This study
0
0 5 10 15 20 25 30 35 40 45 50
Absolute Relative Error, %

Fig. 5. Cumulative frequency of data points versus absolute relative error for different models.

The other result obtained in this figure is that Glasø [28], Labedi— diate and light crude oil systems (i.e., API ≥ 20). This study does not
Libya [32], Labedi—Nigeria and Angola [31], Kaye [29], Al-Khafaji [30], aim to develop a model for heavy crude oils (API < 20) as viscosity
Petrosky [34], Elsharkawy [36], Naseri [13], Hemmati-Sarapardeh [5] of heavy and extra heavy crude oils (API < 20) are strongly affected
correlations cannot satisfactorily predict dead oil viscosity and pro- by heavy components such as resins and asphaltenes, and cannot be
vides RMSEs (cP) more than 50. Also, Beggs and Robinson [27] corre- modeled by using only API and temperature. By comparison of MAPE
lation estimates dead oil viscosity data with an RMSE (cP) more than and MPE, it can be concluded that Beggs and Robinson [27], Labedi—
595,000, which was not shown in this figure. From Figs. 2 and 3, it is Nigeria and Angola [31] correlations overestimate dead oil viscosity,
evident that the proposed models in this study are more robust, re- while Glasø [28], Labedi—Libya [32], Egbogah and Ng [9], Kaye [29],
liable, and accurate than pre-existing correlations, in terms of MAPE Petrosky [34], Kartoatmodjo and Schmidt [33], Bennison [35], Hos-
(%) and RMSE (cP). The other statistical parameters obtained by the sain [38], Naseri [13], Hemmati-Sarapardeh [5] correlations underes-
proposed models in this study and pre-existing correlations are re- timate this property.
ported in Table 3. As illustrated, the proposed models provide the To graphically compare the performance of CSA-LSSVM models
smallest MPE, MAPE, RMSE, and SD, indicating the better efficiency against the existing correlations, two graphical techniques were em-
of the proposed models over the existing correlations. Care should be ployed. The proposed models in this study were compared to three of
taken that the proposed models in this study are only for interme- the most accurate correlations in terms of mean absolute percentage
A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10 7

1000
Experimental data at T=294.94 K

Predicted values at T=294.94 K


Experimental data at T=310.46 K

Predicted values at T=310.46 K


100
Viscosity, cP

10

1
20.00 25.00 30.00 35.00 40.00 45.00 50.00
Oil API gravity

Fig. 6. Experimental viscosity data and the predicted values by the proposed model (Model 1) at two temperatures below 313.15 K.

100
Experimental data at T=333.15 K

Predicted values at T=333.15 K


Experimental data at T=422.03 K

Predicted values at T=422.03 K


10
Viscosity, cP

0.1
20.00 25.00 30.00 35.00 40.00 45.00 50.00
Oil API gravity

Fig. 7. Experimental viscosity data and the predicted values by the proposed model (Model 2) at two temperatures above 313.15 K.

error, namely Al-Khafaji [30] correlation, Kartoatmodjo and Schmidt 50%. Care should be taken that the high errors of the developed mod-
[33] correlation, and Elsharkawy [36] correlation. Error distribution els in predicting viscosity of some crude oils are due to high depen-
curves of these models have been plotted in Fig. 4. As can be observed, dency of oil viscosity on oil nature and source; as already mentioned,
the proposed CSA-LSSVM models provide the least scattering around the data points in this study have been collected from variety of
the zero error line, which reveals the better performance of the pro- geological and geographical locations. Therefore, it is not possible to
posed models against the existing correlations. Furthermore, five of develop very accurate models for these data points as the nature of
the most accurate correlations as well as CSA-LSSVM models were crude oils are different.
selected to compare their cumulative frequency plots. Fig. 5 shows
that the proposed models in this study, as the most accurate models, 5.2. Checking the physically expected trends of viscosity
predict 70% of data points with absolute relative error less than 22%. for the developed models
Elsharkawy [36] correlation predicts 60% of data points with abso-
lute relative error less than 22%. The proposed models in this study To check the validity of the proposed models against the in-
predict only 6% of data points with absolute relative error more than put parameters of the developed models (temperature and API),
8 A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10

The relevancy factor was firstly used by Chen et al. [58] to understand
0.1
the relative effect of their model inputs on the minimum miscibility
0 pressure. Moreover, this factor was also used by other researchers to
-0.1
find the relative impact of the input parameters on their models’ out-
puts [59–62]. The higher the absolute value of r between any input

Relevancy Factor
Temperature API
-0.2 and output, the greater the effect of that input on the output. It is
-0.13
-0.19 -0.3 interesting to note that, r value with directionality provides clearer
understanding about the overall impact; therefore, it was used in this
-0.4
study. Using the following formula, the r values are obtained:
-0.5 n
Above 313.15 K
(Inpk,i − Inpk )(μi − μ)
Below 313.15 K -0.6 r(Inpk , μ) =  i=1
(18)
-0.51 n 2 n
-0.7 i=1 (Inpk,i − Inpk ) i=1 (μi − μ)2
where μi and μ are the ith value of the predicted dead oil viscos-
-0.64

ity and the average value of the predicted dead oil viscosities, re-
Fig. 8. Relative effect of each input parameter on viscosity at temperatures below and spectively. Inpk , i and Inpk denote the ith value and the average value
above 313.15 K. of the kth input variable, respectively (k = temperature and API). In
this study, the relevancy factor was calculated for both of the models
(Model 1 and Model 2). The relative effect of each parameter (temper-
experimental viscosity data and predicted values by the proposed ature and API) on the dead oil viscosity is presented in Fig. 8. As can
models were sketched in Figs. 6 and 7. Fig. 6 shows the experimental be seen, the effect of both temperature and API are negative, mean-
data and predicted values by Model 1 at two temperatures of 294.94 ing that an increase in temperature/API leads to a decrease in dead
and 310.46 K. It is clear that, crude oils with lower API (higher den- oil viscosity. Furthermore, it was found that, for both of the models,
sity) have higher viscosity. However, as oil API gravity increases, oil API has larger effect on dead oil viscosity. In addition, API in Model
viscosity decreases exponentially. Moreover, the higher the temper- 2 has a greater effect on the dead oil viscosity than that in Model 1.
ature, the lower the viscosity. These physical rules exist in crude oil Generally speaking, API plays a key role in viscosity behavior of crude
systems and the proposed model can obey these trends. It is worth oil systems.
noting that the vertical axes in Figs. 6 and 7 are in logarithmic scales;
therefore, in these figures the viscosity decreases linearly as a func- 5.4. Applicability domain of the models and outlier detection
tion of API. Fig. 7 illustrates the experimental data and predicted val-
ues by Model 2 at temperatures of 333.15 and 422.04 K. As can be In the last part of this study, outlier detection, which is of a
seen, the proposed model can capture the physically expected trends vital importance in developing mathematical models/correlations,
with variation of temperature and oil API gravity. Besides, the pre- has been conducted. Some methods have been proposed for
dicted values by Model 2 are in good agreement with the experimen- outlier detection and identifying the applicability domain of a
tal data. model/correlation of which Leverage is known as one of the most
popular ones [63–65]. In this method, residual is defined as the differ-
5.3. Sensitivity analysis ence between the experimental data points and their corresponding
values predicted by the model. Afterward, to calculate Hat or Lever-
To quantitatively assess the relative impact of temperature and oil age indices, the following H matrix should be calculated [63,64]:
API gravity on dead oil viscosity by means of the developed models, a
H = X (X t X )−1 X t (19)
sensitivity analysis was carried out. To this end, the relevancy factor
(r) was employed to evaluate the influence degree of temperature and where X is an N × k matrix in which N is the number of data points
API on both of the proposed models (Model 1 and Model 2) [57,58]. (rows) and k stands for the number of model parameters (columns),

9
Valid Data
Suspected Data
Out of Leverage
6
Upper Suspected Limit
Upper Suspected Limit
Standardized Residuals

Leverage Limit
3

-3

-6

-9
0 0.002 0.004 0.006 0.008 0.01
hat

Fig. 9. Recognition of the probable outlier data and applicability domain of the developed model for crude oil systems with temperatures below 313.15 K.
A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10 9

6
Valid Data
Suspected Data
4 Out of Leverage
Upper Suspected Limit

Standardized Residuals
Lower Suspected Limit
2 Leverage Limit

-2

-4

-6
0 0.005 0.01 0.015 0.02 0.025
hat

Fig. 10. Recognition of the probable outlier data and applicability domain of the developed model for crude oil systems with temperatures above 313.15 K.

and t represents the transpose of matrix X. The diagonal elements of 4. The results demonstrate that both of API and temperature have
the H matrix are called hat in the feasible region of the problem. negative effects on dead oil viscosity, which means that an in-
Williams plot is sketched for graphical presentation of the ap- crease in API/temperature causes a decrease in viscosity.
plicability domain of the used model and suspected data. In fact, 5. The relevancy factor indicates that API has the most important
Williams plot illustrates the correlation of Hat indices and standard- effect on dead oil viscosity.
ized residuals (R). A warning Leverage (H∗ ) is fixed at a value equal 6. The Leverage approach shows that both of Model 1 and Model 2
to 3(k + 1)/N, in which N (Model 1: N = 933; Model 2: N = 564) and are statistically valid and acceptable. In addition, less than 2% of
k (k = 2 for both models) are the number of data points and vari- data points may be regarded as the probable outliers.
ables, respectively. Thus, H∗ of Model 1 and Model 2 are set to be
0.0096 and 0.0158, respectively. Existence of most of the data points
in the range of 0 ≤ H ≤ H∗ and –3 ≤ R ≤ 3 reveals that both the data References
used for the model development and the data predicted by the ap-
plied model are in the applicability domain; and consequently, the [1] El-hoshoudy A, Farag A, Ali O, EL-Batanoney M, Desouky S, Ramzi M. New cor-
model is statistically valid. Williams plots for Model 1 and Model 2 relations for prediction of viscosity and density of Egyptian oil reservoirs. Fuel
2013;112:277–82.
have been sketched in Figs. 9 and 10, respectively. For Model 1, only [2] Ahmed T. Reservoir engineering handbook. Oxford: Gulf Professional Publishing;
11 data points have R > 3 or R < –3, and 4 data points have hat > H∗ . 2006.
These data points (1.6% of data points) may be regarded as the prob- [3] Donnez P. Essentials of reservoir engineering. TECHNIP Ophrys Editions; 2012.
[4] McCain WD. The properties of petroleum fluids. Oklahoma: PennWell Books;
able outlier ones, and are located out of the applicability domain of 1990.
Model 1. For Model 2, only 9 data points have R > 3 or R < –3, and 8 [5] Hemmati-Sarapardeh A, Khishvand M, Naseri A, Mohammadi AH. Toward reser-
data points have hat > H∗ . Therefore, 3% of data points are located out voir oil viscosity correlation. Chem Eng Sci 2013;90:53–68.
[6] Hemmati-Sarapardeh A, Mahmoudi B, Ramazani SA, Mohammadi AH. Experi-
of the applicability domain of Model 2, and may be regarded as the mental measurement and modeling of saturated reservoir oil viscosity. Korean
probable outliers. In conclusion, both of the models are statistically J Chem Eng 2014;31:1253–64.
acceptable and valid. It is worthwhile to note that data points with [7] Hemmati-Sarapardeh A, Shokrollahi A, Tatar A, Gharagheizi F, Mohammadi AH,
Naseri A. Reservoir oil viscosity determination using a rigorous approach. Fuel
lower values of R and hat are recognized as more reliable ones. 2014;116:39–48.
[8] Zéberg-Mikkelsen CK. Viscosity studies of hydrocarbon fluids at reservoir condi-
tions: modeling and measurements. Ph.D. Thesis 2001.
6. Conclusions [9] Egbogah EO, Ng JT. An improved temperature-viscosity correlation for crude oil
systems. J Petrol Sci Eng 1990;4:197–200.
In this study, a supervised learning algorithm, namely least square [10] De Ghetto G, Villa M, editors. Reliability analysis on PVT correlations. European
Petroleum Conference. Society of Petroleum Engineers; 1994.
support vector machine, has been used to develop reliable and accu- [11] Oloso MA, Khoukhi A, Abdulraheem A, Elshafei M, editors. Prediction of crude
rate models for estimation of dead oil viscosity of light and intermedi- oil viscosity and gas/oil ratio curves using recent advances to neural networks.
ate crude oil systems. About 1500 data points from various geological SPE/EAGE Reservoir Characterization & Simulation Conference; 2009.
[12] Al-Marhoun MA. Evaluation of empirically derived PVT properties for Middle East
locations were used to develop these models as a function of temper- crude oils. J Petrol Sci Eng 2004;42:209–21.
ature and oil API gravity. Two distinct models for crude oil systems [13] Naseri A, Nikazar M, Mousavi Dehghani S. A correlation approach for prediction
with temperatures lower than 313.15 K and higher than 313.15 K have of crude oil viscosities. J Petrol Sci Eng 2005;47:163–74.
[14] Naseri A, Yousefi S, Sanaei A, Gharesheikhlou A. A neural network model and an
been developed. According to the results obtained in this study, the updated correlation for estimation of dead crude oil viscosity. Brazil J Petrol Gas
following conclusions can be drawn: 2012;6.
[15] Makinde F, Ako C, Orodu O, Asuquo I. Prediction of crude oil viscosity using feed-
1. The proposed CSA-LSSVM models estimate crude oil viscosity of forward back-propagation neural network (FFBPNN). Petrol Coal 2012;54:120–31.
[16] Sattarin M, Modarresi H, Bayat M, Teymori M. New viscosity correlations for dead
light and intermediate systems with enough accuracy.
crude oils. Petrol Coal 2007;49:33–9.
2. The statistical and graphical error analyses reveal that the devel- [17] Ghorbani B, Ziabasharhagh M, Amidpour M. A hybrid artificial neural network
oped CSA-LSSVM models are more accurate, reliable, and superior and genetic algorithm for predicting viscosity of Iranian crude oils. J Nat Gas Sci
to all of the existing correlations. Eng 2014;18:312–23.
[18] Lashkenari MS, Taghizadeh M, Mehdizadeh B. Viscosity prediction in selected Ira-
3. The CSA-LSSVM models can follow the physically expected trends nian light oil reservoirs: artificial neural network versus empirical correlations.
with variation of temperature and oil API gravity. Petrol Sci 2013;10:126–33.
10 A. Hemmati-Sarapardeh et al. / Journal of the Taiwan Institute of Chemical Engineers 59 (2016) 1–10

[19] Ahmadloo F, Asghari K, Araghi MM, editors. Heavy oil viscosity prediction us- [45] Al-Maamari RS, Houache O, Abdul-Wahab SA. New correlating parameter for the
ing surface response methodology. Canadian International Petroleum Conference. viscosity of heavy crude oils. Energy Fuels 2006;20:2586–92.
Petroleum Society of Canada; 2009. [46] Kamari A, Gharagheizi F, Bahadori A, Mohammadi AH. Determination of the equi-
[20] Isehunwa O, Olamigoke O, Makinde A, editors. A correlation to predict the vis- librated calcium carbonate (calcite) scaling in aqueous phase using a reliable ap-
cosity of light crude oils. Nigeria Annual International Conference and Exhibition. proach. J Taiwan Inst Chem Eng 2014;45:1307–13.
Society of Petroleum Engineers; 2006. [47] Kamari A, Gharagheizi F, Bahadori A, Mohammadi AH, Zendehboudi S. Rigorous
[21] Alomair O, Elsharkawy A, Alkandari H. A viscosity prediction model for modeling for prediction of barium sulfate (barite) deposition in oilfield brines.
Kuwaiti heavy crude oils at elevated temperatures. J Petrol Sci Eng 2014;120:102– Fluid Phase Equilib 2014;336:117–26.
10. [48] Q-f DI, Shuai H, DING W-p, Wei G, CHENG Y-c, Feng Y. Application of support
[22] Abedini R, Abedini A, Yakhfrouzan NE. A new correlation for prediction of under- vector machine in drag reduction effect prediction of nano-particles adsorption
saturated crude oil viscosity. Petrol Coal 2010;52:50–5. method on oil reservoir’s micro-channels. J Hydrodyn, Ser B 2015;27:99–104.
[23] Sánchez-Minero F, Sánchez-Reyna G, Ancheyta J, Marroquin G. Comparison of [49] Dong N, Ma M, Huang H, Ma Y, Zheng L, editors. Support Vector machine in crash
correlations based on API gravity for predicting viscosity of crude oils. Fuel prediction at the level of traffic analysis zones: accessing the spatial proximity
2014;138:193–9. effects. Transportation Research Board 94th Annual Meeting; 2015.
[24] Svrcek W, Mehrotra A. One parameter correlation for bitumen viscosity. Chem [50] Yousefi SH, Azamifard A, Hosseini SA, Shamsoddini MJ, Alizadeh N. Toward a pre-
Eng Res Des 1988;66:323–7. dictive model for predicting viscosity of natural and hydrocarbon gases. J Nat Gas
[25] Mehrotra AK. Generalized one-parameter viscosity equation for light and Sci Eng 2014;20:147–54.
medium liquid hydrocarbons. Ind Eng Chem Res 1991;30:1367–72. [51] Mohammadi K, Shamshirband S, Anisi MH, Alam KA, Petković D. Support vec-
[26] Beal C. The viscosity of air water natural gas crude oil and its associated gases at tor regression based prediction of global solar radiation on a horizontal surface.
oil field temperatures and pressures. Trans AIME. 1946;165:94–115. Energy Convers Manage 2015;91:433–41.
[27] Beggs HD, Robinson J. Estimating the viscosity of crude oil systems. J Petrol Tech- [52] Yan K, Shen W, Mulumba T, Afshari A. ARX model based fault detection and diag-
nol 1975;27(1):140–1 1. nosis for chillers using support vector machines. Energy Build 2014;81:287–95.
[28] Glasø O. Generalized pressure-volume-temperature correlations. J Petrol Technol [53] Zheng B, Myint SW, Thenkabail PS, Aggarwal RM. A support vector machine to
1980;32(05):785–95. identify irrigated crop types using time-series Landsat NDVI data. Int J Appl Earth
[29] Kaye S. Offshore California viscosity correlations. Aug 1985 TS85000940. Observ Geoinf 2015;34:103–12.
[30] Al-Khafaji A, Abdul-Majeed G, Hassoon S. Viscosity correlation for dead, live and [54] Suykens JA, Van Gestel T, De Brabanter J, De Moor B, Vandewalle J, Suykens J, et al.
undersaturated crude oils. J Pet Res 1987;6:1–16. Least squares support vector machines. London: World Scientific; 2002.
[31] Labedi RM. PVT correlations of the African crudes. Leadville, Colorado, USA: Col- [55] Suykens JA, Vandewalle J. Least squares support vector machine classifiers. Neural
orado School of Mines; 1982. Process Lett 1999;9:293–300.
[32] Labedi R. Improved correlations for predicting the viscosity of light crudes. J [56] Pelckmans K, Suykens JA, Van Gestel T, De Brabanter J, Lukas L, Hamers B, et al. LS-
Petrol Sci Eng 1992;8:221–34. SVMlab: a matlab/c toolbox for least squares support vector machines. Belgium:
[33] Kartoatmodjo T, Schmidt Z. Large data bank improves crude physical property Tutorial KULeuven-ESAT Leuven; 2002.
correlations. Oil and Gas J (United States) 1994;4:51–5. [57] Hosseinzadeh M, Hemmati-Sarapardeh A. Toward a predictive model for es-
[34] Petrosky G, Farshad F, editors. Viscosity correlations for Gulf of Mexico crude oils. timating viscosity of ternary mixtures containing ionic liquids. J Mol Liquids
Production operation symposium; 1995. 2014;200:340–8.
[35] Bennison T, editor. Prediction of heavy oil viscosity. Presented at the IBC Heavy [58] Chen G, Fu K, Liang Z, Sema T, Li C, Tontiwachwuthikul P, et al. The genetic al-
Oil Field Development Conference; 1998. gorithm based back propagation neural network for MMP prediction in CO2-EOR
[36] Elsharkawy A, Alikhan A. Models for predicting the viscosity of Middle East crude process. Fuel 2014;126:202–12.
oils. Fuel 1999;78:891–903. [59] Kamari A, Arabloo M, Shokrollahi A, Gharagheizi F, Mohammadi AH. Rapid
[37] Whitson CH, Brulé MR. Phase behavior: Richardson, Tex.: Henry L. Doherty method to estimate the minimum miscibility pressure (MMP) in live reservoir
Memorial Fund of AIME. Society of Petroleum Engineers; 2000. oil systems during CO2 flooding. Fuel 2015;153:310–19.
[38] Hossain MS, Sarica C, Zhang H-Q, Rhyne L, Greenhill K, editors. Assessment and [60] Fathinasab M, Ayatollahi S, Hemmati-Sarapardeh A. A rigorous approach to pre-
development of heavy oil viscosity correlations. SPE International Thermal Oper- dict nitrogen-crude oil minimum miscibility pressure of pure and nitrogen mix-
ations and Heavy Oil Symposium. Society of Petroleum Engineers; 2005. tures. Fluid Phase Equilib. 2015;399:30–9.
[39] Miadonye A., Singh B., Puttagunta V.R. One-parameter correlation in the estima- [61] Esfahani S, Baselizadeh S, Hemmati-Sarapardeh A. On determination of natural
tion of crude oil viscosity. 1992. gas density: least square support vector machine modeling approach. J Nat Gas
[40] Sadeghi MB, Ramazani SAA, Taghikhani V, Ghotbi C. Experimental investigation Sci Eng 2015;22:348–58.
of rheological and morphological properties of water in crude oil emulsions sta- [62] Mohagheghian E, Zafarian-Rigaki H, Motamedi-Ghahfarrokhi Y, Hemmati-
bilized by a lipophilic surfactant. J Dispers Sci Technol 2013;34:356–68. Sarapardeh A. Using an artificial neural network to predict carbon dioxide com-
[41] Degiorgis G, Maturano S, Garay M, Galliano G, Fornes A, editors. Oil mixture pressibility factor at high pressure and temperature. Korean J Chem Eng 2015.
viscosity behavior: use in pipeline design. SPE Latin American and Caribbean doi:10.1007/s11814-015-0025-y.
Petroleum Engineering Conference. Society of Petroleum Engineers; 2001. [63] Rousseeuw P.J., Leroy A.M. Robust regression and outlier detection: Wiley.com;
[42] Everett J, Weinaug CF. Physical properties of eastern Kansas crude oils. Kansas 2005.
Geo Surv, Bull 1955:114. [64] Goodall CR. 13 computation using the QR decomposition. Handbook Stat
[43] De Ghetto G, Paone F, Villa M, editors. Pressure-volume-temperature correlations 1993;9:467–508.
for heavy and extra heavy oils. International heavy oil symposium; 1995. [65] Shateri M, Ghorbani S, Hemmati-Sarapardeh A, Mohammadi AH. Application of
[44] Croft GD, Patzek TW, editors. The Future of California’s Oil Supply. SPE Western Wilcoxon generalized radial basis function network for prediction of natural gas
Regional Meeting. Society of Petroleum Engineers; 2009. compressibility factor. J Taiwan Inst Chem Eng 2015;50:131–41.

Das könnte Ihnen auch gefallen