Fluid Phase Equilibria 363 (2014) 27–31

Contents lists available at ScienceDirect

Fluid Phase Equilibria

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

Are the reservoir fluid compositional grading data reliable?

Amir H. Mohammadi a,b,∗ , Ali Eslamimanesh c ,
Farhad Gharagheizi b,d , Poorandokht Ilani-Kashkouli b,d
a
Institut de Recherche en Génie Chimique et Pétrolier (IRGCP), Paris Cedex, France
b
Thermodynamics Research Unit, School of Engineering, University of KwaZulu-Natal, Howard College Campus, King George V Avenue, Durban 4041, South
Africa
c
Department of Chemical & Biomolecular Engineering, Clarkson University, Potsdam, NY 13699-5705, USA
d
Department of Chemical Engineering, Buinzahra Branch, Islamic Azad University, Buinzahra, Iran

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

Article history: Outlier detection in compositional grading data of a reservoir fluid is the main objective of the present
Received 11 July 2013 study. The experimental data of a petroleum reservoir fluid including the mole fractions of the fluid
Received in revised form 20 October 2013 components at different depths (from around 1000 to about 1400 m) and at constant temperature of
Accepted 30 October 2013
361.15 K are investigated. The utilized algorithm applies the basis of a mathematical approach, in which
Available online 7 November 2013
the statistical Hat matrix, Williams plot, and the residuals of a compositional grading model results
bring about the probable outliers detection. The range of applicability of the applied model and quality
Keywords:
of the existing experimental data are also investigated. The reported results of a previously developed
Evaluation of data
Compositional grading data
model using the Soave–Redlich–Kwong equation of state (SRK EoS) with Peneloux volume correction are
Reservoir fluid employed to evaluate the compositional analysis of the species in different depths of the fluid column. It is
Thermodynamic model interpreted from the obtained results that the applied model for estimation of the compositional gradients
Outlier has wide ranges of applicability. In addition, we may conclude that there is no outlier or probable doubtful
datum in the investigated experimental datasets.
© 2013 Elsevier B.V. All rights reserved.

1. Introduction experimental measurements of the compositions of a real reservoir

Variations in pressure, temperature, composition and PVT
properties of reservoir fluids with depth are normally found in
petroleum reservoirs [1–20]. For instance, the temperature has
been reported to vary 0.02–0.03 K/m and bubble point pressure
can change up to 0.11 MPa/m [2,3]. As the depth of a reservoir
increases, the concentration of lighter components decreases while
the concentration of heavier components increases [1]. Such com-
positional gradients are of much importance when the petroleum
fluid column is deep. It has been observed that the concentration of
methane in an heavy oil reservoir can be decreased up to 20% over
a 81 m depth increase [3]. The compositional grading of a reser-
voir fluid column is of significance for primary depletion of the
reservoir, sampling from the fluid column within the reservoir, and
determination of the depth of its oil–gas contact [1–4].
Gravity and temperature gradients generally play the main role
in compositional variations of the petroleum reservoir fluids. Since
∗ Corresponding author at: Institut de Recherche en Génie Chimique et Pétrolier
(IRGCP), Paris Cedex, France.
E-mail address: a.h.m@irgcp.fr (A.H. Mohammadi).
fluid at different depths are generally difficult, there is a need to
develop reliable thermodynamic models to predict the subsequent
compositional variations.
The compositional grading models can be grouped normally
into isothermal [1,2,21] and non-isothermal [1,22] models. The
starting point of the model development is to evaluate the differ-
ence between the chemical potential of a component in different
depths (locations in the reservoir). Although temperature gradi-
ent brings about heat flow between different points of reservoir,
the thermodynamic equilibrium assumption of gravity segregation
is generally considered as a basis for the compositional grad-
ing thermodynamic modeling. The effects of temperature on the
compositional gradients may be negligible compared with grav-
ity effects. Furthermore, calculations of heat flux, heat diffusions
and their effects on residual enthalpies and finally on chemical
potentials of components throughout the reservoir enhance the
complexity of the non-isothermal model. Therefore, the isothermal
compositional grading model seems to be preferable for practical
purposes.
As mentioned earlier, the corresponding experimental compo-
sitional grading data are scarce in open literature. Furthermore,
there has been no method proposed in open literature so far
to check the reliability of such data. Hence, it is of interest

0378-3812/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.fluid.2013.10.058
28 A.H. Mohammadi et al. / Fluid Phase Equilibria 363 (2014) 27–31
to check the quality of these data as well as the applicability
domain of the isothermal compositional grading model for its
prediction. In this work, a statistical method is applied for this
purpose.
2. Isothermal thermodynamic model
The previously reported results of an available model in the lit-
erature [1] are used in this communication. The elements of the
model are as follows: at isothermal equilibrium conditions, the
total chemical potential of a component is the same at all posi-
tions in a closed system. Therefore, the difference between the
intrinsic chemical potential of components reduces to the effect of
the depth in the fluid column evaluated by the following equation
[1]:
i (h) − i (ho ) = Mi g(h − ho )
where  is the chemical potential, subscript i refers to com-
ponent i in the fluid mixture, M is the molecular weight, and
h denotes the depth. The superscript ‘o’ stands for a reference
depth. The chemical potential can be written in terms of fugac-
ity or fugacity coefficient, pressure and composition as follows
[1]:
di = RTd ln(fi ) = RTd ln(ϕi zi P)
In Eq. (2), R is the universal gas constant, T is temperature, ϕ
is fugacity coefficient, z is the overall mole fraction of the species,
and P stands for pressure. Eq. (2) is only available at isothermal
conditions. Introducing Eq. (2) into Eq. (1) yields [1]:
o Mi g(h − ho )
ln fih − ln fih =
RT
The fugacity of component i can be evaluated using the overall
composition of the species as follows:
fi = Pi zi ϕi
where z is the overall composition of component i. Eq. (3) can be
re-written as follows:
o o o Mi g(h − ho )
ln(ϕih zih P h ) − ln(ϕih zih P h ) = i = 1, 2, . . ., N
RT
Where N is the number of components. An additional equa-
tion can be written regarding the unity of summations of the mole
fractions:

N that the representations/predictions of the model are done in its
zi = 1
i=1
Thus, there are N + 1 variables for a given depth in the fluid col-
umn. Solving N + 1 equations with N + 1 variables leads to obtaining
the molar compositions and pressure with depth. An appropri-
ate equation of state can be applied for determination of the
required variables. In this work, the results of an isothermal
thermodynamic model [1] employing the Soave–Redlich–Kwong
equation of state (SRK EoS) [23] with Peneloux volume correction
[24] have been used for statistical evaluation of the experi-
mental data. Furthermore, the applied model [1] utilizes the
procedure proposed by Pedersen et al. [25] for fluid characteriza-
tion.
3. Experimental data
The experimental data of a petroleum reservoir fluid reported
by Creek and Schrader [26] in 1985 have been herein evaluated.
The data includes the mole fractions of the fluid components at dif-
ferent depths (from around 1000 to about 1400 m) and at constant
temperature of 361.15 K.
4. Theory
4.1. Leverage method
Detection of the outliers is to diagnose individual datum (or
a dataset) that may differ from the bulk of the data (a database)
[27–30]. The proposed methods for this purpose normally consist
of simultaneous numerical and graphical algorithms [27–30]. The
Leverage method [27–30] (employed here) utilizes the values of the
residuals (i.e. the deviations of a model results from the correspond-
ing experimental data) and a matrix (Hat matrix) composed of the
experimental data and the represented/predicted values obtained
from the model [27–30]. Hence, a suitable mathematical model
is also required to pursue the calculation steps of the algorithm
[27–30].
The Leverage or Hat indices are determined as a Hat matrix (H)
(1)
with the following definition [27–30]:
−1
H = X(X t X) Xt (7)
where X is a two-dimensional matrix composed of n data (rows)
and p model parameters (columns) and t stands for the transpose
matrix. The Hat values of the chemicals in the feasible region of
(2) the problem are, as a matter of fact, the diagonal elements of the H
matrix.
The Williams plot is later sketched for graphical identification
of the suspended data or outliers on the basis of the calculated H
values through Eq. (7). This plot shows the correlation of Hat indices
and standardized cross-validated residuals (R ), which are defined
as the difference between the represented/predicted values and the
(3) implemented data as follows [27]:
rep./pred.
R = zi
exp .
− zi (8)
(4) where rep./pred. and exp. represent the represented/predicted and
experimental overall mole fractions, respectively. A warning Lever-
age (H*) is generally fixed at the value equal to 3p/n, where n is
number of training points and p is the number of model input
(5) parameters plus one [27–30]. The leverage of 3 is normally consid-
ered as a “cut-off” value to accept the points within ±3 range (two
horizontal red lines) standard deviations from the mean (to cover
99% normally distributed data) [27–30]. Existence of the major-
ity of data points in the ranges 0 ≤ H ≤ H* and −3 ≤ R ≤ 3 reveals
(6) applicability domain. “Good High Leverage” points are located in
domain of H* ≤ H and −3 ≤ R ≤ 3. The Good High Leverage can be
designated as the ones, which are outside of applicability domain
of the applied model [27–30]. In other words, the model is not able
to represent/predict the following data at all. The points located in
the range of R < −3 or 3 < R (whether they are larger or smaller than
the H* value) are designated as outliers of the model or “Bad High
Leverage” points. These erroneous representations/predictions can
be attributed to the doubtful data [27–30].
5. Results and discussion
The absolute relative deviations (ARDs) of the thermody-
namic model [1] results are presented in Table 1 along with the
experimental data [26]. As can be seen, the deviations of the
predictions from the corresponding experimental data [26] are gen-
erally acceptable to be used for the Leverage statistical approach
[27–30].
To pursue our objectives, the H values have been calculated
through Eq. (7) and the Williams plots have been sketched in Fig. 1.
The whole calculated H and R values are presented in Table 2. The
 A.H. Mohammadi et al. / Fluid Phase Equilibria 363 (2014) 27–31
Fig. 1. Detection of the probable doubtful experimental data [26] for data sets 1–8 in Tables 1 and 2 and the applicability domain of the applied thermodynamic model [1].
The H* value is 1.091. Experimental values: Set 1: *; Set 2: ; Set 3: ; Set 4: ; Set 5: ×; Set 6: •; Set 7: +; Set 8: ♦.
warning Leverages (H*) have been fixed at 3p/n for the whole data.
In addition, the recommended cut-off value of 3 has been applied
[27–30].
The following results are interpreted from application of the
aforementioned methodology:
1. Accumulation of the data points [26] in the ranges 0 ≤ H ≤ H* and
−3 ≤ R ≤ 3 reveals that the applied model [1] is statistically valid
for prediction of the treated experimental values [26].
2. The whole compositional grading data points [26] can be
declared to within the applicability domain of the thermody-
namic model [1]. Furthermore, there are no good high leverage
points that are supposed to be accumulated in the domains of
H* ≤ H and −3 ≤ R ≤ 3. These points may be declared to be out-
side of applicability domain of the applied compositional grading
model [1] though cannot be assigned as doubtful experimental
data. It should be noted that, in the case of facing with good high
leverage points, it is recommended to use/develop other models
on the basis of different theoretical concepts (non-isothermal
[1,22] or continuous thermodynamic ones [2]) for their calcu-
lations/estimations in order to avoid estimation through biased
model calculations [27].
3. The data points located in the range of R < −3 or 3 < R (ignoring
their H values) are generally designated as outliers or bad high
leverage points, as already explained. The results show that all
the treated compositional grading data points [26] are located
outside of the mentioned domain and consequently designated
as valid (not outliers). If there was any datum located within
this domain, they might have been attributed to the doubtful
experimental data.
29
4. The data points in the ranges H* ≤ H and R < −3 or 3 < R may
be designated as neither within the applicability domain of the
applied correlation nor valid data. In other words, these data can-
not be well calculated/estimated by the applied model [1]. There
is no such a point in the treated datasets in this work [26].
5. The quality of the treated data [26] (even different data in the
same dataset) are different. The data with lower absolute R val-
ues (near R = 0 line) and lower H values may be declared as the
more reliable experimental data [26].
In the final analysis, it should be noted that in the present work,
as we were interested in defining only the data quality (and the
range of applicability of the model), consequently, we have focused
on assessment test while our objective has not been conclusion
about the pursued experimental technique.
One point should not be eliminated from our discussion. The
thermodynamic consistency test based on Gibbs–Duhem equation
[31–35] can show to what extent the experimental data are thermo-
dynamically consistent. However, these kinds of tests are inevitably
model-dependent [36]. Therefore, some data could not pass the
consistency test but still be reliable ones from mathematical point
of view. In the case that the results of a thermodynamic consis-
tency test are available, we recommend the users to keep all the
consistent data that can be declared not to be outliers using the
applied statistical approach [27–30] and some of the similar not
fully consistent data, for tuning the thermodynamic models. Later,
on the basis of their experiences, some conclusions should be made
on thermodynamic inconsistent data which can pass the statistical
approach [27–30]. However, we strongly advice the users not to
use the data that cannot fulfill both tests.
 30
Table 1
The absolute relative deviations of the applied model [1] results with respect to the experimental data [26].

Depth (m) Set 1 Set 2 Set3 Set 4 Set 5 Set 6 Set 7 Set 8

1017–1053 1062–1087 1076–1111 1251–1299 1289–1303 1274–1322 1298–1301 1378–1392

a b c
Exp. Pred. ARD % Exp. Pred. ARD % Exp. Pred. ARD% Exp. Pred. ARD % Exp. Pred. ARD % Exp. Pred. ARD % Exp. Pred. ARD % Exp. Pred. ARD %

Components
CO2 0.02 0.19 850 0.16 0.18 13 0.12 0.18 50 0.13 0.15 15 0.14 0.14 0.0 0.14 0.14 0.0 0.10 0.14 40 0.14 0.13 7
N2 1.42 1.21 15 1.31 1.22 6.9 1.20 1.22 1.7 1.42 1.21 15 1.50 1.21 19 1.45 1.2 17 1.28 1.2 6.3 1.04 1.19 14
C1 71.26 71.33 0.1 71.85 70.64 1.7 69.79 70.27 0.7 67.84 63.56 6.3 67.37 62.45 7.3 65.76 62.4 5.1 65.92 62.31 5.5 59.67 58.98 1.2
C2 11.04 11.62 5.3 10.60 11.67 10 11.63 11.70 0.6 11.02 11.96 8.5 11.70 11.96 2.2 11.27 11.96 6.1 11.63 11.96 2.8 11.67 11.89 1.9
C3 5.66 5.78 2.1 5.73 5.86 2.3 5.87 5.91 0.7 5.83 6.55 12 5.92 6.62 12 6.20 6.62 6.8 6.36 6.63 4.2 6.58 6.79 3.2
i-C4 1.39 1.42 2.2 1.34 1.45 8.2 1.41 1.47 4.3 1.45 1.71 18 1.46 1.74 19.2 1.62 1.75 8.0 1.63 1.75 7.4 1.58 1.82 15
n-C4 1.79 1.82 1.7 1.70 1.87 10 1.82 1.90 4.4 1.90 2.28 20 1.75 2.33 33 2.15 2.33 8.4 2.10 2.33 11 2.11 2.46 17
i-C5 0.73 0.72 1.4 0.66 0.74 12 0.77 0.75 2.6 0.80 0.96 20 0.81 0.98 21 0.61 0.98 61 0.88 0.99 13 0.9 1.06 18
n-C5 0.66 0.65 1.5 0.60 0.67 12 0.70 0.68 2.9 0.76 0.88 16 0.70 0.91 30 0.83 0.91 9.6 0.81 0.91 12 0.84 0.99 18

A.H. Mohammadi et al. / Fluid Phase Equilibria 363 (2014) 27–31

C6 0.83 0.82 1.2 0.72 0.85 18 0.96 0.87 9.4 0.96 1.20 25 1.25 1.25 0.0 1.38 1.25 9.4 1.06 1.26 19 1.17 1.39 19
C7+ 5.20 4.45 14 5.33 4.83 9.4 5.73 5.04 12 7.89 9.54 21 7.50 10.4 39 8.59 10.44 22 8.23 10.52 28 14.3 13.32 7
a
Experimental data (mole percent).
b
Predicted value (mole percent).
Exp.
|z −z Pred. |
c
Absolute relative deviation: ARD% = 100 × i i
Exp. .
z
i

Table 2
The results of the statistical approach for checking the quality of experimental compositional data [26] treated in this work.

Depth (m) Set 1 Set 2 Set3 Set 4 Set 5 Set 6 Set 7 Set 8

1017–1053 1062–1087 1076–1111 1251–1299 1289–1303 1274–1322 1298–1301 1378–1392

H R H R H R H R H R H R H R H R

Components
CO2 0.109 −0.6021 0.109 0.323 0.110 −0.572 0.114 0.777 0.115 0.651 0.115 0.631 0.115 0.584 0.118 0.489
N2 0.105 0.576 0.105 0.548 0.106 −0.357 0.109 0.975 0.110 0.834 0.110 0.880 0.110 0.650 0.112 −0.062
C1 0.976 1.271 0.975 2.160 0.975 −0.664 0.958 2.837 0.954 2.398 0.954 2.678 0.953 2.423 0.935 −1.336
C2 0.092 −1.918 0.093 −2.712 0.093 −0.229 0.093 −1.450 0.093 −0.452 0.093 −1.093 0.093 −0.589 0.094 −0.809
C3 0.093 −0.467 0.093 −0.185 0.093 −0.288 0.093 −0.690 0.093 −0.460 0.093 −0.372 0.093 −0.147 0.093 −0.524
i-C4 0.104 −0.209 0.105 0.049 0.105 −0.527 0.107 0.327 0.107 0.276 0.107 0.346 0.107 0.373 0.109 −0.373
n-C4 0.103 −0.206 0.103 −0.116 0.103 −0.602 0.105 0.122 0.105 −0.040 0.105 0.240 0.105 0.201 0.106 −0.744
i-C5 0.107 −0.084 0.107 0.152 0.107 −0.193 0.110 0.523 0.111 0.432 0.111 0.088 0.111 0.439 0.113 −0.087
n-C5 0.107 −0.084 0.108 0.180 0.108 −0.196 0.111 0.582 0.111 0.401 0.111 0.471 0.111 0.457 0.113 −0.052
C6 0.107 −0.083 0.107 0.025 0.107 0.125 0.109 0.399 0.110 0.567 0.110 0.720 0.110 0.313 0.111 −0.289
C7+ 0.096 2.358 0.095 1.391 0.095 2.936 0.091 −2.136 0.091 −2.716 0.092 −2.481 0.092 −2.7976 0.097 2.809
 A.H. Mohammadi et al. / Fluid Phase Equilibria 363 (2014) 27–31 31

6. Conclusion [7] A. Fazlali, M. Nikookar, A.H. Mohammadi, Computational procedure for ther-
modynamic minimum miscibility pressure of reservoir oil, Fuel 106 (2013)
707–711.
A statistical method to evaluate the compositional grading data [8] A.H. Mohammadi, A. Eslamimanesh, D. Richon, Wax solubility in gaseous
[26] of an oil reservoir fluid at different depths and constant tem- system: thermodynamic consistency test of experimental data, Industrial &
perature was proposed on the basis of the Leverage statistical Engineering Chemistry Research 50 (2011) 4731–4740.
[9] A.H. Mohammadi, F. Gharagheizi, A. Eslamimanesh, D.D. Richon, Evaluation
approach. The predicted values using a thermodynamic model [1] of experimental data for wax and diamondoids solubility in gaseous systems,
assuming the isothermal condition in the fluid column were used Chemical Engineering Science 81 (2012) 1–7.
for following the calculation procedure. The results show that: [10] A. Kamari, A. Khaksar-Manshad, F. Gharagheizi, A.H. Mohammadi, S. Ashoori, A
robust model for the determination of wax deposition in oil systems, Industrial
& Engineering Chemistry Research 52 (2013) 15664–15672.
I. The applied thermodynamic model [1] is valid and statistically [11] A.H. Mohammadi, D. Richon, A monodisperse thermodynamic model for esti-
correct for prediction of the compositional grading data [26]. mating asphaltene precipitation, AIChE Journal 53 (2007) 2940–2947.
[12] A.H. Mohammadi, A. Eslamimanesh, D. Richon, Asphaltene precipitation in gas
II. The whole experimental data points [26] are found to be within
condensate system, in: C. James, Taylor (Eds.), Advances in Chemistry Research,
the applicability domain of the employed model [1]. vol. 15, Nova Science Publishers, Inc., Hauppauge, NY, 2012.
III. No compositional grading data [26] may be designated as sus- [13] A.H. Mohammadi, A. Eslamimanesh, D. Richon, Monodisperse thermodynamic
pended (probable doubtful) experimental data. model based on chemical + flory – huggins polymer solution theories for pre-
dicting asphaltene precipitation, Industrial & Engineering Chemistry Research
51 (2012) 4041–4055.
List of symbols [14] A.H. Mohammadi, A. Eslamimanesh, F. Gharagheizi, D. Richon, A novel method
for evaluation of asphaltene precipitation titration data, Chemical Engineering
ARD absolute relative deviations
Science 78 (2012) 181–185.
h depth [15] A. Chamkalani, A.H. Mohammadi, A. Eslamimanesh, F. Gharagheizi, D. Richon,
M molecular weight Diagnosis of asphaltene stability in crude oil through “two parameters” SVM
model, Chemical Engineering Science 81 (2012) 202–208.
g acceleration of gravity
[16] A. Hemmati-Sarapardeh, R. Alipour-Yeganeh-Marand, A. Naseri, A. Safiabadi, F.
d derivative Gharagheizi, P. Ilani-Kashkouli, A.H. Mohammadi, Asphaltene precipitation due
R universal gas constant to natural depletion of reservoir: determination using a SARA fraction based
T temperature intelligent model, Fluid Phase Equilibria 354 (2013) 177–184.
[17] A. Hemmati-Sarapardeh, M. Khishvand, A. Naseri, A.H. Mohammadi, Toward
P pressure reservoir oil viscosity correlation, Chemical Engineering Science 90 (2013)
z overall mole fraction 53–68.
f fugacity coefficient [18] A. Hemmati-Sarapardeh, A. Shokrollahi, A. Tatar, F. Gharagheizi, A.H. Moham-
madi, A. Naseri, Reservoir oil viscosity determination using a rigorous approach,
N number of components Fuel 116 (2014) 39–48.
SRK Soave–Redlich–Kwong [19] A. Farasat, A. Shokrollahi, M. Arabloo, F. Gharagheizi, A.H. Mohammadi, Toward
EoS equation of state an intelligent approach for determination of saturation pressure of crude oil,
Fuel Processing Technology 115 (2013) 201–214.
H Hat matrix [20] M. Arabloo, A.M. Amooie, A. Hemmati-Sarapardeh, M.H. Ghazanfari, A.H.
X two-dimensional matrix Mohammadi, Application of constrained multi-variable search methods for
n number of data prediction of PVT properties of crude oil systems, Fluid Phase Equilibria (2013)
(accepted for publication).
p number of model parameters
[21] A.M. Schulte, Compositional variations within a hydrocarbon column due to
t transpose matrix gravity, in: SPE Annual Technical Conference and Exhibition, American Institute
of Mining, Metallurgical, and Petroleum Engineers, Inc., Dallas, TX, 1980.
[22] K. Ghorayeb, A. Firoozabadi, Molecular, pressure, and thermal diffusion in non-
Greek letters ideal multicomponent mixtures, AIChE Journal 46 (2000) 883–891.
 chemical potential [23] G. Soave, Equilibrium constants from a modified Redlich–Kwong equation of
ϕ fugacity coefficient state, Chemical Engineering Science 27 (1972) 1197–1203.
[24] A. Péneloux, E. Rauzy, R. Fréze, A consistent correction for
Redlich–Kwong–Soave volumes, Fluid Phase Equilibria 8 (1982) 7–23.
Subscripts [25] K.S. Pedersen, A.L. Blilie, K.K. Meisingset, PVT calculations on petroleum reser-
i component i voir fluids using measured and estimated compositional data for the plus
fraction, Industrial & Engineering Chemistry Research 31 (1992) 1378–1384.
[26] J. Creek, M. Schrader, East painter reservoir: an example of a compositional
Superscripts gradient from a gravitational field, in: SPE Annual Technical Conference and
◦ Exhibition, SPE 14441, presented at SPE ATCE, 22–25 September, Las Vegas,
reference depth
NV, 1985.
* warning Leveragesh depth [27] A. Eslamimanesh, F. Gharagheizi, A.H. Mohammadi, D. Richon, A statistical
method for evaluation of the experimental phase equilibrium data of simple
clathrate hydrates, Chemical Engineering Science 80 (2012) 462–468.
References [28] P.J. Rousseeuw, A.M. Leroy, Robust Regression and Outlier Detection, John
Wiley & Sons, New York, 1987.
[1] K.S. Pedersen, P.L. Christensen, S.J. Azeem, Phase Behavior of Petroleum [29] C.R. Goodall, Computation Using the QR Decomposition, Handbook in Statistics,
Reservoir Fluids, Taylor & Francis Group, CRC Press, Boca Raton, FL, USA, vol. 9, Elsevier/North-Holland, Amsterdam, 1993.
2007. [30] P. Gramatica, Principles of QSAR models validation: internal and external, QSAR
[2] C. Lira-Galeana, A. Firoozabadi, J.M. Prausnitz, Computation of compositional & Combinatorial Science 26 (2007) 694–701.
grading in hydrocarbon reservoirs. Application of continuous thermodynamics, [31] J.M. Prausnitz, R.N. Lichtenthaler, E.G. de Azevedo, Molecular Thermodynamics
Fluid Phase Equilibria 102 (1994) 143–158. of Fluid-Phase Equilibria, Prentice Hall, Inc., Upper Saddle River, NJ, 1998.
[3] A. Hirschberg, Role of asphaltenes in compositional grading of a reser- [32] J. Smith, Van ness, M.M. Abbott, Introduction to Chemical Engineering Thermo-
voir’s fluid column, SPE Journal of Petroleum Technology 40 (1988) dynamic, vol. 329, McGraw-Hill, Boston, 2001, pp. 354.
89–94. [33] H.C. Van Ness, M.M. Abbott, Classical Thermodynamics of Nonelectrolyte Solu-
[4] A. Danesh, PVT and Phase Behaviour of Petroleum Reservoir Fluids, Elsevier, tions: With Applications to Phase Equilibria, McGraw-Hill, New York, 1982.
Amsterdam, The Netherlands, 1998. [34] J.D. Raal, A.L. Mühlbauer, Phase Equilibria: Measurement and Computation,
[5] A. Fazlali, M. Nikookar, A. Agha-Aminiha, A.H. Mohammadi, Prediction of Taylor & Francis, Washington, D.C., 1998.
minimum miscibility pressure in heavy oil reservoirs using a modified SAFT [35] B.E. Poling, J.M. Prausnitz, J.P. O’Connell, The Properties of Gases and Liquids,
equation of state, Fuel 108 (2013) 675–681. 5th ed., McGraw-Hill, New York, 2001.
[6] A. Shokrollahi, M. Arabloo-Nareh, F. Gharagheizi, A.H. Mohammadi, Intelligent [36] A. Eslamimanesh, A.H. Mohammadi, D. Richon, Thermodynamic consistency
model for prediction of CO2 – reservoir oil minimum miscibility pressure, Fuel test for experimental data of water content of methane, AIChE Journal 57 (2011)
112 (2013) 375–384. 2566–2573.