Beruflich Dokumente
Kultur Dokumente
Faculty of Engineering and Applied Science, Memorial University of Newfoundland, St. Johns, NL, Canada
Southern Cross University, School of Environment, Science and Engineering, Lismore, NSW, Australia
a r t i c l e
i n f o
Article history:
Received 26 December 2014
Received in revised form 22 March 2015
Accepted 23 March 2015
Available online 31 March 2015
Keywords:
Carbon dioxide
Compressibility factor
Supercritical uids
Least square support vector machine
a b s t r a c t
Owing to the demanding applications and wide uses of supercritical carbon dioxide in oil, gas and
chemical industries, fast and precise estimation of carbon dioxide compressibility factor is of a vital
signicance in order to be imported into the relevant industrial simulators. In this study, a data bank covering wide range of temperature and pressure was gathered from open literature. Afterwards, a rigorous
novel approach, namely least square support vector machine (LSSVM) optimized with coupled simulated
annealing (CSA) was proposed to develop a reliable and robust model for the prediction of compressibility factor of carbon dioxide. Reduced temperature and pressure are the inputs of the model. 80% of
the dataset was used for training the model and the remaining 20% was used to evaluate its accuracy
and reliability. Statistical and graphical error analyses have been conducted to investigate the performance of the model and the obtained results from the proposed model have been compared with those
of six equations of state, REFPROP package and two correlations. It was demonstrated that the proposed
CSALSSVM model is more efcient and reliable than all of the studied empirical correlations, equations
of state and the software package, hence it can be utilized condently for the prediction of carbon dioxide
compressibility factor.
2015 Elsevier B.V. All rights reserved.
1. Introduction
Supercritical carbon dioxide, as the most commonly used supercritical uid, has been used in polymer processing as a solvent,
anti-solvent or plasticizer which reduces the system viscosity
and allows lower temperatures of the process. The low temperature sterilization technology by means of supercritical CO2 can
be incorporated in removing hazardous contaminants in food and
pharmaceutical industries. It can be utilized as a cleaning agent
in micro-electronics, for continuous hydrogenation, as a solvent in
oxidation processes and also can be combined with biocatalysts to
enhance their activity and stability [1].
Moreover, it has been used for more than thirty years in
enhanced oil recovery which can be combined with carbon
sequestration in mature oil elds [2,3] and also has found some
applications as a refrigerant [2] and a working uid in gas turbines
providing high efciency [4].
The compressibility factor (Z-factor) of CO2 is vital in most
chemical engineering calculations and design of processing units
due to its effect on mathematical models and also in oil and gas
industry for CO2 compression, design of pipeline, material balance
calculations and surface facilities design [5].
Compressibility factor as a thermodynamic property is obtained
via experimental laboratory procedures. Experimental measurements are usually expensive, time consuming and cumbersome [6],
and while there is lack of experimentally measured data, engineers
would have to determine them through equations of state (EOSs)
or experimentally derived correlations [7].
The challenge with equations of state is that they are all implicit
in terms of Z-factor. The compressibility factor has to be determined as the EOS root which yields the minimum free Gibbs energy.
The mathematical approach to determine Z-factor is lengthy and
requires proper root selection which adds to numerous engineering
calculations [8]. An extremely large number of carbon dioxide Zfactors is required in the simulation of CO2 injection as an enhanced
oil recovery method in oil reservoirs or for the purpose of CO2
sequestration and the large number of root selections adds to the
simulations run time. Correlations are usually faster and much easier to use [6]; however, they sometimes involve multiple steps
and complicated calculations and are limited to the data for which
the correlation was formulated [9]. A small error in Z-factor will
be propagated in the prediction of other properties such as gas
isothermal compressibility, gas formation volume factor for carbon storage calculations and thermal conductivity of supercritical
carbon dioxide [10].
According to the above, the need is felt to nd a fast and reliable
method which can predict the compressibility factor of supercritical carbon dioxide quickly and accurately. Recently, support vector
machine (SVM) as a supervised learning algorithm has been a tool
for classication and regression analysis. It has found applications
in PVT properties estimation, determination of porosity and permeability from log data, text categorization, protein classication
in medical science, etc. In a modied version of support vector
machine, namely least square support vector machine (LSSVM), the
inequality constraint is replaced by an equality, which eliminates
the necessity of solving a difcult quadratic programming problem. The LSSVM has shown high capability in solving nonlinear and
complex problems and it has been applied for the prediction of several chemical properties of oil and gas [1117]. To the best of the
authors knowledge, this method has not yet been applied for the
prediction of the Z-factor of CO2 .
In this study, a collection of 178 data points for the compressibility factor of supercritical carbon dioxide as a function of
temperature and pressure up to 1273.15 K and 50 MPa has been
gathered from the literature [18]. Afterwards, least square support vector machine (LSSVM) has been applied to predict the
Z-factor of CO2 and coupled simulated annealing (CSA) has been
utilized as an optimization technique to determine the LSSVM
hyper-parameters. Next, statistical and graphical error analyses
have been conducted to measure the accuracy and reliability of
the model, and nally, the results have been compared with six
equations of state, namely PengRobinson EOS, RedlichKwong
EOS, SoaveRedlichKwong EOS, SchmidtWenzel EOS, PatelTeja
EOS, and LawalLakeSilberberg EOS, a software package named
REFPROP developed by NIST as well as two correlations, namely
BahadoriVuthaluru correlation and the Virial coefcients correlation.
2. Compressibility factor
Gas compressibility or deviation factor (Z-factor) is dened
as the ratio of the actual volume to the ideal volume and is an
indication of the gas deviation from ideal behaviour. Some thermodynamic properties such as density, isothermal compressibility
and viscosity can be calculated by means of Z-factor.
According to the law of corresponding states proposed initially
by van der Waals, all gases behave the same at the same reduced
pressure Pr and reduced temperature Tr [6]. In the mathematical
form:
Z = f (Tr , Pr )
(1)
where
Tr =
T
Tc
(2)
Pr =
P
Pc
(3)
141
RT
vb
(4)
v2 + uv w2
The above general equation can be rewritten in terms of compressibility factor as a cubic equation for calculation of volumetric
behaviour:
Z 3 (1 + B U) Z 2 + A BU U W 2 Z
AB BW 2 W 2 = 0
(5)
aP
(RT )2
(6)
B=
bP
RT
(7)
U=
uP
RT
(8)
W=
wP
RT
(9)
R2 Tc2
(Tr , )
Pc
(10)
b = b
RTc
Pc
(11)
where is the temperature dependency term, is the acentric factor, a and b are the generalized coefcients of the
parameters in the original van der Waals equation which can
be constant or functions of acentric factor and critical compressibility factor in the modied equations of state [19]. Six
of widely used equations of states consisting of RedlichKwong
(RK) [20], SoaveRedlichKwong (SRK) [21], PengRobinson
(PR) [22], SchmidtWenzel (SW) [23], PatelTeja (PT) [24] and
LawalLakeSilberberg (LLS) [25] equations of state are reviewed
in Section A in the Supplementary information.
In 1996, Span and Wanger derived an EOS specically for CO2
by applying modern strategies for the optimization of the mathematical form of the EOS and for the simultaneous nonlinear t
to the experimental data. This EOS covers the wide temperature
and pressure range of 216 K (triple point of CO2 ) to 1100 K and 0
to 800 MPa, respectively, is an empirical representation of the fundamental equation explicit in Helmholtz free energy as a function
of density and temperature and is claimed to be accurate [26]. The
drawback of this EOS for calculation of Z-factor making it improper
for engineering purposes is rst that it is implicit in density. As the
result, to calculate density and then Z-factor, the function should
be inverted and since there is no analytic form for that, an iterative method such as Newtons method should be employed and an
initial guess for density should be made for the pressure to converge in an iterative procedure. Apart from that, the derivative
of the Helmholtz energy must be calculated and there are many
coefcients and exponents in the equation (more than 180). These
features make the computational procedures lengthy and unreasonable to be employed as an engineering tool. The fundamental
equations and the equations for density calculation can be viewed
in Section A in the Supplementary information as well.
142
3. Model development
3.1. Data gathering
In order to develop the proposed model and perform the task
mentioned, 178 data points from Perrys chemical engineers handbook [18] consisting of temperatures up to 1273.15 K and pressures
up to 50 MPa as the input and experimentally derived compressibility factors for carbon dioxide as the output of the model were
gathered. This dataset was randomly divided into two subsets of
training and testing in a way that data points are accumulated
homogenously and different operating conditions are covered in
each subset. The data distribution between the subsets is obtained
by means of trial and error. A proper distribution should give optimal statistical error measures (explained later) for each subset as
well as nearest mean and standard deviations of the subsets for
the purpose of homogeneous accumulation. 80% of the dataset has
been used for training purpose; i.e., for constructing the model,
and the rest has been utilized to test the prediction capability of
the model for unused data. Other distributions such as 5050%,
6040%, 7030% and 9010% respectively for training and testing,
have also been examined during the trial and error procedure of the
model development and the 8020% conguration was known to
yield the best results. It is worth to note that the model developed
in this study is only valid for the mentioned range of temperature
and pressure and any extrapolation should be avoided.
(12)
where wT and (x) are the transposed output layer vector and kernel function, respectively, and b is the bias. The model input (x)
is an N n matrix, where N and n represent the number of data
points and the number of input parameters, respectively. Vapnik
proposed the minimization of the following objective function in
order to calculate w and b [32]:
1
Objective function = wT + c
k k
2
N
i=1
(13)
y wT (xk ) b + k ,
k = 1, 2, . . ., N
wT (xk ) + b yk k ,
k = 1, 2, . . ., N
k , k 0,
k = 1, 2, . . ., N
(14)
where xk is the kth input and yk is the kth output data point.
stands for the xed precision of the function approximation, and k
and k are slack variables. Slack variables are normally utilized to
specify the allowed error bounds. c > 0 is the tuning parameter of
the objective function which determines the amount of deviation
from the preferred .
A difcult quadratic programming problem has to be solved in
order to nd the parameters of SVM. Therefore, to decrease the
complication of the SVM method, Suykens and Vandewalle [31,33]
reformulated and modied it to LSSVM, which can be solved by
means of linear programming and is faster than the conventional
SVM while maintains its advantages [11]. The LSSVM method is
expressed in the following format [32]:
1 T
1 2
ek
w w+
2
2
N
Objective function =
(15)
k=1
(16)
1 T
1 2 T
ek
ak w (xk ) + b + ek yk
w w+
2
2
N
k=1
k=1
(17)
N
w
=
ak (xk )
k=1
N
ak = 0
k=1
ak = ek ,
(18)
k = 1, 2, . . ., N
yk = w (xk ) + b + ek ,
k = 1, 2, . . ., N
xk x
2
(19)
143
Fig. 1. A typical owchart for the CSALSSVM algorithm used in this study.
square error (MSE) of the outputs resulted from the LSSVM model
is calculated through the following formula:
n
MSE =
Zi,exp Zi,pred
i=1
2
(20)
where Z is the output compressibility pred and exp for the predicted
and experimental values, respectively, and n represents the number
of samples from the initial population. The LSSVM model developed
by Suykens and Vandewalle [31] has been utilized in this study.
The LSSVM tuning parameters can be optimized by means
of coupled simulated annealing (CSA) whose detailed explanations can be found elsewhere [38]. In this study, CSALSSVM has
been applied for the prediction of compressibility factor of carbon dioxide. A schematic owchart of obtaining this model and
its parameters has been shown in Fig. 1.
4. Error analysis
To evaluate the accuracy and validity of a model, several statistical parameters such as average percent relative error (Er %), average
absolute percent relative error (Ea %), standard deviation (SD), root
mean square error (RMSE) and coefcient of determination (R2 )
are normally used [3943]. The equations for these parameters are
reviewed in Section C in the Supplementary information.
Fig. 2. Average absolute percent relative error for all the investigated methods.
144
Table 1
Statistical parameters for all the investigated methods.
Method
Er%
Ea%
SD
RMSE
R2
-1.52
1.53
0.0351
0.1354
0.9118
Redlich-Kwong EOS
Bahadori-Vuthaluru correlation
Lawal-Lake-Silberberg EOS
Soave-Redlich-Kwong EOS
Schmidt-Wenzel EOS
Patel-Teja EOS
Peng-Robinson EOS
REFPROP package
CSA-LSSVM, testing set
CSA-LSSVM, training set
1.45
-0.07
1.39
-1.05
-1.01
-0.51
-0.15
-0.19
-0.04
-0.01
1.51
1.3
1.49
1.05
1.08
0.65
0.44
0.41
0.19
0.16
0.0216
0.0258
0.0224
0.0186
0.0175
0.0118
0.0081
0.0112
0.0034
0.0049
0.087
0.073
0.0672
0.0465
0.0437
0.0176
0.0074
0.0115
0.003
0.003
0.9433
0.9524
0.9562
0.9697
0.9715
0.9886
0.9952
0.9853
0.9983
0.9991
Fig. 3. Error distribution graphs for REFPROP package, PR and PT EOSs and the CSALSSVM proposed model.
145
Fig. 4. Cumulative frequency vs. absolute percent relative error for all the investigated methods.
dispersion around the zero error line, which means that its prediction is the closest to the experimental values. In addition, as can
be seen, all of the models have better capability for prediction of
CO2 Z-factor at higher temperatures.
Afterwards, in order to carry on the objectives of this study, the
cumulative frequency versus the absolute percent relative error for
all the investigated methods and the proposed model have been
depicted in Fig. 4. As shown in Fig. 4, the proposed CSALSSVM
model predicts around 70% of the data points with absolute relative error smaller than 0.1% and more than 90% of the data points
with absolute relative error smaller than 0.5%. Even the second
most accurate method, REFPROP package, only predicts 40% of the
data points with absolute relative error smaller than 0.1%, which is
an additional proof of the better accuracy and performance of the
proposed model.
A real gas deviates from ideality at high pressures following
an observed experimental trend depending on its temperature.
As a matter of fact, at lower temperatures the compressibility
factor of the gas initially decreases and after reaching a minimum
increases with increasing pressure, while at higher temperatures
the minimum is not observed and the compressibility factor
increases monotonically with increasing pressure. This trend has
to be and is observed for the predicted results of the proposed
model as an evidence of its accuracy and reliability. Therefore, the
prediction of the model in the form of trend lines at three distinct
Fig. 5. CO2 Z-factor at three distinct reduced temperatures vs. reduced pressure for the CSALSSVM model and experimental (Exp) data.
146
Table 2
Predicted CO2 Z-factor for the testing set of data points of the CSALSSVM model.
P (MPa)
T (K)
Pr
Tr
Z-Exp.
Z-Pred.
0.1
0.5
0.5
0.5
0.5
673.15
373.15
423.15
573.15
823.15
0.0135
0.0676
0.0676
0.0676
0.0676
2.2124
1.2264
1.3907
1.8838
2.7055
1.003
1.0056
1.1067
1.0015
1.0019
1.0026
1.0075
1.1067
1.0016
1.0016
0.036
-0.197
-0.003
-0.013
0.025
0.5
1
1
1
4
4
4
4
973.15
523.15
573.15
973.15
473.15
623.15
673.15
1173.15
0.0676
0.1353
0.1353
0.1353
0.5412
0.5412
0.5412
0.5412
3.1985
1.7194
1.8838
3.1985
1.5551
2.0481
2.2124
3.8558
1.0083
0.9416
0.9983
0.9508
1.0093
1.0458
0.9253
0.9938
1.0082
0.9408
0.9973
0.9618
1.0097
1.0517
0.9214
0.9948
0.008
0.077
0.09
-1.161
-0.043
-0.564
0.419
-0.105
4
6
6
6
6
6
6
8
1273.15
573.15
623.15
723.15
773.15
823.15
1173.15
1273.15
0.5412
0.8119
0.8119
0.8119
0.8119
0.8119
0.8119
1.08254
4.1845
1.8838
2.0481
2.3768
2.5411
2.7055
3.8558
4.1845
1.0171
1.0079
0.9896
1.0002
0.9195
0.7264
0.9984
1.0128
1.0172
1.0109
0.9918
0.9998
0.9189
0.7172
0.9968
1.0127
-0.013
-0.301
-0.229
0.037
0.063
1.264
0.152
0.007
10
10
10
20
20
30
30
30
673.15
873.15
1273.15
573.15
1373.15
823.15
1073.15
1273.15
1.3531
1.3531
1.3531
2.7063
2.7063
4.0595
4.0595
4.0595
2.2124
2.8698
4.1845
1.8838
4.5132
2.7055
3.5271
4.1845
1.0017
0.9822
0.9883
0.9621
1.0218
1.1318
0.8995
1.0166
1.002
0.9849
0.988
0.962
1.0218
1.1341
0.9026
1.0172
-0.031
-0.283
0.025
0.001
-0.005
-0.206
-0.348
-0.061
40
40
40
40
40
50
50
573.15
623.15
673.15
823.15
1173.15
823.15
873.15
5.4127
5.4127
5.4127
5.412
5.4127
6.765
6.7659
1.8838
2.0481
2.2124
2.7055
3.8558
2.7055
2.8698
0.9982
0.9773
1.0266
1.0709
0.9658
1.0107
1.0062
0.9976
0.9796
1.0258
1.0718
0.9631
1.0136
1.006
0.051
-0.236
0.075
-0.087
0.273
-0.295
0.014
6. Conclusion
LSSVM has been used in this work to predict the Z-factor of carbon dioxide in the temperature and pressure range which mostly
includes the supercritical CO2 . The tuning parameters of LSSVM
have been optimized by means of coupled simulated annealing
(CSA). A dataset consisting of 178 data points for the Z-factor of CO2
was collected from open literature. Statistical analysis provided an
average absolute percent relative error of 0.19% and R2 of 0.9983
for the previously unused testing dataset. The comparison with six
equations of state, a software package and two empirical correlations showed the better capability of this model for predicting
the Z-factor of CO2 and proved it to be more accurate. The validity
of the model was also shown as it follows the expected physical
trend for compressibility factor as a function of reduced pressure
and temperature. The model has the potential to be imported and
implemented in chemical engineering and reservoir engineering
software and offers increased accuracy and less computational time
for the prediction of CO2 compressibility factor.
147
0.55
0.15
m1 + 2( 0.4)m2
(38)
(39)
(40)
where
Appendix A.
A.1. Section A
a = 0.42747,
(22)
b = 0.08664
1
=
Tr
(23)
2
(24)
Tr
where
m=
(42)
(41)
c = 1 3
= 1+m 1
RTc
Pc
c = c
(25)
1.18
(26)
a = 0.457235,
(27)
b = 0.077796
(28)
(29)
a = [1 (1 q)]3
(30)
(43)
(44)
(45)
(46)
(47)
=
1 + w 3 Zc
(48)
w Zc
w =
3
Zc2 w 1
(49)
2Z +
+ 2w
(1 3Zc ) w
c
(50)
2Z
w
c
3
a = 1 + w 1 Zc
(51)
b = w Zc
(52)
(6 + 1)q3 + 3q2 + 3q 1 = 0
(31)
then
1
=
3(1 + q)
(32)
and
b = q
(33)
for
for
0.4
0.55
for
for 0.3671
> 0.3671
(54)
(55)
T
Tc
(56)
(35)
(36)
0 (,
) = ln() + a01 + a02
+ a03 ln(
) +
(34)
where
2
(53)
8
i=4
(37)
a0i ln 1 exp
i0
(57)
148
r (,
)
7
di ti
ni
+
i=1
2
(1
) + Ai
ci
2 1/(2i ) 2
+ Bi
C = C0 + C1
i (
i )2
ni bi exp Ci ( 1) Di (
1)2
i=40
ni
exp
ni di
ti exp i i
i=35
42
di ti
i=8
39
34
C0 = 0.01407 +
(58)
2 ai
= A1 (A ( , T ))
P = RT 1 +
C1 = 0.02676 +
(59)
+ 2 + 2
Tr
Tr
Tr
(62)
Ei =
Zi,exp Zi,pred
Zi,exp
B1
C1
D1
+ 2 + 3
Pr
Pr
Pr
(63)
= A2 +
B2
C2
D2
+ 2 + 3
Pr
Pr
Pr
(64)
= A3 +
B3
C3
D3
+ 2 + 3
Pr
Pr
Pr
(65)
= A4 +
B4
C4
D4
+ 2 + 3
Pr
Pr
Pr
(66)
(67)
where
B1 = 0.139
(68)
0.172
Tr4.2
100
i = 1, 2, 3, . . ., n
(75)
1
Ei
n
(76)
A.3.3. Root
mean square error
n
1
2
RMSE =
Zi,exp Zi,pred
(69)
(70)
(77)
i=1
A.3.4.
Standard deviation
2
n
1
Zi,exp Zi,pred
SD =
n1
i=1
Zi,exp
(78)
= A1 +
(74)
where
Tr1.6
(73)
i=1
0.422
Tr10.5
i=1
A.2. Section B
B0 = 0.083
0.00242
1
Ei
n
B = B0 + B1
(61)
So, an initial guess should be made for density and then pressure
should be found knowing the temperature and density. The iteration continues until the calculated and known P become sufciently
close.
ln(Z) = +
Tr2.7
Er =
(60)
0.05539
(72)
0.02432
0.00313
Tr
Tr10.5
A.3. Section C
The coefcients and exponents of the above equations are tabulated in [26].
Since the Helmholtz EOS is a function of density, it should be
inverted:
(71)
R2 = 1
Zi,exp Zi,pred
2
i=1
n
Zi,pred Z
2
(79)
i=1
References
149