COVER PAGE

Gas-condensate Reservoir Performance

Modelling.

PhD Thesis

Foad Faraji

A thesis submitted in partial fulfilment of the requirements of the

Teesside University for the degree of Doctor of Philosophy

December 2020
 ABSTRACT
Accurate prediction of gas-condensate reservoir performance below the saturation
pressure is an inherent problem. This is due to the compositional variation and phase
change during the depletion process. In doing so, ensuring accurate reservoir
performance modelling for various pressure – volume – temperature (PVT) properties
such as two phase “gas/condensate” viscosities and compressibility factor (Z factor)
in desired reservoir conditions are particularly important. However, the existing
viscosities and Z factor models cannot capture fluid flow complexity of gas-condensate
reservoirs below the saturation pressure for modelling purposes.
The major contribution of the thesis is development of new gas/condensate viscosity
and two-phase Z factor models using comprehensive experimental data sets. The
data sets are representing downhole and reservoir condition. In the development
process, an investigation on the use of soft computing techniques such as Support
Vector Machine (SVM), Artificial Neural Network (ANN) and fuzzy logic (Mamdani &
TSK) has been carried out. It is found that developed TSK fuzzy logic approaches offer
the most accurate viscosity and two-phase Z factor prediction. The developed models
can predict viscosity and two-phase Z factor of gas-condensate reservoirs in high
pressure high temperature (HPHT) conditions with variety of non-hydrocarbon
contents and they are not limited within geographical location.
The impact of viscosity and two-phase Z factor models towards the production
calculation was the ultimate interest of this research. This led to further contribution on
proposing the new method for computation of gas-condensate reservoir production
rate performance, which involves integrating pseudopressure integral with volumetric
material balance. For the computation of production rate, dynamic three-phase
effective permeability has also been adopted. Distinctively, the proposed method
provides better level of accuracy to compositional commercial simulation software in
term of production forecast and economic impact of gas-condensate wells.
Furthermore, the proposed method offers simpler computational procedures, where
less input parameters are required.

Keywords: Gas-condensate, gas-condensate PVTs, Viscosity, Compressibility factor

(Z factor), Soft Computing approach, Artificial Neural Network (ANN), Fuzzy Logic.

i
 ACKNOWLEDGEMENT

I would like to express my sincere appreciation to my supervisory team, Dr. Johnson

Ugwu, Professor. Farhad Nabhani and Dr. Perk Lin Chong for their continuous support
and guidance throughout this research. I extend my appreciation to Teesside
University for providing the fund under graduate tutor scheme. I also would like to
thanks all my friends for their support during my study.

ii | P a g e
 LIST OF PUBLICATIONS

Published Articles:
1. Development of inflow performance model in high temperature gas-condensate
reservoirs published in Journal of Petroleum Science and Engineering (JPSE)
https://www.sciencedirect.com/science/article/abs/pii/S0920410519305819

2. Modelling viscosity of liquid dropout near wellbore region in gas condensate

reservoirs using modern numerical approaches published in Journal of Petroleum
Science and Engineering (JPSE).
https://www.sciencedirect.com/science/article/abs/pii/S0920410519310253

3. Development of new gas-condensate viscosity correlation using TSK fuzzy logic

approach (Under review)

4. Modelling of gas condensate viscosity using fuzzy logic based on different

membership functions (Under review).

iii | P a g e
 TABLE OF CONTENTS

COVER PAGE ........................................................................................................... 0

ABSTRACT ................................................................................................................. i

ACKNOWLEDGEMENT ..............................................................................................ii

LIST OF PUBLICATIONS .......................................................................................... iii

TABLE OF CONTENTS .............................................................................................iv

LIST OF TABLES .......................................................................................................ix

LIST OF FIGURES .....................................................................................................xi

NOMENCLATURE ................................................................................................... xvi

CHAPTER 1 INTRODUCTION ................................................................................... 1

1.1 Background ...................................................................................................... 1

1.2 Gas-condensate characteristics........................................................................ 3

1.3 Gas-condensate flow behaviour ....................................................................... 5

Phase behaviour ........................................................................................ 5

Drawdown behaviour ................................................................................. 6

1.4 Problem Statement ........................................................................................... 8

1.5 Research aim and objectives .......................................................................... 10

1.6 Proposed approach ........................................................................................ 11

1.7 Organization of the thesis ............................................................................... 11

CHAPTER 2 WELL DELIVERABILITY MODELLING OF GAS-CONDENSATE

RESERVOIRS .......................................................................................................... 15

2.1 Introduction ..................................................................................................... 15

2.2 Gas Rate Theory ............................................................................................ 16

2.3 Well Deliveraility Theory ................................................................................. 19

2.3.1 Modelling Fluid Properties ........................................................................ 26

2.3.2 Compressibility Factor (Z factor) .............................................................. 31

iv | P a g e
 2.3.3 Density of the mixture .............................................................................. 32

2.3.4 Formation Volume Factor of Gas-Condensate Mixture ............................ 34

2.3.5 Solution Gas to Oil Ratio (Rs) .................................................................. 40

2.3.6 Viscosity (Gas/Condensate) ..................................................................... 42

2.3.7 Drawdown Behaviour on modelling .......................................................... 46

2.4 Summary ........................................................................................................ 48

CHAPTER 3 METHODOLOGY AND CONCEPTUAL FRAMEWORK ..................... 50

3.1 Introduction ..................................................................................................... 50

3.2 Pressure – Volume –Temperature (PVT) modelling ....................................... 50

3.3 Approaches of Well Deliverability Modelling ................................................... 51

3.4 Current Study Approach ................................................................................. 53

3.5 Major Work Steps ........................................................................................... 54

3.6 Data Acquisition and Validation ...................................................................... 55

3.7 Summary ........................................................................................................ 56

CHAPTER 4 IMPROVEMENT OF GAS-CONDENSATE FLUID VISCOSITY

PREDICTION ........................................................................................................... 57

4.1 Introduction ..................................................................................................... 57

4.2 Assessment of gas phase viscosity models .................................................... 58

Lohrenz – Bray – Clark (LBC), 1964 ........................................................ 58

Lee – Gonzalez – Eakin (LGE)................................................................. 60

Londono – Archer – Blasingame (LAB) .................................................... 61

Sutton, (2005) .......................................................................................... 62

Elsharkawy (2006) ................................................................................... 63

4.3 Assessment of condensate (oil) phase viscosity models ................................ 66

Data acquisition........................................................................................ 67

Beggs and Robinsons, (1975) .................................................................. 68

De Ghetto et al., (1994) ............................................................................ 69

v|Page
 Elsharkawy and Alikhan (1999) ................................................................ 69

Bergman and Sutton, (2007) .................................................................... 70

Kartoatmodjo and Schmidt, (1991) ........................................................... 70

Results of existing models........................................................................ 71

4.4 Machine Learning (ML) Approach................................................................... 75

Support Vector Machine (SVM)................................................................ 77

Artificial Neural Network (ANN) ................................................................ 84

Results and discussion ............................................................................ 90

4.5 Fuzzy Logic Approach .................................................................................... 94

4.6 Takagi – Sugeno – Kang (TSK) Fuzzy approach............................................ 97

Fuzzy Clustering ...................................................................................... 99

Setting the membership function ............................................................ 101

Determination of constants..................................................................... 103

Development of new correlation ............................................................. 105

Results and discussions ......................................................................... 106

4.7 Mamdani Fuzzy Approach ............................................................................ 114

Fuzzy Engine process ............................................................................ 114

Membership Functions (MFs) fuzzification ............................................. 116

Mamdani Fuzzy Inference System ......................................................... 119

Defuzzification in Mamdani approach .................................................... 121

Results and Discussion .......................................................................... 122

4.8 Computational efficiency of the developed models ....................................... 127

4.9 Summary ...................................................................................................... 128

CHAPTER 5 DEVELOPMENT OF TWO-PHASE Z FACTOR OF GAS-CONDENSATE

MIXTURE ............................................................................................................... 130

5.1 Introduction ................................................................................................... 130

5.2 Preliminary theory of Z factor calculation ...................................................... 131

vi | P a g e
 5.3 Assessment of Z factor models..................................................................... 133

Hall-Yarborough (1973) .......................................................................... 135

Dranchuk – Abu – Kassem (1975) ......................................................... 136

Beggs and Brill (1973) ............................................................................ 137

Rayes et al., (1992) ................................................................................ 137

Azizi et al., (2010) .................................................................................. 138

Data bank ............................................................................................... 139

Results of literature models .................................................................... 141

5.4 Development of new Z factor models ........................................................... 148

Cascade Forward Neural Network (CFNN) ............................................ 149

Feed Forward Backpropagation Neural Networks .................................. 152

Adaptive Neuro Fuzzy Inference System (ANFIS) ................................. 157

Proposed ANFIS structure ..................................................................... 161

Clustering (Partitioning) of the data ........................................................ 166

Results and discussion .......................................................................... 172

5.5 Summary ...................................................................................................... 186

CHAPTER 6 PRODUCTION PROFILE FORECAST USING WELL TEST DATA .. 188

6.1 Introduction ................................................................................................... 188

6.2 Pressure Transient Test (PTA) and two-phase pseudopressure approach .. 189

6.3 Computation of production profile in relation to gas/condensate viscosity and Z

factor in HPHT gas-condensate well................................................................... 196

Construction of pressure – volume – temperature (PVT) relationship .... 199

Mathematical manipulation of pseudopressure integral ......................... 200

Material balance calculation ................................................................... 207

Validation of new production profile ....................................................... 210

Results and discussion .......................................................................... 213

6.4 Summary ...................................................................................................... 219

vii | P a g e
CHAPTER 7 CONCLUSIONS AND RECOMMENDATIONS ................................. 221

7.1 Introduction ................................................................................................... 221

7.2 Improve accuracies of gas-condensate viscosity .......................................... 221

Improvement of two-phase gas-condensate Z factor ............................. 222

Effectiveness of viscosities and two-phase Z factor models for producing

accurate production profile .............................................................................. 224

7.3 Recommendation for future work .................................................................. 225

REFERENCES ....................................................................................................... 226

APPENDIX A: Sample of collected data for development of condensate viscosity

models.................................................................................................................... 265

APPENDIX B: The developed Matlab codes for prediction of Two-phase Z factor 266

B1. Cascade Forward Neural Network (CFNN) code ......................................... 266

B2. Feed Forward Neural Network (FFNN) code ................................................ 266

B3. The Developed ANFIS code for prediction of two-phase Z factor................. 267

APPENDIX C: Sample calculation of inflow performance relationship (IPR) for HTHP

gas condensate well ............................................................................................... 273

APPENDIX D: The developed Eclipse – 300 code for prediction of well inflow
performance of HPHT gas condensate well ........................................................... 274

viii | P a g e
 LIST OF TABLES

Table 1.1. Compositions and properties of several reservoir fluids (Whitson and Brulé,
2000, p. 2). ................................................................................................................. 4
Table 4.1. Statistical information of the data bank. ................................................... 68
Table 4.2. The original and tuned form of the employed literature correlations for
predicting condensate liquid viscosity. ..................................................................... 74
Table 4.3. The optimum values of the LSSVM parameters. ..................................... 82
Table 4.4. Statistical error performance of the LSSVM. ........................................... 84
Table 4.5. Statistical parameters of developed ML based models and utilized
correlations for prediction of condensate viscosity. .................................................. 92
Table 4.6. The range of constants and input parameters of proposed condensate
viscosity model. ...................................................................................................... 106
Table 4.7. Statistical accuracy of condensate phase viscosity models. ................. 108
Table 4.8. Fuzzy rules defined for estimation of condensate viscosity in gas-
condensate reservoirs. ........................................................................................... 121
Table 4.9. Statistical error comparison between Mamdani fuzzy approach with three
MFs and existing literature models for prediction of condensate viscosity. ............ 123
Table 5.1. Coefficients of Azizi et al., (2010) Z factor correlation. .......................... 139
Table 5.2. Statistical description of the data used for development of compressibility
factor. ..................................................................................................................... 141
Table 5.3. Performance of employed literature models in predicting two-phase Z factor
of gas-condensate systems.................................................................................... 142
Table 5.4. The relation between the number of neurons and coefficient of
determination for 1 hidden layer CFNN, using Levenberg – Marquardt (LM) as training
algorithm. ............................................................................................................... 151
Table 5.5. Statistical parameters and convergence time of FFNN with one hidden layer
and using LM algorithm. ......................................................................................... 155
Table 5.6. Statistical parameters and convergence time of FFNN with one hidden layer
and using BR algorithm for training. ....................................................................... 156
Table 5.7. The description and details of the utilized membership functions for
prediction of gas-condensate two-phase Z factor using ANFIS.............................. 163
Table 5.8. Performance of the developed ANFIS model using various membership
function................................................................................................................... 166

ix | P a g e
Table 5.9. Effect of clustering radius on number of fuzzy rules and the performance of
the model (1st iteration)........................................................................................... 169
Table 5.10. Effect of clustering radius on number of fuzzy rules and the performance
of the model (2st iteration)....................................................................................... 170
Table 5.11. Statistical error computation of the developed models in prediction of gas-
condensate’s two-phase Z factor below the saturation pressure............................ 173
Table 5.12. Constant Volume Depletion (CVD) results on North Sea gas-condensate
sample at temperature of 394K (Danesh, 1998, pp. 53–56)....................................... 1
Table 5.13. Constant Volume Depletion (CVD) results on sour gas-condensate sample
(Elsharkawy, 2002)..................................................................................................... 2
Table 5.14. Statistical measurement of the developed models and literature models
for prediction two-phase Z factor of two gas-condensate samples as a function of Ppr.
............................................................................................................................... 182
Table 6.1. Reservoir and fluid property of studied gas-condensate well after
(Economides et al., (1989). .................................................................................... 197
Table 6.2. Pressure build up data for (KAL – 5) obtained from (Economides et al.,
(1989). .................................................................................................................... 198
Table 6.3. Statistical error analysis of the developed analytical technique for predicting
gas and condensate rates in (KAL – 5). ................................................................. 217

x|Page
 LIST OF FIGURES
Figure 1.1. USA Natural Gas withdrawals (EIA, 2020). .............................................. 1
Figure 1.2. Breakdown of Gazprom Group’s explored hydrocarbon reserves in
Russian regions (Gazprom, 2019).............................................................................. 2
Figure 1.3. Typical phase diagram of gas condensate systems (Modified from Fan et
al., 2005). ................................................................................................................... 6
Figure 1.4. Gas/condensate flow in three regions. ..................................................... 8
Figure 1.5. The flow chart indicates organization of the thesis for each chapter. ..... 14
Figure 2.1. Typical plot of gas pressure function vs pressure where the area between
Pwf and PR representing the integral in 2.2. .............................................................. 19
Figure 2.2. Schematic diagram of traditional black oil formulation of PVT model. .... 27
Figure 2.3 schematic diagram of modified black oil (MBO) model. .......................... 29
Figure 2.4. Typical gas-condensate formation volume factor as a function pf pressure.
................................................................................................................................. 36
Figure 2.5. Solution gas to oil ratio (Rs) for North Sea gas-condensate sample at
T=280⁰F (Modified from Whitson and Torp, 1983). .................................................. 40
Figure 4.1. The prediction performance of existing literature models in estimating gas
phase viscosity of gas-condensate reservoirs below the saturation pressure. ......... 65
Figure 4.2. Plot of calculated viscosity using new developed model against measured
viscosity data. ........................................................................................................... 66
Figure 4.3. Cross plot of the experimental condensate viscosity versus predicted
condensate viscosity using employed literature correlations and refined results. .... 73
Figure 4.4 Machine Learning Block Diagram. .......................................................... 76
Figure 4.5. Performance of LSSVM in predicting condensate (oil) viscosity in training
and testing stage. ..................................................................................................... 83
Figure 4.6. Graphical error analysis of LSSVM performance for predicting condensate
(oil) viscosity. ............................................................................................................ 83
Figure 4.7. Schematic illustration of the ANN structure and computational steps to
measure output. ....................................................................................................... 85
Figure 4.8. Developed ANN model architecture for prediction of condensate viscosity.
................................................................................................................................. 88
Figure 4.9. Prediction performance of developed ANN network for condensate liquid
viscosity in training (A) and testing (B) stage. .......................................................... 89

xi | P a g e
Figure 4.10. Prediction performance of the developed LSSVM condensate viscosity
model as a function of pressure for three independent samples. ............................. 90
Figure 4.11. Prediction performance of developed ANN condensate viscosity model
as a function of pressure for three independent samples. ........................................ 90
Figure 4.12. Statistical performance comparison of tuned literature models against
developed ANN and LSSVM models (AARD% and RMSE) for prediction of
condensate viscosity. ............................................................................................... 92
Figure 4.13. Statistical performance comparison of tuned literature models against
developed ANN and LSSVM models (R2) for prediction of condensate viscosity. .... 93
Figure 4.14. Outline of the TSK fuzzy algorithm. ...................................................... 98
Figure 4.15. Optimum number of cluster for the training data sets using validity function
(VRC), for condensate viscosity input data. ........................................................... 101
Figure 4.16. Gaussian membership function for detecting fuzzy clusters............... 103
Figure 4.17. Gaussian membership function for input 1 “pressure”. ....................... 104
Figure 4.18. Architecture of constructed TSK fuzzy model for predicting condensate
viscosity. ................................................................................................................. 106
Figure 4.19. Cross plot of estimated against condensate viscosity measurements of
existing literature correlation, ANN, LSSVM and TSK Fuzzy method. ................... 109
Figure 4.20. Residual plot of relative error percentage for different viscosity models as
a function of reservoir pressure. ............................................................................. 111
Figure 4.21. Residual plot of relative error percentage for different viscosity models as
a function of reservoir temperature......................................................................... 112
Figure 4.22. Experimental prediction capability of the developed TSK fuzzy model as
a function pressure (A), solution gas to oil ratio (B) and temperature (C)............... 113
Figure 4.23. The relation between pressure and solution gas to oil ratio (Rs) for the
utilized databank. ................................................................................................... 114
Figure 4.24. General architecture of a Fuzzy Logic System (FLS). ........................ 115
Figure 4.25. The constructed gaussian membership function representing (A)
pressure (B) temperature (C) solution gas to oil ratio and (D) condensate viscosity.
............................................................................................................................... 117
Figure 4.26. The constructed triangular membership function representing (a)
pressure (b) temperature (c) solution gas to oil ratio and (d) condensate viscosity.
............................................................................................................................... 118

xii | P a g e
Figure 4.27. The constructed trapezoid membership functions (A) pressure (B)
temperature (C) solution gas to oil ratio and (D) the condensate viscosity............. 119
Figure 4.28. Architecture of the fuzzy inference (FIS) system for predicting condensate
viscosity using Gaussian membership function. ..................................................... 122
Figure 4.29. Performance prediction of Gaussian based MFs Mamdani fuzzy model in
predicting condensate viscosity. ............................................................................. 124
Figure 4.30. Performance prediction of triangular based MFs Mamdani fuzzy model in
predicting condensate viscosity. ............................................................................. 124
Figure 4.31. Performance prediction of trapezoidal based MFs Mamdani fuzzy model
in predicting condensate viscosity. ......................................................................... 125
Figure 4.32. Performance of utilized literature correlations and Mamdani fuzzy model
with gaussian MFs for predicting condensate viscosity as a function of pressure;
sample from Coats, (1986). .................................................................................... 126
Figure 4.33. Performance of literature correlations along with gaussian fuzzy approach
in predicting condensate viscosity as a function of solution gas to oil ratio Rs; sample
from Coats, (1986). ................................................................................................ 127
Figure. 4.34. Computational efficiency of the developed AI models (00:00:00 stands
for Hour: Minute: Second). ..................................................................................... 128
Figure 5.1. Scatter plot of five literature models in prediction of two-phase
compressibility factor of gas condensate reservoirs below the saturation pressure.
............................................................................................................................... 143
Figure 5.2. Graph of relative error utilizing literature models for prediction of two-phase
Z factor. .................................................................................................................. 144
Figure 5.3. Residual plot of relative error percentage of utilized literature correlations
in predicting gas-condensate two-phase Z factor as a function of pseudo reduced
pressure (Ppr). ........................................................................................................ 147
Figure 5.4. Residual plot of relative error percentage of literature models in predicting
two-phase Z factor of gas-condensate systems as a function of pseudo reduced
temperature (Tpr). ................................................................................................... 148
Figure 5.5. Architecture of the cascade forward (CFNN) neural network used in this
study....................................................................................................................... 149
Figure 5.6. Cascade forward neural network with three hidden layer constructed in
Matlab for prediction of two-phase Z factor. ........................................................... 150

xiii | P a g e
Figure 5.7. Feed Forward Artificial Neural Network structure used for prediction of Z
factor. ..................................................................................................................... 153
Figure 5.8. Comparison of constructed feed-forward neural network (FFNN)
performance with different number of neurons in hidden layer. .............................. 157
Figure 5.9. Proposed FFNN for prediction gas condensate two-phase compressibility
factor. ..................................................................................................................... 157
Figure 5.10. Schematic illustration of type – 3 fuzzy reasoning (a); ANFIS structure
type – 3 (b). ............................................................................................................ 159
Figure 5.11. Proposed ANFIS architecture for prediction of gas condensate two-phase
Z factor. .................................................................................................................. 162
Figure 5.12. Structure of proposed ANFIS model for prediction of two-phase Z factor
of gas condensate reservoirs. ................................................................................ 171
Figure 5.13. Architecture of Takagi – Sugeno fuzzy model for prediction of gas
condensate two-phase Z factor. ............................................................................. 171
Figure 5.14. Cross plot and relative error of utilized three machine learning models in
predicting two-phase Z factor at training, validation and testing stage. .................. 175
Figure 5.15. Residual plot of relative error percentage of intelligent models in
predicting two-phase Z factor as a function of pseudo reduced pressure (Ppr). ..... 175
Figure 5.16. Residual plot of relative error percentage of intelligent models in
predicting two-phase Z factor as a function of pseudo reduced temperature (T pr). 176
Figure 5.17. Error comparison of all utilized literature models and the developed ML
based approached in prediction of gas-condensate two-phase Z factor. ............... 176
Figure 5.18. Error comparison of all utilized literature models and the developed ML
based approaches in prediction of two-phase Z factor. .......................................... 177
Figure 5.19. Testing different methods in predicting experimental Z factor as a function
of Ppr for sour gas-condensate sample.................................................................. 184
Figure 5.20. Prediction of Z factor as a function of P pr for sweet gas condensate
sample.................................................................................................................... 185
Figure 5.21. Relevancy factor (r) for each input variables. ..................................... 185
Figure 6.1. Horner plot for pressure build up test (Earlougher, 1977). ................... 195
Figure 6.2. construction of PVT table for KAL – 5 gas-condensate well. ................ 200
Figure 6.3. Horner plot for KAL – 5 gas-condensate well. ...................................... 203
Figure 6.4. Plot of gas flow rate vs total ∆𝑚𝑃𝑔𝑡 for KAL – 5................................... 207
Figure 6.5. Schematic illustration of 3D created model in Eclipse 300. .................. 211

xiv | P a g e
Figure 6.6. Condensate saturation curve a result of the CVD experiment of the
reservoir fluid. ......................................................................................................... 211
Figure 6.7. Phase diagram of the reservoir fluid in standard pressure and temperature.
............................................................................................................................... 212
Figure 6.8. Variation of gas and oil (condensate) voscosity with pressure in KAL – 5.
............................................................................................................................... 213
Figure 6.9. Gas phase effective permeability. ........................................................ 215
Figure 6.10. Relative permeability ratio (gas to oil) as a function of pressure for high
temperature (354⁰F) gas condensate well. ............................................................. 215
Figure 6.11. Gas production profile for KAL – 5 using analytical method incorporated
with various ML techniques in generating PVT properties. ..................................... 216
Figure 6.12. Gas production profile for KAL – 5 using analytical method incorporated
with various ML techniques in generating PVT properties. ..................................... 216
Figure 6.13. Statistical error results between analytical pseudopressure approach and
compositional simulation results for perdition of gas phase production profile in KAL –
5. ............................................................................................................................ 218
Figure 6.14. Statistical error results between analytical pseudopressure approach and
compositional simulation results for perdition of oil phase production profile in KAL –
5. ............................................................................................................................ 218

xv | P a g e
 NOMENCLATURE

Symbols
A Cross sectional area of fluid flow, m2
API American Petroleum Institute, degree
ANN Artificial Neural Network
ANFIS Adaptive Neuro Fuzzy Inference System
b Bias unit of SVM function
Β Turbulent flow parameter
Bgd Dry gas formation volume factor , ft3/scf
Bg Gas formation volume factor, ft3/scf
BHFP Bottom-hole flowing pressure, psia
Bo Oil formation volume factor, bbl/STB
BO Black oil
Bob* Correlating parameter of oil formation volume factor
c Number of clusters
ct Total compressibility of fluid
CVD Constant Volume Depletion
D Non-Darcy flow multiplier in gas rate equation
dp/dt Pressure rate of change in time, psia
dp/dx Pressure gradient in x direction
dp/dr Pressure gradient in radial flow
EOS Equation of state
FL Fuzzy Logic
fX Regression function of SVM
g Sigmoid function of ANN network
HPHT high pressure high temperature
J Cost function in ANN and SVM
K absolute permeability, millidarcy
Krg gas phase relative permeability, millidarcy (md)
Kro oil phase relative permeability, millidarcy (md)
Kg/Ko Gas to oil permeability ratio
LSSVM Lease Square Support Vector
Mg Gas molecular weight, lbm/lbm mol
Mo Oil molecular weight, lbm/lbm mol
ML Machine Learning
MBO Modified black oil
Mi Molecular weight of each component
mp Pseudopressure function, MMpsi2/cp
MC7+ Molecular weight of hydrocarbon plus fractions
Mo Surface oil (condensate liquid) molecular weight
Deliverability exponent in well deliverability equation;
n
Mole of gas in ideal gas law, lb-mol
Nc Dimensionless capillary number
Pdew Dew-point pressure, psia
Pwf Bottom-hole flowing pressure, psia
Pr Average reservoir pressure, psia

xvi | P a g e
 Psc Pressure at standard condition, psia
Psep Separator pressure, psia
Ppc Pressure pseudocritical properties
PVT Pressure – Volume – Temperature
Pressure at interface between region 1 and region 2,
P*
psia
qg Surface gas flow rate, scf/day
qo Surface oil flow rate, STB/day
re External drainage radius, ft
rs Solution oil to gas ratio (OGR)
rw Wellbore radius, ft
rp Total producing gas to oil ratio for calculating well
stream gravity, scf/STB
R Universal gas constant
Rpi Producing gas to oil ratio, scf/STB
Rs Solution gas to oil ratio, scf/STB
S Skin factor
Sor Critical oil saturation,
SVM Support Vector Machine
SS Overall cluster variable
Swi Irreducible water saturation
Sg Gas saturation
So Oil saturation
Stotal Total skin factor
Tsc Temperature at standard condition, Kelvin
Tsep Separator temperature, Rankine R
Tpc Temperature pseudocritical properties
Tr Reduced temperature
vpc Volume pseudocritical properties, ft3/lbm mol
WT Transposed matrix of SVM
Z Z factor, compressibility factor, deviation factor

Greek letters
𝜇𝑔 Gas viscosity, centipoise (cp)
𝜇𝑜 Oil (condensate liquid) viscosity, cp
𝜇𝑜𝑑 Dead oil viscosity, cp
𝜇𝑜𝑏 Live oil viscosity, cp
𝜇𝑐 Condensate liquid viscosity, cp
𝜇𝐴 (𝑥) Ambiguous membership of component x in universe 𝜇

𝛾𝑔 Specific gravity of gas

𝛾𝑜 Specific gravity of oil
𝛾𝑤 Well stream gravity
𝛾̅𝑤̅ Average well stream gravity
𝛾̅𝑜̅ Average surface oil gravity
𝛾𝐴𝑃𝐼 API gravity
𝜌𝑔 Density of gas phase, lbm/ft3
𝜌𝑜 Density of oil (condensate liquid) phase, lbm/ft3

xvii | P a g e
 𝜇 Viscosity of mixture, cp
ζ Viscosity reducing parameter, cp-1
𝜇∗ Law pressure gas mixture viscosity, cp
𝜌𝑟 Reduced density, lbm/ft3
𝛾 Tuning parameter of SVM
𝜑(𝑥) Kernel function in SVM
𝜉𝑘 Slack variable of SVM
Δmpgt Total gas pseudopressure function, psi/cp

xviii | P a g e
 CHAPTER 1 INTRODUCTION
INTRODUCTION
1.1 Background

Increasing energy demand worldwide and quest for cheaper and cleaner source of
energy place natural gas and gas-condensate reservoirs in a unique position in today’s
energy market. The global move to use natural gas is an evolution in energy market
that is changing economic and the environment for the better (Tillerson, 2015). Figure
(1.1) shows the continuous increase of demand for natural gas in United State since
1980. Figure (1.2) highlites the explored hydrocarbon reserves in Russia, which
include substantial amount of gas-condensate fields. This also demonestrate the
strong share market of gas-condensate reservoirs in the world. Gas-condensate
reservoirs are playing an important role in respond to the huge global energy market.
It has been reported that gas-condensate reservoirs accounts for 68% of all giant gas
reservoirs (reservoirs larger than 1012m3 ) worldwide (Zhang et al., 2019). Examples
of some of the giant’s gas-condensate fields around the globe are Arun field in
Indonesia, Shtokmanovskoye field in Russia, North field in Qatar (South Pars field in
Iran) and Cupiagua field in Colombia.

Figure 1.1. USA Natural Gas withdrawals (EIA, 2020).

1|Page
Figure 1.2. Breakdown of Gazprom Group’s explored hydrocarbon reserves in Russian
regions (Gazprom, 2019).

Another example of a giant gas-condensate reserve is Khuff gas-condensate reservoir

known as north field in Qatar and south pars field in Iran. Khuff is the world biggest
gas-condensate reservoir in the world, holds between 1000 – 2000 trillion cubic feet
initial gas in place and 30 to 70 billion barrels of condensate in place. However, these
important source of energy are suffering from productivity decline and loosing its most
valuable hydrocarbon source in the depth of the reservoir. This is due to the reduction
of bottom-hole flowing pressure (Pwf) to below the saturation (dew point) pressure,
which triger condensate (liquid) drop out from the gas phase. This liquid drop out has
a significant impact on gas relative permeability and lead to production decline. For
example well productivity in the Arun field, in North Sumatra, Indonesia, declined
significantly 10 years after the production began and it creates serious problem as the
reservoir could not produce the required gas to meet contractual obligation (Fan et al.,
2005). In Arun field, which its opertor is ExxonMobile, the production loss in some
wells were greater than 50% (Afidick, Kaczorowski and Bette, 1994;
Ayyalasomayajula, Silpngarmlers and Kamath, 2005). Another example is Britannia
gas-condensate field where the productivity declined between 50 – 60% in the first
year of production even before stabilization of the fluid flow occurs (Göktaş, Macmillan
and Thrasher, 2010). Shell and Petroleum development Oman reported 60% loss of
productivity of the wells in one of the fields (Smits, Van der Post and Al Shaidi, 2001).

2|Page
This reduction in deliverability and losing productiuon was due to the accumulaion of
the liquid phase (condensate) near the wellbore region. To understand the well
deliverability reduction in gas-condensate reservoirs the phenomenon will be
explained in the following section.
1.2 Gas-condensate characteristics

Gas-condensate reservoirs are a class of hydrocarbon reservoirs that charcterized by

production of surface gas and varying quantities of stock-tank-oil (STO). The STO
usually known as “condensate” or “distillate”. In this study the “condensate” phrase is
used for stock tank oil. Classification of the hydrocabon reservoirs are important for
modelling and also selecting appropriate engineering practices. As illustrated in Table
1.1 hydrocarbon reservoirs are classified in terms of composition and other fluid
properties to dry gas, wet gas, gas-condensate, near-critical oil, volatile oil and black
oil. The composition and reservoir condition (temperature and pressure) would
determine the category the reservoir fluid. For a given composition, a reservoir fluid
can be classisfied to different category with different pressure and temperature. Gas-
condensate fluids typically exhibit condensate (oil) to gas ratio (OGR’s) ranging from
5 to 350 barrel of liquid per million standard cubic feet (STB/MMscf). The API gravity
of gas-condenste fluid, which is measure of weight or density, is between 40⁰ to 60⁰
API (Whitson and Brulé, 2000). Most well known gas-condensate reservoirs are found
in the range of 5000 to 10000 feet deep, at pressure of 3000 to 8000 psia and
temperature of 200⁰F to 400⁰F (Moses, 1986; Moses and Donohoe, 1987). These wide
range of pressure, temperature and compositions of gas-condensate reservoirs lead
to wider difinitions in literature. However, one of the widely accepted defintion is in
terms of temperature as if the reservoir temperature located between critical
temperaure (𝑇𝑐 = 127°𝐹) and cricondentherm temperature (𝑇𝑐𝑡 = 250°𝐹) the
reservoir classified as gas-condensate reservoirs (Whitson and Brulé, 2000; Thomas,
Bennion and Andersen, 2009; Ahmed, 2010). To further settle wide definition of gas-
condenste reservoirs two main characteristics that distinguish all gas-condensate
reservoirs from other type of hydrocarbon systems are introduced by Raghavan and
Jones, (1996) as follow:
 The condensation of the gas at reservoir conditions during isothermal depletion.
 The retrograde and revaporization of the condensate liquid by further decline in
pressure (Raghavan and Jones, 1996).

3|Page
The above characterisctics of gas-condensate reservoirs can be related to fluid
behaviour below the saturation pressure. The fluid behaviour is divided into two
sections of phase behaviour and drawdown behaviour. The fluid behaviour is major
contributor to productivity decline in gas-condensate reservoirs below the dew point
pressure. The fluid behaviour is explained in the following section and in wider context
it will be related to the productivity decline.

Table 1.1. Compositions and properties of several reservoir fluids (Whitson and Brulé, 2000).

Compositions (mol%)
Componenet Dry Gas Wet Gas Gas- Critical Volotile Black Oil
Condensate Oil Oil
CO2 0.10 1.41 2.37 1.30 0.93 0.02
N2 2.07 0.25 0.31 0.56 0.21 0.34
C1 86.12 92.46 73.19 69.44 58.77 34.62
C2 5.91 3.18 7.80 7.88 7.57 4.11
C3 3.58 1.01 3.55 4.26 4.09 1.01
i-C4 1.72 0.28 0.71 0.89 0.91 0.76
n-C4 ------ 0.24 1.45 2.14 2.09 0.49
i-C5 0.50 0.13 0.64 0.90 0.77 0.43
n-C5 ------ 0.08 0.68 1.13 1.15 0.21
C6 ------ 0.14 1.09 1.46 1.75 1.61
C7+ ------ 0.82 8.21 10.04 21.76 56.40
Properties
MC7+ ------ 130 184 219 228 274
𝛾𝐶7+ ------ 0.763 0.816 0.839 0.858 0.920
𝐾𝑤𝐶7+ ------ 12 11.95 11.98 11.83 11.47
GOR,
∞ 105000 5450 3650 1490 300
scf/STB
OGR,
0 10 180 275 ------ ------
STB/MMscf
API gravity ------ 57 49 45 38 24
𝛾𝑔 ------ 0.61 0.70 0.71 0.70 0.63
𝑃𝑠𝑎𝑡 , psia ------ 3430 5650 7015 5420 5810
𝐵𝑠𝑎𝑡 , ft3/scf ------ 0.0051 0.0039 2.78 1.73 1.16
𝜌𝑠𝑎𝑡 , lbm/ft3 ------ 9.61 26.7 30.7 38.2 51.4

4|Page
1.3 Gas-condensate flow behaviour

Phase behaviour

Gas-condensate reservoir behaviour is a function of two parameters of phase

envelope of the fluid and condition of the reservoir (Roussennac, 2001). A typical
phase envelope (P-T diagram) of gas-condensate fluid shown in Figure 1.3. These
phase envelop diagram consists of bubble point line (where first bubbles of gas
vaporizes from the liquid content) and dew point line (where first droplet of liquid
condenses from the gas phase). The bubblepoint line and dew point line meet at the
mixture critical point. The critical point is representing a state where all intensive
properties of gas and liquid phases are equal (Ahmed, 2010; Craft and Hawkin, 2015).
The cricondentherm and cricondenbar are maximum temperature and pressure
respectivly that above them the mixture is only in the form of gas or liquid (only one
phase).
As the reservoir produces, the formation temperature normally doesnot change
(isothermal behaviour) but the average reservoir pressure and flowing bottom hole
pressure varies. In gas-condensate reservoirs the fluid is initialy in single phase (point
B on the graph Figure. 1.3), which consists of predominantly methane “C1” and other
short chain hydrocarbons, called heavy ends. Isothermal pressure depletion to below
the dew point line cause heavy end hydrocarbons drop out of the solution and form
the liquid inside the reservoir (Between point B1 and B2). The liquid phase known as
condensate liquid has zero mobility ratio to the associated gas between B1 and B2. At
this point only gas flows where the heavy end hydrocarbons left behind in the reservoir.
This phenomenon causes a compositional changes of the reservoir mixture.
Further reservoir pressure decline lead to further accumulation of the liquid to the
maximum level at point B2 (known as critical oil saturation). At this point condensate
liquid have enough energy to overcome the gravity segregation in porous media and
move towards the wellbore simultaneously with gas phase. The dashed lines on the
phase diagram represents the percentage of the vaopur phase (gas phase) in the
mixture. Some gas-condensate phase diagrams shown the amount of liquid
percentage inside the two phase region. However in Figure 1.3 the amount of gas
phase shown in two-phase region.
Additional reduction of reservoir pressure, move down point B2 towards point B3.
Between point B2 and point B3 the accumulated condensate liquid vaporizes and turn

5|Page
back to gases state. At point B3 most of the condensate liquid revaporizes and the
reservoir fluid in only in the form of gas (100% vapour). This thermodinamically
anomalous phenomenon first noticed by Kuenen, (1892) and he called it “retrograde
condensation”.

Figure 1.3. Typical phase diagram of gas condensate systems (Modified from Fan et al.,
2005).

Condensate drop out significantly alter the permeability of the reservoir formation to
the gas flow and fundementally change the drawdown behaviour of the reservoir near
the wellbore region. The amount of liquid phase not only depends on compostions of
the mixture shown in Table 1.1, also to the other reservoir behaviour and production
strategy (depletion model of recovery). This amount of generated condensate normally
determines if the reservoir is lean or rich gas-condensate. If the reservoir generates
small amount of liquid normally less than 100 barels per million square cubic feet the
gas-condensate known as lean whereas if the amount of liquid between 150 to 300
barels per million square cubic feet its known as rich gas-condensate reservoir.
Drawdown behaviour

If depletion drive system is selected as mode of recovery for any hydrocarbon

reservoirs, the energy for producing hydrocarbon on the surface comes from the
difference between formation pressure gradient (reservoir pressure) and bottomhole
flowing pressur (Pwf) of the wells. The Pwf is the pressure at the formation of the
wellbore of the producing well (Ahmed, 2010, p. 354). In gas-condensate reservoirs
with depletion mode of recovery when reservoir pressure declined due to the

6|Page
production, Pwf of the wells need to be changed to compensate reservoir pressure
decline, and keep the pressure gradient to meet predetermined rate on the surface.
The relation between a well constrained Pwf and the rate of the production known as
well deliverability (Fevang, 1995).
When the Pwf keeps decreasing to reach the point known as dew point pressure (the
point where first drop of liquid evolve from gas), physics of the flow inside the reservoir
is changing and three flow regions are established.These three main flow regions
proposed by Fevang, (1995) extending from wellbore outward and illustrated in Figure
1.4. These regions are gradually changing during lifetime of a gas-condensate
reservoir. The Pwf controls the production of hydrocarbon fluids on the surface. If Pwf
is above the dew point pressure, the fluid in the reservoir is single phase and region 3
exist. If Pwf goes down to below the dew point pressure, region 2 starts to grow and
two-phases of gas and condensate (oil) are exist as shown in Figure 1.4. In region 2
only gas flows toward the wellbore and condensate (oil) phase is imobile. The
saturation of condensate phase is increasing with time and reaches critical condensate
saturation (Sor). When the maximum Sor is reached the transition is starting from region
2 to region 1, where both phases are flowing toward the wellbore.
Region 1 is the main source of deliverabiity loss in gas-condensate wells because of
higher pressure drop in this region caused by condensate accumulation. Condensate
accumulation in region 1, would decrease gas phase permeability to flow sharply.
The amount of condensate saturation in region 1 is a function of fluid properties that
entering this region and the production rate (Fevang, 1995; Roussennac, 2001).
These properties consist of viscosity of the original mixture, formation volume factor
and solution oil to gas ratio. Condensate drop out further apart the behaviour of the
gas-condensate reservoir mixture from ideal gas law. This deviation is deterimined by
compressibility factor (Z factor). Among the fluid properies condensate (oil) viscosity
in each depletion stage has the highest uncertainity for the purpose of the modelling
in such reservoirs. Accurate Z factor prediction also plays a key role for reliable
modelling of gas-condensate well deliverability (Fevang, 1995; Fevang and Whitson,
1996; Mott, 2002; Whitson and Kuntadi, 2005)
The main objective of this research is to develop several models for acurate prediction
of the gas/condensate (oil) viscosity as well as Z factor. The developed models will be
used for reliable prediction of gas-condensate well production forecast.

7|Page
 Figure 1.4. Gas/condensate flow in three regions.

1.4 Problem Statement

A reliable model that accurately simulate phase and drawdown behaviour and their
effect on gas-condensate well performance is highly desirable for financial projection
and field production planning. The focus of this research is in developing well
deliverability (productivity) model and condensate blockage effect with an emphasis
on accurate prediction of phase behaviour. Although extensive research have been
conducted in general well deliverability modelling of gas-condensate reservoirs, there
is still outstanding issues and shortfall in this area. We highlights some of these
difficulties in following.
Developing accurate and reliable gas-condensate system well deliverability model,
governed by accurate fluid phase behaviour, which require well define pressure –
volume – temperature (PVT) model. A PVT model describes key phase, viscosity and
volumetric behaviour of hydrocarbon mixture that are dictating the recovery of gas and
oil at the surface (Whitson et al., 1999). For gas-condensate mixture, a PVT model for
all reservoir pressure and temperature conditions cannot accurately define certain
properties (e.g., viscosity, compressibility factor, hydrocarbon plus characterization).
For example a PVT model such as equation of state (EOS) often have difficulty of
matching compositional variation of gas and condensate dropout in near critical gas-
condensate systems. Among the PVT properties condensate viscosities are difficult to
predict and has the largest prediction uncertainty using any PVT models (Fevang,

8|Page
1995; Whitson, Fevang and Yang, 1999; Mott, 2002). This is mainly due to inadequacy
of the available methods in prediction of condensate viscosity and unavailability of
experimental condensate viscosity data to tune the existing models (Fevang, 1995;
Whitson, Fevang and Yang, 1999; Mott, 2002; Al-Meshari et al., 2007; Yang et al.,
2007).
Unavailability of experimental data is due to the absence of reservoir sample to carry
out laboratory experiments for gas-condensate systems. This is because obtaining
representative sample of gas-condensate reservoirs are extremely difficult than other
type of hydrocarbon systems (e.g., black oil). The difficulty of sampling is due to
condensation of liquid during sampling process, which leads to incorrect estimation of
gas/condensate percentage and erroneous estimation of reservoir composition
(Ahmed, 2010, p. 171).
Compressibility factor (Z factor) is another PVT property that its accurate estimation is
always needed to develop an accurate PVT model, and consequently for a reliable
gas-condensate well deliverability model (Whitson, Fevang and Yang, 1999). Accurate
estimation of Z factor cannot be overemphasized in gas reserve evaluation, material
balance calculation, reservoir simulation studies, analytical models, well testing and
gas/condensate processing calculations (Rayes et al., 1992; Heidaryan, Moghadasi
and Rahimi, 2010; Sun et al., 2012).
Ideally, condensate viscosity and Z factor should be estimated using laboratory
experiments, however lack of enough experimental data for the reasons mentioned
earlier, motivated to use correlations and EOS extensively in current literature. Using
correlations and EOS for estimation of these PVT properties has several problems
that will be explained in the following.
There are many EOSs in literature for determination of PVT properties (e.g., Peng and
Robinson, 1976; Redlich and Kwong, 1949; Soave- Redlich-Kwong, 1972; Van der
Waals, 1873). However, all EOSs are implicit in nature; this means for instance Z factor
is estimated as a root of EOSs. This makes using EOSs for estimation of gas-
condensate Z factor computationally inconvenient for engineering purposes.
Furthermore their performance deteriorate for estimation of gas-condensate mixture
because of convergence problem they face for near critical point when the phase
change occurs (Sarkar, Danesh and Todd, 1991; Elsharkawy, 2006; Shokir, 2008; Sun
et al., 2012).

9|Page
For estimation of condensate viscosity, the industry heavily rely on correlations, which
their prediction accuracy is very limited especially for near critical gas-condensate
fluids and all thermodynamic conditions. A few issues of using correlations for
prediction of condensate viscosity are as follow.
 Existing condensate viscosity correlations in current literature were developed
with simple assumption of single-phase dry gas and their applicability for
estimation of gas-condensate viscosity below the dew point pressure, where
the two-phase flow exist is arguable.
 Most of available correlations in literature have limited applicability to specific
range (e.g., pressure, temperature and viscosity).
 For high pressure high temperature (HPHT) conditions, which majority of gas-
condensate reservoirs are located, using correlations for estimation of viscosity
is uncertain due to lack of measured data (Al-Meshari et al., 2007; Sun et al.,
2012).
Based on the above shortfalls in literature in relation to sufficient models that can well
define phase behaviour of gas-condensate fluid below the saturation pressure, the aim
and objectives of this study are defined and presented in following section.

1.5 Research aim and objectives

The ultimate aim of this study is to develop an accurate gas-condensate well

deliverability (productivity) model with an emphasis on accurate determination of
phase behaviour.
Based on the gaps in current literature highlighted in section 1.4 for accurate modelling
of PVT properties the following objectives are defined.
I. Investigate the accuracy and applicability of current existing literature models
for estimation of gas/condensate viscosity and Z factor in lean/rich gas-
condensate reservoirs.
II. Development of reliable and accurate gas/condensate viscosity, that can cope
with non-linearity of gas-condensate mixture below the saturation pressure.
III. Developing several models to accurately estimate Z factor of gas-condensate
reservoirs below the saturation pressure.

10 | P a g e
IV. Investigate the effect of dynamic of condensate build up near well bore region
on well inflow performance and production profile through utilizing the
developed viscosity and Z factor models.
1.6 Proposed approach

The following methodologies and approaches were adopted to achieve aim and
objectives of the study.
I. An extensive gas-condensate laboratory experimental and field data were
collected to investigate the accuracy of existing gas/condensate PVT models
for prediction of gas/condensate viscosity and two-phase Z factor.
II. Non-linear regression has performed to optimize several literature models in
order to characterize phase behaviour of gas-condensate mixture through
accurate estimation of viscosity and two-phase Z factor.
III. Sophisticated algorithms of Machine Learning (ML) including Support Vector
Machine (SVM), Artificial Neural Network (ANN), Fuzzy Logic (TSK and
Mamdani) and Adaptive Neuro Fuzzy Inference System (ANFIS) were
employed to develop several gas/condensate viscosity and two-phase Z factor
models.
IV. The optimized and the developed gas/condensate viscosity and two-phase Z
factor models in this study employed for better characterization of gas-
condensate mixture phase behaviour below the saturation pressure and in
establishing an accurate PVT model.
V. A well-defined PVT model in this study is used for generating reliable production
profile of gas-condensate reservoirs experiencing liquid drop out, through
utilizing three regions pseudopressure approach for well inflow calculation and
volumetric material balance for well production forecast.

1.7 Organization of the thesis

The thesis comprises of seven chapters as follow:

Chapter 1 provides the background information and highlights the challenges of gas-
condensate performance modelling. It also covers aims, objectives and proposed
approaches in this study.

11 | P a g e
Chapter 2 covers the literature and relevant theory of fluid flow in porous media in
relation to gas-condensate well deliverability modelling. Various factors that are
influencing the modelling were also discussed.

Chapter 3 outlines methodologies and current modelling approaches of gas-

condensate wells undergoing depletion. The approaches and methodologies
undertaken in this study have been discussed in this chapter.

Chapter 4 investigates the applicability of existing literature models (correlations and

equation of state based models) for prediction of gas-condensate viscosity below the
saturation pressure. Several existing condensate viscosity models in literature have
been optimized using experimental data and presented in this chapter. Machine
learning techniques including regression, Artificial Neural Network (ANN), Least
Square Support Vector Machine (LSSVM) and Fuzzy Logic [Takagi-Sugeno-Kang
(TSK) and Mamdani] were extensively used in developing new models. The new
models can be used for reliable prediction of gas/condensate viscosity of gas-
condensate reservoirs below the saturation pressure as an alternative approach to
EOSs.

Chapter 5 assesses the accuracy of the existing literature models for prediction of
gas-condensate two-phase Z factor below the saturation pressure using
comprehensive data bank. Then several new two-phase Z factor models based on
Cascade Forward Neural Network (CFNN), Feed Forward Neural Network (FFNN) and
Adaptive Neuro Fuzzy Inference System (ANFIS) were developed. Development of
new models including structure of each network are described in details.

Chapter 6 presents well deliverability forecast of gas-condensate reservoirs in tight

formation using back pressure equation and three regions pseudopressure integral for
computation of well inflow performance. Pressure Transient Analysis (PTA) test of a
gas-condensate well is used for computation of effective permeability of
gas/condensate phases. Three regions pseudopressure integral was incorporated
with the developed gas/condensate viscosity as well as the developed two-phase Z
factor model. The detail description of volumetric material balance used for reservoir
depletion is provided. A new production forecast for a gas-condensate field case study

12 | P a g e
is generated utilizing three regions pseudopressure integral incorporated with material
balance method. Furthermore, the detail of compositional reservoir simulation using
Eclipse 300 is given in this chapter.

Chapter 7 presents the conclusion and recommendations of the research and

provides future research avenue.

The flow chart in Figure 1.5 illustrates the organization of the thesis and the tasks that
have been carried out throughout each chapters.

13 | P a g e
 Chapter 2
Background Aim and Density, Gas/oil Formation volume
Information Objective Literature review: Performance
Problem factor, Solution gas to oil ratio (Rs),
s modelling; PVT properties
statement Two-phase Z factor,
Chapter 1 Gas/condensate viscosities

Gap in literature Gas/condensate viscosity; Two-phase Z factor

Gas/condensate Utilizing/Optimization of existing

Viscosity models for prediction of viscosity. Chapter 3
Data collection Utilize existing models (assess the
(Viscosity, Z factor) performance)
Validate the models using field
Developing ANN, SVM, data.
TSK & Mamdani Fuzzy Compare the results.

14
New approach
for prediction of viscosity (AI or ML) ANN, SVM, Fuzzy Logic, ANFIS
Chapter 4

Chapter 5
Two-phase Z factor Assess existing models of DAK, Gas-condensate Generate PVT tables (MBO) using
HY, Beggs & Brill & etc., for performance model developed viscosity and two-phase Z
prediction of two-phase Z factor factor models
Incorporate MOBPVT
Developing FFNN, with pseudopressure
CFNN, ANFIS models Validate the models using field Integrate IPR with material balance to
and generate IPR generate production profile
for predicting Z factor data. Compare the results.
Chapter 6

Conclusion & recommendations, Chapter 7

future studies

Figure 1.5. The flow chart indicates organization of the thesis for each chapter.
 CHAPTER 2 WELL DELIVERABILITY
MODELLING OF GAS-CONDENSATE RESERVOIRS
WELL DELIVERABILITY MODELLING OF GAS
CONDENSATE RESERVOIRS

2.1 Introduction

This chapter provides a critical review towards the modelling of gas-condensate

reservoirs below the dew point pressure. Various modelling approaches have been
critically reviewed and evaluated. The influence of condensate drop out when the
reservoir pressure depleted to below the dew point pressure on modelling approaches
are also discussed.
Modelling and calculation of well deliverability in gas-condensate reservoirs is a
historical issue without simple solution (Fevang and Whitson 1996). Reliable
calculation of well deliverability requires great understanding of phase and drawdown
behaviour in reservoir condition. Gravity segregation of the heavy hydrocarbon
components in gas-condensate reservoirs, trigger liquidation of these heavy ends
below the saturation pressure. Increasing the saturation of the hydrocarbon liquid in
the reservoir will create condensate blockage near the wellbore region. Consequently,
the productivity of the wells in gas-condensate field will dramatically reduce due to the
condensate blockage (Thornton, 1946; Kniazeff and Naville, 1965; Daltaban, 1985;
Vo, Jones and Raghavan, 1989; Raghavan and Jones, 1996). In gas-condensate
reservoirs, the initial phase is gas but typically, the fluid of interest is condensate (oil)
because of higher profitability that produced condensate can bring to the project. The
condensate loss is one of the greatest economical concerns because of its valuable
heavier components of the original reservoir fluid, which trapped in the depth of the
reservoir (Roussennac, 2001).
This chapter provides fundamental of fluid flow inside porous media in relation to the
gas-condensate flow behaviour and well deliverability modelling. First gas flow theory
and fundamentals are discussed then the parameters that required for well
deliverability modelling and phase behaviour are highlighted. Critical review of each
parameter is provided in respective chapter.

15 | P a g e
2.2 Gas Rate Theory

The fluid flow equations for describing fluid behaviour in hydrocarbon reservoir are in
many forms depending on combination of variables such as type of flow, type of fluid
and number of mobile phases in the reservoirs (Ahmed, 2010). By combining the fluid
transport equation (Darcy’s law) with continuity equation for material balance, the
appropriate fluid flow equation can be developed for hydrocarbon reservoirs.
Fundamental principle of fluid flow in porous media defined by Darcy’s law proposed
by Henry Darcy (1856). Darcy equation in the form of 2.1 states that the apparent
velocity (𝑣) of a homogeneous fluid in a porous media with absolute permeability of k
is directly proportional to the pressure difference (dp) and inversely proportional to fluid
viscosity (μ) in a radial flow (dr). The negative sign represents the inverse direction of
pressure gradient with respect to direction of flow.

𝑞 𝑘 𝜕𝑝
𝑣= =− 2.1
𝐴 𝜇 𝜕𝑟

Integrating Darcy’s law yields gas flow rate (qg) equation in a pseudosteady state for
any well geometry (e.g. vertically fractured, radial and horizontal wells) as shown in
2.2. Gas flow rate in equation 2.2 is written in terms of gas formation volume factor
(Bg) or equivalent gas compressibility factor (Z).
𝑃𝑅
1
𝑞𝑔 = 𝐶 × ∫ 𝑑𝑝
𝑃𝑤𝑓 𝜇𝑔 𝐵𝑔
2.2
𝑜𝑟
𝑃𝑅
∗
𝑃
𝑞𝑔 = 𝐶 × ∫ 𝑑𝑝
{ 𝑃𝑤𝑓 𝜇𝑔 𝑍 }
C includes basic reservoir properties and for vertical/horizontal unfractured wells, is
defined as follow:
2𝜋𝑎𝑘ℎ
𝐶= 𝑟𝑒
ln (𝑟 ) − 0.75 + 𝑠 + 𝐷𝑞𝑔 2.3
𝑤
𝑇𝑠𝑐
𝐶∗ = 𝐶
{ 𝑃𝑠𝑐 𝑇𝑅 }
Where 𝑎 represents conversion factor of 1/(2𝜋. 141.2) for field units, 𝑎𝑛𝑑 𝑎 = 1 for SI
units, s is skin factor, k is absolute permeability of the formation, h is reservoir
thickness, re is drainage radius, rw is wellbore radius, 𝑇𝑠𝑐 and 𝑃𝑠𝑐 are temperature and
16 | P a g e
pressure at standard condition (sc) and TR is temperature at reservoir condition. The
unique relationship between pressures and gas properties in pressure integral of
equation 2.2 proposed by Al-Hussainy et al., (1966) and commonly known as
pseudopressure integral.
Gas rate equation in 2.2 is valid within laminar flow region where Reynolds number is
less than 2000. However, in a radial flow system the flow velocity increases
significantly near the wellbore region and flow becomes turbulent. Therefore, gas rate
equation shown in 2.2 no longer follows a linear relationship between velocity and
pressure drop of Darcy’s law in higher gas velocity. Subsequently the gas rate
changes to Forchheimer, (1901) quadratic form as follow.

𝑃𝑅
2
𝑃
𝐴𝑞𝑔 + 𝐵𝑞𝑔 = ∫ 𝑑𝑝 2.4
𝑃𝑤𝑓 𝜇𝑔 𝑍

Where A and B are,

𝑟
ln (𝑟𝑒 ) − 0.75 + 𝑠
𝑤 2.5
𝐴=
2𝜋𝑎𝑘ℎ

𝐷
𝐵= 2.6
2𝜋𝑎𝑘ℎ

D is non-Darcy or turbulent flow factor and can be calculated by the following equation.

𝛽𝑇𝛾𝑔
3.161 × 10−12 [ ] 𝑘ℎ
𝜇𝑔 ℎ2 𝑟𝑤 2.7
𝐷=
1422𝑇

Where the turbulent parameter 𝛽 = 1.88(10−10 )𝐾 −1.47 𝛷−0.53, T is temperature in

Rankine, 𝛾𝑔 is gas specific gravity and 𝛷 is the porosity of the reservoir formation. The
turbulent flow is contributing to additional pressure drop around the wellbore region.
This is particularly important in gas-condensate reservoirs because accumulation of
the liquid around the wellbore in region 1 introduces additional skin damage, which in
return add extra pressure drop into the system.
Plotting the function inside the integral “𝑃/( 𝜇𝑔 𝑍) or 1/(𝜇𝑔 𝐵𝑔)” against the reservoir
pressure yields a graph shown in Figure. 2.1 with three distinct regions. Substituting

17 | P a g e
2.5 and 2.6 in Forchheimer gas rate equation 2.4 resulted a pseudo-psteady state
equation for estimating gas flow in porous media in three forms representing three
areas of the graph in 2.1. The low-pressure region known as region 1, where pressure
is usually less than 2000psia and pressure function 𝑃/( 𝜇𝑔 𝑍) or 1/(𝜇𝑔 𝐵𝑔) exhibit a
linear relationship with pressure. The following equation can be used for exact
calculation of gas flow rate when 𝑃 < 2000 psia.

𝑘ℎ(𝑃𝑅 2 − 𝑃𝑤𝑓 2 )
𝑞𝑔 = 𝑟 2.8
1422𝑇(𝜇𝑔 𝑍) [ln ( 𝑒 ) − 0.75 + 𝑠 + 𝐷𝑞𝑔 ]
𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑟𝑤

The above method known as pressure squared approximation method. The product
of (𝜇𝑔 𝑍) in 2.8 is assumed to be constant for pressures below 2000psia.
In second region the relationship between 𝑃/( 𝜇𝑔 𝑍) or 1/(𝜇𝑔 𝐵𝑔) and pressure is a
curvature shape. In this region, both bottom-hole flowing pressure (Pwf) and average
reservoir pressure (PR) are between 2000 – 3000psia. To estimate gas flow rate when
pressure is between aforementioned values, the following equation known as pressure
approximation method is used.

7.08 × (10−6 )𝑘ℎ(𝑃𝑅 − 𝑃𝑤𝑓 )

𝑞𝑔 = 𝑟 2.9
(𝜇𝑔 𝛽𝑔 )𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑇 [ln (𝑟 𝑒 ) − 0.75 + 𝑠 + 𝐷𝑞𝑔 ]
𝑤

The third region is devoted to high pressure region with pressures (Pwf and PR)
between 3000 – 5000psia. In this pressure region the relationship between pressure
and 𝑃/( 𝜇𝑔 𝑍) or 1/(𝜇𝑔 𝐵𝑔) become almost constant as this can be observed from the
graph. The gas flow rate in this region can be estimated by real gas potential or
pseudopressure form in pressure region between 3000 < 𝑃 < 5000𝑝𝑠𝑖𝑎.

𝑘ℎ(𝑚𝑝𝑟 − 𝑚𝑝𝑤𝑓 )
𝑞𝑔 = 𝑟 2.10
1422𝑇 [ln (𝑟 𝑒 ) − 0.75 + 𝑠 + 𝐷𝑞𝑔 ]
𝑤

Where mp is pseudopressure integral shown in Forchheimer gas rate equation of 2.4.

18 | P a g e
Figure 2.1. Typical plot of gas pressure function vs pressure where the area between Pwf
and PR representing the integral in 2.2.

The derived gas flow rate equations in either forms of 2.8 to 2.10 are implicit, where
direct determination of qg is not possible. Furthermore for determination of gas flow
rate in high pressure region because the product of 𝑃/( 𝜇𝑔 𝑍) is almost constant, the
pressure integral has simple analytical solution, which is not suitable for rate-time
forecasting of a gas well above 3000psia.
Rawlins and Schellhardt, (1936) propsed a simple back presure equation which relate
gas flow rate to bottom-hole flowing pressure (Pwf) regardless of value of Pwf. The
Rawlins and Schellhardt, (1936) is known as well deliverability equation and is
discussed in following section.
2.3 Well Deliveraility Theory

Rawlins and Schellhardt, (1936) developed an empirical equation known as

backpressure equation, that defines relationship between the gas flow rate and some
constraint average reservoir pressure (PR) and bottom hole flowing pressure (Pwf).
Their empirical equation shown in 2.11 is widely accepted in petroleum industry for
estimating gas flow rate (Eilerts, Sumner and Potts, 1965; Gondouin, Iffly and Husson,
1967; Fevang, 1995; Akhimiona and Wiggins, 2005; Al-Attar and Al-Zuhair, 2009;
Ogunrewo, 2014). The equation also referred as backpressure equation or well
deliverability equation.

19 | P a g e
 𝑛
𝑞𝑔 = 𝐶(𝑃𝑅 2 − 𝑃𝑤𝑓 2 ) 2.11

Where qg, PR, Pwf and C are defined in 2.4 to 2.6 and n is deliverability exponent.
Well deliverability equation in 2.12 was developed, as a result of testing several
hundred gas rates of gas wells in different fields. The square terms in backpressure
equation accounts for pressure dependent of the pseudopressure integral. The
exponent n represents degree of flow velocity in non-Darcy flow and is depending on
flowing conditions. This exponent is between 0.5 – 1, where 1 representing completely
laminar flow and 0.5 for fully turbulent flow. Pressure squared terms (𝑃𝑅 2 − 𝑃𝑤𝑓 2) can
be replaced by ΔP2.
The well deliverability equation is valid for calculating gas flow rate for reservoir
pressure of less than 2500 psia. If the average reservoir pressure (PR) is greater than
2500 psia, then ΔP2 should be replaced by ΔP. If the coefficients of n and C are known,
the gas flow rate at the surface for any bottom-hole flowing pressure (Pwf) can be
estimated from well deliverability equation. Plotting qg vs. Pwf results in constructing
Inflow Performance Relationship (IPR) curves. IPR curve is demonstrating ability of
the reservoir to produce gas/condensate to the wellbore (Sousa, Garcia and Waltrich,
2017).
In gas-condensate reservoirs undergoing depletion, accumulation of condensate drop
out near wellbore region creates condensate blockage, which introduce extra pressure
drop. This would results in well productivity reduction at the surface. Condensate
blockage and its effect on well deliverability should be considered for reliable and
accurate prediction of gas-condensate well performance. Introducing condensate
blockage into well deliverability equation in 2.11 has been an active area of research
for many years. Some of the attempts by many scholars will be reviewed here.

Muskat, (1949) was first that addressed condensate blockage in gas-condensate

reservoirs operating below the dew point pressure. He introduced a simple method for
estimating the radius of the condensate blockage as a function of time, gas rate,
reservoir rock and fluid properties. Fetkovich, (1973) used Muskat’s results and
highlighted condensate blockage problem by introducing a skin factor as a function of
gas rate and time for use in standard dry gas equation of 2.4. Eilerts, (1964) was

20 | P a g e
among the first that numerically solved partial differential equation for transient flow in
gas-condensate reservoir. Then Kniazeff and Naville, (1965) continued numerical
solution of Eilerts, (1964) for a radial gas-condensate’s well deliverability. Both studies
confirmed reduction in well deliverability as a results of condensation blockage.
Gondouin et al., (1967) extended the work of Kniazeff and Naville, (1965) by
performing black oil simulations, showing the effect of condensate blockage and non-
Darcy flow on well deliverability. They used deliverability equation to measure gas flow
rate. The work of Gondouin et al., (1967) is a significant contribution in understanding
of gas-condensate reservoirs and well deliverability improvement. They showed
change of formation permeability in near wellbore region and also reservoir fluid
characteristic due to the condensate blockage.

In all above studies, the concept of single-phase pseudopressure approach shown in

equation 2.8 was utilized. However, gas-condensate reservoirs experiencing two-
phase flow near wellbore region and gas rate equation representing two phases are
more adequate. Therefore, the concept of two-phase pseudopressure approach later
introduced for gas-condensate well deliverability modelling, which will be explained in
following.

Pseudopressure (𝑚𝑝 ) is an integral over pressure presented in 2.8, which expresses

gas rate through reservoir to the wellbore. O’Dell and Miller, (1967) presented the first
gas rate equation using pseudopressure function (𝑚𝑝 ) to describe the effect of
condensate blockage. Their results were indicated a significant reduction in well
deliverability for even small condensate blockage. Their equation is valid when the
radius of the blockage around the wellbore is small, which means the Pwf is
considerably above the dew point pressure. The pseudopressure integral used by
O’Dell and Miller, (1967) is based on two regions consisting gas and condensate
phase. They were among the first that discovered the existing of the various regions
during fluid flow in gas condensate reservoirs. Fussell, (1973) proposed an Equation
of State (EOS), for prediction of gas-condensate well productivity using compositional
simulator. The simulator tracks the compositional changes below the dew point
pressure. Fussell, (1973) showed that O’Dell and Miller, (1967) theory cannot predict
the saturation profile in two-phase region correctly. He showed that the productivity of
gas-condensate well is much higher than the results showed by O’Dell and Miller,

21 | P a g e
(1967). Then Jones and Raghavan, (1988) showed that pressure responses from gas-
condensate drawdown test can be correlated with classical liquid solution, if pressures
were transformed into two-phase pseudopressure. They also showed that steady-
state two-phase pseudopressure integral can be used for estimation of reservoir flow
capacity (kh) in equation 2.7. Similar results were developed for pressure build up test
in gas-condensate wells by Jones et al., (1989). Using steady-state pseudopressure
Raghavan et al., (1995) studied several gas-condensate field and concluded that their
method is working best when the reservoir pressure is much higher than the saturation
pressure. Aforementioned scholars utilized the concept of steady state (SS)
pseudopressure method in interpreting and developing the results. The SS model
assume fluid flow in the reservoir with two regions without the transition zone. These
two regions are near wellbore region, where oil and gas are present and both are
mobile toward the wellbore, and an outer region containing only gas phase where the
oil saturation is zero. Fevang, (1995) added another region known as transition zone
where both gas and oil present, but oil (condensate) is not moving towards the
wellbore. In his well deliverability study Fevang, (1995) divided the pseudopressure
integral into three parts representing three flow regions. Xu and Lee, (1999) showed
that using three regions pseudopressure integral of Fevang, (1995) is more accurate
than previous steady-state concept for estimation of flow capacity (kh). Fevang, (1995)
pseudopressure integral required prior knowledge of relative permeability curves as a
function of saturation, correct fluid properties and accurate knowledge of producing
gas to oil ratio (Rp). One of the difficulties of Fevang’s pseudopressure integral is its
dependency on using simulator for accurate estimation of producing gas to oil ratio
(Rp).
To tackle this issue Mott, (2002) proposed simpler technique for calculating Rp based
on growing of region 1 as a function of time. His method does not need simulator
software for estimation of Rp and can be implemented in spreadsheet format. Similar
concept was applied by Raghavan and Jones, (1996), where the size of two-phase
region for gas-condensate reservoir calculated for deliverability of the well. Chowdhury
et al., (2004) proposed a semi-analytical model similar to Mott, (2002) with the effect
of capillary number and non-Darcy flow that was missed in previous analytical
approaches. Their method provides accurate results in prediction of lean/rich gas-
condensate well deliverability, verified by compositional numerical simulation.

22 | P a g e
Jokhio and Tiab, (2002) also used Fevang, (1995) method for gas-condensate well
pressure transient test and predict well deliverability. They used effective permeability
concept in using the pseudopressure integral, which eliminates prior knowledge of
relative permeability curves.
The analytical method (e.g., three regions pseudopressure method) are the efficient
and quick way to analyse the impact of the various factor such as viscosity,
compressibility or rock properties on gas-condensate well deliverability modelling
(Dake, 2001; Fan et al., 2005). Many scholars have used three regions
pseudopressure approach as a predictive tool and modelled well deliverability of gas-
condensate reservoirs (Raghavan and Jones, 1996; Dehane, Tiab and Osisanya,
2000; Guehria, 2000; Jokhio, Tiab and Escobar, 2002; Maravi, 2003; Penula, 2003;
Xiao and Al-Muraikhi, 2004; Wheaton and Zhang, 2007; Vo, 2010; Bonyadi,
Rahimpour and Esmaeilzadeh, 2012; Al-Shawaf, Aramco and Kelkar, 2014; Arabloo,
Heidari Sureshjani and Gerami, 2014; Behmanesh, Hamdi and Clarkson, 2015, 2017;
Rahimzadeh et al., 2016; Khanal, Khoshghadam and Lee, 2016; Hekmatzadeh and
Gerami, 2018). Hence, also in this study analogy of three-flow regions pseudopressure
approach is employed for well deliverability modelling of gas-condensate reservoirs.
Fevang, (1995) pseudopressure integral for estimation of total gas flow rate in terms
of black oil PVT properties is as follow.

𝑃𝑅𝑘𝑟𝑔 𝑘𝑟𝑜
𝑞𝑔,𝑡𝑜𝑡𝑎𝑙 = 𝐶 ∫ + 𝑅 𝑑𝑝 2.12
𝑃𝑤𝑓 𝐵𝑔 𝜇𝑔 𝐵𝑂 𝜇𝑂 𝑆

Where C is defined in equation 2.3; Krg and Kro are representing gas and oil
(condensate) relative permeabilities respectively; Bg and Bo are gas and oil formation
volume factor (both are function of compressibility factor); μg and μo are gas and
condensate (oil) viscosity; Rs is solution gas to oil ratio.
If the bottom-hole flowing pressure (Pwf) falls below the saturation pressure then the
reservoir contains three flow regions as suggested by Fevang, (1995), shown in Figure
1.4. To represent these flow regions the pseudopressure integral in equation 2.12
splits into three parts, representing three flow regions as previously explained in 1.1.
Existence of the three flow regions are solely function of pressure and will be
discussed in details in the following.

23 | P a g e
Region 1: is near wellbore region and the main source of well deliverability reduction
due to condensate blockage. In this region both gas and oil (condensate) flow
simultaneously toward the wellbore at different velocity rate. If bottom-hole flowing
pressure (Pwf) is less than the dew point pressure (𝑃𝑤𝑓 < 𝑃𝑑𝑒𝑤), region 1 will always
exist and can be represented by the following integral.

𝑃∗
𝑘𝑟𝑔 𝑘𝑟𝑜
𝑞𝑔,𝑟𝑒𝑔𝑖𝑜𝑛 1 = 𝐶 ∫ + 𝑅 𝑑𝑝 2.13
𝑃𝑤𝑓 𝐵𝑔 𝜇𝑔 𝐵𝑂 𝜇𝑂 𝑆

Where 𝑃 ∗ is the pressure in the interface between region 1 and region 2.

Condensate build up Region 2: in this region condensate is dropping out of the gas
but the mobility ratio is zero or very low, which is not enough for the condensate phase
to flow toward the wellbore. The first droplet of the liquid dropped out from the original
gas at the outer edge of region 2 (at the boundary with region 3). Hence, the pressure
at the boundary of the region 2 with region 3 is equal to the dew point pressure. Since
in this region only gas flows the pressure integral with the pressure limits between dew
point pressure (Pdew) and pressure at the interface between region 1 and region 2 (𝑃∗ )
is as follow.

𝑃𝑑𝑒𝑤 𝑘𝑟𝑔
𝑞𝑔,𝑟𝑒𝑔𝑖𝑜𝑛 2 = 𝐶 ∫ 𝑑𝑝 2.14
𝑃∗ 𝐵𝑔 𝜇𝑔

Roussennac, (2001) in his experimental study observed that region 2 is initially

expands from the well outwards as soon as the bottom-hole flowing pressure drops to
below the dew point pressure. Then region 2 moved away from the wellbore and region
1 developed next to the wellbore.
Single – phase gas Region 3: this region exist when bottom-hole flowing pressure
(Pwf) is above the dew point pressure (Pdew). If 𝑃𝑤𝑓 > 𝑃𝑑𝑒𝑤 then the whole reservoir
is in single dry gas phase and equation 2.12 turns back to standard gas rate equation,
where 𝐾𝑟𝑜 = 0 and Krg is function of irreducible water saturation (Swi) as follow.

𝑃𝑅 𝑘𝑟𝑔 (𝑆𝑤𝑖 )
𝑞𝑔,𝑟𝑒𝑔𝑖𝑜𝑛 1 = 𝐶 ∫ 𝑑𝑝 2.15
𝑃𝑑𝑒𝑤 𝐵𝑔 𝜇𝑔

24 | P a g e
It should be noted if the pressure interface between region 1 and 2 (𝑃∗ ) is bigger than
average reservoir pressure (PR) [𝑃∗ > 𝑃𝑅 ], then integration of region 1 pressure
function should be only from 𝑃𝑤𝑓 to 𝑃𝑅 . In this case region 2 and 3 don’t exist (Fevang,
1995; Fevang and Whitson, 1996). This is the case in highly saturated gas-condensate
reservoirs (Fevang, 1995; Fevang and Whitson, 1996; Jokhio and Tiab, 2002).
Adding up all three-flow region pressure integrals yield an equation for estimating total
gas flow rate in gas-condensate reservoirs as follow.

𝑃∗𝑘𝑟𝑔 𝑃𝑑𝑒𝑤 𝑘 𝑃𝑅 𝑘 (𝑆 )
𝑘𝑟𝑜 𝑟𝑔 𝑟𝑔 𝑤𝑖
𝑞𝑔,𝑡𝑜𝑡𝑎𝑙 = 𝐶 (∫ + 𝑅𝑆 𝑑𝑝 + ∫ 𝑑𝑝 + ∫ 𝑑𝑝) 2.16
𝑃𝑤𝑓 𝐵𝑔 𝜇𝑔 𝐵𝑂 𝜇𝑂 𝑃∗ 𝐵𝑔 𝜇𝑔 𝑃𝑑𝑒𝑤 𝐵𝑔 𝜇𝑔

Prior and accurate knowledge of phase and drawdown behaviour is essential for
accurate estimation of well deliverability using three-flow regions pseudopressure
approach. The knowledge of phase behaviour is essential for reservoir engineer to
plan optimum production strategy for gas-condensate field development (Ugwu,
2011).
In concept of well deliverability modelling using equation 2.16, accurate estimation of
fluid phase behaviour determines the reliability of the developed model. The phase
behaviour of gas-condensate mixture is one of the most complex due to existing of 10
to 15% of heptane and heavier hydrocarbon components in the mixture. Estimating
fluid properties of such system to develop phase envelope require advance knowledge
of each composition as a function of pressure and temperature. A PVT model (e.g.,
equation of state or black oil) defines the relation between phase behaviour,
compositional variation and fluid properties. To emphasise the importance of accurate
PVT model for gas-condensate mixture, Whitson et al., (1999) highlighted that in
engineering treatment of gas-condensate reservoirs, the extra issues that must be
addressed are:
1. “how yielding the condensate will change during the life of the reservoir;
2. how two-phase gas and oil (condensate) flow near the wellbore, effect the
productivity”.
They suggested both aforementioned issues are strongly related to PVT properties of
the fluid. The knowledge of PVT properties and accurate estimation of each
parameters is paramount factor for accurate well deliverability modelling of gas-
condensate reservoirs.

25 | P a g e
Although the three-flow regions pseudopressure approach extensively used for
deliverability modelling of gas-condensate reservoirs in recent decade, but there is a
gap in current literature to show how inaccuracy in estimation of governing parameters
including PVT properties and relative permeabilities will affect the performance of the
equation. Main objective of this research is to identify the most important parameters
that effect well-deliverability modelling of gas-condensate reservoirs with an emphasis
to PVT properties. The effect of inaccurate estimation of PVT properties on
performance of three regions pseudopressure approach in equation 2.16 will be
investigated. Modified black oil (MBO) PVT properties were used in development of
three-flow region pseudopressure integral in 2.16. Hence, our analysis is based on
MBO approach and its parameters for estimation of three-flow regions
pseudopressure method. In following section, first we introduce two main PVT
modelling approach then the parameters that are contributing to three-flow regions
pseudopressure approach in equation 2.16 is discussed in details.

2.3.1 Modelling Fluid Properties

Currently there are two PVT models including black oil (BO) approach and
compositional approach. BO is based on simple interpolation of PVT properties as a
function of pressure (Spivak and Dixon, 1973; Coats, 1985; Coats and Smart, 1986;
Fevang and Whitson, 1996); and compositional model based on thermodynamically
consistent model such as cubic equation of state (Coats et al., 1995; Gomes and
Corrêa, 1992; Rubin and Buchanan, 1985).
BO PVT model is a fluid characterization formulation that represents multi-components
reservoir hydrocarbons in only two pseudo-components; “surface gas” and stock tank
oil (Fevang, Singh and Whitson, 2000; Walsh and Lake, 2003). Black oil model
quantifies gas and oil phase from a reservoir condition to standard surface condition
(pressure 14.69psia and temperature 60⁰F). The calculation is based on volumetric
estimation (conversion) of gas and oil in the reservoir condition to the surface
condition. This conversion also known as geologic condition of hydrocarbon fluid to
sellable value at the surface. Schematic diagram in Figure 2.2 shows the traditional
BO PVT formulation where the model accounts for surface gas and surface oil through
reservoir to the surface.

26 | P a g e
In order to estimate the volumetric values using BO PVT, the knowledge of how much
gas dissolved in oil phase at reservoir condition and how much of that oil would shrink
at surface condition is necessary. In addition, the knowledge of expansion of free gas
to several hundred times when it brought to the surface is required. To relate this
surface volume to reservoir volume and vice versa several factors were defined in BO
PVT formulation. Three main properties that serve the computations are wet gas
formation volume factor (Bg), oil formation volume factor (Bo) and solution gas to oil
ratio (Rs) shown in equation 2.17 to 2.19 respectively (Whitson and Brulé, 2000; Walsh
and Lake, 2003).

𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑔𝑎𝑠

𝐵𝑔 = 2.17
𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑢𝑟𝑓𝑎𝑐𝑒 𝑔𝑎𝑠 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑔𝑎𝑠

𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑜𝑖𝑙

𝐵𝑜 = 2.18
𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑜𝑐𝑘 𝑡𝑎𝑛𝑘 𝑜𝑖𝑙 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑜𝑖𝑙

𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑢𝑟𝑓𝑎𝑐𝑒 𝑔𝑎𝑠 𝑑𝑖𝑠𝑠𝑜𝑙𝑣𝑒𝑑 𝑖𝑛 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑜𝑖𝑙

𝑅𝑠 = 2.19
𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑜𝑐𝑘 𝑡𝑎𝑛𝑘 𝑜𝑖𝑙 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑔𝑎𝑠

Figure 2.2. Schematic diagram of traditional black oil formulation of PVT model.

These three properties constitute of traditional black oil model and are developed
based on the following assumptions.
A. Reservoir oil consists of only two components of stock-tank oil and surface gas

27 | P a g e
 B. Reservoir gas does not yields liquid when brought to the surface (see diagram
in Figure 2.2)
C. The property of surface gas released from reservoir oil is the same as
properties of reservoir gas
D. Properties of oil and gas are constant during pressure depletion
For gas-condensate reservoirs the assumptions of A and B are implicit because they
ignore the amount of oil (condensate) evolves from gas phase at the surface.
Neglecting the amount of produced oil (condensate) cause severe underestimation of
condensate recovery prediction at the surface. Therefore, to account for the amount
of produced condensate from reservoir gas at the surface another term called solution
oil to gas ratio (rs) was added to the traditional black oil PVT formulation by Kniazeff
and Naville, (1965).
Another assumption in BO PVT model is that the compositional variation is constant
during pressure depletion. This is not the true for gas-condensate reservoirs as the
composition of the mixture is changing with pressure depletion (Evinger and Muskat,
1942; Vo, 2010). Solution gas to oil ratio (rs) that introduced by Kniazeff and Naville,
(1965) also mimics the effect of the compositional variation on condensate properties
in BO PVT model (ECLIPSE, 2014). Solution oil to gas ratio (rs) various with respect
to pressure change in the reservoir.

Additional modification of traditional black oil PVT is reformulation of gas formation

volume factor to only include dry gas at the surface, which known as dry gas formation
volume factor (Bgd). Dry gas formation volume factor (Bgd) is the ratio between volume
of produced gas and volume of its gas components (Whitson, Da Silva and Soreide,
1988; Coats, Thomas and Pierson, 1995; Nassar, El-Banbi and Sayyouh, 2013).
Adding two aforementioned terms to black oil PVT model introduced modified black
oil model (MBO) PVT model. The schematic diagram of MBO PVT model is shown in
Figure 2.3 where four hydrocarbon components can be quantified in terms of volume
ratios defined in following.

28 | P a g e
 Figure 2.3 schematic diagram of modified black oil (MBO) model.

𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑔𝑎𝑠 𝑉𝑔

𝐵𝑔𝑑 = = 2.20
𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑢𝑟𝑓𝑎𝑐𝑒 𝑔𝑎𝑠 𝑝𝑟𝑜𝑑𝑢𝑐𝑒𝑑 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑔𝑎𝑠 𝑉𝑔̅𝑔

𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑜𝑖𝑙 𝑉𝑜

𝐵𝑜 = = 2.21
𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑜𝑐𝑘 𝑡𝑎𝑛𝑘 𝑜𝑖𝑙 𝑝𝑟𝑜𝑑𝑢𝑐𝑒𝑑 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑜𝑖𝑙 𝑉𝑜̅𝑜

𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑢𝑟𝑓𝑎𝑐𝑒 𝑔𝑎𝑠 𝑝𝑟𝑜𝑑𝑢𝑐𝑒𝑑 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑜𝑖𝑙 𝑉𝑔̅𝑜

𝑅𝑠 = = 2.22
𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑜𝑐𝑘 𝑡𝑎𝑛𝑘 𝑜𝑖𝑙 𝑝𝑟𝑜𝑑𝑢𝑐𝑒𝑑 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑜𝑖𝑙 𝑉𝑜̅𝑜

𝑐𝑜𝑛𝑑𝑒𝑛𝑠𝑎𝑡𝑒 𝑝𝑟𝑜𝑑𝑢𝑐𝑒𝑑 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑔𝑎𝑠 𝑉𝑜̅𝑔

𝑟𝑠 = = 2.23
𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑠𝑢𝑟𝑓𝑎𝑐𝑒 𝑔𝑎𝑠 𝑝𝑟𝑜𝑑𝑢𝑐𝑒𝑑 𝑓𝑟𝑜𝑚 𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑔𝑎𝑠 𝑉𝑔̅𝑔

MBO PVT mode successfully used for simulation of gas-condensate reservoirs PVT
characterization below the saturation pressure in previous studies (Whitson and Torp,
1983; Fevang and Whitson, 1996; Fevang, Singh and Whitson, 2000; Jokhio and Tiab,
2002; Mott, 2002; Izgec and Barrufet, 2005; Nassar, El-Banbi and Sayyouh, 2013;
Khamis and Fattah, 2019).
Fevang et al., (2000) in their simulation study compared MBO method with full
compositional PVT model using 22 components. They showed there is no significant
difference between MBO and compositional model in relation to 10 years of production

29 | P a g e
of a lean gas-condensate fluid. A comprehensive guideline for choosing compositional
and black oil models for gas-condensate reservoirs is provided by Fevang et al.,
(2000). For the case of gas injection using MBO for PVT simulation is not
recommended due to extra non-linearity that the injected gas (e.g., CO2) can added
to the mixture properties. Despite some deficiencies of the MBO PVT model, it is very
popular in industry due to its simplicity and faster CPU running time in reservoir
simulation studies.

Alternative approach for quantifying PVT properties of gas-condensate reservoirs are

compositional simulation studies (Coats, Thomas and Pierson, 1995). The
development of the compositional models is started due to increasing occurrence of
gas-condensate and volatile oil reservoirs (Coats, 1985; Rubin and Buchanan, 1985).
In compositional PVT model, the phase equilibrium and fluid properties such as
compressibility, density, and viscosity are determined by equation of state (EOS). An
equation of state presents a theoretical relationship between pressure, volume and
temperature of each individual components at various pressure and temperature
condition (Khanal, Khoshghadam and Lee, 2016). The cubic Peng and Robinson,
(1976) and Soave- Redlich-Knowng (1972) equations of state are commonly used in
the petroleum industry. In this technique, the phase composition is determined by
flashing the fluid over wide range of conditions (pressure and temperature). During the
computation process, the mass balance equation for each composition is used, where
sum of the saturations should be 1 (100%). The number of mass balance equations
increases as the number of compositions in the system increase, which require extra
computation time. Recently, full compositional simulation become more feasible with
advancement in computational techniques, however for large number of cells it is still
impractical.
The main advantage of the compositional method over black oil is in determining and
understanding relative effects of variable and parameters controlling/governing well
deliverability. Despite advantages of the compositional modelling of PVT, it requires
more computational effort than black oil model due to its great complexity. This can
restrict the model application in reservoir studies where significant compositional
variations occur (Gomes and Corrêa, 1992). Using either black oil or compositional
model for simulating PVT properties the PVT quantities required by reservoir simulator
are essentially the same (Whitson, 2006).
30 | P a g e
Well deliverability equation of 2.16 for gas-condensate reservoirs is presented in terms
of MBO PVT model. In following sections, the detail calculation of each parameter in
equation 2.16 is discussed. Some of the parameters like density and Z factor are not
appear in equation 2.16, however they are required for generating PVT table.

2.3.2 Compressibility Factor (Z factor)

The properties of gas mixtures are well defined with many graphical charts and
numerous equation of state in literature (Beggs and Brill, 1973; Standing, 1981;
McCain and Cawley, 1991). The behaviour of the gases at lower pressure originally
correlated based on experimental study of gases by Charles and Boyle, which is the
basis for ideal gas law. Ideal gas law is a thermodynamic equation shown in 2.24 that
allows correlating pressure, volume, temperature (PVT) and number of molecules
(moles) that are present in a gas sample.

𝑃𝑉 = 𝑛𝑅𝑇 2.24

Where p is pressure in pound square inch absolute (psia), V is occupied volume by

gas in cubic feet (ft3), n is mole of gas in pound-mole (lb-mole); T is temperature in
Rankine (R) and R is universal gas constant, which in customary unit is
𝑝𝑠𝑖×𝑓𝑡 3
10.73146 . Using ideal gas equation of state is suitable for gases in very low
𝑅 𝑙𝑏𝑚 𝑚𝑜𝑙𝑒

pressure. However, in higher pressure it leads to about 500% error for establishing the
PVT relationship. This is because ideal gas law assume no attraction or impulsive
forces between the molecules and all collisions between the molecules are assumed
perfectly elastic. Nevertheless, in reality, this is not the case for real gases and
intermolecular forces between mixture components strongly affect volumetric
behaviour. The deviation from ideal gas behaviour can be expressed as a factor. This
factor expressed as compressibility factor, deviation factor or Z factor. Throughout this
report, we used Z factor term.
Z factor is a dimensionless quantity and defined as actual volume of real gas to ideal
volume at specific pressure and temperature 𝑍 = 𝑉(𝑎𝑐𝑡𝑢𝑎𝑙 )/𝑉𝑖𝑑𝑒𝑎𝑙. Following is the

31 | P a g e
real gas law equation, which is standard description of volumetric behaviour of
hydrocarbon gas.

𝑃𝑉 = 𝑍𝑛𝑅𝑇 2.25

The real gas equation is valid for engineering calculations of most reservoir gases.
From real gas law in equation 2.25, all other volumetric PVT properties of gases can
be derived. In following sections, PVT properties that required for estimation of gas
and condensate production using three regions pseudopressure approach will be
reviewed.

2.3.3 Density of the mixture

Density is defined as a mass per unit volume in a specified pressure and temperature
(McCain and Cawley, 1991). In MBO PVT model the density of mixture gases is
defined as a function of pressure, well stream gravity, Z factor, universal gas constant,
temperature and some numerical constant as follow.

𝑃𝛾𝑤
𝜌𝑔 = 28.97 2.26
𝑍𝑅𝑇

Where pressure is in (psia), temperature in (R), density is in (lbm/ft3), 𝛾𝑤 is well stream

gravity. In gas-condensate reservoirs, it is important to use well stream gravity in
estimating density of the mixture. This is because well stream gravity represents the
average molecular weight of the mixture (produced gas and condensate) at the
surface condition (standard condition) and can be estimated from following equation
proposed by Standing, (1974).

𝛾̅𝑔̅ + 4580𝑟𝑝 𝛾̅𝑜̅

𝛾𝑤 = 𝛾 2.27
1 + 133,000𝑟𝑠 (𝑀)
𝑜̅

Where rp is total producing oil (condensate) to gas ratio at the surface in STB/scf; 𝛾̅𝑔̅
is average surface gas gravity; 𝛾̅𝑜̅ represents surface condensate gravity; 𝑀𝑜̅ is surface
condensate molecular weight. Average surface gas gravity can be calculated from the
following equation:

32 | P a g e
 ∑𝑁
𝑖=1 𝑅𝑝𝑖 𝛾𝑔𝑖
𝛾̅𝑔̅ = 2.28
∑𝑁𝑖=1 𝑅𝑝𝑖

Where Rpi is producing gas to oil ratio (GOR) at separator stage i in scf/STB and 𝛾𝑔𝑖 is
gas specific gravity at separator stage (Whitson and Brulé, 2000). The relation
𝛾
between surface specific gravity and condensate molecular weight (𝑀) in equation
𝑜̅

2.27, is defined by following equation proposed by Eilerts, (1947), which is valid for all
types of gas-condensate fluids (Whitson and Brulé, 2000).

𝛾
( ) = 0.001892 + 0.0000735𝛾𝐴𝑃𝐼 − (4.52 × 10−8 )𝛾𝐴𝑃𝐼 2 2.29
𝑀 𝑜̅

Surface condensate gravity 𝛾̅𝑜̅ in 2.27 is ratio of density of condensate (oil) to density
of water measured at standard conditions. 𝛾̅𝑜̅ is dimensionless quantity and is
calculated accurately by the following relation.

141.5
𝛾̅𝑜̅ = 2.30
131.5 + 𝛾𝐴𝑃𝐼

The major issue with engineering calculation of gas-condensate reservoirs are

unavailability of all the data. In practice only first-stage separator data including
solution gas to oil ratio (Rs1), gas specific gravity (𝛾𝑔1 ), stock tank oil gravity (𝛾̅𝑜̅ ),
pressure (Psp1) and temperature (Tsp1) are available (Gold, McCain and Jennings,
1989; Whitson, 2006). However, the total producing oil (condensate) to gas ratio (rp)
is still needed for calculation of well stream gravity using equation 2.27. This is inverse
of solution gas to oil ratio (Rs1) and additional gas that is released from condensate
phase at first-stage separator (Rs+).

1
𝑟𝑝 = 2.31
𝑅𝑠1 + 𝑅𝑠+

Where 𝑅𝑠+ can be calculated from the following correlation proposed by (Whitson,
1989).

33 | P a g e
 1.205
𝑃𝑝𝑠1 (0.0125𝛾𝐴𝑃𝐼 −0.00091𝑇𝑠𝑝1 )
𝑅𝑠+ = [( + 1.4) × 10 ] × 𝛾𝑔+ 2.32
18.2

Where 𝑃𝑝𝑠1 is in psia, 𝑇𝑠𝑝1 is in ⁰F and 𝑅𝑠+ is in scf/STB. 𝛾𝑔+ is a gas gravity of additional
solution (gas released from separator oil). 𝛾𝑔+ can be estimated from Katz, (1942)
correlation as follow.

𝛾𝑔+ = 0.25 + 0.02𝛾𝐴𝑃𝐼 − (3.57 × 10−6 )𝛾𝐴𝑃𝐼 × 𝑅𝑠+ 2.33

Substituting equation 2.31 in 2.32 yields the 𝑅𝑠+ in the following form.

𝐴1 𝐴2
𝑅𝑠+ = 2.34
(1 − 𝐴1 𝐴3 )

Computation of gas phase density from aforementioned procedure normally gives a

reasonable estimation if accurate Z factor is provided in gas-condensate fluid. Gas
density is used for estimation of gas viscosity in MBO PVT model.
To estimate condensate phase density, condensate formation volume factor is
required. Hence, we discuss formation volume factor of gas-condensate mixture in
next section.

2.3.4 Formation Volume Factor of Gas-Condensate Mixture

Formation volume factor is a PVT property of black oil model to convert volume of
hydrocarbon mixture at elevated pressure and temperature to the volume of product
phase at surface pressure and temperature. Surface pressure and temperature in
standard condition is defined as pressure of 14.7 psia and temperature of 520
°Rankine or 60°Fahrenheit.

𝑉𝑚𝑖𝑥𝑡𝑢𝑟𝑒 (𝑃, 𝑇)
𝐵𝑔 = 2.35
𝑉𝑝𝑟𝑜𝑑𝑢𝑐𝑡 (𝑃𝑠𝑐 , 𝑇𝑠𝑐 )

Where Psc and Tsc are pressure and temperature at standard condition. For gas-
condensate mixture, because reservoir fluid produce liquid (condensate) at the surface
as shown in Figure 2.3, two forms of formation volume factor are used for PVT

34 | P a g e
calculation as discussed before in equations 2.17 and 2.20. These include wet gas
formation volume factor (Bg) and dry gas formation volume factor (Bgd).
Using real gas law in equation 2.25, mathematically the wet gas formation volume
factor in ft3/scf is defined as follow.

𝑍𝑛𝑅𝑇
𝑃 𝑃𝑠𝑐 𝑍𝑇
𝐵𝑔 = = 2.36
𝑍𝑠𝑐 𝑛𝑅𝑇𝑠𝑐 𝑇𝑠𝑐 𝑃
𝑃𝑠𝑐

Where Z factor at standard condition (Zsc) is unity and substituting standard condition
of 14.7 psia for Psc and 520°Rankine for Tsc, following relationship yields for calculating
wet gas formation volume factor.

𝑍𝑇
𝐵𝑔 = 0.02827 2.37
𝑃

It should be noted that wet gas formation volume factor (B g) is fully defined from real
gas law as shown in equation 2.37, therefore no correlations are available in literature
for computation of Bg.
In MBO PVT model for gas-condensate fluid, to take the phase change during
depletion into consideration, dry gas formation volume factor is normally used.
The dry gas formation volume factor (Bgd) at standard pressure and temperature (14.7
psia and 520 °R or 60 °F) is defined as follow (Katz, 1942; Standing, 1974; Whitson
and Brulé, 2000).

𝑃𝑠𝑐 𝑍𝑇 𝑍𝑇
𝐵𝑔𝑑 = (1 + 𝐶𝑜̅𝑔 𝑟𝑠 ) = 0.02827 (1 + 𝐶𝑜̅𝑔 𝑟𝑠 ) = 𝐵𝑔 (1 + 𝐶𝑜̅𝑔 𝑟𝑠 ) 2.38
𝑇𝑠𝑐 𝑃 𝑃

Where temperature (T) is in ⁰R, pressure (P) is in psia, solution oil to gas ratio (rs) is in
STB/scf and dry gas formation volume factor (Bgd) is in ft3/scf. 𝐶𝑜̅𝑔 is conversion factor
from surface oil volume in Stock Tank Barrel (STB) to equivalent surface gas in
Standard Cubic Foot (scf) and estimated as follow.

𝛾𝑜̅𝑔
𝐶𝑜̅𝑔 = 133000 2.39
𝑀𝑜̅𝑔

35 | P a g e
 𝛾𝑜̅𝑔
Where can be estimated from equation 2.29. If composition of the mixture is not
𝑀𝑜̅𝑔

available condensate phase molecular weight can be estimated from Cragoe, (1929)
recommended by Whitson et al., (2000).

6084
𝑀𝑜̅𝑔 = 2.40
𝛾𝐴𝑃𝐼 − 5.9

Because formation volume factor is inversely proportional to pressure, reciprocal of

wet (1/Bg) and dry (1/Bgd) gas formation volume factors are used in reservoir
simulation. The typical graph of (1/Bg) and (1/Bgd) against reservoir pressure is shown
in Figure. 2.4 for gas-condensate reservoirs. Accurate estimation of wet and dry gas
formation volume factor is directly affected by accurate determination of Z factor at
desired pressure and temperature.

Figure 2.4. Typical gas-condensate formation volume factor as a function pf pressure.

Condensate (oil) phase formation volume factor (Bo) is another PVT property that
should be considered in computation of black oil PVT model. As shown in equation
2.21 the oil formation factor determines the volumetric ratio of oil (condensate) at
reservoir pressure and temperature to standard (surface) pressure and temperature.
Condensate (oil) formation volume factor is usually measured using constant volume
depletion (CVD) test if the reservoir sample is available. However, in many cases, the
samples are not always available and Bo is estimated using correlations. Developing
accurate correlation for estimating Bo at various pressure and temperature has

36 | P a g e
received extensive attention among the researchers. As a result, many correlations
are available for estimation of oil formation volume factor. Majority of the methods
developed based on experimental data from reservoir samples or measurement of
separator oil at the surface as a function of solution gas to oil ratio (Rs), specific gravity
of gas (𝛾𝑔 ), oil specific gravity (𝛾𝑜 ) and temperature (T) (Ahmed, 2010, p. 96). To select
the appropriate correlation for gas-condensate reservoirs the characteristics of
condensate fluid should be considered. The gravity of produced oil (condensate) can
be used as a selection criteria. Produced condensate is a very light oil with normally
specific gravity between 40 ≤ 𝐴𝑃𝐼 ≤ 60°𝐴𝑃𝐼. Six popular methods that are valid within
aforementioned API limit and frequently used in industry for estimation of Bo are
discussed in following.
Standing, (1947) presented a graphical correlation that correlated oil formation volume
factor as function of Rs, 𝛾𝑜 , 𝛾𝑔 , and T that later in 1981 its mathematical form presented
as follow by the author (Standing, 1981).

1.2
𝛾𝑔 0.5
𝐵𝑜 = 0.9759 + 0.000120 [𝑅𝑠 ( ) + 1.25(𝑇)] 2.41
𝛾𝑜

Where T is in ⁰F, Bo is in bbl/STB and 𝛾𝑜 is specific gravity of stock tank oil. The above
correlation is developed based on 105 experimental data points obtained from 22
Californian hydrocarbon reservoirs. Subsequently many other correlations have been
developed based on original Standing method.

Vazquez and Beggs, (1980) developed an oil formation volume factor correlation
based on 6000 data collected from world wide oil samples at various pressure. Using
regression analysis their correlation is valid for light oils with 𝛾𝐴𝑃𝐼 ≥ 30 and presented
as follow.
𝛾𝐴𝑃𝐼
𝐵𝑜 = 1 + 4.670 × 10−4 (𝑅𝑠 ) + 0.11 × 10−4 (𝑇 − 60) × ( )
𝛾𝑔𝑐
2.42
𝛾𝐴𝑃𝐼
− (1.8106 × 10−8 )𝑅𝑠 (𝑇 − 60) × ( )
𝛾𝑔𝑐
Where 𝛾𝑔𝑐 is the corrected gas specific gravity, included for the effect of separator
conditions and can be estimated from following correlation proposed by the same
authors.

37 | P a g e
 𝑃𝑆𝑒𝑝
𝛾𝑔𝑐 = 𝛾𝑔 [1 + 5.912(10−5 )𝛾𝐴𝑃𝐼 (𝑇𝑆𝑒𝑝 − 460)𝑙𝑜𝑔 ( )] 2.43
114.7

Where Tsep is separator temperature in Rankine and Psep stands for separator pressure
in psia.

From studying 45 oil PVT data Glaso, (1980) proposed his oil formation volume factor,
that can be used for saturated oil (condensate) Bo estimation. Comparative study of
Bo correlations by Sutton and Farshad, (1990) indicates Glaso’s correlation is
accurate for 𝐵𝑜 ≥ 1.4. Glaso’s correlation is presented in following form.

∗ −0.27683𝑙𝑜𝑔𝐵 ∗ 2
𝐵𝑜 = 1 + 10−6.58511+2.91329𝑙𝑜𝑔𝐵𝑜𝑏 𝑜𝑏 2.44

Where 𝐵𝑜𝑏 ∗is a correlating parameter that can be estimated as follow.

𝛾𝑔 0.526
𝐵𝑜𝑏 ∗ = 𝑅𝑠 ( ) + 0.968(𝑇 − 460) 2.45
𝛾𝑜

In above equation T is in ⁰F.

Based on 160 experimental data point of Middle Eastern oil reservoirs Al-Marhoun,
(1990) developed his saturated formation oil factor correlation as a function of Rs, 𝛾𝑜 ,
𝛾𝑔 and T. He recommended a linear relationship between Rs and Bo as follow.

𝐵𝑜 = 0.497069 + (0.8629 × 10−3 𝑇) + (0.18259 × 10−2 𝐹)

2.46
+ (0.31809 × 10−5 𝐹 2 )

Where F can be estimated as

𝐹 = 𝑅𝑠 0.742390 𝛾𝑔 0.323294 𝛾𝑜 −1.202040 2.47

Different approach for estimation of oil formation volume factor is recently developed
based on material balance relation (Ahmed, 2010, p. 97). The correlation is function

38 | P a g e
of solution gas to oil ratio (Rs), gas specific gravity (𝛾𝑔 ), oil specific gravity (𝛾𝑜 ) and oil
density and presented as follow.

62.4𝛾𝑜 + 0.0136𝑅𝑠 𝛾𝑔
𝐵𝑜 = 2.48
𝜌𝑜

Where 𝜌𝑜 is density of oil (condensate) at specified pressure and temperature in lb/ft3.

Ahmed (2010) shows that using equation 2.48 for estimation of Bo is the best among
aforementioned methods with the least average absolute error. More comprehensive
review on validity of oil formation volume factor correlations by Al-Shammasi, (2001)
and Aamir Mahmood and Ali Al-Marhoun, (1996) showed that the studied methods
return the measured Bo with less than 2% error for large data bank. The accuracy of
above Bo correlations are acceptable for PVT modelling of gas-condensate fluid in
three-flow region pseudopressure integral of 2.16. Accurate determination of Bo is also
important for reliable estimation of condensate density.

Condensate (oil) density is mass of a unit volume at specified pressure and

temperature usually estimated from the following relation proposed by Standing,
(1981).

62.4𝛾𝑜 + 0.0136𝑅𝑠 𝛾𝑔
𝜌𝑜 = 2.49
𝐵𝑜

Substituting equation 2.41 for Bo in above correlation yields the following relation that
is widely used in industry for estimation of density of oil at or below the bubble-point
pressure.

62.4𝛾𝑜 + 0.0136𝑅𝑠 𝛾𝑔
𝜌𝑜 = 1.2
𝛾𝑔 0.5 2.50
0.9759 + 0.000120 [𝑅𝑠 (𝛾 ) + 1.25(𝑇)]
𝑜

For gas-condensate mixture, gas specific gravity 𝛾𝑔 should be gravity of gas released
from separator oil (condensate). Standing correlation for density of the oil is superior
to other methods as it does not require correction term for pressure and temperature
(Whitson and Brulé, 2000; Ahmed, 2010). In a recent comparative study by Mmata
and Onyekonwu, (2014) it has been shown that using Standing correlation of 2.50 for

39 | P a g e
estimation oil density return the experimental values with reasonable accuracy of 2.4%
average absolute error.

2.3.5 Solution Gas to Oil Ratio (Rs)

Another important PVT property that needs accurate estimation in MBO PVT model is
solution gas to oil ratio (Rs). Rs defined as the number of standard cubic feet (scf) of
gas that dissolved in one stock tank barrel (STB) of oil at certain pressure and
temperature. This ratio also known as gas solubility (Ahmad, 2020). Rs is an important
volumetric property of MBO, normally measured using laboratory tests of constant
volume depletion (CVD) or constant composition expansion (CCE). Initial value of R s
remains constant until the reservoir pressure drops to below the dew point pressure
and increases afterward. The relationship between Rs and reservoir pressure is
demonstrated for a North Sea gas-condensate sample in following figure.

Figure 2.5. Solution gas to oil ratio (Rs) for North Sea gas-condensate sample at T=280⁰F
(Modified from Whitson and Torp, 1983).

Accurate estimation of Rs is important and directly affect accuracy of other PVT

properties such as formation volume factor (Bo), density (𝜌𝑜 ) and well stream gravity
(𝛾𝑤 ). If compositional data of the gas-condensate fluid is available solution gas to oil

40 | P a g e
ratio can be estimated through material balance method of Whitson and Torp, (1983)
or more complicated cubic equation of state (e.g., Peng and Robinsons).
There are also many empirical correlations for estimating Rs when composition of the
mixture is not available. Different authors developed their correlations based on the
data of certain localities, hence their applications are limited. Many comparative
studies are also conducted to see the superiority of these correlations (Ostermann and
Owolabi, 1983; McCain and Cawley, 1991; De Ghetto et al., 1994; Sutton, 2007). A
few more widely accepted correlations for estimation of Rs is discussed in following.

Standing, (1947) proposed a graphical correlation for solution gas to oil ratio as a
function of pressure, temperature, API gravity and specific gravity. Standing used
same oil samples discussed previously (105 Californian oil samples) in developing
following correlation for estimating Rs. The correlation is valid in the range of 20 – 1425
scf/STB (Danesh, 1998).

1.2048
𝑃
𝑅𝑠 = 𝛾̅𝑔̅ [( + 1.4) × 100.0125𝐴𝑃𝐼−0.00091(𝑇−460) ] 2.51
18.2

Where temperature is in Rankine and pressure is in psia. The Standing correlation

return the experimental solution gas to oil ratio with 4.8% error.
Following Standing’s method Vazquez and Beggs, (1980) developed a correlation for
estimating solution gas to oil ratio based on 5008 measured data points. The
correlation is valid for oil gravity of larger than 30 API and presented in the following
fashion.

𝛾𝐴𝑃𝐼
𝑅𝑠 = 0.178𝛾𝑔𝑐 𝑃1.1870 𝑒𝑥𝑝 [23.931 ( )] 2.52
𝑇

To adjust the gravity to reference separator pressure corrected gas specific gravity
should be used. This is because specific gravity of gas depends on the conditions that
gas separated from the oil phase. Corrected gas specific gravity (𝛾𝑔𝑐 ) in equation 2.43
can be used in Vazquez and Beggs, (1980) solution gas to oil ratio correlation. Sutton
and Farshad, (1990) showed that using Vazquez and Beggs, (1980) for estimating Rs,
predicts the experimental measurements with 12.7% error within the range of 0 – 2199
scf/STB.

41 | P a g e
To improve accuracy of previous models Glaso, (1980) proposed another solution gas
to oil ratio correlation based on 45 oil samples from North Sea, which is valid for Rs in
the range of 90 – 2637 scf/STB. The correlation reported as follow.

1.2255
𝛾𝐴𝑃𝐼 0.989 0.5
𝑅𝑠 = 𝛾̅𝑔̅ [( 0.712
) 102.8869−[14.1811−3.3093log(𝑃)] ] 2.53
(𝑇 − 460)

The author stated an average error of 1.28% for estimating solution gas to oil ratio.
Other correlations are also available in literature for computation of Rs, however most
of them followed Standing, (1947)’s correlation in their development process.
Another well received correlation was proposed by Petrosky and Farshad, (1998) for
estimation of Rs. They used 81 oil samples obtained from Gulf of Mexico in developing
their correlation. Petrosky and Farshad, (1998) proposed the following expression for
Rs.
1.73184
𝑃 0.8439 𝑋
{ 𝑅𝑠 = [(112.727 + 12.340) 𝛾̅𝑔̅ 10 ] } 2.54
−4 1.5410 −5 1.3911
𝑋 = 7.916 × 10 𝛾𝐴𝑃𝐼 − 4.561 × 10 𝑇

The above Rs correlation is valid within the range of 217 ≤ 𝑅𝑠 ≤ 1406𝑠𝑐𝑓/𝑆𝑇𝐵. The
authors reported that their correlation predicts the experimental Rs in specified range
with -0.05 average relative error.
Modelling solution gas to oil ratio (Rs) has received extensive attention among the
research community in recent decades. Many correlations are available that can be
used for accurate estimation of Rs in gas-condensate reservoirs. For instance Petrosky
and Farshad, (1998) predicts Rs with absolute error of -0.05. Also De Ghetto et al.,
(1994) showed that Rollins, McCain and Creeger, (1990) solution gas to oil ratio
correlation predicts the experimental Rs with 4.3% error for light oil with API >31⁰.

2.3.6 Viscosity (Gas/Condensate)

Two types of viscosities are usually applied for engineering calculation of hydrocarbon
reservoirs including dynamic viscosity (μ) and kinematic viscosity (ν). Dynamic
viscosity is defined as measuring resisting of fluid (gas and liquid) to flow under
external forces with the unit of centipoise (cp). However, kinematic viscosity is
resistance to flow of a fluid under gravitational affect (weight of fluid due to gravity

42 | P a g e
effect) with the unit of centistokes (cSt). Two viscosities are related by density of the
fluid (𝜇 = 𝜈𝜌).
Most petroleum engineering applications are using dynamic viscosity, which is for
Newtonian fluid (fluid that flows regardless of external forces with predictable viscosity
when pressure and temperature changes) mathematically defined as follow.

𝜏𝑔𝑐
𝜇= 2.55
𝑑𝑢⁄
𝑑𝑦

Where 𝜏 is shear stress per unit area in a shear plane parallel to the direction of flow,
𝑑𝑢⁄ is velocity gradient perpendicular to plane of shear and 𝑔 is a unit conversion
𝑑𝑦 𝑐

from mass to force.

Developing condensate phase below the saturation pressure in gas–condensate
reservoirs necessitate calculation of gas and condensate phase separately. Gas
viscosities of hydrocarbon reservoirs are generally within the range of 0.01 – 0.03cp
at standard reservoir conditions, increasing to 0.1cp for near critical gas-condensate
systems (Lohrenz, Bray and Clark, 1964; Whitson and Brulé, 2000). Measurement of
gas viscosities are rare because most of laboratories do not have the facility to conduct
the experiments and the prediction is normally through using correlations. Viscosity of
gases are usually correlated as a function of pressure, temperature and mixture
composition [𝜇𝑔 = 𝑓(𝑃, 𝑇, 𝑦𝑖)]. Based on this relationship many empirical and semi-
empirical correlations have been developed for estimating of gas viscosity. The
fundamental principle of using any gas viscosity correlations are as follow:
 Low pressure gas viscosity (𝜇𝑔𝑠𝑐 ) at standard pressure and temperature
condition should be estimated
 Corresponding state principle should be used to estimate actual value of
viscosity (𝜇𝑔 )
This would relate the actual gas viscosity (𝜇𝑔 ) at P and T to low pressure viscosity by
a ratio of 𝜇𝑔 ⁄ 𝜇𝑔𝑠𝑐 as a function of pseudoreduced properties (Ppr) and (Tpr) or as a
function of pseudoreduced density (ρpr). In following a few existing literature
correlations that are using this principles reviewed.
Carr et al., (1954) developed a graphical chart for estimation of gas viscosity using the
𝜇
relationship of 𝑔⁄𝜇𝑔𝑠𝑐 = 𝑓(𝑇𝑝𝑟, 𝑃𝑝𝑟). Later Dempsey, (1965), presented polynomial

43 | P a g e
relation of Carr et al., (1954) viscosity method with 15 constants. Dempsey, (1965) is
valid in the range of 1.2 ≤ 𝑇𝑟 ≤ 3 and 1 ≤ 𝑃𝑟 ≤ 20. In 1962, Jossi et al., (1962)
developed an empirical viscosity correlation for estimation of pure component. Then
Lohrenz et al., (1964) extended Jossi et al., (1962) correlation for estimation of mixture
viscosity. Lohrenz et al., (1964) mixture viscosity correlation usually referred to
Lohrenz – Bray – Clark (LBC) method. The LBC become standard compositional
reservoir simulator correlation for estimating gas/liquid viscosities. The LBC method is
very sensitive to the value of reduced density as the viscosity is estimated based on
forth degree polynomial reduced density.
𝜇
Lucas, (1981) used 𝑔⁄𝜇𝑔𝑠𝑐 relationship similar to Carr et al., (1954) and developed his

gas viscosity correlation that is valid for wider range of 1 ≤ 𝑇𝑟 ≤ 40 and 0 ≤ 𝑃𝑟 ≤

100.
Using corresponding state principle Pedersen and Fredenslund, (1987) developed a
viscosity model where the methane was reference fluid based on work of Christensen
and Fredenslund, (1980). In this method, viscosity of reference fluid (gas, liquid) is
estimated based on pressure, temperature, molecular size and density effect. Last two
parameters can be estimated empirically from fluid molar mass and reduced density.
The Pedersen and Fredenslund, (1987) is one of the fundamental viscosity methods
used in reservoir simulators such as Eclipse (Schlumberger) and Landmark’s Nexus
(Halliburton). Baled et al., (2018) in their review study showed that using Pedersen
and Fredenslund, (1987) correlation return the mixture viscosities with 5 – 15% error
and 15 – 30% for heavy alkane and binary mixtures.

Later in 1966 the authors Lee – Gonzalez – Eakin (LGE) (Lee, Gonzalez and Eakin,
1966) proposed simpler semi-empirical viscosity correlation irrespective the
knowledge of pseudo critical properties (Tpr and Ppr). The correlation is based on
density, molecular weight for estimation of gas viscosity. The correlation developed as
a function of pressure, gas density, temperature and molecular weight. Since its
publication, LGE become very popular due to its simplicity for estimating gas viscosity
in compare to corresponding state based methods. The LGE correlation is used by
most PVT laboratories for estimating gas viscosity and its application extended for
reservoirs simulators. Following LGE correlation, many other correlations have been
developed in similar fashion (Elsharkawy, 2006; Londono et al., 2002; Sutton, 2005).

44 | P a g e
These correlations are computationally simple and effective for quick estimation of gas
viscosity. For lower range of gas viscosities (0.01 – 0.03cp) the accuracy of gas
viscosity correlations are within 5 – 10% error, which is adequate for most applications
(e.g., dry gas, black oil reservoirs). However, in higher range of gas viscosities (0.03
– 0.1cp) the error of 20 – 30% may be expected (McCain and Cawley, 1991; Whitson,
2006). This error would proportionally effect the production forecast and well
deliverability estimation for gas-condensate reservoirs.

The difficulty of viscosity estimation even become harder for condensate phase. The
developed condensate inside the reservoir is a form of light oil with API gravity
between 40 – 60 (0.74 – 0.82 condensate specific gravity). The viscosity of
condensate fluid almost never measured or very rare because of several reasons
highlighted in flowing. Measurement of condensate viscosities are very difficult to
obtain due to unavailability of the samples, lack of high pressure high temperature
(HPHT) facilities, small volume cell viscometers and time and cost required for the
measurements (Whitson, Fevang and Yang, 1999; Al-Meshari et al., 2007; Hemmati-
Sarapardeh et al., 2014).
Therefore using correlations for estimation of condensate viscosity is become a norm
in industry. Nevertheless, using correlations for estimation of condensate viscosity are
exceedingly inaccurate. The best performance of the prevalent existing correlations
are within 10 – 30% error, and often increase to 50% (Whitson, 2006; Yang et al.,
2007).
Unacceptable accuracy of condensate viscosity may cause a serious problem in
developing a reliable well deliverability model, especially when condensate blockage
has significant impact on well deliverability (e.g., rich gas-condensate fluid in tight
formation) (Fevang, Singh and Whitson, 2000; Whitson, 2006). Previous studies show
1% error in reservoir fluid viscosity resulted in a 1% error in cumulative production
(total production of gas and oil) (Whitson, Fevang and Yang, 1999; Al-Meshari, 2004;
Yang et al., 2007).
Sutton, (2005) further investigate using existing literature models for estimation of gas-
condensate viscosity below the saturation pressure and highlights the following
drawbacks of existing literature models:
 They have a limited range of application.

45 | P a g e
  Smooth transition near critical point cannot be achieved.
 Accuracy of many viscosity models are function of accurate density estimation,
in other word they are density dependant.
 Many correlations are required heavy tuning of their constants to match the
experimental data.
 The performance of almost all viscosity correlations are notorious for estimation
of condensate phase viscosity.
Among the MBO PVT properties that associated with three regions pseudopressure
approach in 2.16 for estimation of well deliverability, viscosities (gas/condensate) has
highest prediction error as mentioned in this section. The aforementioned challenges
for viscosity (gas/condensate) prediction and gap in current literature motivated this
study to further focus on developing new approaches for estimation of this important
PVT parameter. The detail description of each viscosity model along with current
modelling approaches is discussed in chapter 4.

2.3.7 Drawdown Behaviour on modelling

The drawdown behaviour of gas-condensate reservoirs undergoing depletion

discussed in 1.3.2. This behaviour is directly affect permeability of the developed
phases as a results of pressure depletion inside the reservoir. Gas rate equation in 2.4
includes absolute permeability (k), which is the capacity and ability of the formation to
transmit fluid in a porous medium (Ahmed, 2010). In a single phase system using
absolute permeability is enough for well deliverability modelling. However, when
several phases (e.g., gas, condensate, water) flow simultaneously in a porous
structure like a hydrocarbon reservoirs, the permeability of the formation to each phase
should be estimated. Respective permeability of each phase known as relative
permeability (Kr). Relative permeability is one of the governing parameters that
influence gas and condensate flow rate at the surface using pseudopressure approach
in equation 2.16.
The flow in gas-condensate below the dew point is a combination of several phases
(at least two phases of gas and condensate), moving towards the wellbore region
simultaneously. Relative permeability characterizes how easily each fluid phase flows
through a porous medium in a multi-phase system. Fevang and Whitson, (1996)

46 | P a g e
investigated the effect of condensate banking on well deliverability by introducing
relative permeability into pseudopressure integral in equation 2.16.
Laboratory measurements are most common methods in determination of gas-
condensate relative permeabilities (Sumnu-Dindoruk and Jones, 1998). The data
obtained from measurements are the base of constructing relative permeability curves
as a function of phase saturation. Conceptual modelling of gas-condensate systems
suggests that relative permeability curves dictate the magnitude of gas productivity
loss (Afidick, Kaczorowski and Bette, 1994; Bourbiaux, 1994; Altug, Mo-Yuen Chen
and Trussell, 1999; Behmanesh, Hamdi and Clarkson, 2017). The use of inaccurate
relative permeability curves may result in over or underestimation of gas-condensate
reservoir performance prediction. Although laboratory measurements are standard
methods for estimating gas and condensate relative permeabilities using core flood
experiment, they involved several challenges as follow:
 The laboratory measurements of relative permeabilities are very expensive.
 The experiment is very difficult to conduct due to complexity of the gas-
condensate fluid below the dew point pressure.
 Limitation of core samples makes the obtained data from the experiments
limited where the whole saturation range cannot be determined (Sumnu-
Dindoruk and Jones, 1998).
Deficiencies of experimental procedure motivate engineers to rely on available existing
relative permeability correlations. Many empirical relative permeability methods have
been developed in past several decades to predict relative permeability under different
conditions. Some of these correlations are Brooks and Corey, (1964), Chierici, (1984);
Corey et al., (1956), Stone, (1973, 1970) and Wyllie, (1951).
In gas-condensate reservoirs when multi-phase flow occurs complex surface tension
(effect of the forces on the interface), and flow velocity effect near wellbore region,
makes relative permeability estimation even more difficult. This would influence the
accuracy of the correlational approach in estimating relative permeability.

The aforementioned difficulties of laboratory measurements make relative

permeability curves scarce and contradictory. Therefore, other approaches have been
proposed in literature for determination of relative permeabilities based on field
performance and or well pressure test data. Fetkovich et al., (1986) proposed a

47 | P a g e
method to determine relative permeabilities for solution gas drive reservoirs using oil
and gas flow rate time data. Al-Khalifa et al., (1987) proposed in-situ method to
estimate three-phase relative permeabilities using pressure transient analysis.
Hatzignatiou and Reynolds, (1996) proposed a procedure to estimate the effective
permeability (absolute permeability multiply by relative permeability) curve from
drawdown well test pressure data. Serra et al., (1990) introduced a procedure to
estimate effective permeability from transient radial drawdown data.
Sumnu-Dindoruk and Jones, (1998) modified Fetkovich et al., (1986) for estimation of
gas-condensate relative permeabilities. Their method require gas and condensate
production rate, PVT properties, calculated initial gas in place (IGIP) from material
balance, average reservoir pressure and critical water saturation. Jokhio, (2002) and
Jokhio and Tiab, (2002) developed a scheme to estimate gas and condensate
effective permeabilities from well pressure buildup test for gas-condensate reservoirs.

In this study mathematical manipulation of three regions pseudopressure approach

developed by Jokhio, (2002) is used for estimation of effective permeability of each
phase. The method relies on actual well test data during pressure build up/drawdown.
This would eliminate uncertainty of the correlational approach in generating relative
permeability curve, which is one of the main factors for inaccurate estimation of three-
flow regions pseudopressure approach in 2.16. The method is further discussed in
chapter 6 of this study.

2.4 Summary

Gas-condensate wells undergoing depletion are experiencing substantial productivity

lose at the surface due to condensate blockage near wellbore region. Condensate
blockage is a result of condensate drop out (heavy hydrocarbon drop out from gas
phase) below the saturation pressure. Condensate drop out effect phase and
drawdown behaviour of such system significantly. A model that accurately predict the
phase and drawdown behaviour with the effect of condensate blockage in gas-
condensate reservoirs are highly desirable. Such model require accurate estimation
of PVT properties and reliable treatment of phase permeabilities. The emphasis of this
study is on accurate estimation of gas and condensate PVT properties. Among the
PVT properties, accurate estimation of Z factor is the key because almost all other

48 | P a g e
PVT properties such as gas density, gas and oil formation volume factor (Bg and Bo),
well stream gravity are function of Z factor (Whitson and Brulé, 2000). Although there
are extensive research on accurate estimation of Z factor for natural gas systems
(Chamkalani et al., 2013; Elsharkawy, 2006; Rostami et al., 2018; Standing and Katz,
1942), however investigations for developing accurate Z factor models of gas-
condensate systems have not received enough attention.
Viscosity is another PVT property that need accurate prediction, yet it has highest
uncertainty as highlighted in 2.3.6 among the PVT properties. This is highlight the fact
that current literature approaches for estimation of gas-condensate viscosity are not
adequate in various conditions. The fundamental PVT properties of any PVT models
are Z factor and viscosity. The two properties of viscosity and Z factor are strongly
dependant on pressure and temperature. The performance of almost all existing
viscosity models are with very high error between 10 – 50% for gas-condensate
reservoirs (Whitson, 2006). Z factor is another crucial PVT property that always should
be estimated with high accuracy. Nevertheless previous studies (Elsharkawy, Hashem
and Alikhan, 2000; Ghiasi et al., 2014; Saghafi and Arabloo, 2018) showed that using
different mixing rules and hydrocarbon plus characterization the accuracy of current
existing methods for prediction of Z factor is with deviation of 8 – 56% from actual
values.
Hence, throughout the remaining of the thesis, the focus is on accurate modelling of Z
factor and viscosity for better PVT representation of gas-condensate fluid below the
saturation pressure. The effect of accurate determination of gas and condensate
viscosities and Z factor on well production profile is ultimate interest of this study. For
this purpose, three regions pseudopressure integral in equation 2.16 incorporated with
volumetric material balance to generate the production profile of studied gas-
condensate reservoir. The result of this study will be compared with actual field data
as well as compositional reservoir simulator.

49 | P a g e
 CHAPTER 3 METHODOLOGY AND
CONCEPTUAL FRAMEWORK
METHODOLOGY AND CONCEPTUAL FRAMEWORKM

3.1 Introduction

This chapter introduces the methodologies and framework that have been taken to
conduct this research study. First, a brief discussion of the existing literature
approaches for characterizing PVT models is provided. Then various well deliverability
models are discussed, where the advantages and drawbacks of each method are
highlighted. The justification behind the selected approach is also provided.

3.2 Pressure – Volume –Temperature (PVT) modelling

Traditionally PVT properties of hydrocarbon mixtures are estimated using black oil
(two components) model and compositional (several components) model. The later
fluid modelling provides better accuracy for gas-condensate reservoirs, as it has the
ability to monitor each component’s saturation at all reservoir pressure and
temperature stages (Khanal, Khoshghadam and Lee, 2016). However, the
computational procedure is more complex and require more time to run in compare to
black oil PVT model.
The development of the compositional models are started due to increasing
occurrence of gas-condensate and volatile oil reservoirs. The phase equilibrium and
fluid properties such as compressibility, density, and viscosity are determined by
equation of state (EOS). An equation of state represents a theoretical relationship
between pressure, volume and temperature. The cubic Peng and Robinson, (1976)
and Soave - Redlich - Knowng (1972) equations of state are commonly used in the
petroleum industry.
In this technique, the phase composition is determined by flashing the fluid over wide
range of pressures. During the computation process, the mass balance equation for
each composition is used, where sum of the saturations should be 1 (100%). The
phase behaviour of the fluid has to be consistent with pressure, temperature and
composition in each grid cell in reservoir simulations.

50 | P a g e
The number of mass balance equations increase as the number of compositions in
the system increase. This would add extra computation time in simulation studies.
Recently, full compositional simulation become more feasible with advancement in
computational techniques, however for large number of cells it is still impractical.
Despite advantages of the compositional modelling of PVT properties, it requires more
computational effort than black oil model due to its great complexity. The prediction
accuracy of equation of state in compositional simulation is deteriorate in near critical
regions (e.g., rich gas-condensate fluid and highly volatile oil) (Elsharkawy, 2006). This
can restrict the model application in reservoir studies where significant compositional
variations occur (Gomes and Corrêa, 1992).
To attempt for modifying black oil PVT model an approach was developed known as
modified black oil (MBO), where the knowledge of expansion of gas and shrinkage of
oil in the surface due to the amount of dissolved gas is added to the MBO (Izgec and
Barrufet, 2005; Nassar, El-Banbi and Sayyouh, 2013). MBO is discussed in details in
2.3.1. Many studies show excellent agreement of MBO with compositional models
(EOS) for simulating gas-condensate fluid (Whitson and Torp, 1983; Coats, Thomas
and Pierson, 1995; Mott, 2002; Chowdhury et al., 2004; Fan et al., 2005; Behmanesh,
Hamdi and Clarkson, 2017). MBO is much simpler than compositional formulation
using EOS. Three-flow regions pseudopressure equation in 2.16 proposed by
Fevang, (1995) written in terms of MBO PVT model. Therefore, in this study MBO PVT
approach also adopted to generate PVT properties of gas-condensate fluid.

3.3 Approaches of Well Deliverability Modelling

Existing of condensate liquid below the saturation pressure in gas-condensate

reservoir is making well deliverability modelling of such reservoir a challenging task.
This is because a comprehensive model that account for condensate blockage and
effect of capillary, viscous and inertial forces is a significant task for reservoir
engineers to implement.
Two fundamental approaches to study fluid flow in any hydrocarbon reservoirs are
including “classical” and “modern” techniques. Classical approach based on utilizing
analytical solutions of linear differential equations while modern methods consist of
numerical simulation models. Use of analytical technique can be regarded as more

51 | P a g e
specialized and difficult to apply, since it requires considerable knowledge to use in
particular equation to describe a physical situation in a reservoir (Dake, 2001).

The most common approach of modelling well deliverability of gas-condensate

reservoir and predicting well performance is through reservoir simulation studies. The
simulation models incorporate the rock and fluid properties to predict the dynamic
influence of condensate blockage over gas and condensate production (Fan et al.,
2005). The standard compositional industry simulator is Eclipse – 300 that enables
better prediction of well deliverability by using small grids near wellbore region, where
condensate blockage is exist (Fan et al. 2005). The small grids can be constructed
using Local Grid Refinment (LGR). The disadvantage of LGR for modelling near
wellbore region is a significant increase in computation time especially for reservoirs
with tight formation. One of the major issues of gas-condensate resevoirs undergoing
depletion is compositional changes during depletion. Compositional simulation studies
enable engineers to detect compositional changes below the saturation pressure using
advanced cubic equation of state (EOS).
Although compositional simulation studies are very popular in industry, however there
are many applications that these type of analaysis is not justifiable (Mott, 2002; Fan et
al., 2005; Bonyadi, Rahimpour and Esmaeilzadeh, 2012). This is due to their
computational efficiency and considerable data required to start the study. The typical
problem with engineering calculation of gas-condensate reservoirs is unavailability of
all the data to run the simulation (Whitson, 2006). Therefore, other techniques so
called analystical methods (e.g,. steady-state pseudopressure or three-flow regions
pseudopressure approach) are currently exist in literature that allow such studies with
good accuracy comparable to compositional simulation studies (Fevang and Whitson,
1996; Mott, 2002; Chowdhury et al., 2004; Fan et al., 2005). The
For gas-condensate well deliverability modelling, three-flow regions pseudopressure
approach, introduced in 2.16, has become a standard choice in recent years. This
method can be implemented in spreadsheet and its very useful for quick estimation of
well deliverability where many sensitivity runs are required (Fan et al., 2005).
Comparison of fully compositional simulation and local grid refinement with
pseudopressure approach of equation 2.16 incorporated with non-Darcy flow show
that pseudopressure method captured all near wellbore condensate blockage effect
accurately (Mott, 2002; Chowdhury et al., 2004; Fan et al., 2005; Behmanesh, Hamdi
52 | P a g e
and Clarkson, 2017). The main advantage of pseudopressure approach is easy to
implement and there is no increase in computational time (Mott, 2002; Chowdhury et
al., 2004; Fan et al., 2005). To successfully implement three-flow regions
pseudopressure approach reliable relative permeability curves, accurate estimation of
GOR and accurate determination of PVT properties are required (Fevang and
Whitson, 1996; Roussennac, 2001). The aim of this study is to investigate the
accuracy of three-flow regions pseudopressure approach for gas-condensate
performance modelling with an emphasis on accurate estimation of PVT properties
including gas/condensate viscosity and Z factor.

3.4 Current Study Approach

One of the major issues in modelling of well deliverability in gas-condensate reservoirs

using any discussed approaches is accurate determination of PVT properties. In
relation to MBO PVT, used in three-flow regions pseudopressure integral in equation
2.16, the main PVT properties are gas and oil formation volume factor (Bg, Bo) (they
are function of two-phase Z factor), gas and condensate viscosities (𝜇𝑔 , 𝜇𝑜 ) and
solution gas to oil ratio (Rs). The estimation of MBO PVT properties mainly rely on
correlations when the compositional data is not available. The problems and limitation
of existing literature correlations for estimating different PVT parameters (Bg, Bo,
Rs,𝜇𝑔 , 𝜇𝑜 ) are reviewed in 2.3.
In this research, extensive attempt has been made to eliminate some of the difficulties
that PVT systems encounter for modelling gas-condensate fluid properties below the
saturation pressure. In doing so for better and more accurate PVT representation of
gas-condensate fluid the Artificial Intelligent (AI) or known as machine learning
techniques were utilized. This approach was taken due to availability of the data in
recent years for gas-condensate reservoirs. The AI methods are data driven
techniques, that making prediction based on the real data rather than physics of the
problem. There are many successful applications of AI techniques for reservoir
modelling in recent years. Some of the examples of the machine learning techniques
for different aspect of reservoir modelling in literature can be found in (Ahmadi et al.,
2014; Ghiasi et al., 2014; Hemmati-Sarapardeh et al., 2014; Kamari et al., 2013;
Naderi and Khamehchi, 2019; Shokir, 2008).

53 | P a g e
Various algorithms of Artificial Intelligent manipulated for better modelling of gas-
condensate PVT properties. The AI methods applied in this study listed as follow.
 Artificial Neural Network (Feed Forward Neural Network, Cascade Forward
Neural Network)
 Support Vector Machine (SVM)
 Fuzzy Logic [including Takagi – Sugeno – Kang (TSK) and Mamdani]
 Adaptive Neuro Fuzzy Inference System (ANFIS)
Based on the above AI methods several models have been developed, which offer
accuracy, mathematical efficiency and simplicity in compare to existing literature
correlations for estimation of viscosity and Z factor of gas-condensate reservoirs.
In order to see the effectiveness of the developed viscosity and Z factor models in
relation to well deliverability of gas-condensate reservoirs; three regions
pseudopressure method also was employed. Three regions pseudopressure method,
incorporated with results of PVT properties (gas/condensate viscosity, Z factor) for
calculation of well inflow performance. The ultimate interest of this study is to improve
gas-condensate well production performance modelling. Hence, the results obtained
from pseudopressure method incorporated with volumetric material balance to
generate production profile of gas and condensate on the surface.

Our approach for computation of well inflow performance and generating production
profile is similar to previous study performed by (Fevang and Whitson, 1996; Jokhio
and Tiab, 2002; Mott, 2002; Chowdhury et al., 2004; Arukhe and Mason, 2012;
Bonyadi, Rahimpour and Esmaeilzadeh, 2012; Behmanesh, Hamdi and Clarkson,
2015, 2017; Hekmatzadeh and Gerami, 2018). However, the advantage of the current
approach is that the properties that require for computing three-flow regions
pseudopressure integral estimated using AI techniques to ensure better accuracy of
the model is achieved.

3.5 Major Work Steps

The conceptual framework of this investigation can be summarized and major work
steps that have been taken are highlighted as follow.
i. Critical review of the relevant literature was carried out and the gaps and
shortfalls in current literature are identified. This literature review assisted to

54 | P a g e
 shape aims and objectives of the study in improving well deliverability
modelling of gas-condensate reservoirs.
ii. Extensive published literature used to develop a data bank to improve PVT
modelling of gas-condensate reservoir through accurate modelling of
gas/condensate viscosity and two-phase Z factor.
iii. Several dynamic gas/condensate viscosity models have been developed using
regression and AI methods including ANN, SVM and Fuzzy Logic.
iv. Several two-phase Z factor models have been developed using smart AI
methods including Cascade-Forward Neural Network (CFNN), Feed-Forward
Neural Network (FFNN) and Adaptive Neuro Fuzzy Inference System (ANFIS).
v. A dynamic Inflow Performance Relationship (IPR) curve has been developed
using two-phase pseudopressure integral for high temperature gas-
condensate reservoirs in tight formation.
vi. The developed models in this study (viscosity, Z factor) were combined with
well deliverability equation of pseudopressure integral and material balance
equation to establish production profile of a single gas-condensate well.

3.6 Data Acquisition and Validation

Practical application of the developed models were verified by using several case
studies. Required data for this study was sourced from published open literature. PVT
reports, experimental studies, published papers were source of our data bank. An
extensive data bank for each phase of the study was collected and utilized in
developing data driven (AI) methods. Availability of the data to some extent in recent
years motivated using the current approach in this study.
The gathered data covered wide range of gas-condensate reservoir conditions
(pressure, temperature and compositional differences). An example of this data bank
shown in Appendix A.
To ensure the validity and accuracy of the data used for this study the following steps
were taken,
 The statistical and graphical analysis of the gathered data against current
methods in literature were performed in order to verify the data bank.
 The physical trend of the data were performed and the consistency of the data
sets have been confirmed.

55 | P a g e
3.7 Summary

The methodologies in this chapter have been implemented to ensure accurate well
deliverability modelling for gas-condensate reservoir below the dew point pressure is
achieved. The data bank for development of models in this study is from published
literature. This data source partially used for development, testing and validating of
each model. To ensure robustness of the developed models the random selection of
the data in each step has been implemented. The current chapter provides a road map
of the methodologies that have been adopted in this study.

56 | P a g e
 CHAPTER 4 IMPROVEMENT OF GAS-
CONDENSATE FLUID VISCOSITY PREDICTION
IMPROVEMENT OF GAS-CONDENSATE FLUID VISCOSITY
PREDICTION

4.1 Introduction

Viscosity is one of the governing parameters for modelling gas-condensate well

deliverability (Whitson, Fevang and Yang, 1999; Hernandez; et al., 2002; Yang et al.,
2007; Arukhe and Mason, 2012). To emphasise the importance of viscosity on
production the research shows 1% error in calculating reservoir fluid viscosity resulted
in 1% error in cumulative production (Al-Meshari et al., 2007; Fevang and Whitson,
1996; Hernandez; et al., 2002; Sutton, 2005; Whitson et al., 1999). Davani et al.,
(2013) also showed that small error in viscosity prediction would have big impact on
production forecast. This lead to significant error in financial evaluation of the project.
Although the accurate estimation of viscosity is critical for well deliverability modelling,
yet it has highest uncertainty prediction using existing literature models.
Behmanesh et al., (2017) found that using single dry gas viscosity and Z factor effect
the performance prediction of gas-condensate reservoirs. Using single-phase
viscosity models for estimation of gas-condensate viscosity are valid above saturation
pressure, as condensate liquid yet to develop. However, below the saturation pressure
using single-phase viscosity models are not valid as the fluid become two phases of
gas and condensate. The developed phases of gas-condensate mixture should be
computed individually for modelling purposes due to gravity segregation of the phases.
The aim of this chapter is to improve several models for better prediction of gas-
condensate fluid viscosity below the saturation pressure.
This chapter provides the detail modelling of gas-condensate fluid viscosity below the
saturation pressure. Initially the viscosity of gas phase in gas-condensate reservoirs
is studied using existing viscosity models in literature. The actual gas phase viscosity
data is used to optimize a well-known viscosity model using non-linear regression.
Several models have been developed for better and accurate prediction of condensate
viscosity in gas-condensate reservoirs using ANN, SVM and Fuzzy Logic approaches.

57 | P a g e
4.2 Assessment of gas phase viscosity models

Gas viscosity models in literature are categorized to two types based on availability of
the data and other reservoir information. First type are semi-empirical correlations that
correlate gas viscosity as a function of pressure, temperature and mixture
composition. These type of correlations initially are estimating the fluid (gas & liquid)
viscosity at lower pressure (𝜇𝑔𝑠𝑐 ) and relate this value to actual viscosity (𝜇𝑔 ) using
corresponding state principles (CSP). The CSP states that two substances at the
same conditions of critical pressure and critical temperature will have similar
properties. These conditions are known as reduced temperature (Tr) and reduced
pressure (Pr) (Katz, 1959).
To name a few of these correlations such as (Carr, Kobayashi and Burrows, 1954;
Jossi, Stiel and Thodos, 1962; Lohrenz, Bray and Clark, 1964; Lucas, 1981; Pedersen
and Fredenslund, 1987). In compositional reservoir simulators (e.g., Eclipse),
(Pedersen and Fredenslund, 1987) and Lohrenz et al., (1964) known as LBC are very
popular for estimating viscosity of hydrocarbon fluid (gas/liquid).
Second catagory of the correlations are less computationally intensive and they are
very popular for reservoir engineering PVT calculations. Some of these methods
specifically developed to address gas-condensate issues in estimating viscosity.
The applicability of the most popular literature methods in prediction of gas viscosity
will be assessed in following using actual viscosity data obtained from the open
literature.
Lohrenz – Bray – Clark (LBC), 1964

Jossi et al., (1962) proposed a correlation for estimating viscosity of gas and liquid
pure components. Later Lohrenz et al., (1964) used same equation of Jossi et al.,
(1962) for estimating hydrocarbon mixture viscosity. The equation usually referred to
Lohrenz – Bray – Clark (LBC) correlation and originally developed based on residual
viscosity concept and the theory of corresponding state principle (CSP). LBC
correlation shown in following relates mixture viscosity (𝜇) to fourth degree polynomial
of the reduced density.

1
[(𝜇 − 𝜇 ∗ )𝜁 + 10−4 ]4 = 𝐴 0 + 𝐴 1 𝜌𝑟 + 𝐴 2 𝜌𝑟 2 − 4.1
𝐴 3 𝜌𝑟 3 + 𝐴 4 𝜌𝑟 4

58 | P a g e
Where ζ, 𝜌𝑝𝑟 and 𝜇 ∗ represent viscosity reducing parameter, reduced density and low
pressure gas mixture viscosity respectively, A0 – 4 are LBC coefficients of 0.1023,
0.023364, 0.058523, -0.040758 and 0.0093324 respectively.
1/6
𝑇𝑝𝑐
𝜁 = 5.35 ( 3 4 )
𝑀𝑖 𝑃𝑝𝑐
𝜌 𝜌
𝜌𝑝𝑟 = = 𝜈𝑝𝑐 4.2
𝜌𝑝𝑐 𝑀
∑𝑁𝑖=1 𝑧𝑖 𝜇𝑖
𝜇∗ =
{ ∑𝑁𝑖=1 𝑧𝑖 √𝑀𝑖

Where pseudocritical properties of temperature 𝑇𝑝𝑐, pressure 𝑃𝑝𝑐 and volume 𝜈𝑝𝑐 can
be calculated from Kay’s mixing rule (Kay, 1936). In equation 4.2, 𝑧𝑖 stands for mole
fraction of each pure component 𝑖 and 𝑀𝑖 is molecular weight of each component.
To establish special relation between 𝐶7+ fractions of gas-condensate mixture and
critical volume, LBC suggested the following relationship.

𝑣𝑐𝐶7+ = 21.573 + 0.015122 𝑀𝐶7+ − 27.65𝛾𝐶7+ + 0.070615 𝑀𝐶7+ 𝛾𝐶7+ 4.3

Where 𝑣𝑐𝐶7+ is the critical molar volume, 𝑀𝐶7+ is molecular weight of C7+ fraction and
𝛾𝐶7+ is specific gravity of C7+ fraction.
The low pressure, pure component gas viscosity for each composition 𝜇𝑖 ∗ , can be
calculated using Stiel and Thodos, (1962) expression as follows.

𝜇𝑖 ∗ 𝜁𝑖 = (34 × 10−5 )𝑇𝑟 0.94 𝑓𝑜𝑟 𝑇𝑟 ≤ 1.5

{ } 4.4
𝜇𝑖 ∗ 𝜁𝑖 = (17.78 × 10−5 )(4.58𝑇𝑟 − 1.67)5/8 𝑓𝑜𝑟 𝑇𝑟 > 1.5

In LBC mixture viscosity correlation ‘𝜇’ is in centipoise (cp), viscosity reducing

parameter ‘ζ’ is in cp-1, density ‘ρ’ is in lbm/ft3, specific volume ‘𝑣𝑐 ’ is in ft3/lbm mol,
temperature ‘T’ is in Rankine (°R), pressure ‘P’ is in psia, and molecular weight of each
component ‘Mi’ is in lbm/lbm mol. The accuracy of the LBC in predicting oil viscosity
reported with average absolute deviation of 16 percent. They reported that their
correlation is best performed in the range of 8.33 – 73.49⁰F operating temperature and
4.94 – 7014psia reservoir pressure.

59 | P a g e
We selected LBC correlation for this study because it is widely implemented in
compositional reservoir simulators and it is computationally simpler and faster than
other CSP based viscosity correlations such as Pedersen and Fredenslund, (1987).
LBC method is sensitive to the density value since the method is developed with fourth
degree polynomial of reduced density.
LBC can be used for estimating viscosity of gas and liquid at various reservoir pressure
and temperature. However to match the measured viscosity values, the LBC
coefficients are normally tuned. The tuning is generally through critical volume of C 7+
components or LBC original constantans.

Lee – Gonzalez – Eakin (LGE)

Unlike the LBC correlation, which correlates low-pressure gas viscosity with
temperature, molecular weight, pseudocritical temperature, and pseudocritical
pressure, Lee et al., (1966) proposed a correlation that relates low-pressure gas
viscosity with gas gravity (or molecular weight) and temperature. Lee et al., (1966)
correlation commonly referred to Lee – Gonzalez – Eakin (LGE). In both LBC and LGE
method gas viscosity is estimated by multiplying low pressure viscosity by a ratio. The
ratio is correlated with density in LGE and reduced density in LBC. Sutton, (2007)
showed that the performance of both corresponding state principle based correlations
and LGE method is the same for estimation of gas viscosity.
In LGE method the authors modified theoretical viscosity expression of Starling and
Ellington, (1964) and proposed a semi-empirical relation for estimation of viscosity of
natural gas as a function of reservoir temperature, gas density (ρ) and fluid
composition (yi). The LGE correlation shown in 4.5 is based on measured data of pure
component of 8 natural gases with specific gravity of less than 1.
𝜌𝑔 𝑌
𝜇𝑔 = 10−4 𝐾 × 𝑒𝑥𝑝 [𝑋 ( ) ]
62.4
(9.379 + 0.016𝑀𝑎)𝑇 1.5
𝐾= 4.5
209.2 + 19.26𝑀𝑎 + 𝑇
986.4
𝑋 = 3.448 + + 0.01009𝑀𝑎
𝑇
{ 𝑌 = 2.4 − 0.2𝑋

60 | P a g e
Where T is temperature in Rankine (⁰R) and 𝜌𝑔 is density at reservoir pressure and
temperature in lb/ft3, Ma is apparent molecular weight. Apparent molecular weight is
calculated as follow (Ahmed, 2010).

𝑀𝑎 = ∑ 𝑦𝑖 𝑀𝑖 4.6
𝑖=1

Where Mi is molecular weight of each component in the mixture and can be obtained
from property tables; 𝑦𝑖 is mole fraction of each composition in the mixture. LGE is one
of the main viscosity correlations in any compositional reservoir simulator for
estimating of gas viscosity. In addition, the LGE method is very simple to use and is
very popular for quick estimation of gas viscosity.
The method is limited for temperature range of 100 – 340⁰F and pressure range
between 100 – 8000psia. The authors reported +/-2.7 error for estimating low pressure
gas viscosity and 4% error for high pressure, provided density and molecular weight
of compositions. The LGE correlation becomes less accurate for prediction of gas
viscosity above the specific gravity of 1.
Success of LGE method for estimating gas viscosity motivated many researchers to
present their viscosity correlations in similar fashion to LGE. Some of these methods
will be discussed in following sections.

Londono – Archer – Blasingame (LAB)

Londono – Archer – Blasingame (LAB) (Londono, Archer and Blasingame, 2002)

optimized LGE method to make it more accurate for estimating viscosity of pure
component and light-natural gas mixture. LAB used large database of 4909 data points
and cast the LGE method in the following form.
𝜇𝑔 = 10−4 𝐾𝑒𝑥𝑝[𝑋𝜌𝑔 𝑌 ]
(16.7175 + 0.0419188𝑀𝑤)𝑇 1.40256
𝐾=
212.209 + 18.1349𝑀𝑤 + 𝑇

𝑌 = 1.09809 − 0.0392581𝑋 4.7

2063.71
𝑋 = 2.12575 + + 0.011926𝑀𝑤
𝑇
−2
𝑀𝑤. 𝑃
{ 𝜌𝑔 = 1.601846 × 10
𝑅𝑇

61 | P a g e
Where 𝜌𝑔 is density (g/cm3); Mw is molecular weight of the gas, T is temperature, 𝜇𝑔 is
gas viscosity, R is universal gas constant and P is reservoir pressure. The average
absolute error (AAE) using optimized form of LGE in 4.7 was reported 2.29% for LAB
data. The main difference between original and optimized form of LGE in 4.7 is that
the original form of LGE generated using smaller database of mixture component,
whether the optimized version used extensive database.

Sutton, (2005)

Sutton, (2005) used low pressure gas viscosity (𝜇𝑔𝑠𝑐 ) of Lucas, (1981) and modified
original LGE correlation for estimating viscosity of gas and gas-condensate reservoirs.
Sutton correlation is as follow,
𝜌𝑔 𝑌
𝜇𝑔 = 𝜇𝑔𝑠𝑐 × 𝑒𝑥𝑝 [𝑋 ( ) ]
62.4
1588 4.8
𝑋 = 3.47 + + 0.0009𝑀𝑎
𝑇
{ 𝑌 = 1.66378 − 0.04679
Where 𝜇𝑔𝑠𝑐 is defined as a function of pseudo critical properties of temperature (Tpc),
pressure (Ppc) and can be estimated as follow.

−4 0.807𝑇𝑟 0.618 − 0.357 exp(−0.449𝑇𝑟 ) +

𝜇𝑔𝑠𝑐 𝜉 = 10 [ ]
0.34𝑒𝑥𝑝(−4.058𝑇𝑟 ) + 0.018
1/6
4.9
𝑇𝑝𝑐
𝜉 = 0.9490 ( 3 4 )
{ 𝑀 𝑃𝑝𝑐

The relationship between pseudo critical properties and reduced properties can be
defined as follow.
𝑇
𝑇𝑟 =
𝑇𝑝𝑐
4.10
𝑃
𝑃𝑟 =
{ 𝑃𝑝𝑐

The pseudo-critical properties of temperature and pressure related to the gas specific
gravity (𝛾𝑔 ) and following two correlations were introduced by Sutton, (2005).

62 | P a g e
 𝑇𝑝𝑐 = 164.3 + 357.7𝛾𝑔 − 67.7𝛾𝑔 2
{ 4.11
𝑃𝑝𝑐 = 744 + 125.4𝛾𝑔 + 5.9𝛾𝑔 2

The above two correlations are based on 634 hydrocarbon sample compositions
obtained from 275 constant volume depletion (CVD) report for gas-condensate and
associate gas samples. Sutton, (2005) method is valid for gas specific gravity of
between 0.554 – 2.819, temperature of 0 – 460⁰F and pressure of 12 – 17065psia.
Sutton method is widely accepted among the research community for estimating gas
viscosity using pseudocritical properties in 4.11 based on specific gravity of the
mixture, knowing the compositions.

Elsharkawy (2006)

Elsharkawy, (2006) added calculation of non-hydrocarbon impurities including carbon

dioxide (CO2) and hydrogen sulphite (H2S) to original LGE gas viscosity correlation.
He also added another expression to LGE to account for existing of heptane plus
fraction (C7+), which is particularly important in estimating viscosity of gas-condensate
fluid. Existing of non-hydrocarbon impurities in natural gas and gas-condensate
mixtures influence the accuracy of PVT properties such as viscosity and Z factor.
Elsharkawy, (2006) used 2400 data point collected from experimental and published
literature in developing his method. Elsharkawy, (2006) method is as follow.
𝜇𝑔𝑐 = 𝜇𝑔 + ∆𝜇𝐶7+ + ∆𝜇𝐶𝑂2 + ∆𝜇𝐻2𝑆
𝜌𝑔 𝑌
𝜇𝑔 = 10−4 𝐾 × 𝑒𝑥𝑝 [𝑋 ( ) ]
62.4
(9.379 + 0.016𝑀𝑎)𝑇 1.5
𝐾=
209.2 + 19.26𝑀𝑎 + 𝑇
986.4 4.12
𝑋 = 3.448 + + 0.01009𝑀𝑎
𝑇
𝑌 = 2.4 − 0.2𝑋
∆𝜇𝐶7+ = 𝑦𝐶7+ (−3.2875 × 10−1 𝑙𝑜𝑔𝛾𝑔 + 1.2885 × 10−1 )
∆𝜇𝐻2𝑆 = 𝑦𝐻2𝑆 (−3.2268 × 10−3 𝑙𝑜𝑔𝛾𝑔 + 2.1479 × 10−3 )
{ ∆𝜇𝐶𝑂2 = 𝑦𝐶𝑂2 (6.4366 × 10−3 𝑙𝑜𝑔𝛾𝑔 + 6.7255 × 10−3 )

Elsharkawy, (2006) was also optimized pseudo-critical properties of Sutton, (2005)

shown in 4.13 based on 1200 data sets of gas-condensate mixtures compositions.
Pseudo-critical properties recommended by Elsharkawy, (2006) is as follow.

63 | P a g e
 𝑇𝑝𝑐 = 149.18 + 358.14𝛾𝑔 − 66.976𝛾𝑔 2
{ 4.13
𝑃𝑝𝑐 = 787.06 + 147.34𝛾𝑔 + 7.916𝛾𝑔 2

Elsharkawy, (2006) viscosity and critical properties methods are valid for the
temperature range of 94 to 327⁰F and pressure between 200 to 11,830psia. The
author reported 8.9% absolute average deviation for estimating gas viscosity using his
method.

In this study, the applicability of the above gas viscosity models for prediction of gas-
condensate fluid is assessed. For this purpose two sets of viscosity experimental data
by Al-Meshari et al., (2007) and Yang et al., (2007) have been selected. These studies
carried out in elevated pressure and temperature in laboratory condition similar to the
reservoir temperature and pressure condition. The fluids used in these experimental
studies are from gas-condensate reservoirs in Saudi Arabia and North Sea. The
collected fluids (gas and liquid) recombined in laboratory and gas viscosity
measurements were made (Al-Meshari et al., 2007; Yang et al., 2007). Statistical
accuracy of the existing literature models against experimental data examined using
percentage of Absolute Average Relative Deviation (AARD%), presented in 4.14.
Figure 4.1 illustrates the performance of each model in predicting gas phase viscosity.

The correlation proposed by Londono – Archer – Blasingame (LAB) provides best

performance in predicting experimental viscosity data with lowest AARD%, hence it
has been selected for further modification.
𝑁
1 𝜇𝑖 𝐸𝑥𝑝𝑒𝑟𝑖𝑚𝑒𝑛𝑡 − 𝜇𝑖 𝐶𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑
𝐴𝐴𝑅𝐷% = ∑| | × 100
𝑁𝑝 𝜇𝑖 𝐸𝑥𝑝𝑒𝑟𝑖𝑚𝑒𝑛𝑡 4.14
𝑖=1

64 | P a g e
Figure 4.1. The prediction performance of existing literature models in estimating gas phase
viscosity of gas-condensate reservoirs below the saturation pressure.

In an attempt to minimize the error between experimental data and the LAB correlation
a non-linear regression model on MATLAB was employed. Then the LAB model was
cast in the following form:

𝜌𝑔 𝑌
{𝜇𝑔 = 2.469 × 10−4 𝐾𝑒𝑥𝑝 [𝑋 ( ) ] 4.15
27.6718

The parameters of K,Y and X are same as the original Lee – Gonzalez – Eakin (LGE)
equation shown in 4.5. Half of the experimental data were used for optimizing LAB
model and the other half used to test the performance of the model. The performance
of proposed model in 4.15 plotted against the experimental data and shown in Figure
4.2. New optimized LAB model is predicting experimental data with 5.2% average
absolute relative deviation (AARD%). This model will be used for prediction of gas
phase viscosity of gas-condensate reservoirs, required for well deliverability modelling
in chapter 6 of this study.

65 | P a g e
Figure 4.2. Plot of calculated viscosity using new developed model against measured
viscosity data.

The viscosity of liquid phase (condensate) is another factor that need accurate
determination for reliable modelling of gas-condensate reservoirs undergoing
depletion. Next section focuses on studying condensate viscosity of gas-condensate
reservoir below the saturation pressure.

4.3 Assessment of condensate (oil) phase viscosity models

Accurate and reliable estimation of condensate viscosity is required for reliable

estimation of PVT properties using any available PVT modelling approaches (e.g.,
Modified black oil or equation of state). Less attention is given for accurate estimation
of condensate viscosity in existing literature. Yet condensate viscosity has the highest
prediction uncertainty (Whitson, 2006; Yang et al., 2007).
In this section, initially the validity of the existing literature models for prediction of
condensate viscosity is examined, and then several models will be proposed for
prediction of this crucial PVT property.
The existing literature models are divided into two categories as follow:
 Corresponding state principle (CSP) based model (also known as
compositional modelling)

66 | P a g e
  Gas-saturated oil viscosity based models (also known as live-oil viscosity
correlations)
The example of compositional models are LBC (1964) introduced in 4.2.1 and
Pedersen and Fredenslund, (1987). LBC is originally proposed for estimating viscosity
of hydrocarbon mixture (gas/liquid). For estimating condensate viscosity in this study
we employed LBC method due to its simplicity and its extensive use in numerical
reservoir simulation. The compositional model of LBC is used when compositional
data is available (see 4.2.1). However, the typical problem in engineering calculation
of gas-condensate reservoirs are unavailability of all required data (Gold, McCain and
Jennings, 1989; Whitson, 2006). In this case, to estimate condensate viscosity
engineers rely on other purely empirical correlations known as live-oil or gas-saturated
oil correlations.
These correlations are function of solution gas to oil ratio Rs, reservoir pressure,
reservoir temperature, fluid API gravity, gas specific gravity (𝛾𝐺𝑎𝑠 ) and dead oil
viscosity “𝜇𝑜𝑑 "(viscosity that contain no dissolved gas and relatively thick oil).
There are many live-oil (oil contains dissolved gas) correlations in literature that can
be used for prediction of condensate phase viscosity. We defined some standard
criteria to pick the most adequate models for our purpose. It is known that condensate
liquid is a light oil mixture with viscosity, ranging from 0.1 to 1cp, in the near wellbore
region (Whitson, Fevang and Yang, 1999; Al-Nasser and Al-Marhoun, 2012) and also
the API gravity between 40 – 60 ⁰API. The aforementioned two conditions were our
constraints in selecting existing live oil literature correlations that can be used for
prediction of condensate viscosity in this study. These correlations are useful when
compositional data of the gas mixture is not available. Five well-known live oil literature
models have been selected for estimating condensate viscosity and will be discussed
in this chapter.
Furthermore, in order to assess applicability of existing literature models for estimating
condensate viscosity the compositional model of LBC which introduced in 4.2.1 is
utilized.

Data acquisition

To assess the validity of existing literature models as well as developing AI models

comprehensive published experimental data sets was collected. The source of the

67 | P a g e
data is from (Fevang, 1995; Guo et al., 1997; Audonnet and Pádua, 2004; Gozalpour
et al., 2005; Al-Meshari et al., 2007; Yang et al., 2007; Thomas, Bennion and
Andersen, 2009; Kariznovi, Nourozieh and Abedi, 2012; Kashefi et al., 2013; Khorami
et al., 2017; Strand and Bjørkvik, 2019). In aforementioned studies, various methods
of rolling ball viscometer, electromagnetic pulse technology viscometer, capillary
viscometer and vibrating-wire sensor have been used for measurement of condensate
phase viscosity. Binary mixtures of methane and n-decane in different temperature
and pressure considered as gas-condensate fluid in most of previously mentioned
experimental studies. This data bank consists of 335 data sets, which covers API
gravity, gas specific gravity, reservoir fluid compositions, reservoir pressure, reservoir
temperature and initial gas to oil ratio.
Statistical distribution of the data is summarized in Table 4.1. Complete description of
the data bank is provided in Appendix A.
Table 4.1. Statistical information of the data bank.

Property Minimum Maximum Average

Pressure, (MPa) 0.25 75.84 25.25

Reservoir Temperature, 303 443.15 353.15

(°K)
Solution GOR, Rs, 41.7 13496 3628
(scf/STB)
API gravity 39.7 65 57.3
Gas Specific gravity 0.57 1.49 0.89
(𝛾𝑔 )
Condensate viscosity, 0.0404 0.982 0.232
(cp)

Beggs and Robinsons, (1975)

Almost all live oil (gas-saturated oil) viscosity correlations are in a similar form of
original Chew and Connally, (1959). Live oil viscosity correlated by Chew and
Connally, (1959) as a function of solution gas to oil ratio and dead oil viscosity at the
reservoir condition. They showed that in fixed solution gas to oil ratio (Rs) the relation
between the live oil and corresponding dead oil viscosity is a straight line with
logarithmic scale (Chew and Connally, 1959). They propose this relationship in the
following simple mathematical form.

68 | P a g e
 𝜇𝑜𝑏 = 𝐴(𝜇𝑜𝑑 )𝐵 4.16

Where A and B can be estimated from the original chart of Beal, (1946). Beggs and
Robinson, (1975) developed a functional relation for A and B based on 2073 live oil
viscosity measurements. The viscosity values measured with -1.83% uncertainty.
10.715
𝐴=
(𝑅𝑠 + 100)0.515
4.17
5.44
𝐵=
{ (𝑅𝑠 + 150)0.338
Using live oil viscosity relation shown in 4.16 along with Beggs and Robinsons
correlation in 4.17 is limited to Rs within the range of 20 to 2070 scf/STB, oil gravity of
16 to 58°API, pressure range of 0 to 5250 and temperature of 70 to 295°F (Beggs and
Robinson, 1975; Aily et al., 2019).

De Ghetto et al., (1994)

De Ghetto et al., (1994) used similar form of Chew and Connally, (1959) live oil
viscosity correlation and optimized the parameters of the A and B functions in 4.17,
proposed originally by Beggs and Robinsons, (1975). They used 195 oil samples
collected worldwide to optimize the values of A and B and indicate that the correlation
is applicable to viscosity of live oil with API > 31.1. De Ghetto et al., (1994) present the
following relation for A and B as a function of Rs.
25.192
𝐴=
(𝑅𝑠 + 100)0.6487
4.18
2.7516
𝐵=
{ (𝑅𝑠 + 150)0.2135
They showed that using above A and B values, the live oil viscosity can be estimated
with less than 10% error within the temperature range of 80.6 to 334.6 °F, Rs of 8.61
to 3299scf/STB and 0.07< 𝜇𝑜𝑏 < 295.9cp (De Ghetto et al., 1994).

Elsharkawy and Alikhan (1999)

Elsharkawy and Alikhan, (1999) optimized the value of A and B to be used for
estimation of live oil viscosity in 4.16 not similar to Beggs and Robinsons, (1975) and
De Ghetto et al., (1994). They used 254 datasets from Middle Eastern oil samples and

69 | P a g e
proposed two new functions for A and B. They concluded that estimating A and B
values using their proposed equation shown in 4.19, would predicts live oil viscosity
with 18.6% average absolute relative error. Their correlation is valid for the range of
10 to 3600 for (Rs) and 0.05 to 20.89cp (𝜇𝑜𝑏 ) (Elsharkawy and Alikhan, 1999).

𝐴 = 1241.932(𝑅𝑠 + 641.026)−1.12410
{ 4.19
𝐵 = 1768.84(𝑅𝑠 + 1180.335)−1.06622

Bergman and Sutton, (2007)

Bergman and Sutton, (2007) used 2048 live oil measured viscosity data sets collected
from worldwide and proposed equation 4.20 to estimate values of A and B. They
showed that live oil viscosity can be estimated using their proposed equation in 4.20
incorporated with 4.16 with absolute average error of 9%. Their method best fit the
value of live oil viscosity within the range of 5 to 2890scf/STB solution gas to oil ratio
(Rs) and range of 0.125 to 123cp live oil viscosity (𝜇𝑜𝑏 ).

𝐴 = 𝑒 [4.768−0.8359 ln(𝑅𝑠 +300)]

{ 133.5 4.20
𝐵 = 0.555 +
𝑅𝑠 + 300

Kartoatmodjo and Schmidt, (1991)

Kartoatmodjo and Schmidt, (1991) proposed a live oil viscosity correlation as a

function of solution gas to oil ratio (Rs) and dead oil viscosity. Their correlation is
slightly different from original live oil viscosity of Chew and Connally, (1959) shown in
4.16 as shown in 4.21. They used 5321 gas-saturated-oil samples collected globally
to develop their live oil correlation. Their correlation can be applied to crude oils in the
range of 14.4 to 59°API gravity, temperature range of 80 to 320°F, Rs range of 0 to
2890scf/STB and live oil viscosity range of 0.098 to 586cp (Kartoatmodjo and Schmidt,
1991).

𝜇𝑜𝑏 = −0.06821 + 0.9824𝑋1 + 4.034 × 10−4 𝑋2 2

−4
{ 𝑋1 = 0.43 + 0.5165 × 10(−8.1×10 𝑅𝑠 ) 4.21
−4
𝑋2 = [0.2001 + 0.8428 × 10(−8.1×10 𝑅𝑠 ) ]𝜇𝑜𝑑 𝑋1

All live oil viscosity methods in literature and seen above are function of solution gas
to oil ratio (Rs) and dead oil viscosity (𝜇𝑜𝑑 ). Accurate estimation of Rs and 𝜇𝑜𝑑 is

70 | P a g e
required for reliable estimation of live oil viscosity. Measured values of Rs and 𝜇𝑜𝑑 are
preferable for calculation; however, in the absence of experimental values,
correlational estimation of viscosities are very popular (Chew and Connally, 1959;
Beggs and Robinson, 1975). Rs is correlated with gas gravity, saturation pressure,
stock tank oil gravity and temperature (Beal, 1946; Standing, 1947). Further discussion
about important determination of Rs is given in 2.3.5. Among the most well-known
methods for calculating Rs, we employed equation 2.50 proposed by Standing, (1947).
This equation is still widely used in industry and recommended by Whitson et al.,
(2000) and Yang et al., (2007) as one of the most accurate correlation in literature.
Considerable number of dead oil viscosity correlations are also exist in literature (Beal,
1946; Beggs and Robinson, 1975; Bergman and Sutton, 2007; Egbogah and Jack,
1990; Elsharkawy and Alikhan, 1999; Kartoatmodjo and Schmidt, 1994; Labedi, 1992;
Naseri et al., 2005; Standing, 1981). Bergman and Sutton, (2007) reviewed the
performance of 23 dead oil viscosity correlations with an extensive data bank. They
found their proposed dead oil correlations is superior to all other techniques in
literature. Whitson and Brulé, (2000, p. 36) also recommend Bergman and Sutton,
(2007) for estimating dead oil viscosity for engineering calculation of gas-condensate
reservoirs.
Therefore in this study Bergman and Sutton, (2007) is employed for estimation of dead
oil viscosity. Bergman and Sutton, (2007) correlation is as follow.

ln ln(𝜇𝑜𝑑 + 1) = 𝐴0 + 𝐴1 ln(𝑇 + 310)

{𝐴0 = 22.33 − 0.194𝛾𝐴𝑃𝐼 + 0.00033𝛾𝐴𝑃𝐼 2 4.22
𝐴1 = −3.20 + 0.0185𝛾𝐴𝑃𝐼

Where 𝛾𝐴𝑃𝐼 is oil (condensate) stock tank gravity, T is temperature in Fahrenheit. All
discussed live oil correlations as well as compositional model of LBC is applied for
estimating experimental condensate viscosity. In following section the results of
existing literature models are discussed.

Results of existing models

Three statistical metrics of coefficient of determination (R2), root mean square error
(RMSE) and absolute average relative deviation percentage (AARD%), shown in

71 | P a g e
equations 4.23 to 4.25 respectively, used for assessing the performance of existing
literature models in prediction of condensate viscosity.
To visualise the accuracy of each model the graphical error analysis in a form of cross
plot was employed. The results of statistical and graphical error analysis show that the
performance of existing literature models in predicting condensate viscosity is very
poor. Hence, to improve the accuracy of each models, non-linear regression was
applied to refine the models. Table 1 depicts the original and tuned form of the utilized
correlations for estimating condensate viscosity. Graphical error analysis of the
original and refined literature correlations in predicting condensate viscosity are
presented in Figure. 4.3a-f. The slope line of 45° in aforementioned figures
representing zero error line in matching between measured and calculated values.
The reason for poor performance of LBC viscosity models might be due to sensitivity
of LBC method to mixture density and critical volumes of the heavy components
(Whitson and Brulé, 2000; Yang et al., 2007). Live oil viscosity correlations are function
of deal oil viscosity and solution gas to oil ratio. It should be noted that dead oil
viscosity is one of the most difficult properties to be estimated by correlations due to
its dependency to paraffin, aromatic, naphthalene and asphaltene content (Whitson
and Brulé, 2000; Hemmati-Sarapardeh et al., 2014). This can be a reason for poor
performance of live oil viscosity correlations for estimating condensate viscosity.

The procedure for tuning of the LBC, (1964) correlation recommended by Yang et al.,
(2007) followed in this study, where coefficients of A0 – 4 of the LBC correlation are
optimized to match the experimental data. The new coefficients of LBC correlation are
shown in Table 4.2. Coefficients of live oil viscosity models were also optimized using
regression to match the experimental condensate viscosity data. The original and new
tuned correlations of live oil viscosity are shown in Table 4.2.

∑𝑁 ̂
𝑖 (𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐶𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑(𝑖) − 𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐴𝑐𝑡𝑢𝑎𝑙(𝑖) )
2
𝑅2 = 1 − 2 4.23
∑𝑁 ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
𝑖 (𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐶𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑(𝑖) − 𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐴𝑐𝑡𝑢𝑎𝑙(𝑖) )

0.5
∑𝑁
𝑖 (𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐶𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑(𝑖) − 𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐴𝑐𝑡𝑢𝑎𝑙(𝑖) )
2
𝑅𝑀𝑆𝐸 = ( ) 4.24
𝑁

72 | P a g e
 𝑁
100 |(𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐶𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑(𝑖) − 𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐴𝑐𝑡𝑢𝑎𝑙(𝑖) )|
𝐴𝐴𝑅𝐷% = ∑ 4.25
𝑁 𝑐𝑜𝑛𝑑 𝑣𝑖𝑠𝐴𝑐𝑡𝑢𝑎𝑙(𝑖)
𝑖

(a) (b)

(c) (d)

(e) (f)
Figure 4.3. Cross plot of the experimental condensate viscosity versus predicted condensate
viscosity using employed literature correlations and refined results.

73 | P a g e
 Table 4.2. The original and tuned form of the utilized literature correlations.

Author Original Correlation Tuned Correlation

1 1
[(𝜇 − 𝜇∗ )𝜁 + 10−4 ]4 [(𝜇 − 𝜇∗ )𝜁 + 10−4 ]4
= 0.1023
+ 0.023364𝜌𝑟
= 0.11364
+ 0.058523𝜌𝑟 2 − + 0.02173𝜌𝑟
LBC, (1964)
0.040758𝜌𝑟 3 + 0.0093324𝜌𝑟 4 − 0.20666𝜌𝑟 2
+ 0.06283𝜌𝑟 3
+ 0.17139𝜌𝑟 4

𝜇𝑜𝑏 𝜇𝑐
Beggs & 5.44 4.056
10.715 (𝑅 +150) 0.338 17.99 (𝑅 +150) 0.338
= (𝜇 ) 𝑠 = (𝜇𝑜𝑑 ) 𝑠
Robinson, (𝑅𝑠 + 100)0.515 𝑜𝑑 (𝑅𝑠 + 100) 0.515

(1975)

𝜇𝑜𝑏 = −0.06821 + 0.9824𝑋1 + 4.034 𝜇𝑐 = −0.30612 + 1.174𝑋1 + 4.034

× 10−4 𝑋2 2 × 10−4 𝑋2 2

Kartoatmodj 𝑋1 = 0.43 + 0.5165 × 10(−8.1×10

−4 𝑅 )
𝑠 𝑋1 = 0.43 + 0.5165 × 10(−8.1×10
−4 𝑅 )
𝑠

o & Schmidt,
𝑋2 𝑋2
(1991)
= [0.2001 + 0.8428 = [0.2001 + 0.8428
−4 −4
× 10(−8.1×10 𝑅𝑠 ) ]𝜇𝑜𝑑 𝑋1 × 10(−8.1×10 𝑅𝑠 ) ]𝜇𝑜𝑑 𝑋1

𝜇𝑐
De Ghetto et 𝜇𝑜𝑏 2.7516 62.96 2.1334
25.192 = (𝜇 ) (𝑅𝑠 +150)0.2135
al, (1994) = (𝜇 )(𝑅𝑠 +150)0.2135 (𝑅𝑠 + 100)0.6487 𝑜𝑑
(𝑅𝑠 + 100)0.6487 𝑜𝑑
For (°API >
31.1)

𝜇𝑜𝑏 = 𝐴(𝜇𝑜𝑑 )𝐵 𝜇𝑐 = 𝐴(𝜇𝑜𝑑 )𝐵

Elsharkawy 𝐴 = 1241.932(𝑅𝑠 + 641.026)−1.12410 𝐴 = 3978.167(𝑅𝑠 + 641.026)−1.12410

& Alikhan,
𝐵 = 1768.84(𝑅𝑠 + 1180.335)−1.06622 𝐵 = 1361.93(𝑅𝑠 + 1180.335)−1.06622
(1999)
133.5 133.5
0.555+ 0.555+
Bergman & 𝜇𝑜𝑏 = 𝐴(𝜇𝑜𝑑 ) 𝑅𝑠 +300 𝜇𝑐 = 𝐴(𝜇𝑜𝑑 ) 𝑅𝑠 +300

Sutton ,
𝐴 = 𝑒 [4.768−0.8359 ln(𝑅𝑠 +300)] 𝐴 = 𝑒 [4.6792−0.7772 ln(𝑅𝑠 +300)]
(2000)

74 | P a g e
The accuracy of condensate viscosity using existing literature models depends on
accurate estimation of many parameters (e.g., solution gas to oil ratio, reduced
density, dead oil viscosity, well stream gravity for condensate). Although the attempt
to improve the accuracy of the viscosity models were made by optimizing associated
constants od each model, yet the accuracy is still less than satisfactory. Hence, in this
study we took different approach known as Machine Learning (ML) for accurate
estimation of condensate viscosity without going through tedious computational
procedure of existing literature models. These methods are also known as Artificial
Intelligent (AI) techniques. In next section, a brief overview of AI techniques for
application of reservoir modelling is discussed.

4.4 Machine Learning (ML) Approach

In order to use computers for solving and processing applications, mathematical

models are required (Forsyth, 1989). The fundamental approach to acquiring these
models in natural sciences and engineering disciplines is by applying fundamental
laws and theories in physics, chemistry and other related science to determine usually
the relationships between variables (Forsyth, 1989). The input variabes are called the
independent variables, while the output varaibles are the dependant variables. For
some tasks, however the relationship between dependant and independent variables
are not defined and there are no governing mathematical equations between them.
Therefore the transformation between the input and the output can not be done directly
(Flach, 2012). For instance physics of gas-condenste fluid below the saturation
pressure is becoming more thermodanamically complex, which makes it very difficult
to establish a robust relationship between the input and the output varaibles for
instance in the case of condensate viscosity.
Because of that lack of knoweledge or complexities of the problems, alternative
methods to build models have been introduced by scientists. These methods rely on
using sets of measured input and output data that have been obtained during
measurement process or from past saved data (Flach, 2012). The computer which is
being trained uses certain algorithms and statistical analysis to build a model which
can; for instance; predict the output values (condensste viscosity). This type of process
is known as machine learning (ML).

75 | P a g e
Arthur Samuel has used the term machine learning for first time in 1959. He defined
ML as “a field of study that gives computers the ability to learn without being explicitly
programed” (Samuel, 1959). Figure 4.4 shows the main structure of modelling by
machine learning.

Figure 4.4 Machine Learning Block Diagram.

Based on the method selected to feed the input and output data for learning process,
there are two main class of machine learnings known as supervised and unsupervised
learning (Louridas and Ebert, 2016). In supervised machine learning, the nature and
the label of the output are known. For example in this chapter using experimental
database, we know the output is condensate viscosity. In the other word the data that
provided for training of the network include sets of input and output parameters.
In unsupervised learning, the training set is containing the data but no solutions; and
the computer must find the solution on its own by relating the pattern on the data set
(Louridas and Ebert, 2016). Regardless of classes of ML (supervised or unsupervised)
the process is that the computer first learns to perform a task by studying a training
set. Then the computer performs the same task with the data it hasnot encountered
before.
The recent development and success of machine learning (ML) techniques in solving
complex engineering problems have drawn attention to their various application in
petroleum industry. Shell recently announced using new artificial intelligence (AI)
platform to drive its effort in predictive maintenance and spread AI-powered
applications across the company. British Petroleum (BP) has invested heavily on using
AI techniques aims at accelerating project lifecycle from exploration through reservoir
modelling (Norton, 2018; BP, 2019). In modelling aspect, which is the scope of this
research, availability of the data in recent years motivated to utilize the data driven
techniques for reliable and accurate forecast of outputs (e.g., flow rate, Z factor and

76 | P a g e
GOR). Some of these example can be found in the literature include (Ahmadi et al.,
2014; Ghiasi et al., 2014; Hemmati-Sarapardeh et al., 2014; Kamari et al., 2013;
Naderi and Khamehchi, 2019; Shokir, 2008).
For gas-condensate reservoirs Ahmadi and Ebadi (2014), Elsharkawy and Foda
(1998), Jalali et al. (2007) and Nowroozi et al. (2009) were using machine learning
(ML) approaches for predicting dew point pressure. Zendehboudi et al. (2012) used
ML approach to model condensate-to-gas ratio (CGR). Recently Ghiasi et al. (2014)
employed least square support vector machine (LSSVM) to predict compressibility
factor of gas-condensate reservoirs. Although the aforementioned studies utilized ML
techniques for modelling gas-condensate reservoirs such as dew point pressure, CGR
and compressibility factor, however there is still significant potential to use various
approaches of ML for modelling this type of reservoir. There is not any models in
current literature for prediction of condensate viscosity using any ML methods. Hence,
the aim is to use ML based techniques to develop a robust and accurate models for
prediction of condensate viscosity in gas-condensate reservoirs below the saturating
pressure. For this purpose, three ML based techniques including Support Vector
Machine (SVM), Artificial Neural Network (ANN) and Fuzzy logic methods employed
and will be discussed in details in the following sections.

Support Vector Machine (SVM)

The support vector machine (SVM) has been identified as an efficient and powerful
strategy developed from the machine-learning community (Cortes and Vapnik, 1995;
Suykens et al., 2002). SVM is a tool for a set of related supervised learning methods
that analyse data and recognize pattern using regression analysis and it is identified
as a non-probabilistic binary linear classifier.
SVM method has many advantages over other machine learning techniques as
follows: they are more likely to converge to the global optima, prior determination of
the network is not required in this model and can be automatically determined as the
training ends (Suykens et al., 2002; Eslamimanesh et al., 2012). Furthermore, the
number of hidden layers and hidden nodes should not be determined and this
algorithm has fewer adjustable parameters compared to ANN network (Ahmadi and
Ebadi, 2014; Eslamimanesh et al., 2012; Hemmati-Sarapardeh et al., 2014; Kamari et

77 | P a g e
al., 2013). SVM is mapping the predictor’s data using kernel function and initial formula
indicates that any function can be regressed as follow:

𝑓𝑥 = 𝑤 𝑇 𝜑(𝑥) + 𝑏 4.26

Where 𝑤 𝑇 and 𝜑(𝑥) and b are transposed output layer vector, kernel function and
bias, respectively. The input of the model x is of a dimension of 𝑁 × 𝑛 in which N and
n denotes to number of data points and number of inputs parameters, respectively. In
order to calculate w in above function, Cortes and Vapnik, (1995) proposed
minimization of the cost function, defined in 4.26 subject to some constraint shown in
4.27 (Ahmadi and Ebadi, 2014; Eslamimanesh et al., 2012, 2011; Hemmati-
Sarapardeh et al., 2014; Suykens et al., 2002).

𝑁
1
𝐽 = 𝑤 𝑇 + 𝐶 ∑(𝜉𝑘 − 𝜉 ∗ 𝑘 ) 4.27
2
𝑘=1

Where J is the cost function, c is the tuning parameter of the SVM and 𝜉𝑘 and 𝜉 ∗ 𝑘 are
the slack variables.

𝑦𝑘 − 𝑤 𝑇 𝜑(𝑥𝑘 ) − 𝑏 ≤ 𝜀 + 𝜉𝑘 , 𝑘 = 1.2, … , 𝑁
{𝑤 𝑇 𝜑(𝑥𝑘 ) + 𝑏 − 𝑦𝑘 ≤ 𝜀 + 𝜉 ∗ 𝑘 , 𝑘 = 1,2, … , 𝑁 } 4.28
𝜉𝑘 , 𝜉 ∗ 𝑘 ≤ 0 𝑘 = 1,2, … , 𝑁

Where xk is input of kth data point and yk is output of kth data points. ε is the fixed
precision of the function approximation. Choosing a small ε leads to developing an
accurate model, however some data points may be outside of ε precision, and may be
result in infeasible solution. Therefore, slack parameters can be used to define
allowable margin error. C in equation 4.26 is one of the tuning parameters, which
determine the amount of deviation of the model from the desired ε. The optimum value
of cost function in SVM can be reached by minimization process, where Lagrangian
approach can be used as follow.

78 | P a g e
 𝑁 𝑁
1
𝐿(𝑎, 𝑎∗ ) = − ∑ ( 𝑎𝑘 − 𝑎∗ 𝑘 )(𝑎𝑙 − 𝑎∗ 𝑙 )𝐾(𝑥𝑘 , 𝑥𝑙 ) − 𝜀 ∑( 𝑎𝑘 − 𝑎 ∗ 𝑘 ) …
2
𝑘,𝑙=1 𝑘=1
𝑁 𝑁

… + ∑ 𝑦𝑘 (𝑎𝑘 − 𝑎)∗ 𝑘 ) ∑(𝑎𝑘 − 𝑎∗ 𝑘 ) = 0, 𝑎𝑘 , 𝑎∗ 𝑘 ∈ [0, 𝑐] 4.29

𝑘=1 𝑘=1
{ 𝐾(𝑥𝑘 , 𝑥𝑙 ) = 𝜑(𝑥𝑘 )𝑇 𝜑(𝑥𝑙 ), 𝑘 = 1,2, … , 𝑁 }

Where 𝑎𝑘 and 𝑎∗ 𝑘 are Lagrangian multipliers, and the final form of least square SVM
function is obtained as follow.

𝑓(𝑥) = ∑ ( 𝑎𝑘 − 𝑎 ∗ 𝑘 )𝐾(𝑥, 𝑥𝑘 ) + 𝑏 4.30

𝑘,𝑙=1

To solve the above equation and find 𝑎𝑘 , 𝑎∗ 𝑘 and b, quadratic programming problem
can be used which is very difficult to implement. Thus the least square modification
proposed by Suykens and Vandewalle, (1999) to facilitate original SVM algorithm.
They reformulated the SVM cost function J as follow.

𝑁
1 1
𝐽 = 𝑤 𝑇 𝑤 + 𝛾 ∑ 𝑒 2𝑘 4.31
2 2
𝑘=1

𝑌𝑘 = [𝑤 𝑇 𝜑(𝑥𝑘 ) + 𝑏 + 𝑒𝑘 ], 𝑘 = 1, … , 𝑁.
4.32

Where 𝛾 is a tuning parameter, ek is error variable, Yk is constraint of the cost function

(J). Equation 4.31 is new form of SVM equation known as least square SVM (LSSVM).
The Lagrangian solution for this equation is calculated as follow.
𝑁 𝑁
1 1
𝐿(𝑤, 𝑏, 𝑒; 𝑎) = 𝑤 𝑇 𝑤 + 𝛾 ∑ 𝑒 2 𝑘 − ∑ 𝑎𝑘 (𝑤 𝑇 𝜑(𝑥𝑘 ) + 𝑏 + 𝑒𝑘 − 𝑦𝑘 )
2 2 4.33
𝑘=1 𝑘=1

Where 𝑎𝑘 denotes to the Lagrange multiplier, that may be either positive or negative,
since LSSVM has equality restrictions. The above equation can be solved by equate
the derivate of each parameter (w, b, e, and a) to 0 according to Karush Kuhn-Tucher’s
(KKT) conditions (Fletcher, 1987). These conditions are demonstrated in equation
4.34 for cost function to achieve optimum goal (Suykens and Vandewalle, 1999;

79 | P a g e
Pelckmans et al., 2002; Suykens et al., 2002). Hence, the following equations are
obtained for optimal solution of the cost function.
𝑁
𝜕𝐿
= 0 ⟹ 𝑤 = ∑ 𝑎𝑘 𝜑(𝑥𝑘 )
𝜕𝑤
𝑘=1
𝑁
𝜕𝐿
= 0 ⟹ ∑ 𝑎𝑘 = 0
𝜕𝑏 4.34
𝑘=1
𝜕𝐿
= 0 ⟹ 𝑎𝑘 = 𝛾𝑒𝑘 , 𝑘 = 1,2, … , 𝑁
𝜕𝑒
𝜕𝐿 𝑇 )
{𝜕𝑎 = 0 ⟹ 𝑤 𝜑(𝑥𝑘 + 𝑏 + 𝑒𝑘 − 𝑦𝑘 = 0 𝑘 = 1,2, … , 𝑁}

To turn the above equation to a linear form the following expression can be
demonstrated.

0 −1𝑇
[ 1 ] [ 𝑏 ] = [0 ] 4.35
1 𝛺 + 𝐼𝑁 𝛼 𝑦
𝛾

while 𝑦 = (𝑦1 , … , 𝑦𝑛 )𝑇 , 1𝑛 = (1, … ,1)𝑇 , 𝑎 = (𝑎1 , … , 𝑎𝑛 )𝑇 and 𝛺𝑖𝑙 = 𝜑(𝑥𝑖)𝑇 𝜑(𝑥𝑙) for 𝑖, 𝑙 =
1, … , 𝑛. By Mercer’s theorem, the resulting LSSVM model for function approximation
turns to the following equation (Cortes and Vapnik, 1995; Suykens and Vandewalle,
1999).

𝑓(𝑥) = ∑ 𝛼𝑘 𝐾(𝑥, 𝑥𝑘 ) + 𝑏 4.36

𝑘,𝑙=1

Where 𝛼 and b are the routes to equation 4.36 and can be determined as bellow:
1 −1
1𝑇𝑛 (𝛺 + 𝛾 𝐼𝑛 ) 𝑦
𝑏=
1 −1
1𝑇𝑛 (𝛺 + 𝛾 𝐼𝑛 ) 𝐼𝑛 4.37
1 −1
𝛼 = (𝛺 + 𝐼𝑛 ) (𝑦 − 1𝑛 𝑏)
{ 𝛾

Knowing a and b from 4.37, then 𝑓(𝑥) in 4.36 may be executed as choice of nonlinear
regression and utilize the Kernel function. (Cortes and Vapnik, 1995; Suykens and
Vandewalle, 1999; Suykens et al., 2002; Eslamimanesh et al., 2012; Fazeli et al.,

80 | P a g e
2013). In equation 4.35, 𝐾(𝑥, 𝑥𝑘) stands for dependency of Kernel function to the inner
values of two vectors x and xi in the feasible area referred to the inner products of
vectors ɸ(𝑥)𝑇 ɸ(𝑥𝑖 ) as shown in below equation:

𝐾(𝑥, 𝑥𝑘) = ɸ(𝑥)𝑇 ɸ(𝑥𝑘 )

4.38

Where the radial basis function (RBF) Kernel may executed as following formulation
(Cortes and Vapnik, 1995; Eslamimanesh et al., 2012; Pelckmans et al., 2002;
Suykens et al., 2002).

‖𝑥𝑘 − 𝑥‖2
𝐾(𝑥, 𝑥𝑘 ) = exp (− )
𝜎2 4.39

Where 𝛾 and 𝜎 2 are tuning parameters of LSSVM, that they can be indicated by
performing of any optimization algorithm such as Simulated Annealing algorithm or
Genetic Algorithm (GA). In order to achieve the optimum values of aforementioned
parameters in LSSVM method, our objective function is root mean square error
(RMSE) between the experimental values and the output of the LSSVM.
Our task in this chapter was to develop a smart model for prediction of condensate
viscosity using LSSVM, explained in details earlier. To implement the LSSVM for
prediction of condensate viscosity database, initially the data sets divided into three
subsets of “Training”, “Optimization” and “Testing”. Training set is used for generating
the model structure, optimization is used for minimization of the error in trained model
and test data is used to investigate the prediction capability of the developed model.
Optimization sets also known as the validation set. The database was randomly split
into three sub data sets of 80% training, 10% testing and 10% validation. We also
ensure the homogeneous accumulations distribution of the data for the domain of the
three data sets (Eslamimanesh et al., 2011; Gharagheizi et al., 2014).The allocation
percentage of the data is selected according to the recommendations by Ahmadi and
Ebadi, (2014) and Eslamimanesh et al., (2012) to enhance the performance of the
developed model.
During the training of the model cross validation has been performed where, the
training data sets into several folds and accuracy of each fold was checked. The
LSSVM implemented in MATLAB and powerful Simulated Annealing (SA) algorithm

81 | P a g e
was utilized to minimize the error between the experimental viscosity and predicted
values by the model. The objective function of minimization procedure was root mean
square error as defined as follow.

2
∑𝑛𝑖=1(𝑉𝑖𝑠𝑒𝑠𝑡𝑖 − 𝑉𝑖𝑠𝑒𝑥𝑝𝑖 ) 4.40
𝑅𝑀𝑆𝐸 = √
𝑛𝑠

Where 𝑉𝑖𝑠 represents condensate viscosity, subscripts est and exp represent the
predicted and actual value, ns is number of data points from the initial assigned
population of 144 data sets. The optimized values of 𝛾 and 𝜎 2 using SA optimization
method for predicting the condensate liquid viscosity presented in the following table.

Table 4.3. The optimum values of the LSSVM parameters.

LSSVM model Input parameters Model parameters
𝛾 𝜎2
Condensate phase Reservoir pressure,
viscosity Temperature, API, gas SG, Rs 5625.256 23.65

The algorithm reduces the error between the experimental values of viscosity, and the
model to reach the optimum solution. The input variables for this model are as
pressure, temperature, API gravity, gas specific gravity and solution gas to oil ratio
“Rs”.
The prediction capability of the trained model was tested for new data sets using the
generated LSSVM code. The cross plot of in Figure. 4.5A-B illustrates the performance
of developed LSSVM model in training stage and testing stage, respectively. The
majority (73%) of the data points in this study are within lower viscosity range of 0 –
0.4cp. Therefore, the performance of the LSSVM in testing new data is toward lower
viscosity region, which is more realistic characterisation of condensate viscosity below
the dew point near wellbore region (Whitson, Fevang and Yang, 1999; Yang et al.,
2007). The viscosity of condensate liquid in near wellbore region (Region 1), where
condensate liquid is mobile is very low. This is due to the existence of more lighter C7+
fractions in mobile condensate liquid composition in aforementioned region (Fevang,
1995, p. 44).

82 | P a g e
 A B

Figure 4.5. Performance of LSSVM in predicting condensate (oil) viscosity in training and
testing stage.

Graphical representation of the error analysis of the LSSVM in predicting the

condensate viscosity depicted in Figure. 4.6. Ability of the trained LSSVM in predicting
new data sets are also analysed by presenting graph of standard deviation error in
Figure. 4.6A and standard error from the mean in Figure. 4.6B.

A B
Figure 4.6. Graphical error analysis of LSSVM performance for predicting condensate (oil)
viscosity.

Three statistical error analysis including coefficient of determination (R2), root mean
square error (RMSE) and average absolute relative deviation percentage (AARD%)
shown in equations 4.22 to 4.24 respectively, were adapted for estimating the
performance of the LSSVM. Table 4.4 shows the result of this analysis in training,
validation and testing stage of the LSSVM.

83 | P a g e
 Table 4.4. Statistical error performance of the LSSVM.

Stage of the process 𝑅2 RMSE AARD%

Training set 0.9139 0.10845 13.96
Validation set 0.87256 0.111121 14.12
Testing set 0.7723 0.121037 14.25

The results obtained from LSSVM approach indicate a promising outcome for reliable
and more accurate prediction of condensate viscosity. The results indicate that
LSSVM is performing better than tuned literature correlations. However, the error is
still high, approximately about 23% in testing stage, where the capability of the model
assessed using new data sets. Therefore, to certify the effectiveness and accuracy of
the suggested LSSVM model for estimation of condensate viscosity among smart
approaches in another attempt an Artificial Neural Network (ANN) was developed,
which is presented in following section.

Artificial Neural Network (ANN)

Artificial Neural Network (ANN) is a type of machine learning and artificial intelligence
model that mimics human central nervous system, in particular the brain in human
body (Bell, 2014). An ANN network consists of several organized layers containing
hundreds or even more single units and artificial neurons that are connected through
weight functions (Agatonovic-Kustrin and Beresford, 2000; Giri Nandagopal and
Selvaraju, 2016). There are many types of neural networks but the main distinguish
features between them are transfer functions of their neurons, learning rules and
connection formula.
In ANN network complex computation system is performed for predicting the output
responses. ANNs are inspired by biological networks, performing in a massive parallel
connection between nonlinear, parametrized, and bounded functions called neurons
(Cios and Shields, 1997; Mesbah, Soroush and Rostampour Kakroudi, 2017). Such a
network is a massively parallel-distributed processor that has a natural tendency for
storing experimental knowledge and making it available for future use. In ANN system
knowledge is acquired by the network through a learning process and synaptic weights
will store this knowledge (Haykin, 1994). Hence, mathematical interpretation of the
problem does not required. Neurons in such system coordinate their work, and they

84 | P a g e
transfer information by using synapses “electromagnetic communications” (Ghaffari et
al., 2006a). Through a set of known input and output data the network will be trained.
Through a learning process the network monitors the error between the predicted and
desired outputs and continue to adjust the weights until the optimization criteria are
reached. This process is usually carried out in two following stages:
 first the input variables are linearly combined
 the result is used as an argument of non-linear activation function (a)
The activation function must be non-decreasing and differentiable function; the most
common choices are either the identity function (𝑦 = 𝑥), or bounded sigmoid (s-
shaped) function, as the logistic [𝑦 = 1/(1 + 𝑒 −𝑥 )] (Haykin, 1994; Hippert, Pedreira
and Souza, 2001; Ghaffari et al., 2006b; Eslamimanesh et al., 2011).
The neurons are organized in a way that define the network architecture. In this
chapter we applied multilayer perception (MLP) type, in which the neurons are
organized in continuous layers.
In MLP architecture, the neurons in each layer may share same inputs, but they are
not connected to each other. As shown in Figure 4.7, typical neural networks consist
of hidden layers, output layer, inputs and bias units. Number of hidden layers and
number of neuron of each layers can be arbitrary (Khosrojerdi et al., 2016). However,
increasing number of neurons may cause overfitting while decreasing their numbers
may result on poor performance of the network. The main advantage of ANN is ability
to process large amount of data sets (Ghaffari et al., 2006; Khosrojerdi et al., 2016;
Mesbah et al., 2017; Hippert et al., 2001).

Figure 4.7. Schematic illustration of the ANN structure and computational steps to measure
output.

85 | P a g e
The mathematical definition of the typical MLP neural network shown in Figure 4.7 is
presented in the following section where superscripts are values associated with each
layer.
In this mathematical definition 𝑎𝑖 (𝑗) = activation" 𝑜𝑓 𝑢𝑛𝑖𝑡 𝑖 𝑖𝑛 𝑙𝑎𝑦𝑒𝑟 𝑗 and
Ɵ(𝑗) = matrix of weights controlling function mapping from layer j to layer j+1. In order
to calculate each activation function (𝑎) a sigmoid function (g) is multiplied by sum of
linear combination of inputs for each neuron; these inputs in Figure 4.7 include
(𝑥1 , 𝑥2 , 𝑥3 𝑎𝑛𝑑 𝑏𝑖𝑎𝑠 𝑢𝑛𝑖𝑡 𝑥0 ) in hidden layer. The activation function (a) then can be
calculated from the following equation.
(2) (1) (1) (1) (1)
𝑎1 = 𝑔(Ɵ10 𝑥0 + Ɵ11 𝑥1 + Ɵ12 𝑥2 + Ɵ13 𝑥3 )
(2) (1) (1) (1) (1)
𝑎2 = 𝑔(Ɵ20 𝑥0 + Ɵ21 𝑥1 + Ɵ22 𝑥2 + Ɵ23 𝑥3 ) 4.41
(2) (1) (1) (1) (1)
{𝑎3 = 𝑔(Ɵ30 𝑥0 + Ɵ31 𝑥1 + Ɵ32 𝑥2 + Ɵ33 𝑥3 )

Then the output function ℎ𝜃 (𝑥) in Figure 4.7, is sum of each neuron’s weight multiplied
by activation function of same neuron in layer 2 defined by equation 4.42.The neurons
of the output layer have linear transfer functions.

(3) (2) (2) (2) (2) (2) (2) (2) (2)

ℎƟ (𝑥) = 𝑎1 = 𝑔( Ɵ10 𝑎0 + Ɵ11 𝑎1 + Ɵ12 𝑎2 + Ɵ13 𝑎3 ) 4.42

Where g is a sigmoid function can be evaluated from the following equation.

1
𝑔(𝑧) = 4.43
(1 + 𝑒 −𝑧 )

In order for neural network to perform, the activation function of each neuron should
be vectorised in a matrix form of z as follow.
(1) (1) (1) (1) (2)
Ɵ10 𝑥0 + Ɵ11 𝑥1 + Ɵ12 𝑥2 + Ɵ13 𝑥3 = 𝑍1
{Ɵ(1) (1) (1) (1) (2)
20 𝑥0 + Ɵ21 𝑥1 + Ɵ22 𝑥2 + Ɵ23 𝑥3 = 𝑍2
} 4.44
(1) (1) (1) (1) (2)
Ɵ30 𝑥0 + Ɵ31 𝑥1 + Ɵ32 𝑥2 + Ɵ33 𝑥3 = 𝑍3

Substituting 4.44 into 4.42 transforms the activation function into the following form.

86 | P a g e
 (2) (2)
𝑎1 = 𝑔(𝑍1 )
(2) (2)
𝑎2 = 𝑔(𝑍2 ) 4.45
(2) (2)
{𝑎3 = 𝑔(𝑍3 )}
And if x and z represented by the following two matrices,
𝑥0
𝑥1
𝑥= [𝑥 ]
2
𝑥3
(2) 4.46
𝑧1
𝑧 (2) = [𝑧2(2) ]
(2)
{ 𝑧3 }

The value of the output layer is a sigmoid function of z(3), and can be determined by
the following equation.

ℎƟ (𝑥) = 𝑔(𝑧 (3) ) 4.47

Where the value of x are considered as input of activation function. The above
calculation procedure was implemented using MATLAB software in this study. The
detailed computation procedure is for simple ANN network with only one hidden layer
shown in Figure. 4.7. The complexity of a neural network is influenced by its size, i.e.,
the number of neurons and hidden layers as previously mentioned. However, the
network should be designed with enough level of complexity, so that it does not start
to over fit the data (Hagan et al., 2014; Soroush et al., 2015).
Figure 4.8 depicts the schematic diagram of ANN structure for prediction of
condensate viscosity. This design has one input layer consists of five parameters, one
hidden layer, two bias units and one output unit. This architecture recommended by
Hagan et al. (2014), Hagan and Menhaj (1994) and Hippert et al. (2001) as an efficient
and the most popular multilayer feed-forward architecture. Nevertheless, there are
large number of other designs, which might be considered suitable for other
applications. To select the best architecture in terms of number of neurons in a hidden
layer a trial and error procedure was implemented. The performance of each structure
was assessed by comparing coefficient of determination (R2) and root mean square
error (RMSE). We came up with the proposed structure shown in Figure 4.8, which

87 | P a g e
comprises of five neurons in layer two as the best topology. The input parameters are
API gravity, solution gas to oil ratio (Rs), pressure, temperature and gas specific gravity
(SG). The output layer is viscosity of condensate fluid calculated by the ANN network.

Figure 4.8. Developed ANN model architecture for prediction of condensate viscosity.

There are many algorithms available to train the network. The role of any algorithm is
to minimize the error between the output of the network and target values
(experimental condensate viscosity). This can be done by finding the optimum values
of the weights and biases in an iterative procedure. The most well-known training
algorithms are Levenberg–Marquardt (LM), scaled conjugate gradient (SCG),
Bayesian Regularization (BP) and resilient back propagation (Hippert, Pedreira and
Souza, 2001; Soroush et al., 2015). The LM backpropagation algorithm introduced by
Kenneth, (1944) and recommended by Behera and Chattopadhyay, (2012) as one of
the fastest and most popular backpropagation algorithm was used for adjusting the
weights in this study. The tangent sigmoid transfer functions selected as an activation
function for the neurons in hidden layer. For training of the model 70% of whole data
bank (210 data points) randomly selected and split into three subsets of 80% (168
data points) for training, 10% (21 data points) for validation and 10% (21 data points)
for testing. The ANN network is trained to map input data by iterative adjustment of
the weight functions. Information from the input layer feed forwarded through the
network to optimize the weight between the neurons. Optimization of the weight
function is carried out by back propagation of the error during training or learning stage.

88 | P a g e
The ANN reads the inputs and output values in training stage and changes the value
of weight functions to minimise the difference in predicted and the target (observed)
values. The error in prediction is minimized across training iterations (epochs) and
training continues to the point that the network reaches a specified level of accuracy
(Ghaffari et al., 2006a). Once the model has reached satisfactory accuracy or the
model is converged, the training will stop. The performance of the ANN network in
predicting the condensate viscosity has been shown in following graphical figures.

A B
Figure 4.9. Prediction performance of developed ANN network for condensate liquid viscosity
in training (A) and testing (B) stage.

To ensure that the developed model of LSSVM and ANN follow the physical trend of
condensate viscosity as a function of pressure, their performance have been tested
using three independent condensate viscosity samples adopted from the literature.
These samples are binary fluids representing gas-condensate mixture that show liquid
drop out below the saturation pressure. The samples are from (Al-Meshari et al., 2007;
Yang et al., 2007; Kashefi et al., 2013).
The results shown in Figure 4.10 and Figure 4.11 indicate that ANN and LSSVM are
following physical trend of condensate viscosity as a function of pressure with good
accuracy.

89 | P a g e
Figure 4.10. Prediction performance of the developed LSSVM condensate viscosity model
as a function of pressure for three independent samples.

Figure 4.11. Prediction performance of developed ANN condensate viscosity model as a

function of pressure for three independent samples.

Results and discussion

The prediction performance of several existing literature models against experimental

condensate viscosity data were tested. All existing literature models failed to predict
condensate viscosity with acceptable accuracy.
The cross plot of the results shown in Figure 4.3a confirms the poor performance of
widely used compositional based LBC, (1964) model in prediction of condensate
90 | P a g e
viscosity. The reason for this is due to sensitivity of LBC method to mixture density
and critical volumes of the heavy components. Hence, in this study the coefficients of
the LBC correlation were optimized using least-square approach to match the
experimental condensate viscosity data. The tuned LBC correlation shown on Table
4.1. Figure 4.3a represents the prediction performance of LBC, (1964) with default and
adjusted coefficients.
Furthermore, the coefficients of five well-known live oil viscosity correlations were
regressed to match the condensate experimental data. The results of these
regressions presented in Figure 4.3b – f. These empirical correlations are function of
dead oil viscosity and solution gas to oil ratio (Rs). It should be noted that dead oil
viscosity is one of the most “difficult” properties to be estimated by correlations due to
its dependency to paraffin, aromatic, naphthalene and asphaltene content (Whitson
and Brulé, 2000; Hemmati-Sarapardeh et al., 2014). This might be one of the reasons
for poor performance of the default empirical live oil viscosity correlations. Moreover,
these correlations were originally developed using crude oil samples, which its
properties are fundamentally different from condensate liquid. The API gravity of crude
oil sample is between 15 – 45⁰API while API gravity of condensate liquid is normally
above 50⁰API.
Poor performance of the published literature correlations in predicting condensate
viscosity, motivated to develop two machine leaning models of LSSVM and ANN
network in this study. The performance of the newly proposed models of LSSVM and
ANN were compared against refined previously published correlations through
graphical and statistical error analysis. The statistical error analysis carried out using
coefficient of determination (R2), root mean square error (RMSE) and average
absolute relative deviation percentage (AARD%). The result of this error analysis is
tabulated in the Table 4.5. Graphical representation of statistical analysis is provided
in Figure 4.12 and Figure 4.13. The results indicate that ANN model outperformed
other methods with AARD of 16.20%, R2 of 0.8423 and RMSE of 0.1144. ANN followed
by LSSVM, Kartoatmodjo and Schmidt (1994), Elsharkawy and Alikhan (1999),
Bergman and Sutton (2007), LBC (1964), De Ghetto et al. (1994) and Beggs and
Robinson (1975).

91 | P a g e
Table 4.5. Statistical parameters of developed ML based models and utilized correlations for
prediction of condensate viscosity.

Method R2 RMSE AARD%

LBC, (1964) 0.7241 0.1240 27.07
Bergman & Sutton, (2007) 0.7297 0.1236 26.29
Beggs & Robinson, (1975) 0.7207 0.1244 27.84
Elsharkawy & Alikhan, 0.7344 0.1228 24.87
(1999)
De Ghetto et al, (1994) 0.7243 0.1240 27.56
Kartoatmodjo & Schmidt, 0.7412 0.1220 23.89
(1994)
LSSVM 0.7738 0.1208 17.22
ANN 0.8423 0.1144 16.20

RMSE AARD%
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0

Figure 4.12. Statistical performance comparison of tuned literature models against developed
ANN and LSSVM models (AARD% and RMSE) for prediction of condensate viscosity.

92 | P a g e
Figure 4.13. Statistical performance comparison of tuned literature models against
developed ANN and LSSVM models (R2) for prediction of condensate viscosity.

The results confirm using either compositional model of LBC or live oil viscosity
correlations require significant tuning of coefficients for prediction of condensate
viscosity of gas-condensate reservoirs below the saturation pressure. Whereas
developed two intelligent approaches were able to monitor condensate viscosity with
appropriate precision and integrity.
Non-linear relationship between the available experimental data and the desired
outputs created using developed LSSVM model. The optimum values of two important
tuning parameters of LSSVM include 𝜎 2 and 𝛾 are presented in Table 4.3. Simulated
Annealing optimization (SA) algorithm was applied to achieve these two optimum
values.
The ability of proposed LSSVM and ANN models for calculating condensate liquid
viscosity as a function of changing pressure have been investigated for three gas
condensate samples from the literature. Figure 4.10 and Figure 4.11 are
demonstrating experimental and predicted condensate viscosities using LSSVM and
ANN models respectively. The results show that both models are able to forecast
physical trend of experimental condensate viscosity. The accuracy of the models for
predicting condensate viscosity of two independent samples determined by AARD%.

93 | P a g e
The error analysis show that both AI models perform well with acceptable level of
accuracy.
From Figure 4.10 and Figure 4.11, it is evident that increasing pressure would
decreases the condensate viscosity. The pressure changes due to depletion in gas-
condensate reservoirs can have significant effect on condensate viscosity variation
near wellbore region (Fevang and Whitson, 1996). These changes might due to the
complex behaviour of gas-condensate reservoir below the dew points pressure, which
violate thermodynamic laws. The developed LSSVM and ANN models successfully
captured the trend of condensate viscosity while utilized correlations were not accurate
enough in tracking these changes.
Although the prediction performance of the LSSVM was better than published
literature correlations, however the error was still high with R2 of 0.7738 and AARD of
17.22%. Therefore, Artificial Neural Network (ANN) method was used aiming to
improve the accuracy. Performance prediction of ANN network is a function of number
of neurons that is used in hidden layer. A trial and error approach were implemented
to find the optimum number of neurons. For this study the ANN architecture with five
neurons provides the most satisfying results with the least RMSE and the highest R2.
Both machine learning models are providing promising results in prediction of
condensate viscosity.
Although the developed ML based models are performing better than existing literature
models, however the main deficiencies of both ML models is that they sacrifice the
physics of the problem for accurate outcome. Hence, in another attempt of this study,
different machine learning method known as Fuzzy Logic (FL) employed in order to
compensate the neglected physics by ANN and LSSVM. The output of the TSK fuzzy
logic is in a form of linear equation. This approach will be explained in the following
sections of this chapter.

4.5 Fuzzy Logic Approach

In previous section of this chapter, it has been demonstrated that the accuracy of
existing literature models for predicting condensate viscosity are inadequate. Hence,
an alternative two soft computing approaches namely LSSVM and ANN models have
been developed for better estimation of condensate viscosity. Although
aforementioned approaches provide satisfactory estimation of condensate viscosity,

94 | P a g e
however the major criticism is that they are black box methods and visual relationship
between input and output parameters cannot be established (AlQuraishi and Shokir,
2011; Fayazi et al., 2014). These approaches are based on statistical truth of the
utilized data rather than scientific relationship between the input/output parameters.
As a result, not any meaningful correlation can be derived from these approaches . Our
aim is to develop accurate yet mathematically efficient condensate viscosity
correlation. The proposed correlation has to cope with nonlinearity of the condensate
viscosity that occurs due to various stage of pressure depletion in gas-condensate
reservoirs. For this purpose, another machine learning approach known as Fuzzy
Logic was utilized. There are two typical Fuzzy Logic approaches in literature as follow:
 Takagi – Sugeno – Kang (TSK) rule based fuzzy logic
 Mamdani approach
The TSK fuzzy gives accuracy of the ML based method without compromising the
physics of the problem.
Mamdani approach is purely linguistic; means interpret the mathematical problem in a
language understandable to the human logic. Both approaches utilized for accurate
modelling of condensate viscosity in gas-condensate reservoirs. In this section, first,
we give background information to the fuzzy logic approach in general and then TSK
and Mamdani techniques will be explained in details.

The science today is based on Aristotle’s crisp logic, formed more than 2000 year ago.
In Aristotle’s logic, every problem has a bivalent manner, means black and white, yes
or no and 0 and 1. The Aristotle’s bivalent logic made accessible to modern science
by German mathematician Cantor, who developed set theory for this purpose. In the
Cantor’s theory sets are defined as a collection of distinguishable objects. In 1937 the
first paper published on the theory of vague sets by Black, (1937). The vague sets
basically is similar to probability theory in mathematics. In vague sets, the objects no
longer are distinguishable as introduced by Cantor’s theory. In 1965, Zadeh continued
the work on vague sets and he introduced well known theory of fuzzy sets, which we
know it as fuzzy logic in today’s science literature (Zadeh, 1965). He developed many,
key concepts, including the idea of membership function and provide comprehensive
framework to apply the theory for solving engineering and scientific problem.
Fuzzy theory deals with vagueness, imprecision and uncertainty in a system. Unlike
the classical bivalent logic, that only admits (e.g., black and white, true (1) or false (0))

95 | P a g e
of an occurrence, fuzzy logic covers degree of truth of a factor between 0 and 1. In
another word in contrast to the classical crisp set, where an object either belongs to a
set or it does not, everything is a matter of degree in fuzzy set. The belonging of an
object to a set can be done by the membership function. Fuzzy logic provides a
linguistic solution for a problem, means it lifts the restriction of computer language (0
&1) for a problem.
Fuzzy logic approach has been used in several petroleum engineering application
including petro physics, reservoir characterization, enhanced oil recovery, drilling,
decision-making analysis and well-stimulation (Chen et al., 1993; Zhou, Wu and
Cheng, 1993; Zhanggui et al., 1998; Shokir, 2006, 2008; Ahmadi et al., 2015; Zhao et
al., 2018; Khazali, Sharifi and Ahmadi, 2019). Shokir, (2006) used rule based fuzzy
logic for modelling density and viscosity of natural gas reservoirs. Liao et al., (2008)
utilized fuzzy approach in separation process of oil; Khazali et al., (2019) applied fuzzy
decision tree for enhance oil recovery screening; Ahmadi and Ebadi, (2014) used
fuzzy logic method for calculation of minimum miscible pressure of injected gas and
reservoir oil.

Gas-condensate reservoirs modelling with condensate blockage effect, exhibits all

characteristics that a fuzzy logic system proposed to deal with. For instance the well
deliverability-modelling associate with inaccuracy in PVT and rock properties
estimation. Some properties like condensate (oil) viscosity has the largest uncertainty
both experimentally and in prediction (Fevang, 1995; Whitson, Fevang and Yang,
1999). Accurate determination of PVT properties is challenging task due to the
complex fluid behaviour of gas-condensate reservoirs below the dew point that violate
thermodynamic laws. The thermodynamic abnormality is because of the condensation
of gas phase “become liquid hydrocarbon (condensate)” phase instead of vaporization
as it is expected when pressure reduced and expansion induced in an isothermal
system. After many years of research, this behaviour is still not understood accurately
and we call it vague in this study. Such behaviour directly affect the accurate
estimation of PVT properties (condensate viscosity, compressibility factor).
These vagueness behaviour and difficulties motivated us in this study to take fuzzy
logic approach, hoping to tackle if not all some of these challenges.
Two most well-known methods of fuzzy logic in literature are Mamdani and Takagi –
Sugeno – Kang (TSK) approach. In this study, both methods have been adopted for
96 | P a g e
developing intelligent models for prediction of condensate viscosity below the
saturation pressure.

4.6 Takagi – Sugeno – Kang (TSK) Fuzzy approach

The most well know rule based fuzzy inference systems (FIS) are linguistic Mamdani-
type and Takagi – Sugeno – Kang (TSK) (Mamdani and Assilian, 1975; Takagi and
Sugeno, 1985). Both antecedents and consequence are fuzzy sets in Mamdani
approach, whereas in TSK model antecedent contains of fuzzy sets and the
consequence is a linear equation. Therefore, the outcome of TSK fuzzy model can be
interpreted as a linear equation. This linear equation would benefit the accuracy of the
intelligent model as well as representing the physics of the problem.
This linear relationship between input/output is defined by set of fuzzy so called IF –
THEN rules as follow.

𝑅1 : 𝑖𝑓 𝑥𝑖 𝑖𝑠 𝐴𝑖1 𝑎𝑛𝑑 … 𝑎𝑛𝑑 𝑥𝑚 𝑖𝑠 𝐴𝑖𝑚 𝑡ℎ𝑒𝑛

4.48
𝑦𝑖 = 𝑎𝑖1 𝑥1 + ⋯ + 𝑎𝑖𝑚 𝑥𝑚 + 𝑎𝑖0

Where 𝑅1 = (1, 2, … , 𝑛) is number of fuzzy rules, 𝑥𝑖 = (1,2, … , 𝑚) are the input

variables, 𝑦𝑖 are the output variables whose values are inferred, 𝐴𝑖1 , … , 𝐴𝑖𝑚 are
membership functions of the fuzzy sets in the premises and 𝑎𝑖0 , 𝑎𝑖1 , … , 𝑎𝑖𝑚 are the
model parameters in the consequence (Takagi and Sugeno, 1985). To determine
these three items using input-output data of a respective system “gas-condensate
viscosity data” the design procedure of TSK fuzzy model can be summarized in three
steps as follow:
1. Fuzzy clustering
2. Setting the membership functions
3. Parameters estimation (Takagi and Sugeno, 1985; Passino and Yurkovich,
1998; Shokir, 2006, 2008).
Partitioning set of input variables into some fuzzy sub sets can be carried out in first
two steps, while the relation of input/output of each fuzzy sub sets is defined in third
step (Takagi and Sugeno, 1985). The outline of the TSK fuzzy algorithm is shown in
Figure 4.14. From the diagram, it can be seen that TSK algorithm requires combination
of input/output variable called premise variables, where in this study the input variables
are pressure (P), temperature (T) and solution-gas to oil ratio (Rs), and the output

97 | P a g e
variable is condensate viscosity (𝜇𝑐 ). As previously mentioned temperature has a
significant effect on viscosity and it has given a special attention in this study because,
strictly speaking, according to Ahmed, (2010) and Whitson et al., (2000) gas
condensate reservoirs are defined by only temperature. If temperature located
between critical and cricondentherm temperature (326°K – 395°K) the reservoir
classed as gas-condensate reservoir.

After preparing the dataset the partitioning of the data is required, which any clustering
technique can be used for this step. Subsequently, the membership functions are
defined by Gaussian distribution, which essentially relates the premise variables to
each cluster with a certain degree of membership. Least square approach can be used
to determine the value of constants in 4.42.

Figure 4.14. Outline of the TSK fuzzy algorithm.

The optimum number of clusters is fundamental for efficiency of the TSK algorithm.
This also known as optimum number of rules for data sets. The optimum number of
rules, for condensate viscosity utilized data were determined using Calinski and
Harabasz, (1974) cluster evaluation method. The data bank introduced in 4.3.1 is used
for developing the fuzzy models in this section. Having determine optimum number of
clusters (rules), K-mean fuzzy clustering approach, as one of the most popular
classification techniques (Bezdek and Pal, 1992; Klawonn et al., 2015) for partitioning
set of data has been used to develop the TSK fuzzy model.
Then Gaussian membership function as an efficient technique over other methods
such as triangular or trapezoidal function (Ahmadi and Ebadi, 2014) was utilized to
determine the degree of membership of an object to a set. Hence, membership degree
of input/output data in respect to each cluster has been defined (Zadeh, 1965; Ross,
2017). The details description of each step highlighted in Figure 4.14, is given in the
following section.

98 | P a g e
 Fuzzy Clustering

Partitioning set of data points ‘input/output’ into several groups (clusters) in a way that
the points in the same group are highly similar and dissimilar from points of other
clusters is the aim of fuzzy clustering (Bai et al., 2017). In this study 250 data sets
were selected as a training data sets and divided into several clusters and then
interpreted as rules.
There are several fuzzy clustering methods in literature such as fuzzy c-means (FCM),
Gustafson-Kessel (GK), K-means clustering and subtractive clustering.
Comprehensive introduction to each clustering technique can be found in (Bezdek and
Pal, 1992).
In this study K-means clustering method was used as one of the most popular
classification algorithm for the data without any defined categories or unlabelled data.
This algorithm, is an iterative, hill climbing data-partitioning algorithm, where “𝑁”
observations can be partitioned into “𝑐 “ clusters. Each observation in the process
belongs to a cluster with nearest mean. Each cluster is represented by a so-called
prototype, which has to be in the centre of the corresponding cluster. K-means
clustering algorithm is based on an objective function “J” that can be determined from
the following equation (Klawonn, Kruse and Winkler, 2015).

𝑐 𝑛

𝐽 = ∑ ∑ 𝑢𝑖𝑗 𝑑𝑖𝑗 4.49

𝑖=1 𝑗=1

J also known as cost function and should be minimized under the following constraints:

∑ 𝑢𝑖𝑗 = 1 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑗 ∈ {1, … , 𝑛} 4.50

𝑖=1

Where 𝑢𝑖𝑗 ∈ {0,1} indicates whether data vector 𝑥𝑗 is assigned to a cluster 𝑖 (𝑢𝑖𝑗 = 1)
2
or not (𝑢𝑖𝑗 = 0); 𝑑𝑖𝑗 = ‖𝑥𝑗 − 𝑣𝑖 ‖ is squared Euclidean distance between data vector
𝑥𝑗 and cluster prototype 𝑣𝑖 . In this method the number of cluster “𝑐 “must be known in
advance. Our criteria for assigning initial number of clusters is based on the
assumption that nonlinearity in the data can be approximated by 12 clusters (Shokir,
2006, 2008).

99 | P a g e
In general, there is no specific rule for defining optimum number of clusters “𝑐 “
however several number of techniques such as elbow method, the silhouette method
G-means algorithm and Calinski-Harabasz are exist in literature. In this study Calinski-
Harabasz cluster evaluation method was used as an efficient technique (Calinski and
Harabasz, 1974). The criteria of Calinski-Harabasz also called variance ratio criterion
(VRC) and its defined as follow.

𝑆𝑆𝐵 (𝑁 − 𝑐)
𝑉𝑅𝐶𝑐 = × 4.51
𝑆𝑆𝑤 (𝑐 − 1)

Where 𝑆𝑆𝐵 and 𝑆𝑆𝑤 are between and within overall cluster variance respectively,
𝑐 stands for number of clusters, and N represents the data points. The 𝑆𝑆𝐵 and 𝑆𝑆𝑤
are defined as follow.
𝑐

𝑆𝑆𝐵 = ∑ 𝑛𝑖 ‖𝑚𝑖 − 𝑚‖2

𝑖=1
𝑐 4.52
𝑆𝑆𝑊 = ∑ ∑ ‖𝑥 − 𝑚𝑖 ‖2
𝑖=1 𝑥∈𝑐𝑗

Where 𝑛𝑖 is number of observations in cluster 𝑖, 𝑚𝑖 is centroid of cluster 𝑖, 𝑚 stands

for mean of the data, 𝑥 is number of data samples, 𝑐𝑗 is the ith cluster and ‖𝑚𝑖 − 𝑚‖2,
‖𝑥 − 𝑚𝑖 ‖2 is Euclidean distances between two vectors. Large 𝑆𝑆𝐵 and a smaller 𝑆𝑆𝑤
are representing well-grouped clusters, which means the larger 𝑉𝑅𝐶𝑐 ratio, the better
data partition (Calinski and Harabasz, 1974). Therefore to achieve the optimum
number of clusters, the validity measure of 𝑉𝑅𝐶𝑐 is maximized with respect to number
of clusters 𝑐. Hence, highest Calinski-Harabasz index is the optimum number of
clusters (Calinski and Harabasz, 1974; Bandyopadhyay and Maulik, 2002). Figure
4.15 illustrates the obtained results using validity criterion function 𝑉𝑅𝐶𝑐 using equation
4.51 for training data set.

100 | P a g e
Figure 4.15. Optimum number of cluster for the training data sets using validity function
(VRC), for condensate viscosity input data.

Having defined the optimum number of clusters for our training data, k-means
algorithm, presented in equation 4.49 was proceed for assumed finite set of points 𝐴 =
{𝑥1 , 𝑥2 … , 𝑥𝑛 } in n-dimensional space ℝ𝑛 as follow.
Step 1: Choose an initial cluster centres 𝑧1 , 𝑧2 , 𝑧3 , … , 𝑧𝐾 randomly from the 𝑛 points
{𝑥1 , 𝑥2 … , 𝑥𝑛 }.
Step 2: assign data points 𝑎 = 𝐴 to its closest centre and obtain 𝑘-partition of 𝐴.
Step 3: Recalculate centres for the new partition and go to step 2 until no more data,
change their clusters, or the algorithm is converged.
K-means algorithm is minimizing distances of sum of point-to-centroid, over all defined
clusters. This is an iterative procedure where centre of each cluster k, and membership
values of each cluster (maximum and minimum) can be obtained at the end (Jain,
Murty and Flynn, 1999; Bandyopadhyay and Maulik, 2002; Bagirov, 2008; Klawonn,
Kruse and Winkler, 2015).

Setting the membership function

To determine membership degree of an object to a certain set, 𝐴𝑖1 , … , 𝐴𝑖𝑚 in equation

4.48 the membership function has to be set (Zadeh, 1965; Mohaghegh, 2000). The
membership degree is usually between 0 and 1.
Mathematical expression in equation 4.53 states the membership (1) or non-
membership (0) of a component x to a set (universe U). This is a binary issue, which
state an object is either belong to a set or not (Zadeh, 1965).
101 | P a g e
 1, 𝑥∈𝐴
𝜇𝐴 (𝑥) = { 4.53
0, 𝑥∉𝐴

Where 𝜇𝐴 (𝑥) represents an ambiguous membership of component x in set A, and ∈

and ∉ represent contained or not contained in set A, respectively. Zadeh, (1965)
extended classical binary membership of only 0 and 1 to a real continuous intervals
where the numbers between 0 and 1 can represent various degree of a membership
of an object to a set. This mathematically can be represented as follow.

𝜇𝐴 : 𝑈 → [0,1] 4.54

Where U represents a universal set defined for specific problem in fuzzy set A. For
instance if 𝑈 = {𝑥1 , 𝑥2 , … , 𝑥𝑛 }, then the degree of membership of 𝑥1 , 𝑥2 , … , 𝑥𝑛 in U can
be defined by the following equation.

𝐴 = {(𝜇𝐴 (𝑥1 )|𝑥1 ), (𝜇𝐴 (𝑥2 )|𝑥2 ), … , (𝜇𝐴 (𝑥𝑛 )|𝑥𝑛 )} 4.55

The relation between the input/output are defined by fuzzy if/then rules, where
conclusion can be achieved based on the hypothesis. This explain the principle of an
inference mechanism that states if a fact of a hypothesis is known then another fact
or conclusion can be reached (Shokir, 2008).
The information that how the data points are distributed in the input space would
provide the guideline for creating number of fuzzy clusters and their detection. Cluster
centres and eigenvalues of fuzzy covariant matrices can be used for capturing this
information (Takagi and Sugeno, 1985; Shokir, 2006, 2008). For example if the input
space of ith cluster centre is 𝑣𝑖1 , 𝑣𝑖2 , … , 𝑣𝑖𝑚 , then the antecedent fuzzy sets of TSK
model can be defined by any membership function. In this study gaussian membership
function is employed to define the antecedent fuzzy sets as follow.

(𝑣𝑖 − 𝑐𝑖 )2
𝜇𝐴𝑖 (𝑣𝑖 ) = 𝑒𝑥𝑝 (− ) 4.56
2𝜎𝑖 2

102 | P a g e
Where 𝑣𝑖𝑗 is scalar values of inputs, 𝜎𝑖 is standard deviation and 𝑐𝑖 is the mean of the
ith fuzzy set 𝐴𝑖. The centre 𝑐𝑖 and variance 𝜎𝑖 of the membership function is shown in
following figure.

Figure 4.16. Gaussian membership function for detecting fuzzy clusters.

Determination of constants

In order to develop condensate viscosity correlation the data bank collected from open
literature, divided to two parts, where 242 of these data sets randomly selected to build
the TSK model and 84 data sets utilized to test the developed model. The data sets
N=1,2,…., 242 have been organized in a matrix form with dimension of a 242 × 4, as
shown in equation 4.57. Having determined optimum number of clusters using Calinski
and Harabasz, (1974) method, K-mean algorithm has been applied to define the range
of each cluster. Gaussian membership function then used to determine how a data
point belongs to a defined cluster by means of membership degree between 0 and 1.
Figure 4.17 illustrates the output respond of Gaussian membership functions to one
of the inputs (reservoir pressure).
𝑃1 𝑇1 𝑅𝑠1 𝜇𝑐1
𝑃 𝑇2 𝑅𝑠2 𝜇𝑐2
𝑁=[ 2 ] 4.57
⋮ ⋮ ⋮ ⋮
𝑃𝑁 𝑇𝑁 𝑅𝑠𝑁 𝜇𝑐𝑁

103 | P a g e
 Figure 4.17. Gaussian membership function for input 1 “pressure”.

The consequent parameters of the TSK model 𝑎𝑖0 , 𝑎𝑖1 , … , 𝑎𝑖𝑚 term in equation 4.48
are computed using least-square approximation method in the following fashion.
If X denotes to a matrix having 𝑖𝑡ℎ row in the input vector of 𝑥𝑖 (inputs), Y represents
a vector column with 𝑦𝑖 (output) as its 𝑖𝑡ℎ component and 𝑤𝑖 denotes to 𝑁 × 𝑁 real
matrix, then the degree of firing 𝛽𝑖𝑗 is defined by equation 4.57. The 𝛽𝑖𝑗 is jth diagonal
of the real matrix. (Takagi and Sugeno, 1985; Shokir, 2008).

𝛽𝑖 (𝑥𝑖 )
𝛽𝑖𝑗 = 𝑐 4.58
∑𝑧=1 𝛽𝑖 (𝑥𝑗 )

And if 𝜃𝑖 = [𝑎𝑖1 , … , 𝑎𝑖𝑚 , 𝑎𝑖0 ] represents consequent parameters of ith rule in each
vector, in order to determine 𝑎𝑖0 in 𝜃𝑖 , a unitary column 𝐼 is added to the matrix X, 𝑋𝑒 =
[𝑋, 𝐼]. This is an extended matrix for the input values, then 𝜃𝑖 is calculated by the
following equation.

−1
𝜃𝑖 = [𝑋𝑒 𝑇 . 𝑊𝑖 . 𝑋𝑒 ] 𝑋𝑒 𝑇 . 𝑊𝑖 . 𝑌 4.59

Where 𝑋𝑒 𝑇 is transpose of matrix 𝑋𝑒 . The obtained parameters 𝜃𝑖 for each matrix,

substituted in equation 4.60 to approximate output value Y.

104 | P a g e
 𝑌 ≈ 𝑋. 𝜃𝑖 4.60

Y represents condensate viscosity in this study, which can also be represented as a

linear function that is explained in following section.

Development of new correlation

In this part of the study, a simple yet accurate correlation has been developed for
prediction of condensate (oil) viscosity in gas-condensate reservoirs below the dew
point pressure, using TSK fuzzy approach. The proposed correlation is function of
pressure, temperature and solution gas to oil ratio. Our reasoning for aforementioned
selection of the factors explained in following. Condensate viscosity is direct function
of reservoir pressure as changing in pressure due to reservoir depletion has a major
impact on this important property (Fevang, 1995; Whitson, 2000). The effect of
pressure on condensate viscosity is demonstrated in Figure 4.10 and Figure 4.11. In
many gas-condensate reservoirs, condensate liquid developed only when BHFP
depleted below the dew point pressure. Moreover, it is well-known that temperature
has a significant effect on hydrocarbon viscosity (Gozalpour et al., 2005; Craft and
Hawkin, 2015). Another factor that has been considered in developing the viscosity
function is solution gas to oil ratio (Rs). Solution gas to oil ratio determine how much
gas is dissolved in the liquid, which indicate if the mixture behave more like liquid or
gas. This in return directly influence the viscosity of gas-condensate mixture. The
recommended linear function in this study has the following form.

𝜇𝑐 = 𝐴𝑃 + 𝐵𝑇 + 𝐶𝑅𝑠 + 𝐷 4.61

Where P, T and Rs stand for reservoir pressure, temperature and solution gas to oil
ratio respectively and A, B, C and D are arbitrary constants of the model. Pressure
and temperature are directly were imported from the data bank as an input. Rs is
calculated from Standing and Katz, (1942) shown in 4.22.
The TSK fuzzy approach is employed to determine and optimize the parameters of A,
B, C and D in 4.60. The architecture of fuzzy system using TSK fuzzy with Gaussian

105 | P a g e
MFs shown in Figure 4.18. Numerical value of 9, next to the inputs, are devoting to the
number of rules.

Figure 4.18. Architecture of constructed TSK fuzzy model for predicting condensate viscosity.

The range of each cluster for input variables (P, T, Rs) and the values of constants
parameters (A, B, C, D) in equation 4.60 is presented in Table 4.6. For instance, the
function introduced in 4.60 can be used in the following fashion:
𝜇𝑐 = 𝐴𝑃 + 𝐵𝑇 + 𝐶𝑅𝑠 + 𝐷
{ 𝑅1: 𝑖𝑓 44.99 < 𝑃 < 75.15 𝑎𝑛𝑑 303.15 < 𝑇 < 405.37 𝑎𝑛𝑑 5245 < 𝑅𝑠 < 6101
𝑇ℎ𝑒𝑛 𝜇𝑐 = −0.0063𝑃 + 0.0025𝑇 + 0.0452𝑅𝑠 + 0.0032

Table 4.6. The range of constants and input parameters of proposed condensate viscosity
model.

Pressure Temperature Rs A B C D
(MPa) (°K) (scf/STB)
44.99 – 75.15 303.15 – 405.37 5245 – 6101 -0.0063 0.0025 0.0452 0.0032
11.29 – 26.93 348 – 404.6 714 – 9732 0.0003 0.0025 -0.0123 -0.0063
19.91 – 20.24 390.72 – 393.15 1167 – 1465 0.0024 0.0022 0.000124 -7.007
50.07 – 75.44 303.15 – 315.92 971 – 3646 0.0011 -0.0056 -0.0017 23.84
7.99 – 11.29 348.15 – 353.92 1160 – 9869 0.0008 -0.0019 0.000012 -0.5807
28.52 – 31.16 323.15 – 353.15 4955 – 6267 0.0006 -0.0034 0.0001 -0.7915
3.55 – 10.49 255.37 – 303.15 4695 – 5425 0.0007 0.0014 0.0001 0.2219
21.90 – 60.08 315.28 – 405.37 1235 – 9186 0.0001 0.00004 - 0.001 0.6126
11.29 – 31.16 323.15 – 443.15 2222 – 2618 0.00001 0.0036 -0.0053 13.97

Results and discussions

The new developed model was compared with well-known literature correlations of
LBC (1964); Beggs and Robinsons, (1975); Kartoatmodjo and Schmidt, (1991), De
Ghetto et al. (1999) and Elsharkawy and Alikhan, (1999). These existing literature

106 | P a g e
correlations were introduced previously in this chapter. The developed correlation was
also compared with other tow developed machine learning (ML) based models of ANN
and LSSVM in this study. Statistical error parameters of root mean square error
(RMSE), mean absolute error (MAE) and average absolute relative deviation
percentage (AARD%) shown in Table 4.7, and were used for comparison of the
results. RMSE and AARD% were estimated using equations 4.24 and 4.25
respectively. MAE is determined from the following equation.

𝑛
1
𝑀𝐴𝐸 = ∑|𝜇𝑒𝑥𝑝𝑒𝑟𝑖𝑚𝑒𝑛𝑡𝑎𝑙 − 𝜇𝑐𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑 | 4.62
𝑛
𝑗=1

To visualize the performance of the new developed model against other methods
graphical error analysis include error distribution and cross plot have been utilized. We
tested the prediction performance of the developed correlation as well as other models
for condensate viscosity in the range of (0 – 0.5cp). This is because the viscosity of
condensate liquid is normally between 0 – 1cp (Fevang, 1995; Whitson and Brulé,
2000) and in lower range, this liquid can flow toward the wellbore along with the gas
phase. In higher viscosity range the interfacial tension of the condensate liquid is high,
which not allowed the liquid flow towards the wellbore. In higher viscosity range, most
of the liquid drop out not able to flow simultaneously with gas phase and trapped inside
the porous structure (become non-recoverable condensate). This is due to stronger
intermolecular forces that imposed to the liquid droplets in lower viscosity region.
The results indicate that the developed condensate viscosity correlation yields good
agreement between the predicted and measured condensate viscosity with the lowest
RMSE of 0.0194, MAE of 0.0163, and AARD % of 7.123. The cross plot of estimated
against experimental condense viscosity is presented in Figure 4.19A-H. From the
results we can see that Kartoatmodjo and Schmidt, (1994) has the highest scattering
around zero error line while Beggs and Robinson, (1975) provides least spreading for
prediction of condensate viscosity. The reason for the high error by utilized literature
models is might due to individual limitation of each model and also different oil type
(e.g., oil with lower API gravity) that was used in their development. Both ANN and
LSSVM performance for prediction of condensate viscosity are good in lower range of
(0 – 0.3cp) as it shown in Figure 4.19F and Figure 4.19G. However, in higher viscosity
range between 0.3 – 0.5cp, the data scatter and their prediction performance

107 | P a g e
deteriorate. This is because the majority of the utilized data bank (73%) that used in
development of ANN and LSSVM models are within the range of 0 – 0.4cp. Hence,
their performance is better within aforementioned range are more accurate. Although
the error is high in higher viscosity range, but still the performance of these two ML
based models are better than existing literature correlations. This inconsistency of
condensate viscosity prediction by ANN and LSSVM was tackled in this study using
TSK fuzzy approach. Figure 4.19H depicts the performance of the developed
condensate viscosity model using TSK fuzzy approach in the range of (0 – 0.5cp). The
result confirms that TSK fuzzy model well predicts experimental viscosity in
aforementioned range.
Table 4.7. Statistical accuracy of condensate phase viscosity models.
Method RMSE MAE AARD%

Lohrenz-Bray-Clark 0.1826 0.1493 54.98

(1964)
Beggs and Robinson 0.1205 0.0926 39.14
(1975)
Elsharkawy and Alikhan 0.1468 0.1199 47.75
(1999)
De Ghetto (1994) ) 0.1427 0.1149 45.19
Kartoatmodjo and 0.1865 0.1666 63.49
Schmidt, (1994)
ANN 0.0656 0.0474 20.11
LSSVM 0.074 0.0631 26.50
TSK Fuzzy 0.0194 0.0163 7.123

108 | P a g e
 Lohrenz-Bray-Clark, (1964) [A] Beggs and Robinson, (1975) [B]

De Ghetto et al., (1994) [C] Kartoatmodjo and Schmidt, (1994) [D]

Elsharkawy and Alikhan, (1999) [E] LSSVM [F]

ANN [G] TSK Fuzzy method [H]

Figure 4.19. Cross plot of estimated against condensate viscosity measurements of existing
literature correlation, ANN, LSSVM and TSK Fuzzy method.

109 | P a g e
It is important to know how the developed models behave in prediction of condensate
viscosity as a function of pressure and temperature. This is because condensate
viscosity is strong function of pressure and temperature. Therefore, Figure 4.20A – H
and Figure 4.21A – H were generated, which illustrate how each model performs in
predicting condensate viscosity using relative errors percentage based on reservoir
pressure and temperature. Both figures confirm that the developed correlation using
TSK fuzzy approach scatter less than other models around the zero error line. In
addition, the results show that compositional model of LBC, (1964) underestimates the
condensate viscosity while other correlations are overestimating the condensate
viscosity in specified pressure range of (0.25 – 75.84 MPa) and temperature range of
(303 – 443.15°K).
Figure 4.20H and Figure 4.21H both are showing that the new condensate viscosity
model is responding very well to the pressure and temperature change in prediction
of condensate viscosity, while existing literature models cannot cope with non-linearity
of the gas-condensate system due to pressure/temperature change in the system.
This confirms by poor performance of the existing literature models depicted in Figure
4.20A – E.

LBC, (1964) [A] Beggs and Robinson, (1975) [B]

De Ghetto et al., (1994) [C] Kartoatmodjo and Schmidt, (1994) [D]

110 | P a g e
 Elsharkawy and Alikhan, (1999) [E] LSSVM [F]

ANN [G] TSK Fuzzy [H]

Figure 4.20. Residual plot of relative error percentage for different viscosity models as a
function of reservoir pressure.

Lohrenz-Bray-Clark, (1964) [A] Beggs and Robinson, (1975) [B]

De Ghetto et al., (1994) [C] Kartoatmodjo and Schmidt, (1994) [D]

111 | P a g e
 Elsharkawy and Alikhan, (1999) [E] LSSVM [F]

ANN [G] TSK Fuzzy [H]

Figure 4.21. Residual plot of relative error percentage for different viscosity models as a
function of reservoir temperature.

In order to see the effect of temperature, pressure and solution gas to oil ratio on
condensate viscosity, the trend analysis of the developed model has also been
studied. Figure 4.22A – C show the effect of reservoir pressure in the range of (0 –
78MPa) on condensate viscosity. As the pressure increases condensate oil viscosity
is also increasing. The results also comply with the physics of the problem as
explained in following.
In gas-condensate reservoirs with depletion mode of recovery in the beginning of the
production, the initial reservoir pressure is high and condensate liquid behave like a
gas phase, due to existing very high amount of (Rs) in the mixture. The relationship
between Rs and reservoir pressure in this study is illustrated in Figure 4.23. As the
reservoir pressure depleted, it reaches the dew point pressure where hydrocarbon
liquid phase developed, the amount of dissolved gas (Rs) is decreasing proportionally
with pressure reduction. Hence, in lower reservoir pressure there is less amount of
dissolved gas, which in return increase the viscosity of the condensate liquid (Bergman
and Sutton, 2007a; Hemmati-Sarapardeh et al., 2014). The result in Figure 4.22B also

112 | P a g e
confirms the aforementioned criterion. As it can be seen from these Figure 4.22 and
Figure 4.23, the developed condensate viscosity correlations are following the physical
trend of the experimental data with very good accuracy.
The effect of temperature on condensate viscosity within the temperature range of
(338 – 420°K) illustrated in Figure 4.22C. It is well known that increasing temperature
will decrease the viscosity of the liquid either in isobaric condition or as saturated liquid
condition (Craft and Hawkin, 2015, p. 511). According to the particle theory when
temperature of a liquid increases this would increase distance between the molecules,
which reduce the binding forces that hold the molecules of the liquid. This
proportionally decrease the viscosity of the liquid. The result presented in Figure 4.22C
satisfies the particle theory very well and confirms the validity of the developed TSK
model for prediction of condensate viscosity as a function of temperature.

(A) (B)

(C)
Figure 4.22. Experimental prediction capability of the developed TSK fuzzy model as a
function pressure (A), solution gas to oil ratio (B) and temperature (C).

113 | P a g e
Figure 4.23. The relation between pressure and solution gas to oil ratio (Rs) for the utilized
databank.

4.7 Mamdani Fuzzy Approach

In previous section, it has been observed that using TSK fuzzy approach given
promising results for prediction of condensate viscosity. In this section another well
known fuzzy logic approach known as Mamdani fuzzy is applied for better modelling
of condensate viscosity in gas-condensate reservoirs.
Mamdani fuzzy approach has been applied for many engineering applications. The
main reason for using this method was to use completely new methodology based on
linguistic approach in modelling gas-condensate PVT property (condensate viscosity)
and simplify the computational procedure.
In this section, first we introduce the nature of fuzzy algorithm and show how Mamdani
method embedded in the algorithm. Then the detail description of Mamdani approach
for modelling condensate viscosity will be provided.

Fuzzy Engine process

It is concluded by Biezma et al., (2018) that the fuzzy system introduced by Zadeh,
(1973) has three main following features.
 Linguistic variables instead of or in addition to numerical variables
 An inference mechanism that uses approximate reasoning algorithms;
 Simple relations between the variables in terms of IF-THEN rules to formulate
complex relationships.

114 | P a g e
These aforementioned characteristics of FL allow us to deal with any problem that has
imprecise, ambiguous and vague information to be dealt with conventional binary
logic. Figure 4.24 depicts the relationship between aforementioned three features
introduced by Zadeh, (1973) in a typical fuzzy logic system.

Figure 4.24. General architecture of a Fuzzy Logic System (FLS).

The input data is the condensate viscosity data including pressure, temperature and
solution gas to oil ratio (Rs). The output of the system would be condensate viscosity.
In the fuzzification step, the considered value of the input variables converted into
grades of memberships of fuzzy sets. Fuzzy sets are defined by membership functions
(e.g., Triangular, Gaussian, Trapezoid, etc.,). These MFs utilizes the spectrum of
logical values between 0 and 1. To achieve the spectrum membership functions are
divided into certain number of subsets identified with linguistic terms (e.g., Very Low,
Low, Medium, High and Very High).
In next step having generated number of fuzzy rules either manually or automatically,
these so called IF-THEN rules transferred into the inference fuzzy engine for fuzzy
reasoning and establishing input, output relationship. During the process several fuzzy
rules are fired in parallel-rule firing to allow simultaneous consideration of all the
information (Ahmadi and Ebadi, 2014; Mohaghegh, 2000; Ross, 2017). The
information generated in fuzzy inference system should be defuzzified to be able to
see a numerical values in the form of a crisp sets. Mathematically, the defuzzification
of a fuzzy sets are the process of “rounding it off” from its location in the unit hypercube
to the nearest. The last step is implementation of the defuzzified fuzzy sets to obtain
the produced output from fuzzy inference system (FIS) in the form of crisp values.
The detail of each step for modelling condensate viscosity is explained in following
sections.

115 | P a g e
 Membership Functions (MFs) fuzzification

The role of MFs in any fuzzy inference system discussed in details in section 4.6.2,
here the focus is on using different type of MFs in fuzzy inference systems.
All the information in fuzzy set is described by its membership function, which they can
be symmetrical or asymmetrical. This can specify the degree of membership of input
variables to different subsets (e.g., Very Low, Low, Medium, High, and Very High).
Different kind of MFs with the variety of the detected advantages and disadvantages
have been proposed in literature. It is very important to select right type of membership
function suitable to the relevant problem as success of one MF in one problem cannot
guarantee its applicability to another.
Several geometrical shapes proposed in literature for defining MFs. In this study we
choose three most widely used MFs in literature include gaussian, triangular and
trapezoidal due to their flexibility, transmissibility and eye catching ability (Ahmadi and
Ebadi, 2014; Hameed, 2011; Khazali et al., 2019). In this part we examine the
performance of these three well-known membership functions for determination of
condensate viscosity in gas-condensate reservoirs.
Gaussian MFs, defined by equation 4.63, are suitable for problems that require
continuously differentiable curves and smooth transition (Hameed, 2011). Gaussian
MFs initially utilized to relate the three inputs (pressure, temperature and RS) to one
output (condensate viscosity) of FIS system.

𝑥−𝑐)2
−(( ) 4.63
𝑓(𝑥; 𝜎, 𝑐) = 𝑒 2𝜎 2

Where c is the centre (i.e., mean) of ith fuzzy set and 𝜎 is the width (i.e., standard
deviation) of the ith fuzzy set.
Based on the above formula the three inputs of pressure, temperature and solution
gas to oil ratio (Rs) and one output of condensate viscosity were defined. Then several
main linguistic subsets known as Very Low (VL), Low (L), Medium (M), High (H) and
Very High (VH) used to interrelate the inputs and output. Figure 4.25A-D is illustrating
the input and output parameters of Mamdani approach using Gaussian MF.

116 | P a g e
 (A) (B)

(C) (D)

Figure 4.25. The constructed gaussian membership function representing (A) pressure (B)
temperature (C) solution gas to oil ratio and (D) condensate viscosity.

The triangular shape MFs are the second most important and widely used function in
fuzzy inference systems. The triangular shape is vector of x and depends on three
scalar parameters a, b and c. The scalar parameters are basically representing the
three sides of triangle. Mathematically triangular MFs are defined as follow.
0, 𝑥≤𝑎
𝑥−𝑎
, 𝑎≤𝑥≤𝑏
𝑓(𝑥; 𝑎, 𝑏, 𝑐) = 𝑐𝑏 −
−𝑎
𝑥 4.64
, 𝑏≤𝑥≤𝑐
𝑐−𝑏
{ 0, 𝑥≥𝑐 }

Graphical representation of triangular MFs for condensate viscosity estimation

depicted in following figure.

117 | P a g e
 (A) (B)

(C) (D)

Figure 4.26. The constructed triangular membership function representing (a) pressure (b)
temperature (c) solution gas to oil ratio and (d) condensate viscosity.

In the last step of this level to relate the inputs and output parameters of condensate
viscosity within linguistic subintervals, trapezoidal membership functions have been
utilized. Mathematical definition of trapezoidal MF is given in equation 4.65. The
trapezoidal shape is a function of a vector, x, and four scaler parameters of a, b, c and
d. The parameters of a and d represent the “feet” of the trapezoid and b and c
represent the “shoulder” (MathWorks, 2019). Figure 27A-D illustrates the trapezoidal
MFs for three inputs of pressure, temperature and solution gas to oil ratio and the
output of condensate viscosity.
0, 𝑥≤𝑎
𝑥−𝑎
, 𝑎≤𝑥≤𝑏
𝑏−𝑎
𝑓(𝑥; 𝑎, 𝑏, 𝑐, 𝑑) = 1, 𝑏≤𝑥≤𝑐
𝑑−𝑥 4.65
, 𝑐≤𝑥≤𝑑
𝑑−𝑐
{ 0, 𝑥≥𝑑 }

118 | P a g e
 (A) (B)

(D)
(C)

Figure 4.27. The constructed trapezoid membership functions (A) pressure (B) temperature
(C) solution gas to oil ratio and (D) the condensate viscosity.

Mamdani Fuzzy Inference System

Having defined the MFs in previous section, the expert knowledge has been
introduced into the fuzzy inference system, defining IF-THEN rules. The degree of
membership in one fuzzy set is determined by IF portion of the rule and the
consequence associated with system output is determined by THEN portion of the
rule.
To implement IF-THEN rule procedure, Mamdani FIS is proposed by Mamdani and
Assilian, (1975) and widely accepted among the fuzzy community due to the following
advantages (Asklany et al., 2011; MathWorks, 2019).
 Very fast computational procedure
 Easy to implement
 It is intuitive
 Well suited to human input

119 | P a g e
The applicability of Mamdani fuzzy approach for modelling condensate viscosity is
examined by this study.
In this method the fuzzy sets from the consequent of each rule are combined through
the aggregation operator and the resulting fuzzy set is defuzzified to yield the output
of the system (Mamdani and Assilian, 1975; Asklany et al., 2011).The typical IF-THEN
rule of the Mamdani algorithm is shown in the following equation.

𝐼𝐹 𝑥1 𝑖𝑠 𝐴1 𝐴𝑁𝐷 𝑥2 𝑖𝑠 𝐴2 … 𝐴𝑁𝐷 𝑥𝑛 𝑖𝑠 𝐴𝑛 𝑇𝐻𝐸𝑁 𝑦 𝑖𝑠 𝐵 4.66

Where: 𝑥1,2,...,𝑛 are input variables; 𝐴1,2,…,𝑛 are value of a certain linguistic input (VL, L,
M, H or VH); y is output and B is value of the output (VL, L, M, H or VH).
In this study the parameters of pressure, temperature and solution gas to oil ratio (Rs)
are input variables and condensate viscosity is the output variable. The relationship
between considered input variables and condensate viscosity (𝜇𝑐 ) is defined by set of
15 rules shown in Table 4.8. These set of rules are including all input, output
relationship possibilities.
This table is read from left to right for instance, rule no 1 can be read as follow:
IF Pressure is VL AND Temperature is VLLMHVH AND Rs is VLLMHVH THEN
Condensate Viscosity is VL.
Where VLLMHVH = Very Low, Low, Medium, High or Very High.

120 | P a g e
Table 4.8. Fuzzy rules defined for estimation of condensate viscosity in gas-condensate
reservoirs.

no Pressure Temperature Rs Condensate

(psia) (F) (scf/STB) Viscosity (cp)
1 VL VLLMHVH VLLMHVH VL
2 L VLLMHVH VLLMHVH L
3 M VLLMHVH VLLMHVH M
4 H VLLMHVH VLLMHVH H
5 VH VLLMHVH VLLMHVH VH
6 VLLMHVH VL VLLMHVH VL
7 VLLMHVH L VLLMHVH L
8 VLLMHVH M VLLMHVH M
9 VLLMHVH H VLLMHVH H
10 VLLMHVH VH VLLMHVH VH
11 VLLMHVH VLLMHVH VL VL
12 VLLMHVH VLLMHVH L L
13 VLLMHVH VLLMHVH M M
14 VLLMHVH VLLMHVH H H
15 VLLMHVH VLLMHVH VH VH

Maximum – minimum (max-min) composition method was utilized to implement

Mamdani FIS model as one of the most common type in literature (Ross, 2017). The
max-min method is initially used by Zadeh, (1965) in definition of approximate
reasoning using natural language IF-THEN rules. In max-min method minimum
operator is used to model the implication and the resulting output membership
functions are combined by using maximum operator (Hameed, 2011; Ross, 2017;
Biezma, Agudo and Barron, 2018).

Defuzzification in Mamdani approach

Defuzzification is required to convert the fuzzified quantity of inputs to a precise

quantity of output. There are many methods in literature for defuzzification of FIS
output. In this study we used Centroid or centre of gravity (COG) method as one of the
most prevalent and physically attractive among all the defuzzification methods (Takagi
and Sugeno, 1985; Leekwijck and Kerre, 1999; Ross, 2017, p. 121). Centroid

121 | P a g e
defuzzification returns the centre of area under the curve using integration technique
as expressed by the following equation.

∫ 𝑥. 𝜇(𝑥)𝑑𝑥
𝜇𝑐𝑒𝑛𝑡𝑟𝑒 = 4.67
∫ 𝜇(𝑥)𝑑𝑥

Where 𝜇(𝑥) is representing values of membership functions obtained from output of

FIS engine, 𝜇𝑐𝑒𝑛𝑡𝑟𝑒 is the centre of the fuzzified membership functions. The integrals
are taken over the entire range of the FIS output (condensate viscosity) and present
the 𝜇𝑐𝑒𝑛𝑡𝑟𝑒 as a crisp value.
This crisp value is the numerical value of condensate viscosity returned by the FIS
system. All calculation procedure of the Mamdani approach carried out in MATLAB.
Three Mamdani fuzzy using the above discussed procedure was constructed with
different membership functions for prediction of condensate viscosity. Figure 4.28
represents the constructed Mamdani fuzzy model using Gaussian MFs. Another two
Mamdani fuzzy systems using Triangular and Trapezoidal membership function were
constructed in similar fashion. Numerical values of 5 next to the inputs and output of
the system represent utilized 5 linguistic terms of Very Low (VL), Low (L), Medium (M),
High (H) and Very High (VH).

Figure 4.28. Architecture of the fuzzy inference (FIS) system for predicting condensate
viscosity using Gaussian membership function.

Results and Discussion

The accuracy of the developed Mamdani fuzzy model using three types of MFs in
prediction of condensate viscosity was quantified by metric quantifiers of root mean
square error (RMSE), average absolute relative deviation percentage (AARD%) and

122 | P a g e
mean absolute error (MAE) shown in equations 4.24. 4.25 and 4.61 respectively. The
performance of the developed models compared to four existing live oil viscosity
models of Bergman and Sutton, (2007), De Ghetto et al., (1994), Elsharkawy and
Alikhan, (1999) and Kartoatmodjo and Schmidt, (1994). Summary of the statistical
error analysis is presented in Table 4.9.

Table 4.9. Statistical error comparison between Mamdani fuzzy approach with three MFs and
existing literature models for prediction of condensate viscosity.

Method RMSE MAE AARD%

Bergman (2000) 0.3287 0.2635 104.65

Elsharkawy and Alikhan 0.2744 0.1935 60.64

(1999)
De Ghetto (1994) 0.2656 0.1842 57.47
Kartoatmodjo and 0.2924 0.2185 62.92
Schmidt (1994)
Mamdani approch 0.3265 0.1389 73.99
(Triangular MF)
Mamdani approch 0.2057 0.1644 46.45
(Trapezoid MF)
Mamdani method 0.0556 0.0443 17.22
(Gaussian MF)

To visualize the performance of the developed Mamdani fuzzy model for prediction of
condensate viscosity, graphical representation of the results in terms of cross plot
were generated. The aim of this graphical error analysis is to demonstrate how the
prediction deviate from zero error line or 45°, the slope. This slope line is representing
zero error and make perfect match between measured and calculated values
(Mansour et al., 2013; Aily et al., 2019). Figure 4.29A, Figure 4.30A and 4.31A depict
the cross plot of developed Mamdani fuzzy model with Gaussian, Triangular and
Trapezoidal MFs for prediction of condensate viscosity. Mamdani fuzzy models with
gaussian MFs, Figure 4.29A, predicted values of condensate viscosity close to the
diagonal line (zero error line) for all condensate viscosity ranges between 0 – 1cp.
This model outperformed other utilized techniques with lowest RMSE of 0.0556, MAE
of 0.0443 and AARD% of 17.22%.
The statistical error analysis for implemented triangular based fuzzy model, shows that
this model predicts condensate viscosity with RMSE of 0.3265, MAE of 0.1389 and
AARD% of 73.99. As can be seen from Figure 4.30A cross plot of triangular based

123 | P a g e
Mamdani fuzzy model the estimated values are not in good match with diagonal line
especially in higher condensate viscosity range (𝜇𝑐𝑜𝑛𝑑𝑒 ≥ 0.35𝑐𝑝).
Figure 4.31A demonstrates graphical error estimation of trapezoidal MF’s based
Mamdani fuzzy model for forecasting condensate viscosity. This model predicts
condensate viscosity with RMSE of 0.2057, MAE of 0.1644 and AARD% of 46.45.
Comparing the prediction performance of all three fuzzy models verify that using the
developed Mamdani fuzzy models incorporated with triangular and trapezoidal MFs
predict condensate viscosity with uncertainty. The large uncertainty of these two
models is because the triangular and trapezoidal MFs are not able to include all ranges
of condensate viscosity or all ranges of input variables (pressure, temperature and
solution gas to ratio).

(A) (B)
Figure 4.29. Performance prediction of Gaussian based MFs Mamdani fuzzy model in
predicting condensate viscosity.

(A) (B)

Figure 4.30. Performance prediction of triangular based MFs Mamdani fuzzy model in
predicting condensate viscosity.

124 | P a g e
 (A) (B)

Figure 4.31. Performance prediction of trapezoidal based MFs Mamdani fuzzy model in
predicting condensate viscosity.

To further illustrate the performance of the developed Mamdani fuzzy models for
prediction of condensate viscosity the time series graphs are generated and shown in
Figure 4.29B, Figure 4.30B and Figure 31B for gaussian, triangular and trapezoidal
based Mamdani fuzzy models respectively. These graphs highlight the contrast
between the developed Mamdani fuzzy model outputs and condensate viscosity
measured experimental values versus relevant data index. From these figures, it can
be seen that the Gaussian-based Mamdani fuzzy approach forecast the condensate
viscosity in all data indexes while other two models deviate from the experimental data.
From the graphical and statistical error analysis of this work, it can be concluded that
the fuzzy model with Gaussian membership function can predict the condensate
viscosity in depleted gas-condensate reservoirs with acceptable accuracy in compare
to triangular and trapezoidal fuzzy models and also live oil literature correlations.

To make sure the developed Mamdani fuzzy models are physically following the trend
of the condensate viscosity a set of data from Coats, (1986) was selected. The
performance of the Mamdani fuzzy with gaussian MFs as well as existing literature
correlations was tested using experimental data as a function of pressure. Figure 4.32
depicts this comparison and show satisfactory match that achieved between
developed Gaussian MFs based fuzzy model and target condensate viscosity data.
Variation of condensate viscosity with pressure can be observed from Figure 4.32,
where reducing the reservoir pressure would decrease the condensate viscosity. This
is a true characterization of condensate liquid viscosity, hence the developed fuzzy
model is valid physically for prediction of condensate liquid viscosity.

125 | P a g e
Figure 4.32. Performance of utilized literature correlations and Mamdani fuzzy model with
gaussian MFs for predicting condensate viscosity as a function of pressure; sample from
Coats, (1986).

The effect of solution gas to oil ratio (Rs) on condensate viscosity is also investigated
and shown in Figure 4.33. Confirmed by the results of Figure 4.33, it can be concluded
that Rs has negative impact on condensate viscosity in gas-condensate reservoirs,
means increasing Rs, would decrease viscosity. This is also true characteristic of
condensate liquid viscosity, as increasing solution gas to oil ratio would increase the
amount of gas phase in liquid phase, which in return make the mixture lighter with
lower viscosity. Both Figure 4.32 and Figure 4.33 indicate the validity of the developed
condensate viscosity using Mamdani fuzzy approach in following physical trend of the
experimental data as other literature correlations do.

126 | P a g e
Figure 4.33. Performance of literature correlations along with gaussian fuzzy approach in
predicting condensate viscosity as a function of solution gas to oil ratio Rs; sample from
Coats, (1986).

4.8 Computational efficiency of the developed models

The developed ML models in this chapter for prediction of condensate (oil) viscosity
of gas-condensate reservoirs are favourable to the existing literature models as they
require less parameters. ANN and LSSVM are only function of P, T, Rs, API gravity
and Gas Specific gravity. Both developed TSK and Mamdani models are function of
P, T and Rs. In terms of computational efficiency (the required time for each model to
return the output), the developed AI models predict condensate viscosity very fast as
shown in Figure 4.34. All developed AI methods predict the condensate viscosity with
less than a minute. However, Mamdani fuzzy is faster than other AI techniques with
only required 5 seconds to return the output values. Mamdani Fuzzy followed by ANN,
TSK fuzzy and LSSVM.

127 | P a g e
 00:01:09
ANN LSSVM TSK Fuzzy Mamdani Fuzzy
00:01:00
00:00:52
00:00:43
Time

00:00:35
00:00:26
00:00:17
00:00:09
00:00:00
Developed AI model

Figure. 4.34. Computational efficiency of the developed AI models (00:00:00 stands for
Hour: Minute: Second).

4.9 Summary

Viscosity is one of the governing parameters that influence well deliverability modelling
of gas-condensate reservoirs. This important PVT property should be treated
separately and with high precision for both phases of gas and condensate (liquid)
below the saturation pressure.
In this chapter, extensive effort has been made for accurate modelling of viscosity of
both phases (gas and condensate). The applicability of existing literature models for
prediction of viscosity of gas and condensate (oil) has been examined using published
experimental data. The results show that the performance of existing literature models
for prediction of gas and condensate (oil) viscosity are inadequate. In fact, for
prediction of condensate (oil) viscosity existing literature models show significant error.
To improve the performance of the existing literature models in prediction of gas phase
viscosity non-linear regression was performed and a well-known method of Londono
et al., (2002) was modified. New optimized Londono et al., (2002) model is predicting
experimental data with 5.2% average absolute relative deviation (AARD%).
Variety of machine learning (ML) as well as non-linear regression approaches were
utilized for better modelling of condensate viscosity.

128 | P a g e
Initially non-linear regression employed for optimizing several existing literature
models of Beggs and Robinson, (1975); Bergman and Sutton, (2007); De Ghetto et
al., (1994); Elsharkawy and Alikhan, (1999); Kartoatmodjo and Schmidt, (1991) and
LBC, 1964). Then, several smart approaches including Support Vector Machine
(SVM), Artificial Neural Network (ANN), TSK fuzzy logic, and Mamdani fuzzy logic
were employed for better modelling of condensate viscosity in gas-condensate
reservoirs.
Statistical and graphical error analysis confirmed the superiority of ML based
approaches in prediction of condensate viscosity in comparison to the existing
literature models. A unique condensate viscosity correlation as a function of pressure,
temperature and solution gas to oil ratio was proposed using TSK fuzzy approach.
Physical accuracy of the developed machine learning models in this chapter for
prediction of condensate viscosity is verified using trend analysis of independent data
sets.
The results of present work could be utilized for quick estimation of the gas-
condensate viscosity below the dew point pressure. These models offers
computational efficiency, high accuracy and require less parameters for predicting
condensate viscosity. The developed models can be embedded inside PVT packages
and other simulation software currently in use in industry for better well deliverability
modelling of gas-condensate reservoirs.

The developed ANN and LSSVM are predicting the experimental condensate viscosity
with 0.1144 and 0.1208 RMSE, respectively as well as AARD% of 16.20 and 17.20.

129 | P a g e
 CHAPTER 5 DEVELOPMENT OF TWO-
PHASE Z FACTOR OF GAS-CONDENSATE MIXTURE
DEVELOPMENT OF TWO-PHASE Z FACTOR OF GAS-
CONDENSATE MIXTURE

5.1 Introduction

The development and optimization of gas-condensate recovery and accurate well

deliverability modelling require accurate determination of PVT properties. One of the
most important factors that need accurate determination by engineers in calculation of
well flow rate through reservoir rock, material balance calculations, evaluation of gas
reserves, design of production equipment, and planning the development of a gas-
condensate reservoir is the fluid Z factor (Elsharkawy and Foda, 1998; Sun et al.,
2012; Zendehboudi et al., 2012). As explained in chapter 2, Z factor is one of the
fundamental PVT properties in gas rate equations of 2.2 and 2.4. Furthermore, as
described in chapter 2, Z factor is influencing the well deliverability calculation in gas
condensate reservoirs.
In calculation of gas-condensate well deliverability accurate estimation of Z factor, is
critical for the following reasons. Firstly, accurate and consistent estimation of initial
hydrocarbon fluid in place require accurate estimation of compressibility factor.
Secondly, correct estimation of Z factor dictates accurate prediction of recovery factor
(gas and condensate liquid) as a function of pressure in depletion drive gas-
condensate reservoirs (Whitson, Fevang and Yang, 1999).
Furthermore, at average reservoir pressure above the dew point condensate recovery
is exactly equal to the gas recovery. Consequently, condensate recovery is strongly
dependant on accurate description of Z factor both above and below the dew point
pressure. To emphasise the important of accurate estimation of compressibility factor
in gas-condensate well deliverability modelling Whitson et al., (1999) quoted; “A “+5%
error in initial compressibility factor and a –5% error in compressibility factor at the
dewpoint will result in (a) a +5% error in initial gas and initial condensate in place, and
(b) a +5 to +10% error in recovery of gas and condensate at the dewpoint”.

130 | P a g e
In this chapter, primarily the fundamental theories behind the Z factor will be
discussed. Then the performance accuracy of existing literature models for estimation
of gas-condensate Z factor will be reviewed. Finally, several Z factor models are
proposed to ensure accurate estimation of gas-condensate reservoir Z factor below
the saturation pressure.

5.2 Preliminary theory of Z factor calculation

In 1873, Johannes Diderik van der Waals recognized a unique relationship between Z
factor, reduced pressure (Pr) and reduced temperature (Tr), which known as two
parameters corresponding state principle. The corresponding state principle is
generalization of properties of gases (pressure and temperature) and are related to
the critical properties in universal way. It describes that the two substances at the
same condition of reduced pressure (Pr) and reduced temperature (Tr) have similar
properties. In other word Z factor of any pure gases at same Pr and Tr are the same.
For mixture compositions such as gas-condensate fluid, the reduced properties
replaced by pseudocritical properties and become pseudo-reduced pressure (Ppr) and
pseudo-reduced temperature (Tpr). These values are serving as a correlating
parameters for corresponding state principle and the way they are estimated effect the
accuracy of the Z factor (Sutton, 2005).
Prior to discuss computation of Z factor, it is important to understand the calculation
of pseudoreduced properties. The pseudo-reduced pressure and pseudo-reduced
temperature are expressed by the following equations:
𝑃
𝑃𝑝𝑟 =
𝑃𝑝𝑐
5.1
𝑇
𝑇𝑝𝑟 =
{ 𝑇𝑝𝑐
Where Ppc is pseudocritical pressure and Tpc stands for pseudocritical temperature.
These pseudocritical properties can be estimated with gas compositions and mixing
rules or from correlations based on gas gravity (Whitson and Brulé, 2000, pp. 23–24;
Sutton, 2005b). Several correlations have been proposed in literature for calculation
of pseudocritical properties (Ppc and Tpc) if the composition of the mixture is available
(Kay, 1936; Stewart, Burkhardt and Voo, 1959; Sutton, 1985; Piper, McCain and
Corredor, 1993; Elsharkawy, 2006). Based on gas specific gravity there are other
correlations in literature for estimation of pseudocritical properties (Matthews and
131 | P a g e
Roland, 1942; Standing, 1981; Elsharkawy, Hashem and Alikhan, 2000; Elsharkawy
and Elkamel, 2001; Londono, Archer and Blasingame, 2005; Sutton, 2007).
The Z factor data bank that is collected and used in this chapter include all
compositional details of gas-condensate mixtures. This would allow us to use widely
accepted Kay’s mixing rule (Kay, 1936) shown in following equation for estimating
critical properties of hydrocarbon components.
𝑁

𝑃𝑝𝑐 = ∑ 𝑦𝑖 𝑃𝑐𝑖
𝑖=1,2,..
𝑁 5.2

𝑇𝑝𝑐 = ∑ 𝑦𝑖 𝑇𝑐𝑖
{ 𝑖=1,2,..

Where y is mole fraction of component i in the mixture. Gas-condensate mixture

frequently contains non-hydrocarbon impurities such as hydrogen sulphide (H2S),
nitrogen (N2) and carbon dioxide (CO2). These impurities influence the accuracy of the
PVT properties including Z factor estimation and need to be considered in calculation
of pseudocritical properties. Wichert and Aziz, (1972) proposed a correction to include
non-hydrocarbon impurities for calculation of pseudocritical properties. Wichert and
Aziz’s correlation has widely been accepted in industry; hence, in this study we
employed same method for estimating non-hydrocarbon impurities in gas-condensate
mixture. Wichert and Aziz, (1972) presented their equation in following form.
𝑇𝑝𝑐 ∗ = 𝑇𝑝𝑐 −∈
𝑃𝑝𝑐 (𝑇𝑝𝑐 −∈)
𝑃𝑝𝑐 ∗ = 5.3
𝑇𝑝𝑐 ∗ + 𝑦𝐻2 𝑆 (1 − 𝑦𝐻2 𝑆 ) ∈
0.9 1.6
{∈= 120 [(𝑦𝐶𝑂2+ 𝑦𝐻2 𝑆 ) − (𝑦𝐶𝑂2 + 𝑦𝐻2 𝑆 ) ] + 15(𝑦𝐶𝑂2 0.5 − 𝑦𝐻2 𝑆 4 )

Where Tpc and Ppc are mixture pseudocritical properties based on Kay’s mixing rule.
In addition to non-hydrocarbon impurities one of the important properties that influence
the calculation of Z factor in gas-condensate reservoirs below the saturation pressure,
is accurate determination of hydrocarbon plus (C7+) pseudocritical properties. Many
correlations such as Lee and Kesler, (1975); Kesler and Lee, (1976); Rowe, (1978);
Whitson, (1983); Pedersen, Thomassen and Fredenslund, (1984); Riazi and Daubert,
(1987); Elsharkawy, Hashem and Alikhan, (2000); Sutton, (2005b) have been
proposed for this purpose. Many of these methods are applicable for characterization

132 | P a g e
of C7+ fractions when compositional data is not available. If compositional data is
available Matthews and Roland, (1942) is conventional method for estimating
pseudocritical properties of C7+. This method is practical and widely used in industry
for estimation of hydrocarbon plus pseudocritical properties (Sutton, 1985; Whitson
and Brulé, 2000; Maravi, 2003).
Due to availability of compositional data in this study Matthews and Roland, (1942)
shown in equation 5.4 utilized for estimating pseudocritical properties of C7+.Their
method calculates hydrocarbon plus pseudocritical properties as a function of
hydrocarbon plus specific gravity (𝛾𝐶7+ ) and molecular weight (𝑀𝐶7+ ).

𝑇𝑐𝐶7+ = 608 + 364𝑙𝑜𝑔(𝑀𝐶7+ − 71.2) + (2450𝑙𝑜𝑔𝑀𝐶7+ − 3800)𝑙𝑜𝑔𝛾𝐶7+

{ 𝑃𝑐𝐶7+ = 1188 − 431𝑙𝑜𝑔(𝑀𝐶7+ − 61.1) + ⋯ 5.4
… [2319 − 852𝑙𝑜𝑔(𝑀𝐶7+ − 53.7)](𝛾𝐶7+ − 0.8)

5.3 Assessment of Z factor models

Based on the theory of corresponding state principle Standing and Katz, (1942)
introduced a chart for estimating Z factor of natural dry and sweet gases as a function
of pseudo reduced pressure (Ppr) and pseudo reduced temperature (Tpr). Standing
and Katz (SK) chart is one of the most widely accepted practical correlation in
petroleum engineering for calculating gas-phase Z factor. The data of binary mixture
of methane with propane, ethane, butane and natural gases with wide range of
composition was used in developing SK chart. The molecular weight of the mixtures
used in SK chart not exceeding 40 (Elsharkawy, Hashem and Alikhan, 2000). The SK
chart is suitable for estimating Z factor of dry gases. Several corrective methods
introduced to SK chart to account for existence of high molecular weight gases and
hydrocarbon plus fraction C7+ (Stewart, Burkhardt and Voo, 1959; Sutton, 1985,
2005b; Elsharkawy, Hashem and Alikhan, 2000). The importance of SK chart in
industry motivated scholars to develop mathematical representation of the chart.
Some of these methods are Dranchuk, Purvis and Robinson, (1973); Hall and
Yarborough, (1973); Dranchuk and Abou-Kassem, (1974); Yarborough and Hall,
(1974). Engineering community typically uses the published methods by Hall and
Yarborough, (1973) and Dranchuk and Abou-Kassem, (1975) as they fit best the SK
chart (Takacs, 1976). These methods are form of an equation of state that have been
derived in an iterative procedure to estimate the Z factor as a function of P pr and Tpr.

133 | P a g e
The Ppr and Tpr are determined using one of the available mixing rules for instance
Kay, (1936).
Other scholars also tried to develop methods for direct calculation of SK chart, which
using fitting techniques (e.g., regression). Some of these methods in literature are as
follow (Papay, 1968; Beggs and Brill, 1973; Gopal, 1977; Azizi, Behbahani and
Isazadeh, 2010; Heidaryan, Moghadasi and Rahimi, 2010; Sanjari and Lay, 2012).
These correlations are explicit, which means they don’t need numerical iterative
procedure. These correlations recently become popular and they have been utilized
by many researchers in the field for calculation of Z factor due to their simplicity and
acceptable accuracy in wide range of pressure and temperature condition.

Another methodology for estimation of Z factor is through implementation of cubic

equation of state (EOS). Cubic EOS are simple computation procedure that define
relationship between pressure, volume and temperature of pure hydrocarbon
components and mixtures components. First cubic EOS proposed by Van der Waals,
(1873). Many equations of state after Van der Waals have been proposed in literature
to model reservoir fluid phase behaviour and Z factor (Carnahan and Starling, 1969;
Fuller, 1976; Lawal, 1999; Nasrifar and Moshfeghian, 2001). Redlich and Kwong,
(1949) improved initial Van der Waals EOS for more accurate calculation of vapour
phase Z factor. Evolutionary path of van der Waals EOS lead to proposing an excellent
form of EOS by Soave-Redlich and Kwong (SRK) in 1972. SRK and later Peng and
Robinson, (1976) EOSs become the industry leader for calculation of PVT properties
(e.g., Z factor). Generally performance of EOS are good for simple hydrocarbon
systems like natural gas and black oil reservoirs, however their performance
deteriorates for more complex systems like volatile oil and gas-condensate reservoirs
(Sarkar, Danesh and Todd, 1991; Elsharkawy and Foda, 1998; Ghiasi et al., 2014).
This is because the interaction between the gas and liquid molecules is not well
simulated in critical conditions.
Availability and accessibility of the data in oil and gas industry motivated many
scholars to use different approach so called intelligent models for prediction of various
fluid properties. Recently giant oil and gas companies such as British Petroleum (BP)
and Shell have supported using artificial intelligent (AI) techniques in different area of
their operations. Many researchers are also using AI techniques for prediction of PVT
properties in recent years. A few examples are (Heidaryan, Moghadasi and Rahimi,
134 | P a g e
2010; Chamkalani et al., 2013; Hatampour and Ghiasi-Freez, 2013; Ghiasi et al.,
2014). The intelligent methods provide promising breakthrough in accurate prediction
of PVT properties.
The core objectives of this chapter can be separated to two parts. Part 1 is to assess
previously developed methods for estimation of two-phase Z factor in gas-condensate
reservoirs. Second part is to develop smart, yet accurate models for prediction of gas-
condensate two-phase Z factor for wide range of reduced pressure and temperature
when composition of the mixture is available. An extensive data bank from the
literature was collected to achieve the objective of this chapter. In the following
sections, the most frequent methods in literature for computation of Z factor are briefly
introduced.

Hall-Yarborough (1973)

Success of Standing and Katz (SK) chart in petroleum industry for estimating Z factor,
motivated many scholars to reproduce the SK chart in a mathematical (equation) form.
Hall and Yarborough, (1973) and Yarborough and Hall, (1974) were among the first
researchers that start this journey. They modified Carnahan and Starling, (1969)
equation of state that accurately describes SK Z factor chart. The coefficient of their
equation estimated by fitting the data obtained from SK chart. Hall and Yarborough,
(1973) and Yarborough and Hall, (1974) propose the following equation for estimating
of Z factor.

0.06125𝑃𝑝𝑟 𝑡
𝑍=[ ] 𝑒𝑥𝑝[−1.2(1 − 𝑡 2 )] 5.5
𝑌

Where Ppr is pseudo-reduced pressure; t is reciprocal of pseudo-reduced temperature

(e.g., Tpc/T) and Y is the product of the Van der Walls co-volume and reduced-density
and can be obtained by solving F(Y) as follow.

𝑌 + 𝑌2 + 𝑌3 + 𝑌4
𝐹(𝑌) = 𝑋1 + − 𝑋2 𝑌 2 + 𝑋3 𝑌𝑋4 5.6
(1 − 𝑌)3

Where the parameters of X are determined as follow.

135 | P a g e
 𝑋1 = −0.01625Ppr 𝑡[−1.2(1 − t 2 )]
𝑋2 = (14.76𝑡 − 9.76𝑡 2 + 4.58𝑡 3 )
5.7
𝑋3 = (90.7𝑡 − 242.2𝑡 2 + 42.48𝑡 3 )
{ 𝑋4 = (2.18 + 2.82𝑡)

To solve function F(Y) an iterative technique of Newton-Raphson with the initial guess
of Y=0.001 is used. The method returned the experimental Z factor within 3 to 10
iterations.
Hall – Yarborough method is representing SK chart Z factor with very good accuracy
in wide range of pressure and temperature. This method is valid for the reduced
temperature in the range of 1 ≤ 𝑇𝑟 ≤ 3 and reduced pressure in the range of 0.2 ≤
𝑃𝑟 ≤ 25 − 30 (Yarborough and Hall, 1974; Whitson and Brulé, 2000; Ahmed, 2010).

Dranchuk – Abu – Kassem (1975)

Unlike Hall and Yarborough Z factor equation, which is only function of pseudoreduced
pressure and temperature, Dranchuk and Abou-Kassem, (1975) or known as DAK
proposed an equation of state for calculation of Z factor that also include reduced gas
density. Reduced gas density (𝜌𝑟 ) is a ratio of gas density at specific pressure and
temperature over gas density at critical pressure and temperature. DAK used 1500 Z
factor data points obtained from SK chart in developing their method. DAK proposed
Z factor equation with 11 constants as follow.
𝐴2 𝐴3 𝐴4 𝐴5 𝐴7 𝐴8
𝑍 = [𝐴1 + + 3
+ 4
+ 5
] 𝜌𝑟 + [𝐴6 + + 2
] 𝜌𝑟2 …
𝑇𝑝𝑟 𝑇𝑝𝑟 𝑇𝑝𝑟 𝑇𝑝𝑟 𝑇𝑝𝑟 𝑇𝑝𝑟
2
𝐴7 𝐴8 𝜌𝑟
… − 𝐴9 [ + 2
] 𝜌𝑝𝑟 5 + 𝐴10 (1 + 𝐴11 𝜌𝑟 2 ) 3𝑒𝑥𝑝[−𝐴11 𝜌𝑟 2 ] + 1 5.8
𝑇𝑝𝑟 𝑇𝑝𝑟 𝑇𝑝𝑟 3
0.27𝑃𝑃𝑟
𝜌𝑟 =
{ 𝑍𝑇𝑝𝑟

Where the constants of A1 to A11 are as follow, 𝐴1 = 0.3265, 𝐴2 = −1.0700, 𝐴3 =

−0.5339, 𝐴4 = 0.01569, 𝐴5 = −0.05165, 𝐴6 = 0.5475, 𝐴7 = −0.7361, 𝐴8 = 0.1844, 𝐴9 =
0.1056, 𝐴10 = 0.6134, 𝐴11 = 0.7210. 𝜌𝑟 is the reduced density of the mixture.
A1 to A11 constants obtained by the authors using non-linear regression using SK chart
data points. In implementing DAK method for estimating Z factor, Newton – Raphson
iterative technique has been used (Lee and Wattenbarger, 1995). For high-density
gases DAK method estimates Z factor with high error. Hence the modification has

136 | P a g e
been introduced to the original DAK code by Borges, (1991), where he sets the
reduced density from value of 2.2 to 3 or greater. The DAK method predicts Z factor
for reduced temperature in the range of 1 ≤ 𝑇𝑟 ≤ 3 and reduced pressure range of
0.2 ≤ 𝑃𝑟 ≤ 25 − 30.

Beggs and Brill (1973)

Although Hall – Yarborough and DAK method simulate SK Z factor chart with good
accuracy, however the computation procedure is tedious and iterative techniques are
needed. For easing the computation of the Z factor without using iterative procedure
Beggs and Brill, (1973) proposed a direct correlation. For many petroleum engineering
applications Beggs and Brill, (1973) give satisfactory representation of the SK chart
with 1 – 2% deviation in reduced temperature range of 1.2 < 𝑇𝑟 < 2. The main
limitations of Beggs and Brill are that reduced temperature must be greater than 1.2
(≈80°F) and less than 2.0 (≈340°F) and reduced pressure should be less than 15
(≈10,000psia). The Beggs and Brill is presented as follow for calculation of Z factor.

1−𝐴
𝑍 =𝐴+ + 𝐶𝑃𝑝𝑟 𝐷 5.9
𝑒𝐵

Where the parameters of A, B C and D can be determined as follow:

0.5
𝐴 = 1.39(𝑇𝑝𝑟 − 0.92) − 0.36𝑇𝑝𝑟 − 0.10
0.066 0.32𝑃𝑝𝑟 6
𝐵 = (0.62 − 0.23𝑇𝑝𝑟 )𝑃𝑝𝑟 + ( − 0.037) 𝑃𝑝𝑟 2 +
𝑇𝑝𝑟 − 0.86 10𝐸
𝐶 = 0.132 − 0.32log(𝑇𝑝𝑟 ) 5.10
𝐷 = 10𝐹
𝐸 = 9(𝑇𝑝𝑟 − 1)
{ 𝐹 = 0.3106 − 0.49𝑇𝑝𝑟 + 0.1824𝑇𝑝𝑟 2

The main advantage of Beggs and Brill is that the equation is explicit and easy to use
in many engineering calculation.

Rayes et al., (1992)

SK chart and subsequent models of Hall – Yarbrough, DAK and Beggs and Brill,
(1973) are good representation of Z factor as for natural gas with single-phase

137 | P a g e
behaviour. Gas-condensate reservoirs undergoing depletion are developing
condensate phase, which make the flow at least two-phase or even more if connate
water enter the pay zone. Therefore using single-phase Z factor for this type of
reservoirs is arguable.
Evolving condensate (liquid) phase below the saturation pressure in gas-condensate
reservoirs motivated Rayes et al., (1992) to develop a two-phase Z factor as a function
of pseudocritical properties of temperature and pressure. They used 131 constant
volume depletion (CVD) studies of gas-condensate reservoirs for developing their
correlation. Their correlation is valid for reduced temperature range between 1.1 <
𝑇𝑟 < 2.1 and reduced pressure range of 0.7 < 𝑃𝑟 < 20. They propose the following
equation for computation of two-phase Z factor.

1 2
1 2 𝑃𝑟
𝑍2−𝑝ℎ𝑎𝑠𝑒 = 𝐴0 + 𝐴1 (𝑃𝑟) + 𝐴2 ( ) + 𝐴3 (𝑃𝑟) + 𝐴4 ( ) + 𝐴5 ( ) 5.11
𝑇𝑟 𝑇𝑟 𝑇𝑟

Where 𝐴0 = 2.24353, 𝐴1 = −0.0375281, 𝐴2 = −3.56539, 𝐴3 = 0.000829231, 𝐴4 =

1.53428 and 𝐴5 = 0.131987.
Rayes et al., (1992) is the only two-phase Z factor model that specifically developed
for gas-condensate reservoirs. The equation can be calculated explicitly for
determining two-phase Z factor. However, the equation is limited for high pressure and
high temperature (HPHT) condition.

Azizi et al., (2010)

The methods that developed to representing SK chart such as DAK and Hall –
Yarbrough were based on limited data from the original SK chart. In another attempt
to simulate the SK chart by a mathematical equation, Azizi et al., (2010) used a large
Z factor data bank of 3038 points obtained from SK chart and proposed a direct Z
factor correlation. Their proposed method include 14 constants obtained from fitting
with SK chart Z factor data. They proposed their Z factor correlation in the following
fashion.

𝐵+𝐶
𝑍 =𝐴+ 5.12
𝐷+𝐸

138 | P a g e
Where A, B, C and D are determined from the following equations.
𝐴 = 𝑎𝑇𝑟 2.16 + 𝑏𝑃𝑟 1.028 + 𝑐𝑃𝑟 1.58 𝑇𝑟 −2.1 + 𝑑𝑙𝑛(𝑇𝑟 )−0.5
𝐵 = 𝑒 + 𝑓(𝑇𝑟 )2.4 + 𝑔𝑃𝑟 1.56 + ℎ𝑃𝑟 0.124 𝑇𝑟 3.033
𝐶 = 𝑖𝑙𝑛(𝑇𝑟 )−1.28 + 𝑗𝑙𝑛(𝑇𝑟 )1.37 + 𝑘𝑙𝑛(𝑃𝑟) + 𝑙𝑙𝑛(𝑃𝑟 )2 + 𝑚𝑙𝑛(𝑃𝑟 )𝑙𝑛(𝑇𝑟 ) 5.13
𝐷 = 1 + 𝑛(𝑇𝑟 )5.55 + 𝑜(𝑃𝑟 )0.68 (𝑇𝑟 )0.33
{ 𝐸 = 𝑝𝑙𝑛(𝑇𝑟 )1.18 + 𝑞𝑙𝑛(𝑇𝑟 )2.1 + 𝑟𝑙𝑛(𝑃𝑟) + 𝑠(𝑃𝑟 )2 + 𝑡𝑙𝑛(𝑃𝑟 )𝑙𝑛(𝑇𝑟 )

The tuned values of 𝑎 − 𝑡 in equation 5.13 are given in following table.

Table 5.1. Coefficients of Azizi et al., (2010) Z factor correlation.

Coefficient Value Coefficient Value

a 0.0373142485385592 k −24449114791.15
b −0.0140807151485369 l 19357955749.32
c 0.0163263245387186 m −126354717916.60
d −0.0307776478819813 n 623705678.38
e 13843575480.94 o 17997651104.33
e −16799138540.76 p 151211393445.06
g 1624178942.649 Q 139474437997.17
h 13702270281.08 r −24233012984.09
i −41645509.89 s 18938047327.52
j 237249967625.01 t −141401620722.68

The advantage of Azizi et al., (2010) method over other DAK and Hall-Yarborough for
simulating SK chart is that it does not need an iterative computational method for
estimating Z factor. It is a direct correlation that developed to represents SK chart as
simple as possible. This model is valid for reduced pressure between 0.2 < 𝑃𝑟 < 11
and reduced temperature between 1.2 < 𝑇𝑟 < 2.
To assess the applicability of all studied literature model for estimating two-phase Z
factor of gas-condensate fluid below the saturation pressure a large data bank has
been collected, which is discussed in following section.

Data bank

To examine the performance of the discussed literature methods in prediction of gas-

condensate Z factor and also developing new methods for this important PVT property,

139 | P a g e
a data bank was collected. The data bank include 1084 data sets (19518 data points).
This is largest and one of the most comprehensive data bank to assess and improve
accuracy of gas-condensate reservoirs Z factor. Experimental studies of gas-
condensate fluids and also the PVT reports are source of our data bank. The following
references is used in sourcing the Z factor data bank: (Olds, Sage and Lacey, 1945,
1949; Whitson and Torp, 1983; Coats, 1985; Coats and Smart, 1986; Moses, 1986;
Kenyon and Behie, 1987; Drohm, Goldthorpe and Trengove, 1988; Yang, Chen and
Guo, 1997; Elsharkawy and Foda, 1998; Sun et al., 2012; Liu et al., 2013; Bonyadi et
al., 2014). In most of these studies constant volume depletion (CVD) has been
employed in order to estimate the phase behaviour of gas-condensate samples. The
data bank includes hydrocarbon compositions, molecular weight of C 7+, pressure,
temperature, gas specific gravity and two-phase Z factor. The data covers wide range
of reservoir conditions as well as reservoir compositions from lean gas-condensate to
very rich obtained from worldwide. The statistical description of the data is provided in
Table 5.2.

140 | P a g e
Table 5.2. Statistical description of the data used for development of two-phase Z factor.

Property Minimum Maximum Average

Temperature, ⁰F 86 351 207
Tpr 1.1631 4.261 2.901
Pressure, psia 200 23244 8305
Ppr, (mole %) 0.2878 33.525 11.917
H2S, (mole %) 0.000 0.5137 0.0075
CO2, (mole %) 0.0001 0.749 0.0472
N2, (mole %) 0.000 0.525 0.0326
C1, (mole %) 0.0687 96.003 5.1115
C2, (mole %) 0.0024 2.101 0.1529
C3, (mole %) 0.0007 0.202 0.028
IC4, (mole %) 0.0002 0.0638 0.0065
NC4, (mole %) 0.00 0.0638 0.0084
IC5, (mole %) 0.00 0.06 0.0037
NC5, (mole %) 0.00 0.0431 0.0032
C6, (mole %) 0.0002 0.0592 0.0047
C7+, (mole %) 0.0004 0.228 0.0249
SG gas 0.046717 1.410137 0.26218
MW C7+ 102.30 253.00 186.78
Two-phase Z-factor 0.553 2.162 1.287

Results of literature models

Using the data bank shown in Table 5.2 applicability and accuracy of the existing
literature models for prediction of gas-condensate Z factor is assessed. Statistical and
graphical error analysis carried out to see the performance of the literature models in
prediction of two-phase Z factor. The statistical parameters of mean average error
(MAE), root mean square error (RMSE) and absolute average relative deviation
percentage (AARD%) have been employed to see the accuracy of the utilized
methods. The results of this statistical error analysis shown in Table 5.3. These
analyses indicate DAK method best performed among existing literature models with
MAE of 0.1300, RMSE of 0.1671 and AARD% of 2.4945. DAK method followed by
Hall-Yarborough, (1973), Rayes et al., (1992), Beggs and Brill, (1973) and Azizi et al.,
(2010) for estimation of gas-condensate Z factor below the saturation pressure.

141 | P a g e
Table 5.3. Performance of employed literature models in predicting two-phase Z factor of
gas-condensate systems.
Model MAE RMSE AARD%
Hall-
Yarborough, 0.1421 0.1846 3.0906
(1973)

DAK, (1975) 0.1300 0.1671 2.4945

Beggs and Brill,

0.2047 0.2769 17.0566
(1973)

Azizi et al., 0.3851 0.4822 28.6958

(2010)

Rayes et al., 0.1516 0.1904 10.0522

(1992)

In order to visualize the performance of literature models several graphical error

analysis were performed. The graphs in Figure 5.1 are presenting cross plot of the
obtained results against the experimental values of two-phase Z factor. The diagonal
line is representing zero error line, means scattering the data over or under this line
showing poor performance of the model. The results of cross plot show that DAK,
(1975), Hall-Yarborough, (1973) and Rayes et al., (1992) predict the expariemntal
values of two-phase Z factor with some level of accuracy while Beggs and Brill, (1973)
and Azizi et al., (2010) performance is unsatisfactory.

(A) (B)

142 | P a g e
 (C) (D)

(E)

Figure 5.1. Scatter plot of five literature models in prediction of two-phase compressibility
factor of gas condensate reservoirs below the saturation pressure.

To investigate the performance of each method in predicting various range of

experimental gas-condenste two-phase Z factor the graph of relative error was
generated and presented in Figure 5.2. As the graph shows prediction performance of
DAK, (1975) and Hall-Yarborough, (1973) methods are poor especially in the specific
range of 1.5 < 𝑍𝑡𝑤𝑜−𝑝ℎ𝑎𝑠𝑒 < 0.7. The reasons for poor performance of both
aforementioned methods in predicting gas-condensate two-phase Z factor could be
behind their initial development. Hall-Yarborough (HY), (1973) and DAK, (1975) are
both iterative techniques that are representing mathematical description of well-known
SK (Standing and Katz, 1942) Z factor chart. The associated constants in DAK and
HY equations were tuned using the data obtained from SK chart. SK chart is suitable
for prediction of single phase sweet and dry gas systems, nevertheless not for
prediction of two-phase Z factor of gas-condensate fluid (Rayes et al., 1992;
Elsharkawy, Hashem and Alikhan, 2000). Both DAK and HY also shown similar

143 | P a g e
behaviour for prediction of two-phase Z factor, as they initially were developed to
simulate the SK chart.

(A) (B)

(C) (D)

(E)
Figure 5.2. Graph of relative error utilizing literature models for prediction of two-phase Z
factor.

Azizi et al., (2010) performance is also associated with high error as indicated by the
results in Figure 5.1D and Figure 5.2D. In developing their correlation Azizi et al.,
(2010) used 3038 data points obtained from SK chart. The poor performance of this

144 | P a g e
correlation also indicates that SK chart is not applicable for estimating two-phase Z
factor of gas-condensate reservoir below the saturation pressure (two-phase Z factor).

Beggs and Brill, (1973) correlation also predicts the experimental data with large error
as can be seen from Figure 5.1C and Figure 5.2C. Rayes et al., (1992) method is the
only method that specifically developed for prediction of gas-condensate Z factor
below the dew point pressure. This method also return the experimental data with high
error shown in Figure 5.1E and Figure 5.2E.
For fair evaluation of each literature models it should be highlighted that each method
developed for specific range of Tpr and Ppr. Hence, the performance of the each
method plotted against various range of Tpr and Ppr and shown in Figure 5.3 and Figure
5.4.
Hall-Yarborough and DAK methods well predict Z factor of reservoir gases (dry and
sweet) within the range of 1 ≤ 𝑇𝑝𝑟 ≤ 3 and 0.2 ≤ 𝑃𝑝𝑟 ≤ 25 − 30 (Hall and Yarborough,
1973; Dranchuk and Abou-Kassem, 1974; Whitson and Brulé, 2000). The results in
Figure (5.3A-B) and Figure (5.4A-B) show that even within aforementioned ranges of
reduced pressure and temperature Hall-Yarbrough and DAK predict two-phase Z
factor of gas-condensate reservoirs with high error.

Beggs and Brill Z factor correlation is only valid for specific range of pseudo reduced
temperature of 1.2 < 𝑇𝑝𝑟 < 2 and pressure of 𝑃𝑝𝑟 < 15. This correlation forecasts
the two-phase gas-condensate Z factor within the aforementioned T pr and Ppr with
large deviation as can be seen from Figure 5.3C and Figure 5.4C.

Azizi, et al, (2010) model is valid for reduced temperature between 0.2 < 𝑃𝑟 < 11 and
reduced temperature between 1.2 < 𝑇𝑟 < 2. Looking at the results in Figure 5.12D
and Figure 5.3D where performance of Azizi’s model in specified reduced pressure
and temperature depicted respectively, the two-phase Z factor experimental data is
highly overestimated.
Rayes et al., (1992) performs very well if the Z factor is within the range of 0.704 <
𝑍𝑡𝑤𝑜−𝑝ℎ𝑎𝑠𝑒 < 1.775, and deviate under and beyond these range. In terms of relative
error as a function of Tpr and Ppr, Rayes et al., (1992) method returns two-phase

145 | P a g e
experimental Z factor with unsatisfactory accuracy, as illustrated by Figure 5.3E and
Figure 5.4E for reduced pressure and reduced temperature respectively.
We should also highlight the fact that another reason for high error by utilized methods
are using different reservoirs fluid samples in their development. Based on the
presented results it can be concluded that using existing literature methods for
calculation of two-phase Z factor of gas-condensate reservoirs are not adequate and
associated with large error. The unsatisfactory performance and applicability to only
limited range of data of literature models motivated to focus on developing methods
for prediction of two-phase Z factor of gas-condensate reservoirs. There is a need for
robust models that cope with non-linearity of gas-condensate systems below the
saturation pressure in relation to prediction of Z factor. Availability of the data in
literature and success of using machine learning techniques in recent years, motivated
us to use several methods known as machine learning (ML) techniques. A
comprehensive background information for ML techniques are provided in chapter 4.4
of this thesis.

(A)
(B)

(D)
(C)

146 | P a g e
 (E)
Figure 5.3. Residual plot of relative error percentage of utilized literature correlations in
predicting gas-condensate two-phase Z factor as a function of pseudo reduced pressure (Ppr).

(A) (B)

(C) (D)

147 | P a g e
 (E)
Figure 5.4. Residual plot of relative error percentage of literature models in predicting two-
phase Z factor of gas-condensate systems as a function of pseudo reduced temperature (Tpr).

5.4 Development of new Z factor models

One of the primary aim of this study is to use innovative and different approach for
modelling various aspect of gas-condensate well deliverability modelling. As it can be
seen from chapter 4 of this thesis, a new modern numerical techniques known as
machine learning (ML) extensively used to achieve our primary aim of the study. For
modelling gas-condensate two-phase Z factor below the saturation pressure, various
ML based techniques were implemented. The reason for implementing ML based
approaches in estimating PVT properties are given as follow.
 The correlations for estimating properties such as Z factor are limited to various
range of operational conditions (temperature and pressure).
 Based on the presented figures and plots in previous section it is concluded
that using existing literature methods for calculating two-phase gas-condensate
Z factor associated with large error.
The developed ML techniques should tackle the above mentioned challenges that
exist using correlations for estimating Z factor. For this purpose two artificial neural
networks models namely feed forward neural network (FFNN) and cascade forward
neural network (CFNN) have been employed for better estimation of gas-condensate
Z factor. Furthermore, another form of TSK fuzzy model known as adaptive neuro
fuzzy inference system (ANFIS) also has been utilized. Based on ANFIS modelling a
set of gas-condensate two-phase Z factor correlations were proposed. In this part the
methodology of each utilized ML based approaches is discussed in details.

148 | P a g e
 Cascade Forward Neural Network (CFNN)

One of the class of neural networks alongside feed forward neural network (FFNN),
are cascade forward neural network (CFNN). This network also known as cascade
forward back propagation (CFBP). CFBP algorithm resembles FFNN and the only
different is that each individual neurons in input layer attached to each neuron in
hidden layer and to each neuron in the output layer (Warsito, Santoso and Yasin,
2018). This means each subsequent neuron is bound to the previous one and training
is performed accordingly. The typical CFNN for one input and one hidden layer is
shown in Figure 5.5, where the neurons in input layer is connected to the output layer
activation function. This additional connection of the input and output is the main
difference between the conventional FFNN and CFNN.

Figure 5.5. Architecture of the cascade forward (CFNN) neural network used in this study.

The relation between input and output layer of CFNN with one bias unit shown in
Figure. 5.5 can be defined as follow.

𝑛 𝑘 𝑛

𝑦(𝑥) = ∑ 𝑔𝑖 Ɵ𝑖𝑖 𝑥𝑖 + 𝑔𝑜 ( Ɵ𝑏 + ∑ Ɵ𝑗𝑜 𝑔ℎ (Ɵ𝑗𝑏 + ∑ Ɵ𝑗𝑖ℎ 𝑥𝑖 )) 5.14

𝑖=1 𝑗=1 𝑖=1

Where 𝑥𝑖 is representing inputs; Ɵ𝑖 (𝑡) is connection weight function between input and
output layer; 𝑔𝑖 is the input activation function (usually sigmoid type), 𝑔𝑜 is output

149 | P a g e
activation function, Ɵ𝑏 is the weight function of the bias unit, Ɵℎ is the weight function
of the hidden layer and 𝑔ℎ is hidden layer activation function.
The input layer of CFNN has 16 parameters include reservoir pressure (P),
temperature (T), hydrogen sulphide (H2S), nitrogen (N2), carbon dioxide (CO2),
hydrocarbon compositions (C1 – C7), molecular weight of C7+ (MWC7+), specific gravity
of gas on the surface (SGg). The constructed network aims at predicting two-phase Z
factor as output value. The inputs of the networks are independent variable and output
is dependant. In another word gas-condensate two-phase Z factor is a function of all
the inputs and can be written as follow.
𝑍2−𝑝ℎ𝑎𝑠𝑒 = 𝑓(𝑃, 𝑇, 𝐻2 𝑆, 𝑁2 , 𝐶𝑂2 , 𝐶1 , 𝐶2 , 𝐶3 , 𝐼𝐶4 , 𝑁𝐶4 , 𝐼𝐶5 , 𝑁𝐶5 , 𝐶6 , 𝐶7+ , 𝑀𝑊𝐶7+ , 𝑆𝐺𝑔 )
In preparing the data, the data bank divided into three different data sets of training,
validation and testing. The random selection of the data is employed, where 70% of
the data bank assigned for training, 15% for validation and 15% for testing. Then
several architecture of CFNN with various hidden layer in terms of size and
combination of different neurons in each hidden layer examined. This is to ensure in
developing a network with acceptable topology.
The objective for choosing the best CFNN network was the least mean square error
(MSE) and highest coefficient of determination R (1). An example of the constructed
CFNN in Matlab with three hidden layers and various neurons in each layer shown in
following figure.

Figure 5.6. Cascade forward neural network with three hidden layer constructed in Matlab for
prediction of two-phase Z factor.

Table 5.4 illustrates the results of this comparison for 1 hidden layer in constructed
CFNN with different neurons. Same procedure was repeated to see the effect of
number of hidden layers on improving the performance of the CFNN by means of
increasing R. Among many algorithms available to train the network Levenberg –

150 | P a g e
Marquardt (LM) was used due to its stability and rapid convergence (Tanasa et al.,
2013; Mohamadi-Baghmolaei et al., 2015; Rostami, Hemmati-Sarapardeh and
Shamshirband, 2018). The LM algorithm is utilized for optimizing weights and biases
of constructed CFNN, where the weights are computed as follow.

−1
𝑊𝑘+1 = 𝑊𝑘 − [𝐽𝑇 𝑊𝑘 𝐽𝑊𝑘 − Ƞ𝐼] × 𝐽𝑇 𝑊𝑘 𝑒𝑊𝑘 5.15

Where the weight matrix is 𝑊𝑘+1 and 𝑊𝑘 during 𝑘 + 1𝑡ℎ and 𝑘 𝑡ℎ repetitions, J stands
for Jacobian matrix, e is accumulated errors vector, 𝐼 is the identity matrix and Ƞ is the
parameter to express the ability of LM algorithm for altering the searching method
(Hagan and Menhaj, 1994; Rostami, Hemmati-Sarapardeh and Shamshirband, 2018).

Table 5.4. The relation between the number of neurons and coefficient of determination for 1
hidden layer CFNN, using Levenberg – Marquardt (LM) as training algorithm.
No. Neurons Training R Validation R Testing R Convergence
time
5 0.9931 0.9829 0.9919 00:00:00
10 0.9945 0.9953 0.9871 00:00:09
15 0.9951 0.9865 0.9744 00:00:00
20 0.9978 0.9926 0.9330 00:00:01
25 0.9972 0.9900 0.9778 00:00:02
30 0.9957 0.9931 0.9854 00:00:11
35 0.9965 0.9981 0.9869 00:00:04
40 0.9957 0.9953 0.9926 00:00:01
45 0.9943 0.9943 0.9686 00:00:02
50 0.9956 0.9923 0.9651 00:00:01
55 0.9967 0.9923 0.9651 00:00:02
60 0.9959 0.9962 0.9862 00:00:03
65 0.9973 0.9803 0.9974 00:00:13
70 0.9949 0.9858 0.9822 00:00:04

After attempting several trial and error to come up with the best CFNN network for
prediction of gas-condensate two-phase Z factor, the topology of CFNN with one

151 | P a g e
hidden layer with only five neurons return the output with the lowest MSE and highest
coefficient of determination (R). The developed code for this network provided in
Appendix B.

Feed Forward Backpropagation Neural Networks

There are many different types of ANN networks that can be used for different
applications. Feed-forward neural network is the first and simplest type of the neural
network. In this type the information only moves in one direction forward from input to
the hidden layer and then output (Giri Nandagopal and Selvaraju, 2016). In
implementing ANN, it is important to recognize the different between the network
structure (the network’s arrangement) and ANN algorithm (the computation that
eventually produce the output of the network). Once the ANN is structured for
prediction of particular application (in our case Z factor), the network is ready to be
trained. Two approaches of supervised and unsupervised learning are currently exist
in literature for training of the data. Supervised learning is working with a set of labelled
data where an output response exist for each input data (training stage). In
unsupervised learning the algorithm would find any hidden relations or pattern among
the input and output data (Bell, 2014, p. 4).
The most widely used ANN architecture is fully connected, supervised network with
backpropagation algorithm. Selecting the type of network depends on the nature of
the data that is used for developing the network. Our data bank allows us to use
supervised learning ANN method incorporated with backpropagation algorithm for
training of the data. This type of networks are excellent for prediction task (Agatonovic-
Kustrin and Beresford, 2000). The architecture of feedforward backpropagation ANN
network that constructed for prediction of two-phase Z factor is shown in Figure 5.7.
This architecture is for one hidden layer, the number of hidden layers are arbitrary and
trial and error can be used to get the optimum structure to achieve the highest
performance of the network.

152 | P a g e
 Figure 5.7. Feed Forward Artificial Neural Network structure used for prediction of Z factor.

In FFNN structure if 𝑥 represents the input variables, 𝜔 represents weight function of

each neuron and y stands for output of the network, the mathematical equation can
be written as follow.

𝑘 𝑛
𝑦 = 𝑓 𝑜 (𝜔𝑏 + ∑ 𝜔𝑗𝑜 𝑓 ℎ (𝜔𝑗𝑏 + ∑ 𝜔𝑗𝑖ℎ 𝑥𝑖 )) 5.16
𝑗=1 𝑖=1

Where 𝜔𝑏 is the bias of input layer to output and 𝜔𝑗𝑏 is the weight from bias to hidden
layer.
The accuracy of the neural network performance is influenced by network’s
architecture such as number of neurons, hidden layers and the training algorithm
(Haykin, 1994; Gharagheizi et al., 2014).
Despite high popularity of the FFNN, its main defect is correct determination of the
neurons required in each layer. Using fewer neurons leads to poor performance of the
model and higher number of neurons results in overfitting (Soroush et al., 2015;
Mesbah, Soroush and Rostampour Kakroudi, 2017). Defining the optimal architecture
that simulate the actual behaviour of experimental two-phase Z factor is not an easy
task. Hence, a trial and error scheme has been developed to find the size of hidden
layers and magnitude of the neurons in each individual layer.

153 | P a g e
Initially the collected data bank 1084 data sets (19518 data points) were partitioned
into three parts of training (70%) validation (15%) and testing (15%). The data
selection for each developed networks is random, which means training, validation
and testing in one network is not necessarily similar to other network. This gives high
capability of the trained networks to predict the output with various input.
Furthermore, to train the network diverse algorithms can be used to minimise the error
during the training. Some of these algorithms are including Levenberg–Marquardt
(LM), Bayesian – Regularization (BR), Resilient Back Propagation (RBP) and Scaled
Conjugate Gradient (SCG) (Majidi et al., 2014). The most frequently used algorithms
of Levenberg – Marquardt (LM) and Bayesian – Regularization (BR) have been
implemented in our scheme to achieve the optimum network performance. LM
algorithm is shown in equation 5.15. Here the formulation of BR algorithm is presented.

BR algorithm minimizes the combination of squared error and weights and then
generalized the network by determination of correct weights and biases. The BR
function uses Jacobian matrix, which performance of the network will be assessed by
sum of mean square error (Mackay’, 1992). BR takes place within LM algorithm, where
backpropagation can be used to calculate Jacobian matrix with respect to the weights
and biases of the network (Louridas and Ebert, 2016). The BR is formulated in the
following.

[𝐽𝑇 𝑊𝑘 𝐽𝑊𝑘 + Ƞ𝐼]

𝑊𝑘+1 = −( ) 5.17
𝐽𝑇 𝑊𝑘 𝑒𝑊𝑘

Both LM and BR algorithm are used in order to find the best neural network to perform
for prediction of Z factor. Our criteria for selecting the best FFNN was chosen by
monitoring network performance through mean square error (MSE) and coefficient of
determination (R). The results of these comparisons for one hidden layer are given in
Table 5.5 and Table 5.6 for FFNN with LM and FFNN with BR algorithm respectively.
The networks are constructed in Matlab and the codes are presented at Appendix B.

154 | P a g e
Table 5.5. Statistical parameters and convergence time of FFNN with one hidden layer using
LM algorithm.
No. Neurons R2 MSE Convergence time
(Second)

5 0.9864 0.0074 00:00:02

10 0.9891 0.0034 00:00:00
15 0.9892 0.0033 00:00:01
20 0.9605 0.0072 00:00:00
25 0.9914 0.0026 00:00:01
30 0.9947 0.0013 00:00:01
35 0.9854 0.0039 00:00:02
40 0.9922 0.0021 00:00:02
45 0.9953 0.0013 00:00:02
50 0.9741 0.0061 00:00:02
55 0.9917 0.0018 00:00:03
60 0.9537 0.0115 00:00:06
65 0.9861 0.0032 00:00:09
70 0.9503 0.0120 00:00:06

155 | P a g e
Table 5.6. Statistical parameters and convergence time of FFNN with one hidden layer using
BR algorithm for training.
No. Neurons R2 MSE Convergence time
(Hours)

5 0.9756 0.0060 00:00:05

10 0.9895 0.0026 00:00:47
15 0.7559 0.0838 00:01:43
20 0.9909 0.0024 00:09:01
25 0.9965 0.0023 00:00:55
30 0.9726 0.0079 00:08:01
35 0.9746 0.00557 00:08:33
40 0.9956 0.00106 00:08:33
45 0.8411 0.0442 00:03:08
50 0.9877 0.0031 00:05:27
55 0.9834 0.00478 00:05:50
60 0.9876 0.003375 00:11:20
65 0.9872 0.003272 00:09:54
70 0.8442 0.004076 00:10:04

The performance of the FFNN in terms of MSE with two different algorithms of LM and
BR is presented in Figure 5.8. As the results in Figure 5.8 shows the constructed FFNN
incorporated with BR algorithm consists of 40 neurons returns the output values with
the least MSE of 0.001062 than other networks. This followed by constructed network
using LM algorithm with 30 neurons with the MSE of 0.001346. Taking the
convergence time (computational efficiency) into consideration we propose the
network with LM algorithm for prediction of gas-condensate two-phase Z factor as BR
needs more time (00:08:30) than LM algorithm (00:00:01) in the training stage. The
proposed FFNN is shown in Figure 5.9. This network consists of 16 inputs, one hidden
layer with 30 neurons, one bias in each layer and one output (two-phase Z factor).
This network is representation of the ANN structure in Figure 5.7. The FFNN will be
used for prediction of gas-condensate two-phase Z factor below the saturation
pressure.

156 | P a g e
 0.014

0.012
LM BR
0.01

0.008
MSE

0.006

0.004

0.002

0
5 10 15 20 25 30 35 40 45 50 55 60 65 70
Number of Neurons
Figure 5.8. Comparison of constructed feed-forward neural network (FFNN) performance with
different number of neurons in hidden layer.

Figure 5.9. Proposed FFNN for prediction gas condensate two-phase compressibility factor.

Adaptive Neuro Fuzzy Inference System (ANFIS)

Another intelligent technique that is utilized in this study for prediction of two-phase Z
factor in gas-condensate reservoirs known as Adaptive Neuro Fuzzy Inference System
(ANFIS). The method is another form of TSK fuzzy algorithm implemented in 4.6. The
method is combination of artificial neural network and fuzzy rule based system. The
main reason for adopting ANFIS to estimate two-phase Z factor is to have the benefits
of artificial neural network as well as fuzzy reasoning.
As explained in section 4.5 and 4.6 fuzzy inference system is employing fuzzy IF-
THEN rules that can model the qualitative aspects of human knowledge of a system

157 | P a g e
or process without any need of precise quantitative analysis. The ability of fuzzy sets
for modelling and fuzzy identification initially was identified by Takagi and Sugeno,
(1985). Although there has been numerous application of fuzzy system in various
fields of engineering after Takagi and Sugeno, (1985) discovery, there was not any
standard method for transforming human knowledge into rule base fuzzy inference
system (FIS). Furthermore there was not an effective method available for tuning the
membership function of Takagi and Sugeno, (1985) FIS system.
Adaptive Neuro Fuzzy Inference System (ANFIS) was developed following TSK fuzzy
method to fulfil the aforementioned deficiencies of TSK fuzzy approach. In early
development of TSK algorithm type 1 and type 2 of ANFIS were shaped. In these two
types calculation of the errors in ANFIS algorithm was through using back-propagation
gradient decent, which calculate the error signals from output layer backward to the
input nodes. This learning rule is same as backpropagation gradient descent of feed-
forward neural network. The difficulty with this approach is in tuning the membership
functions through rigorous trial and error, which is not an easy task to implement.
Furthermore, the gradient decent learning rule is notorious for slowness and has
tendency to trap in local minima. Hence, Jang, (1993) introduced type – 3 Adaptive
Neuro Fuzzy Inference System (ANFIS) in order to tackle the aforementioned
deficiencies and enhance previous version of ANFIS model. He proposed an adaptive
hybrid neural network based on combination of gradient decent algorithm and least
square method for better optimization of the parameters of Takagi and Sugeno FIS
system. This optimization of FIS system makes type 3 ANFIS a powerful tool for
prediction and forecast of various applications. The architecture of type 3 fuzzy
reasoning and ANFIS with two input variables of 𝑥 and 𝑦 is shown in Figure 5.10 and
will be explained in following.

158 | P a g e
 b

Figure 5.10. Schematic illustration of type – 3 fuzzy reasoning (a); ANFIS structure type – 3
(b).

Figure. 5.10a represents the fuzzy reasoning behind the ANFIS model with two input
variables and triangular membership function. This is essentially explains Takagi and
Sugeno’s fuzzy type with introducing number of rules in this case two as follow:
Rule 1: If 𝑥 is A1 and 𝑦 is B1, then 𝑓1 = 𝑝1 𝑥 + 𝑞1 𝑦 + 𝑟1
Rule 2: If 𝑥 is A2 and 𝑦 is B2, then 𝑓2 = 𝑝2 𝑥 + 𝑞2 𝑦 + 𝑟2
Where A1, A2, B1 and B2 are membership functions x and y are inputs, p1, q1 and r1 are
constant parameters of linear function.
The typical ANFIS structure shown in Figure 5.10b consists of five layers, each
processing different operations in the structure. Two square and circular nodes have
been proposed in Figure 5.10b, where square nodes reflect adaptive (having constant
parameters) and circular nodes are fixed (without parameters) (Jang, 1993). The
operation of ANFIS algorithm in each layer is as follow.
Layer 1: every node 𝑖 in this layer represents by a function of square node as follow:

𝑂𝑖1 = 𝜇𝐴𝑖 (𝑥) 5.18

Where 𝑥 is input of the node 𝑖, 𝐴𝑖 is a linguistic term of fuzzy sets (very low, low, etc.,…)
associated with the node function. The main role of this layer is to turn the raw data
into linguistic terms based on previously selected membership functions (node
functions). Many membership functions can be selected such as triangular, gaussian,
trapezoidal, sigmoidal, s-shape and z-shape in this layer.

159 | P a g e
Layer 2: the nodes in this layer labelled with ∏, where the results are calculated as
multiplication of incoming signals from layer 1. Effectiveness of the network is
determined by the results of this layer, as the outcome is a weight function. Following
equation represents the network operation in layer 2.

𝜔𝑖 = 𝜇𝐴𝑖 (𝑦) × 𝜇𝐵𝑖 (𝑦) = 𝜔𝑖 , 𝑖 = 1,2 5.19

Where 𝐴𝑖 and 𝐵𝑖 are linguistic terms of y input in universe of 𝜇. The output of the above
equation is the firing weight of each node (𝜔𝑖 ).
Layer 3: Incoming weight’s strengths calculated in layer 2 are normalized in this step
to identify the difference between the firing strength of each rule from the total firing
strength of entire rules. This normalization in layer 3 labelled by N and following
equation can be used to determine each weight’s strength.

𝜔𝑖
𝜔𝑖 = 5.20
𝜔1 /𝜔2

Layer 4: This layer characterizes the linguistic terms of the output model. This means
the influence of each rule on output is determined in this layer with some degree. A
square node with the following function carries out the computation in layer 4.

𝑂𝑖4 = ̅̅̅𝑓
𝜔𝑖 𝑖 = 𝜔 1 1 𝑥 + 𝑞1 𝑦 + 𝑟1 )
̅̅̅̅(𝑝 5.21

Where 𝑝1, 𝑞1 and 𝑟1are linear variables. These variables are optimized by ANFIS
algorithm to provide the accuracy between the model outcome and the target value.
Layer 5: A circular signal node in this layer is used to sum up all the incoming signals
(rules) from previous layer. The computation transfers the rules to numeric values
using weighted average sum approach as follow.

∑𝑖 ̅̅̅𝑓
𝜔𝑖 𝑖
𝑂𝑖5 = ∑ ̅̅̅𝑓
𝜔𝑖 𝑖 = 𝜔
̅̅̅̅𝑓
1 1 + ̅̅̅̅𝑓
𝜔2 2 = 5.22
∑𝑖 ̅̅̅
𝜔𝑖
𝑖

Where 𝑂𝑖5 stands for output of the network, 𝑓1,2 represent output functions of rule 1 and
rule 2.

160 | P a g e
The above procedure also has been adopted for calculation and prediction of two-
phase Z factor for gas-condensate reservoirs below the dew point pressure. For this
purpose an ANFIS structure is proposed and will be explained in following section.

Proposed ANFIS structure

In this part a new ANFIS structure based on Jang, (1993) was developed for prediction
of two-phase Z factor of gas-condensate reservoirs. A Matlab code has been
developed and used for this purpose (refer to Appendix B for details of the developed
code). The inputs variables consist of 16 parameters including, pressure (P),
temperature (T), carbon dioxide (CO2), hydrogen sulphide (H2S), nitrogen (N2), HCC1,
HCC2, HCC3, HCIC4, HCNC4, HCIC5, HCNC5, HCC6, HCC7+, molecular weight of C7+
(MWC7+) and gas specific gravity (SG). To distinguish between the parameters of the
fuzzy rule and English letters HC is added to each hydrocarbon components. The
output parameter of the network is two-phase Z factor. To be consistent with other
intelligent model CFNN and FFNN, the data bank divided to three parts of training
(70%), validation (15%) testing (15%). These proportions were randomly selected to
ensure suitable coverage of all the data ranges of the variables.

For simplicity if we assume the proposed structure with only two rules and using Takagi
– Sugeno type FIS system, then the two-phase Z factor can be presented as follow.
Rule 1:
if T is A1, P is B1, CO2 is C1, H2S is D1, N2 is E1, HCC1 is F1, HCC2 is G1, HCC3 is H1,
HCIC4 is I1, HCNC4 is J1, HCIC5 is K1, HCNC5 is L1, HCC6 is M1, HCC7+ is N1, MWC7+
is O1 SG is P1; Then
𝒇𝒖𝒏𝒄𝒕𝒊𝒐𝒏𝟏 = 𝑎1 𝑇 + 𝑏1 𝑃 + 𝑐1 𝐶𝑂2 + 𝑑1 𝐻2𝑆 + 𝑒1 𝑁2 + 𝑓1 𝐻𝐶𝐶1 + 𝑔1 𝐻𝐶𝐶2 + ℎ1 𝐻𝐶𝐶3
+ 𝑖1 𝐻𝐶𝐼𝐶4 + 𝑗1 𝐻𝐶𝑁𝐶4 + 𝑘1 𝐻𝐶𝐼𝐶5 + 𝑙1 𝐻𝐶𝑁𝐶5 + 𝑚1 𝐻𝐶𝐶6 + 𝑛1 𝐻𝐶𝐶7+
+ 𝑜1 𝑀𝑊𝐶7+ + 𝑝1 𝑆𝐺 + 𝑞1
Rule 2:
if T is A2, P is B2, CO2 is C2, H2S is D2, N2 is E2, HCC1 is F2, HCC2 is G2, HCC3 is H2,
HCIC4 is I2, HCNC4 is J2, HCIC5 is K2, HCNC5 is L2, HCC6 is M2, HCC7+ is N2, MWC7+
is O2 SG is P2; Then

161 | P a g e
 𝒇𝒖𝒏𝒄𝒕𝒊𝒐𝒏𝟐 = 𝑎2 𝑇 + 𝑏2 𝑃 + 𝑐2 𝐶𝑂2 + 𝑑2 𝐻2𝑆 + 𝑒2 𝑁2 + 𝑓2 𝐻𝐶𝐶1 + 𝑔2 𝐻𝐶𝐶2 + ℎ2 𝐻𝐶𝐶3
+ 𝑖2 𝐻𝐶𝐼𝐶4 + 𝑗2 𝐻𝐶𝑁𝐶4 + 𝑘2 𝐻𝐶𝐼𝐶5 + 𝑙2 𝐻𝐶𝑁𝐶5 + 𝑚2 𝐻𝐶𝐶6 + 𝑛2 𝐻𝐶𝐶7+
+ 𝑜2 𝑀𝑊𝐶7 + + 𝑝2 𝑆𝐺 + 𝑞2

In above rules T, P, CO2, H2S, N2, HCC1 – HCC7+, MWC7+ and SG are input
parameters of the model, 𝐴𝑖 to 𝑃𝑖 is representing the membership function (gaussian,
triangular, trapezoidal and etc.,). In the developed functions, 𝑎𝑖 to 𝑞𝑖 are the constant
parameters of the ANFIS model (optimized by training algorithm). In order to estimate
these constant parameters, least square approach was used. The corresponding
ANFIS architecture for prediction of gas-condensate two-phase Z factor is shown in
Figure 5.11.

Figure 5.11. Proposed ANFIS architecture for prediction of gas condensate two-phase Z
factor.

The effectiveness of the developed ANFIS model for prediction of two-phase Z factor
influenced by the type of membership function, cluster range and number of fuzzy
rules. Several membership functions are exist to relate the input variables to the
output. The ability of 9 membership functions and their effectiveness in prediction of
two-phase Z factor have been investigated. These membership functions are including
gaussian, triangular, trapezoidal, generalized-bell shaped, gaussian combination

162 | P a g e
membership function, pi (∏) shaped, sigmoidal, S shaped and Z shaped membership
functions. The description and mathematical representation of gaussian, triangular
and trapezoidal membership functions previously given in section (4.7.2). The shape
and mathematical formula of other studied membership functions presented in Table
5.7. Primarily the effectiveness of all membership functions have been studied using
constant number of fuzzy rules (for instance 8 rules). The statistical parameters of
mean square error (MSE) and coefficient of determination (R2) have been recorded to
see the best performance of the ANFIS using various membership functions. As the
results in Table 5.8 shows gaussian membership function has the best performance
for prediction of the output (two-phase Z factor). This is proven by lower mean square
error (MSE) and highest coefficient of determination (R2). Hence, gaussian
membership function has been chosen to incorporate with ANFIS algorithm in this
study for modelling gas-condensate two-phase Z factor.

Table 5.7. The description and details of the utilized membership functions for prediction of
gas-condensate two-phase Z factor using ANFIS.

Name Description Shape

The generalized-bell shaped

membership function has three
parameters of a, b and c. The
parameter of c representing centre of
generalized- the curve and a and b represent the
bell shaped
feet of the shape. Parameter b is
usually positive.
1
𝑓(𝑥; 𝑎, 𝑏, 𝑐) = 𝒈𝒃𝒍𝒎𝒇 = 𝒇(𝒙; 𝟐, 𝟒, 𝟔)
𝑥 − 𝑐 2𝑏
1+| 𝑎 |

163 | P a g e
 This membership function is a
combination of sigmoid function and
parameter; 𝑦 =
𝑔𝑎𝑢𝑠𝑠2𝑚𝑓(𝑥: 𝑠𝑖𝑔1, 𝑐1, 𝑠𝑖𝑔2, 𝑐2).
The first function sig1, c1 determine
Gaussian
combinatio the shape of the most left curve.
n The second function sig2, c2
membershi
p function represents the shape of the most
right curve shape. The function can
be written mathematically as follow. 𝑔𝑎𝑢𝑠𝑠2𝑚𝑓 = 𝑓(𝑥; 0.1, 10)
−(𝑥−𝑐)2
𝑓(𝑥; 𝜎, 𝑐) = 𝑒 2𝜎2

The name of this membership

function is because of its pi (∏)
shape. Four parameters of a, b,
c and d is used in construction of
the shape. The parameters of a
and d locate the feet of the
curve, while b and c locate its
shoulders.
pi (∏)
𝑓(𝑥; 𝑎, 𝑏, 𝑐, 𝑑)
shaped 0, 𝑥≤𝑎
𝑥−𝑎 2 𝑎+𝑏
2( ) , 𝑎≤𝑥≤
𝑏−𝑎 2
2
𝑥−𝑏 𝑎+𝑏 𝒑𝒊𝒎𝒇 = 𝒇(𝒙; 𝟏, 𝟒, 𝟓, 𝟏𝟎)
1 − 2( ) , ≤𝑥≤𝑏
𝑏−𝑎 2
= 1, 𝑏≤𝑥≤𝑐
𝑥−𝑐 2 𝑐+𝑑
1 − 2( ) , 𝑐≤𝑥≤
𝑑−𝑐 2
𝑥−𝑐 2 𝑐+𝑑
2( ) , ≤𝑥≤𝑑
𝑑−𝑐 2
{ 0, 𝑥≥𝑑 }

164 | P a g e
 The sigmoidal membership function
is depending on two parameters of a
and c. Depending on the sign of
these parameters the sigmoidal
membership function is open to the
Sigmoidal right or to the left. This would give
this function ability to cover very
large or very negative data points.
1
𝑓(𝑥; 𝑎, 𝑐) = 𝒔𝒊𝒈𝒎𝒇 = 𝒇(𝒙; 𝟐, 𝟒)
1+ 𝑒 −𝑎(𝑥−𝑐)

The S shape membership function is

spline based curve on vector x. the
parameters of a and b represents the
extreme slope portion of the curve.
The mathematical description of the
function is:
𝑓(𝑥; 𝑎, 𝑏)
S shape 0, 𝑥≤𝑎
𝑥−𝑎 2 𝑎+𝑏
2( ) , 𝑎≤𝑥≤
= 𝑏−𝑎 2
𝑥−𝑎 2 𝑎+𝑏
1−( ) , ≤𝑥≤𝑏
𝑏−𝑎 2
{ 1, 𝑥≥𝑐 } 𝒔𝒎𝒇 = 𝒇(𝒙; 𝟏, 𝟖)

The shao of this membership

function is Z shape and it is function
of two parameters of a and b.
mathematically this function can be
written as follow.
Z shape 𝑓(𝑥; 𝑎, 𝑏, 𝑐, 𝑑)
1, 𝑥≤𝑎
𝑥−𝑎 2 𝑎+𝑏
2( ) , 𝑎≤𝑥≤
= 𝑏−𝑎 2
𝑥−𝑎 2 𝑎+𝑏
2( ) , ≤𝑥≤𝑏 𝒁𝒎𝒇 = 𝒇(𝒙; 𝟑, 𝟕)
𝑏−𝑎 2
{ 0, 𝑥≥𝑏 }

165 | P a g e
Table 5.8. Performance of the developed ANFIS model using various membership function.

Membership No. IF-THEN R2 MSE

function rules Training Testing Training Testing
Gaussian 8 0.9956 0.8454 0.001057 0.0486
Triangular 8 0.9265 0.76672 0.00253 0.1163
Trapezoidal 8 0.9512 0.8077 0.00105 0.0482
generalized-bell 8 0.9365 0.79521 0.001253 0.0576
shaped
Gaussian
8 0.9486 0.80549 0.001059 0.0486
combination
pi (∏) shaped 8 0.87562 0.74372 0.002365 0.01087
sigmoidal 8 0.86256 0.73243 0.002563 0.11784
S shape 8 0.85236 0.72376 0.002623 0.12060
Z shape 8 0.86325 0.73301 0.002585 0.11886

Clustering (Partitioning) of the data

Prior selection of adequate membership functions for ANFIS modelling need

partitioning of the training data set to certain number of clusters. For partitioning of the
data to number of clusters several clustering methods are available in literature. The
examples are look up table method, fuzzy c-means (FCM), Gustafson-Kessel (GK),
K-means clustering and subtractive clustering (SC). Using the look up table method is
not recommended for the large scale problem due to increasing the parameters of the
model and number of MFs, which consequently leads to the complexity of the model
(Lee, 2005; Safari et al., 2014) . Further details about different clustering techniques
is given in section 4.6.1. FCM and SC are two methods that have been used
extensively in the literature for data organization and also data compressions in ANFIS
modelling. Both methods shown high efficiency over wide range of data for various
applications, however many studies preferred SC over FCM due to its simplicity in
calculation procedure and also its independency of problem dimensionality (Jang, Sun
and Mizutani, 1997; Nikravesh, Zadeh and Aminzadeh, 2003; Lee, 2005).
FCM clustering algorithm determines the degree of belonging of each data point to
each cluster. The centre of each cluster is found in an iterative procedure by
minimizing an objective function called mountain function. If the data is highly non-
linear and multidimensional the computation time is growing exponentially. This is
because the algorithm needs to evaluate the mountain function over all of the grid
points (Bezdek and Pal, 1992; Jang, Sun and Mizutani, 1997). Nevertheless,
subtractive clustering is more adaptable to non-linear more dimensional data and

166 | P a g e
return the number of clusters quickly and efficiently in a simple computational
procedure. SC is an appropriate choice for the data sets without knowing how many
clusters are required (Chiu, 1994). Hence, in this study we utilized SC method
proposed by Chiu, (1994) for partitioning two-phase Z factor training data bank. SC
assumes that each data point is a potential cluster centroid. The results of SC can be
used for initializing centroid of the FCM algorithm. The SC proposed by Chiu, (1994)
is an extension of mountain clustering method of Yager and Filev, (1994). Based on
density index (𝐷𝑖 ) of potential centroid data, the SC algorithm evaluates the density
index of each data point 𝑥𝑖 as follow (Chiu, 1994; Chen et al., 2015).

𝑛
‖𝑥𝑖 − 𝑥𝑘 ‖2
𝐷𝑖 = ∑ 𝑒𝑥𝑝 [− 2 ] 5.23
𝑟
𝑘 ( 𝑎⁄2)

Where 𝑥𝑖 and 𝑥𝑘 are data points, 𝑟𝑎 is clustering radius and can be determined by the
following equation.

𝑟𝑎 = 0.5𝑚𝑖𝑛{𝑚𝑎𝑥‖𝑥𝑖 − 𝑥𝑘 ‖} 5.24

The data beyond the radius has little or no effect on the density index (𝐷𝑖 ). To
implement equation 5.24, first, the data index of cluster 1 (𝑥c1 ) should be selected.
This data index has the highest density index (Dc1) at the first cluster centroid. Then
the data in the radius ra is removed from potential centroid data set, and the
neighbourhood with lower density index is defined by the following equation.

‖𝑥𝑖 − 𝑥𝑘 ‖2
𝐷𝑖 = 𝐷𝑖 − 𝐷𝑐1𝑒𝑥𝑝 [− 2 ] 5.25
𝑟
( 𝑏⁄2)

Where 𝑟𝑏 > 𝑟𝑎 , this would distinguish between two clusters. The above mathematical
procedure of SC can be summarized in five steps as follow:
1. Using equation 5.23 based on the density of surrounding data points; calculate
the likelihood that each data point would define a cluster centre.
2. Choose the data points with the highest potential to be the first cluster centre.
3. Using equation 5.24 remove all data points near the first cluster centre.
4. Choose the remaining points with the highest potential as the next cluster
centre using equation 5.25.
167 | P a g e
 5. Repeat steps 3 and 4 until all data points are within the influence range (ra) of
a cluster centre.

Accurate determination of clustering radius (𝑟𝑎 ) using above procedure is directly

influence the performance of the developed ANFIS model. Changing the value of r a
would change the number of membership functions, consequently number of IF-THEN
fuzzy rules. To obtain optimum number of rules the value of clustering radius changed
between 0 < 𝑟𝑎 < 1. The performance of the ANFIS system evaluated by the statistical
metric of RMSE in the training and testing stage.
In this study a new iterative technique is developed and carried out to find the optimum
number of fuzzy rules and determine the optimum number of clusters. The
implementation of the method is as follow:
1. Calculate ra values between 0 < 𝑟𝑎 < 1 in ten intervals using equation 5.24; and
record the RMSE of training and testing of ANFIS model.
2. Select two intervals with lowest RMSE (for example 5 and 6 in table 15), and
eliminate other clustering radius with high error.
3. Calculate ra values for two new selected intervals (0.45 < 𝑟𝑎 < 0.5); and record
RMSE of the ANFIS model.
4. Repeat step 2 to 3 until the RMSE reduced to the point where number of IF-
THEN rules converged or no longer changes.
5. Use the calculated clustering radius (ra) value as an optimum choice for
developing number of rules in ANFIS model.
The results in Table 5.9 and 5.10 illustrate the values of RMSE for training and testing
data sets using the above procedure. As it can be seen from Table 5.9 the number of
IF-THEN rules is converged at 11, with the cluster radius of (𝑟𝑎 ≈ 0.45). Therefore, the
proposed ANFIS model for prediction of gas-condensate two-phase Z factor consists
of 11 fuzzy IF-THEN rules.

168 | P a g e
Table 5.9. Effect of clustering radius on number of fuzzy rules and the performance of the
model (1st iteration).

No. ANFIS Clustering radius No. IF-THEN RMSE

(ra) rules Training Testing
1 0.25 35 852 4451
2 0.30 23 3.60 277
3 0.35 17 43 1286
4 0.40 12 0.07252 1.13
5 0.45 11 0.04224 0.08657
6 0.50 9 0.04773 0.12197
7 0.55 8 0.04734 0.07299
8 0.60 6 0.05029 0.26412
9 0.65 4 0.05077 0.49331
10 0.70 4 0.049461 0.39360
11 0.75 4 0.05025 0.28450
12 0.80 3 0.05869 0.09796
13 0.85 2 0.06338 0.05266
14 0.90 2 0.06289 0.05207
15 0.95 2 0.06228 0.05146
16 1 2 0.06224 0.05696

169 | P a g e
Table 5.10. Effect of clustering radius on number of fuzzy rules and the performance of the
model (2st iteration).

No. ANFIS Clustering radius No. IF-THEN RMSE

(ra) rules Training Testing
1 0.45 11 0.04224 0.08657
2 0.451 11 0.04158 0.09991
3 0.4515 11 0.04121 0.14248
4 0.452 11 0.03984 0.11151
5 0.4525 11 0.04085 0.15212
6 0.453 11 0.03980 0.10294
7 0.4535 11 0.04201 0.10277
8 0.454 11 0.04206 0.09990
9 0.4545 11 0.04203 0.20143
10 0.455 11 0.039483 0.20952
11 0.4555 11 0.04198 0.18015
12 0.4560 11 0.04206 0.18578
13 0.4565 11 0.04055 0.12805
14 0.457 11 0.04199 0.15657
15 0.4575 11 0.04209 0.10474
16 0.4580 11 0.04209 0.18251
17 0.4585 11 0.04037 0.15201
18 0.4590 11 0.04007 0.09991
19 0.4595 11 0.03979 0.10813
20 0.46 11 0.03953 0.09709

The ANFIS structure with 11 fuzzy IF-THEN rules constructed in Matlab and shown in
Figure 5.12. The output of ANFIS is a linear function with several optimized
parameters for each rule. More clear graphical visualization of constructed type 3
ANFIS for prediction of gas-condensate two-phase Z factor is presented in Figure
5.13.

170 | P a g e
Figure 5.12. Structure of proposed ANFIS model for prediction of two-phase Z factor of gas
condensate reservoirs.

Figure 5.13. Architecture of Takagi – Sugeno fuzzy model for prediction of gas condensate
two-phase Z factor.

171 | P a g e
 Results and discussion

Z factor of gas-condensate reservoirs is one of the most important PVT properties that
its accurate determination leads to accurate well deliverability modelling of such
reservoirs below the saturation pressure. In this study extensive data bank from
various published sources have been collected to investigate the accuracy of the
current literature models for prediction of two-phase Z factor.
Five literature correlations of Hall-Yarborough (1973) and Dranchuk-Abu-Kassem
(1975), Beggs and Brill (1973), Azizi et al., (2010) and Rayes et al., (1992) have been
utilized for this purpose.
In utilizing the existing literature models initially pseudo critical properties of
temperature (Tpc) and pressure (Ppc) of each gas-condensate system were computed,
using mixture composition and Kay’s mixing rule. For computing critical properties of
hydrocarbon plus fraction because of availability of compositional data the method of
Matthews and Roland, (1942) was used. To include non-hydrocarbon impurities in
computation of pseudocritical properties Wichert and Aziz, (1972) correlation was
employed. From calculated Tpc and Ppc, pseudo reduced temperature (Tpr) and
pressure (Ppr) were calculated. Finally, two-phase Z factor of each gas-condensate
system was estimated. The prediction accuracy of existing literature models for
estimating gas-condensate two-phase Z factor were not acceptable and associated
with high error. Therefore other techniques known as machine learning based models
have been utilized for the prediction of gas-condensate two-phase Z factor.

Accuracy and robustness of any ML based method is depending on the diversity of

the data bank that is used for their development (Ahmadi and Ebadi, 2014; Hajirezaie
et al., 2015; Hemmati-Sarapardeh et al., 2020; Kamari et al., 2019). The detail
statistical description of the data bank is presented in 5.3.6 of this chapter. Two neural
network based models known as feed forward neural network (FFNN) and cascade
forward neural network (CFNN) and one fuzzy logic based model know as adaptive
neuro fuzzy inference system (ANFIS) adopted for prediction of two-phase Z factor in
gas-condensate reservoirs. The data divided into three parts of training (70%),
validation (15%) and testing (15%). Consistent proportion of the data used for all three
models. The random selection of the data was performed to ensure homogeneous
distribution of the data bank. These models consist of 16 inputs of T, P, H2S, CO2, N2,

172 | P a g e
C1 – C7+, MW C7+ and gas specific gravity (SG), and a single output of gas-condensate
two-phase Z factor.
Statistical error analysis of the developed machine learning methods were performed
using metrics of MAE, RMSE and AARD% and illustrated in Table 5.11. This table
shows the error measurements of all data sets at training, validation and testing stage
for three developed ML based techniques. The results confirms very good
performance of all three ML based models. ANFIS model outperforms the CFNN and
FFNN with lowest MAE of 0.0025124, RMSE of 0.0025367 and AARD% of 0.2191478
followed by FFNN and CFNN for prediction of tow-phase Z factor of gas-condensate
reservoirs. Furthermore, computational efficiency of all three developed models are
demonstrated in table 5.11. As the results indicate the developed ML based models
predicts the two-phase Z factor very fast with almost 0 seconds, nevertheless the
CFNN is slightly perform better than FFNN and ANFIS.

Table 5.11. Statistical error computation of the developed models in prediction of gas-
condensate’s two-phase Z factor below the saturation pressure.

Convergence time
Model MAE RMSE AARD%
(hour,minute,second)

FFNN 0.0047 0.0058 0.3029

00:00:00

CFNN 0.0053 0.0062 0.5117

00:00:01

ANFIS 0.0025 0.0025 0.2191

00:01:25

To visualize the performance of ML based models Figure 5.14 shows the cross plot
and the graph of relative error percentage of three models. The results of the cross
plot show all three ML models predict experimental two-phase Z factor with high
accuracy in all three stages of training, validation and testing. The statistical metrics in
Table 5.11 and the results of cross plots/relative error in Figure 5.14E-F indicate that
the ANFIS model has the best performance among other two utilized ML based
methods for prediction of two-phase Z factor. This confirms high capability of ANFIS

173 | P a g e
for relating non-linear inputs and outputs of a problem, once correct optimization of its
membership functions have obtained.
In order to see the performance of the ML based models as a function of T pr and Ppr
for prediction of two-phase Z factor, the graph of relative error against T pr and Ppr were
generated and presented in Figure 5.15 and Figure 5.16. The results show an
excellent agreement between the obtained values of ML based models and
experimental two-phase Z factor as a function of Tpr and Ppr.
Comparison of statistical error metrics of MAE, RMSE and AARD% of all utilized
literature models and the developed ML based approaches presented in Figure 5.17
and Figure 5.18. The results show that ANFIS model has superiority over all other
techniques with the least MAE, RMSE and AARD% of 0.00251, 0.00254 and 0.219
respectively.

A CFNN B CFNN Relative error %

C FFNN D FFNN Relative error %

174 | P a g e
 E ANFIS F ANFIS Relative error %

Figure 5.14. Cross plot and relative error of utilized three machine learning models in
predicting two-phase Z factor at training, validation and testing stage.

(A) (B)

(C)

Figure 5.15. Residual plot of relative error percentage of intelligent models in predicting two-
phase Z factor as a function of pseudo reduced pressure (Ppr).

175 | P a g e
 (A) (B)

(C)
Figure 5.16. Residual plot of relative error percentage of intelligent models in predicting two-
phase Z factor as a function of pseudo reduced temperature (Tpr).

0.600

0.500
MAE & RMSE

0.400

0.300

0.200

0.100

0.000
Hall-
Beggs Rayes Azizi,
Yarbor DAK FFNN CFNN ANFIS
& Brill et al et al
ough
MAE 0.14219 0.13008 0.20476 0.15165 0.38510 0.00472 0.00533 0.00251
RMSE 0.18470 0.16718 0.27691 0.19043 0.48225 0.00576 0.00620 0.00254

Figure 5.17. Error comparison of all utilized literature models and the developed ML based
approached in prediction of gas-condensate two-phase Z factor.

176 | P a g e
 35.00

30.00 28.696

25.00
AARD%
20.00 17.057

15.00
10.052
10.00

5.00 3.091 2.495

0.303 0.512 0.219
0.00

Figure 5.18. Error comparison of all utilized literature models and the developed ML based
approaches in prediction of two-phase Z factor.

The ability of the developed ML based models in following the physical trend of the
two-phase Z factor of gas-condensate reservoirs as a function of pressure and
temperature has to be tested. This is important to ensure the developed ML based
models are following physical trend of two-phase Z factor similar to literature
correlations. For this purpose, two independent samples of gas-condensate fluids from
literature were collected. First sample is from a rich gas-condensate reservoir in North
Sea obtained from Danesh, (1998, pp. 53–56). This sample is from sweet gas-
condensate with no hydrogen sulphide (H2S) content. Constant Volume Depletion
(CVD) test was performed and compositional changes of each components were
detected with pressure depletion. The result of CVD test for this sample is provided in
Table 5.12.
Second sample obtained from Elsharkawy, (2002), is a very rich sour gas-condensate
reservoir contain considerable amount of hydrogen sulphide (H2S=28.16% at the dew
point pressure). CVD test also was performed on this sample, where the compositional
changes of the sample during the CVD test was recorded as shown in Table 5.13.

Estimation of Z factor using literature models requires determination of pseudocritical

properties (Ppc and Tpc). Kay, (1936) mixing rule is used for calculation of critical

177 | P a g e
properties. For characterization of hydrocarbon plus (C7+) properties and
determination of pseudocritical properties (Ppc7+, Tpc7+) the method proposed by
Matthews and Roland, (1942) is used. Non – hydrocarbon properties were also
determined by utilizing the method of Wichert and Aziz, (1972).
Performance of utilized literature models and also the developed ML based techniques
in prediction of aforementioned independent samples examined by means of statistical
parameters of MAE, RMSE and AARD%. These statistical parameters are determined
for each model as a function of pseudo reduced pressure (Ppr). The results of this error
analysis shown in Table 5.14 for both gas-condensate fluid samples.

Graphical performance of all methods for prediction of rich sour gas-condensate

sample is plotted as a function of Ppr and shown in Figure 5.19. As the result in Figure
5.19 shows, among the literature models only Rayes et al., (1992) is close to the
experimental two-phase Z factor in the range of (𝑃𝑝𝑟 ≤ 4). Other literature techniques
of Azizi et al., (2010); Beggs and Robinson, (1975); DAK, (1973) and Hall and
Yarborough, (1973) estimate the experimental two-phase Z factor with high error
below (𝑃𝑝𝑟 ≤ 4).
The graph of experimental Z factor in Figure 5.19 also indicates that below the
saturation pressure the Z factor has a direct correlation with reservoir pressure. As the
pressure increases, the magnitude of two-phase Z factor also increases. The result
show that the conventional methods cannot follow the physical trend of the change in
Z factor of gas-condensate reservoirs below the saturation pressure. Statistical
measurements of different approach as well as the developed models are illustrated
in Table 5.14. The results of statistical error metrics show that the developed methods
predict experimental two-phase Z factor with very good accuracy in compare to
conventional techniques.
In sample 1, that high content of hydrogen sulphide (H2S) exist, the conventional
techniques return the experimental two-phase Z factor with very high error as shown
in Figure 5.19 and Table 5.14. ANFIS outperforms other methods with MAE of 0.01764
and RMSE of 0.02994 followed by FFNN, CFNN, Rayes et al., (1992), Beggs and Brill,
(1973), DAK, (1974), Hall and Yarborough, (1973) and Azizi et al., (2010). The results
show reliabilty and robustness of the developed models for predction of two-phase Z
factor in critical gas-condensate fluid below the saturation pressure. The developed

178 | P a g e
ML based methods follow the physical trend of the experimental Z factor with minimum
deviation.
For the second sample which the H2S content is zero, although the conventional
methods perform well in terms of MAE and RMSE, however the ANFIS model
retuened the experimental Z factor with the least AARD% of 0.4506 followd by CFNN,
FFNN, Beggs and Brill, (1975), DAK, (1975), Hall and Yarborough, (1973), Rayes et
al., (1992) and Azizi et al., (2010). Figure 5.20 depicts the trend of experimental two-
phase Z factor of sample 2 as a function of pressure and the performance of utilized
literature methods and developed ML based models. As can be seen the ML based
models are following the physical trend of the two-phase Z factor with very good
accuracy.
The problem with conventional methods is that their accuracy are limited for specific
range and their performance depends on the type of mixing rule that has been used
in the calculation of critical properties (Ppr and Tpr). The ML based approaches
developed in this study don’t have aformentioned limitations. Their performance is
highly accurate for prediction of two-phase Z factor of gas-condensate reservoirs
regradless of the compositional variation that occurs in such reservoirs below the
saturation pressure.

179 | P a g e
Table 5.12. Constant Volume Depletion (CVD) results on North Sea gas-condensate sample at temperature of 394K (Danesh, 1998, pp. 53–
56).

T(F) P H2S N2 CO2 C1 C2 C3 IC4 NC4 IC5 NC5 C6 C7+ MW SG Z 2-

(psia) C7+ Phase

249.5 6822 0.0 0.30 1.72 79.17 7.48 3.29 0.52 1.25 0.36 0.55 0.62 3.74 231 0.943 1.1718

249.5 5800 0.0 0.30 1.71 79.93 7.44 3.22 0.51 1.23 0.35 0.54 0.58 4.19 207 0.889 1.0867

249.5 4930 0.0 0.31 1.71 80.77 7.41 3.21 0.50 1.21 0.34 0.52 0.55 3.47 202 0.845 1.0056

180
249.5 3915 0.0 0.32 1.72 81.61 7.46 3.20 0.50 1.18 0.33 0.50 0.52 2.66 190 0.797 0.9479

249.5 3045 0.0 0.32 1.73 82.33 7.54 3.19 0.49 1.15 0.32 0.48 0.49 1.96 180 0.760 0.9176

249.5 2030 0.0 0.33 1.75 82.71 7.64 3.22 0.49 1.15 0.32 0.48 0.46 1.45 174 0.737 0.9171
 Table 5.13. Constant Volume Depletion (CVD) results on sour gas-condensate sample (Elsharkawy, 2002).

T(F) P H2S N2 CO2 C1 C2 C3 IC4 NC4 IC5 NC5 C6 C7+ MW SG Z 2-

(psia) C7+ Phase

250 4204 28.16 3.83 6.08 40.33 4.48 2.48 0.6 1.32 0.79 0.81 1.21 9.91 165 0.818 0.838

250 3614 27.67 4.55 6.44 43.82 4.71 2.43 0.55 1.2 0.68 0.69 0.96 6.3 121 0.778 0.788

250 3014 27.22 4.76 6.69 46.41 4.81 2.39 0.51 1.11 0.6 0.6 0.78 4.12 116 0.773 0.75

181
250 2414 26.95 4.73 6.85 48.07 4.87 2.37 0.49 1.06 0.55 0.54 0.66 2.86 112 0.768 0.718

250 1814 27.32 4.61 6.94 48.44 4.93 2.39 0.49 1.06 0.53 0.52 0.6 2.17 109 0.764 0.686

250 1214 28.92 4.34 6.99 46.88 4.96 2.52 0.55 1.14 0.58 0.57 0.63 1.92 107 0.762 0.639

250 714 31.82 3.94 6.79 43.31 4.94 2.77 0.67 1.4 0.74 0.71 0.77 2.14 107 0.762 0.553
Table 5.14. Statistical measurement of the developed models and literature models for
prediction two-phase Z factor of two gas-condensate samples as a function of Ppr.

Sample 1 MAE RMSE AARD%

Hall-Yarborough, 0.1753 0.2154 24.04

(1973)
0.1681 0.2064 22.95
DAK, (1974)
0.1680 0.2072 22.78
Beggs & Brill, (1975)
0.2374 0.2720 34.36
Azizi, et al, (2010)
0.0648 0.0707 6.91
Rayes et al., (1992)
0.0271 0.0450 -2.25
FFNN
0.0420 0.0469 -0.0929
CFNN
0.0176 0.0299 1.49
ANFIS

Sample 2 MAE RMSE AARD%

Hall-Yarborough, 0.0752 0.0861 7.41
(1973)
0.0640 0.0731 6.38
DAK, (1974)
0.0299 0.0394 -2.75
Beggs & Brill, (1975)
0.7888 0.7913 79.35
Azizi, et al, (2010)
0.2176 0.2267 22.34
Rayes et al., (1992)
0.0232 0.0384 -1.64
FFNN
0.0752 0.1014 -1.17
CFNN
0.0116 0.0210 0.3764
ANFIS

The proposed approaches in this study are particularly of great interest for prediction
of gas-condensate two-phase Z factor at high temperature and high pressure (HTHP)
operational conditions 𝑇𝑝𝑟 ≥ 3 and 𝑃𝑝𝑟 ≥ 30.
In order to see the impact of each input variables on output of the developed models
in this study, a sensitivity analysis has been carried out. The impact of all input
variables, which are independent variables, (T, P, H2S, CO2, N2, C1 – C7+, MW C7+,
SG) on output parameter, which is dependant variable, (two-phase Z factor) was
investigated. This sensitivity analysis also further reveals gas-condensate fluid

182 | P a g e
behaviour below the dew point pressure. To conduct sensitivity analysis, Pearson
relevancy factor (r) was employed. In this method the relevance importance of each
input variable will be detected by assigning a quantitative scale. The scale is a
normalized values between -1 and 1, where the negative value show the negative
effect of the input variable on output and positive show the positive effect (highly
relevant). If the (r) is zero, it signifies that there is no relation between the input and
output variable. The Pearson relation is defined by the following equation:

∑𝑛𝑖=1(𝐼𝑛𝑝𝑘,𝑖 − 𝐼𝑛𝑝𝑎𝑣𝑒,𝑘 )(𝑍𝑇𝑃,𝑖 − 𝑍𝑇𝑃,𝑎𝑣𝑒 )

𝑟(𝐼𝑛𝑝𝑘 , 𝑍𝑇𝑃 ) =
2 2 5.26
√∑𝑛𝑖=1(𝐼𝑛𝑝𝑘,𝑖 − 𝐼𝑛𝑝𝑎𝑣𝑒,𝑘 ) ∑𝑛𝑖=1(𝑍𝑇𝑃,𝑖 − 𝑍𝑇𝑃,𝑎𝑣𝑒 )

Where, 𝐼𝑛𝑝𝑘,𝑖 and 𝐼𝑛𝑝𝑎𝑣𝑒,𝑘 are ith value and the average value of kth input variables
respectively (k=T, P, H2S, CO2, N2, C1 – C7+, MW C7+, SG); 𝑍𝑇𝑃,𝑖 represents ith values
of the predicted two-phase Z factor and 𝑍𝑇𝑃,𝑎𝑣𝑒 stands for average values of predicted
two-phase Z factor. The relative impact of each parameter on two-phase Z factor
presented in Figure 5.21. The result in Figure 5.21 shows that pressure (P) has the
highest impact on two-phase Z factor of gas-condensate reservoirs below the dew
point pressure. The positive impact of P on Z factor means increasing P would directly
increase the Z factor. Physical law of gas-condensate fluid supports the positive
impact of pressure on Z factor. It is well known from real gas law equation of state for
mixture gases (𝑍 = 𝑃𝑉/𝑛𝑅𝑇), that pressure (P) has positive impact on Z factor and
temperature should have negative impact. These relationships can be explained
further in following and the obtained results are related to the physics of the problem.

The intermolecular connection of a gas mixture diminishes with rising temperature (T)
and cause the gas behaves more like an ideal gas. According to ideal gas law the
temperature should have negative impact on Z factor, means increasing (T) should
decrease the Z factor. The result in Figure 5.21 for temperature impact is against the
aforementioned theory, as the temperature (T) has a positive impact on Z factor of
gas-condensate fluid. The theory is correct as long as the mixture remains as a single-
phase gas. However, below the saturation pressure, the liquid phase evolved from the
gas phase and gas-condensate mixture is no longer single phase. Hence, the mixture
behaves more like a liquid rather than gas especially for rich gas-condensate

183 | P a g e
reservoirs with high amount of condensate (oil) developed inside the pores in the
reservoirs. Hence increasing (T) like pressure contributes to higher Z factor.
The compositional contents of C1 to C7+ has different influence on Z factor. C1, C2 and
C3 do not have considerable effect on Z factor, whereas IC4 to C7+ have positive impact
on Z factor. It is interesting to relate the molecular weight of each composition to the
relevancy factor. For heavier hydrocarbon fraction with higher molecular weight, the
impact is greater in positive direction. IC4 to C7+ have higher molecular weight than C1
– C3, hence they have a higher positive impact on Z factor. This can be concluded that
the higher amount of heavy hydrocarbon (IC4, NC4, IC5, NC5, C6 and C7+) would
results in higher Z factor in the modelling.
Non-hydrocarbon components of H2S, CO2 and N2, all have negative impact on Z
factor. Increasing the molar fraction of H2S, CO2 and N2 would result in decreasing Z
factor.
After pressure the molecular weight of C7+ (MWC7+) has a highest positive impact on
Z factor. This property causes gas-condensate mixture behave like a liquid (oil) phase,
below the saturation pressure. This would increase the deviation from ideal gas
behaviour, which results in larger Z factor.

Figure 5.19. Testing different methods in predicting experimental Z factor as a function of

Ppr for sour gas-condensate sample.

184 | P a g e
Figure 5.20. Prediction of Z factor as a function of Ppr for sweet gas condensate sample.

Figure 5.21. Relevancy factor (r) for each input variables.

185 | P a g e
5.5 Summary

Compressibility factor (Z factor) is one of the most important PVT properties that its
accurate determination is required for reliable well deliverability modelling of gas-
condensate reservoirs above and below the saturation pressure. In this chapter
reliability and accuracy of existing literature models for prediction of two-phase Z factor
below the dew point pressure is initially examined using extensive data bank obtained
from open literature. The data bank is mainly from Constant Volume Depletion (CVD)
tests performed on lean/rich gas-condensate fluids worldwide.
Two iterative techniques of Hall and Yarborough, (1973) and DAK, (1975) as well as
three direct methods of Beggs and Brill, (1973), Rayes et al., (1992), and Azizi et al.,
(2010) utilized for prediction of Z factor below the dew point. Graphical and statistical
error measurements have been carried out and indicate that DAK, (1975) best
performed among literature models with MAE of 0.1300, RMSE of 0.1671 and AARD%
of 2.4945 for estimating gas-condensate two-phase Z factor. DAK method followed by
Hall-Yarborough, (1973), Rayes et al., (1992), Beggs and Brill, (1973) and Azizi et al.,
(2010). The problem with the current literatutre apporoaches is that they are not
representing all range of operational conditions and limited to specific range of
pseudo-reduced pressure (Ppr) and pseudo-reduced temperature (Tpr). Almost all
literature mdoels are limited for determination of two-phase Z factor in high pressure
high temperature (HPHT) conditions.
In this study different approach was taken for modelling two-phase Z factor of gas-
condensate reservoirs below the saturation pressure. Three machine learning (ML)
based models including FFNN, CFNN and ANFIS were implemented for accurate
modelling of two-phase Z factor. Various optimization algorithms such as Levenberg
– Marquardt (LM) and Bayesian regularization (BR) coupled with different number of
hidden layers and neurons have been adopted to achieve best optimum performance
of FFNN and CFNN. Integrity and robustness of FFNN and CFNN examined using
independent data sets. The statistical and graphical error measurements of the
obtained results show high accuracy of both methods for prediction of two-phase Z
factor. Another ML based approach known as ANFIS was also utilized for prediction
of two-phase Z factor of gas-condensate reservoirs below the saturation pressure.
Subtractive clustering technique was utilized and an iterative scheme was developed
for determination of optimum number of IF–THEN fuzzy rules in ANFIS modelling.
186 | P a g e
ANFIS is predicting two-phase Z factor with the best accuracy among other intelligent
models and literature correlations. The developed ANFIS model returned the
experimental values of two-phase Z factor with MAE of 0.00251, RMSE of 0.00254
and AARD% of 0.5119. The developed MA based models overcome the limitation of
existing literature models and they can be used for prediction of gas-condensate two-
phase Z factor below the dew point in all ranges of pressure and temperature
operational conditions. Furthermore the ML based models has an excellent
computational efficiency.
The developed ML based models can be used as an alternative tools instead of
existing literature models in reservoir simulation packages. This would ensure
accuracy and reliability of the production forecast for gas-condensate reservoirs
experiencing condensate drop out.

187 | P a g e
 CHAPTER 6 PRODUCTION PROFILE
FORECAST USING WELL TEST DATA
PRODUCTION PROFILE FORECAST USING WELL TEST
DATA

6.1 Introduction

Accurate and reliable well production forecast of gas-condensate reservoirs are very
important for field development planning, projection of cash flow and future
hydrocarbon recoveries.
Reliable modelling of this type of reservoirs require understanding the effect of
condensate blockage and other contributing factors. Among many factors phase
behaviour, absolute/relative permeability and how the well is being produced are the
most important (Fevang and Whitson, 1996; Shi, 2009).
Phase behaviour of gas-condensate below the saturation pressure is related to
variations in fluid properties (e.g., viscosity, Z factor). These variations would influence
modelling of well deliverability and also production profile forecast. From gas rate
equations of 2.1, 2.2 and 2.16, it is apparent that the cumulative production (Gp) and
flow rate (q) are direct function of permeability, viscosity, solution gas to oil ratio (Rs),
and Z factor. Hence, accurate determination of the aforementioned parameters will
affect the estimation of gas and condensate production rate at the surface. Many
studies in literature have shown that root cause of the unreliable modelling of gas-
condensate well deliverability is due to inaccurate estimation of PVT properties. Some
of these studies are as follow (Rayes et al., 1992; Fevang and Whitson, 1996; Whitson,
Fevang and Yang, 1999; Elsharkawy, Hashem and Alikhan, 2000; Mott, 2002; Sutton,
2005b, 2007; Arukhe and Mason, 2012; Yan et al., 2013; Ghiasi et al., 2014;
Behmanesh, Hamdi and Clarkson, 2017; Khazali, Sharifi and Ahmadi, 2019).

Hence, to address the PVT issues in gas-condensate reservoirs, Artificial Intelligent

algorithms were manipulated extensively in chapter 4 and 5. Several Artificial
Intelligence (AI) or known as machine learning approaches were developed for better
prediction of gas-condensate viscosity and two-phase Z factor. Although the machine
188 | P a g e
learning based models in chapter 4 and 5 predict PVT properties (viscosity and Z
factor) very well, the reliability of the developed models in relation to production profile
of gas-condensate reservoirs is unclear. Therefore, the aim of this chapter is to
investigate the reliability of the developed PVT models in computation of production
profile of gas-condensate wells, which is ultimate aim of this study. In doing so three
regions pseudopressure integral has been utilized and coupled with material balance
equation to generate production profile of a gas-condensate well. To verify the
effectiveness of the developed PVT properties (e.g., viscosity and two-phase Z factor)
in computation of production profile compositional simulation of the studied well
performed using Eclipse 300.
Permeability is another important parameter that would influence the accuracy of well
deliverability modelling in gas-condensate reservoirs. In this chapter effective
permeability of each phase (gas and condensate) will be used for computation of
pseudopressure integral. The detail reservoir and fluid data has obtained from well
pressure test data.

6.2 Pressure Transient Test (PTA) and two-phase pseudopressure approach

Detail reservoir information are essential for reservoir engineers to predict the current
and future performance of the reservoir. Pressure transient analysis (PTA) test is
available in industry to provide engineers with detail reservoir information required for
the analysis. PTA is essentially measuring pressure changes in a well as a function of
time (Lee and Wattenbarger, 1995, pp. 111–113). It is well known that pressure
behaviour of the reservoir following a rate change reflect the geometry and flow
properties of the reservoir. From PTA test the information such as effective
permeability, formation damage/stimulation, volumetric average reservoir pressure,
drainage pore volume, fracture properties and communication between wells can be
obtained.

Classical interpretation of PTA test results for determination of the aforementioned

parameters are based on linear diffusion equation. The combination of Darcy’s law,
continuity equation and equation of state forms the diffusivity equation. The equation
is a differential equation that governs the transient flow of the fluid through a porous

189 | P a g e
medium (Sanni, 2018). The diffusivity equation determines variation of pressure with
time and position in the reservoir as follow.

∅𝜇𝑐𝑡 𝜕𝑝
∇2 𝑝 − =0 6.1
𝑘 𝜕𝑡

Where ∇ devoted to gradient, P is pressure, ∅ is porosity of the medium, 𝜇 is viscosity

𝜕𝑝
of the fluid, 𝑐𝑡 is total compressibility, k is permeability and 𝜕𝑡 is pressure rate of change

with respect to time.

Equation 6.1 is derived using several assumption as follow (Horne, 1995; Stewart,
2011)
 Darcy’s law apply
 Single-phase flow
 Porosity, permeabilities, viscosity and compressibility are constant
 Fluid compressibility is very low
𝜕𝑝
 Pressure gradient in the reservoir is small
𝜕𝑡

 Gravity and thermal effect are negligible

With the above assumptions equation 6.1 is applicable for single-phase slightly
compressible flow (oil reservoir) when average reservoir pressure is above bubble
point pressure (𝑃𝑟 > 𝑃𝑏𝑢𝑏𝑏𝑙𝑒). However, for gas-condensate reservoirs as liquid
condensation may occurs in the reservoir and both gas and liquid phase coexist below
the saturation pressure, equation 6.1 is no longer valid. Furthermore, the properties of
gas-condensate fluid are strong function of pressure and they are changing with time.
Hence, the flow equation for gas-condensate reservoir is non-linear and equation 6.1
is not applicable for such reservoir. A classical approach to account for this non-
linearity is defining a variable known as pseudopressure function (mp). Therefore, the
governing pressure transmission for gas-condensate reservoirs is changing to the
following equation.

1 𝜕𝑚𝑝
∇ 2 𝑚𝑝 − =0 6.2
𝜂 𝜕𝑡

Where 𝜂 is constant of diffusivity equation and it is function of pressure and time.

Equation 6.2 can further be simplified using pseudotime. Computation of
190 | P a g e
pseudopressure and pseudotime allow all solutions (gas-condensate fluid) defined for
standard well test analysis, using pseudopressure and pseudotime instead of pressure
and time (Roussennac, 2001). In the case of multiphase flow for instance gas-
condensate below the saturation pressure the pseudopressure function in 6.2 must
take into account the relative permeability data, at reservoir condition, which is very
difficult to obtain. In this situation, effective permeability can be calculated using
pseudopressure function.

Pseudopressure function (𝑚𝑝 ) is defined in chapter 2 by pressure integral in equation

2.2 after Al-Hussainy et al., (1966). For gas-condensate fluid flow, pseudopressure
function can be combined with analogy of three flow regions introduced by Fevang,
(1995), explained in 2.3 to estimate gas and condensate flow rates at the surface.
Three-flow regions pseudopressure approach is very quick and easy to implement for
well performance modelling of gas-condensate reservoirs. Using PTA data combined
with three regions pseudopressure method resulted in estimating gas flow rate (qg) as
a function of pressures. The established relationship between pressure and flow rates
resulted in generating Inflow Performance Relationship (IPR). IPRs are important
element for reservoir engineers in the design of new wells and also for monitoring and
optimizing existing wells. In this section we discuss how to establish IPRs for gas-
condensate reservoirs using three regions pseudopressure approach when PTA data
is available. Then generating production profile of gas-condensate reservoirs using
results of pseudopressure integral combined with volumetric material balance will be
discussed.

Gilbert, (1954) introduced Inflow Performance Relationship (IPR) for oil wells. IPRs
are important tool in understanding the reservoir/well behaviour and quantifying
production rate (Guo, Sun and Ghalambor, 2008; Fattah et al., 2014). IPRs are
essentially quantifying gas or condensate (oil) flow rates in respect to specific bottom-
hole flowing pressure (Pwf). Rawlins and Schellhardt, (1936) proposed an equation to
relate gas flow rate (qg) to Pwf, their equation is known as back – pressure equation or
deliverability equation and shown in 2.11. To account for non-linearity of the flow (e.g.,
variation of viscosity and Z factor) due to liquid drop out below the saturation pressure

191 | P a g e
in gas-condensate reservoirs, pseudopressure function (𝑚𝑝 ) is added to well
deliverability equation.
Substituting 𝑚𝑝 in 2.11 yields equation 6.3. This equation is effectively gas-
condensate well deliverability equation that can be used for construction of IPR curve.

𝑛
𝑞𝑔𝑡 = 𝐶 (∆𝑚𝑝 𝑔𝑡 )
{ } 6.3
∆𝑚𝑝 𝑔𝑡 = 𝑚𝑝 𝑔𝑎𝑣𝑔 − 𝑚𝑝 𝑔𝑤𝑓

Where C is performance coefficient in Mscf/day/psi2, ∆𝑚𝑝 𝑔𝑡 is total gas

pseudopressure function in psi, 𝑚𝑝 𝑔𝑎𝑣𝑔 is average pseudopressure function, 𝑚𝑝 𝑔𝑤𝑓

pseudopressure function at the bottom-hole flowing pressure, n is an exponent and

𝑞𝑔𝑡 is total gas flow rate on the surface. Depending on flowing velocity, the exponent
n can be vary between 1 for completely laminar flow and 0.5 for fully turbulent flow.
The coefficient of C is depending on well and reservoir geometry, permeability and
fluid properties and its mathematically has been defined in chapter 2 equation 2.3.
The coefficient values of C and n in equation 2.3 can also be determined using well
test data.
By taking the logarithm of both sides of equation 6.3 and solving for logarithm of
pseudopressure function, the expression can be rewritten as:

1 1
𝑙𝑜𝑔 (∆𝑚𝑝 𝑔𝑡 ) = 𝑙𝑜𝑔𝑞𝑔 − log 𝐶 6.4
𝑛 𝑛

The above equation suggests that plotting ∆𝑚𝑃𝑔𝑡 versus 𝑞𝑔 in log-log scale should yield
a straight line with slope of 1/n and intercept of C. The deliverability exponent n can
be determined from any two points on a straight line as follow.

∆𝑙𝑜𝑔𝑞𝑔
𝑛= 6.5
∆log(∆𝑚𝑝𝑔𝑡 )

Also coefficient C can be estimated by rearranging equation 6.3 when pressure test
data is available as follow.

192 | P a g e
 𝑞𝑔𝑡
𝐶= 6.6
∆𝑚𝑝 𝑔 𝑛

The above procedure is a conventional well deliverability estimation of any gas well in
industry using PTA test data. Establishing Inflow Performance Relationship (IPR)
curve using the above procedure and equation 6.3, require calculation of
pseudopressure function ∆𝑚𝑝 𝑔𝑡 .

In this study three-flow regions pseudopressure approach of Fevang, (1995) adopted

for estimation of ∆𝑚𝑝 𝑔𝑡 . ∆𝑚𝑝 𝑔𝑡 is calculated for three flow regions of gas-condensate

reservoirs. Substitution of Fevang’s approach in 6.3, yields the following expression in

terms of gas and condensate effective permeabilities.
𝑃∗ 𝑘. 𝑘𝑟𝑔 𝑘. 𝑘𝑟𝑐 𝑃𝑑𝑒𝑤 𝑘. 𝑘
𝑟𝑔
𝑞𝑔𝑡 = 𝐶 {∫ ( + 𝑅𝑠 ) 𝑑𝑝 + ∫ ( ) 𝑑𝑝
𝑃𝑤𝑓 𝐵𝑔 𝜇𝑔 𝐵𝑐 𝜇𝑐 𝑃∗ 𝐵 𝜇
𝑔 𝑔

𝑛
6.7
𝑃𝑅
𝑘. 𝑘𝑟𝑔 (𝑆𝑤𝑖 )
+∫ ( ) 𝑑𝑝 }
𝑃𝑑𝑒𝑤 𝐵𝑔 𝜇𝑔

This is true representation of gas-condensate flow below the saturation pressure.

Equation 6.7 includes three pressure integrals representing three flow regions inside
gas-condensate reservoirs as discussed in 1.3. Calculation of pressure integrals in 6.7
depends on co-existing of the regions in the reservoirs as explained in section 2.3.

Pressure integrals in 6.7 has written in terms of effective permeability instead of

relative permeability as previously shown in 2.16. This is particularly useful if the well
pressure transient test data is available, where the effective permeability can be
calculated from the well test data. This analogy would be more explored by using a
gas-condensate case study in this chapter.

Using equation 6.7 requires information of pressures, permeabilities, fluid properties

and coefficients of n and C. Well pressure test analysis (PTA) are conventional method
to obtain pressure and permeability of the formation. Several well pressure transient
test methodologies exist in the literature such as pressure build up, pressure
drawdown, multirate, pulse, interference, fall off, injectivity, step rate and fall off. Our
intention is not to discuss different well test methods, as this is beyond the scope of

193 | P a g e
this study. Nevertheless, the aim is to introduce how effective permeability (k.krg) and
skin factor and pressure information that are required for computation of 6.7 can be
calculated from well test data.

Pressure build up test is usually used to provide the reservoir and other properties of
the formation. This test describes the build up in wellbore as a function of time after
the well shut-in for certain time. The general equation of the pressure build up and
drawdown come from solution of diffusivity equation and principle of superposition
theory. Pressure build up equation originally proposed by Horner, (1951) and modified
by Earlougher, (1977). The pressure build up test in terms of pseudopressure function
and surface flow rate for slightly compressible fluid is as follow.
𝑞𝑔,𝑚𝑒𝑎𝑠 𝑘. 𝑘𝑟𝑔(𝑝)
𝑚𝑃𝑟 − 𝑚𝑃𝑤𝑓 = 162.6 ( ) (𝑙𝑜𝑔(𝑡) + 𝑙𝑜𝑔 ( ) − 3.23
ℎ ∅𝜇𝑔 𝑐𝑡 𝑟𝑤 2
6.8
+ 0.87𝑠𝑡𝑜𝑡𝑎𝑙 )

Where 𝑞𝑔,𝑚𝑒𝑎𝑠 is measured gas flow rate at surface during the test; 𝑡 is recorded
pressure test time; h is reservoir thickness; k.krg is effective permeability of gas phase;
∅ is porosity of the media; 𝜇𝑔 is gas viscosity; ct is total compressibility factor; rw is
wellbore radius; and stotal is total skin factor include the skin from formation damage
and stimulation (Horner, 1951; Al-Hussainy and Ramey, 1966; Agarwal, 1979; Lee
and Wattenbarger, 1995)

∆𝑚𝑝 𝑔,1ℎ𝑜𝑢𝑟 𝑘. 𝑘𝑟𝑔

𝑠𝑡𝑜𝑡𝑎𝑙 = 1.153 [ − 𝑙𝑜𝑔 ( ) + 3.227] 6.9
𝑞𝑔,𝑚𝑒𝑎𝑠 𝑚𝑔,1 ℎ𝑜𝑢𝑟 ∅𝜇𝑔 𝑐𝑡 𝑟𝑤 2

Where m is slope of the straight line. Rearranging 6.9 and comparing with straight-line
equation of 𝑦 = 𝑚𝑥 + 𝑏 suggests an analysis technique with the arrangement of the
following terms.
𝑦~𝑚𝑃𝑤𝑓
𝑥~log(𝑡)
𝑞𝑔,𝑚𝑒𝑎𝑠
𝑚~ − 162.6 ( ) 6.10
ℎ
𝑞𝑔,𝑚𝑒𝑎𝑠 𝑘. 𝑘𝑟𝑔(𝑝)
𝑏~ − 162.6 ( ) (𝑙𝑜𝑔 ( ) − 3.23 + 0.87𝑠𝑡𝑜𝑡𝑎𝑙 )
{ ℎ ∅𝜇𝑔 𝑐𝑡 𝑟𝑤 2 }

194 | P a g e
Equation 6.8 indicates that the plot of bottom-hole flowing pressure (Pwf), against
𝑞𝑔,𝑚𝑒𝑎𝑠
logarithm of time (t) exhibits a straight line with the slope of 𝑚~ − 162.6 ( ) and
ℎ
𝑞𝑔,𝑚𝑒𝑎𝑠 𝑘.𝑘𝑟𝑔(𝑝)
intercept of 𝑏~ − 162.6 ( ) (𝑙𝑜𝑔 (∅𝜇 2 ) − 3.23 + 0.87𝑠𝑡𝑜𝑡𝑎𝑙 ). The slope “m”
ℎ 𝑔 𝑐𝑡 𝑟𝑤

allows us to calculate the permeability of the formation (k). The plot commonly known
as Horner plot and the typical semi log plot of Pwf against logarithm of time (t) is shown
in Figure 6.1.

Figure 6.1. Horner plot for pressure build up test (Earlougher, 1977).

An early time deviation from the graph in Figure 6.1 is due to the wellbore storage
effect and skin factor (Horner, 1951; Lee and Wattenbarger, 1995; Ahmed, 2010). This
deviation is large if permeability is low and fluid compressibility is high. This is the case
in heavy gas-condensate reservoir when the liquid build up starts from the beginning
of the production.
In this study Horner plot type analysis also has been adopted for calculation of effective
permeability when PTA data is available. Pseudopressure integral in 6.7 rewritten in
terms of effective permeabilities of the phases instead of relative permeabilities. This
is due to the fact that using relative permeability data, obtained from empirical models
such as Corey et al., (1956) would produce misleading results for very low permeability
gas-condensate reservoirs (Ojha et al., 2017; Hassan et al., 2019).
195 | P a g e
The detail computation of pseudopressure integral is given in following section through
demonstration of a case study. Pressure-volume-temperature (PVT) properties
(viscosity and Z factor) are also an important factors for estimating pseudopressure
function. The objective of this chapter is to investigate how the production of gas (qg)
and condensate (qc) are influenced using different PVT properties and two-phase Z
factor models developed in chapter 4 and 5.

6.3 Computation of production profile in relation to gas/condensate viscosity

and Z factor in HPHT gas-condensate well

The example that we analyse in this section is a vertical well known as (KAL – 5). The
is in Yugoslavia and located in very high temperature and tight formation (365 °F at
11,500 ft [180°C at 3500m]. The compact formation makes the production impossible
without stimulation. Pressure build up test was run in the well and bottom-hole flowing
pressure (Pwf) was recorded against the shut in period (time in hours). The pressure
test data, reservoir and fluid properties of the well is obtained from Economides et al.,
(1989) and presented in Table 6.1 and Table 6.2. The reservoir fluid contains high
content of non-hydrocarbon impurities of 10.9% carbon dioxide (CO2). Initial condition
of the reservoir fluid coincides with most gas-condensate reservoirs with GOR of 9470
scf/STB (1706 std m3/stock-tank m3). The initial reservoir pressure (6750psi
[46.5Mpa]) is identical to dew point pressure and condensation starts from the
beginning of the production. This means the only region that developed during
pressure depletion is region 1 (𝑃∗ > 𝑃𝑅 ). Hence, the pseudopressure integral with
pressure limit between Pwf and PR, shown in 2.12, can be used for calculation of well
deliverability. The obtained data from pressure transient test indicate the well is
located in a very tight formation with the permeability of 0.0035 millidarcy.

196 | P a g e
Table 6.1. Reservoir and fluid property of studied gas-condensate well after (Economides et
al., (1989).

Well & Reservoir data Fluid Data (%mole fraction)

Pi, (psia) 6750 H2S 0.006
Pdew (psia) 6750 N2 1.452
qg, Mscf/day 75.4 CO2 10.931
qc, STB/day 2.8 C1 72.613
GOR scf/STB 9470 C2 6.242
𝛾𝑔 (to air) 0.94 C3 1.631
T, °F 354 i – C4 0.553
h, ft 216.5 n – C4 0.693
hp,ft 36 i – C5 0.442
∅ 0.062 n – C5 0.379
rw, ft 0.54 C6 0.516
Sw 0.3 C7 0.644
Rp,scf/STB 9470 C8 0.541
API [Assumed] 50 C9 0.388
Skin - 4.52 C10+ 2.979
K (md) 0.0035

197 | P a g e
Table 6.2. Pressure build up data for (KAL – 5) obtained from (Economides et al., (1989).

Time (Hours) Pressure (Psi)

0 1083.1
0.167 1174.5
0.333 1226.7
0.5 1303.6
1.0 1490.6
2.0 1751.6
3.0 2046.0
4.0 2279.4
6.0 2759.4
8.0 3246.5
12.0 4221.0
16.0 5162.0
22.0 6161.0
28.0 6336.0
34.0 6406.1
42.0 6452.5
50.0 6487.3
58.0 6507.6
68.0 6526.5
82.0 6556.9
97.0 6587.3
112.0 6587.3
141.0 6601.8

In order to calculate gas and oil maximum flow rates (qg, qo) using pseudopressure
integral and construct the IPR curves the following data is needed.
 The PVT properties of each phase in region 1, (construction of PVT tables)
 Producing gas/oil ratio, RP (obtained from well test data)
 Knowledge of pressure, Pwf and PR (obtained from well test data)
 Effective permeabilities of each phase (k.krg) and (k.kro)

198 | P a g e
In the following section, first we discuss the construction of PVT tables and then
mathematical solution of pseudopressure integral will be discussed. Effective
permeability of each phase will be calculated using well pressure test data.

Construction of pressure – volume – temperature (PVT) relationship

To estimate maximum gas and oil (condensate) flow rate at the surface using
psuedopressure approach, knowledge of PVT data such as formation volume factor,
viscosity, Z factor and solution gas to oil ratio is required. Viscosity and Z factor are
governing parameters in computation of pseudopressure integral and determinining
the performance of the well (Whitson, Fevang and Yang, 1999; Hernandez; et al.,
2002; Yang et al., 2007; Arukhe and Mason, 2012). In chapter 4 and 5 accurate
determination of viscosity and Z factor of gas-condensate reservoirs have been
comprehensivly discussed.
It also has been shown that using current literature approaches for modelling gas-
condensate viscosity and Z factor resulted in high error. Subsequently several
machine learning (ML) based models have been developed and presented in chapters
4 and 5. The developed ML based models were adopted for computation of fluid
properties and constructing PVT tables for studied gas-condenste well. These models
were developed to address the non-linearity of fluid flow below the saturation pressure
in gas-condensate reservoirs undergoing depletion.
Other two PVT properties that have major influence for accurate determination of the
pseudopressure integrals are solution GOR (Rs) and oil to gas ratio (rs) (Fevang and
Whitson, 1996; Mott, 2002). Whitson and Torp, (1983) MBO method proved to be one
of the most accurate techniques for calculation of Rs and rs for gas-condensate fluid

199 | P a g e
(Borthne, 1986; Coats and Smart, 1986; Guo, Du and School, 1989; Khamis and
Fattah, 2019).
Therefore, Whitson and Torp, (1983) utilized for computation of Rs and rs in region 1.
The steps shown in following figure have been taken for construction of PVT table.

Figure 6.2. construction of PVT table for KAL – 5 gas-condensate well.

The constructed PVT table in this part will be used for computation of gas and
condensate flow rates utilizing three regions pseudopressure approach.

Mathematical manipulation of pseudopressure integral

The well deliverability equation for this reservoir rewritten only for region 1 as
described in 6.3 and depicted in 6.11 for gas phase and 6.12 for oil (condensate)
phase.

𝑛
𝑃𝑅
𝑘. 𝑘𝑟𝑔 𝑘. 𝑘𝑟𝑜
𝑛
𝑞𝑔 = 𝐶(∆𝑚𝑃𝑔1 ) = 𝐶 {∫ ( + 𝑅 ) 𝑑𝑝 } 6.11
𝑃𝑤𝑓 𝐵𝑔 𝜇𝑔 𝐵𝑜 𝜇𝑜 𝑠

𝑛
𝑃𝑅
𝑘. 𝑘𝑟𝑜 𝑘. 𝑘𝑟𝑔
𝑞𝑜 = 𝐶(∆𝑚𝑃𝑐1 )𝑛 = 𝐶 {∫ ( + 𝑅 ) 𝑑𝑝 } 6.12
𝑃𝑤𝑓 𝐵𝑜 𝜇𝑜 𝐵𝑔 𝜇𝑔 𝑜

Mathematical solution of pseudopressure integrals in 6.11 and 6.12 required

producing gas to oil ratio, (Rp). Rp is a ratio of total gas production to total oil production

200 | P a g e
on the surface and can be estimated from following equation (Fetkovich et al., 1986;
Fevang and Whitson, 1996; Guehria, 2000; Jokhio and Tiab, 2002; Ahmed, 2010).
𝑘𝑟𝑔 𝑘
𝐶 [(𝐵 𝜇 ) + (𝐵 𝑟𝑜𝜇 ) 𝑅𝑠 ]
𝑞𝑔𝑡 𝑔 𝑔 𝑜 𝑜
𝑅𝑝 = = 6.13
𝑞𝑜𝑡 𝑘𝑟𝑔 𝑘
𝐶 [(𝐵 𝜇 ) 𝑅𝑜 + (𝐵 𝑟𝑜𝜇 )]
𝑔 𝑔 𝑜 𝑜

Where, RO is oil to gas ratio in (STB/scf), 𝑞𝑔𝑡 and 𝑞𝑜𝑡 are total gas flow rate and total
oil flow rate at the surface respectively. Simplificaion of 6.13 yeilds the following
equation.

𝑘𝑟𝑔 𝐵𝑜 𝜇𝑜
𝑅𝑝 = 𝑅𝑠 + ( )( ) (1 − 𝑅𝑜 𝑅𝑝 ) 6.14
𝑘𝑟𝑜 𝐵𝑔 𝜇𝑔

Fetkovich et al., (1986), rearranged and solved 6.14 for relative permeabilities ratio
(𝑘𝑟𝑔/𝑘𝑟𝑜 ) and proposed the following equation.

𝑘𝑟𝑔 (𝑅𝑝 − 𝑅𝑠 ) 𝐵𝑔 𝜇𝑔
(𝑃) = ( ) 6.15
𝑘𝑟𝑜 (1 − 𝑅𝑜 𝑅𝑝 ) 𝐵𝑜 𝜇𝑜

Relative permeabilities Krg and Kro can be expressed directly as a function of ratio
(𝐾𝑟𝑔/𝐾𝑟𝑜), when both phases are mobile (Evinger and Muskat, 1942). Using 6.15, for
a given Rp, relative permeabilities of gas (Krg) and condensate (Kro) in region 1 can be
evaluated directly as a function of pressure, 𝐾𝑟𝑔(𝑝) = 𝑓[𝑘𝑟𝑔/𝑘𝑟𝑜(𝑝)] and 𝐾𝑟𝑜(𝑝) =
𝑓[𝑘𝑟𝑔/𝑘𝑟𝑜(𝑝)]. Presenting 6.14 in terms of effective permeabilities (k.krg) and (k.kro)
and rearranging the equation allow us to calculate effective permeability of one phase
as a function of other phase. The effective permeabilities of gas and oil phase in region
1, then can be calculated using following expressions.

(𝑅𝑝 − 𝑅𝑠 ) B𝑔 μg {𝑘. 𝑘𝑟𝑜 }

𝑘. 𝑘𝑟𝑔 = ( ) 6.16
(1 − 𝑅𝑝 ) B𝑜 μ𝑜

(1 − 𝑅𝑜 𝑅𝑝 ) B𝑜 μ𝑜 {𝑘. 𝑘𝑟𝑔 }
𝑘. 𝑘𝑟𝑜 = ( ) 6.17
(𝑅𝑝 − 𝑅𝑠 ) B𝑔 μg

201 | P a g e
Substituting 6.16 in pseudopressure integral of 6.11 and simplifying yields
pseudopressure integral in terms of gas effective permeability as follow (Fetkovich et
al., 1986; Fevang, 1995; Fevang and Whitson, 1996; Guehria, 2000; Jokhio, 2002).

𝑃𝑅 𝑅𝑝 (1 − 𝑅𝑜 𝑅𝑠 )
𝑘. 𝑘𝑟𝑔
∆𝑚𝑝 𝑔1 = ∫ ( ) (𝑃)𝑑𝑝 6.18
𝑝𝑤𝑓 𝐵𝑔 𝜇𝑔 (𝑅𝑝 − 𝑅𝑠 )

Pseudopressure integrals can now be computed if effective permeability integral as

well as other properties of respective phase are known. Gas phase pseudopressure
integral in 6.18 can be rewritten without effective permeability term (Jokhio, 2002) as
shown in 6.19. This permits to estimate the effective permeability integral using Horner
plot and theory of well testing.

𝑃𝑅 𝑅𝑝 (1 − 𝑅𝑜 𝑅𝑠 )
1
∆𝑚𝑝 𝑔1 = [∫ ( ) (𝑃)𝑑𝑝] = ∆𝑚𝑝 /𝑀𝑔
𝑝𝑤𝑓 𝐵𝑔 𝜇𝑔 (𝑅𝑝 − 𝑅𝑠 ) 𝑔1 6.19

The above integral is function of bottom hole flowing pressure (Pwf), average reservoir
pressure (PR) and other PVT properties. Pwf and PR obtained from well test data and
PVT properties were calculated as discussed in previous section. The integral is
assumed to be equal to a term in right hand side (∆𝑚𝑝 𝑔1 /𝑀𝑔). The trapezoidal rule of

integration is utilized for integration part of 6.19.

From theory of the well testing during pressure transient period for dimensionless
𝑡𝑖𝑚𝑒 (𝑡) > 50 gas phase effective permeability integral is determined from the
following equation. The effective permeability integral is also assumed to be equal to
a special term known as Mg.

𝑃𝑅 𝑞𝑔,𝑚𝑒𝑎𝑠
∫ 𝑘. 𝑘𝑟𝑔 (𝑃) 𝑑𝑝 = 162.6 = 𝑀𝑔
𝑃𝑤𝑓
𝑑∆𝑚𝑃𝑔1
ℎ( ) 6.20
𝑑𝑙𝑛(𝑡)

The above equation indicates that (K.Krg) integral in two-phase system is inversely
proportional to the derivate of gas pseudopressure (𝑚𝑝 𝑔 ) with natural logarithmic of

time (Horner, 1951; Serra, Peres and Reynolds, 1990, 2007; Dake, 2001; Jokhio,
202 | P a g e
2002). On semi log plot of time against pseudopressure, the rate of change of
pseudopressure is a slope of straight line. The plot of pseudopressure (𝑚𝑝 ) against
the recorded time during well test for (KAL – 5) is shown in Figure 6.3. This is a Horner
plot, where the straight portion of the line represents the effective permeability of the
formation defined by 6.20. This equation is valid for fully developed flow where
pressure waves have crossed the skin and the wellbore storage effect. This means
the enough time should be devoted to the pressure transient test to allow pressure
waves are fully developed.
Where, 𝑑∆𝑚𝑃/𝑑𝑙𝑛(𝑡) is the derivative function of each phase that can be estimated
from 6.21 (Jokhio, 2002; Jokhio, Tiab and Escobar, 2002).

Figure 6.3. Horner plot for KAL – 5 gas-condensate well.

𝑑∆𝑚𝑃𝑖−1 𝑑∆𝑚𝑃𝑖+1
𝑑∆𝑚𝑃 ( ) ∆𝑙𝑛(𝑡)𝑖+1 + ( ) ∆𝑙𝑛(𝑡)𝑖−1
𝑑𝑙𝑛(𝑡)𝑖−1 𝑑𝑙𝑛(𝑡)𝑖+1 6.21
( ) =
𝑑𝑙𝑛(𝑡) 𝑖 [∆𝑙𝑛(𝑡)𝑖+1 + ∆𝑑𝑙𝑛(𝑡)𝑖−1 ]

The point i in above equation is the point, where the derivative is calculated and point
(i-1) is the point before it and (i+1) is the point after it. Pseudopressure difference
(𝑑∆𝑚𝑃) is the difference between the (𝑚𝑝 ) at each pressure and initial (𝑚𝑝 ) at zero

203 | P a g e
hour (𝑑∆𝑚𝑝 = 𝑚𝑝 − 𝑚𝑝 (𝑡=0) ). This is the difference in pseudopressure of any given

pressure and pseudopressure at the beginning of the pressure build up test.

Computing pseudopressure function and effective permeability by the above

procedure allow us to estimate the accurate value of ∆𝑚𝑃𝑔 in region 1 as follow.

∆𝑚𝑝 𝑔1
∆𝑚𝑃𝑔1 = ( ) × 𝑀𝑔
𝑀𝑔 6.22

For modelling condensate (oil) phase in KAL – 5, gas-condensate well similar

approach as explained for gas phase is employed. Condensate (oil) phase flow rate
is determined from the following equation (Penula, 2003).

𝑛
𝑃𝑅
𝑘. 𝑘𝑟𝑜 𝑘. 𝑘𝑟𝑔
𝑞𝑜𝑡 = 𝐶(∆𝑚𝑃𝑜1 )𝑛 = 𝐶 {∫ ( + 𝑅 ) 𝑑𝑝 } 6.23
𝑃𝑤𝑓 𝐵𝑜 𝜇𝑜 𝐵𝑔 𝜇𝑔 𝑜

Where 𝑞𝑜𝑡 is the total condensate flow rate at the surface in (STB/day); ∆𝑚𝑝 𝑜1is two-

phase pseudopressure function in psi2/cp; k.kro is effective permeability to oil in Darcy

unit; 𝑘. 𝑘𝑟𝑔 is effective permeability to gas in Darcy unit; Bo is condensate (oil) formation
volume factor in Barrel/ STB; 𝜇𝑜 is condensate (oil) viscosity in centipoise; Bg is gas
formation volume factor in cubic feet (ft3)/standard cubic feet (scf); 𝜇𝑔 is gas viscosity
in centipoise and Ro is oil to gas ratio in (STB/scf).

Effective permeability to condensate (oil) phase in Region 1 has previously defined in

6.17. Substituting 6.17 into 6.23 and simplifying yields ∆𝑚𝑝 𝑜1, which represents

condensate (oil) phase pseudopressure function in terms of effective permeability

(Evinger and Muskat, 1942; Fetkovich et al., 1986; Fevang, 1995; Guehria, 2000;
Penula, 2003).

𝑃𝑅
𝑘. 𝑘𝑟𝑜 (1 − 𝑅𝑜 𝑅𝑠 )
∆𝑚𝑝 𝑜1 = ∫ ( ) (𝑃)𝑑𝑝 6.24
𝑝𝑤𝑓 𝐵𝑜 𝜇𝑜 (1 − 𝑅𝑜 𝑅𝑝 )

204 | P a g e
Oil phase pseudopressure integral can be rewritten without oil effective permebility
term in the folllowing form (Jokhio, 2002).

𝑃𝑅
1 (1 − 𝑅𝑜 𝑅𝑠 )
∆𝑚𝑝 𝑜1 = [∫ ( ) (𝑃)𝑑𝑝] = ∆𝑚𝑝 /𝑀𝑜 6.25
𝑝𝑤𝑓 𝐵𝑜 𝜇𝑜 (1 − 𝑅𝑜 𝑅𝑝 ) 𝑜1

The above psudopressure integral is assumed to be equal to a special term

∆𝑚𝑝 𝑜1 /𝑀𝑜, that later can be used to simplify and determine ∆𝑚𝑝 𝑜1.

Oil effective permeability integral can be defined according to well test thory. Oil phase
effective permeability integral in two-phase system is inversely proportional to the
derivative of oil phase pseudopressure (𝑚𝑝𝑜 ) with natural logarithmic of time and can
be defined as follow.
𝑃𝑅
𝑞𝑜,𝑚𝑒𝑎𝑠
∫ 𝑘. 𝑘𝑟𝑜 (𝑃) 𝑑𝑝 = 162.6 = 𝑀𝑜
𝑃𝑤𝑓
𝑑∆𝑚𝑝 𝑜1
ℎ( ) 6.26
𝑑𝑙𝑛(𝑡)

𝑑∆𝑚𝑝
Where pseudopressure funtion and its derivative ( 𝑑𝑙𝑛(𝑡)𝑜1 ) in above equation can also

be calculated by utilizing 6.21. Mo is a special term that used for simplifying the
equation. Using special terms at the right hand side of equations 6.25 and 6.26, the
oil (condensate) phase pseudopressure integral is determined as follow.

∆𝑚𝑝 𝑜1
∆𝑚𝑃𝑜1 = ( ) × 𝑀𝑔
𝑀𝑔 6.27

The production rate for well in low permeability formation is usually low that pressure
drop caused by non-Darcy flow can be neglected. Reservoirs in compacted formation
with very low permeability are usually stimulated (fractured), which eliminates the
effect of non-Darcy flow (Fevang, 1995). Hence, the effect of non-Darcy flow on
production rate has been excluded and the effect of condensate blockage on well
production is our primary concern. Another assumption in our calculations is steady

205 | P a g e
state flow in the reservoir, meaning mass flow rate would be constant and there is no
net accumulation of any fluid in the reservoir.
Estimation of pseudopressure functions (∆𝑚𝑝 𝑔1 ) and (∆𝑚𝑝 𝑜1 ) in region 1 by above

procedure, permits to calculated total gas and condensate (oil) flow rate on the surface
using 6.11 and 6.12 respectively. The coefficients of n and C in aforementioned
equations can be estimated using semi-log plot of ∆𝑚𝑝 𝑜1 versus measured surface

flow rates (gas and oil). The semi-log plot is a straight line with the slope of ‘n’ and
intercept of ‘C’. These relationship demonstrated in 6.4 to 6.6 (Forchheimer, 1901;
Rawlins and Schellhardt, 1936). The graph in Figure 6.4 depicts plot of measured gas
flow rate against total gas pseudopressure function. Using straight portion of the line
on the graph, n is 0.8 and coefficient C is 0.0948. Consequently maximum gas flow
rate can be calculated using 6.11 for gas phase and 6.12 for condensate (oil) phase
and IPR curves can be established.

Our ultimate goal in this study is to see the effect of variation in viscosity and two-
phase Z factor estimation using various approaches on computation of production
profile of gas-condensate wells. In order to generate production profile of the well,
pseudopressure approach incorporated with volumetric material balance equation.
Production profile is forecasting expected flow rate of wells as a function of time. This
is very important for assessing economic attractiveness of any reservoir engineering
projects. Moreover it is a reflection of gas-condensate reservoir performance
modelling.

206 | P a g e
 Figure 6.4. Plot of gas flow rate vs total ∆𝑚𝑃𝑔𝑡 for KAL – 5.

Material balance calculation

To relate production rate, calculated from pseudopressure integral, to time the

volumetric material balance equation can be used (Guehria, 2000; Mott, 2002).
Volumetric material balance provides estimation of hydrocarbon reserves in any stage
of a reservoir depletion based on conservation of mass. It is also a tool to indicate
cumulative gas and oil production from a reservoir with a stepwise pressure depletion.
Tarner, (1944) developed a volumetric material balance that originally proposed for
solution – gas drive reservoirs. Tarner volumetric material balance is utilized to relate
the rate of production to time. This method is an iterative technique to find the
acceptable value of producing gas to oil ratio ‘Rp’ during pressure depletion process
(Tarner, 1944). The method adhere the following assumptions (Standing, 1979).
 The reservoir formation is homogeneous with respect to permeability, porosity,
fluid saturation and relative permeability.
 Gravity forces affecting fluid flow are negligible.
 Pressure is uniform throughout the reservoir.
 Equilibrium exists at all time between the saturated gas and the liquid drop out
that has been evolved from saturated gas.
 Reservoir hydrocarbon pore volume remains constant.

207 | P a g e
Original form of Tarner’s material balance method simplified by Dake, (1978, pp. 81–
82) as shown in following.

𝑁𝑃 [(𝐵𝑜 − 𝑅𝑠 𝐵𝑔 ) + 𝐺𝑝 𝐵𝑔 ]
𝑁=
(𝐵𝑜 − 𝐵𝑜𝑖 ) + (𝑅𝑠𝑖 − 𝑅𝑠 )𝐵𝑔 6.28
Where N is cumulative production, Np is oil cumulative production and Gp stands for
gas cumulative production. Multiplying components in 6.28 are groups of
thermodynamic variables (defined in pseudopressure integral) that they can be
summarized by two distinct multipliers for condensate (oil) production ɸ𝑐 and gas ɸ𝑔 .
𝐵𝑜 − 𝑅𝑠 𝐵𝑔
ɸ𝑐 =
(𝐵𝑜 − 𝐵𝑜𝑖 ) + (𝑅𝑠𝑖 − 𝑅𝑠 )𝐵𝑔
6.29
𝐵𝑔
ɸ𝑔 =
{ (𝐵𝑜 − 𝐵𝑜𝑖 ) + (𝑅𝑠𝑖 − 𝑅𝑠 )𝐵𝑔 }

Substituting ɸ𝑐 and ɸ𝑔 in 6.28 yields the following equation.

𝑁 = 𝑁𝑝 ɸ𝑛 + 𝐺𝑝 ɸ𝑔 6.30

Where Np is condensate (oil) cumulative production and Gp is cumulative gas

production. The instantaneous, producing gas-oil ratio, Rp, given as
∆𝐺𝑝
𝑅𝑃 =
∆𝑁𝑝 6.31
The value of Rp is the average of produced gas to oil ratio “𝑅𝑃,𝑎𝑣𝑒𝑟𝑎𝑔𝑒 ” within each
pressure interval. In a stepwise pressure reduction between pressure intervals 𝑖 and
𝑖 + 1, the ∆𝑁𝑝𝑖→𝑖+1 for cumulative production of 1STB oil can be calculated as follow.

1 − 𝑁𝑝𝑖 ɸ𝑛,𝑎𝑣𝑒 − 𝐺𝑝𝑖 ɸ𝑔,𝑎𝑣𝑒

∆𝑁𝑝𝑖→𝑖+1 =
ɸ𝑛,𝑎𝑣𝑒 − 𝑅𝑃,𝑎𝑣 ɸ𝑔,𝑎𝑣𝑒 6.32
For each subsequent pressure intervals, the calculation steps is summarized in
following steps:
1. Define the pressure intervals, ΔP (200psia – 500psia).
2. Determine ɸ𝑐 and ɸ𝑔 using 6.29 and updated PVT properties obtained for
average pressure within each pressure intervals.

208 | P a g e
 3. Assume a value of 𝑅𝑝, 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 = 𝑅𝑝, 𝑔𝑢𝑒𝑠𝑠 in the interval (as an initial guess
the Rp assumed between 9000 – 20000 scf/STB).
4. Calculate ∆𝑁𝑝𝑖→𝑖+1 from 6.32.
5. Calculate cumulative gas production 𝐺𝑝𝑖→𝑖+1 in each pressure intervals
from (∆𝐺𝑝𝑖→𝑖+1 = ∆𝑁𝑝𝑖→𝑖+1 × 𝑅𝑝, 𝑔𝑢𝑒𝑠𝑠).
6. Calculate the Condensate (oil) saturation from

𝑁𝑝 𝐵𝑜
𝑆𝑜 = (1 − ) (1 − 𝑆𝑤 ) 6.33
𝑁 𝐵𝑜𝑖
7. Knowing absolute permeability (Table 6.1) and the effective permeability of
each phase, the relative permeability of gas and condensate (oil) phase can be
calculated from the following.
𝑘. 𝑘𝑟𝑔
𝑘𝑟𝑔 =
{ 𝑘 }
𝑘. 𝑘𝑟𝑜
𝑘𝑟𝑜 = 6.34
𝑘
8. Estimate produced gas to oil ratio 𝑅𝑝,𝑐𝑎𝑙𝑐 from the following equation.

𝐾𝑟𝑔𝜇𝑜 𝐵𝑜
𝑅𝑝,𝑐𝑎𝑙𝑐 = 𝑅𝑠 +
𝐾𝑟𝑐𝜇𝑔 𝐵𝑔 6.35

9. Compare two values of 𝑅𝑝,𝑔𝑢𝑒𝑠𝑠 in step 3 and 𝑅𝑝,𝑐𝑎𝑙𝑐 in step 8. If they do not
converge, repeat steps 3 to 8 with a new value of the 𝑅𝑝,𝑔𝑢𝑒𝑠𝑠 until convergence
occurs.
10. With average reservoir pressure (PR) from each pressure intervals calculate
gas/condensate flow rates from pseudopressure integral.
11. The gas production rates 𝑞𝑔 and 𝑞𝑐 then can be combined with obtained gas
cumulative production (Gp) and condensate (oil) cumulative production (Np).
12. Gas and condensate production profile can be calculated as a function of time.
13. The time of depletion for each pressure intervals is calculated from the
following equation.

∆𝐺𝑃
𝑡=
𝑞𝑔 6.36
The explained computation of material balance incorporated with well inflow
performance calculation using pseudopressure approach to generate production
209 | P a g e
profile of the well. In computation of pseudopressure integral the viscosity and two-
phase Z factor of gas and condensate phase were calculated using existing literature
models and the developed AI models in chapter 4 and 5.

Validation of new production profile

To validate the gas and condensate production results, obtained from three-flow
regions pseudopressure approach for the studied gas-condensate (KAL – 5)
commercial numerical reservoir simulator Eclipse 300 was performed.
Initially Constant – Volume – Depletion (CVD) test on reservoir fluid compositions
(table 6.1) was performed using PVTi (Schlumberger) and Peng – Robinson cubic
equation of state. The Peng – Robinson is industry standard choice for
characterization of gas-condensate fluid properties. This is also known as full
compositional simulation to ensure highest accuracy of the PVT properties.
The well (KAL – 5) is located in a compact formation with absolute permeability of
0.0035 millidarcy, which indicate without stimulation the production is impossible.
Therefore, in our model multiple conductive fractures were created in two X and Y
directions inside the reservoir. The conductivity of each fracture is 500 millidarcy per
foot, which allow gas flows towards the wellbore.
Then a vertical well (KAL– 5) placed in middle of the created geometry. The reservoir
and fluid details are obtained from Economides et al., (1989) as shown in Table 6.1.
This gas-condensate well is in a high temperature (354°F [179°C]) formation of
Pannonian basin in Yugoslavia. The field covers a large area of 131234 ft2 (40 km2).
The outer well drainage area of the model is 2400 ft2 (731.52m2) where a vertical well
placed in the middle of the model with the pay zone of 36ft (10.97m). The schematic
illustration of the created model is shown in Figure 6.5. The details of the developed
Eclipse reservoir simulator code and created fractures can be found in Appendix D.

210 | P a g e
 Figure 6.5. Schematic illustration of 3D created model in Eclipse 300.

Initial condition of the reservoir fluid coincides with most gas-condensate systems with
producing GOR of 9470 scf/STB (1706 std m3/stock-tank m3). The initial reservoir
pressure (6750psi [46.5Mpa]) is identical to dew point pressure and condensation
starts from the beginning of the production. The result of CVD in Figure 6.6 indicates
the reservoir fluid is a rich gas-condensate with maximum condensate saturation of
25.2%. Three parameters Peng – Robinson cubic equation of state utilized for full
compositional simulation of the reservoir fluid properties. Multi-flash experiment of the
reservoir fluid in standard condition of 14.696 psi and 60˚F has been performed and
fluid phase diagram was generated and shown in Figure 6.7. The simulation was run
and generated results were recorded in terms of production profile. The production
profile indicates how much gas or condensate is produced as a function of time.

Figure 6.6. Condensate saturation curve a result of the CVD experiment of the reservoir
fluid.
211 | P a g e
 Figure 6.7. Phase diagram of the reservoir fluid in standard pressure and temperature.

The result of compositioanl simulation will be compared to the results of three-flow

regions psudopressure approach using optmized MBO model and the developed
viscosities and Z factor models.
In order to optimize MBO model, that can be used in three regions pseudopressure
approach, several condensate viscosity models of optimized form of Beggs and
Robinson, (1975); De Ghetto et al., (1994); Elsharkawy and Alikhan, (1999);
Kartoatmodjo and Schmidt, (1991); LBC, (1964) and four machine learning based
models of ANN, LSSVM, TSK fuzzy and Mamdani fuzzy developed in chapter 4 were
used. Among the developed oil (condensate) viscosity methods TSK fuzzy approach,
Artificial Neural Network (ANN) and least square support vector machine (LSSVM)
outperformed other optimized methods. Therefore, aforementioned models have been
used for prediction of oil (condensate) viscosity in generating MBO PVT table. In order
to compute gas phase viscosity the optimized version of Londono et al., (2002) shown
in equation 4.15 was used.
For better computation of two-phase Z factor three developed machine learning based
models of cascade forward neural network (CFNN), feed forward neural network
(FFNN) and adaptive neuro fuzzy inference system (ANFIS) in chapter 5 were
considered. ANFIS model performed better than other two methods for prediction of
two-phase Z factor. Hence, ANFIS was utilized for computation of two-phase Z factor
of KAL – 5 gas-condensate well in MBO PVT table. Once the gas and condensate
MBO PVT tables were generated, the results used for computation of gas and
condensate flow rates using three regions pseudopressure integral. The results from
212 | P a g e
three regions pseudopressure integral were combined with material balance equation
and production profile of gas and condensate were established. This would allow the
comparison of the obtained results to compositioanl reservoir simulation results using
Eclipse 300.

Results and discussion

In presenting the results first the behaviour of studied gas-condenste fluid in the
reservoir will be explored and then the effect of inaccurate PVT data and two-phase Z
factor in relation to production profile will be discussed.
The graph in Figure 6.8 shows how the fluid behaves in terms of viscosity variation as
a function of pressure. Increasing reservoir pressure would proportioanlly increase the
gas phase viscosity as intermolecular connection of the gas phase become stronger
and the gas behave like a liquid in very high pressure. Contrary to the gas phase the
liquid phase (condensate) viscosity is decreased with increasing the reservoir
pressure as the liquid behave like a gas in very high temperature due to reduction in
intermolecular forces (obsorption and repulsion). The result of Figure 6.8 confirms the
aformentiond criterion in regards to physical behaviour of the fluid as a function of
pressure. It also shows that there is a strong relation between presure and viscosity
of gas/condensate and viscosity is extremly non-linear. Before analysisng the affect of
various gas/condensate viscosity and Z factor on the production results, we discuss
permeability affect.

Figure 6.8. Variation of gas and oil (condensate) voscosity with pressure in KAL – 5.

213 | P a g e
Permeability alteration of the reservoir is a critical issue because of the condensate
drop out. Figure 6.9 shows that gas effective permeability is sharply declined after the
production begins. When reservoir pressure is reduced from initial reservoir pressure
of 6750psia to 5850psia (the red line on the graph), the effective permeability of
formation to the gas phase stabilized. The relative permeability of each phase
calculated using equation 6.32 and the ratio of the krg to kro (krg/kro) is estimated and
plotted against pressure in Figure 6.9. Changes in relative permeability of each phase
due to the condensation drop out divided into five stages as it demonstrated in Figure
6.10. Each stage are discussed in following.
 Stage 1: The ratio of krg to kro (krg/kro) is high at the beginning of the production
(pi=6750psia), however as the production commence, severe decline in gas
phase relative permeability occurs, where krg lost its ratio by 230 times to kro.
 Stage 2: The gas phase relative permeability continuous to decline further with
a fast rate between 4200 𝑝𝑠𝑖𝑎 < 𝑃 < 6050𝑝𝑠𝑖𝑎. This is due to further
accumulation of liquid drop out in region 1.
 Stage 3: At this stage the gas relative permeability (krg) reduction would almost
stabilizes. At the beginning of this stage, the liquid drop out might reaches its
maximum saturation and after this point, it starts to decrease due to
vaporization (this also known as retrograde behaviour). After vaporization
condensate liquid produced as a gas form at the surface (Danesh, 1998, pp.
26–27).
 Stage 4: Further decline in pressure would increase the gas phase relative
permeability. This means that the gas phase saturation inside the reservoir
increases. This behaviour might imply the fact that the further reduction of
reservoir pressure contributes to revaporization of the condensate phase that
dropped out during pressure depletion.

214 | P a g e
 Figure 6.9. Gas phase effective permeability.

Figure 6.10. Relative permeability ratio (gas to oil) as a function of pressure for high
temperature (354⁰F) gas condensate well.

Figure 6.11 and Figure 6.12 depict the generated production profile for gas and
condensate rate of the studied well. From both graphs, it is evident that three-flow
regions pseudopressure approach incorporated with optimized MBO PVT properties
(e.g., TSK fuzzy for condensate viscosity and ANFIS for two-phase Z factor), is
following the results of compositional simulation very well. In presenting production
profiles, only various viscosity models have been labelled on the graphs for
comparison, however in the computation process, the developed ANFIS used to
generate two-phase Z factor in all models.

215 | P a g e
Statistical accuracy of the obtained results examined using two statistical metrics of
average absolute relative deviation percentage (AARD %) and root mean square error
(RMSE) shown in 6.37.

Figure 6.11. Gas production profile for KAL – 5 using analytical method incorporated with
various ML techniques in generating PVT properties.

Figure 6.12. Gas production profile for KAL – 5 using analytical method incorporated with
various ML techniques in generating PVT properties.

𝑁
100 |(𝑆𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛. (𝑖) − 𝐴𝑛𝑎𝑙𝑦𝑡𝑖𝑐𝑎𝑙. (𝑖))|
𝐴𝐴𝑅𝐷% = ∑ × 100
𝑁 𝑆𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛. (𝑖)
𝑖−1 6.37
2 0.5
∑𝑁𝑖(𝐴𝑛𝑎𝑙𝑦𝑡𝑖𝑐𝑎𝑙. (𝑖) − 𝑆𝑖𝑚𝑢𝑙𝑎𝑡𝑖𝑜𝑛. (𝑖))
𝑅𝑀𝑆𝐸 = ( )
{ 𝑁 }

216 | P a g e
The results of statistical error analysis presented in Table 6.3, Figure 6.13 and Figure
6.14. From the results it has been observed that pseudopressure integral incorporated
with TSK fuzzy condensate viscosity model and ANFIS model for calculating two-
phase Z factor outperform other two methods with the least RMSE of 0.0224 and
0.1441 for gas phase and oil phase respectively in first 12 months of production
forecast. In terms of AARD% the analytical model incorporated with TSK fuzzy
approach condensate viscosity outperformed other methods with 0.856% for gas
phase and 2.398% for condensate phase. The results confirm using accurate
prediction of PVT properties (viscosity and two-phase Z factor) below the saturation
pressure is critical for forecasting gas-condensate wells using three-flow regions
pseudopressure integral and MBO model. The optimized MBO PVT model using
intelligent methods ensure the accurate production profile of gas-condensate wells in
compact formation.
Gas and condensate production from the studied well (KAL – 5) declined rapidly after
45 days. This is due to accumulation of the condensate liquid inside the reservoir,
which directly effect the gas phase relative permeability (Figure 6.10 and Figure 6.11).
This has a direct impact on forecast of the future performance of such reservoirs.
Successful application of three regions pseudopressure integral is depending on
reliable prediction of viscosity and two-phase Z factor in generating PVT tables.
Machine learning based approaches provide a promising results for better prediction
performance accuracy of gas-condensate wells in tight formation.

Table 6.3. Statistical error analysis of the developed analytical technique for predicting gas
and condensate rates in (KAL – 5).

Method RMSE AARD%

Pseudopressure (ANN. 0.0829 2.358
ANFIS)
Gas
Pseudopressure (LSSVM, 0.0302 1.098
phase
ANFIS)
Pseudopressure (TSK
0.0224 0.856
Fuzzy, ANFIS)
Pseudopressure (ANN, 0.1945 3.375
ANFIS)
Oil Pseudopressure (LSSVM, 0.1538 2.656
phase ANFIS)
Pseudopressure (TSK
0.1441 2.398
Fuzzy, ANFIS)

217 | P a g e
Figure 6.13. Statistical error results between analytical pseudopressure approach and
compositional simulation results for perdition of gas phase production profile in KAL – 5.

Figure 6.14. Statistical error results between analytical pseudopressure approach and
compositional simulation results for perdition of oil phase production profile in KAL – 5.

Using analytical method of three-flow regions pseudopressure approach combined

with developed MBO PVT models in this study has the following advantage over fully
compositional simulation (Eclipse 300).
 The method is less data demanding in comparison to Eclipse reservoir
simulation.
218 | P a g e
  Provide accurate and fast deliverability predictions when many sensitivity
analyses are required particularly at the beginning of the production.
 The pseudopressure approach does not need knowledge of using reservoir
simulators (e.g., Eclipse) and can be implemented in spreadsheet format.
 PVT properties required for computation of pseudopressure integral can be
calculated using simple modified black oil model, where gas/condensate
viscosity and two-phase Z factor can be estimated using developed smart
models in this study.
The results of this chapter is very close to the existing literature work of Fevang,
(1995), Mott, (2002), Jokhio, (2002), Chowdhury et al., (2004), Behmanesh et al.,
(2015) and Hekmatzadeh and Gerami, (2018). The major different between this work
and aforementioned studies are in treating PVT properties. They mostly used modified
black oil method (MBO) to generate and compute the PVT properties, then validate
their result with compositional reservoir simulators, where EOS used for computing
PVT properties. The highest error between two methods were related in uncertainty of
gas/condensate viscosity and two-phase Z factor below the saturation pressure. In this
study we bridge the gap between the results of MBO and EOS models in treating PVT
properties by ensuring accurate estimation of gas/condensate viscosity and two-phase
Z factor using Al techniques. The justification for our approach is with an improved
viscosity and two-phase Z factor in PVT calculation, better accuracy of gas and
condensate liquid can be achieved.
Our approach in this study is more efficient and assure a higher performance for
computation of gas and oil flow rate of gas-condensate reservoirs. Inaccuracy of MBO
model for establishing PVT properties of complex gas-condensate fluid in critical
conditions was compensated by using AI techniques proposed in this study.
This has been confirmed by comparing the results of this study with compositional
simulation of a HTHP gas-condensate well in compact formation.

6.4 Summary

Accurate prediction of production profile in gas-condensate wells are very important

for financial evaluation and well production planning. The aim of this chapter was to
investigate the effect of accurate estimation of gas/condensate viscosity and two-
phase Z factor on reliable estimation of production profile. For this purpose, MBO PVT
219 | P a g e
model, where gas/condensate viscosity and two-phase Z factor were estimated using
AI methods, incorporated with pseudopressure integral to calculate well inflow
performance. Then to relate the results obtained from pseudopressure integral to time,
volumetric material balance utilized and production profiles of a HTHP gas-condensate
well were generated. The results show that pseudopressure integral incorporated with
gas/condensate viscosity and two-phase Z factor using TSK fuzzy and ANFIS
approach respectively has the best performance and predicts the production rate very
close to compositional reservoir simulator. Gas phase flow rate at the surface is
predicted using pseudopressure integral incorporated with TSK fuzzy and ANFIS for
computation of gas viscosity and two-phase Z factor respectively with RMSE of 0.0224
and AARD% of 0.856 from compositional simulation results. Oil (condensate) phase
flow rate at the surface is also estimated using pseudopressure integral combined with
TSK fuzzy and ANFIS model in calculation of oil viscosity and two-phase Z factor
respectively with RMSE of 0.1441 and AARD% of 2.398 in compare to simulator
results.
Accurate prediction of PVT properties including viscosity and two-phase Z factor
ensure accurate prediction of production profile. Using developed AI algorithms in this
study for estimating governing parameters of PVT properties such as gas/condensate
viscosity and Z factor is eliminating the uncertainty in prediction and ensure accurate
gas-condensate reservoir performance modelling.
The generated production profile in this chapter is a simple analytical three-flow
regions pseudopressure approach that can be used for quick and reliable estimation
of gas-condensate wells. The method is an excellent alternative to fully compositional
simulators, where considerable reservoir and fluid data is required to start the
simulation.

220 | P a g e
 CHAPTER 7 CONCLUSIONS AND
RECOMMENDATIONS
CONCLUSIONS AND RECOMMENDATIONS

7.1 Introduction
This chapter consists of two parts. In the first section, general conclusions and brief
discussion of various methods utilized in this study are shown. Furthermore, it has
been highlighted that the aim and objectives of the study have been achieved.
In the second part, the possible improvement of the current study is discussed. More
conclusions that are specific were presented at the end of each chapter.

7.2 Improve accuracies of gas-condensate viscosity

One of the main achievement of this study is in developing several models for
prediction of gas-condensate fluid viscosities through using numerical approaches.
The current methodology for estimation of gas-condensate viscosities below the
saturation pressure is based on empirical and semi-empirical correlations. The
correlations are embedded in black oil model or compositional model (equation of
state) for computation of fluid viscosity.
Primarily in chapter 4, accuracy of the most frequently used methods of Lohrenz –
Bray – Clark (1964), Lee – Gonzalez – Eakin, (1966), Londono – Archer – Blasingame
(2002), Sutton, (2005), Elsharkawy, (2006), Beggs and Robinsons, (1975),
Kartoatmodjo and Schmidt, (1991), De Ghetto et al., (1994), Elsharkawy and Alikhan
(1999), Bergman and Sutton, (2007), for estimation of gas-condensate viscosities
were assessed using experimental data. The results of the statistical error analysis
show the performance of these models for estimating gas-condensate viscosity below
the saturation pressure is as minimum as 15% and as high as 99% absolute average
relative deviation. Both compositional and gas-saturated oil viscosity models were
failed to predict the condensate viscosity of gas-condensate fluid below the saturation
pressure. Non-linear regression was performed to optimize the coefficients of the
existing models to be applicable for estimation of gas-condensate viscosities.
Several ML based approaches have been utilized for better estimation of condensate
viscosity below the saturation pressure. Initially two ML based models including a
221 | P a g e
Least Square Support Vector Machine (LSSVM) and an Artificial Neural Network
(ANN) were developed for estimation of condensate viscosity. The accuracy of the
developed ML based models are compared with the tuned utilized literature viscosity
models. The comparison of the results indicate that the developed ANN model predict
condensate viscosity with coefficient of determination of 84.23 and root mean square
error of 0.1144, which is superior to all other utilized viscosity models.
Furthermore, in this context two well-known fuzzy logic inference systems including
Mamdani and TSK fuzzy logic were manipulated for modelling condensate phase
viscosity. Mamdani based fuzzy approach performed to develop 15 IF-THEN rules to
interrelate condensate viscosity to pressure, temperature and solution gas to oil ratio.
Utilizing TSK fuzzy logic approach, an unique relationship between condensate
viscosity, pressure, temperature and solution gas to oil ratio was established through
proposing a linear correlation (equation 4.61). The superiority of the developed
correlation in 4.61 over other methods was verified by root mean square error of
0.0194, mean average error of 0.0163 and absolute average relative deviation of
7.123%.
The main advantages of the developed viscosity models over existing literature
models is that they are less data demanding, which make them particularly useful for
prediction of condensate viscosity when the compositional data are not available. In
addition, the developed models are computationally efficient and they can be used as
an effective tool for quick and reliable determination of gas-condensate viscosities.

Improvement of two-phase gas-condensate Z factor

In the context of PVT modelling improvement of gas-condensate fluids, another major

achievement of the study is in developing several smart approaches for determination
of two-phase Z factor below the saturation pressure. Accurate prediction of two-phase
Z factor is a critical importance for reliable deliverability modelling in gas-condensate
reservoirs. This important PVT property determines how much gas-condensate
deviate from ideal gas law and can be used for all engineering calculation of gas-
condensate reservoirs. Comprehensive data bank was used to assess the accuracy
and reliability of the existing two-phase Z factor models including Hall – Yarborough,
(1973), Dranchuk – Abu –Kassem, (1975), Beggs and Brill, (1973), Rayes et al.,
(1973) and Azizi et al., (2010). The statistical and graphical error analysis indicate the
222 | P a g e
performance of employed literature models for prediction of gas-condensate two-
phase Z factor below the saturation pressure are not satisfactory for various ranges of
pseudoreduced pressure and pseudoreduced temperature especially for high
pressure high temperature (HPHT) conditions.
Several machine-learning (ML) based approaches was developed for accurate
estimation of gas-condensate two-phase Z factor. To this end, two types of neural
networks known as cascade forward neural network (CFNN) and feed-forward neural
network (FFNN) as well as Adaptive Neuro Fuzzy Inference System (ANFIS) were
implemented. Performance of the developed machine-learning models are better than
utilized existing models for prediction of gas-condensate two-phase Z factor, verified
by statistical indicators of root mean square error (RMSE), mean average error (MAE)
and average absolute relative deviation percentage (AARD%). ANFIS model
outperformed all other Z factor models in various range of pseudoreduced pressure
and pseudoreduced temperature with root mean square error of 0.0025 and absolute
average relative deviation of 0.2191%. An iterative procedure was developed within
ANFIS algorithm to ensure faster convergence of the algorithm for computation of two-
phase Z factor. The developed ML based approaches in this study do not limit to
certain range of Ppr and Tpr like existing literature models. The developed models are
able to predict the gas-condensate reservoirs in highly critical conditions (HTHP) and
with variety of non-hydrocarbon contents (e.g., CO2, H2S and N2). Another advantage
of the developed models is that they are not limited within geographical location of the
reservoirs (e.g., North Sea, Middle East etc.,) like existing literature models.

Analysing the impact of each input variable used for development of the ML based
models on two-phase Z factor using Pearson relevancy factor (r), revealed that
pressure, molecular weight of C7+ and C7+ content, C6 content and temperature have
the highest positive impact on two-phase Z factor. This means increasing
aforementioned parameters would directly increase the magnitude of two-phase Z
factor.
Moreover, non-hydrocarbon components of H2S, N2 and CO2 have negative impact on
two-phase Z factor, which indicate increasing these variables resulted in decreasing
two-phase Z factor.

223 | P a g e
The developed models in this study provide very good improvement in prediction of
gas-condensate two-phase Z factor over previous correlations with broader
applications in terms of pressure, temperature and compositional variations. These
models can be implemented in any reservoir simulation software and provide superior
accuracy and performance for prediction of two-phase gas-condensate Z factor. The
limitation of the developed model is that the compositional data of gas-condensate
reservoir should be available and insert into the models for perdition of Z factor.

Effectiveness of viscosities and two-phase Z factor models for producing

accurate production profile

In chapter 6, the effect of the developed viscosity and two-phase Z factor models in
this study for generating reliable production profile of a gas-condensate reservoir was
investigated. The common industry approach for establishing production profile of gas-
condensate reservoirs is through either implementing fully compositional reservoir
simulations or analytical approach using pseudopressure integral.
The effectiveness of viscosities and two-phase Z factor models on production profile
of gas-condensate wells have been investigated through implementing three-flow
regions pseudopressure integral that combined with volumetric material balance. In
order to calculate the PVT properties within three-flow regions pseudopressure
integral optimized modified black oil (MBO) model was implemented. The novelty of
the new approach in this study is to embed the developed viscosities and two-phase
Z factor using machine – learning approaches for generating PVT table in MBO model.
For computation of condensate viscosity and two-phase Z factor the developed TSK
fuzzy model and developed ANFIS model were utilized respectively. The results show
that the new developed production profile matched with fully compositional model
using Eclipse 300 with RMSE of 0.0224 and AARD% of 0.856 for gas phase and
RMSE of 0.1441 and AARD% of 2.398 for condensate (oil) phase.
The generated production profile is less data demanding and has an advantage over
fully compositional simulations (Eclipse 300), which normally require extensive data to
run. The developed analytical method is easy to implement in spreadsheet format and
does not require knowledge of reservoir simulator. The results of the developed
production profile indicate the dependency of gas-condensate reservoir performance
modelling on accurate estimation of viscosities and two-phase Z factor.
224 | P a g e
The results of effective permeabilities of gas and condensate phase is calculated using
pressure transient test data. The results of permeability indicate severe decline in gas
permeability for the reservoirs where the condensation occurs from the beginning of
the production.
The utilized computational procedure in chapter 6 is well able to predict single well
production (gas and condensate) at the surface.

7.3 Recommendation for future work

In this study, the critical issues associated for accurate PVT properties of gas-
condensate reservoirs, required for well deliverability modelling, have been
investigated. Extensive attempt has been made for accurate modelling of gas-
condensate reservoirs PVT properties including development of several smart
modelling approaches for computation of gas-condensate viscosities and two-phase
Z factor. The developed approaches in this study can be further improved as follow.
 The evolved ANN and LSSVM and Mamdani Fuzzy models that were used for
prediction of condensate viscosity in this study can be further improved with
other optimization algorithms such as Genetic Algorithm (GA) or Coupled
Simulated Annealing (CSA).
 The developed intelligent models in this study for prediction of gas-condensate
viscosities and two-phase Z factor can be embedded inside PVT packages
such as Schlumberger PVTi in order to generate reliable fluid properties of such
reservoirs.
 The intelligent models such as fuzzy approach can also be utilised for
categorization of various type of gas-condensate fields around the world (e.g.,
low, medium, high and very high). This is to ensure better recovery planning
and production optimization of such fields.
 The developed production profile in chapter 6 is applicable for single vertical
well. The method can be applied to horizontal and inclined gas-condensate
wells.

225 | P a g e
 REFERENCES

Aamir Mahmood, M. and Ali Al-Marhoun, M. (1996) Evaluation of empirically derived

PVT properties for Pakistani crude oils, Journal of Petroleum Science and
Engineering.

Afidick, D., Kaczorowski, N. J. and Bette, S. (1994) ‘Production performance of a

retrograde gas reservoir: a case study of the Arun Field’, in SPE - Asia Pacific Oil &
Gas Conference. Melbourne: Society of Petroleum Engineers, pp. 73–80. doi:
10.2523/28749-ms.

Agarwal, R. G. (1979) ‘“Real Gas Pseudo-Time” - A New Function For Pressure

Buildup Analysis Of MHF Gas Wells’, in SPE Annual Technical Conference and
Exhibition. Las Vegas, Nevada: Society of Petroleum Engineers. doi: 10.2118/8279-
MS.

Agatonovic-Kustrin and Beresford, R. (2000) ‘Basic concepts of artificial neural

network (ANN) modeling and its application in pharmaceutical research’, Journal of
Pharmaceutical and Biomedical Analysis, 22, pp. 717–727. Available at:
https://www.sciencedirect.com/science/article/abs/pii/S0731708599002721.

Ahmadi, M. et al. (2015) ‘Phase Equilibrium Modeling of Clathrate Hydrates of Carbon

Dioxide + 1,4-Dioxine Using Intelligent Approaches’, Journal of Dispersion Science
and Technology, 36(2), pp. 236–244. doi: 10.1080/01932691.2014.904792.

Ahmadi, M. and Ebadi, M. (2014a) ‘Evolving smart approach for determination dew
point pressure through condensate gas reservoirs’, Fuel. Elsevier, 117, pp. 1074–
1084. doi: 10.1016/J.FUEL.2013.10.010.

Ahmadi, M. and Ebadi, M. (2014b) ‘Fuzzy Modeling and Experimental Investigation of

Minimum Miscible Pressure in Gas Injection Process’, Fluid Phase Equilibria. Elsevier
B.V., 378, pp. 1–12. doi: 10.1016/j.fluid.2014.06.022.
226 | P a g e
Ahmadi, M. and Ebadi, M. (2014c) ‘Fuzzy Modeling and Experimental Investigation of
Minimum Miscible Pressure in Gas Injection Process’, Fluid Phase Equilibria. Elsevier,
378(2014), pp. 1–12. doi: 10.1016/J.FLUID.2014.06.022.

Ahmadi, M., Ebadi, M. and Hosseini, S. M. (2014) ‘Prediction breakthrough time of

water coning in the fractured reservoirs by implementing low parameter support vector
machine approach’, Fuel. Elsevier, 117(PART A), pp. 579–589. doi:
10.1016/j.fuel.2013.09.071.

Ahmed, T. H. (2010) Reservoir engineering handbook. 4th edn. Oxford: Gulf

Professional Pub.

Aily, M. et al. (2019) ‘Modeling viscosity of moderate and light dead oils in the presence
of complex aromatic structure’, Journal of Petroleum Science and Engineering.
Elsevier, 173, pp. 426–433. doi: 10.1016/J.PETROL.2018.10.024.

Akhimiona, N. and Wiggins, M. L. (2005) ‘An Inflow-Performance Relationship for

Horizontal Gas Wells’, in SPE Eastern Regional Meeting. Society of Petroleum
Engineers. doi: 10.2118/97627-MS.

Al-Attar, H. and Al-Zuhair, S. (2009) ‘Simplified approach for predicting gas well
performance’, Journal of Petroleum Science and Engineering, 65(1–2), pp. 51–61. doi:
10.1016/j.petrol.2008.12.016.

Al-Hussainy, R. and Ramey, H. J. (1966) ‘Application of Real Gas Flow Theory to Well
Testing and Deliverability Forecasting’, Journal of Petroleum Technology. Society of
Petroleum Engineers (SPE), 18(05), pp. 637–642. doi: 10.2118/1243-b-pa.

Al-Hussainy, R., Ramey, H. J. and Crawford, P. B. (1966) ‘The Flow of Real Gases
Through Porous Media’, Journal of Petroleum Technology. Society of Petroleum
Engineers, 18(05), pp. 624–636. doi: 10.2118/1243-A-PA.

227 | P a g e
Al-Khalifa, A. A., Horne, R. N. and Aziz, K. (1987) ‘In-place determination of reservoir
relative permeability using well test analysis’, in SPE Annual Technical Conference
and Exhibition. Dallas: Society of Petroleum Engineers, pp. 459-472 17973. doi:
10.2523/17973-ms.

Al-Marhoun, M. A. (1990) New Correlations for FVFs of Oil and Gas Mixtures. King
Fahd Univerity of Petroleum & Minerals.

Al-Meshari, A. et al. (2007) ‘Measurement of Gas Condensate, Near-Critical and

Volatile Oil Densities and Viscosities at Reservoir Conditions’, in Proceedings of SPE
Annual Technical Conference and Exhibition. California: Society of Petroleum
Engineers. doi: 10.2523/108434-ms.

Al-Meshari, A. A. (2004) New strategic method to tune equation-of-state to match

experimental data for compositional simulation. Texas. Available at:
http://repository.tamu.edu/handle/1969.1/1388.

Al-Nasser, K. S. and Al-Marhoun, M. A. (2012) ‘Development of New Gas Viscosity

Correlations’, in SPE International Production and Operations Conference &
Exhibition. Doha: Society of Petroleum Engineers. doi: 10.2118/153239-ms.

Al-Shammasi, A. A. (2001) ‘A review of bubblepoint pressure and oil formation volume

factor correlations’, SPE Reservoir Evaluation and Engineering. Society of Petroleum
Engineers, 4(2), pp. 146–160. doi: 10.2118/71302-PA.

Al-Shawaf, A., Aramco, S. and Kelkar, M. (2014) ‘A New Method To Predict the
Performance of Gas-Condensate Reservoirs’. doi: 10.2118/161933-PA.

AlQuraishi, A. A. and Shokir, E. M. (2011) ‘Artificial neural networks modeling for

hydrocarbon gas viscosity and density estimation’, Journal of King Saud University -
Engineering Sciences. King Saud University, 23(2), pp. 123–129. doi:
10.1016/j.jksues.2011.03.004.

228 | P a g e
Altug, S., Mo-Yuen Chen and Trussell, H. J. (1999) ‘Fuzzy inference systems
implemented on neural architectures for motor fault detection and diagnosis’, IEEE
Transactions on Industrial Electronics, 46(6), pp. 1069–1079. doi: 10.1109/41.807988.

Arabloo, M., Heidari Sureshjani, M. and Gerami, S. (2014) ‘A new approach for
analysis of production data from constant production rate wells in gas condensate
reservoirs’, Journal of Natural Gas Science and Engineering. Elsevier, 21, pp. 725–
731. doi: 10.1016/j.jngse.2014.09.028.

Arukhe, I. N. and Mason, W. E. (2012) ‘The Use of Two Phase Compressibility Factors
in Predicting Gas Condensate Performance’, in SPE Annual Technical Conference
and Exhibition. San Antonio: Society of Petroleum Engineers. doi: 10.2118/159080-
ms.

Asklany, S. A. et al. (2011) ‘Rainfall events prediction using rule-based fuzzy inference
system’, Atmospheric Research. Elsevier B.V., 101(1–2), pp. 228–236. doi:
10.1016/j.atmosres.2011.02.015.

Audonnet, F. and Pádua, A. A. . (2004) ‘Viscosity and density of mixtures of methane

and n-decane from 298 to 393 K and up to 75 MPa’, Fluid Phase Equilibria. Elsevier,
216(2), pp. 235–244. doi: 10.1016/J.FLUID.2003.10.017.

Ayyalasomayajula, P., Silpngarmlers, N. and Kamath, J. (2005) ‘Well Deliverability

Predictions for a Low-Permeability Gas/Condensate Reservoir’, in SPE Annual
Technical Conference and Exhibition. Dallas, Texas: Society of Petroleum Engineers,
pp. 9–12.

Azizi, N., Behbahani, R. and Isazadeh, M. A. (2010) ‘An efficient correlation for
calculating compressibility factor of natural gases’, Journal of Natural Gas Chemistry,
19, pp. 642–645. doi: 10.1016/S1003-9953(09)60081-5.

Bagirov, A. M. (2008) ‘Modified global k-means algorithm for minimum sum-of-squares

clustering problems’, Pattern Recognition. Pergamon, 41(10), pp. 3192–3199. doi:

229 | P a g e
10.1016/J.PATCOG.2008.04.004.
Bai, L. et al. (2017) ‘Fast density clustering strategies based on the k-means
algorithm’, Pattern Recognition. Pergamon, 71, pp. 375–386. doi:
10.1016/J.PATCOG.2017.06.023.

Baled, H. O. et al. (2018) ‘Viscosity models for pure hydrocarbons at extreme

conditions: A review and comparative study’, Fuel. Elsevier Ltd, 218, pp. 89–111. doi:
10.1016/j.fuel.2018.01.002.

Bandyopadhyay, S. and Maulik, U. (2002) ‘An evolutionary technique based on K-

Means algorithm for optimal clustering in RN’, Information Sciences. Elsevier, 146(1–
4), pp. 221–237. doi: 10.1016/S0020-0255(02)00208-6.

Beal, C. (1946) ‘The Viscosity of Air, Water, Natural Gas, Crude Oil and Its Associated
Gases at Oil Field Temperatures and Pressures’, Transactions of the AIME, 165(01),
pp. 94–115. doi: 10.2118/946094-g.

Beggs, H. D. and Brill, J. R. (1973) ‘Study of two-phase flow in inclined pipes’, JPT,
Journal of Petroleum Technology. Society of Petroleum Engineers, 25(05), pp. 607–
617. doi: 10.2118/4007-PA.

Beggs, H. D. and Robinson, J. R. (1975) ‘Estimating the Viscosity of Crude Oil

Systems’, Journal of Petroleum Technology. Society of Petroleum Engineers, 27(09),
pp. 1140–1141. doi: 10.2118/5434-PA.

Behera, S. S. and Chattopadhyay, S. (2012) ‘A Comparative Study of Back

Propagation and Simulated Annealing Algorithms for Neural Net Classifier
Optimization’, Procedia Engineering. Elsevier, 38, pp. 448–455. doi:
10.1016/J.PROENG.2012.06.055.

Behmanesh, H., Hamdi, H. and Clarkson, C. R. (2015) ‘Production data analysis of

tight gas condensate reservoirs’, Journal of Natural Gas Science and Engineering.
Elsevier, 22, pp. 22–34. doi: 10.1016/j.jngse.2014.11.005.

230 | P a g e
Behmanesh, H., Hamdi, H. and Clarkson, C. R. (2017) ‘Production data analysis of
gas condensate reservoirs using two-phase viscosity and two-phase compressibility’,
Journal of Natural Gas Science and Engineering. Elsevier B.V, 47, pp. 47–58. doi:
10.1016/j.jngse.2017.07.035.

Bell, J. (2014) Machine Learning : Hands-On for Developers and Technical

Professionals. 1st edn. Canada: John Wiley & Sons, Incorporated.

Bergman, D. F. and Sutton, R. P. (2007a) ‘A Consistent and Accurate Dead-Oil-

Viscosity Method’, in SPE Annual Technical Conference and Exhibition. Anaheim:
Society of Petroleum Engineers, p. 30. doi: 10.2118/110194-MS.

Bergman, D. F. and Sutton, R. P. (2007b) ‘An Update to Viscosity Correlations for

Gas-Saturated Crude Oils’, in SPE Annual Technical Conference and Exhibition.
Anaheim: Society of Petroleum Engineers. doi: 10.2118/110195-MS.

Bezdek, J. and Pal, S. K. (1992) Fuzzy models for pattern recognition, Fuzzy models
for pattern recognition. New York: IEEE Press.

Biezma, M. V., Agudo, D. and Barron, G. (2018) ‘A Fuzzy Logic method: Predicting
pipeline external corrosion rate’, International Journal of Pressure Vessels and Piping,
163, pp. 55–62. doi: 10.1016/j.ijpvp.2018.05.001.

Black, M. (1937) ‘Vagueness. An Exercise in Logical Analysis’, Philosophy of Science,

4(4), pp. 427–455.

Bonyadi, M. et al. (2014) ‘Theoretical and experimental determination of initial

reservoir fluid in a lean gas condensate reservoir’. doi: 10.1016/j.petrol.2014.01.003.

Bonyadi, M., Rahimpour, M. R. and Esmaeilzadeh, F. (2012) ‘A new fast technique for
calculation of gas condensate well productivity by using pseudopressure method’,
Journal of Natural Gas Science and Engineering. Elsevier, 4, pp. 35–43. doi:

231 | P a g e
10.1016/J.JNGSE.2011.07.012.
Borges, P. R. (1991) ‘Correction improves z-factor values for high gas density’, The
Oil and Gas Journal, 89(9), pp. 54–56.

Borthne, G. (1986) Development of a materia balance and inflow performane model

for oil and gas-condensate reservoirs. The Norwegian Institute of Technolodgy.

Bourbiaux, B. J. (1994) ‘Parametric Study of Gas-Condensate Reservoir Behavior

During Depletion: A Guide for Development Planning’, in European Petroleum
Conference. London: Society of Petroleum Engineers. doi: 10.2118/28848-MS.

BP (2019) BP invests in new artificial intelligence technology | News and insights |

Home. Available at: https://www.bp.com/en/global/corporate/news-and-
insights/press-releases/bp-invests-in-new-artificial-intelligence-technology.html
(Accessed: 8 April 2020).

Brooks, R. H. and Corey, A. T. (1964) Hydraulic Properties of Porous Media.

Calinski, T. and Harabasz, J. (1974) ‘A dendrite method for cluster analysis’,
Communications in Statistics - Theory and Methods, 3(1), pp. 1–27. doi:
10.1080/03610927408827101.

Carnahan, N. F. and Starling, K. E. (1969) ‘Equation of state for nonattracting rigid

spheres’, The Journal of Chemical Physics. American Institute of PhysicsAIP, 51(2),
pp. 635–636. doi: 10.1063/1.1672048.

Carr, N. L., Kobayashi, R. and Burrows, D. B. (1954) ‘Viscosity of Hydrocarbon Gases

Under Pressure’, Journal of Petroleum Technology. Society of Petroleum Engineers
(SPE), 6(10), pp. 47–55. doi: 10.2118/297-g.

Chamkalani, A. et al. (2013) ‘Utilization of support vector machine to calculate gas

compressibility factor’, Fluid Phase Equilibria, 358, pp. 189–202. doi:
10.1016/j.fluid.2013.08.018.

232 | P a g e
Chen, H. C. et al. (1993) ‘Novel approaches to the determination of Archie parameters
II: Fuzzy regression analysis’, Society of Petroleum Engineers of AIME, (Paper) SPE.
Society of Petroleum Engineers, 3(01), pp. 71–94. doi: 10.2118/26288-pa.

Chen, Z. et al. (2015) ‘Self-Adaptive Prediction of Cloud Resource Demands Using

Ensemble Model and Subtractive-Fuzzy Clustering Based Fuzzy Neural Network’,
Computational and Mathematical Methods in Medicine. Hindawi Limited, 2015(1), pp.
2–4. doi: 10.1155/2015.

Chew, J. . and Connally, C. (1959) ‘A viscosity correlation for gas-saturated crude oils’,
Pet. Trans. AIME. Society of Petroleum Engineers, 216, pp. 23–25.

Chierici, G. L. (1984) ‘Novel relation for drainage and imbibition relative permeabilities’,
Society of Petroleum Engineers journal. Society of Petroleum Engineers, 24(3), pp.
275–276. doi: 10.2118/10165-PA.

Chiu, S. L. (1994) ‘Fuzzy model identification based on cluster estimation’, Journal of

Intelligent and Fuzzy Systems. IOS Press, 2(3), pp. 267–278. doi: 10.3233/IFS-1994-
2306.

Chowdhury, N. et al. (2004) ‘A Semi-Analytical Method to Predict Well Deliverability in

Gas-Condensate Reservoirs’, in SPE Annual Technical Conference and Exhibition.
Houston: Society of Petroleum Engineers. doi: 10.2118/90320-MS.

Christensen, P. L. and Fredenslund, A. A. (1980) ‘A corresponding states model for

the thermal conductivity of gases and liquids’, Chemical Engineering Science.
Pergamon, 35(4), pp. 871–875. doi: 10.1016/0009-2509(80)85073-1.

Cios, K. J. and Shields, M. E. (1997) ‘The handbook of brain theory and neural
networks: By Micheal A. Arbib (Ed.), MIT Press, Cambridge, MA, 1995, ISBN 0-262-
01148-4, 1118 pp’, Neurocomputing. Elsevier, 16(3), pp. 259–261. doi:
10.1016/S0925-2312(97)00036-2.

233 | P a g e
Coats, K. H. (1985) ‘Simulation of gas-condensate reservoirs performance’, JPT,
Journal of Petroleum Technology, 37(11), pp. 1870–1886. doi: 10.2118/10512-pa.

Coats, K. H. and Smart, G. T. (1986) ‘Application of a regression-based EOS PVT

program to laboratory data’, SPE Reservoir Engineering (Society of Petroleum
Engineers), 1(3), pp. 277–299. doi: 10.2118/11197-PA.

Coats, K. H., Thomas, L. K. and Pierson, R. G. (1995) ‘Compositional and Black Oil
Reservoir Simulation’, in SPE Reservoir Simulation Symposium. San Antonio: Society
of Petroleum Engineers. doi: 10.2118/50990-pa.

Corey, A. T. et al. (1956) ‘Three-Phase Relative Permeability’, Journal of Petroleum

Technology. Society of Petroleum Engineers (SPE), 8(11), pp. 63–65. doi:
10.2118/737-g.

Cortes, C. and Vapnik, V. (1995) ‘Support-vector networks’, Machine Learning. Kluwer

Academic Publishers, 20(3), pp. 273–297. doi: 10.1007/BF00994018.

Craft, B. and Hawkin, M. (2015) Applied Petroleum Reservoir Engineering. Third.

Edited by T. Ronald and R. Brandon. New Jersy: Prentice Hall.

Cragoe, C. S. (1929) Thermodynamic Properties of Petroleum Products. Washington,

DC.

Curilem, M. et al. (no date) ‘Neural networks and support vector machine models
applied to energy consumption optimization in semiautogeneous grinding’,
folk.ntnu.no. Available at: http://folk.ntnu.no/skoge/prost/proceedings/pres2011-and-
icheap10/PRES11/85Curilem.pdf (Accessed: 12 March 2019).

Dake, L. (1983) Fundamentals of reservoir engineering. 19th edn. Oxford: Elsevier.

Dake, L. (2001) The practice of reservoir engineering. 3rd edn. Amsterdam: Elsevier.

234 | P a g e
Daltaban, T. (1985) Numerical modelling of recovery processes from gas condensate
reservoirs. University og London.

Danesh, A. (1998) PVT and phase behaviour of petroleum reservoir fluids. 1st edn.
Elsevier.

Davani, E. et al. (2013) ‘HPHT viscosities measurements of mixtures of

methane/nitrogen and methane/carbon dioxide’, Journal of Natural Gas Science and
Engineering, 12, pp. 43–55. doi: 10.1016/j.jngse.2013.01.005.

Dehane, A., Tiab, D. and Osisanya, S. O. (2000) ‘Comparison of the Performance of

Vertical and Horizontal Wells in Gas-Condensate Reservoirs’, in SPE Annual
Technical Conference and Exhibition. Society of Petroleum Engineers. doi:
10.2118/63164-MS.

Dempsey, J. R. (1965) ‘Computer routine treats gas viscosity as a variable’, Oil and
Gas Journal, 63(8), pp. 141–143.

Dranchuk, P. M. and Abou-Kassem, J. H. (1974) ‘Calculation of Z factors for natural

gases using equations of state.’, Journal of Canadian Petroleum Technology.
Petroleum Society of Canada, 14(3), pp. 34–36. doi: 10.2118/75-03-03.

Dranchuk, P. M., Purvis, R. A. and Robinson, D. B. (1973) ‘Computer Calculation Of

Natural Gas Compressibility Factors Using The Standing And Katz Correlation’, in.
Edmonton: Society of Petroleum Engineers (SPE). doi: 10.2118/73-112.

Drohm, J. K., Goldthorpe, W. H. and Trengove, R. (1988) ‘Enhancing the Evaluation

of PVT Data’, in Offshore South East Asia Show. Singapore: Society of Petroleum
Engineers. doi: 10.2118/17685-MS.

Earlougher, R. C. (1977) Advances in well test analysis. SPE. Available at:

https://store.spe.org/Advances-in-Well-Test-Analysis-P8.aspx (Accessed: 14 May
2019).

235 | P a g e
ECLIPSE (2014) Eclipse Reservoir Simulation Reference Manual.

Economides, M. J. et al. (1989) ‘The Stimulation of a Tight, Very-High- Temperature

Gas-Condensate Well’, SPE Formation Evaluation. Society of Petroleum Engineers,
4(01), pp. 63–72. doi: 10.2118/15239-PA.

Economldes, M. J. et al. (1989) ‘The Stimulation of a Tight , Very-High- Temperature

Gas-Condensate Well’, SPE Formation Evaluation, Economldes(March), pp. 63–72.

Egbogah, E. and Jack, N. (1990) ‘An improved temperature-viscosity correlation for

crude oil systems’, Journal of Petroleum Science and Engineering, 5, pp. 197–200.
Available at: https://vdocuments.mx/an-improved-temperature-viscosity-correlation-
for-crude-oil-systems.html.

EIA (2020) Open Data, eai.gov. Available at: www.eia.gov (Accessed: 28 June 2020).
Eilerts, C. K. (1964) ‘Integration of Partial Differential Equation for Transient Linear
Flow of Gas-Condensate Fluids in Porous Structures’, Society of Petroleum Engineers
Journal. Society of Petroleum Engineers (SPE), 4(04), pp. 291–306. doi: 10.2118/111-
pa.

Eilerts, C. K., Sumner, E. F. and Potts, N. L. (1965) ‘Integration of Partial Differential

Equation for Transient Radial Flow of Gas-Condensate Fluids in Porous Structures’,
Society of Petroleum Engineers Journal. Society of Petroleum Engineers (SPE), 5(02),
pp. 141–152. doi: 10.2118/716-pa.

Eilerts, K. C. (1947) ‘Gas-condensate reservoir engineering. 1. The reservoir fluid, its

composition and phase behavior’, Oil Gas J., 45(39).

Elsharkawy, A. M. (2002) ‘Predicting the Properties of Sour Gases and Condensates’,

in 2002 SPE International Petroleum Conference and Exhibition. Society of Petroleum
Engineers. doi: 10.2118/74369-MS.

236 | P a g e
Elsharkawy, A. M. (2006) ‘Efficient methods for calculations of compressibility, density,
and viscosity of natural gases’, in Journal of Canadian Petroleum Technology.
Petroleum Society of Canada, pp. 55–61. doi: 10.2118/06-06-04.

Elsharkawy, A. M. and Alikhan, A. A. (1999) ‘Models for predicting the viscosity of

Middle East crude oils’, Fuel. Elsevier, 78(8), pp. 891–903. doi: 10.1016/S0016-
2361(99)00019-8.

Elsharkawy, A. M. and Elkamel, A. (2001) ‘The accuracy of predicting compressibility

factor for sour natural gases’, Petroleum Science and Technology. Taylor & Francis
Group, 19(5–6), pp. 711–731. doi: 10.1081/LFT-100105285.

Elsharkawy, A. M. and Foda, S. G. (1998) ‘EOS simulation and GRNN modeling of the
constant volume depletion behavior of gas condensate reservoirs’, in SPE Asia Pacific
Conference on Integrated Modelling for Asset Management. Kuala Lumpur: Society of
Petroleum Engineers. doi: 10.1021/ef970135z.

Elsharkawy, A. M., Hashem, Y. S. K. S. and Alikhan, A. A. (2000) ‘Compressibility

Factor for Gas Condensates’, in SPE Permian Basin Oil and Gas Recovery
Conference. Midland: Society of Petroleum Engineers. doi: 10.2118/59702-MS.

Eslamimanesh, A. et al. (2011) ‘Phase Equilibrium Modeling of Structure H Clathrate

Hydrates of Methane + Water &quot;Insoluble&quot; Hydrocarbon Promoter Using
Group Contribution-Support Vector Machine Technique’, Ind. Eng. Chem. Res, 50, pp.
12807–12814. doi: 10.1021/ie2011164.

Eslamimanesh, A. et al. (2012) ‘Phase equilibrium modeling of clathrate hydrates of

methane, carbon dioxide, nitrogen, and hydrogen+water soluble organic promoters
using Support Vector Machine algorithm’, Fluid Phase Equilibria. Elsevier, 316, pp.
34–45. doi: 10.1016/j.fluid.2011.11.029.

Evinger, H. H. and Muskat, M. (1942) ‘Calculation of Theoretical Productivity Factor’,

Transactions of the AIME. Society of Petroleum Engineers, 146(01), pp. 126–139. doi:

237 | P a g e
10.2118/942126-g.
Fan, L. et al. (2005) ‘Understanding Gas-Condensate Reservoirs’, Oilfield Review,
17(4), pp. 14–27. Available at:
https://pdfs.semanticscholar.org/947f/0f6128844c3334638325ae38efe17bb802c7.pd
f (Accessed: 10 April 2020).

Fattah, K. A. et al. (2014) ‘New Inflow Performance Relationship for solution-gas drive
oil reservoirs’, Journal of Petroleum Science and Engineering. Elsevier, 122, pp. 280–
289. doi: 10.1016/j.petrol.2014.07.021.

Fayazi, A. et al. (2014) ‘State-of-the-art least square support vector machine

application for accurate determination of natural gas viscosity’, Industrial and
Engineering Chemistry Research, 53(2), pp. 945–958. doi: 10.1021/ie402829p.

Fazeli, H. et al. (2013) ‘Experimental Study and Modeling of Ultrafiltration of Refinery

Effluents Using a Hybrid Intelligent Approach’, Energy & Fuels. American Chemical
Society, 27(6), pp. 3523–3537. doi: 10.1021/ef400179b.

Fetkovich, M. D. et al. (1986) ‘Oil and Gas Relative Permeabilities Determined From
Rate-Time Performance Data’, in SPE Annual Technical Conference and Exhibition.
New Oreleans: Society of Petroleum Engineers, pp. 1–44.

Fetkovich, M. J. (1973) ‘The Isochronal Testing of Oil Wells’, in Fall Meeting of the
Society of Petroleum Engineers of AIME. Las Vegas, Nevada: Society of Petroleum
Engineers, p. SPE 4529. doi: 10.2118/4529-MS.

Fevang, Ø. (1995) Gas Condensate Flow Behavior and Sampling, October. PhD
Thesis. Petroleum Engineering and Applied Geophysics.

Fevang, Ø., Singh, K. and Whitson, C. H. (2000) ‘Guidelines for choosing

compositional and black-oil models for volatile oil and gas-condensate reservoirs’, in
SPE Reservoir Engineering (Society of Petroleum Engineers). Dallas: Society of
Petroleum Engineers. doi: 10.2118/63087-MS.

238 | P a g e
Fevang, Ø. and Whitson, C. H. (1996) ‘Modeling Gas-Condensate Well Deliverability’,
SPE Reservoir Engineering. Society of Petroleum Engineers, 11(04), pp. 221–230.
doi: 10.2118/30714-PA.

Flach, P. A. (2012) Machine learning: the art and science of algorithms that make
sense of data. Cambridge: Cambridge University Press.

Fletcher, R. (1987) Practical Methods of Optimization. 2nd edn. Chichester: Wiley.

Forchheimer, P. (1901) ‘Wasserbewegung durch Boden [Movement of Water through

Soil]’, Zeitschrift des Vereins deutscher Ingenieur, 45, pp. 1736–1749.

Forsyth, R. (1989) Machine Learning. 1st edn. London: Chapman and Hall.

Fuller, G. G. (1976) ‘A Modified Redlich-Kwong-Soave Equation of State Capable of

Representing the Liquid State’, Industrial & Engineering Chemistry Fundamentals,
15(04), pp. 254–257.

Fussell, D. D. (1973) ‘Single-well performance predictions for gas condensate

reservoirs.’, JPT, Journal of Petroleum Technology. Society of Petroleum Engineers,
25(07), pp. 860–870. doi: 10.2118/4072-PA.

Gazprom (2019) Gas and oil reserves. Available at:

https://www.gazprom.com/about/production/reserves/ (Accessed: 9 March 2020).

Ghaffari, A. et al. (2006a) ‘Performance comparison of neural network training

algorithms in modeling of bimodal drug delivery’, International Journal of
Pharmaceutics. Elsevier, 327(1–2), pp. 126–138. doi:
10.1016/J.IJPHARM.2006.07.056.

Ghaffari, A. et al. (2006b) ‘Performance comparison of neural network training

algorithms in modeling of bimodal drug delivery’, International Journal of

239 | P a g e
Pharmaceutics. Elsevier, 327(1–2), pp. 126–138. doi:
10.1016/J.IJPHARM.2006.07.056.

Gharagheizi, F. et al. (2014) ‘Development of a LSSVM-GC model for estimating the

electrical conductivity of ionic liquids’, Chemical Engineering Research and Design.
Elsevier, 92(1), pp. 66–79. doi: 10.1016/J.CHERD.2013.06.015.

De Ghetto, G. et al. (1994) ‘Reliability Analysis on PVT Correlations’, in European

Petroleum Conference. London: SPE.

Ghiasi, M. M. et al. (2014) ‘Robust modeling approach for estimation of compressibility

factor in retrograde gas condensate systems’, Industrial and Engineering Chemistry
Research. American Chemical Society, 53(32), pp. 12872–12887. doi:
10.1021/ie404269b.

Gilbert, W. E. (1954) ‘Flowing and gas-lift well performance’, in API Drilling and
Production Practice. Los Angeles: American Petroleum Institute. doi: 10.2118.

Giri Nandagopal, M. S. and Selvaraju, N. (2016) ‘Prediction of Liquid−Liquid Flow

Patterns in a Y-Junction Circular Microchannel Using Advanced Neural Network
Techniques’, Industrial & Engineering Chemistry Research, 55(43), pp. 11346–11362.
doi: 10.1021/acs.iecr.6b02438.

Glaso, O. (1980) ‘Generalized Pressure-Volume-Temperature Correlations’, Journal

of Petroleum Technology, 32(05), p. 785.

Göktaş, B., Macmillan, N. A. and Thrasher, T. S. (2010) ‘A systematic approach to

modeling condensate liquid dropout in Britannia reservoir’, in SPE Latin American and
Caribbean Petroleum Engineering Conference Proceedings. Lima: Society of
Petroleum Engineers, pp. 965–973. doi: 10.2118/139056-ms.

Gold, D. K., McCain, W. D. and Jennings, J. W. (1989) ‘Improved method for the
determination of the reservoir-gas specific gravity for retrograde gases’, Journal of

240 | P a g e
Petroleum Technology. Society of Petroleum Engineers, 41(7), pp. 747–752. doi:
10.2118/17310-pa.

Gomes, H. P. and Corrêa, A. C. F. (1992) ‘Fully implicit compositional modelling of

gas condensate and volatile oil reservoirs’, in SPE Latin American and Caribbean
Petroleum Engineering Conference Proceedings. Caracas: Society of Petroleum
Engineers, pp. 1–6. doi: 10.2523/23700-ms.

Gondouin, M., Iffly, R. and Husson, J. (1967) ‘An Attempt to Predict the Time
Dependence of Well Deliverability in Gas Condensate Fields’, Society of Petroleum
Engineers Journal. Society of Petroleum Engineers, 7(02), pp. 113–124. doi:
10.2118/1478-pa.

Gopal, V. (1977) ‘Gas Z-Factor Equation Developed for computer.’, Oil and Gas
Journal, 75, pp. 58–60.

Gozalpour, F. et al. (2005) ‘Viscosity, density, interfacial tension and compositional

data for near critical mixtures of methane + butane and methane + decane systems at
310.95 K’, Fluid Phase Equilibria. Elsevier, 233(2), pp. 144–150. doi:
10.1016/J.FLUID.2005.03.032.

Guehria, F. M. (2000) ‘Inflow Performance Relationships for Gas Condensates’, in

SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers.
doi: 10.2118/63158-MS.

Guo, B., Sun, K. and Ghalambor, A. (2008) Well Productivity Handbook: Vertical,
Fractured, Horizontal, Multilateral, and Intelligent Wells. 1st edn, Well Productivity
Handbook: Vertical, Fractured, Horizontal, Multilateral, and Intelligent Wells. 1st edn.
Suite: Gulf Publishing Company. doi: 10.1016/C2013-0-15529-8.

Guo, T. M., Du, L. and School, B. G. (1989) A New Three-Parameter Cubic Equation
Of State For Fluids III. Application To Gas Condensates.

241 | P a g e
Guo, X. et al. (1997) ‘Viscosity model based on equations of state for hydrocarbon
liquids and gases’, Fluid Phase Equilibria. Elsevier, 139(1–2), pp. 405–421. doi:
10.1016/S0378-3812(97)00156-8.

Hagan, M. T. et al. (2014) Neural network design. 2nd edn. Frisco. Available at:
https://capitadiscovery.co.uk/teesside-ac/items/630382 (Accessed: 11 March 2019).

Hagan, M. T. and Menhaj, M. B. (1994) ‘Training feedforward networks with the

Marquardt algorithm’, IEEE Transactions on Neural Networks, 5(6), pp. 989–993. doi:
10.1109/72.329697.

Hajirezaie, S. et al. (2015) ‘A smooth model for the estimation of gas/vapor viscosity
of hydrocarbon fluids’, Journal of Natural Gas Science and Engineering. Elsevier, 26,
pp. 1452–1459. doi: 10.1016/j.jngse.2015.07.045.

Hall, K. R. and Yarborough, L. (1973) ‘A new equation of state for Z-factor

calculations’, Oil and Gas Journal, 71(25), pp. 82–92.

Hameed, I. A. (2011) ‘Using Gaussian membership functions for improving the

reliability and robustness of students’ evaluation systems’, Expert Systems with
Applications, 38(6), pp. 7135–7142. doi: 10.1016/j.eswa.2010.12.048.

Hassan, A. et al. (2019) ‘Gas condensate treatment: A critical review of materials,

methods, field applications, and new solutions’, Journal of Petroleum Science and
Engineering. Elsevier, 177, pp. 602–613. doi: 10.1016/J.PETROL.2019.02.089.

Hatampour, A. and Ghiasi-Freez, J. (2013) ‘A fuzzy logic model for predicting dipole
shear sonic imager parameters from conventional well logs’, Petroleum Science and
Technology, 31(24), pp. 2557–2568. doi: 10.1080/10916466.2011.603005.

Hatzignatiou, D. G. and Reynolds, A. C. (1996) ‘Determination of effective or relative

permeability curves from well tests’, SPE Journal. Publ by Soc of Petroleum Engineers
of AIME, 1(01), pp. 89–100. doi: 10.2118/20537-pa.

242 | P a g e
Haykin, S. S. (1994) Neural networks : a comprehensive foundation. 1st edn. New
York: Macmillan.

Heidaryan, E., Moghadasi, J. and Rahimi, M. (2010) ‘New correlations to predict

natural gas viscosity and compressibility factor’, Journal of Petroleum Science and
Engineering, 73(1–2), pp. 67–72. doi: 10.1016/j.petrol.2010.05.008.

Hekmatzadeh, M. and Gerami, S. (2018) ‘A new fast approach for well production
prediction in gas-condensate reservoirs’, Journal of Petroleum Science and
Engineering. Elsevier Ltd, 160(September 2017), pp. 47–59. doi:
10.1016/j.petrol.2017.10.032.

Hemmati-Sarapardeh, A. et al. (2014) ‘Reservoir oil viscosity determination using a

rigorous approach’, Fuel. Elsevier, 116, pp. 39–48. doi: 10.1016/J.FUEL.2013.07.072.

Hemmati-Sarapardeh, A. et al. (2020) ‘Modeling natural gas compressibility factor

using a hybrid group method of data handling’, Engineering Applications of
Computational Fluid Mechanics, 14(1), pp. 27–37. doi:
10.1080/19942060.2019.1679668.

Hernandez;, J. C. et al. (2002) ‘Sensitivity of Reservoir Simulations to Uncertainties in

Viscosity’, in SPE/DOE Improved Oil Recovery Symposium. Oklahoma: Society of
Petroleum Engineers, pp. 2–10. doi: 10.2118/75227-ms.

Hippert, H. S., Pedreira, C. E. and Souza, R. C. (2001) ‘Neural networks for short-term
load forecasting : a review and evaluation’, IEEE Transactions on Power Systems,
16(1), p. 4333.

Horne, R. N. (1995) Modern Well Test Analysis A Computer-Aided Approach. Palo

Alto: Petroway.

Horner, D. R. (1951) ‘Pressure Build-up in Wells’, in 3rd World Petroleum Congress.

243 | P a g e
The Hague: World Petroleum Congress. Available at:
https://www.onepetro.org/conference-paper/WPC-4135 (Accessed: 14 May 2019).

Izgec, B. and Barrufet, M. A. (2005) ‘Performance analysis of compositional and

modified black-oil models for a rich gas/condensate reservoir’, in Offshore Europe
Conference - Proceedings. California: Society of Petroleum Engineers, pp. 1–14. doi:
10.2118/93374-ms.

Jain, A. K., Murty, M. N. and Flynn, P. J. (1999) ‘Data Clustering: A Review’, ACM
Computing Surveys, 31(3), pp. 264–323.

Jalali, farhang, Abdy, Y. and Akbari, M. (2007) ‘Dewpoint Pressure Estimation of Gas
Condensate Reservoirs, Using Artificial Neural Network (ANN)’, in Proceedings of
EUROPEC/EAGE Conference and Exhibition. London: Society of Petroleum
Engineers. doi: 10.2523/107032-MS.

Jang, J. S. R. (1993) ‘ANFIS: Adaptive-Network-Based Fuzzy Inference System’, IEEE

Transactions on Systems, Man and Cybernetics, 23(3), pp. 665–685. doi:
10.1109/21.256541.

Jang, J. S. R., Sun, C. T. and Mizutani, E. (1997) ‘Neuro-Fuzzy and Soft Computing-
A Computational Approach to Learning and Machine Intelligence’, IEEE Transactions
on Automatic Control. Institute of Electrical and Electronics Engineers (IEEE), 42(10),
pp. 1482–1484. doi: 10.1109/tac.1997.633847.

Jokhio, S. A. (2002) Production Performance of Horizontal Wells In Gas-Condensate

Reservoirs. University of Oklahoma.

Jokhio, S. A. and Tiab, D. (2002) ‘Establishing Inflow Performance Relationship (IPR)

for Gas Condensate Wells’, in SPE Gas Technology Symposium. Alberta: Society of
Petroleum Engineers, pp. 1–20. doi: 10.2118/75503-MS.

Jokhio, S. A., Tiab, D. and Escobar, F. (2002) ‘Forecasting Liquid Condensate and

244 | P a g e
Water Production In Two-Phase And Three-Phase Gas Condensate Systems’, in. San
Antonio: Society of Petroleum Engineers, pp. 1–13. doi: 10.2118/77549-ms.

Jones, J. R. and Raghavan, R. (1988) ‘Interpretation of flowing well response in gas-

condensate wells’, SPE Formation Evaluation. Society of Petroleum Engineers, 3(3),
pp. 578–594. doi: 10.2118/14204-PA.

Jones, J. R., Vo, D. T. and Raghavan, R. (1989) ‘Interpretation of pressure-buildup

responses in gas-condensate wells’, SPE Formation Evaluation. Society of Petroleum
Engineers, 4(1), pp. 93-104 15535. doi: 10.2118/15535-pa.

Jossi, J. A., Stiel, L. I. and Thodos, G. (1962) ‘The viscosity of pure substances in the
dense gaseous and liquid phases’, AIChE Journal. John Wiley & Sons, Ltd, 8(1), pp.
59–63. doi: 10.1002/aic.690080116.

Kamari, A. et al. (2013) ‘Prediction of sour gas compressibility factor using an

intelligent approach’, Fuel Processing Technology. Elsevier, 116, pp. 209–216. doi:
10.1016/J.FUPROC.2013.06.004.

Kamari, A., Mohammadi, A. H. and Ramjugernath, D. (2019) ‘Petroleum Science and

Technology Characterization of C7+ fraction properties of crude oils and gas-
condensates using data driven models. doi: 10.1080/10916466.2019.1570254.

Kariznovi, M., Nourozieh, H. and Abedi, J. (2012) ‘Experimental and thermodynamic

modeling study on (vapor+liquid) equilibria and physical properties of ternary systems
(methane+n-decane+n-tetradecane)’, Fluid Phase Equilibria, 334, pp. 30–36. doi:
10.1016/j.fluid.2012.06.028.

Kartoatmodjo, T. R. S. and Schmidt, Z. (1991) ‘New Correlations For Crude Oil

Physical Properties’. USA: Society of Petroleum Engineers, p. 30.

Kartoatmodjo, T. and Schmidt, Z. (1994) ‘Large data bank improves crude physical
property correlations’, The Oil & gas journal, 92(27), pp. 51–55.

245 | P a g e
Kashefi, K. et al. (2013) ‘Viscosity of binary and multicomponent hydrocarbon fluids at
high pressure and high temperature conditions: Measurements and predictions’,
Journal of Petroleum Science and Engineering. Elsevier, 112, pp. 153–160. doi:
10.1016/J.PETROL.2013.10.021.

Katz, D. L. (1942) ‘Prediction Of The Shrinkage Of Crude Oils’, Drilling and Production
Practic. American Petroleum Institute, 137.

Katz, D. L. (1959) Handbook of Natural Gas Engineering. McGraw-Hill.

Kay, W. B. (1936) ‘Density of Hydrocarbon’, Industrial and Engineering Chemistry,

28(9), pp. 1014–1019. doi: 10.1021/ie50321a008.

Kenneth, L. (1944) ‘A Method for the Solution of Certain Non-Linear Problem in Least
Squares’. Chcago: American Mathematical Society, pp. 536–538. doi:
10.1090/qam/10666.

Kenyon, D. E. and Behie, G. A. (1987) ‘Third SPE Comparative Solution Condensate

Reservoirs Project: Gas Cycling of Retrograde’, Journal of Petroleum Technology.
Society of Petroleum Engineers, 39(8), pp. 981–997. doi: 10.2118/12278-PA.

Kesler, M. . and Lee, B. . (1976) ‘Improved prediction of enthalpy fractions’,

Hydrocarbon Processing, 55, pp. 153–158.

Khamis, M. A. and Fattah, K. A. (2019) ‘Estimating oil–gas ratio for volatile oil and gas
condensate reservoirs: artificial neural network, support vector machines and
functional network approach’, Journal of Petroleum Exploration and Production
Technology. Springer Verlag, 9(1), pp. 573–582. doi: 10.1007/s13202-018-0501-0.

Khanal, A., Khoshghadam, M. and Lee, W. J. (2016) ‘Accurate forecasting of liquid

rich gas condensate reservoirs with multiphase flow’, in SPE/AAPG/SEG
Unconventional Resources Technology Conference 2016. Unconventional Resources

246 | P a g e
Technology Conference (URTEC). doi: 10.15530/urtec-2016-2426222.

Khazali, N., Sharifi, M. and Ahmadi, M. A. (2019) ‘Application of fuzzy decision tree in
EOR screening assessment’, Journal of Petroleum Science and Engineering, 177, pp.
167–180. doi: 10.1016/J.PETROL.2019.02.001.

Khorami, A. et al. (2017) ‘Density, viscosity, surface tension, and excess properties of
DSO and gas condensate mixtures’, Applied Petrochemical Research. Springer
Nature, 7(2–4), pp. 119–129. doi: 10.1007/s13203-017-0183-4.

Khosrojerdi, S. et al. (2016) ‘Thermal conductivity modeling of graphene

nanoplatelets/deionized water nanofluid by MLP neural network and theoretical
modeling using experimental results’, International Communications in Heat and Mass
Transfer. Science Direct, 74, pp. 11–17. doi:
10.1016/J.ICHEATMASSTRANSFER.2016.03.010.

Klawonn, F., Kruse, R. and Winkler, R. (2015) ‘Fuzzy clustering: More than just
fuzzification’, Fuzzy Sets and Systems. North-Holland, 281, pp. 272–279. doi:
10.1016/J.FSS.2015.06.024.

Kniazeff, V. J. and Naville, S. A. (1965) ‘Two-Phase Flow of Volatile Hydrocarbons’,

Society of Petroleum Engineers Journal. Society of Petroleum Engineers, 5(01), pp.
37–44. doi: 10.2118/962-pa.

Kuenen, J. . (1892) On Retrograde Condensation and the Critical Phenomena of

Mixtures of Two Substances, Communications in Physics. Netherlands.

Labedi, R. (1992) ‘Improved correlations for predicting the viscosity of light crudes’,
Journal of Petroleum Science and Engineering. Elsevier, 8(3), pp. 221–234. doi:
10.1016/0920-4105(92)90035-Y.

Lawal, A. S. (1999) Application of the Lawal-Lake-Silberberg equation-of-state to

thermo- dynamic and transport properties of fluid and fluid mixtures, Technical Report

247 | P a g e
TR-4-99. Lubbock.

Lee, A. L., Gonzalez, M. H. and Eakin, B. E. (1966) ‘The Viscosity of Natural Gases’,
Journal of Petroleum Technology. Society of Petroleum Engineers, 18(08), pp. 997–
1000. doi: 10.2118/1340-PA.

Lee, B. I. and Kesler, M. G. (1975) ‘A Generalized Thermodynamic Correlation Based

on Three-Parameter Corresponding States’, AIChE Journal. John Wiley & Sons, Ltd,
21(3), pp. 981–983. doi: 10.1002/aic.690300615.

Lee, J. and Wattenbarger, R. (1995) Gas Reservoir Engineering. Richardson: Society

of Petroleum Engineers (SPE).

Lee, K. H. (2005) First Course on Fuzzy Theory and Applications. Taejon: Springer
Nature.

Leekwijck, W. Van and Kerre, E. E. (1999) ‘Defuzzification: criteria and classification’,

Fuzzy Sets and Systems, 108(2), pp. 159–178. doi: 10.1016/s0165-0114(97)00337-
0.

Liao, R. F. et al. (2008) ‘Fuzzy logic control for a petroleum separation process’,
Engineering Applications of Artificial Intelligence, 21(6), pp. 835–845. doi:
10.1016/j.engappai.2007.09.006.

Liu, H. et al. (2013) ‘Phase behavior and compressibility factor of two China gas
condensate samples at pressures up to 95MPa’, Fluid Phase Equilibria, 337, pp. 363–
369. doi: 10.1016/j.fluid.2012.10.011.

Lohrenz, J., Bray, B. G. and Clark, C. R. (1964) ‘Calculating Viscosities of Reservoir

Fluids From Their Compositions’, Journal of Petroleum Technology. Society of
Petroleum Engineers, 16(10), pp. 1171–1176. doi: 10.2118/915-pa.

Londono, F. E., Archer, R. A. and Blasingame, T. A. (2002) ‘Correlations for

248 | P a g e
Hydrocarbon Gas Viscosity and Gas Density - Validation and Correlation of Behavior
Using a Large-Scale Database’, SPE Reservoir Evaluation & Engineering. Society of
Petroleum Engineers, 8(06), pp. 561–572. doi: 10.2118/75721-pa.

Londono, F. E., Archer, R. A. and Blasingame, T. A. (2005) ‘Correlations for

hydrocarbon-gas viscosity and gas density-validation and correlation of behavior using
a large-scale database’, SPE Reservoir Evaluation and Engineering. Society of
Petroleum Engineers, 8(6), pp. 561–572. doi: 10.2118/75721-PA.

Louridas, P. and Ebert, C. (2016) ‘Machine Learning’, IEEE Software. IEEE Computer
Society, 33(5), pp. 110–115. doi: 10.1109/MS.2016.114.

Lucas, K. (1981) ‘Die Druckabhängigkeit der Viskosität von Flüssigkeiten – eine

einfache Abschätzung’, Chemical Engineering Technology, 53(12), pp. 959–960.

Mackay’, D. J. C. (1992) ‘Bayesian Interpolation’, Neural computation, 4(3), pp. 415–

447.

Majidi, S. M. J. et al. (2014) ‘Evolving an accurate model based on machine learning

approach for prediction of dew-point pressure in gas condensate reservoirs’, Chemical
Engineering Research and Design. Institution of Chemical Engineers, 92(5), pp. 891–
902. doi: 10.1016/j.cherd.2013.08.014.

Mamdani, E. H. and Assilian, S. (1975) ‘An experiment in linguistic synthesis with a

fuzzy logic controller’, International Journal of Man-Machine Studies. Academic Press,
7(1), pp. 1–13. doi: 10.1016/S0020-7373(75)80002-2.

Mansour, E. M. et al. (2013) ‘Predicting PVT properties of Egyptian crude oils by a

modified Soave–Redlich–Kowng equation of state’, Egyptian Journal of Petroleum.
Elsevier, 22(1), pp. 137–148. doi: 10.1016/J.EJPE.2012.09.005.

Maravi, Y. D. C. (2003) New Inflow Performance Relationship for Gas-condensate

reservoirs. Texas A&M University.

249 | P a g e
MathWorks (2019) Comparison of Sugeno and Mamdani Systems. Available at:
https://uk.mathworks.com/help/fuzzy/comparison-of-sugeno-and-mamdani-
systems.html (Accessed: 31 July 2019).

Matthews, T. A. and Roland, C. H. (1942) ‘High Pressure Gas Measurement’,

Petroleum Refiner, 20(6), p. 58.

McCain, W. D. and Cawley, G. (1991) ‘Reservoir-Fluid Property Correlations-State of

the Art’, SPE Reservoir Engineering, 6(02), pp. 266–272. doi: 10.2118/18571-pa.

Mesbah, M., Soroush, E. and Rostampour Kakroudi, M. (2017) ‘Predicting physical

properties (viscosity, density, and refractive index) of ternary systems containing 1-
octyl-3-methyl-imidazolium bis(trifluoromethylsulfonyl)imide, esters and alcohols at
298.15 K and atmospheric pressure, using rigorous classification’, Journal of
Molecular Liquids. Elsevier, 225, pp. 778–787. doi: 10.1016/J.MOLLIQ.2016.11.004.

Mmata, B. and Onyekonwu, M. (2014) ‘Estimation of anhydrous oil density for accurate
multiphase flow measurement: A comparative case study’, in SPE Nigeria Annual
International Conference and Exhibition. Lagos, Nigeria: Society of Petroleum
Engineers, pp. 276–284. doi: 10.2118/172369-ms.

Mohaghegh, S. (2000) ‘Virtual-Intelligence Applications in Petroleum Engineering’,

Journal of Petroleum Technology, 52(9)(November 2000), pp. 64–71. doi:
10.2118/62415-MS.

Mohamadi-Baghmolaei, M. et al. (2015) ‘Prediction of gas compressibility factor using

intelligent models’, Natural Gas Industry B. Elsevier B.V., 2(4), pp. 283–294. doi:
10.1016/j.ngib.2015.09.001.

Moses, P. . (1986) ‘Engineering Applications of Phase Behavior of Crude Oil and

Condensate Systems’, Journal of Petroleum Technology, pp. 715–723.

250 | P a g e
Moses, P. . and Donohoe, C. (1987) ‘Gas-Condensate Reservoirs’, in In Petroleum
Engineering Handbook. Richardson: Society of Petroleum Engineers (SPE), p. 28.
Available at: https://www.onepetro.org/book/peh/spe-1987-39-peh (Accessed: 6 July
2020).

Mott, R. (2002) ‘Engineering Calculations of Gas-Condensate-Well Productivity’, SPE

Reservoir Evaluation & Engineering, 6(05), pp. 298–306. doi: 10.2118/86298-PA.

Muskat, M. (1949) Physical principles of oil production. 1st ed. New York, McGraw-Hill
Book Co.

Naderi, M. and Khamehchi, E. (2019) ‘Fuzzy logic coupled with exhaustive search
algorithm for forecasting of petroleum economic parameters’, Journal of Petroleum
Science and Engineering. Elsevier, 176, pp. 291–298. doi:
10.1016/J.PETROL.2019.01.049.

Naseri, A., Nikazar, M. and Mousavi Dehghani, S. A. (2005) ‘A correlation approach

for prediction of crude oil viscosities’, Journal of Petroleum Science and Engineering.
Elsevier, 47(3–4), pp. 163–174. doi: 10.1016/J.PETROL.2005.03.008.

Nasrifar, K. and Moshfeghian, M. (2001) ‘A new cubic equation of state for simple
fluids: pure and mixture’, Fluid Phase Equilibria, 190(1–2), pp. 73–88.

Nassar, I. S., El-Banbi, A. H. and Sayyouh, M. H. (2013) ‘Modified black oil PVT
properties correlations for volatile oil and gas condensate reservoirs’, in Society of
Petroleum Engineers - North Africa Technical Conference and Exhibition 2013, NATC
2013. Cairo: Society of Petroleum Engineers, pp. 998–1016. doi: 10.2118/164712-ms.

Nikravesh, M., Zadeh, A. and Aminzadeh, F. (2003) Soft Computing and Intelligent
Data Analysis in Oil Exploration. 1st edn, Developments in Petroleum Science. 1st
edn. Edited by M. Nikravesh L.A. Zadeh Fred Aminzadeh. Elsevier Science. doi:
10.1016/S0376-7361(03)80005-5.
Norton, S. (2018) Shell Announces Plans to Deploy AI Applications at Scale - CIO

251 | P a g e
Journal. - WSJ, CIO Journal. Available at: https://blogs.wsj.com/cio/2018/09/20/shell-
announces-plans-to-deploy-ai-applications-at-scale/ (Accessed: 8 April 2020).

Nowroozi, S. et al. (2009) ‘Development of a neural fuzzy system for advanced

prediction of dew point pressure in gas condensate reservoirs’, Fuel Processing
Technology, 90(3), pp. 452–457. doi: 10.1016/j.fuproc.2008.11.009.
O’Dell, H. and Miller, R. (1967) ‘Successfully Cycling a Low-Permeability, High-Yield
Gas Condensate Reservoir’, Journal of Petroleum Technology, 19(01), pp. 41–47. doi:
10.2118/1495-PA.

Ogunrewo, O. (2014) Imperial College London. Imperial College London.

Ojha, S. P. et al. (2017) ‘Relative permeability estimates for Wolfcamp and Eagle Ford
shale samples from oil, gas and condensate windows using adsorption-desorption
measurements’, Fuel. Elsevier Ltd, 208, pp. 52–64. doi: 10.1016/j.fuel.2017.07.003.

Olds, R. H., Sage, B. H. and Lacey, W. N. (1945) ‘Volumetric and Phase Behavior of
Oil and Gas from Paloma Field’, Transactions of the AIME. Society of Petroleum
Engineers (SPE), 160(01), pp. 77–99. doi: 10.2118/945077-g.

Olds, R. H., Sage, B. H. and Lacey, W. N. (1949) ‘Volumetric and Viscosity Studies of
Oil and Gas from a San Joaquin Valley Field’, Transactions of the AIME. Society of
Petroleum Engineers (SPE), 179(01), pp. 287–302. doi: 10.2118/949287-g.

Ostermann, R. D. and Owolabi, O. O. (1983) ‘Correlations for the Reservoir Fluid

Properties of Alaskan Crudes’, in SPE California Regional Meeting. Ventura: Society
of Petroleum Engineers. doi: 10.2118/11703-MS.

Papay, P. (1968) ‘A Termelestechologiai Parametrrek Valtozasa a Gazlelepk

Muvelese Soran’, in Ogil Musz. Budapest, pp. 267–273.

Passino, K. and Yurkovich, S. (1998) Fuzzy Control: The Basics, Fuzzy Control.
California: Addison Wesley Longman, Inc.
Pedersen, K. S. and Fredenslund, A. (1987) ‘An improved corresponding states model

252 | P a g e
for the prediction of oil and gas viscosities and thermal conductivities’, Chemical
Engineering Science. Pergamon, 42(1), pp. 182–186. doi: 10.1016/0009-
2509(87)80225-7.

Pedersen, K. S., Thomassen, P. and Fredenslund, A. (1984) ‘Thermodynamics of

Petroleum Mixtures Containing Heavy Hydrocarbons. 1. Phase Envelope Calculations
by Use of the Soave-Redlich-Kwong Equation of State’, Industrial and Engineering
Chemistry Process Design and Development. American Chemical Society, 23(1), pp.
163–170. doi: 10.1021/i200024a027.

Pelckmans, K. et al. (2002) LS-SVMlab: a MATLAB/C toolbox for Least Squares

Support Vector Machines, Tutorial. KULeuven- ESAT. Leuven-Heverlee. Available at:
ftp://ftp.esat.kuleuven.be/stadius/kpelckma/kp02-44.pdf.

Peng, D. Y. and Robinson, D. B. (1976) ‘A New Two-Constant Equation of State’,

Industrial and Engineering Chemistry Fundamentals. American Chemical Society,
15(1), pp. 59–64. doi: 10.1021/i160057a011.

Penula, G. (2003) ‘Gas-Condensate Well-Test Analysis with and without the Relative
Permeability Curves’, in Annual Technical Conference and Exhibition held in Dallas.
Dallas: SPE. doi: 10.1306/a9673950-1738-11d7-8645000102c1865d.

Petrosky, G. E. and Farshad, F. F. (1998) ‘Viscosity correlations for gulf of Mexico

crude oils’, SPE Reservoir Evaluation & Engineering, 01(05), pp. 249–258. doi:
10.2523/29468-ms.

Piper, L. D., McCain, W. D. and Corredor, J. H. (1993) ‘Compressibility Factors for

Naturally Occurring Petroleum Gases’, in SPE Annual Technical Conference and
Exhibition. Houston: Society of Petroleum Engineers, pp. 661–671. doi:
10.2523/26668-ms.

Raghavan, R., Chu, W. C. and Jones, J. R. (1995) ‘Practical Considerations in the

Analysis of Gas-Condensate Well Tests’, in SPE Annual Technical Conference and

253 | P a g e
Exhibition. Dallas, Texas: Society of Petroleum Engineers. doi: 10.2118/30576-MS.

Raghavan, R. and Jones, J. R. (1996) ‘Depletion performance of gas-condensate

reservoirs’, JPT, Journal of Petroleum Technology. Society of Petroleum Engineers
(SPE), 48(8), pp. 725–731. doi: 10.2118/36352-JPT.

Rahimzadeh, A. et al. (2016) ‘Condensate blockage study in gas condensate

reservoir’, Journal of Natural Gas Science and Engineering. Elsevier, 33, pp. 634–643.
doi: 10.1016/J.JNGSE.2016.05.048.

Rawlins, E. and Schellhardt, M. (1936) ‘Back-Pressure Data on Natural Gas Wells and
Their Application to Production Practices’, in Bureau of Mines Monograph. Bureau of
Mines Monograph.

Rayes, D. G. et al. (1992) ‘Two-Phase Compressibility Factors for Retrograde Gases’,

SPE Formation Evaluation. Society of Petroleum Engineers, 7(01), pp. 87–92. doi:
10.2118/20055-PA.

Redlich, O. and Kwong, J. N. (1949) ‘On the Thermodynamics of solutions. V An

equation of state. Fugacities of gaseous solutions’, Chemical Reviews, 44(1), pp. 233–
244.

Riazi, M. R. and Daubert, T. E. (1987) ‘Characterization Parameters for Petroleum

Fractions’, Industrial and Engineering Chemistry Research. American Chemical
Society, 26(4), pp. 755–759. doi: 10.1021/ie00064a023.

Rollins, J. B., McCain, W. D. and Creeger, J. T. (1990) ‘Estimation of solution GOR of

black oils’, JPT, Journal of Petroleum Technology. Society of Petroleum Engineers,
42(1), pp. 92–94. doi: 10.2118/18602-pa.

Ross, T. J. (2017) Fuzzy Logic with Engineering Applications. 4th edn. Chichester:
Wiley.
Rostami, A., Hemmati-Sarapardeh, A. and Shamshirband, S. (2018) ‘Rigorous

254 | P a g e
prognostication of natural gas viscosity: Smart modeling and comparative study’, Fuel,
222, pp. 766–778. doi: 10.1016/j.fuel.2018.02.069.

Roussennac, B. (2001) Gas condensate well test analysis, Department of Petroleum

Engineering. Standford University. Available at:
https://pangea.stanford.edu/ERE/pdf/pereports/MS/Roussennac01.pdf (Accessed: 30
April 2019).

Rowe, A. . (1978) Internally Consisitent Corrrelation for Predicting phase Composition

of Heptane and Heavier Fractions. Tulsa.

Rubin, B. and Buchanan, W. L. (1985) ‘General purpose compositional model’, Society

of Petroleum Engineers journal. Society of Petroleum Engineers, 25(2), pp. 202–214.
doi: 10.2118/11713-PA.

Safari, H. et al. (2014) ‘Assessing the Dynamic Viscosity of Na−K−Ca−Cl−H 2 O

Aqueous Solutions at High-Pressure and High-Temperature Conditions’, Industrial &
Engineering Chemistry Research, 53, pp. 11488–11500. doi: 10.1021/ie501702z.

Saghafi, H. and Arabloo, M. (2018) ‘Development of genetic programming (GP)

models for gas condensate compressibility factor determination below dew point
pressure’, Journal of Petroleum Science and Engineering. Elsevier B.V., 171, pp. 890–
904. doi: 10.1016/j.petrol.2018.08.020.

Samuel, A. L. (1959) ‘Some Studies in Machine Learning Using the Game of

Checkers’, IBM Journal of Research and Development, 3(3), pp. 535–554.

Sanjari, E. and Lay, E. N. (2012) ‘An accurate empirical correlation for predicting
natural gas compressibility factors’, Journal of Natural Gas Chemistry, 21, pp. 184–
188. doi: 10.1016/S1003-9953(11)60352-6.

Sanni, M. (2018) Petroleum engineering : principles, calculations, and workflows. John

Wiley & Sons, Incorporated. doi: 10.1002/9781119387985.

255 | P a g e
Sarkar, R., Danesh, A. S. and Todd, A. C. (1991) ‘Phase Behavior Modeling of Gas-
Condensate Fluids Using an Equation of State’, in SPE Annual Technical Conference
and Exhibition. Dallas: Society of Petroleum Engineers. doi: 10.2118/22714-MS.

Serra, K., Peres, A. M. and Reynolds, A. C. (1990) ‘Well-Test Analysis for Solution-
Gas-Drive Reservoirs: Part 1-Determination of Relative and Absolute Permeabilities’,
SPE Formation Evaluation, 05(02), pp. 124–132.

Serra, K. V., Peres, A. M. M. and Reynolds, A. C. (2007) ‘Well-Test Analysis for

Solution-Gas-Drive Reservoirs: Part 2-Buildup Analysis’, SPE Formation Evaluation.
Society of Petroleum Engineers, 5(02), pp. 133–140. doi: 10.2118/17048-pa.

Shi, C. (2009) Flow Behavior of gas-condenate wells. Stanford University.

Shokir, E. M. (2006) ‘A Novel Model for Permeability Prediction in Uncored Wells’,

SPE Reservoir Evaluation & Engineering. Society of Petroleum Engineers, 9(03), pp.
266–273. doi: 10.2118/87038-pa.

Shokir, E. M. (2008) ‘Novel Density and Viscosity Correlations for Gases and Gas
Mixtures Containing Hydrocarbon and Non-Hydrocarbon Components’, Journal of
Canadian Petroleum Technology. Petroleum Society of Canada, 47(10). doi:
10.2118/08-10-45.

Smits, R. M. M., Van der Post, N. and Al Shaidi, S. M. (2001) ‘Accurate Prediction of
Well Requirements in Gas Condensate Fields’, in Proceedings of the Middle East Oil
Show. Manama: Society of Petroleum Engineers, pp. 611–621. doi: 10.2118/68173-
ms.

Soave, G. (1972) Equilibrium constants from a modified Redkh-Kwong equation of

state, Chemical Engineering Science. Pergamon Press.

Soroush, E. et al. (2015) ‘Evolving a robust modeling tool for prediction of natural gas

256 | P a g e
hydrate formation conditions’, Journal of Unconventional Oil and Gas Resources.
Elsevier, 12, pp. 45–55. doi: 10.1016/J.JUOGR.2015.09.002.

Sousa, P. C. De, Garcia, A. P. and Waltrich, P. J. (2017) ‘Analytical Development of a

Dynamic IPR for Transient Two-Phase Flow in Reservoirs’, in SPE Annual Technical
Conference and Exhibition. San Antonio: Society of Petroleum Engineers. doi:
10.2118/187232-MS.

Spivak, A. and Dixon, T. N. (1973) ‘Simulation Of Gas-Condensate Reservoirs’, in SPE

Symposium on Numerical Simulation of Reservoir Performance. Houston: Society of
Petroleum Engineers. doi: 10.2118/4271-MS.

Standing, M. B. (1947) ‘A Pressure-Volume-Temperature Correlation For Mixtures Of

California Oils And Gases’, Drilling and Production Practice. American Petroleum
Institute, 275.

Standing, M. B. (1974) Petroleum Engineering Data Book. Trondheim: Norwegian Inst.

of Technology.

Standing, M. B. (1979) Calculating solutions gas drive reservoirs performance by

Tarner’s method. doi: 10.1017/CBO9781107415324.004.

Standing, M. B. (1981) Volumetric and Phase Behavior of Oil Field Hydrocarbon

Systems. 1st edn. Texas: SPE Richardson.

Standing, M. B. and Katz, D. L. (1942a) ‘Density of Crude Oils Saturated with Natural
Gas’, Transactions of the AIME. Society of Petroleum Engineers (SPE), 146(01), pp.
159–165. doi: 10.2118/942159-g.

Standing, M. B. and Katz, D. L. (1942b) ‘Density of Natural Gases’, Transactions of

the AIME. Society of Petroleum Engineers, 146(01), pp. 140–149. doi:
10.2118/942140-g.
Starling, K. E. and Ellington, R. T. (1964) ‘Viscosity correlations for nonpolar dense

257 | P a g e
fluids’, AIChE Journal. John Wiley & Sons, Ltd, 10(1), pp. 11–15. doi:
10.1002/aic.690100112.

Stewart, G. (2011) Well Test Design and Analysis. Tulsa: PennWell Corporation.
Available at: https://capitadiscovery.co.uk/teesside-ac/items/570433 (Accessed: 1
May 2019).

Stewart, W., Burkhardt, S. and Voo, D. (1959) ‘Prediction of pseudocritical parameters

for mixtures’, in AIChE Meeting. Kansas City.

Stiel, L. I. and Thodos, G. (1962) ‘The viscosity of polar gases at normal pressures’,
AIChE Journal, 8(2), pp. 229–232. doi: 10.1002/aic.690080220.

Stone, H. L. (1970) ‘Probability model for estimating three-phase relative permeability’,

Journal of Petroleum Technology. Society of Petroleum Engineers, 22(2), pp. 214–
218. doi: 10.2118/2116-pa.

Stone, H. L. (1973) ‘Estimation of three -phase relative permeability and residual oil
data’, Journal of Canadian Petroleum Technology. Petroleum Society of Canada,
12(4), pp. 53–61. doi: 10.2118/73-04-06.

Strand, K. A. and Bjørkvik, B. J. A. (2019) ‘Interface light-scattering on a methane–

decane system in the near-critical region at 37.8 °C (100 °F)’, Fluid Phase Equilibria.
Elsevier B.V., 485, pp. 168–182. doi: 10.1016/j.fluid.2018.12.016.

Sumnu-Dindoruk, M. D. and Jones, J. R. (1998) ‘Determination of Gas-Condensate

Relative Permeabilities from Field Production Data’, in SPE Annual Technical
Conference and Exhibition. New Orleans: Society of Petroleum Engineers. doi:
10.2118/49267-MS.

Sun, C. Y. et al. (2012) ‘Experiments and modeling of volumetric properties and phase
behavior for condensate gas under ultra-high-pressure conditions’, Industrial and
Engineering Chemistry Research, 51(19), pp. 6916–6925. doi: 10.1021/ie2025757.

258 | P a g e
Sutton, R. P. (1985) ‘Compressibility Factors for High-Molecular-Weight Reservoir
Gases’, in SPE Annual Technical Conference and Exhibition. Las Vegas: Society of
Petroleum Engineers. doi: 10.2118/14265-ms.

Sutton, R. P. (2005a) ‘Fundamental PVT Calculations for Associated and Gas-

Condensate Natural Gas Systems’, in SPE Annual Technical Conference and
Exhibition. Dallas: Society of Petroleum Engineers. doi: 10.2118/97099-MS.

Sutton, R. P. (2005b) ‘Fundamental PVT Calculations for Associated and

Gas/Condensate Natural-Gas Systems’, SPE Reservoir Evaluation & Engineering.
Society of Petroleum Engineers, 10(03), pp. 270–284. doi: 10.2118/97099-pa.

Sutton, R. P. (2007) ‘Fundamental PVT calculations for associated and

gas/condensate natural-gas systems’, SPE Reservoir Evaluation and Engineering.
Society of Petroleum Engineers, 10(3), pp. 270–284. doi: 10.2118/97099-pa.

Sutton, R. P. and Farshad, F. (1990) ‘Evaluation of Empirically Derived PVT Properties

for Gulf of Mexico Crude Oils’, SPE Reservoir Engineering (Society of Petroleum
Engineers), 5(01).

Suykens, J. A. K. et al. (2002) ‘Weighted least squares support vector machines:

robustness and sparse approximation’, Neurocomputing. Elsevier, 48(1–4), pp. 85–
105. doi: 10.1016/S0925-2312(01)00644-0.

Suykens, J. A. K. and Vandewalle, J. (1999) ‘Least Squares Support Vector Machine

Classifiers’, Neural Processing Letters, 9(June), pp. 293–300. doi: 10.1023/A.

Takacs, G. (1976) ‘Comparison Made for Computer Z-Factor Calculation’, Oil and Gas
Journal, pp. 64–66.

Takagi, T. and Sugeno, M. (1985) ‘Fuzzy Identification of Systems and Its Applications
to Modeling and Control’, IEEE Transactions on Systems, Man and Cybernetics. IEEE,

259 | P a g e
SMC-15(1), pp. 116–132. doi: 10.1109/TSMC.1985.6313399.

Tanasa, D. E. et al. (2013) ‘Photodegradation process of Eosin y using ZnO/SnO2

nanocomposites as photocatalysts: Experimental study and neural network modeling’,
Journal of Materials Science. Springer, 48(22), pp. 8029–8040. doi: 10.1007/s10853-
013-7617-x.

Tarner, J. (1944) ‘How different size gas caps and pressure maintenance programs
affect amount of recoverable oil’, Oil Weekly, 144, pp. 32–34.

Thomas, F. B., Bennion, D. B. and Andersen, G. (2009) ‘Gas Condensate Reservoir

Performance’, Journal of Canadian Petroleum Technology, 48(07), pp. 18–24. doi:
10.2118/09-07-18.

Thornton, F. (1946) ‘Gas-Condensate Reservoirs-a Review’, Production Practice, pp.

150–159.

Tillerson, R. (2015) Natural Gas and the Policies of the Future. Available at:
https://corporate.exxonmobil.com/News/Newsroom/Speeches/2015/0602_Natural-
Gas-and-the-Policies-of-the-Future (Accessed: 9 March 2020).

Ugwu, J. O. (2011) A Semi-Eemerical Approach To Modelling Well Deliverability In

Gas Condensate Reservoirs. Robert Gordon University Aberdeen. doi:
10.1177/0266242610391936.

Vazquez, M. and Beggs, H. D. (1980) ‘Correlations for fluid physical property

prediction’, Journal of Petroleum Technology. Society of Petroleum Engineers (SPE),
32(06), pp. 968–970. doi: 10.2118/6719-pa.

Vo, D. T., Jones, J. R. and Raghavan, R. (1989) ‘Performance predictions for gas-
condensate reservoirs’, SPE Formation Evaluation. Society of Petroleum Engineers,
4(4), pp. 576-584 16984. doi: 10.2118/16984-pa.
Vo, H. X. (2010) Composition Variation During Flow of Gas-Condensate Wells.

260 | P a g e
Standford University.

Van der Waals, J. . (1873) Over de Continuiteit van den Gas-en Vloeistoftoestand (On
the Continuity of the Gas and Liquid State). University of Leiden.

Walsh, M. and Lake, L. (2003) A generalized approach to primary hydrocarbon

recovery. 1st edn. Amsterdam: Elsevier Ltd.

Warsito, B., Santoso, R. and Yasin, H. (2018) ‘Cascade Forward Neural Network for
Time Series Prediction’, Journal of Physics: Conference Series, 1025, p. 12097. doi:
10.1088/1742-6596/1025/1/012097.

Wheaton, R. J. and Zhang, H. R. (2007) ‘Condensate Banking Dynamics in Gas

Condensate Fields: Compositional Changes and Condensate Accumulation Around
Production Wells’, in. doi: 10.2118/62930-ms.

Whitson, C. and Brulé, M. (2000) Phase Behavior. 1st edn, Society of Petroleum
Engineers. 1st edn. Richardson: Society of Petroleum Engineers. doi:
10.1021/ma00080a014.

Whitson, C. H. (1983) ‘Characterizing hydrocarbon plus fractions’, Society of

Petroleum Engineers journal. Society of Petroleum Engineers, 23(4), pp. 683–694.
doi: 10.2118/12233-PA.

Whitson, C. H. (1989) ‘Discussion of An Improved Method for the Determination of the

Reservoir-Gas Specific-Gravity for Retrograde Gases’, Journal of Petroleum
Technology, 1216.

Whitson, C. H. (2006) ‘Pressure/Volume/Temperature (PVT) treatment’, in Petroleum

engineering handbook Volume V, Reservoir engineering and petrophysics. Texas:
Society of Petroleum Engineers, pp. 1458–1465.

Whitson, C. H., Fevang, Ø. and Yang, T. (1999) ‘Gas Condensate PVT – What’s Really

261 | P a g e
Important and Why?’, in Optimisation of Gas Condensate Fields. London: Norwegian
U. of Science and Technology. doi: 10.2118/117930-PA.

Whitson, C. H. and Kuntadi, A. (2005) ‘Khuff gas condensate development’, in 2005

International Petroleum Technology Conference Proceedings. Doha, pp. 1191–1209.
doi: 10.2523/iptc-10692-ms.

Whitson, C. H., Da Silva, F. and Soreide, I. (1988) ‘Simplified compositional

formulation for modified black-oil simulators’, in SPE Annual Technical Conference
and Exhibition. Houston: Society of Petroleum Engineers. doi: 10.2118/18315-MS.

Whitson, C. H. and Torp, S. B. (1983) ‘Evaluating Constant Volume Depletion Data’,

JPT, Journal of Petroleum Technology, 35(3), pp. 610–620. doi: 10.2118/10067-PA.

Wichert, E. and Aziz, K. (1972) ‘Calculation of Z’s for Sour Gases’, Hydrocarbon
Processing, 51(5), pp. 119–122.

Wyllie, M. R. J. (1951) ‘A Note on the Interrelationship Between Wetting and

Nonwetting Phase Relative Permeability’, Journal of Petroleum Technology. Society
of Petroleum Engineers (SPE), 3(10), pp. 17–17. doi: 10.2118/951381-g.

Xiao, J. J. and Al-Muraikhi, A. J. (2004) ‘A New Method for the Determination of Gas
Condensate Well Production Performance’, in SPE Annual Technical Conference and
Exhibition. Texas: Society of Petroleum Engineers. doi: 10.2118/90290-MS.

Xu, S. and Lee, W. J. (1999) ‘Two-phase well test analysis of gas condensate
reservoirs’, in Proceedings - SPE Annual Technical Conference and Exhibition.
Houston, Texas: Soc Pet Eng (SPE). doi: 10.2118/56483-ms.

Yager, R. R. and Filev, D. P. (1994) ‘Generation of Fuzzy Rules by Mountain

Clustering’, Journal of Intelligent & Fuzzy Systems: Applications in Engineering and
Technology, 2(3), pp. 209–219.
Yan, K. Le et al. (2013) ‘Measurement and calculation of gas compressibility factor for

262 | P a g e
condensate gas and natural gas under pressure up to 116 MPa’, Journal of Chemical
Thermodynamics. Academic Press, 63, pp. 38–43. doi: 10.1016/j.jct.2013.03.025.

Yang, T. et al. (2007) ‘LBC Viscosity Modeling of Gas Condensate to Heavy Oil’, in
SPE Annual Technical Conference and Exhibition. Anaheim: Society of Petroleum
Engineers. doi: 10.2523/109892-ms.

Yang, T., Chen, W. . and Guo, T. . (1997) ‘Phase behavior of a near-critical reservoir
fluid mixture’, Fluid Phase Equilibria, 128, pp. 183–197.

Yarborough, L. and Hall, K. R. (1974) ‘How to Solve EOS for Z-factors’, Oil and Gas
Journal, 86.

Zadeh, L. . (1965) ‘Fuzzy Sets’, Information and control, 8(3), pp. 338–353.

Zadeh, L. A. (1973) ‘Outline of A New Approach to the Analysis of Complex Systems’,

IEEE Trans. Syst. Man, Cybern., SMC-3(1), pp. 28–44.

Zendehboudi, S. et al. (2012) ‘Prediction of Condensate-to-Gas Ratio for Retrograde

Gas Condensate Reservoirs Using Artificial Neural Network with Particle Swarm
Optimization’, Energy & Fuels, 26(6), pp. 3432–3447. doi: 10.1021/ef300443j.

Zhang, W. et al. (2019) ‘Production performance analysis for horizontal wells in gas
condensate reservoir using three-region model’, Journal of Natural Gas Science and
Engineering. Elsevier B.V., 61, pp. 226–236. doi: 10.1016/j.jngse.2018.11.004.

Zhanggui, L. et al. (1998) ‘Integration of Fuzzy Methods into Geostatistics for

Petrophysical Property Distribution’, in SPE Asia Pacific Oil and Gas Conference and
Exhibition. Perth: Society of Petroleum Engineers (SPE). doi: 10.2118/49964-ms.

Zhao, X. et al. (2018) ‘Reservoir geostatistical estimation of imprecise information

using fuzzy kriging approach’, in SPE Western Regional Meeting Proceedings.
California: Society of Petroleum Engineers. doi: 10.2118/190051-pa.

263 | P a g e
Zhou, C. D., Wu, X. L. and Cheng, J. A. (1993) ‘Determining reservoir properties in
reservoir studies using a fuzzy neural network’, in Proceedings - SPE Annual
Technical Conference and Exhibition. Houston: Society of Petroleum Engineers, pp.
141–150. doi: 10.2523/26430-ms.

264 | P a g e
APPENDIX A: Sample of collected data for development of
condensate viscosity models

Table A1: Sample data used for development of condensate viscosity of gas-condensate
reservoirs.

Pressure Temperature API 𝜸𝑮𝒂𝒔 Rs Standing Condensate

(psia) (F) Gravity viscosity
(cp)
2393 167 60 0.845 2661 0.04
2537 109 60 1.2514 4312 0.0404
2730 189 60 1.2514 4510 0.0433
2320 167 60 0.845 2574 0.045
2248 167 60 0.845 2488 0.0475
2175 167 60 0.845 2401 0.05
2103 167 60 0.845 2315 0.055
2030 167 60 0.845 2229 0.0575
1958 167 60 0.845 2143 0.06
2283 109 60 1.2514 3845 0.0649
1885 167 60 0.845 2057 0.065
1813 167 60 0.845 1972 0.0675
2184 189 65 1.2514 3573 0.0678
1740 167 60 0.845 1887 0.07
1667 167 60 0.845 1802 0.0725
1775 109 65 1.2514 2950 0.0739
1595 167 60 0.845 1717 0.075
1522 167 60 0.845 1633 0.0785
2457 189 60 1.2514 4026 0.0794
1450.4 167 60 0.845 1549 0.08
1522.2 109 65 1.2514 2493 0.0821
1377.88 167 60 0.845 1465 0.0825
2029.6 109 65 1.2514 3384 0.0849
1305.36 167 60 0.8451 1382 0.085
5801 248 62 0.5757 4577 0.086
1232.84 167 60 0.845 1299 0.0875
1268.5 109 65 0.7581 2043 0.0899

265 | P a g e
 APPENDIX B: The developed Matlab codes for prediction
of Two-phase Z factor
B1. Cascade Forward Neural Network (CFNN) code
% Calculating Two-phase compressibility factor of gas condensate reservoirs
% Input data:
% (Tem,Pres,H2S,CO2,N2,SGgas,C1,C2,C3,IC4,NC4,IC5,NC5,C6,C7+,MWC7+,SGc7+)
% Target data is 2-Phase compressibility factor
% One hidden layer /the x and t data should be defined for running the code
% Created by Foad Faraji on 30-Apr-2020 22:19:05
% This script assumes these variables are defined:
%
% x - input data.
% t - target data.

x = input';
t = target';

% Choose a Training Function

% For a list of all training functions type: help nntrain
% 'trainlm' is usually fastest.
% 'trainbr' takes longer but may be better for challenging problems.
% 'trainscg' uses less memory. Suitable in low memory situations.
trainFcn = 'trainlm'; % Levenberg-Marquardt backpropagation.

% Create a Fitting Network

hiddenLayerSize = 5;
net = cascadeforwardnet(hiddenLayerSize,trainFcn);

% Setup Division of Data for Training, Validation, Testing

net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;

% Train the Network

[net,tr] = train(net,x,t);

% Test the Network

y = net(x);
e = gsubtract(t,y);
performance = perform(net,t,y)

% View the Network

view(net)
view(plotregression)

% Plots
% Uncomment these lines to enable various plots.
%figure, plotperform(tr)
%figure, plottrainstate(tr)
%figure, ploterrhist(e)
%figure, plotregression(t,y)
%figure, plotfit(net,x,t)

B2. Feed Forward Neural Network (FFNN) code

% Solve an Input-Output Fitting problem with a Neural Network

266 | P a g e
% Script generated by Neural Fitting app
% Created by Foad Faraji on 30-Apr-2020 22:01:22
%
% This script assumes these variables are defined:
%
% data - input data.
% data_1 - target data.

x = input';
t = target';

% Choose a Training Function

% For a list of all training functions type: help nntrain
% 'trainlm' is usually fastest.
% 'trainbr' takes longer but may be better for challenging problems.
% 'trainscg' uses less memory. Suitable in low memory situations.
trainFcn = 'trainlm'; % Levenberg-Marquardt backpropagation.

% Create a Fitting Network

hiddenLayerSize = 30;
net = fitnet(hiddenLayerSize,trainFcn);

% Setup Division of Data for Training, Validation, Testing

net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;

% Train the Network

[net,tr] = train(net,x,t);

% Test the Network

y = net(x);
e = gsubtract(t,y);
performance = perform(net,t,y)

% View the Network

view(net)

% Plots
% Uncomment these lines to enable various plots.
%figure, plotperform(tr)
%figure, plottrainstate(tr)
%figure, ploterrhist(e)
%figure, plotregression(t,y)
%figure, plotfit(net,x,t)

B3. The Developed ANFIS code for prediction of two-phase Z factor

[Developed by Foad Faraji]
Name='Gas-condensate two-phase Zfactor'
Type='ANFIS'
Version=3.0
NumInputs=16
NumOutputs=1
NumRules=8
AndMethod='prod'
OrMethod='probor'
267 | P a g e
ImpMethod='prod'
AggMethod='sum'
DefuzzMethod='wtaver'

[Input1]
Name='in1'
Range=[86 325]
NumMFs=8
MF1='in1cluster1':'gaussmf',[43.939616492805 212.000000668321]
MF2='in1cluster2':'gaussmf',[43.9396154549174 104.000000079057]
MF3='in1cluster3':'gaussmf',[43.9396152970173 235.000000099111]
MF4='in1cluster4':'gaussmf',[43.9396155607974 139.999999943725]
MF5='in1cluster5':'gaussmf',[43.9396153693023 279.999999687274]
MF6='in1cluster6':'gaussmf',[43.9396154076292 175.999999970267]
MF7='in1cluster7':'gaussmf',[43.9396157558023 255.000000125218]
MF8='in1cluster8':'gaussmf',[43.9396152975791 251.000000132237]

[Input2]
Name='in2'
Range=[200 23244]
NumMFs=8
MF1='in2cluster1':'gaussmf',[4236.58785306734 10581.9999999971]
MF2='in2cluster2':'gaussmf',[4236.58785307214 11039]
MF3='in2cluster3':'gaussmf',[4236.58785307423 2250.00000000649]
MF4='in2cluster4':'gaussmf',[4236.58785307181 7263.00000000187]
MF5='in2cluster5':'gaussmf',[4236.58785308286 14342.0000000042]
MF6='in2cluster6':'gaussmf',[4236.58785307428 15368.9999999975]
MF7='in2cluster7':'gaussmf',[4236.5878530787 5952.0000000046]
MF8='in2cluster8':'gaussmf',[4236.58785307252 3014.9999999937]

[Input3]
Name='in3'
Range=[0 0.3182]
NumMFs=8
MF1='in3cluster1':'gaussmf',[0.0585008072527943 2.93425098980129e-06]
MF2='in3cluster2':'gaussmf',[0.0585003547682657 -4.35711822099071e-08]
MF3='in3cluster3':'gaussmf',[0.0585125232360066 3.920654716736e-06]
MF4='in3cluster4':'gaussmf',[0.0585003585012129 3.57358628042432e-09]
MF5='in3cluster5':'gaussmf',[0.0585003749078678 2.37195865262904e-08]
MF6='in3cluster6':'gaussmf',[0.0585003559359281 -3.53122863778846e-08]
MF7='in3cluster7':'gaussmf',[0.0585003532042919 -4.60732216844032e-08]
MF8='in3cluster8':'gaussmf',[0.0584877379626363 -6.75724218907177e-06]

[Input4]
Name='in4'
Range=[0.0001 0.2084]
NumMFs=8
MF1='in4cluster1':'gaussmf',[0.0401613083121935 0.00458163909699585]
MF2='in4cluster2':'gaussmf',[0.0382700153253149 0.00419517750559465]
MF3='in4cluster3':'gaussmf',[0.0418651236512032 0.00983790276288075]
MF4='in4cluster4':'gaussmf',[0.0382976043771252 0.00560357238162204]
MF5='in4cluster5':'gaussmf',[0.0383076571498975 0.00419856705369914]
MF6='in4cluster6':'gaussmf',[0.0382724375195043 0.00309742082476367]
MF7='in4cluster7':'gaussmf',[0.0382677423396701 0.0055808663568813]
MF8='in4cluster8':'gaussmf',[0.0330126699595674 0.00780999629460519]

268 | P a g e
[Input5]
Name='in5'
Range=[0 0.1284]
NumMFs=8
MF1='in5cluster1':'gaussmf',[0.0238315335879429 0.00627438049210001]
MF2='in5cluster2':'gaussmf',[0.0236044365109744 0.00621246344533544]
MF3='in5cluster3':'gaussmf',[0.0234589814159074 0.000438829119254106]
MF4='in5cluster4':'gaussmf',[0.023606925598594 0.00610490708916884]
MF5='in5cluster5':'gaussmf',[0.023617800021654 0.00630146893508594]
MF6='in5cluster6':'gaussmf',[0.0236032858024176 0.00618305313871095]
MF7='in5cluster7':'gaussmf',[0.0236003856299571 0.00608143238616504]
MF8='in5cluster8':'gaussmf',[0.0238839586551994 0.0116440844337897]

[Input6]
Name='in6'
Range=[0.0687 0.9668]
NumMFs=8
MF1='in6cluster1':'gaussmf',[0.165150553187308 0.884578858736474]
MF2='in6cluster2':'gaussmf',[0.165111896601757 0.88459081776779]
MF3='in6cluster3':'gaussmf',[0.165210031746201 0.825402307244853]
MF4='in6cluster4':'gaussmf',[0.165113507316405 0.964900845674523]
MF5='in6cluster5':'gaussmf',[0.165107839743975 0.884578157602831]
MF6='in6cluster6':'gaussmf',[0.165110322343173 0.966804385058771]
MF7='in6cluster7':'gaussmf',[0.165121263844434 0.964881358891908]
MF8='in6cluster8':'gaussmf',[0.165066896995706 0.755766146908575]

[Input7]
Name='in7'
Range=[0.0164 0.944]
NumMFs=8
MF1='in7cluster1':'gaussmf',[0.170537156326957 0.0709951928989352]
MF2='in7cluster2':'gaussmf',[0.170535976525295 0.0710045817804966]
MF3='in7cluster3':'gaussmf',[0.170823693963334 0.0901375904915624]
MF4='in7cluster4':'gaussmf',[0.170537007754508 0.0163991704194279]
MF5='in7cluster5':'gaussmf',[0.170534925678062 0.0710112470165836]
MF6='in7cluster6':'gaussmf',[0.170536114129939 0.0174970343925587]
MF7='in7cluster7':'gaussmf',[0.170541037663105 0.0164129045437485]
MF8='in7cluster8':'gaussmf',[0.170250455876678 0.0994769386419807]

[Input8]
Name='in8'
Range=[0.0022 0.1018]
NumMFs=8
MF1='in8cluster1':'gaussmf',[0.0182785732196822 0.0131446902790636]
MF2='in8cluster2':'gaussmf',[0.0183285883784814 0.0133242186889857]
MF3='in8cluster3':'gaussmf',[0.0178949230235043 0.0454564001090557]
MF4='in8cluster4':'gaussmf',[0.0183113629133724 0.00249286416737207]
MF5='in8cluster5':'gaussmf',[0.0182357688175312 0.0134669529963677]
MF6='in8cluster6':'gaussmf',[0.0182850669469996 0.0021667899903619]
MF7='in8cluster7':'gaussmf',[0.0184157915219734 0.00270102537643144]
MF8='in8cluster8':'gaussmf',[0.0180152760310817 0.0477535134282819]

[Input9]
Name='in9'
Range=[0.0005 0.0404]
NumMFs=8
MF1='in9cluster1':'gaussmf',[0.00731487820979324 0.00248705844697268]
MF2='in9cluster2':'gaussmf',[0.007333262442539 0.00273834693161455]
MF3='in9cluster3':'gaussmf',[0.00604639249004752 0.00766948272983621]
MF4='in9cluster4':'gaussmf',[0.00733666244176128 0.000595313425476853]
269 | P a g e
MF5='in9cluster5':'gaussmf',[0.00729450032120393 0.00295481243543017]
MF6='in9cluster6':'gaussmf',[0.00732367459622212 0.000474267243179914]
MF7='in9cluster7':'gaussmf',[0.00738900334100898 0.00082534739516243]
MF8='in9cluster8':'gaussmf',[0.0076818000584948 0.0125581351570796]

[Input10]
Name='in10'
Range=[0 0.0392]
NumMFs=8
MF1='in10cluster1':'gaussmf',[0.00747277691174156 0.00274451654371034]
MF2='in10cluster2':'gaussmf',[0.00738008536525681 0.0028322815061444]
MF3='in10cluster3':'gaussmf',[0.00666993738573774 0.0125045152448831]
MF4='in10cluster4':'gaussmf',[0.00721139358378071 0.000699342876967532]
MF5='in10cluster5':'gaussmf',[0.00717791263011896 0.00300936669590871]
MF6='in10cluster6':'gaussmf',[0.00721183693677959 0.000580641875893435]
MF7='in10cluster7':'gaussmf',[0.00724359398144546 0.000906893188813673]
MF8='in10cluster8':'gaussmf',[0.00510787308580824 0.020523686456587]

[Input11]
Name='in11'
Range=[0.0003 0.06]
NumMFs=8
MF1='in11cluster1':'gaussmf',[0.010989628253451 0.00102876181349276]
MF2='in11cluster2':'gaussmf',[0.0109804851472784 0.00103494165778055]
MF3='in11cluster3':'gaussmf',[0.0110944296216656 0.00365004369558179]
MF4='in11cluster4':'gaussmf',[0.0109758926313568 0.000299451581006178]
MF5='in11cluster5':'gaussmf',[0.010975064190915 0.00106588632379741]
MF6='in11cluster6':'gaussmf',[0.010975989920257 0.000298791894919426]
MF7='in11cluster7':'gaussmf',[0.0109769927203072 0.000334011467299146]
MF8='in11cluster8':'gaussmf',[0.0109001501554013 0.00885089077594512]

[Input12]
Name='in12'
Range=[0 0.0216]
NumMFs=8
MF1='in12cluster1':'gaussmf',[0.00457778955022597 0.00111196522778592]
MF2='in12cluster2':'gaussmf',[0.00423289799883626 0.00101248197634151]
MF3='in12cluster3':'gaussmf',[0.00619416038926742 0.00558639453008147]
MF4='in12cluster4':'gaussmf',[0.00397464927220307 0.000189386843120908]
MF5='in12cluster5':'gaussmf',[0.00398812565111719 0.00110093043615154]
MF6='in12cluster6':'gaussmf',[0.00398861318747511 0.000285119443082563]
MF7='in12cluster7':'gaussmf',[0.00397097261582544 0.000396707083667242]
MF8='in12cluster8':'gaussmf',[0.00398179998065765 0.00849152556357554]

[Input13]
Name='in13'
Range=[0.0002 0.0592]
NumMFs=8
MF1='in13cluster1':'gaussmf',[0.0108684508337383 0.00126140042095949]
MF2='in13cluster2':'gaussmf',[0.010870902974869 0.00126233616513022]
MF3='in13cluster3':'gaussmf',[0.0116305743579675 0.00316351153430631]
MF4='in13cluster4':'gaussmf',[0.0108468687521291 0.000191464708564138]
MF5='in13cluster5':'gaussmf',[0.0108488403333567 0.0012516185013373]
MF6='in13cluster6':'gaussmf',[0.0108492287135814 0.000596009458114832]
MF7='in13cluster7':'gaussmf',[0.0108511139108379 0.000269921606950057]
MF8='in13cluster8':'gaussmf',[0.0110647815223455 0.0082766796912394]

[Input14]
Name='in14'
Range=[0.0019 0.1719]
270 | P a g e
NumMFs=8
MF1='in14cluster1':'gaussmf',[0.0313895504555309 0.0128126493856118]
MF2='in14cluster2':'gaussmf',[0.0312690910203329 0.0127399065893526]
MF3='in14cluster3':'gaussmf',[0.033945335994664 0.00369439822954486]
MF4='in14cluster4':'gaussmf',[0.0312530103063453 0.00249646567033026]
MF5='in14cluster5':'gaussmf',[0.0312385040941791 0.0127593524892205]
MF6='in14cluster6':'gaussmf',[0.0312515251352922 0.00199422853454202]
MF7='in14cluster7':'gaussmf',[0.0312739158522125 0.00256866844036951]
MF8='in14cluster8':'gaussmf',[0.0307937319746233 0.0224414354898865]

[Input15]
Name='in15'
Range=[102 225]
NumMFs=8
MF1='in15cluster1':'gaussmf',[22.6132755686906 224.999999792627]
MF2='in15cluster2':'gaussmf',[22.6132753096855 224.999999820843]
MF3='in15cluster3':'gaussmf',[22.6132738389611 137.000002857337]
MF4='in15cluster4':'gaussmf',[22.6132747363538 225.000000097399]
MF5='in15cluster5':'gaussmf',[22.6132743305156 225.000000254138]
MF6='in15cluster6':'gaussmf',[22.6132753260657 195.999999803551]
MF7='in15cluster7':'gaussmf',[22.6132753149952 224.999999724766]
MF8='in15cluster8':'gaussmf',[22.613271651713 122.999996497212]

[Input16]
Name='in16'
Range=[0.046716897 1.410136517]
NumMFs=8
MF1='in16cluster1':'gaussmf',[0.250719711771256 0.175492332239836]
MF2='in16cluster2':'gaussmf',[0.250661991137857 0.175478576150735]
MF3='in16cluster3':'gaussmf',[0.25091461673315 0.245266327423517]
MF4='in16cluster4':'gaussmf',[0.250662064038314 0.0499995769632495]
MF5='in16cluster5':'gaussmf',[0.250657890609538 0.17548566069849]
MF6='in16cluster6':'gaussmf',[0.250705369797355 0.0467291163640506]
MF7='in16cluster7':'gaussmf',[0.250671096084732 0.0500141702177388]
MF8='in16cluster8':'gaussmf',[0.250615492824929 0.757005180901017]

[Output1]
Name='out1'
Range=[0.553 2.1627]
NumMFs=8
MF1='out1cluster1':'linear',[-0.00755946821024183 8.86822083638194e-05 -
0.412276537299596 0.745026305035173 9.56344806960528 -2.89047524642926 -
7.5797641664173 2.94441542464119 0.679505589228421 1.86942166050559
0.560716546924377 0.855813168079827 0.429499405754188 -0.815630074279709 -
0.00965304624602736 1.60498727899211 6.88143509795696]
MF2='out1cluster2':'linear',[-0.00460409002885835 0.000108459087557863 -
0.0062375138460695 3.74855041077772 2.54470760261085 2.15879531146028 -
16.3324009260478 0.376883823395303 0.015018869404537 1.41208249305524
0.462287501267406 0.906777719212053 0.882684655316704 2.83452350392611 -
0.00428859761793081 9.55855338304278 -0.708738887464017]
MF3='out1cluster3':'linear',[0.000654969920065858 3.15514989381458e-06 -
1.38187325019845 -0.484591779456197 -4.59980740406345 -0.961475341734974
7.79296678187562 -11.8652879601992 -3.10545212298483 -24.2770009386534 -
35.412087739472 22.2766624459308 -34.3371337101806 7.24961947123729 -
0.00383584950248734 -0.388429258351758 2.48328480288172]
MF4='out1cluster4':'linear',[-0.00184849889883234 9.49185556072603e-05 -
0.00014095174767368 2.25058904103199 -16.1831749939464 5.51839161925737
11.4546457747793 -4.77166490765641 -1.50880268005852 -2.32377763088734 -
0.909164647778335 -0.927806311399684 -0.225828088770019 4.84780947280657 -
0.00150769455910119 -1.389192049236 -4.30002498392912]

271 | P a g e
MF5='out1cluster5':'linear',[-0.00489848807731454 1.67091758006954e-05 -
0.0887295942773967 1.68487226515761 -12.6858383339192 7.78537683563561
5.0887221862383 -3.95034999014721 -1.21509477194293 -1.75856087026495 -
0.682572244937583 -0.714591222236499 -0.152713713423265 3.62373006285843 -
0.00246435574546424 3.20612915974109 -4.22263078753113]
MF6='out1cluster6':'linear',[-0.00180609861597681 9.52588235927679e-05 -
0.00204264406659315 2.84360244172998 -14.8735594657765 8.73949213298495
3.88166839803435 -4.46950500956205 -1.40535689468844 -1.95793684868143 -
0.813050133643518 -0.846995138006763 -0.361348044746124 3.05368444917823 -
0.00211678625950648 3.62802080920173 -7.50951241023598]
MF7='out1cluster7':'linear',[-0.00219204652439519 5.10001229244977e-05 -
0.01773153949996 -1.28015744630374 7.72617783348421 -2.11658645471097 -
4.42934013531728 2.19127473716969 0.709237954037057 1.0509957945385
0.424597447499027 0.44089453206662 0.128646474136584 -2.14285603725373 -
3.37436828469696e-05 0.697452124271445 3.43216202064018]
MF8='out1cluster8':'linear',[0.00022197675241711 5.66551503828279e-05 -
1.41094164824755 4.63301335859171 -0.102152414135072 -0.0208196609353636 -
0.244315490381672 -0.0203120179125655 0.740931989631539 -2.0461159783671 -
0.242012620863792 -15.5766598399363 4.08841008676089 0.885455014808267
0.00088095922505857 -0.330946185837141 0.872730470001803]

[Rules]
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1, 1 (1) : 1
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2, 2 (1) : 1
3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3, 3 (1) : 1
4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4, 4 (1) : 1
5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5, 5 (1) : 1
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6, 6 (1) : 1
7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7, 7 (1) : 1
8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8, 8 (1) : 1

272 | P a g e
 APPENDIX C: Sample calculation of inflow performance
relationship (IPR) for HTHP gas condensate well

In this appendix the sample calculation of PVT properties for KAL – 05 is provided.

Please refer to the published paper for obtaining sample calculation of the IPR.
https://www.sciencedirect.com/science/article/abs/pii/S0920410519305819

273 | P a g e
 APPENDIX D: The developed Eclipse – 300 code for
prediction of well inflow performance of HPHT gas
condensate well

The details of developed Eclipse 300 code is given in following.

===========================================================================
-- Study : Gas-condensate HTHP/tight formation (KAL – 5)
-- Author : F. Faraji
-- Simulator : Eclipse-300
-- DATE : 24 March 2020
===========================================================================

--RUNSPEC section--------------------------------------------------

RUNSPEC

--Request the FIELD unit set

FIELD

--Water is present

WATER

--AIM solution method

AIM

--Nine components in study (plus water )

COMPS
9/

--Peng-Robinson equation of state to be used

EOS
PR /

DIMENS
994/

--ICP NTABSS=2
TABDIMS
2 1 40 40 /

--Is a gas condensate study

ISGAS

MULTSAVE
0/

--ICP Debug
DEBUG3
81* 1 /

274 | P a g e
--ICP
SCFDIMS
4/

FMTOUT

--Grid section--------------------------------------------------------

GRID

INIT

--Basic grid block sizes

DX
324*293.3 /

DY
324*293.3 /

DZ
162*30 162*50 /

--Cell top depths - only for first layer specified

TOPS
81*7400 /

PORO
324*0.13 /

PERMX
81*130 81*40 81*20 81*150 /

PERMY
81*130 81*40 81*20 81*150 /

PERMZ
81*13 81*4 81*2 81*15 /

CONDFRAC
'SCF3' 2 10.0 500.0 /
6 6 4 5 1 4 'X' /
6 7 5 5 1 4 'Y' /
7 7 5 6 1 4 'X' /
7 8 6 6 1 4 'Y' /
8 8 6 8 1 4 'X' /
8 9 8 8 1 4 'Y' /
9 9 8 9 1 4 'X' / <-- Intersects P1
/

--ICP
CONDFRAC
'SCF1' 2 10.0 500.0 /
4 7 2 2 1 4 'Y' /
/

CONDFRAC
'SCF2' 2 10.0 500.0 /
2 2 2 5 1 4 'X' /
/
275 | P a g e
--Properties section-----------------------------------------------

PROPS

NCOMPS
9/

EOS
PR /

--ICP
RPTPROPS
SWFN SGFN SOF3 /

-- Peng-Robinson correction

PRCORR

-- Standard temperature and pressure in Deg F and PSIA

STCOND
60.0 14.7 /

-- Component names

CNAMES
CO2 N2 C1 C2 C3 C4-6 C7+1 C7+2 C7+3 /

-- Critical temperatures Deg R

TCRIT
548.46000 227.16000 343.08000 549.77400 665.64000
806.54054 838.11282 1058.03863 1291.89071 /

-- Critical pressures PSIA

PCRIT
1071.33111 492.31265 667.78170 708.34238 618.69739
514.92549 410.74956 247.56341 160.41589 /

-- Critical Z-factors

ZCRIT
.27408 .29115 .28473 .28463 .27748
.27640 .26120 .22706 .20137 /

-- Acentric factors

ACF
.22500 .04000 .01300 .09860 .15240
.21575 .31230 .55670 .91692 /

-- Molecular Weights

MW
44.01000 28.01300 16.04300 30.07000 44.09700
66.86942 107.77943 198.56203 335.19790 /

276 | P a g e
-- Omega_A values

OMEGAA
.4572355 .4572355 .5340210 .4572355 .4572355
.4572355 .6373344 .6373344 .6373344 /

-- Omega_B values

OMEGAB
.0777961 .0777961 .0777961 .0777961 .0777961
.0777961 .0872878 .0872878 .0872878 /

-- Default fluid sample composition

ZMFVD
1.00000 .01210 .01940 .65990 .08690
.05910 .09670 .04745 .01515 .00330
10000.00000 .01210 .01940 .65990 .08690
.05910 .09670 .04745 .01515 .00330 /

-- Boiling point temperatures Deg R

TBOIL
350.46000 139.32000 201.06000 332.10000 415.98000
523.33222 689.67140 958.31604 1270.40061 /

-- Reference temperatures Deg R

TREF
527.40000 140.58000 201.06000 329.40000 415.80000
526.05233 519.67000 519.67000 519.67000 /

-- Reference densities LB/FT3

DREF
48.50653 50.19209 26.53189 34.21053 36.33308
37.87047 45.60035 50.88507 55.89861 /

-- Parachors (Dynes/cm)

PARACHOR
78.00000 41.00000 77.00000 108.00000 150.30000
213.52089 331.78241 516.45301 853.48860 /

-- Binary Interaction Coefficients

BIC
-.0200
.1000 .0360
.1300 .0500 .000000
.1350 .0800 .000000 .000
.1277 .1002 .092810 .000 .000
.1000 .1000 .130663 .006 .006 .0
.1000 .1000 .130663 .006 .006 .0 .0
.1000 .1000 .130663 .006 .006 .0 .0 .0 /

-- Reservoir temperature in Deg F

RTEMP
200.0 /
277 | P a g e
--Water saturation functions

SWFN
0.16 0 50
0.18 0 40
0.20 0.002 32
0.24 0.010 21
0.28 0.020 15.5
0.32 0.033 12.0
0.36 0.049 9.2
0.40 0.066 7.0
0.44 0.090 5.3
0.48 0.119 4.2
0.52 0.150 3.4
0.56 0.186 2.7
0.60 0.227 2.1
0.64 0.277 1.7
0.68 0.330 1.3
0.72 0.390 1.0
0.76 0.462 0.7
0.8 0.540 0.5
0.84 0.620 0.4
0.88 0.710 0.3
0.92 0.800 0.2
0.96 0.900 0.1
1.00 1.000 0.0
/
0.2 0 0
1.0 1 0
/

--Gas saturation functions

SGFN
0.00 0.000 0.0
0.04 0.005 0.0
0.08 0.013 0.0
0.12 0.026 0.0
0.16 0.040 0.0
0.20 0.058 0.0
0.24 0.078 0.0
0.28 0.100 0.0
0.32 0.126 0.0
0.36 0.156 0.0
0.40 0.187 0.0
0.44 0.222 0.0
0.48 0.260 0.0
0.52 0.300 0.0
0.56 0.349 0.0
0.60 0.400 0.0
0.64 0.450 0.0
0.68 0.505 0.0
0.72 0.562 0.0
0.76 0.620 0.0
0.80 0.680 0.0
0.84 0.740 0.0
/
/

278 | P a g e
--Oil saturation functions

SOF3
0.00 0.000 0.000
0.04 0.000 0.000
0.08 0.000 0.000
0.12 0.000 0.001
0.16 0.000 0.002
0.20 0.000 0.003
0.24 0.000 0.004
0.28 0.005 0.005
0.32 0.012 0.012
0.36 0.024 0.024
0.40 0.040 0.040
0.44 0.060 0.060
0.48 0.082 0.082
0.52 0.112 0.112
0.56 0.150 0.150
0.60 0.196 0.196
0.68 0.315 0.315
0.72 0.400 0.400
0.76 0.513 0.513
0.80 0.650 0.650
0.84 0.800 0.800
/
/

--Rock and water pressure data

ROCK
3550 0.000004 /

PVTW
3550 1.0 0.000003 0.31 0.0 /

--Surface density of water

DENSITY
1* 63.0 1* /

--Solution section------------------------------------------------------

SOLUTION

--Equilibration data - initial pressure 3500 psi at 7500, which is

--the oil-water and the oil-gas contact depth

EQUIL
--Dep Pref Dow Pcow Dgo Pcog
7500 3550 7500 0 5000 0 / -- 1 1 0 /

RPTRST
BASIC=2 PRESSURE SOIL SWAT SGAS /

SUMMARY =============================================================

ALL

RUNSUM
279 | P a g e
FOPR
FOPT
FGOR
FPR

--Schedule section------------------------------------------------------

SCHEDULE

WELSPECS
I FIELD 1 1 7500 WATER /
P FIELD 9 9 7500 OIL /
/

COMPDAT
I 1 1 3 4 1* 1 /
P 9 9 1 2 1* 1 /
/

--ICP
-- Must come after COMPDAT
-- PI mult only used if NOCUT
-- <Well name> <CUT/NOCUT> <PI_Mult> <I> <J> <K> <C1> <C2>
WELLCF
P CUT 1.0 5* /
/

WCONPROD
P OPEN ORAT 2000 1* 1* 2* 500 /
/

WCONINJE
I WATER OPEN RATE 2000 /
/

TSTEP
36*10 /

END

280 | P a g e