You are on page 1of 210

Numerical Simulation and Optimization of Fracturing-Stimulated Reservoir Volume

by

Hao Zhang, M.S.

A Dissertation

In

Petroleum Engineering

Submitted to the Graduate Faculty


of Texas Tech University in
Partial Fulfillment of
the Requirements for
the Degree of

DOCTOR OF PHILOSOPHY

Approved

James J. Sheng
Chair of Committee

Hossein Emadibaladehi

Lloyd Heinze

Qingwang Yuan

Mark Sheridan
Dean of the Graduate School

August 2021
Copyright 2021, Hao Zhang
Texas Tech University, Hao Zhang, August 2021

ACKNOWLEDGMENTS
First and foremost, I would like to express my sincerest gratitude to my advisor
Dr. James J. Sheng. His guidance and support have been immensely instrumental for
my doctorate program. He has spent a lot of time providing valuable insight into my
research topic and helping me to further develop my research skills. He guided me to
the right track of research, suggesting valuable and critical ideas. His ingenious
perspectives into the research as well as optimistic and friendly characters have set up
an example that I wish to follow in my future career. I have been very lucky to have a
mentor who is so deeply invested in his students.

I would like to extend my appreciation to my committee members: Dr. Hossein


Emadibaladehi, Dr. Lloyd Heinze, and Dr. Qingwang Yuan, for spending their time
and effort reviewing my dissertation and providing valuable comments, as well as
serving on my defense committee.

This work would not have been complete without the help of my colleagues and
friends, Xiukun Wang, David Tu, Nur Wijaya, Kerui Liu, Junzhe Jiang, Xinyue
Liang, Lingfeng Yang, and Zhe Liu. I appreciate all the help that you have provided
me.

Any amount of gratitude would be never enough for my family, my father Jijun
Zhang, my mother Shurong Ming, and my grandmother Shufen Xu. My family has
always been a source of motivation for me to pursuit my long educational career and
puts up with me this far.

Finally, I would like to express my sincere acknowledgment to Mr. and Mrs. Ed


Whitacre, who generously offered me the Presidential Graduate Fellowship to support
me in completing my doctorate program, to the Bob L. Herd Department of Petroleum
Engineering and the Graduate School for all the service and support they provided
towards my doctorate program. I heartfully thank Texas Tech University for being my
home for three years and giving me a great memory of studying and researching in the
United States.

ii
Texas Tech University, Hao Zhang, August 2021

TABLE OF CONTENTS

ACKNOWLEDGMENTS ........................................................................................... ii

ABSTRACT ................................................................................................................ vii

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

LIST OF FIGURES ...................................................................................................... x

I. INTRODUCTION .................................................................................................... 1

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


1.2 Problem Statement ............................................................................................ 5
1.3 Objective of the Study....................................................................................... 7
1.4 Organization of the Dissertation ....................................................................... 8

II. LITERATURE REVIEW .................................................................................... 10

2.1 Hydraulic fracturing propagation simulation method ..................................... 10


2.1.1 Analytical model ................................................................................... 10
2.1.2 Discrete element method....................................................................... 12
2.1.3 Extended finite element method ........................................................... 13
2.1.4 Displacement discontinuity method ...................................................... 14
2.2 Fractured reservoir production simulation method ......................................... 16
2.2.1 Dual-porosity / dual-permeability model .............................................. 16
2.2.2 Local grid refinement model ................................................................. 17
2.2.3 Discrete Fracture Model ....................................................................... 18
2.2.4 Embedded Discrete Fracture Model ..................................................... 19
2.3 Stimulated reservoir volume estimation method ............................................ 21
2.3.1 Field data monitoring method ............................................................... 22
2.3.2 Matrix permeability increase method ................................................... 23
2.3.3 Complex fracture network method ....................................................... 24
2.4 Hydraulically fractured well optimization method ......................................... 25
2.4.1 Single objective optimization ............................................................... 26
2.4.2 Multi-objective optimization ................................................................ 27

III. STIMULATED RESERVOIR VOLUME ESTIMATION AND


HYDRAULIC FRACTURING OPTIMIZATION ........................................... 30

3.1 Methodologies for SRV estimation................................................................. 30

iii
Texas Tech University, Hao Zhang, August 2021

3.1.1 Induced stress model ............................................................................. 30


3.1.2 The SRV model..................................................................................... 34
3.1.3 The natural fracture distribution model ................................................ 35
3.1.4 Stress dependent permeability .............................................................. 37
3.1.5 Workflow of SRV estimation ............................................................... 38
3.1.6 A case for SRV estimation .................................................................... 40
3.2 Methodologies for SRV optimization ............................................................. 44
3.2.1 Optimization objective function ........................................................... 44
3.2.2 Optimization algorithm ......................................................................... 45
3.2.3 Modified-PSO algorithm ...................................................................... 46
3.2.4 The modified-PSO algorithm validation ............................................... 48
3.2.5 The workflow of hydraulic fracturing well optimization ..................... 51
3.3 Examples ......................................................................................................... 53
3.3.1 Case 1: Single hydraulically fractured well optimization based
on the effect of SRV ............................................................................. 53
3.3.2 Case 2: Multiple hydraulically fractured well optimization based
on the effect of SRV ............................................................................. 59

IV. COMPLEX FRACTURE NETWORK SIMULATION AND


OPTIMIZATION ................................................................................................. 66

4.1 Methodologies for complex fracture network propagation simulation ........... 67


4.1.1 Hydraulic fracture dynamic propagating model ................................... 67
4.1.2 Hydraulic fracture interaction with natural fracture ............................. 69
4.1.3 Workflow of complex fracture network propagation simulation ......... 71
4.1.4 A case for hydraulic fracture propagation and natural fracture
encountering simulation ....................................................................... 74
4.2 Complex fracture networks production simulation ......................................... 77
4.2.1 The embedded discrete fracture model ................................................. 77
4.2.2 The principle of embedded discrete fracture model ............................. 78
4.2.3 Shale gas flow model ............................................................................ 80
4.3 The complex fracture network optimization ................................................... 82
4.3.1 Optimization objective function ........................................................... 82
4.3.2 The neural network algorithm ............................................................... 82
4.3.3 The modified neural network algorithm ............................................... 85
4.3.4 The workflow of complex fracture network optimization .................... 88
4.4 Examples ......................................................................................................... 89
4.4.1 Case 1: Complex fracture network propagation simulation and
sensitivity analysis ................................................................................ 89
4.4.2 Case 2: Complex fracture network productivity analysis ..................... 94
4.4.3 Case 3: Complex fracture network optimization under M-NNA.......... 99
iv
Texas Tech University, Hao Zhang, August 2021

V. UNSTRUCTURED QUAD-GRID BASED EMBEDDED DISCRETE


FRACTURE MODEL ....................................................................................... 106

5.1 The drawback of the structured grid in EDFM ............................................. 107


5.2 The unstructured quadrangular grid .............................................................. 109
5.2.1 The principle of frontal Delaunay quadrangular grid ......................... 109
5.2.2 Local grid refinement for unstructured quadrangular grid ................. 114
5.3 The optimization method .............................................................................. 115
5.3.1 Optimization objective function ......................................................... 115
5.3.2 The workflow of optimization ............................................................ 116
5.4 Examples ....................................................................................................... 117
5.4.1 Case 1: Verification of the unstructured quad-grid based EDFM ...... 118
5.4.2 Case 2: Hydraulic fractured well optimization based on the
unstructured quad-grid based EDFM. ................................................ 133

VI. SURROGATE-ASSISTED MULTI-OBJECTIVE HYDRAULIC


FRACTURE OPTIMIZATION CONSIDERING UNCERTAINTY ........... 141

6.1 The multi-objective optimization algorithm ................................................. 142


6.1.1 The multi-objective optimization problem and Pareto front............... 142
6.1.2 The hybrid multi-objective PSO algorithm......................................... 143
6.2 The Kriging surrogate model ........................................................................ 147
6.2.1 The principle of Kriging model .......................................................... 147
6.2.2 A case for Kriging surrogate model validation ................................... 148
6.3 The surrogate-assisted multi-objective hydraulically fractured well
optimization considering uncertainty ........................................................... 151
6.3.1 The reservoir geological uncertainty description ................................ 151
6.3.2 The multi-objective functions for hydraulically fractured well
optimization ........................................................................................ 153
6.3.3 Workflow of surrogate-assisted multi-objective hydraulically
fractured well optimization................................................................. 154
6.4 Examples ....................................................................................................... 156
6.4.1 Case 1: Validation of the surrogate-assisted HMOPSO algorithm ..... 156
6.4.2 Case 2: Multi-objective hydraulically fractured well optimization
in the uncertain reservoir using the surrogate-assisted HMOPSO ..... 164

VII.CONCLUSION AND DISCUSSION ............................................................ 174

7.1 Summary ....................................................................................................... 174

v
Texas Tech University, Hao Zhang, August 2021

7.1.1 Stimulated reservoir volume estimation and hydraulic fracturing


optimization ........................................................................................ 174
7.1.2 Complex fracture network simulation and optimization .................... 175
7.1.3 Unstructured quad-grid based embedded discrete fracture model ...... 176
7.1.4 Surrogate-assisted multi-objective hydraulic fracture
optimization considering uncertainty ................................................. 177
7.2 Future works ................................................................................................. 178

BIBLIOGRAPHY ..................................................................................................... 180

vi
Texas Tech University, Hao Zhang, August 2021

ABSTRACT
The shale reservoir is making an ever-increasing contribution to today’s
petroleum industry, and the hydraulically fractured horizontal well is one of the most
effective development strategies that enable the economic viability of shale oil and gas
resources. Besides, dozens of evidence demonstrate that the complex fracture network
(CFN) can be generated during hydraulic fracturing in shale reservoirs, which is usually
approximated as a 3D volume and known as the stimulated reservoir volume (SRV).
Since the tight formation’s productivity is significantly stimulated within the SRV
region, the economic success of shale reservoirs’ development highly depends on the
optimal creation of fracturing-stimulated reservoir volume, yet the mechanism of which
has not been fully studied.

Facing the urgent requirement, this dissertation makes a comprehensive


numerical investigation on the hydraulically fractured well construction and production
optimization in the fractured shale reservoir considering the effect of the SRV. First, a
novel SRV numerical model is established using the displacement discontinuity method
(DDM). By the coupling mechanism of the geo-stress and formation pressure change
caused by the hydraulic fracture (HF) on the natural fracture (NF), the scale of SRV is
estimated, and the reservoir stimulation within the SRV is presented by evaluating the
permeability increase. Additionally, by simulating the HF propagating and interacting
with the NF, a CFN propagation model is constructed, which is more realistic to reflect
the geometry character of SRV. The sensitivity analysis of the effects of different
geological and fracturing operational parameters on the configuration of CFN is
investigated. Furthermore, to effectively and accurately simulate the production of the
shale reservoir with complicated distributed fracture networks and irregular boundary
conditions, an unstructured quad grid-based embedded discrete fracture model (UnQ-
EDFM) is proposed, which adopts a frontal Delaunay quad-grid to subdivide the
reservoir domain and overcomes several drawbacks of the current structured grid-based
EDFM. Based on the above SRV estimation and production models, the hydraulically
fractured horizontal well optimization framework is constructed, in which the net

vii
Texas Tech University, Hao Zhang, August 2021

present value (NPV) is employed as the objective function to evaluate the economic
viability of the well designing. Several novel optimization algorithms, such as the
modified particle swarm optimization (M-PSO) and the modified neural network
algorithm (M-NNA) are proposed, by which the hydraulic fracturing parameters (frac-
spacing, well spacing, frac-half length, injection rate, etc.) are optimized considering
the effect of SRV to achieve the optimal NPV. Finally, the optimal design of the
hydraulically fractured horizontal well is studied under the reservoir geological
uncertainty, in which the NF distribution is treated as the most essential uncertain
parameter that threatens the economical return. A series of stochastic NF distribution
models are collected to describe the NF uncertainty, and the multi-objective
optimization (MOO) strategy is applied. A surrogate-assisted hybrid multi-objective
particle swarm optimization algorithm (HMOPSO) is developed to increase the
optimization efficiency and accuracy of this computationally expensive problem. A
Pareto front is obtained from this MOO which provides a set of tread-off solutions that
fully grasp the expected economical revenue and the possible development risk, for the
designer.

At the end of each chapter, case studies based on conceptual and real reservoir
models are presented to validate the methodologies proposed. This dissertation proposes
comprehensive simulation and optimization studies of the fracturing-stimulated
reservoir volume and hydraulically fractured horizontal well design, provides new
insight into the development of unconventional reservoirs.

viii
Texas Tech University, Hao Zhang, August 2021

LIST OF TABLES
3.1. Geological parameters for the reservoir ...................................................... 40
3.2. Geological parameters for the reservoir model ........................................... 54
3.3. Parameters of the reservoir model and their values .................................... 55
3.4. Optimization variables and search space .................................................... 56
3.5. Values of optimization variables during optimization (Objective 1) .......... 58
3.6. Values of optimization variables during optimization (Objective 2) .......... 58
3.7. Optimization variables and search space .................................................... 60
3.8. Values of optimization variables during optimization (Type 1) ................. 63
3.9. Values of optimization variables during optimization (Type 2) ................. 64
4.1. Geological and fluid parameters for the reservoir....................................... 74
4.2. Geological parameters for production simulation ....................................... 97
4.3. Properties of shale and gas .......................................................................... 97
4.4. Economic parameters and their values...................................................... 101
4.5. Optimization variables and their search space .......................................... 101
4.6. Optimization variables during optimization (Aligning fracturing
placement) ................................................................................................. 104
4.7. Optimization variables during optimization (Alternating fracturing
placement) ................................................................................................. 105
5.1. Values of key points .................................................................................. 114
5.2. Geological parameters for the reservoir model ......................................... 119
5.3. Properties of shale and gas ........................................................................ 120
5.4. Values of key points .................................................................................. 128
5.5. Optimization variables and search space for single well optimization ..... 134
5.6. Economic parameters and their values...................................................... 134
5.7. The optimum values after the optimization under different grids for
single well optimization ............................................................................ 137
5.8. Optimization variables and search space for multi-well optimization ...... 138
5.9. The optimum values after the optimization under different grids for
multi-well optimization ............................................................................. 140
6.1. Optimization variables and their search space .......................................... 167
6.2. Economic parameters and their values...................................................... 167
6.3. The values of the optimization variables for different points ................... 170
ix
Texas Tech University, Hao Zhang, August 2021

LIST OF FIGURES
1.1. U.S. oil and gas production growth from the EIA (2019)............................. 1
1.2. The U.S. oil and gas production prediction from the EIA (2021)................. 2
1.3. Illustration of a hydraulic fractured well in the shale reservoir
(Zanganeh et al, 2015)................................................................................... 3
1.4. The statistic of U.S. crude oil and natural gas production. ........................... 4
1.5. The sketch for complex fracture network and stimulated reservoir
volume (Mayerhofer et al., 2010) ................................................................. 5
2.1. The schematic view of KGD and PKN models (Wu, 2014) ....................... 11
2.2. Particle contacts and flow domains for the DEM model (De Pater et
al., 2005) ..................................................................................................... 12
2.3. A schematic view of the XFEM model ....................................................... 13
2.4. Fracture discretized by DDM (Jo and Hurt, 2013) ..................................... 15
2.5. Illustration of the DPDP model ................................................................... 16
2.6. A fracture constructed by the LGR method (Eshkalak et al., 2014) ........... 17
2.7. DFM construction by unstructured mesh .................................................... 18
2.8. Comparison of grids between DFM and EDFM (Yan et al.,2016) ............. 20
2.9. A comprehensive comparison between different fracture production
simulation methods ..................................................................................... 20
2.10. The SRV estimation based on the micro-seismic data monitoring
(Ren et al., 2018) ......................................................................................... 22
2.11. Schematic view of SRV and its corresponding stress-dependent
permeability within a simulator model (Sen et al., 2018) ........................... 23
2.12. SRV represented by complex fracture network (Ren et al., 2016) ............. 24
2.13. The Pareto front for multi-objective optimization ...................................... 28
3.1. Discrete fracture units ................................................................................. 31
3.2. Angles between the different coordinate systems ....................................... 33
3.3. Normal and shear failure of NF .................................................................. 34
3.4. The comparison of real NF pattern and the power-law probability
NF pattern ................................................................................................... 36
3.5. Power-law NF distribution and permeability field...................................... 38
3.6. Workflow for SRV estimation .................................................................... 39
3.7. The induced stress and pressure field around the fracture .......................... 41
3.8. The SRV region and the stress-dependent permeability after
correction (mD) ........................................................................................... 41
x
Texas Tech University, Hao Zhang, August 2021

3.9. The SRV variation with the fracture spacing (fracture half-length
=150 ft) ........................................................................................................ 42
3.10. The stress-dependent permeability (mD) variation with different
fracture spacing (frac half-length=150 ft) ................................................... 43
3.11. The SRV volume vs. fracture half-length and fracture spacing .................. 44
3.12. The workflow for Modified-PSO ................................................................ 47
3.13. The sketch of reassigning position operation for the Modified-PSO .......... 47
3.14. The field for optimization searching ........................................................... 48
3.15. The optimization procedure for 50 particles ............................................... 49
3.16. The optimization procedure for 5 particles ................................................. 50
3.17. The optimization accuracy of the original/modified PSO........................... 51
3.18. Schematic diagram of the SRV optimization procedure ............................. 52
3.19. The concept of the MINC model (Pruess, 1983) ........................................ 53
3.20. The sketch for single horizontal fracturing well optimization .................... 54
3.21. Geology parameters of the reservoir model ................................................ 55
3.22. Initial hydraulically fractured well as well as the production result ........... 56
3.23. Optimal hydraulically fractured well as well as the production result
for objective 1: maximizing SRV volume .................................................. 57
3.24. Optimal hydraulically fractured well as well as the production result
for objective 2: maximizing NPV ............................................................... 57
3.25. The accumulative shale gas production ...................................................... 58
3.26. SRV and NPV comparisons for the two optimization objectives ............... 59
3.27. Two types of multiple horizontal well placement ....................................... 60
3.28. The initial hydraulically fractured well and the corresponding
production result (Type 1: aligning hydraulic fractures) ............................ 61
3.29. The optimal hydraulically fractured well and the corresponding
production result (Type 1: aligning hydraulic fractures) ............................ 62
3.30. The initial hydraulically fractured well and the corresponding
production result (Type 2: alternate hydraulic fractures) ............................ 62
3.31. The optimal hydraulically fractured well and the corresponding
production result (Type 2: alternate hydraulic fractures) ............................ 63
3.32. The accumulative shale gas production ...................................................... 64
3.33. SRV and NPV variations during optimization ............................................ 65
4.1. The fluid flow balance for the fracture unit ................................................ 67
4.2. Illustration of fracture propagation direction .............................................. 68

xi
Texas Tech University, Hao Zhang, August 2021

4.3. Schematic of fracture approaching interface (Gu and Weng, 2010) ........... 69
4.4. Fracture crossing conditions with various intersection angles (Gu
and Weng 2010) .......................................................................................... 71
4.5. Schematic of fracture dynamic propagation ............................................... 71
4.6. Workflow for fracture network propagation ............................................... 73
4.7. Two types of fracturing placement ............................................................. 75
4.8. Hydraulic fracture propagation path and the corresponding induced
stress map (psi)............................................................................................ 75
4.9. The propagating HF and interaction with different oriented NF and
the induced stress map (psi) ........................................................................ 76
4.10. Local grid refinement (LGR) method ......................................................... 77
4.11. Comparison of DFM and EDFM models .................................................... 78
4.12. Schematic diagram of the EDFM (Xu, 2015) ............................................. 79
4.13. Illustration of NNA (Sadollah et al., 2018) ................................................. 83
4.14. Distribution of the 100 optimization results for different algorithms ......... 87
4.15. Flow chart of the CFN optimization ........................................................... 89
4.16. The fracture network propagation with different NF lengths ..................... 90
4.17. The fracture network scale variation with NF length.................................. 91
4.18. The fracture network propagation with different NF angle ........................ 91
4.19. The fracture network scale variation with NF angles ................................. 92
4.20. The fracture network propagation with different HF spacing ..................... 92
4.21. The fracture network scale variation with HF spacing ............................... 93
4.22. The fracture network propagation with different NF density ..................... 93
4.23. The fracture network scale variation with NF density ................................ 94
4.24. A sketch for complex fracture network propagation simulation ................. 95
4.25. The CFN and NF pattern converts to EDFM .............................................. 96
4.26. Correction factors for adsorption and transport mechanisms ..................... 96
4.27. The pressure field (psi) after 10 years of production for the EDFM .......... 98
4.28. The cumulative shale gas production comparison ...................................... 98
4.29. Two types of hydraulic fracturing placement ............................................. 99
4.30. The NF pattern within the reservoir domain ............................................. 100
4.31. The initial and optimal CFN for aligning fracturing placement ............... 102
4.32. The initial and optimal CFN for alternating fracturing placement ........... 102

xii
Texas Tech University, Hao Zhang, August 2021

4.33. The pressure fields (psi) after 10 years of production under initial
and optimal CFNs for aligning fracturing placement ............................... 103
4.34. The pressure fields (psi) after 10 years of production under initial
and optimal CFNs for alternating fracturing placement ........................... 103
4.35. The accumulative gas production under initial and optimal CFNs ........... 104
4.36. The objective function variation with the optimization iterations ............ 104
5.1. Structured grids generated in the reservoir with irregular boundary ........ 108
5.2. The comparison of the structured grid and the triangular grid in the
same irregular reservoir............................................................................. 109
5.3. Comparison of different triangular grids (Remacle et al., 2013) .............. 110
5.4. The grids and the corresponding cross fields (Remacle et al., 2013) ...... 111
5.5. Example of Frontal Delaunay Quad Generation ....................................... 112
5.6. The comparison of the structured grid and the triangular grid in the
same irregular reservoir............................................................................. 113
5.7. Unstructured quad grid with local grid refinement ................................... 115
5.8. Flowchart of M-NNA optimization procedure ......................................... 117
5.9. The reservoir model with fractured well in different simulators .............. 118
5.10. Correction factors for adsorption and transport mechanisms ................... 119
5.11. Reservoir pressure (MPa) after 10 years of production for models in
different simulators ................................................................................... 120
5.12. The cumulative gas production for different models ................................ 121
5.13. The horizontal view of the reservoir ......................................................... 121
5.14. The fracture map of the reservoir .............................................................. 122
5.15. Reservoir models with fractures and different sized structured grids ....... 123
5.16. The reservoir pressure (MPa) after 20 years of production under
different grid lengths ................................................................................. 124
5.17. The comparison of simulation results under different grids ..................... 124
5.18. The statistics of grid number and simulation time for different grids....... 125
5.19. Grid maps and the corresponding reservoir pressure (MPa) after 20
years of production for different sized unstructured quad grids ............... 126
5.20. Cumulative gas production under different grid types and sizes .............. 128
5.21. The reservoir models with varying-sized unstructured quad grid ............. 129
5.22. The reservoir models with different-sized triangular grids ....................... 129
5.23. Reservoir pressure (MPa) after 20 years of production for different
grid-based reservoirs ................................................................................. 130

xiii
Texas Tech University, Hao Zhang, August 2021

5.24. The cumulative gas production under different grids ............................... 131
5.25. The statistics of grid number and simulation time for different grids....... 132
5.26. Heterogeneous permeability field with different types of grids (nD) ....... 133
5.27. The optimum well placement and the corresponding pressure field
(MPa) for single well optimization ........................................................... 135
5.28. The comparison of optimization results under different grids for the
single well optimization ............................................................................ 136
5.29. Comparison of optimization times under different grids for single
well optimization....................................................................................... 136
5.30. The optimum well placement and the corresponding pressure field
(MPa) for multi-well optimization ............................................................ 139
5.31. The comparison of optimization results under different grids for
multi-well optimization ............................................................................. 139
5.32. The comparison of optimization time for single and multi-well
optimization under different grids ............................................................. 139
6.1. Dominated, non-dominated, and Pareto front (Mahesh et al., 2016) ........ 143
6.2. Gaussian distribution reinitialization (Zapotecas and Coello, 2011) ........ 145
6.3. Image of the benchmark function ............................................................. 148
6.4. Image of the approximation function with different sample points .......... 149
6.5. Comparison of the approximate and real function values......................... 150
6.6. Comparison of real NF map and power-law probability NF map ............ 151
6.7. Stochastic NF models generated using the same parameter ..................... 152
6.8. The illustration of NPV distribution ......................................................... 153
6.9. Flowchart of the surrogate-assisted multi-objective optimization ............ 156
6.10. The optimization results for different benchmark functions with
different MOO algorithms......................................................................... 159
6.11. Optimization results for different MOO algorithms in ZDT1 with
different particle population ...................................................................... 162
6.12. The results of surrogate-assisted HMOPSO based on ZDT1 ................... 163
6.13. The Brugge reservoir model...................................................................... 165
6.14. The stochastic NF distribution maps ......................................................... 166
6.15. The optimization results of different surrogate-assisted algorithms ......... 168
6.16. The analysis of the PF obtained by the surrogate-assisted HMOPSO ...... 169
6.17. The hydraulically fractured wells and their corresponding reservoir
pressure map (MPa) after 20 years of production for different points ...... 171
6.18. The comparison of NPV distributions for different points ....................... 172
xiv
Texas Tech University, Hao Zhang, August 2021

CHAPTER I

INTRODUCTION
In this chapter, the background for this research is introduced. Then, the current
problems and the corresponding objectives of this research are presented. The
organization of the contents of this dissertation is briefed at the end.

1.1 Background

The shales are traditionally regarded as potential source rocks and cap rocks for
hydrocarbon resources due to their potentially high organic content and low
permeability; however, with the depletion of conventional hydrocarbon resources and
the advancement of production technologies, unconventional oil and gas resources are
playing an increasingly important role in today’s petroleum industry. According to the
2019 Annual Energy Outlook of the U.S. Energy Information Administration (EIA),
tight oil and shale gas production has experienced rapid growth and surpassed 50%
share of the U.S. total oil and gas production in recent years. As is more intuitively
shown in Figure 1.1, shale gas has reached 70% of total U.S. natural gas production in
2018, while the tight oil accounted for 60% of the U.S. total crude oil production in
2018, respectively (EIA, 2019).

Figure 1.1 U.S. oil and gas production growth from the EIA (2019)

1
Texas Tech University, Hao Zhang, August 2021

Furthermore, the 2021 Annual Energy Outlook of EIA predicted that the future
development of U.S. crude oil and natural gas production will continue to be driven by
the tight oil and shale gas resources for the next 30 years. As is shown in Figure 1.2, by
2050, the U.S. tight oil production is predicted to reach 9 million stock tank barrels per
day (STB/d) and occupies about 70% of the U.S. total crude oil production; while the
shale gas production is forecasted to above 40 trillion cubic feet and contributes to
nearly 90% of the U.S. total natural gas production (EIA, 2021).

Figure 1.2 The U.S. oil and gas production prediction from the EIA (2021)

Though important, the shale reservoir is hard to be developed, compared with


conventional oil and gas formations, shale formations have some distinct properties.
The extremely low permeability (nano-Darcy permeability (Feiyu et al., 2013)), the
super tight formation (less than 10% porosity (Barton and Zoback, 2002)), and also the
natural fracture distribution (Gale and Holder, 2008; King, 2010), all make the
traditional developing methods not suitable for shale oil and gas development. Based
on decays’ of field application on the shale formation, scholars and engineers have
summarized two key technologies for shale oil and gas development, which are
horizontal drilling (Fisher et al., 2004; Wiley et al., 2004) and hydraulic fracturing (Britt
et al., 2006; Waters et al., 2009). Hydraulically fractured well is one of the most
effective stimulation strategies that enable the economic viability of shale oil and gas.

2
Texas Tech University, Hao Zhang, August 2021

To increase the fracture contact area with the formation for improving recovery in shale
reservoirs, tenth or hundreds of hydraulic fractures are employed in a horizontal
wellbore (Figure 1.3) (Durst et al., 2008). Sequential and simultaneous fracturing
methods are also implemented to improve well performance and enhance operation
efficiency (Mutalik and Gibson, 2008; Roussel and Sharma, 2011).

Figure 1.3 Illustration of a hydraulic fractured well in the shale reservoir (Zanganeh et
al, 2015)

Based on the most recent available data from the EIA, the oil and gas production
from hydraulically fractured wells now makes up the majority of total U.S. crude oil
and natural gas production (Figure 1.4). In 2000, approximately 23,000 hydraulically
fractured wells produced 102,000 STB/d of oil in the U.S., making up less than 2% of
the national total. By 2015, the number of hydraulically fractured wells grew to an
estimated 300,000, and oil production from those wells had grown to more than 4.3
million STB/d, making up about 50% of the total oil output of the U.S. (EIA, 2016a).
Also in 2000, there were approximately 3.6 billion cubic feet per day (Bcf/d) of
marketed gas produced from hydraulically fractured wells in the U.S., making up less
than 7% of the national total. By 2015, the gas production from those wells had grown
to more than 53 Bcf/d, making up about 67% of the total natural gas output of the U.S.
(EIA, 2016b). The hydraulically fractured well has allowed the United States to increase
its oil and gas production faster than at any time in its history.

3
Texas Tech University, Hao Zhang, August 2021

(a) The U.S. crude oil production (EIA, 2016a)

(b) The U.S. natural gas production (EIA, 2016b)


Figure 1.4 The statistic of U.S. crude oil and natural gas production.

In conventional reservoirs and tight gas sands, single-plane fracture half-length


and conductivity are the key drivers for stimulation performance; however, ultra-low
permeability shale reservoirs require a large fracture network to maximize well
performance. In recent years, scholars and engineers have found a new hydraulic
fracturing technique, which names volumetric fracturing (Fisher et al., 2005; Maxwell
et al., 2002). Micro-seismic mapping results indicated that in many shale reservoirs, the
propagating hydraulic fractures (HF) could stimulating and connecting the pre-existing

4
Texas Tech University, Hao Zhang, August 2021

natural fractures (NF) and construct a complex fracture network (CFN) (Figure 1.5-a).
The size of the CFN can be approximated as a 3-D volume, which is named stimulated
reservoir volume or SRV (Figure 1.5-b) (Mayerhofer et al., 2010). The SRV
significantly extends the contact surface between fractures and the shale formation and
increases the permeability and conductivity, the well performance in low-permeability
shale reservoirs will be improved dozens or hundreds of times (Ren et al., 2018). Based
on these advantages, the SRV has become a key factor to evaluate the quality of
hydraulic fracturing operations.

(a) Complex fracture network (b) Stimulated reservoir volume


Figure 1.5 The sketch for complex fracture network and stimulated reservoir volume
(Mayerhofer et al., 2010)

1.2 Problem Statement

According to the above statements, the shale reservoir is becoming an ever-


increasing important component for today’s petroleum industry, and the horizontal well
with the multi-hydraulic fractures has been demonstrated to be a highly effective method
for shale reservoir development. However, the current study of the hydraulically
fractured well in shale reservoir is far from enough, the geometry and productivity of
fracturing stimulated reservoir volume are still challenging to be measured precisely
and optimized concisely by current techniques. Some urgent problems still need to be
solved.

5
Texas Tech University, Hao Zhang, August 2021

The interactions of the propagated hydraulic fractures between the adjacent


fracturing stages and between adjacent fracturing wells are not totally understood.
Besides, SRV caused by hydraulic fracturing can significantly improve production but
has not been fully investigated. (Mayerhofer et al., 2010; Wu and Olson, 2016). Nearly
all the SRV estimation methods are post-fracturing and depend on the inversion of field
data, such as micro-seismic and well logging, which is too expensive or computationally
complex. Very little research considers SRV prediction before fracturing, though it is
more valuable for hydraulic fracturing design. How to estimate the scale and geometry
of SRV created by hydraulic fracturing using a more computationally concise and
economically friendly method still need further investigation.

Since the existence of SRV and its improvement to the production in shale
reservoirs is verified by scholars from a large number of field observations (Maxwell et
al., 2002 Fischer et al., 2004; Mayerhofer et al., 2010), optimally design the hydraulic
fracturing parameters, such as frac-spacing, well spacing, injection rate, etc. based on
the geological condition of shale reservoirs will help to create the best SRV and then
obtain the optimum shale oil & gas recovery. Unfortunately, the previous hydraulically
fractured well design and optimization studies seldom considered the effect of SRV;
instead, they mainly depended on the traditional bi-wing planar fractures. However, in
shale reservoirs with complex natural fracture distribution, the conventional fracture
model might fail to capture some key physical mechanisms of the complex fracture
network within SRV.

Besides, the productivity evaluation of SRV also needs further improvement.


Several numerical methods for the fractured reservoir production simulation have been
developed, including the DPDP (Warren and Root, 1963), the LGR (Heinemann et al.,
1983), the DFM (Hoteit and Firoozabadi, 2008), and the EDFM (Moinfar et al., 2013).
However, these methods all possess their own shortcomings for the complex fractured
reservoir production simulation, a more effective fractured reservoir production
simulation method is still needed.

6
Texas Tech University, Hao Zhang, August 2021

Apart from the hydraulic fracturing numerical model, the traditional algorithms
and strategies for hydraulically fractured well optimization are generally out of the date,
which is time-consuming, with low accuracy, and cannot well adapt to the complex
optimization requirements. Usually, the traditional hydraulically fractured well
optimization is performed based on a certain reservoir geological condition and results
in a unique objective. However, in fractured shale reservoir, the geological parameters,
especially the natural fracture distribution is hard to be precisely described, which brings
about unavoidable uncertainty to the reservoir model and causes a huge risk to the
hydraulically fractured well design and the reservoir development (Jansen et al., 2009;
Isebor and Durlofsky 2014). The traditional optimization algorithms and strategies
cannot effectively deal with such uncertain and computationally expensive problems
and bring about a reasonable solution.

All the above problems objectively exist and threaten the future development of
shale reservoirs, thus a targeted study is urgently needed.

1.3 Objective of the Study

Based on the above-mentioned problems, the primary objective of this


dissertation is to construct a computationally efficient and accurate numerical model for
hydraulically-stimulated reservoir volume propagation and production simulation and
use these models to realize the hydraulically fractured well optimal design in fractured
shale reservoir under different kinds of complex geological conditions. The specific
objective of this research mainly consists of the following topics:

i. Develop a coupled complex fracture network simulation model, which


incorporates most of the important physical mechanisms that control the
hydraulic fracture propagation in a shale reservoir, including the geo-stress
in the formation, fluid distribution among multiple fractures, and interaction
with natural fractures. Then based on this model, precisely estimate the SRV
with computationally less demanding.

7
Texas Tech University, Hao Zhang, August 2021

ii. Propose a novel algorithm for hydraulically fractured well optimization,


which presents better performance than the traditional algorithms and is
more accurate and efficient in dealing with complex and computationally
expensive optimization problems.

iii. Improve the current fractured reservoir numerical simulator to make it adapt
to the geological complexity of fractured shale reservoir, and achieve the
SRV production simulation more precisely and efficiently.

iv. Construct a robust optimization framework based on the above SRV


numerical models and novel optimization algorithm, then realize the SRV-
based hydraulically fractured well optimum design in fractured shale
reservoirs, fully considering the complex and uncertain geological
conditions, to obtain the best recovery with the least development risk.

1.4 Organization of the Dissertation

According to the above research objectives, this dissertation is divided into


seven chapters.

Chapter I introduces the background information of this dissertation topic and


explains the research motivation and the objectives.

Chapter II gives a literature review on the concept of hydraulic fracture


propagation simulation, the fractured shale reservoir production simulation, the SRV
estimation, and the hydraulically fractured well optimization. The updated state of the
art is summarized.

Chapter III presents the methodology, workflow, and cases for hydraulic
fracture induced stress-based SRV estimation and optimization. First, the numerical
models of hydraulic fracture induced stress simulation and SRV estimation are
illustrated; then an SRV-based hydraulically fractured well optimization workflow is
constructed, in which a novel M-PSO algorithm is proposed and employed as the

8
Texas Tech University, Hao Zhang, August 2021

optimization algorithm. Reservoir cases of multiple hydraulically fractured wells


optimization based on the above models and workflow are presented for validation.

Chapter IV presents the methodology, workflow, and cases for complex fracture
network (CFN) construction and optimization. First, a DDM-based hydraulic fracture
propagation model is coupled with the natural fracture model to simulate the CFN; then,
a novel M-NNA algorithm is proposed to construct the workflow for CFN optimization.
Additionally, the CFN sensitivity analysis and the CFN optimization under multiple
hydraulic fractures are performed based on reservoir cases.

Chapter V presents the methodology, workflow, and cases for a novel


unstructured grid-based embedded discrete fracture model (UnQ-EDFM). A robust
frontal Delaunay quadrangular grid is combined with the EDFM, to solve the problems
of current fractured reservoir production models in stimulating the production of CFN.
Testified by reservoir cases, the UnQ-EDFM is proved to be more accurate and reliable
for the production simulation of irregularly shaped and complexly fractured shale
reservoirs.

Chapter VI presents the methodology, workflow, and cases of a surrogated-


assisted multi-objective optimization of the hydraulically fractured well considering the
geological uncertainty. A series of stochastic reservoir models are generated to describe
geological uncertainty, and the multi-objective optimization (MOO) strategy is adopted
to realize the optimization. Then, a novel hybrid multi-objective particle swarm
optimization algorithm (HMOPSO) is developed and combined with a Kriging
surrogate model to deal with the computationally expensive problem. The hydraulically
fractured well is optimized using this method in uncertain reservoir cases and a set of
trade-off solutions are provided for designers.

Chapter VII draws conclusions stemming from the above research chapters of
this dissertation and proposes recommendations for future research.

9
Texas Tech University, Hao Zhang, August 2021

CHAPTER II

LITERATURE REVIEW
2.1 Hydraulic fracturing propagation simulation method

Hydraulic fracturing is a common and effective stimulation method to enhance


production in the shale reservoir. So precisely calculate the propagation of hydraulic
fracturing is an essential problem for shale gas and oil development. From the aspect of
economics, the numerical simulation method has been developed by scholars to study
the mechanism of hydraulic fracturing propagation and acted as the guidance for the
hydraulically fractured well design. An acceptable numerical model should at least be
capable to capture major physical processes during hydraulic fracturing. The basic
processes of hydraulic fracturing include (1) fracture deformation induced by internal
pressure in the fracture; (2) fluid flow in the fracture; (3) fluid leak-off into the formation
and (4) fracture propagation (Veatch et al., 1986; Adachi et al., 2007). Based on these
above requirements, a series of excellent analytical and numerical models were
proposed and reinforced to make the hydraulic fracturing simulation more consistent
with the real field application.

2.1.1 Analytical model

Hydraulic fracturing analytical models have evolved from simple two-


dimensional models to complex three-dimensional models. Typical two-dimensional
models are Khristianovich-Geertsma-DeKlerk (KGD) model (Khristianovic and
Zheltov, 1955; Geertsma and De Klerk, 1969) and Perkins-Kern-Nordgren (PKN)
model (Perkins et al., 1961; Nordgren, 1972). The KGD model presumes plane strain
for a fracture in the horizontal plane and assumes no variation of fracture width in the
vertical direction (Figure 2.1-a). The fracture has an elliptical shape in the horizontal
plane, it has a vertical rectangular cross-section perpendicular to fracture propagation
direction. The PKN model assumes plane strain for a fracture in a vertical plane, normal to
the fracture propagation direction, and the vertical cross-section of the fracture is elliptical
in shape (Figure 2.1-b), each vertical cross-section is independent of each other.

10
Texas Tech University, Hao Zhang, August 2021

(a) KGD model (b) PKN model


Figure 2.1 The schematic view of KGD and PKN models (Wu, 2014)

These two analytical models are approximations and have been shown to be
useful, but they have limits to their accuracy. The KGD model is more suitable for
fracture propagation at the early injection time when fracture height is much greater
than fracture length, while the PKN model is more reasonable at late injection time when
fracture length far exceeds fracture height. Based on these classical two-dimensional
analytical models, pseudo-3D and true 3D models were developed to describe the
hydraulic fracturing process with better accuracy. Simonson et al. (1978) developed a
pseudo-3D model to simulate height growth in a symmetric three-layer formation and
investigated the effects of different parameters for the pay zone and the barrier
formations on fracture containment. Fung et al. (1982) extended Simonson et al.’s work
to multi-layered asymmetrical formations using a semi-analytic technique. Settari et al.
(1984) also developed a pseudo-three-dimensional model by describing lateral fluid
flow and crack opening for the fracture and coupling with an efficient scheme to deal
with vertical fracture growth. Pseudo-3D models are efficient to capture the behavior of
fracture height growth by employing special schemes, but most of them do not consider
the fluid flow in the vertical direction, then scholars developed several true 3D models
(Clifton et al., 1981; Abou-Sayed et al., 1984). The general true 3D models are based
on the analysis of three-dimensional rock deformation and two-dimensional fluid flow
through the fractures, the fluid movement is calculated not only in the lateral direction
but also in the vertical direction. The advantage of the analytical model is that the

11
Texas Tech University, Hao Zhang, August 2021

computational cost is usually not too expensive; whereas, the fractures described by the
analytical model are usually simple planer fractures and cannot account for the whole
complexity of the problem.

2.1.2 Discrete element method

Based on the theoretical foundation of analytical models, a lot of excellent


numerical models for hydraulic fracturing propagation simulation were developed. The
discrete element method (DEM) is a typical numerical model used to describe the
mechanical behavior of the hydraulic fracturing process by dis-continuous bodies.
Introduced by Cundall and Strack (1979), the DEM was developed for the analysis of
rock mechanic problems using deformable polygonal-shaped blocks. Today’s DEM is
becoming widely accepted as an effective method of addressing engineering problems
in granular and discontinuous materials.

Figure 2.2 Particle contacts and flow domains for the DEM model (De Pater et al., 2005)

For simulation using DEM, the domain is subdivided into a series of small
distinct particles, the fracture is simulated by calculating the displacement of particles
(Figure 2.2). At the start of each time step, the set of contacts is updated from the known
particle and wall positions, and the force-displacement law is applied to update the
contact forces of the constitutive model; next, the law of motion is applied to each
particle to update its velocity and position. The DEM was widely applied by scholars to
simulate hydraulic fracturing. Garcia et al. (2013) proposed a study on hydraulic fracture

12
Texas Tech University, Hao Zhang, August 2021

vertical propagation through layered formations, in which the DEM was applied to
account for the physic of a T-shaped fracture growth between layers. Zhang et al. (2017)
simulated the proppant embedded hydraulic fracture, the fracture conductivity after the
fracture closing was evaluated by the DEM integrating the CFD model. Tan et al. (2017)
used a DEM-based software PFC2d to establish the hydraulic fracturing model in a
sandstone-mudstone interbedding reservoir, the propagation geometry of the fracture is
simulated and the effect of in-situ stress difference, rock interface properties, and
construction parameters on the propagation mechanism was analyzed. The advantage of
the DEM is that it simulates a wide variety of granular flow and rock mechanics
situations; however, the calculation using DEM is computationally demanding, and not
suitable to simulate large-scaled fractures.

2.1.3 Extended finite element method

The extended finite element method (XFEM) was developed by Moës et al.
(1999), to help alleviate shortcomings of the traditional finite element method (FEM).
This technique allows the entire fracture to be represented independently of the mesh,
so re-meshing is not necessary to the model with fracture growth; besides, additional
degrees of freedom for the enriched elements around the fractures eliminate the need
for mesh refinement. (Figure 2.3).

Figure 2.3 A schematic view of the XFEM model

13
Texas Tech University, Hao Zhang, August 2021

The principle of XFEM is under the standard finite element framework, the
continuous area adopts the standard finite element method, and the displacement
approximation function of the finite element is modified in a relatively narrow area
containing discontinuous boundaries. The level set method is used to describe the
discontinuous interface, making the discontinuous description independent of finite
element grids to avoid grid encryption and reconstruction during the calculation process.
Lecampion (2009) first applied the XFEM into hydraulic fracturing, fracture opening
and fluid pressure simulated are compared at each node with analytical solutions. Ren
et al. (2009) explored numerical modeling for concrete hydraulic fracturing with the
XFEM by assuming that water pressure is constant and imposed on the fracture surfaces.
Keshavarzi et al. (2012) adopted the XFEM to simulate the hydraulic fracturing
interaction with the pre-existing natural fracture and in-situ stress difference and found
that the fracture interaction was strongly controlled by the orientation of the natural
fractures as well as hydraulic fracture net pressure. Zeng et al. (2019) numerically
studied hydraulic fracturing in an anisotropic poroelastic medium via a hybrid XFEM
approach and validate the results against the analytical solutions of the KDG model. The
main advantage of the XFEM method is that the finite element mesh does not need to
be re-meshed to track the fracture path, which simplified the calculation, so it is suitable
to simulate large-scaled fracture propagation. While the limit for the method is that it is
computational consuming especially simulating the complex fracture networks.

2.1.4 Displacement discontinuity method

The displacement discontinuity method (DDM) was first proposed by Crouch


(1976), which is a typical boundary element method. Based on the analytical solution
of stresses and displacements caused by a constant displacement discontinuity over a
finite line segment in an infinite elastic solid body, the displacement discontinuity is
calculated in the surface of fracture or can also be regarded as boundary of the domain,
rather than values throughout the space.

14
Texas Tech University, Hao Zhang, August 2021

Figure 2.4 Fracture discretized by DDM (Jo and Hurt, 2013)

In DDM-based fracture, the elastic rock deformation theory is used to calculate


fracture opening and shearing under given boundary conditions, which are provided by a
fluid mechanics model that computes the distribution of pressure along the fracture path.
The fracture opening generates fracture aperture which allows fluid to flow, and The
shearing controls the fracture propagation direction (Figure 2.4). The DDM is especially
concise in simulating multi-fracture propagation and fracture-crossing, so it was used by
scholars to simulate complex fracture network (CFN) propagation in recent years. Wu and
Olson (2016) coupled the DDM fracture propagation model with the pre-existing NFs
to simulate the CFN for the single-stage HF. Ren et al. (2016) built up a multi-stage
CFN numerical model by coupling the modified-DDM with the pseudo-3D fracture
propagation model and made sensitivity analysis on different geological parameters’
effect, but the stress shadow effect was not fully considered. Zeng et al. (2016)
performed the numerical modeling of hydraulic fracture propagation and fracture
network generation using the DDM and indicated that fracture spacing and stress
anisotropy have a significant influence on the propagation path and geometry of
multiple fractures. The DDM is often more efficient than other methods, for it works by
constructing grids over the fracture surface, not the entire body, which significantly
simplifies the calculation. However, the limit is also obvious, for the BEF is less
accurate than volume-discretization methods.

15
Texas Tech University, Hao Zhang, August 2021

2.2 Fractured reservoir production simulation method

The fracture production simulation is an effective method to evaluate the quality


of hydraulic fracturing. However, it is hard to simulate the fracture production using the
traditional reservoir simulating method; for the fracture is a kind of discrete element.
Facing this problem, scholars developed some outstanding numerical methods for
hydraulic fracturing production simulation.

2.2.1 Dual-porosity / dual-permeability model

In the early decades, scholars derived analytical models for the fluid flow in
fractured porous media, the fracture geometry was simple bi-wing fractures (Barenblatt
et al., 1960; Gringarten et al., 1974). Based on the analytical approaches, scholars then
developed several numerical models to simulate the production of fractured reservoirs.
The dual-porosity / dual-permeability model is the early proposed model, in which the
discrete fractures are transformed into a kind of continuous media and the reservoir
domain is subdivided by the structured grid (Warren and Root, 1963).

Figure 2.5 Illustration of the DPDP model

The concept of the DPDP is that it transforms the discrete fractures into a kind
of continuous media, so one reservoir grid contains two properties, one is for matrix
another for fracture. (Figure 2.5). The crossflow from the matrix to the fracture is
considered. Besides, the flow processes both between the matrix rock, and the matrix
rock with the fractures, are all taken into account. The DPDP model simplifies the
gridding and simulation process and makes it possible to simulate fracture production
with the ordinary grid-based reservoir model. However, the discrete fractures are treated

16
Texas Tech University, Hao Zhang, August 2021

as continuous media, which will inevitably reduce the calculation accuracy, so the
DPDP was often combinedly used with other models. Li et al. (2016) integrated the
DPDP with EDFM to simulated fluid flow in fractured shale reservoirs. Cordero et al.
(2019) integrate the DPDP with DFM to construct a hydro-mechanical model to study
the behavior of fracture networks in deformable fractured media. To simulate the gas
production from shale reservoir with the organic component, scholars extended the
conventional DPDP into a multiple interacting continua model or abbreviated as MINC
model (Pruess, 1983), the matrix grid in the MINC model is subdivided into several
levels of nested sub-grids, which stands for fractures, the organic and non-organic
components, so the transient interaction between the matrix and fractures can be treated
more realistically by the MINC model (Wu and Pruess 1988). Wu, et al. (2014)
constructed a generalized framework model for simulation of gas production in shale
reservoirs, in which the MINC model is employed to simulate the shale gas production
within the SRV region. Farah et al. (2019) performed a flow modeling of shale
reservoirs using a DFM-MINC approximate function, the transient effects and the
different flow regimes occurring during production were well-considered.

2.2.2 Local grid refinement model

Facing the limits of the DPDP model, a local grid refinement method (LGR) is
proposed by scholars. By using this method, some local grids can be refined into much
thinner grids to represent the discrete fractures, and the fracture properties, such as
conductivity and porosity, can be defined differently from the matrix grids.

Figure 2.6 A fracture constructed by the LGR method (Eshkalak et al., 2014)

17
Texas Tech University, Hao Zhang, August 2021

The LGR links two or more different‐sized finite‐difference grids: a coarse


(parent) grid covering a large area that incorporates regional boundary conditions, and
a fine (child) grid covering a smaller area of interest. The grid refinement covers only
the area of interest (Figure 2.6). The LGR preserved the discrete geometry of fractures,
which makes the simulation more reasonable and precise than the DPDP method.
Eshkalak et al. (2014) adopted the LGR model to evaluated the effectivity of the
enhanced gas recovery by CO2 sequestration and re-fracturing treatment in the shale gas
field and found that re-fracturing treatment outperformed the CO2 sequestration. Rubin
(2020) proposed an accurate simulation of Darcy and non-Darcy flow within an
explicitly modeled SRV, with and without primary hydraulic fractures, a LGR was used
to simulate the network fractures. However, the LGR is also not perfect, for fractures
generated by LGR can only parallel to the direction of the original reservoir grid. To
simulate the fractures that are not parallel to the reservoir grid direction, the refinement
gridding will be very complex.

2.2.3 Discrete Fracture Model

To further eliminate the limits by the LGR method, a discrete fracture model
(DFM) was proposed, which provides a more realistic representation of fractured
reservoirs. In DFMs, unstructured grids are generated based on the orientation of
fractures to conform to the oriental geometry and location of each fracture. (Figure 2.7).
The DFM can be used to simulate realistic geometry of the fracture system, thereby
accounting explicitly for the effect of individual fractures on fluid flow.

Figure 2.7 DFM construction by unstructured mesh

18
Texas Tech University, Hao Zhang, August 2021

Compared to DPDP and LGR models, the DFM offers several advantages. It
accurately represents the complex and irregular fracture pattern without simplification;
also, the specification of the exchange between matrix and fracture is relatively
straightforward since it depends directly on fracture geometry. The DFM provides the
flexibility and capability to consider more complex fracture configurations than LGR,
so it has been widely applied by scholars to simulate the production of fracture networks.
Jiang and Younis (2015) implement a lower-dimensional DFM for shale gas reservoir
production simulation with secondary fracture networks, the storage and transport
mechanisms for the gas flow in the fractured porous media were calculated. Liang et al.
(2016) proposed a systematic study of fracture parameters' effect on fracture network
permeability based on the DFM and made sensitivity analysis on various fracture
parameters. However, the shortcoming for DFM is also obvious. When simulating
massive fracture distribution, generating an unstructured grid will become substantially
challenging. A greatly refined grid will be generated near fractures, which not only
complicates the gridding process computationally expensive in production simulation
(Yan et al., 2016).

2.2.4 Embedded Discrete Fracture Model

To increase the adaptivity of simulating reservoirs with complex and massive


fracture distribution, Li and Lee (2008) developed an embedded discrete fracture model
(EDFM) which embedded the discrete fractures into the structured reservoir grids, this
approach was further extended by Moinfar et al. (2013) to model arbitrarily oriented,
dip angled fractures. The EDFM first generates structured grids in the reservoir matrix,
fractures are treated special grids outside the basic model, then embed the discrete
fracture grids into the matrix grids and coupled with the matrix domain by the non-
neighboring transmissibility factors. This approach introduces a robust methodology,
which could precisely model the complex fracture distribution without causing
complicated grids of the reservoir domain and will not lead to computationally
expensive simulation (Figure 2.8).

19
Texas Tech University, Hao Zhang, August 2021

(a) DFM (b) EDFM


Figure 2.8 Comparison of grids between DFM and EDFM (Yan et al., 2016)

Compare with traditional popular fracture simulation methods (Figure 2.9), the
EDFM perfectly solves the shortcomings of the DPDP, LGR, DFM, and takes synergic
advantage of them. It outperforms the traditional fracture simulation methods and
presents better performance in accuracy, computational efficiency, and flexibility,
which provides an alternative for efficient fractured reservoir simulation.

Figure 2.9 A comprehensive comparison between different fracture production


simulation methods
Based on these advantages, the EDFM has been widely used by scholars for
complicated fractured reservoir simulations. Chai et al. (2016) implemented the EDFM
model into a multi-porosity model to establish a dynamic fractured shale reservoir

20
Texas Tech University, Hao Zhang, August 2021

simulation approach that could estimate the SRV region around the fractures. Yan et al.
(2016) improved the numerical solution of EDFM with the mimetic finite difference
method and finite volume method to simulate fractured reservoirs with a complex
geometrical shape. Shakiba et al. (2018) implemented EDFM into an in-house simulator
to study gas production from the fracture network and evaluated the effect of fracture
network geometry. Liu and Reynolds (2019) incorporated the EDFM with the ES-MDA
algorithm to construct a stochastic fractal model and described the complex fracture
network by history match. Wu et al. (2019a) integrated the complex fracture and
numerical reservoir simulation through a non-intrusive EDFM to optimize the well
spacing in the Permian Basin. Recently, Wang (2020) proposed a framework for
naturally fractured shale gas reservoir simulation using EDFM, mechanisms like
adsorption isotherm and fracture deformation were considered. Because of the
outstanding performance in computational efficiency and gridding flexibility, the
EDFM has obtained increasing attention in complicated fractured reservoir simulation.

2.3 Stimulated reservoir volume estimation method

During decades of field fracturing application, scholars and engineers have


found from many evidence that, in naturally fractured shale reservoirs, hydraulic
fracturing could help create large fracture networks. Base on this observation, Maxwell
et al. (2002) and Fisher et al. (2004) first discussed the creation of large fracture
networks in the Barnett shale and show initial relationships between treatment size,
network size and shape, and production response. Additionally, experiments and field
observations all showed that the permeability and productivity within this fracture
network region are much higher than that of the shale matrix (Ozkan et al., 2010; Guo
et al., 2013). Soon after, a new concept of stimulated reservoir volume (SRV) was
proposed by Mayerhofer et al. (2010). It is an approximate 3-D volume to represent the
scale of the reservoir stimulated by complex fracture networks. The SRV is regarded as
the key indicator to evaluate the quality of fracturing operation, so precisely estimate
the scale of SRV, as well as the improvement within the SRV, is becoming an urgent
request.

21
Texas Tech University, Hao Zhang, August 2021

2.3.1 Field data monitoring method

The field data monitoring method is the first method adopted to estimate the
scale of SRV. This method estimates the SRV by recording real-time seismic or well
logging signals, then combining the observed data with surface and downhole tilt
fracture mapping to characterizing the possible distribution of fracture networks and
estimate the scale of SRV (Figure 2.10). Fisher et al. (2005) adopted the micro-seismic
monitoring method to estimate the SRV by recording the micro-seismic mapping to
characterize the created fracture networks. Mayerhofer et al. (2006) combined the
micro-seismic technology with fracture variation of the Barnett Shale and proposed the
relationship between the SRV volume of different reservoirs and the cumulative
production. Arvind et al. (2010) established an SRV estimation model for multi-stage
horizontal fractured well by combining the data obtained from micro-seismic
monitoring and proposed that in addition to the location of the main fracture. Besides,
the electromagnetic imaging method was adopted by LaBrecque et al. (2016), via
imaging the distribution of the electro-magnetic proppants to estimate the remote
fracture distribution.

Figure 2.10 The SRV estimation based on the micro-seismic data monitoring (Ren et
al., 2018)

22
Texas Tech University, Hao Zhang, August 2021

The advantage for field data monitoring is obvious, the SRV obtained usually
matches very well with the micro-seismic event and with high accuracy. it is reliable to
the real SRV distribution. However, the drawback of this method is also obvious, for a
lot of micro-seismic and expensive well-logging equipment should be used to obtain the
result, which costs huge economic investment. Besides, it can only obtain the scale of
SRV, but the formation stimulation within the SRV is hard to be described.

2.3.2 Matrix permeability increase method

Facing the high economic cost for field data monitoring, scholars tend to seek
for more economically affordable SRV estimation method. The SRV expansion mainly
depends on natural fracture failure generating from reservoir pressure lifting and
formation stress changing during shale fracturing, so scholars developed mathematical
models to estimate the scale of SRV by simulating the main process during shale
fracturing propagation and natural fracture failure. Meanwhile, a stress-dependent
permeability (SDP) methodology is used to calculate the permeability increase caused
by nature fracture brake, thus the stimulation within the SRV can also be described (Ren
et al., 2018) (Figure 2.11).

Figure 2.11 Schematic view of SRV and its corresponding stress-dependent


permeability within a simulator model (Sen et al., 2018)

Nassir et al. (2014) predicted the SRV by simulating the formation press and
stress change, the permeability stimulation within the SRV was calculated according to

23
Texas Tech University, Hao Zhang, August 2021

the fracture aperture enhancement. Hu et al. (2017) established a mathematical model


based on full tensor permeability, continuum method, and fluid mass conservation
equation to analyze the scale and permeability increase of SRV. Ren et al. (2018)
estimated the scale of SRV by coupling the induced stress and fluid pressure caused by
hydraulic fracturing and calculated the permeability increase within the SRV by cubic
law. The Matrix permeability increase makes it more economically affordable for SRV
estimation, and the formation improvement can also be evaluated. However, the SRV
description is transformed from discrete fractures into continuous volume, which
reduced the accuracy of the geometry estimation.

2.3.3 Complex fracture network method

With the development of the fracture numerical simulation method, discrete


fractures could be modeled and simulated by numerical approaches more easily, so a
novel SRV estimation method based on the complex fracture networks (CFN) has been
proposed (Ren et al., 2016). According to this method, the nature fractures are described
discretely in the matrix grids; meanwhile, the hydraulic fracture propagates and interacts
with the pre-existing natural fractures to construct the CFN (Figure 2.12). This method
simulates the configuration of SRV closest to the real formation situation.

Figure 2.12 SRV represented by complex fracture network (Ren et al., 2016)

24
Texas Tech University, Hao Zhang, August 2021

The CFN was recognized and studied even before the concept of SRV was
proposed. During decades of field observation, scholars found that the hydraulic fracture
could interact with the natural fracture and construct a fracture network (Maxwell et al.,
2002; Fisher et al., 2004). Then the mechanism of fracture interaction was studied.
Renshaw and Pollard (1995) experimented with a propagating fracture encountering an
orthogonal existing fracture and proposed the crossing criterion to judge the fracture
interaction; then Gu and Weng (2010) improved Renshaw’s crossing criterion into a
propagating fracture with any contacting angles of the existing fracture. Ever since the
CFN based on the crossing criterion was widely applied to evaluate the SRV. Wu and
Olson (2016) established a complex fracture network model to estimate SRV, which
coupled the fracture propagation model with a power-law natural fracture pattern and
used crossing criterion to simulate the HF and NF interaction. Ren et al. (2016) built up
a multi-stage CFN numerical model by coupling the modified-DDM with the pseudo-
3D fracture propagation model and made sensitivity analysis on different geological
parameters’ effect. Sheng et al. (2019) proposed a fractal induced-fracture method to
describe the CFN, in which the distribution of CFN is simulated by fractal dimension.
Liu and Reynolds (2019) combined the EDFM model with the DDM to estimate the
SRV and proposed a novel stochastic fractal model to describe the CFN. Shiqian et al.
(2020) applied history matching to transform the CFN model from the micro-seismic
data and inversely estimated reservoir properties. The CFN method combines the
advantages of both micro-seismic monitoring and permeability increasing, shows the
best performance for SRV estimation.

2.4 Hydraulically fractured well optimization method

Since illustrated in the above sections, the hydraulically fractured well is the
most effective stimulation strategy that enables the economic viability of shale
formation, Optimally designing the hydraulically fractured well to obtain the maximum
recovery with the minimum investment input is an essential research topic for shale
reservoirs’ development.

25
Texas Tech University, Hao Zhang, August 2021

2.4.1 Single objective optimization

Currently, most of the hydraulically fractured well optimization studies aim to


obtain a single optimal objective. Some studies mainly focus on the hydraulic fracturing
process, without considering the hydraulic fracture productivity. Their objective is to
obtain the best hydraulic fracture configuration, and the geo-stress is their main
consideration. Roussel and Sharma (2011) calculated the stress reversal and
reorientation extent for stress interference based on the 3D numerical model, the
optimum fracturing spacing for consecutive and altering fracturing was studied. Rafiee
et al. (2012) proposed an optimization method for zipper fractures, in which the effect
of well spacing on normal stress was evaluated and analytically optimized, and the
contact area together with the fluid production was enhanced. Siddhamshetty et al.
(2018) proposed a closed-loop hydraulic fracturing operation framework, in which a
Kalman filter algorithm together with the predictive control method was applied to
achieve the desired fracture geometry. Additionally, the hydraulically fractured well
was also optimized based on the influence of SRV. Ren et al. (2016) studied the complex
fracture network propagation by coupling the modified-DDM with the pseudo-3D
fracture propagation model. To optimize the multi-well fracturing treatments, Li and
Zhang (2018) numerically investigated the fracture network growth, the effects of
perforation-cluster spacing, well spacing, and the fracturing sequence of multi-well
completion upon fracture network complexity were studied.

Furthermore, some other studies consider the entire life cycle of the hydraulic
fractured well, their purpose is to obtain the best economic revenue, and the net present
value (NPV) is widely applied as the optimization objective. Defined as the difference
between the present values of cash inflow and outflow over a certain time (Ross, 1995),
the NPV is a perfect objective function to evaluate the economic viability of the
hydraulic fracturing design. Yu and Sepehrnoori (2013) employed surface methodology
into multiple horizontal well placement optimization to maximize NPV with
numerically modeling multistage hydraulic fractures in combination with economic
analysis. Jahandideh and Jafarpour (2016) optimized the hydraulic fracturing

26
Texas Tech University, Hao Zhang, August 2021

considering the geospatial variability shale frac-ability, the heterogeneity for shale frac-
ability was taken into consideration, a more economical adaptive horizontal well with
varying fracture spacing and fracture half-length was obtained. Liu and Forouzanfar
(2018) applied an efficient robust optimization approach into naturally fractured
reservoirs, the hierarchical clustering method together with the StoSAG scheme was
proposed to maximize the NPV. Xu et al. (2018) presented an efficient multistage
fractured horizontal well optimization framework that coupled EDFM and intelligent
algorithms to maximize the NPV, four optimization algorithms were compared in their
study. Yu et al. (2018) combined EDFM with LBM to simulate the gas slippage
phenomenon in CFN, complex fracture geometry was predicted by a fracture
propagation model considering natural fractures.

2.4.2 Multi-objective optimization

The above single objective hydraulically fractured well optimization is usually


based on a certain reservoir geological condition. However, since the reservoir
description from well logging and core analysis is limited, the uncertainty of geological
parameters, such as permeability porosity and natural fracture (NF) distribution,
becomes an unavoidable factor and presents a significant impact on optimization. To
enhance the performance of optimization with geological uncertainty, several
approaches have been developed. The first strategy is the mean value method, in which
a series of reservoir models are generated to describe as many geological realizations as
possible under the geological uncertainty; the mean value of metrics for all the
geological realizations is calculated and put into the optimization process (Jansen et al.,
2009). However, this method can only obtain a good expected value but may also cause
a large variance, which leads to a huge development risk. The second strategy is the
mean-variance method, which aggregates the mean value and the standard deviation
into one objective function with corresponding weight coefficients, so both the expected
value and the variance could be considered in the optimization process (Yang et al.,
2011; Siraj et al., 2015; Zhang et al., 2019). Though the robustness of optimization is
significantly improved, the mean-variance method is still not perfect, for the artificially

27
Texas Tech University, Hao Zhang, August 2021

weighted coefficients in the single objective function could not diminish the entire risk
(Liu, 2020; Zhao et al., 2020a). To fully consider the balance between the expected
value and the risk, the third strategy was proposed, which was named the multi-objective
optimization method (MOO). In MOO, the mean value and the variance are regarded as
two separate objectives and optimized simultaneously, a Pareto front (PF) is adopted to
search for the non-dominated results and offer multiple trade-off solutions for the
decision-makers (Figure 2.13) (Isebor and Durlofsky, 2014).

Figure 2.13 The Pareto front for multi-objective optimization

Since the MOO is different from traditional single-objective optimization and


usually leads to a large amount of calculation, professional and efficient optimization
algorithms are urgently required. The gradient-based approaches were first proposed,
which were theoretically concise and computationally efficient (Liu and Reynolds,
2016). However, the gradient-based algorithms are easy to get trapped into local
extreme and susceptible to the continuity and shape of the PF. To overcome these
drawbacks, heuristic algorithms were developed. A nondominated sorting genetic
algorithm (NSGA) was proposed and later updated into NSGA-II (Deb et al., 2002); as
a kind of heuristic algorithm, it searches for the PF by simulating the reproduction and
natural selection processes of creatures. Similarly, a multi-objective particle swarm
optimization algorithm (MOPSO) was also developed by simulating the behaviors of

28
Texas Tech University, Hao Zhang, August 2021

bird swarm (Coello et al., 2004). These approaches avoid the gradient calculation and
are also less susceptible to the nonconvex PF, which gives them a wide application in
MOO of reservoir management (Chang, 2015). Based on the above, several hybrid
algorithms were developed to strengthen the MOO accuracy and efficiency, such as the
NSGA-II-SDR (Tian et al., 2018), CS-NSGA-II (Zhao et al., 2020b), and CMOPSO
(Zhang et al., 2018).

Since scholars usually describe the uncertainty by using large numbers of


stochastic models, the MOO leads to an unavoidable huge amount of calculation even
using the efficient algorithms, how to reduce the amount of calculation while
maintaining the optimization accuracy becomes another urgent problem, one effective
solution to this problem is the surrogate model. By training the relative parameters using
the real simulation data, the surrogate model will gradually converge to the numerical
simulation results, then it could be used to replace the numerical simulation process in
the optimization. Many approximation methods have been applied to build surrogate
models and have revealed excellent performance, such as the radial-basis function (RBF)
(Sun et al., 2017), the Kriging model (Liu et al., 2017), the artificial neural network
(ANN) (Pan et al., 2018), and also the ensemble of surrogate models (Alizadeh et al.,
2019). By replacing the time-consuming numerical simulation with a concise surrogate
model, the efficiency of the optimization process could be increased dozens of times
and the accuracy will not be affected (Ong et al., 2003). However, the research on multi-
objective hydraulically fractured well optimization in uncertain reservoirs is limited,
few scholars have made a comprehensive study in this field; besides, the novel MOO
algorithms, as well as the surrogate model methodologies have not attracted much
attention in the hydraulically fractured well optimization, more studies are still needed.

29
Texas Tech University, Hao Zhang, August 2021

CHAPTER III

STIMULATED RESERVOIR VOLUME ESTIMATION AND


HYDRAULIC FRACTURING OPTIMIZATION
This chapter focuses on developing an effective numerical hydraulic fracturing
and SRV estimation model, which incorporates the stress shadow effects, fluid
distribution among multiple fractures, and the natural fracture stimulation criterion, to
precisely estimate the SRV in a less expansive and computational consuming way.
Furthermore, an effective and robust optimization framework is constructed to predict
and optimize fracturing parameters to obtain the best SRV, which could adapt to the
geological complexity at the field scale. The methodology for the SRV estimation and
hydraulic fracture optimization are organized in the following procedure: First, a DDM
method is employed to simulate the induced stress distribution around the hydraulic
fractures; additionally, a power-law probability distribution method is used to simulate
the natural fracture (NF) distribution, and the NF stimulation is determined according
to the Warpinski’s theory (Warpinski et al., 1987; Warpinski, 1991); then the SRV
region is estimated coupling the effect of NF net pressure and induced stress. Hereafter,
a stress-dependent permeability (SDP) within the SRV region is calculated by a cubic-
law algorithm. After the SRV estimation, a hydraulic fracturing optimization framework
is constructed, in which a Modified-PSO algorithm is proposed and employed to
optimize the NPV.

3.1 Methodologies for SRV estimation

3.1.1 Induced stress model

SRV is generated based on the failure and connection of natural fractures (NF)
(Nassir et al., 2014). Within a certain region, the NFs will be stimulated by the coupling
mechanism of induced stress and net pressure caused by the hydraulic fractures, which
will then connects into an SRV (Su et al., 2015). So the SRV can be estimated by
coupling the induced stress model, the hydraulic pressure model, as well as the NF
failure criterion. In this study, the discontinuity displacement method (DDM) is adopted

30
Texas Tech University, Hao Zhang, August 2021

to model the induced stress caused by hydraulic fractures. First proposed by Crouch
(1976), the DDM is a commonly used method for induced-stress simulation. Since only
the fracture elements need to be discretized, the DDM calculates faster than other
methods, and maintains the simulation accuracy. As is shown in Figure 3.1, a hydraulic
fracture is subdivided into 𝑁 segments, with each length of 2𝑎𝑖 . A local 𝜁, 𝜉 coordinate
system is built for each frac-unit, and the origin is located at the center of the unit, in
which, 𝜉 towards the tangential direction and 𝜁 towards the normal direction. Based on
the DDM, the influence of other frac-units on frac-unit 𝑖 are expressed as following
equations.

Figure 3.1 Discrete fracture units

(𝜎𝑛 )𝑖 = ∑𝑁 𝑁
𝑗=1(𝐺)𝑖𝑗 (𝐴𝑛𝑠 )𝑖𝑗 (𝐷𝑠 )𝑗 + ∑𝑗=1(𝐺)𝑖𝑗 (𝐴𝑛𝑛 )𝑖𝑗 (𝐷𝑛 )𝑗
{ (3-1)
(𝜎𝑠 )𝑖 = ∑𝑁 𝑁
𝑗=1(𝐺)𝑖𝑗 (𝐴𝑠𝑠 )𝑖𝑗 (𝐷𝑠 )𝑗 + ∑𝑗=1(𝐺)𝑖𝑗 (𝐴𝑠𝑛 )𝑖𝑗 (𝐷𝑛 )𝑗

where (𝜎𝑛 )𝑖 and (𝜎𝑠 )𝑖 are the normal and shear stress on the frac-element 𝑖;
(𝐷𝑛 )𝑗 and (𝐷𝑠 )𝑗 are the normal and shear strains on frac-element 𝑗; (𝐴𝑠𝑠 )𝑖𝑗 , (𝐴𝑠𝑛 )𝑖𝑗 ,
(𝐴𝑛𝑠 )𝑖𝑗 , and (𝐴𝑛𝑛 )𝑖𝑗 are the elastic influence coefficients on frac-element 𝑖, which are
induced by the normal and shear strains of frac-element 𝑗, the detailed derivation and
description of these coefficients is given by Crouch and Starfield (1983); the (𝐺)𝑖𝑗 is a
correction factor derived by Olson (2004) for the HF with a finite height:
𝛽
𝑑𝑖𝑗
(𝐺)𝑖𝑗 =1- 𝛽/2 (3-2)
2 +(ℎ/𝛼)2 ]
[𝑑𝑖𝑗

31
Texas Tech University, Hao Zhang, August 2021

where 𝑑𝑖𝑗 is the distance from the centroids of frac-element 𝑖 to that of frac-
element 𝑗; ℎ is the fracture height; 𝛼 and 𝛽 are the empirical constants, where 𝛼 = 1
and 𝛽 = 2.3 shows the best correction effect.

When a hydraulic fracture is opened by injecting fluid, the internal net pressure
𝑃𝑛𝑒𝑡 of the fracture will equal the difference between the hydraulic pressure 𝑃𝑓 inside
the fracture and the minimum principal stress 𝜎ℎ outside the fracture. The stress
equilibrium equations for frac-unit 𝑖 can be calculated by the following boundary
condition:

(𝜎𝑡 )𝑖 = 0
{ (3-3)
(𝜎𝑛 )𝑖 = −𝑃𝑛𝑒𝑡 = −(𝑃𝑓 − 𝜎ℎ )

Given the boundary condition, the equilibrium equations can be calculated, and
the tangential and normal displacement discontinuities (𝑢̂𝑡 )𝑖 and (𝑢̂𝑛 )𝑖 for frac-unit 𝑖
are obtained. Then, the induced stress for any point field can be calculated by the
following equations (Crouch, 1976; Ren et al., 2015).
̂𝑛
𝐺𝑢
𝛥𝜎𝑥𝑥 = [2𝑛𝑙𝐹3 + (𝑛2 − 𝑙2 )𝐹4 + 𝜁(𝑙𝐹5 + 𝑛𝐹6 )]
2𝜋(1−𝜈)
̂𝑡
𝐺𝑢
+ [2𝑛2 𝐹3 − 2𝑛𝑙𝐹4 + 𝜁(𝑛𝐹5 − 𝑙𝐹6 )]
2𝜋(1−𝜈)
̂𝑛
𝐺𝑢
𝛥𝜎𝑦𝑦 = [2𝑛𝑙𝐹3 + (𝑛2 − 𝑙2 )𝐹4 − 𝜁(𝑙𝐹5 + 𝑛𝐹6 )]
2𝜋(1−𝜈)
̂𝑡
𝐺𝑢 (3-4)
+ [2𝑙2 𝐹3 + 2𝑛𝑙𝐹4 + 𝜁(𝑛𝐹5 − 𝑙𝐹6 )]
2𝜋(1−𝜈)
̂𝑛
𝐺𝑢 𝑡 ̂
𝐺𝑢
{𝛥𝜎𝑥𝑦 = 2𝜋(1−𝜈) 𝜁(𝑙𝐹6 − 𝑛𝐹5 ) + 2𝜋(1−𝜈) [𝐹4 + 𝜁(𝑙𝐹5 + 𝑛𝐹6 )]

where ∆𝜎𝑥𝑥 , ∆𝜎𝑦𝑦 and ∆𝜎𝑥𝑦 are the induced normal and shear stresses; 𝐺 and 𝑣
represent the shear modulus and Poisson’s ratio, respectively; 𝑙 represents the cosine
value of the angle between the 𝜁-axis in the local coordinate system and 𝑥-axis in the
global coordinate system, and 𝑛 represents the cosine value of the angle between the 𝜁-
axis in the local coordinate system and 𝑦-axis in the global coordinate system, as shown
in Figure 3.2; 𝐹3 − 𝐹6 are the coefficient equations, the details can be seen in the
following equations (Crouch, 1976).

32
Texas Tech University, Hao Zhang, August 2021

Figure 3.2 Angles between the different coordinate systems

(𝑛2 −𝑙 2 )𝜁−2𝑛𝑙(𝜉+𝑎) (𝑛2 −𝑙 2 )𝜁−2𝑛𝑙(𝜉−𝑎)


𝐹3 = (𝜉+𝑎)2 +𝜁 2
− (𝜉−𝑎)2 +𝜁 2
2𝑛𝑙𝜁+(𝑛2 −𝑙 2 )(𝜉+𝑎) 2𝑛𝑙𝜁+(𝑛2 −𝑙 2 )(𝜉−𝑎)
𝐹4 = (𝜉+𝑎)2 +𝜁 2
+ (𝜉−𝑎)2 +𝜁 2
𝑛(𝑛2 −3𝑙 2 )[(𝜉+𝑎)2 −𝜁 2 ]+2𝑙(3𝑛2 −𝑙 2 )(𝜉+𝑎)𝜁
𝐹5 = [(𝜉+𝑎)2 +𝜁 2 ]2
𝑛(𝑛2 −3𝑙 2 )[(𝜉−𝑎)2 −𝜁 2 ]+2𝑙(3𝑛2 −𝑙 2 )(𝜉−𝑎)𝜁
− [(𝜉−𝑎)2 +𝜁 2 ]2 (3-5)
2
2𝑛(𝑛2 −3𝑙 2 )(𝜉+𝑎)𝜁−𝑙(3𝑛2 −𝑙 2 )[(𝜉+𝑎)2 −𝜁 2 ]
𝐹6 = [(𝜉+𝑎)2 +𝜁 2 ]2
2
2𝑛(𝑛2 −3𝑙 2 )(𝜉−𝑎)𝜁−𝑙(3𝑛2 −𝑙 2 )[(𝜉−𝑎)2 −𝜁 2 ]
{ − [(𝜉−𝑎)2 +𝜁 2 ]2

where 𝜁, 𝜉 are the coordinates of the field point in the local coordinate system of
frac-unit 𝑖.

After the induced stress is obtained, the current stress distribution in the field
can be calculated by:
0
𝜎𝑥𝑥 = 𝜎𝑥𝑥 + 𝛥𝜎𝑥𝑥
0
{𝜎𝑦𝑦 = 𝜎𝑦𝑦 + 𝛥𝜎𝑦𝑦 (3-6)
𝜎𝑥𝑦 = 𝛥𝜎𝑥𝑦

0 0 0
where 𝜎𝑥𝑥 , 𝜎𝑦𝑦 , and 𝜎𝑥𝑦 are the initial formation stress; ∆𝜎𝑥𝑥 , ∆𝜎𝑦𝑦 , and ∆𝜎𝑥𝑦
are the incremental formation stress; 𝜎𝑥𝑥 , 𝜎𝑦𝑦 , and 𝜎𝑥𝑦 are the current formation of
stress.

33
Texas Tech University, Hao Zhang, August 2021

3.1.2 The SRV model

When doing the hydraulic fracturing operation, a huge quantity of fracturing


fluid will be injected into the hydraulic fractures, then leak-off into the reservoir. This
fracturing fluid flow in the reservoir can be represented by the continuity equation
according to pour mechanics theory (Ren et al., 2018):

𝜕2 𝑝 𝜕2 𝑝 𝜕𝑝
𝑘𝑥 + 𝑘𝑦 = 𝜇𝜙𝐶𝑡 (3-7)
𝜕𝑥 2 𝜕𝑦 2 𝜕𝑡

where 𝑘𝑥 , 𝑘𝑦 are the permeabilities in 𝑥, 𝑦 directions, respectively; 𝑝 represents


the formation pressure; ∅ is the porosity; 𝜇 is the viscosity of the injection fracturing
fluid; 𝐶𝑡 is the formation comprehensive compression coefficient.

The initial and boundary conditions for this above equation are:

𝑝(𝑥, 𝑦, 0) = 𝑝𝑖
{𝑝(𝑥𝑖 , 𝑦, 0) = 𝑝𝑓 (𝑥𝑖 ) (3-8)
𝑝(+∞, +∞, 0) = 𝑝𝑖

where 𝑝𝑖 represents the initial formation pressure; 𝑥𝑖 is x-location of the


𝑖th hydraulic fracture; 𝑝𝑓 (𝑥𝑖 ) is the fracturing pressure of the 𝑖 th hydraulic fracture.

The continuity equation Eq.(3-7) can be transformed into a finite difference


equation as follows.
𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1
𝐸𝑖,𝑗 𝑝𝑖−1,𝑗 + 𝑊𝑖,𝑗 𝑝𝑖+1,𝑗 + 𝑄𝑖,𝑗 𝑝𝑖,𝑗 + 𝑆𝑖,𝑗 𝑝𝑖,𝑗−1 + 𝑁𝑖,𝑗 𝑝𝑖,𝑗+1 = 𝑄𝑖,𝑗 (3-9)

where, 𝐸, 𝑊, 𝐴, 𝑆, 𝑁, 𝑄 are the coefficients, the derivation of which can be seen


in Ewing (1983); 𝑖, 𝑗 are grid numbers in 𝑥, 𝑦 direction, independently.

Figure 3.3 Normal and shear failure of NF

34
Texas Tech University, Hao Zhang, August 2021

Under the coupling mechanism of induced stress and hydraulic pressure, the NF
in some regions of formation will endure normal and shear failures (Figure 3.3), and the
failure condition of NF can be judged. The unit normal vector of NF is:

𝒏 = {𝑛𝑖 𝑒𝑖 } = (𝑛𝑥 , 𝑛𝑦 ) (3-10)

where 𝑛𝑖 represents the unit normal vector component of NF; 𝑒𝑖 is the unit
vector of the coordinate system; 𝑛𝑥 = 𝑐𝑜𝑠𝛼, 𝑛𝑦 = 𝑠𝑖𝑛𝛼, they are the normal vector
components in 𝑥, 𝑦 directions, in which 𝛼 is the angle between the tangential direction
of the NF and the maximum principal stress direction (𝑦-direction).

Then the shear and normal stress on the NF surface can be calculated by:
2 2
𝜎𝜏 = √𝑭 • 𝑭 − 𝜎𝑛2 = [(𝜎𝑥𝑥 𝑛𝑥 + 𝜎𝑥𝑦 𝑛𝑦 ) + (𝜎𝑥𝑦 𝑛𝑥 + 𝜎𝑦𝑦 𝑛𝑦 )
1/2 (3-11)
−(𝜎𝑥𝑥 𝑛𝑥 𝑛𝑥 + 𝜎𝑥𝑦 𝑛𝑥 𝑛𝑦 + 𝜎𝑦𝑥 𝑛𝑦 𝑛𝑥 + 𝜎𝑦𝑦 𝑛𝑥 𝑛𝑥 )]
𝜎
{ 𝑛 = 𝑭 • 𝒏 = 𝜎𝑥𝑥 𝑥 𝑥 + 𝜎𝑥𝑦 𝑛𝑥 𝑛𝑦 + 𝜎𝑦𝑥 𝑛𝑦 𝑛𝑥 + 𝜎𝑦𝑦 𝑛𝑥 𝑛𝑥
𝑛 𝑛

where 𝜎𝜏 , 𝜎𝑛 are the shear and normal stress on the surface of the NF.

Therefore, the tensile and shear of the NF caused by the normal and shear stress
can be judged by the Warpinski theory (Warpinski, 1991; Ren et al., 2015).

𝑝𝑛𝑓 > 𝜎𝑛 + 𝑆𝑡 𝑇𝑒𝑛𝑠𝑖𝑙𝑒 𝑓𝑎𝑖𝑙𝑢𝑟𝑒


{ (3-12)
𝜎𝜏 > 𝜏0 + 𝐾𝑓 (𝜎𝑛 − 𝑝𝑛𝑓 ) 𝑆ℎ𝑒𝑎𝑟 𝑓𝑎𝑖𝑙𝑢𝑟𝑒

where 𝑝𝑛𝑓 is the hydraulic pressure inside the NF; 𝑆𝑡 is the tensile strength of
NF; 𝐾𝑓 is friction coefficient of NF; 𝜏0 is the cohesion of NF. If the NF in a certain
region is opened by normal or shear failure, this region is regarded as the SRV region
(Lin et al., 2017).

3.1.3 The natural fracture distribution model

The SRV is generated by HFs opening and connecting the NFs, so accurately
describing the NF distribution is critical to the SRV estimation. Two characteristics of
the NF contribute to the network complexity, one is the fracture length distribution and
another is the fracture-center location. Based on a large quantity of core test and

35
Texas Tech University, Hao Zhang, August 2021

geological analysis, scholars have found that the NFs in unconventional reservoir fit a
power-law probability distribution (Segall and Pollard, 1983; Davy, 1993):

𝑛(𝑙)𝑑𝑙 = 𝛼𝑙−𝑎 (3-13)

where, 𝑛(𝑙)𝑑𝑙 represents the number of NFs with a length in the range
of [𝑙, 𝑙 + 𝑑𝑙]; 𝛼 is a coefficient of proportionality; 𝑎 is the length exponent, which is
usually between 1 and 3.

(a) Real NF pattern (Davy, 1993) (b) Power-law probability NF pattern


Figure 3.4 The comparison of real NF pattern and the power-law probability NF pattern

As is shown in Figure 3.4-a, a real NF pattern is shown and the relationship


between the number and length of NFs is depicted, which matches pretty well with the
power-law distribution trend. Additionally, a power-law probability NF pattern example
is generated in a 100 ft×100 ft reservoir (Figure 3.4-b), with the parameters 𝛼 = 10, 𝑎 =
2, 𝑙 ∈ [1,20] 𝑓𝑡. As can be seen that, the power-law probability NF pattern perfectly
reflects the distribution characters of the NF pattern, the existence probability of long
fractures is much smaller than that of short fractures. It can also refer to as the longer
the fracture is, the smaller the number will be, which is closed to the real NF distribution
(Gale et al., 2007; Wu and Olson, 2016).

36
Texas Tech University, Hao Zhang, August 2021

3.1.4 Stress dependent permeability

Caused by NF stimulation, the permeability within the SRV region will


dramatically increase. This stimulated permeability is named “stress-dependent
permeability (SDP)”, which has been studied by experiments and analytical methods
(Kassis and Sondergeld, 2010; Cho et al., 2013; Lopez et al., 2015; Zhu et al., 2018).
Among all of these SDP calculating methods, the cubic-law method is employed in this
study. First proposed by Bagheri (2006), the cubic law is simple but relatively precise
to estimate equivalent SDP caused by the NF stimulation in discrete reservoir grids (Tao
et al., 2009).
3
𝑛𝑛𝑓 𝑎𝑛𝑓
𝐾𝑛𝑓 = (3-14)
12

where 𝐾𝑛𝑓 is permeability caused by NF stimulation; 𝑛𝑛𝑓 is the NF frequency;


𝑎𝑛𝑓 is the NF aperture and the aperture of the NF can be divided into the tensile aperture
and shear aperture, which can be calculated by Hossain et al. (2002):
𝑝𝑛𝑓 −𝜎𝑛
𝑇𝑒𝑛𝑠𝑖𝑙𝑒 𝑎𝑝𝑎𝑟𝑡𝑢𝑟𝑒
𝐾
𝑎𝑛𝑓 = {𝛥𝜎𝜏 𝑛 (3-15)
𝑡𝑎𝑛( 𝜑𝑑𝑖𝑙 ) 𝑆ℎ𝑒𝑎𝑟 𝑎𝑝𝑎𝑟𝑡𝑢𝑟𝑒
𝐾𝑠

where 𝐾𝑛 , 𝐾𝑠 are the normal and shear stiffness of NF; 𝜑𝑑𝑖𝑙 is the shear dilation
angle of NF.

Usually, to perform the numerical simulation, the reservoir should be discretized


into grids; the NF pattern is a kind of discrete media, which is hard to be recognized and
calculated by the reservoir numerical simulator, so it should also be discretized into
grids. The cubic-law method in Eq.(3-14) is adopted to calculate the permeability
increase for each reservoir grid, caused by the stimulated NFs inside of it. The
permeability of a stimulated NF is transformed into the permeability of the entire
reservoir grid in which the NF is located. For a reservoir grid with 𝑛 NFs inside, the
stimulated permeability in 𝑥, 𝑦 directions can be calculated by the following equation
(Ren, et al., 2018):

37
Texas Tech University, Hao Zhang, August 2021

𝐾𝑥 = ∑𝑛i=1 𝑠𝑖𝑛2 (𝜃𝑖 ) ⋅ 𝐾𝑛𝑓𝑖 + 𝐾0


{ (3-16)
𝐾𝑦 = ∑𝑛i=1 𝑐𝑜𝑠 2 (𝜃𝑖 ) ⋅ 𝐾𝑛𝑓𝑖 + 𝐾0
where 𝜃𝑖 is the angle between the tangential direction of the ith NF and the y-
direction; 𝐾𝑛𝑓𝑖 is the stimulated permeability caused by the 𝑖 th NF; 𝐾0 is the original
formation permeability.

(a) NFs and reservoir grid subdivision (b) NFs transform to permeability field (mD)
Figure 3.5 Power-law NF distribution and permeability field

According to the above permeability calculation algorithm, the reservoir in


Figure 3.4-b is subdivided into 20×20 grids (Figure 3.5-a). We assume that all the NFs
in this reservoir example are stimulated with the same aperture of 0.1in, then the
permeability (mD) for each reservoir grid is shown in Figure 3.5-b. As can be seen that
the more NFs in a grid, the higher the permeability it will have. Using the cubic-law
together with the power-law algorithm, the reservoir, as well as the NFs, are subdivided
into grids, and the SDP is calculated.

3.1.5 Workflow of SRV estimation

Based on the above methods, the workflow for SRV estimation is established
and listed below and shown intuitively in Figure 3.6.

(1) Input the geological and fracturing parameters.


(2) In the first time step, based on the continuity equation mentioned in section 2.2,

38
Texas Tech University, Hao Zhang, August 2021

calculate the net pressure 𝑝𝑛𝑒𝑡 in the hydraulic fracture and the formation pressure
caused by fluid leak-off. After the 𝑝𝑛𝑒𝑡 is obtained, regard it as the boundary
condition for DDM, and calculate hydraulic fracture displacement discontinuities
𝑢̂𝑡 , 𝑢̂𝑛 . Then calculate the induced stress in the reservoir domain using DDM.
(3) Coupling the induced stress and formation pressure, calculate the shear and normal
stress on the surface of NF. Then judge the failure of NF according to the Warpinski
theory.
(4) Calculate the aperture of NF being stimulated, and hereafter, calculate the SDP in
the SRV region.
(5) Renew the reservoir permeability field, and back to step 2 for the next simulation
time step. When all the time step is over, output the final pressure, the stress field,
SRV, and SDP.
Start

Input geological and


fracturing parameters

SRV estimation model


Fluid continuity function DDM function

The formation pressure The induced stress


model model

pnf 

NF stimulation judgement
anf Cubic law

SRV calculation SDP calculation

No
t=tmax ?
Update the
Yes perm-field

Output the
SRV & SDP

Figure 3.6 Workflow for SRV estimation

39
Texas Tech University, Hao Zhang, August 2021

3.1.6 A case for SRV estimation

Based on the SRV and SDP construction methods mentioned above, an SRV
estimation example is established. A reservoir with a dimension of 200ft×500ft is
defined, and is subdivided into 100×100×1 blocks, which dimension is 2ft×5ft×10ft. A
hydraulic fracture is generated in the middle of the reservoir (represented by the black
line), the half-length of which is set as 150m. The fracturing pressure at the bottom of
the hydraulic fracture is set as 5000psi. The geological parameters for this reservoir are
listed in Table 3.1, some of which are derived from real reservoir data (Hu et al., 2017).

Table 3.1 Geological parameters for the reservoir

Parameters Value Parameters Value

Reservoir initial pressure (Pe) 2500 psi The fluid compressibility (𝐶𝑙 ) 5.5-6 psi-1
Reservoir porosity (∅) 0.06 The NF friction coefficient 0.2
Reservoir permeability (k) 0.1 mD The NF length range (𝑙) [1,20] ft
Minimal horizontal stress (𝜎ℎ ) 4000 psi The NF distribution coefficient (𝛼, 𝑎) (10, 2)
Maximal horizontal stress (𝜎𝐻 ) 4500 psi The NF tensile strength (𝑆𝑡 ) 75 psi
The shear modulus (𝐺) 6.46 psi The NF shear dilation angle (𝜑𝑑𝑖𝑙 ) 3°
The Poisson’s ratio (𝑣) 0.3 The NF normal stiffness (𝐾𝑛 ) 9.76 psi/ft
The rock compressibility (𝐶𝑟 ) 2.7-5 psi-1 The NF shear stiffness (𝐾𝑠 ) 4.86 psi/ft

After the simulation, the induced stress and pressure fields around the hydraulic
fractures are obtained (Figure 3.7). Based on the formation pressure together with stress,
the SRV and the SDP are calculated and shown in Figure 3.8. As can be seen that, the
SRV is generated around the hydraulic fracture, and the SDP within the SRV region
increases, which is conformed to the experimental results of previous scholars (Kassis
and Sondergeld, 2010; Zhu et al., 2018).

40
Texas Tech University, Hao Zhang, August 2021

(a) 𝜎𝑥𝑥 (psi) (b) 𝜎𝑦𝑦 (psi)

(c) 𝜎𝑥𝑦 (psi) (d) 𝑝𝑛𝑓 (psi)

Figure 3.7 The induced stress and pressure field around the fracture

Figure 3.8 The SRV region and the stress-dependent permeability after correction (mD)

41
Texas Tech University, Hao Zhang, August 2021

In the actual hydraulic fracturing operation, multiple fractures will be created


along the horizontal well, the hydraulic fracture spacing and fracture half-length will
affect the pressure and the induced-stress field between the adjacent hydraulic fracture,
in turn, affect the SRV. To estimate the influence of fracture spacing and half-length on
the SRV scale, an example with two adjacent hydraulic fractures is established. The
fracture half-length ranges from 20ft to 150ft with a step of 25ft; while the fracture
spacing ranges from 10ft to 150ft with a step of 5ft. After the calculation, the SRV and
the SDP can be seen in Figure 3.9 and Figure 3.10, respectively.

Figure 3.9 The SRV variation with the fracture spacing (fracture half-length=150ft)

42
Texas Tech University, Hao Zhang, August 2021

Figure 3.10 The stress-dependent permeability (mD) variation with different fracture
spacing (frac half-length=150ft)

SRV volumes for different fracture spacing and half-length are calculated and
summarized in Figure 3.11. As can be seen, the SRV volume will increase as the fracture
half-length increases, but the extent of increasing will gradually reduce. On the other
hand, as the fracture spacing increases, the SRV volume will first increase, but then
fracture spacing exceeds a certain distance, the SRV volume will shapely decrease and
then remain steady. The best fracture spacing for SRV volume in this study is around
35ft.

43
Texas Tech University, Hao Zhang, August 2021

Figure 3.11 The SRV volume vs. fracture half-length and fracture spacing

3.2 Methodologies for SRV optimization

3.2.1 Optimization objective function

The ultimate objective of reservoir optimization problems is usually obtaining


the highest economical benefit from the recovery while spending the lowest cost for the
investment, the net present value (NPV) is an effective economical parameter to reflect
this requirement, for it is determined by calculating the negative and positive cash flows
of an investment (Rammay et al., 2016). The NPV is calculated by the difference
between the present values of cash inflow and outflow over a certain time (Ross, 1995),
which is the perfect objective function to evaluate the economic viability of the
hydraulic fracturing design. The NPV for the shale gas production can be defined as:

𝑁𝑡 𝑁𝑝
(𝑟𝑔 ⋅ 𝑄𝑔𝑛 ,𝑗 − 𝑟𝑤 ⋅ 𝑄𝑤
𝑛
,𝑗
) × 𝛥𝑡 𝑛
𝐽𝑜𝑏 = ∑ {∑ 𝑡𝑛
}
𝑛=1 𝑗=1 (1 + 𝑏) 365

𝑁
𝑝 𝑓 𝑁
− ∑𝑗=1 (∑𝑖=1 𝐿𝑓𝑖𝑗 ⋅ 𝐶𝑓𝑖𝑗 + 𝐿𝑤𝑗 ⋅ 𝐶𝑤𝑗 + 𝑟𝑤 ⋅ 𝑄𝑓 ,𝑗 )
(3-17)

where 𝐽𝑜𝑏 is the value of NPV, 𝑛 represents the simulation time step; 𝑁𝑡 is the
whole time step for the shale gas production simulation; 𝑁𝑝 and 𝑁𝑓 represent the

44
Texas Tech University, Hao Zhang, August 2021

numbers of production wells and the number of hydraulic fractures in a certain


production well, respectively; 𝑟𝑔 represents the price the gas in unit volume;
𝑛 𝑛
𝑟𝑤 represents the cost of water disposal in unit volume; 𝑄𝑔,𝑗 , as well as 𝑄𝑤,𝑗 are the gas
and water production rate for the 𝑗th well in the 𝑛th simulation step; 𝐶𝑖𝑗 is the hydraulic
fracturing cost per unit length for the 𝑖 th hydraulic fracture in the 𝑗th well; 𝑄𝑓,𝑗 is the
total amount of water injected for the 𝑗th well, during the hydraulic fracturing; 𝐿𝑓𝑖𝑗 is the
length of the 𝑖 th hydraulic fracture in the 𝑗th well; 𝐶𝑤𝑗 is the well drilling cost per unit
length for the 𝑗th well, and 𝐿𝑤𝑗 is the length of the 𝑗th well; 𝑏 is the annual discount rate.

3.2.2 Optimization algorithm

In this study, the particle swarm optimization (PSO) algorithm is employed to


be the optimization algorithm to search for the optimal values of variables. First
developed by Kennedy and Eberhart (1995), the PSO algorithm is a kind of population-
based stochastic optimization algorithm. Considering a particle 𝑖 in an N-dimensional
searching space, the particle position in the 𝑘 th iteration step is defined as x𝑘𝑖 =
𝑘 𝑘 𝑘
(𝑥𝑖1 , 𝑥𝑖2 ⋯ 𝑥𝑖𝑁 ), and the searching velocity of this particle is v𝑘𝑖 = (𝑣𝑖1
𝑘 𝑘
, 𝑣𝑖2 𝑘
⋯ 𝑣𝑖𝑁 ).
Besides, the particle’s historical best position is PX𝑘𝑖 = (𝑝𝑥𝑖1
𝑘 𝑘
, 𝑝𝑥𝑖2 𝑘
⋯ 𝑝𝑥𝑖𝑁 ), and the
whole swarm’s historical best position is GX𝑘𝑖 = (𝑔𝑥𝑖1
𝑘 𝑘
, 𝑔𝑥𝑖2 𝑘
⋯ 𝑔𝑥𝑖𝑁 ). Accordingly, the
particle’s searching velocity in the (𝑘 + 1)th iteration step is updated by:
𝑘+1 𝑘 𝑘 𝑘 𝑘 𝑘
𝑣𝑖𝑗 = 𝑐0 𝑣𝑖𝑗 + 𝑐1 𝑟1 (𝑝𝑥𝑖𝑗 − 𝑥𝑖𝑗 ) + 𝑐2 𝑟2 (𝑔𝑥𝑖𝑗 − 𝑥𝑖𝑗 ) (3-18)

where 𝑐0 is the weighting coefficient for the searching velocity; 𝑐1 , 𝑐2 are the
weighting coefficients for the particle and global best position, which are usually set as
2; 𝑟1 , 𝑟2 represent two randomly generated numbers, within the limit of [0, 1];

Therefore, the particle’s position in the (𝑘 + 1)th iteration step can be updated
by:

𝑋𝑖𝑘+1 = 𝑋𝑖𝑘 + 𝑉𝑖𝑘+1 (3-19)

45
Texas Tech University, Hao Zhang, August 2021

where v𝑘+1
𝑖
𝑘+1 𝑘+1
= (𝑣𝑖1 𝑘+1
, 𝑣𝑖2 ⋯ 𝑣𝑖𝑁 ) is the searching velocity in the (𝑘 + 1) th
𝑘+1
iteration, in which, 𝑣𝑖𝑗 (𝑗 = 1,2 ⋯ 𝑁) is calculated from Eq.(3-18).

After all the optimization iteration step is finished, the historical best position
GX of the particle swarm will be regarded as the optimum value.

3.2.3 Modified-PSO algorithm

For the PSO algorithm, to obtain a precise result, the number of particles is set
as no less than 50. However, more particles mean a larger calculation quantity. For the
hydraulic fracture optimization problem, each particle needs one time of the numerical
simulation, so 50 particles will be too time-consuming to be used. To reduce the
calculation time, we need to decrease the particle number, which will inevitably reduce
the calculation accuracy, thus we are trapped in a paradoxical situation. Besides, we
found that when searching for the optimum value, the searching is usually led by
particles with better values, and the particles with the worst values cannot make effort
for searching for the optimum value but waste calculation. Facing this situation, we
propose a Modified-PSO algorithm, in which we sort the particles from better to worse
and pick up the worst particles and reassign random position values 𝑥 and velocity
values 𝑣 to them. The workflow for the Modified-PSO can be seen in Figure 3.12 and
the reassigning position operation for the Modified-PSO is intuitively presented in
Figure 3.13.

𝑋𝑖𝑤𝑜𝑟𝑠𝑡 (𝑡 + 1) = 𝐿𝐵 + (𝑈𝐵 − 𝐿𝐵) × 𝑟𝑎𝑛𝑑, 𝑖 = 1,2, ⋯ 𝑁𝑤𝑜𝑟𝑠𝑡 (3-20)

where 𝐿𝐵 and 𝑈𝐵 represent the lower and upper searching boundaries; 𝑟𝑎𝑛𝑑 is
a random number within (0,1); 𝑁𝑤𝑜𝑟𝑠𝑡 is the number of the worst particles to be
selected, by trial and error, 𝑁𝑤𝑜𝑟𝑠𝑡 = 0.3×𝑁𝑝𝑜𝑝 shows the best optimization effect.

46
Texas Tech University, Hao Zhang, August 2021

Start

Initialization of algorithmic
parameter and swarm position

Calculate fitness of each particle

Update PX and GX

Reassign random values to the


worst particles

Update particle veloscity and


position

Maximum value No
Reached?

Yes

End

Figure 3.12 The workflow for Modified-PSO

Figure 3.13 The sketch of reassigning position operation for the Modified-PSO

After reassigning new random values, those particles will get out of the worst
region and have an opportunity to jump to better regions, or even become the best
particles in the swarm. After this, all the particles will make effort searching for the
optimum value. The optimization variables and summarized as a vector X =
(𝑥1 , 𝑥2 ⋯ 𝑥𝑁 ) in which 𝑁 is the total number of optimization variables. For this

47
Texas Tech University, Hao Zhang, August 2021

fracturing horizontal well optimization problem, the hydraulic fracture half-length and
spacing, the horizontal well length, as well as the horizontal well spacing, are chosen as
the optimization variables X.

3.2.4 The modified-PSO algorithm validation

The reliability of the modified-PSO algorithm proposed in this study is testified


using a benchmark function as the objective function, which is shown below:
𝑥
𝐹(𝑥, 𝑦) = 𝑥 𝑠𝑖𝑛 ( ) 𝑐𝑜𝑠(𝑥) − 2𝑥 𝑠𝑖𝑛(2𝑥) (𝑦𝑠𝑖𝑛(𝑦) 𝑐𝑜𝑠(𝑦)) − 2𝑦𝑠𝑖𝑛(2𝑦) (3-21)
2

The image of the function 𝐹(𝑥, 𝑦) can be seen from Figure 3.14, which shows
that this function is a complex multi-local extreme value function with many local
extremes.

Figure 3.14 The field for optimization searching

The original and Modified-PSO algorithms are compared. Different particle


numbers of 50, 20, 10, 5, 3 are set for these two algorithms. The maximum iteration
time is defined as 50. After the optimization, some of the searching processes for
different particles can be seen from Figure 3.15 to Figure 3.16, and the results for all the
particles are shown in Figure 3.17.

48
Texas Tech University, Hao Zhang, August 2021

(a) Original PSO algorithm

(b) Modified-PSO algorithm


Figure 3.15 The optimization procedure for 50 particles

49
Texas Tech University, Hao Zhang, August 2021

(a) Original PSO algorithm

(b) Modified-PSO algorithm


Figure 3.16 The optimization procedure for 5 particles

50
Texas Tech University, Hao Zhang, August 2021

(a) The optimization accuracy of original PSO with different particle numbers

(b) The optimization accuracy of Modified-PSO with different particle numbers


Figure 3.17 The optimization accuracy of the original/modified PSO

It can be seen that, when there are 50 particles, the original PSO algorithm will
find the global optimal value easily; however, as the number of particles decreases, the
original PSO algorithm will be easy to trap in local extremes. When there are 3 or 5
particles, the ultimate optimized values will be far from the true optimum value. While
for the Modified-PSO algorithm, the final optimized value will always reach the global
optimal value as the number of the particles reduces from 50 to 5; even when the particle
number is 3, the result is still far better than that of the original PSO algorithm. The
comparison results are convincing enough to prove the advantage of the Modified-PSO
algorithm over the original one.

3.2.5 The workflow of hydraulic fracturing well optimization

51
Texas Tech University, Hao Zhang, August 2021

In each iteration step of the optimization, after the value for vector X is obtained,
the corresponding fracturing horizontal well with the SRV can be obtained and put into
the numerical simulation. Afterward, the NPV value is calculated by Eq.(3-17). The
detailed workflow is listed as follows and also shown in Figure 3.18.

(1) Randomly define an initial particle swarm 𝐗 ∗ = (X1 , X2 , X3 , ⋯ X𝑁 ), which contains


𝑁 particles in the optimization searching domain.
(2) Based on the geological and fracturing parameters, calculate the SDP field for each
particle 𝑖 according to the DDM and SDP methods.
(3) Put the obtained SDP field for each particle into the shale gas numerical model and
𝑘
start the production simulation, respectively. Then calculate the NPV value 𝐽𝑜𝑏,𝑖 for
𝑘
particle 𝑖, and obtain the best-know NPV value 𝐽𝑜𝑏,𝑏𝑒𝑠𝑡 . Then, use the Modified-
PSO method to update the position 𝑋𝑖𝑘+1 for particle 𝑖.
(4) Back to step 2, calculate a new SDP based on the variable 𝑋𝑖𝑘+1 , and then obtain
𝑘+1
the new NPV value 𝐽𝑜𝑏,𝑖 , accordingly.

(5) Repeat step 2-5 until the best-known NPV value 𝐽𝑜𝑏,𝑏𝑒𝑠𝑡 converges.

Start

Generate Initial
Particle Swarm

Calculate Import SDP Reservoir SDP


SDP model simulator update

NPV Value

No
NPV
Update the optimization Converged ?
variable using Modified-PSO Yes

Finish

Figure 3.18 Schematic diagram of the SRV optimization procedure

52
Texas Tech University, Hao Zhang, August 2021

3.3 Examples

In this section, the reliability of the hydraulic fracturing well optimization


framework consideration SRV is testified according to two examples of hydraulic
fracturing wells in shale gas reservoirs. Facing the complexity of fluid flow in shale gas
reservoirs, the MINC model is adopted to simulate gas production. First developed by
Pruess (1983), the MINC stands for “Multiple interacting continua” which is an
approximate method for simulating fluid flow fractured-porous media. Comparing with
the conventional DPDP model, the matrix grid in the MINC model is subdivided into
several levels of nested sub-grids, which stands for fractures, the organic and non-
organic components (Figure 3.19). The transient interaction between the matrix and
fractures can be treated more realistically by the MINC model (Wu and Pruess, 1988).

Figure 3.19 The concept of the MINC model (Pruess, 1983)

Base on the MINC model, two cases for hydraulic fracturing optimizations are
generated. In case 1, there is one horizontal production well placed in the reservoir, and
two optimizations procedure with different objectives are processed (Maximum SRV
and maximum NPV, respectively), the optimization results of which are compared. In
case 2, multi-wells are drilled in the reservoir; the development effect of two fracturing
placement types are optimized and compared (Aligning and alternating hydraulic
fractures). The detailed statements can be seen as follows.

3.3.1 Case 1: Single hydraulically fractured well optimization based on the effect
of SRV

As shown in Figure 3.20, the reservoir model used in this case is a two-
dimensional rectangular reservoir. For numerical simulation, a 1500ft×600ft field is

53
Texas Tech University, Hao Zhang, August 2021

discretized into 600×60×1 blocks, each one of which is 2.5ft×10ft×20ft. The reservoir
is homogenous and isotropic, geological properties for this reservoir are listed in Table
3.2. The relative permeability, as well as capillary pressure of water and gas in different
media, is shown in Figure 3.21. In the middle of the reservoir, a hydraulic fracturing
horizontal well is drilled, with the wellhead location (150ft, 300ft). The local grid
refinement method (LGR) is adopted to represent the hydraulic fractures, for which the
fracture width is set as 0.1in. The well is drilled along the direction of minimal principle
stress 𝜎ℎ , while the fractures along the maximal principle stress 𝜎𝐻 . The fracturing
pressure is set as 5000psi.

Figure 3.20 The sketch for single horizontal fracturing well optimization

Table 3.2 Geological parameters for the reservoir model

Parameters Value Parameters Value

Reservoir initial pressure (Pe) 2500 psi Hydraulic fracture conductivity(∅𝐻𝐹 ) 0.8
Reservoir matrix permeability (𝑘𝑚 ) 100 nD The NF permeability (𝑘𝑓 ) 0.01 mD
Reservoir matrix porosity (∅𝑚 ) 0.05 The NF porosity (∅𝑓 ) 0.5
Minimal horizontal stress (𝜎ℎ ) 4000 psi The NF friction coefficient 0.2
Minimal horizontal stress (𝜎𝐻 ) 4500 psi The NF length range (𝑙) [1,20] ft
The shear modulus (𝐺) 6.46 psi The NF distribution coefficient (𝛼, 𝑎) (10, 2)
The Poisson’s ratio (𝑣) 0.3 The NF tensile strength (𝑆𝑡 ) 75 psi
Hydraulic fracture conductivity (𝐾𝐻𝐹 ) 250 mD-ft The NF shear dilation angle (𝜑𝑑𝑖𝑙 ) 3°
The rock compressibility (𝐶𝑟 ) 2.7-5 psi-1 The NF normal stiffness (𝐾𝑛 ) 9.76 psi/ft
The fluid compressibility (𝐶𝑙 ) 5.5-6 psi-1 The NF shear stiffness (𝐾𝑠 ) 4.86 psi/ft

54
Texas Tech University, Hao Zhang, August 2021

(a) Relative permeability curve (b) Capillary pressure curve


Figure 3.21 Geology parameters of the reservoir model

The SDP is regarded as the initial permeability in the following production


simulation. The reservoir for production simulation is a closed boundary, with an initial
pressure of 2500psi. The production well is controlled by a constant BHP of 150psi.
The total production time is 10 years. The economical parameters relative to NPV
calculation are listed in Table 3.3, where the cost of hydraulic fracturing per unit length
(ft) was set as 100$/ft according to the study of Schweitzer and Bilgesu (2009).

Table 3.3 Parameters of the reservoir model and their values

Parameters Value

The price of shale gas 3 $/MSCF


The cost of produced water disposal 5 $/STB
The cost of well drilling 250 $/ft
The cost of fracturing 100 $/ft
Annual discount rate 0.1

Two objectives are optimized and compared in this example, objective 1:


maximizing the SRV volume, and objective 2: maximizing the NPV. The Modified-
PSO algorithm is employed as the optimization algorithm, in which the particle number
is set as 10, and the total optimization step is defined as 50. The optimization variables
and their search space are listed in Table 3.4, we define that the hydraulic fractures share
the same spacing and half-length.

55
Texas Tech University, Hao Zhang, August 2021

Table 3.4 Optimization variables and search space

Optimization variables Search space

The fracture spacing (ft) [10, 300]


The fracture half-length (ft) [0, 500]
The well length (ft) [100, 1200]

The initial hydraulic fracturing well with the corresponding SDP field is shown
in Figure 3.22-a, in which the black line is the hydraulic fracture, the red line is the
horizontal well, and the background color represents the SDP (mD); besides, the
corresponding pressure distribution after the production simulation of 10 years is shown
in Figure 3.22-b. Additionally, the optimal hydraulic fracturing well with the
corresponding SDP for both of these two objectives can be seen in Figure 3.23 and
Figure 3.24, respectively. The variations of optimization variables are listed in Table
3.5 and Table 3.6, respectively. The accumulated shale gas productions for both of the
two optimized and initial wells during the production are shown in Figure 3.25. The
variations of SRV and NPV for these two objectives are shown in Figure 3.26.

(a) The hydraulically fractured well with SDP (mD)

(b) The pressure distribution after 10 years’ production (psi)


Figure 3.22 Initial hydraulically fractured well as well as the production result

56
Texas Tech University, Hao Zhang, August 2021

(a) The hydraulically fractured well with SDP (mD)

(b) The pressure distribution after 10 years’ production (psi)


Figure 3.23 Optimal hydraulically fractured well as well as the production result for
objective 1: maximizing SRV volume

(a) The hydraulically fractured well with SDP (mD)

(b) The pressure distribution after 10 years’ production (psi)


Figure 3.24 Optimal hydraulically fractured well as well as the production result for
objective 2: maximizing NPV

57
Texas Tech University, Hao Zhang, August 2021

Table 3.5 Values of optimization variables during optimization (Objective 1)

Optimization Values in Values in Values in Values in Values in


Variables Step 1 Step 5 Step 10 Step 20 Step 50

The frac spacing (ft) 68.23 67.47 63.43 59.56 28.06


The frac half-length (ft) 335.23 332.95 336.56 389.69 459.45
The well length (ft) 939.51 1199.94 931.40 1275.92 1184.33

Table 3.6 Values of optimization variables during optimization (Objective 2)

Optimization Values in Values in Values in Values in Values in


Variables Step 1 Step 5 Step 10 Step 20 Step 50

The frac spacing (ft) 68.23 54.46 31.40 34.53 35.90


The frac half-length (ft) 335.23 392.53 329.75 396.27 412.87
The well length (ft) 939.51 1097.90 1166.78 1160.89 1154.14

Figure 3.25 The accumulative shale gas production

58
Texas Tech University, Hao Zhang, August 2021

(a) The SRV variation (b) The NPV value variation


Figure 3.26 SRV and NPV comparisons for the two optimization objectives

As can be seen from the above results, after the optimization, both of these two
objectives lead to better productions than the initial fracturing well, but the optimization
results show a big difference. According to objective 1, we can get the maximum SRV
volume. However, the maximum SRV volume doesn’t mean the maximum revenue. As
is compared between Figure 3.25 and Figure 3.26, even though the larger SRV volume
brings about higher shale gas production, it also requires higher drilling and fracturing
cost, which inversely leads to a lower ultimate economic income. On the contrary,
choosing the NPV as the objective, all the economical investments will be taken into
consideration; although it brings less ultimate shale gas production, it leads to a more
economical revenue. So that NPV is the more reasonable optimization objective, which
will help to avoid the development risk.

3.3.2 Case 2: Multiple hydraulically fractured well optimization based on the effect
of SRV

In this case, a multi-well optimization proceeds. The reservoir model used in this
case is the same as that used in case 1, except for the reservoir dimension. The reservoir
model is enlarged to 1500ft×1000ft and subdivided into 600×100×1 blocks (Figure
3.27). The geological properties, fracturing, and production parameters are similar to
those used in case 1. Two hydraulic fracturing horizontal wells are drilled in this

59
Texas Tech University, Hao Zhang, August 2021

reservoir, and two types of fracturing placement methods are compared, which are type
1: aligning fracturing, and type 2: alternating fracturing. The hydraulic fracture spacing
and half-length, the horizontal well length, as well as the horizontal well spacing, are
chosen as the optimization variables, the searching space for which are listed in Table
3.7. The Modified-PSO algorithm is applied to search for the optimum value, with the
particle number defined as 10, and the total iteration time step is 50.

(a) Type1: aligning hydraulic fractures (b) Type2: alternating hydraulic fractures
Figure 3.27 Two types of multiple horizontal well placement

Table 3.7 Optimization variables and search space

Optimization variables Search space

The fracture spacing (ft) [10, 100]


The fracture half-length (ft) [0, 500]
The well spacing (ft) [10, 500]
The well length (ft) [100, 1200]

After optimization, the results for two fracturing types are listed as follows. The
initial and optimal wells together with the corresponding SDP for type 1, as well as the
pressure distribution after 10 years of production simulation, are compared in Figure
3.28 and Figure 3.29, respectively. The above simulation results for type 2 are shown in
Figure 3.30 and Figure 3.31, respectively. Furthermore, the change of optimization
variables during iterations for both fracturing types are listed in Table 3.8 and Table 3.9,
respectively. The accumulative shale gas production before and after the optimization

60
Texas Tech University, Hao Zhang, August 2021

for both the two fracturing types can be seen in Figure 3.32. Besides, the variation of
SRV volume, as well as NPV values for these two types, are shown in Figure 3.33.

(a) The hydraulically fractured well with SDP (mD)

(b) The pressure distribution after 10 years’ production (psi)


Figure 3.28 The initial hydraulically fractured well and the corresponding production
result (Type 1: aligning hydraulic fractures)

(a) The hydraulically fractured well with SDP (mD)

61
Texas Tech University, Hao Zhang, August 2021

(b) The pressure distribution after 10 years’ production (psi)


Figure 3.29 The optimal hydraulically fractured well and the corresponding production
result (Type 1: aligning hydraulic fractures)

(a) The hydraulically fractured well with SDP (mD)

(b) The pressure distribution after 10 years’ production (psi)


Figure 3.30 The initial hydraulically fractured well and the corresponding production
result (Type 2: alternate hydraulic fractures)

62
Texas Tech University, Hao Zhang, August 2021

(a) The hydraulically fractured well with SDP (mD)

(b) The pressure distribution after 10 years’ production (psi)


Figure 3.31 The optimal hydraulically fractured well and the corresponding production
result (Type 2: alternate hydraulic fractures)

Table 3.8 Values of optimization variables during optimization (Type 1)

Optimization Values in Values in Values in Values in Values in


Variables Step 1 Step 5 Step 10 Step 20 Step 50

The frac spacing (ft) 52.56 47.75 44.89 38.19 29.65


The frac half-length (ft) 162.37 161.49 161.03 170.48 178.67
The well spacing (ft) 401.19 392.34 440.28 445.83 449.59
The well length (ft) 921.79 1094.03 1162.13 1174.73 1156.82

63
Texas Tech University, Hao Zhang, August 2021

Table 3.9 Values of optimization variables during optimization (Type 2)

Optimization Values in Values in Values in Values in Values in


Variables Step 1 Step 5 Step 10 Step 20 Step 50

The frac spacing (ft) 52.56 55.04 53.53 51.47 45.11


The frac half-length (ft) 162.37 124.06 149.72 158.71 184.51
The well spacing (ft) 401.19 278.11 276.08 276.04 276.67
The well length (ft) 921.79 889.26 1052.45 1088.4 1146.89

(a) Type 1: aligning hydraulic fractures

(b) Type 2: alternating hydraulic fractures


Figure 3.32 The accumulative shale gas production

64
Texas Tech University, Hao Zhang, August 2021

(a) The SRV variation

(b) The NPV value variation


Figure 3.33 SRV and NPV variations during optimization

Combining the fracturing placement in Figure 3.29 and Figure 3.31, together
with the SRV volume in Figure 3.33-a, we can find that for the alternating fracturing
type, the hydraulic fractures for different wells across with each other, which stimulates
the NF and create more SRV in the cross area. Comparing with the aligning fracturing
type, the alternating fracturing type could create the same SRV volume with fewer
fracture numbers, which means less fracturing cost. The NPVs also show that, with less
fracturing cost, the alternating fracturing type leads to better ultimate revenue. These all
illustrate that alternating fracturing is better for multi-well design in shale gas reservoirs.

65
Texas Tech University, Hao Zhang, August 2021

CHAPTER IV

COMPLEX FRACTURE NETWORK SIMULATION AND


OPTIMIZATION
The above chapter numerically estimates the scale of SRV using fracturing
induced stress, and evaluates the improvement within the SRV using stress-dependent
permeability; however, this model converts the discrete fractures into continuous media,
which cannot precisely reflect the real geometry of SRV. The SRV is formed by a
complex fracture network (CFN), which has not been fully studied. How the hydraulic
fractures propagate and interact with natural fractures, as well as the geological and
operation parameters’ effect on the CFN is still not fully clear. So considering these
problems, this chapter aims to build up a CFN numerical model, investigate how the
CFN is constructed and affect by different geological and operational parameters, and
optimize the CFN under complex geological conditions. First, a hydraulic fracture (HF)
dynamic propagation model is established in Section 4.1 coupling the DDM and the
fluid flow functions, the rock deformation, and fluid flow in HF are fully considered.
Based on the HF propagation model the CFN propagation model is constructed in
Section 4.2, in which the power-law probability model is employed to describe the
natural fracture (NF) distribution, a crossing criterion is used to calculate the
interactions between the HFs and NFs. Furthermore, a CFN production simulation is
performed in Section 4.3, in which the embedded discrete fracture model (EDFM)
coupled with the shale gas mathematical model is adopted for the simulation. Then in
Section 4.4, a framework for CFN optimization is established with the net present value
(NPV) as the objective function to fully consider the economic viability. Finally, three
examples are presented in Section 4.5, a series of CFN propagations are simulated under
different NF distributions to analyze the geological and operational parameters’ effects
on the CFN configuration, then the CFN production in a shale gas reservoir is simulated
coupling the effects of geomechanics and fluid properties; and also, the CFN
optimization for multi-fracturing in multi-well is performed, the results of which are
compared and analyzed.

66
Texas Tech University, Hao Zhang, August 2021

4.1 Methodologies for complex fracture network propagation simulation

4.1.1 Hydraulic fracture dynamic propagating model

The displacement discontinuity method (DDM), together with the fluid flow
equation, is used to build up the CFN propagation model, which considers the influence
of the combined stress field distribution, propagation patterns, and criteria during the
fracturing treatments. The DDM has been illustrated in above Section 3.1.1, so it will
not be introduced in this section. Fracture propagation is driven by an incompressible
and non-Newtonian fluid. For a small unit in the hydraulic fracture, the amount of fluid
flow into the unit equals the sum of fluid flow out of the unit together with the unit
volume change, which is also named fluid flow balance (Figure 4.1,).

Figure 4.1 The fluid flow balance for the fracture unit

Fluid flow through a rock fracture is modeled by the Navier-Stokes equations of


fluid mechanics (Schlichting, 1968), together with the leak-off effect, the fluid flow
equation of hydraulic fracture is:

𝜕𝑞(𝑥,𝑡) 𝜕𝐴(𝑥,𝑡)
− 𝑞𝑙 (𝑥, 𝑡) − =0
{𝜕𝑝𝜕𝑥 64𝜇𝑞(𝑥,𝑡) 𝜕𝑡
(4-1)
=− 3
𝜕𝑥 𝜋ℎ𝑤

where 𝑡 is the fracture propagation time; 𝑥 is the distance along the fracture
propagation direction; 𝑞(𝑥, 𝑡) is the fracture internal flow rate; 𝜇 is the fluid viscosity;
𝐴(𝑥, 𝑡) is the cross-sectional area of the fracture; ℎ and 𝑤 are the fracture height and
width; 𝑞𝑙 (𝑥, 𝑡) is the fluid leak-off volume described by Carter function (Carter, 1957):
2ℎ𝐶𝑙
𝑞𝑙 (𝑥, 𝑡) = (4-2)
√𝑡−𝜏(𝑥)

67
Texas Tech University, Hao Zhang, August 2021

where 𝐶𝑙 is the fluid loss coefficient; 𝜏(𝑥) is the time at which fracture reaches
position 𝑥; ℎ is the fracture height. The Carter leak-off model assumes that leak-off is
one dimension in the direction perpendicular to the fracture plane.

The propagation criterion for a hydraulic fracture is dependent on the


propagation regime (Adachi et al., 2007), for which the propagation direction is affected
according to the maximum circumferential stress criterion (Figure 4.2).

Figure 4.2 Illustration of fracture propagation direction

The criterion states that a fracture starts propagating in the plane perpendicular
to the direction of greatest tension when the stress reaches a critical material constantly.
So the propagation steering angle (𝜃) satisfies the following equation:

𝐾𝐼 𝑠𝑖𝑛 𝜃 + 𝐾𝐼𝐼 (3 𝑐𝑜𝑠 𝜃 − 1) = 1 (4-3)

where 𝐾𝐼 is the stress intensity factor of mode I (opening mode), 𝐾𝐼𝐼 is the stress
intensity factor of mode II (shearing mode). The steering angle 𝜃 can be described as:

0 𝐾𝐼𝐼 = 0
𝐾𝐼 𝐾 2
𝜃= ⁄𝐾 −𝑠𝑔𝑛(𝐾𝐼𝐼 )√( 𝐼⁄𝐾 ) +8 (4-4)
𝐼𝐼 𝐼𝐼
2 𝑎𝑟𝑐𝑡𝑎𝑛 ( ) 𝐾𝐼𝐼 ≠ 0
4
{

When 𝐾𝐼 > 0 and 𝐾𝐼𝐼 = 0, the fracture propagation should proceed along 𝜃 =
0; when 𝐾𝐼 = 0 and 𝐾𝐼𝐼 ≠ 0, a pure shear in mode, the result is the strongest possible
kink in fracture direction along 𝜃 = ±75° for positive or negative 𝐾𝐼𝐼 , respectively (Wu,
2014). The stress intensity factors (𝐾𝐼 and 𝐾𝐼 ) can be calculated as (Olson, 2007):

68
Texas Tech University, Hao Zhang, August 2021

0.806𝐸 √𝜋
𝐾𝐼 = 𝐷 𝑂𝑝𝑒𝑛𝑖𝑛𝑔 𝑓𝑎𝑐𝑡𝑜𝑟
4(1−𝜈 2 )√2𝑎 𝑛
{ (4-5)
0.806𝐸 √𝜋
𝐾𝐼𝐼 = 𝐷 𝑆ℎ𝑒𝑎𝑟𝑖𝑛𝑔 𝑓𝑎𝑐𝑡𝑜𝑟
4(1−𝜈 2 )√2𝑎 𝑠

where 𝐸 is the young’s modulus; 𝐷𝑛 and 𝐷𝑠 are the normal and shear
displacement discontinuity; 𝑎 is the half element length; 𝑣 is the Poisson's ratio.

4.1.2 Hydraulic fracture interaction with natural fracture

The complexity of a CFN depends on the behavior of an HF when it encounters


NFs. If the HF propagates across an NF, the HF remains a planar fracture; else if the HF
dilates and propagates along with the NF, a CFN may result. It is important to determine
whether an HF crosses NFs under particular field conditions. In this paper, a power-law
probability NF distribution field is used (Davy, 1993; Segall and Pollard, 1983), the
detailed statement can be seen in Section 3.1.3.

Renshaw and Pollard (1995) proposed a crossing criterion for a propagating


fracture come across an orthogonal existing fracture, then Gu and Weng (2010)
extended the crossing criterion for a hydraulic fracture with any contacting angles of the
existing natural fracture. According to this criterion, NFs were assumed as frictional
interfaces without conductivity in the model. As shown in Figure 4.3. The intersection
angle between the HF and the interface is 𝛽.

Figure 4.3 Schematic of fracture approaching interface (Gu and Weng, 2010)

69
Texas Tech University, Hao Zhang, August 2021

In the above figure, 𝜏𝛽 and 𝜎𝛽𝑦 are the shear and normal stresses acting on the
interface; 𝜏𝛽 and 𝜎𝛽𝑦 can be calculated by projecting the stresses onto the interface (Gu
and Weng 2010; Wu, 2014):

𝜎𝐻 − 𝜎ℎ 𝐾𝐼 𝜃 𝜃 3𝜃
𝜏𝛽 = − 𝑠𝑖𝑛 2 𝛽 − 𝑐𝑜𝑠 ( ) 𝑠𝑖𝑛 ( ) 𝑐𝑜𝑠 (2𝛽 − )
2 √2𝜋𝑟𝑐 2 2 2
𝐾𝐼𝐼 3𝜃 𝜃 𝜃 3𝜃
− [𝑐𝑜𝑠 (2𝛽 − ) + 𝑠𝑖𝑛 ( ) 𝑐𝑜𝑠 ( ) 𝑠𝑖𝑛 (2𝛽 − )] (4-6)
√2𝜋𝑟𝑐 2 2 2 2

𝜎𝐻 − 𝜎ℎ 𝜎𝐻 − 𝜎ℎ 𝐾𝐼 𝜃 𝜃 3𝜃
𝜎𝛽𝑦 = − 𝑐𝑜𝑠 2 𝛽 + 𝑐𝑜𝑠 ( ) [𝑠𝑖𝑛 ( ) 𝑠𝑖𝑛 (2𝛽 − ) − 1]
2 2 √2𝜋𝑟𝑐 2 2 2
𝐾𝐼𝐼 𝜃 3𝜃 𝜃 𝜃 3𝜃
+ [𝑠𝑖𝑛 ( ) + 𝑠𝑖𝑛 (2𝛽 − ) − 𝑠𝑖𝑛 ( ) 𝑐𝑜𝑠 ( ) 𝑐𝑜𝑠 (2𝛽 − )] (4-7)
√2𝜋𝑟𝑐 2 2 2 2 2

where 𝛽 is the contact angle of the existing HF and the NF, 𝜎𝐻 , and 𝜎ℎ are the
maximum and minimum horizontal stress, respectively; 𝑟𝑐 is the critical radius, which
is calculated under two conditions 𝜃 = 𝛽 or 𝜃 = 𝜋 − 𝛽:
𝐾𝐼 𝜃
𝑟𝑐 = 𝑐𝑜𝑠 ( ) (4-8)
√2𝜋𝐾 2

Based on the above equations, slip along the interface will not occur whenever
(Gu and Weng 2010; Wu, 2014; Zeng and Yao, 2016):

|𝜏𝛽 | < 𝑆𝑜 + 𝜇𝜎𝛽𝑦 (4-9)

where 𝜇 is the coefficient of friction for the interface; 𝑆𝑜 is the cohesion of the
interface.

Gu indicated that the HF propagation intersecting with an NF is affected by the


strength of the NF, the relative angle between HFs and NFs, and the remote in-situ stress
ratio. The crossing or slip condition using the smaller stress intensity factor is taken as
the outcome of the criterion (Figure 4.4).

70
Texas Tech University, Hao Zhang, August 2021

Figure 4.4 Crossing conditions with various intersection angles (Gu and Weng 2010)

It can be seen that, as the angle decreases away from 90°, the difficulty for
crossing to occur increases because the stress ratio ≥ 1 becomes more and more
indistinct. In other words, the hydraulic fracture is more likely to turn and propagate
along the natural fracture.

4.1.3 Workflow of complex fracture network propagation simulation

Based on the rock deformation equation, the fluid flow equation, and the fracture
crossing criterion, the coupled numerical model for CFN propagation is constructed. An
iteratively coupled solution procedure is used to solve the coupled model, in which the
rock deformation equation is first solved, and then the fluid flow equation is solved, the
results of the two equations are compared cyclically until the coverage is satisfied.

Figure 4.5 Schematic of fracture dynamic propagation

71
Texas Tech University, Hao Zhang, August 2021

The CFN propagation is simulated by combining the HF propagation model with


the power-law probability NF pattern and solved using an iteratively coupled solution
procedure, the workflow of which is introduced in detail as follows, and the
corresponding flowchart is shown in Figure 4.6. Examples of CFN propagation
simulation based on this model can be seen in Section 4.1.4.

(1) At the first simulation time step, generate the initial frac-element of the HF at the
injection point and set an assumed time increment 𝑑𝑡 and an assumed fluid pressure
𝑃0 . Combine the initial HF with the power-law probability NF pattern generated by
Eq.(3-13).

(2) Put 𝑃0 into Eq. (3-3) to obtain the boundary stress condition 𝜎𝑛 , 𝜎𝑠 , and then solve
the DDM function in Eq.(3-1) to get the induced strains 𝐷𝑛 , 𝐷𝑠 .

(3) Bring the induced strains 𝐷𝑛 , 𝐷𝑠 into Eq.(4-1) to calculate the fluid flow function
and obtain the fluid pressure 𝑃𝑘 , and regard the former fluid pressure as 𝑃𝑘−1 . If 𝑝𝑘
and 𝑝𝑘−1 don’t converge, update the new fluid pressure 𝑃𝑘+1 according to:

𝑃𝑘+1 = 𝛼𝑝 𝑃𝑘 + (1 − 𝛼𝑝 )𝑃𝑘−1 (4-10)

where 𝛼𝑝 is the coefficient of convergence, which is set as 0.4 in this paper.

Then go back to step 2, recalculate the DDM function based on the new fluid
pressure 𝑝𝑘+1 repeat this process until the convergence of fluid pressure 𝑝𝑘 and
𝑝𝑘−1 is reached.

(4) Correct the time increment 𝑑𝑡 based on the fluid balance equation. The entire
volume of fluid injected is equal to the sum of fluid in the HF and the fluid leaks
into the surrounding rock, so the 𝑑𝑡 can be calculated by (Wu, 2014):
𝑁 𝑁 2ℎ𝐶𝑙 𝑥𝑖
𝑄𝑇 (𝑡)(𝑡 + 𝑑𝑡) − ∑1 𝑓 ∑𝑁 𝑓 𝑁 𝑡
𝑖=1 ℎ𝑤𝑖 𝑥𝑖 − ∑1 ∑𝑖=1 ∑𝑘=1 =0 (4-11)
√𝑡−𝜏𝑖𝑘

where 𝑁𝑓 is the total HF number in the CFN, and 𝑁 mean the entire frac-element
number in the 𝑖 th HF; ℎ, 𝑤𝑖 𝑥𝑖 represents the height, width, and length of frac-

72
Texas Tech University, Hao Zhang, August 2021

element 𝑖 , respectively; 𝑡 is the total injection time at the beginning of this


simulation time step; 𝑄𝑇 (𝑡) is the injection rate at time 𝑡.

(5) Start a new simulation time step, if the 𝑖 th HF could continue to propagate, add a
new frac-element on the tip of the HF (Figure 4.2), with the direction along with
the propagation steering angle 𝜃 calculated by Eq.(4-4).

(6) Check if the new frac-element encounters any pre-existing NFs, if yes, check if this
frac-element could open the encountering NF according to the crossing criterion in
Eq.(4-9), and then correct its propagating direction. If the new frac-element could
open both sides of the NF, add a new HF to this CFN with the fracture bottom
located at the intersection point.

(7) Repeat steps 2 to 6 until reaching the designed injection time, the final CFN is
obtained.

Start

A new time step

NF crossing
Add a new element
criteria

Rock deformation
DDM equation P0
calculation

Fluid flow equation


Pk+1=αPk+(1-α)Pk-1

No
Converged?

Yes

Mass balance equation

No Time
finished?
Yes
Output final CFN

Figure 4.6 Workflow for fracture network propagation

73
Texas Tech University, Hao Zhang, August 2021

4.1.4 A case for hydraulic fracture propagation and natural fracture encountering
simulation

Based on the hydraulic fracture propagation simulation methods mentioned


above, a hydraulic fracture propagation numerical model is established. First of all, a
square reservoir domain is defined with a dimension of 500ft×500ft in the horizontal
view and a dimension of 50ft in the vertical view. The in-situ stress distribution within
this domain is under an anisotropic condition, for which the maximum horizontal stress
along the y-direction and the maximum horizontal stress is the x-direction. Other
geological parameters are listed in Table 4.1. The Matlab software is used to construct
the CFN propagation and production numerical model.

Table 4.1 Geological and fluid parameters for the reservoir

Parameters Value

Minimal horizontal stress (𝜎ℎ ) 4500 psi


Maximal horizontal stress (𝜎𝐻 ) 5000 psi
Poisson’s Ratio (𝑣) 0.25
Young’s Modulus (𝐸) 6.56 psi
Fluid viscosity (𝜇) 1 cp
Fluid leak-off coefficient (𝐶𝐿 ) 1-5 ft/min0.5
The rock compressibility (𝐶𝑟 ) 2.7-5 psi-1

First, the hydraulic fracture propagation is simulated. To testify the fracture


interaction, two fractures are generated in the domain and placed as two types, for type
1: parallel fracturing placement (Figure 4.7-a), and case 2: zipper fracturing placement
(Figure 4.7-b). The fracture distance for case 1 and case 2 are the same 100ft, and the
wellbore distance for case 2 is 200ft. The injection rate is 20bbl/min, and the injection
time is 30min. All the numerical simulation framework is constructed using Matlab
software.

74
Texas Tech University, Hao Zhang, August 2021

(a) Parallel fracturing placement (b) Zipper fracturing placement


Figure 4.7 Two types of fracturing placement

After the simulation, the hydraulic fracture together with the induced stress for
the two cases and shown in Figure 4.8 respectively, where the black curves represent
the hydraulic fractures and the color in the field represent the induced stress. As can be
seen that, because of induced stress at the fracture tips caused by mechanical interaction,
the two parallel fractures in case 1 separate apart, and the zipper fractures in case 1 grow
toward each other.

(a) Parallel fracturing placement (b) Zipper fracturing placement


Figure 4.8 Hydraulic fracture propagation path and the corresponding induced stress
map (psi)

75
Texas Tech University, Hao Zhang, August 2021

The model for HF propagation and interaction with the NF is simulated. The
reservoir domain is the same as that used in the above simulation, a HF propagates from
the bottom of the reservoir domain and encounters different oriented NFs. The injection
is set as 20bbl/min and will stop when the HF reaches 300ft in the y-axis.

(a) No NF (b) NF contact angle: 30°

(c) NF contact angle: 50° (d) NF contact angle: 70°


Figure 4.9 The propagating HF and interaction with different oriented NF and the
induced stress map (psi)

After the simulation, the results of HF propagation and encountering different


oriented NFs are shown in Figure 4.9, where the black curve and the blue segment
represent the HF and NF, respectively, and the background color shows the intensity of

76
Texas Tech University, Hao Zhang, August 2021

the induced stress field (psi). It can be seen that the HF causes high induced stress on
the two sides and low induced stress around the frac-tip. Besides, when encountering
the NF with small contact angles, the HF can only stimulate and open the side of the NF
that along with the HF propagation direction; however, as the contact angle enlarges,
the HF will be more likely to open the two sides of the NF, this will help to create more
frac-branches to form the CFN. The simulation results of HF propagation and
interaction with the NF show good agreement with the former scholars’ relative studies
(Wu et al., 2012; Zeng and Yao, 2016; Ren et al., 2018).

4.2 Complex fracture networks production simulation

4.2.1 The embedded discrete fracture model

To evaluate the effect of the fracture network obtained from the numerical
simulation, fracture production numerical simulation is needed. However, the fracture
is a kind of discrete media, which is hard to be recognized and simulated by the
traditional reservoir simulating method. To solve this problem, scholars have developed
several fractured reservoir production simulation methods. The traditional fracture
treatment is local grid refinement (LGR), it subdivides the reservoir with structured
grids and uses refined grids to represent the fracture. This method is theoretically
concise and easy to be used in the commercial numerical simulator. However, this
method is not suitable for complex fracture networks, fractures presented by LGR can
only along the direction of the original reservoir grid, for inclined fractures, the
refinement gridding will be very complex (Figure 4.10).

Figure 4.10 Local grid refinement (LGR) method

77
Texas Tech University, Hao Zhang, August 2021

Facing the drawbacks of LGR, a discrete fracture model (DFM) is proposed to


deal with complex to simulate fracture production. For DFM, the fracture model is first
constructed, and the unstructured grid is adopted to subdivide the reservoir according to
the orientation of the fractures to fit the complex fractures (Figure 4.11-b), it is very
adaptive to arbitrarily distributed fractures. However, when it comes to the complex
fracture network, the grids generated by DFM will be massive and complex and the
gridding error frequently occurs.

(a) Fracture network (b) DFM (c) EDFM


Figure 4.11 Comparison of DFM and EDFM models

To avoid the drawbacks of traditional LGR and DFM methods, Li and Lee (2008)
developed a novel embedded discrete fracture model (EDFM), and later be extended to
deal with arbitrarily distributed fractures (Moinfar et al., 2013). For EDFM, structured
grids are first generated in the reservoir, and then the fractures are embedded into these
grids (Figure 4.11-c). The EDFM could perfectly avoid the drawbacks of the LGR and
DEM but take the advantage of both of them, the reservoir grids are independent of the
fracture geometry and the reservoir can be subdivided by regular structured grids, so the
gridding and the simulation efficiency is significantly improved and the complicate
fracture distribution could be precisely described. The EDFM is used in the paper to
simulate the CFN production.

4.2.2 The principle of embedded discrete fracture model

The basic principle of EDFM is represented in Figure 4.12. Firstly, the matrix in
the physical domain is divided into structured cells, then the fractures that penetrate

78
Texas Tech University, Hao Zhang, August 2021

matrix cells will be divided into several segments by the cell boundaries (Figure 4.12-
a). Secondly, the cells for both matrix and fractures are added to the computational
domain; as can be seen that some neighboring cells in the physical domain will be
separate in the computational domain, so the none-neighboring connections (NNC) are
defined in the computational domain to connect these cells (Figure 4.12-b). Additionally,
a well perforation point will be added to the matrix and fracture cells that are connected
to the wellbore.

(a) The physical domain of EDFM (b) The computational domain of EDFM
Figure 4.12 Schematic diagram of the EDFM (Xu, 2015)

In addition to the fluid flow between the neighboring matrix cells, EDFM also
considers the flow rate between NNC, which is calculated as:

𝑞 = 𝑇𝑁𝑁𝐶 × ∆𝑃 (4-12)
where ∆𝑃 is the pressure difference between NNC; 𝑇𝑁𝑁𝐶 is the transmissibility
factor of the NNC.

There are three NNC types in the EDFM, the first is the matrix-fracture
connection, and the transmissibility factor 𝑇𝑚−𝑓 is (Xu, 2015):

̿ ∙𝑛
2𝐴𝑓 (𝐾 ⃗ )𝑛

𝑇𝑚−𝑓 = (4-13)
𝑑𝑚−𝑓

̿ is the tensor of
where 𝐴𝑓 is the area of the fracture segment on one side; 𝐾
matrix permeability; 𝑑𝑚−𝑓 is the average normal distance from the matrix to the fracture;
𝑛⃗ is the normal vector of the fracture plane.

79
Texas Tech University, Hao Zhang, August 2021

The second is the fracture-fracture connection, the transmissibility factor 𝑇𝑓−𝑓


is (Moinfar et al., 2013):
𝑇1 𝑇2 𝑘𝑓 𝐴𝑐 𝑘𝑓 𝐴𝑐
𝑇𝑓−𝑓 = ; 𝑇1 = , 𝑇2 = (4-14)
𝑇1 +𝑇2 𝑑𝑠𝑒𝑔1 𝑑𝑠𝑒𝑔2

where 𝐴𝑐 is the intersection area of frac-segment 1 and 2; 𝑘𝑓 is the fracture


permeability; 𝑑𝑠𝑒𝑔1 and 𝑑𝑠𝑒𝑔2 give the distances between the intersection and centroids
of frac-segment 1 and 2, respectively.

The third is the fracture-well connection, the transmissibility factor 𝑊𝐼𝑓 is:

2𝜋𝑘𝑓 𝑤𝑓
𝑊𝐼𝑓 = (4-15)
ln (𝑟𝑒 ⁄𝑟𝑤 )

where 𝑤𝑓 is the aperture of the fracture; 𝑟𝑤 is the radius of the production well;
𝑟𝑒 is the radius of the reservoir, which is usually calculated by transforming the reservoir
into a round shape with the same area.

4.2.3 Shale gas flow model

Together with the EDFM, a shale gas flow model in the fractured porous media
is established to realize the CFN production simulation in the shale reservoir. For this
model, we consider only the gas flow, which is regarded as the free phase in the fracture,
and as both free and adsorbed phases in the matrix.

In the matrix system, considering the adsorption and transport mechanisms, the
equation to describe the shale gas flow is (Wang, 2020):

𝜕 ∏𝑖 𝐹𝑎𝑝𝑝,𝑖 𝑘𝑚0
(𝜌𝑔 ∅𝑚 + (1 − ∅𝑚 )𝑚𝑎𝑑 ) + ∇ ∙ (−𝜌𝑔 ∇𝑝𝑚 ) = 𝑞𝑚𝑓 (4-16)
𝜕𝑡 𝜇𝑔

In the fracture system, considering only transport mechanisms, the equation to


describe the shale gas flow is (Wang, 2020):

𝜕 ∏𝑖 𝐹𝑎𝑝𝑝,𝑖 𝑘𝑓0
(𝜌𝑔 ∅𝑓 ) + ∇ ∙ (−𝜌𝑔 ∇𝑝𝑓 ) = −𝑞𝑚𝑓 + 𝑞𝑤 (4-17)
𝜕𝑡 𝜇𝑔

80
Texas Tech University, Hao Zhang, August 2021

where 𝜌𝑔 and 𝜇𝑔 are the mass density and viscosity of the gas, respectively;
𝑚𝑎𝑑 is the storage term caused by adsorption; ∅𝑚 and ∅𝑓 are the matrix and fracture
porosities, respectively; 𝑝𝑚 and 𝑝𝑓 are the matrix and fracture pressures, respectively;
𝑘𝑚0 and 𝑘𝑓0 are the matrix and fracture absolute permeabilities, respectively; 𝐹𝑎𝑝𝑝,𝑖 is
the 𝑖 th permeability correction factor for a specific shale gas transport mechanism; 𝑞𝑚𝑓
represents the gas flow between the matrix and the fracture, and 𝑞𝑤 is the source/sink
term; 𝑡 is the production time.

In the matrix system, the adsorption mechanism together with the shale gas
slippage and diffusion effect is considered. The adsorption mechanism is calculated by
the Langmuir isotherm equation (Olsen and Watanabe, 1957):
𝑝𝑉𝐿
𝑚𝑎𝑑 = 𝜌𝑅 𝜌𝑔𝑠 (4-18)
𝑝+𝑃𝐿

And the apparent permeability of shale gas flow affected by the slippage and
diffusion effect in the matrix system can be implemented as (Liu and Reynolds, 2019):
4𝐾𝑛
𝑘𝑚 = 𝑘𝑚0 𝐹𝑎𝑝𝑝 = (1 + 𝛼𝑘 𝐾𝑛 ) (1 + ) (4-19)
1+𝐾𝑛

where 𝜌𝑔𝑠 is the gas molar density at the standard condition and 𝜌𝑅 is the rock
bulk density; 𝑘𝑚0 and 𝑘𝑚 are the intrinsic and apparent permeabilities of the matrix; 𝑃𝐿
and 𝑉𝐿 are the Langmuir pressure and Langmuir volume, respectively; 𝐾𝑛 and 𝛼𝑘 are
the Knudsen number and rarefaction parameter, respectively, which can be calculated
by (Wang, 2020):

𝜇𝑔 𝜋𝑅𝑇∅
𝐾𝑛 = √2𝑀𝑘
2.8284𝑝
{ 𝑚0 (4-20)
128
𝛼𝑘 = 𝑡𝑎𝑛−1 (4𝐾𝑛0.4 )
15𝜋2

The dynamic fracture permeability for shale gas flow will be decreased by
fracture closure deal to the pressure and geo-stress change, which can be described as:
𝐹𝑐𝑑 (𝑝)
𝑘𝑓 = 𝑘𝑓0 𝐹𝑎𝑝𝑝 = 𝑘𝑓0 (4-21)
𝐹𝑐𝑑 (𝑝0 )

81
Texas Tech University, Hao Zhang, August 2021

where, 𝑘𝑓0 and 𝑘𝑓 are the intrinsic and apparent permeabilities of the fracture;
𝐹𝑎𝑝𝑝 is the permeability correction factor; 𝐹𝑐𝑑 (𝑝0 ) and 𝐹𝑐𝑑 (𝑝) are the normalized
fracture conductivity under the initial and current reservoir conditions, the formulation
can be seen in Alramahi and Sundberg (2012) and Wu et al. (2019b) respectively.

4.3 The complex fracture network optimization

In this section, based on the CFN propagation and production simulation models
mentioned above, a framework for CFN optimization is established. The net present
value (NPV) is adopted as the objective function, and a modified neural network
algorithm (M-NNA) is proposed and employed as the optimization algorithm.

4.3.1 Optimization objective function

For reservoir engineering problems, the ultimate optimization destination is


usually achieving the maximum economical recovery with the lowest cost. Same as
Section 3.2.1, the NPV is chosen as the objective function, but the formulation is a little
bit different. The NPV value 𝐽𝑜𝑏 of CFN optimization can be described as:

𝑁 (𝑟𝑔 ⋅𝑄𝑔𝑛 )×𝛥𝑡𝑛 𝑁 𝑁


𝑡𝑁 𝑝 ,𝑗 𝑝 𝑓
𝐽𝑜𝑏 = ∑𝑛=1 {∑𝑗=1 𝑡𝑛 /365 } − ∑𝑗=1 (𝐿𝑤𝑗 ⋅ 𝐶𝑤 + 𝑟𝑓 ∑𝑖=1 𝑄𝑓𝑖 ) (4-22)
(1+𝑏)

where, 𝑁𝑡 is the total simulation time step of the CFN production; 𝑡 is the total
simulation time, and 𝛥𝑡𝑛 is the length of the nth simulation time step; 𝑏 is the annual
𝑛
discount rate; 𝑟𝑔 is the gas price, and 𝑄𝑔,𝑗 represents the gas production rate of the 𝑗th
well in the 𝑛th simulation time step, respectively; 𝑁𝑝 is the number of production wells,
and 𝑁𝑓 is the number of HFs on the 𝑗th well; 𝐿𝑤𝑗 is the length of the 𝑗th well, and 𝐶𝑤 is
the drilling cost per unit well length; 𝑟𝑓 is the cost of fracturing fluid per unit volume,
and 𝑄𝑓𝑖 is the volume of fracturing fluid injected into the 𝑖 th HF in the 𝑗th well.

4.3.2 The neural network algorithm

The M-PSO proposed in Section 3.2.3 is an excellent heuristic algorithm and


presents good performance for optimization, but still in the traditional architecture and

82
Texas Tech University, Hao Zhang, August 2021

difficult to adapt to complex optimization problems. Nowadays, with the fierce


development of artificial intelligence, a lot of novel machine learning algorithms were
proposed and applied to deal with complex optimization problems in a wide variety of
research fields and presented a great superiority over the traditional algorithms
(Tripoppoom et al., 2020). Facing the great superiority of machine learning algorithms
over the traditional optimization algorithms in dealing with complex optimization
problems, a state-of-the-art neural network algorithm (NNA) is introduced as the
optimization algorithm in this study. Recently proposed by Sadollah et al. (2018), the
NNA is a kind of unsupervised machine learning algorithm (Hinton et al. 1999), it
searches for the target value by adjusting the weight values of connections between the
elements of the input layer and hidden layer, without prior labeled data.

(a) Schematic view of NNA

(b) Generating new pattern solution of NNA


Figure 4.13 Illustration of NNA (Sadollah et al., 2018)

83
Texas Tech University, Hao Zhang, August 2021

The workflow for NNA optimization searching is listed as follows:

(1) First, the pattern solution 𝑋 = (𝑥1 , 𝑥2 , 𝑥3 ⋯ 𝑥𝐷 ) is defined based on the


optimization variable 𝑥𝑖 , and 𝐷 represents the total number of optimization
variables. At the beginning of the optimization, 𝑁𝑝𝑜𝑝 initial pattern solutions are
randomly generated in the searching domain to form a 𝑁𝑝𝑜𝑝 × 𝐷 solution matrix
for the input layer in Figure 4.13-a.

𝑋1 𝑥11 , 𝑥21 , 𝑥31 ⋯ 𝑥𝐷1


𝑋2 𝑥12 , 𝑥22 , 𝑥32 ⋯ 𝑥𝐷2
𝐗=[ ⋮ ]= (4-23)
⋱ ⋮
𝑋𝑁𝑝𝑜𝑝 𝑁𝑝𝑜𝑝 𝑁𝑝𝑜𝑝 𝑁
[𝑥 ,𝑥1 ⋯ 𝑥 𝑝𝑜𝑝 ]
2 𝐷

At the same time, the initial weights for the connections between the input layer
and the hidden layer are randomly defined, a 𝑁𝑝𝑜𝑝 × 𝑁𝑝𝑜𝑝 weight matrix is
constructed:
𝑤11 , 𝑤21 , 𝑤31 ⋯ 𝑤𝑁𝑝𝑜𝑝 1
𝑤12 , 𝑤22 , 𝑤32 ⋯ 𝑤𝑁𝑝𝑜𝑝 2
𝐖(𝑡) = [𝑊1 , 𝑊2 , 𝑊2 ⋯ 𝑊𝑁𝑝𝑜𝑝 ] = ⋱ ⋮ (4-24)
[𝑤1𝑁𝑝𝑜𝑝 , 𝑤2𝑁𝑝𝑜𝑝 , ⋯ 𝑤𝑁𝑝𝑜𝑝 𝑁𝑝𝑜𝑝 ]

where the summation of weights for a pattern solution should be one:


𝑵𝒑𝒐𝒑
∑𝒋=𝟏 𝑤𝑖𝑗 = 1, 𝑖 = 1,2,3 ⋯ 𝑁𝑝𝑜𝑝 (4-25)

(2) During the optimization, pattern solutions 𝑋𝑗 (𝑡 + 1) for the next iteration (𝑡 + 1)
are calculated using pattern solutions in the current iteration (𝑡), each element 𝑋𝑗𝑛𝑒𝑤
in the hidden layer is calculated by the weighted sum of all the pattern solutions
𝑋𝑗 (𝑗 = 1,2,3 ⋯ 𝑁𝑝𝑜𝑝 ) in the input layer (Figure 4.13-b).

𝑁
𝑋𝑗𝑛𝑒𝑤 (𝑡 + 1) = ∑𝑖=1
𝑝𝑜𝑝
𝑤𝑖𝑗 (𝑡) × 𝑋𝑖 (𝑡) , 𝑗 = 1,2,3 ⋯ 𝑁𝑝𝑜𝑝
{ (4-26)
𝑋𝑗 (𝑡 + 1) = 𝑋𝑗 (𝑡) + 𝑋𝑗𝑛𝑒𝑤 (𝑡 + 1), 𝑗 = 1,2,3 ⋯ 𝑁𝑝𝑜𝑝

(3) After obtaining new pattern solutions, the best pattern solution is regarded as the
“target solution” and the corresponding weight value is regarded as the “target

84
Texas Tech University, Hao Zhang, August 2021

weight”, the weight matrix is updated according to the “target weight”:

𝑤𝑖 (𝑡 + 1) = 𝑤𝑖 (𝑡) + 2 × 𝑟𝑎𝑛𝑑 × (𝑤 𝑡𝑎𝑟𝑔𝑒𝑡 (𝑡) − 𝑤𝑖 (𝑡)), 𝑖 = 1,2, ⋯ 𝑁𝑝𝑜𝑝 (4-27)

During the iteration, the summation of new weights for a pattern solution should
always be one (Eq.(4-25)).

(4) After each iteration, a bias factor 𝛽 is applied to further adjust the pattern solution
away from the premature convergence. The bias operator 𝛽 initially equal to 1, and
will reduce during the iteration:

𝛽(𝑡 + 1) = 𝛽(𝑡) × 0.9, 𝑖 = 1,2, ⋯ 𝑀𝑎𝑥_Iteration (4-28)

Then a random value is set for every pattern solution, if the random value is larger
than the bias factor 𝛽, the corresponding pattern solution will be transferred towards
the target solution according to the following equation:

𝑋𝑖∗ (𝑡 + 1) = 𝑋𝑖 (𝑡 + 1) + 2 ∙ 𝑟𝑎𝑛𝑑(𝑋𝑡𝑎𝑟𝑔𝑒𝑡 (𝑡) − 𝑋𝑖 (𝑡 + 1)), 𝑖 = 1,2, ⋯ 𝑁𝑝𝑜𝑝 (4-29)

(5) Back to step 2, repeat this process until the target solution converges.

4.3.3 The modified neural network algorithm

The NNA is proved to be with higher calculation accuracy comparing with the
traditional optimization algorithms (Sadollah et al., 2018; AbouOmar et al., 2019);
however, there are still drawbacks. The NNA starts the optimization searching from the
random initial pattern solutions, which show a significant influence for the ultimate
result; bad initial pattern solutions possess a high risk for the ultimate result trapped in
local extremes. Besides, the bias operation transfers the selected pattern solutions
towards the “target solution”; however, if the “target solution” locates in the local
extreme, it will attract other solutions to trap in the local extremes. Facing these
drawbacks, we propose a modified neural network algorithm (M-NNA). We found that
in the NNA the “target solution” is always found by pattern solutions with better values,
while the worst pattern solutions seldom have a chance to become the “target solution”,
in other words, the calculation for the worst pattern solutions is wasted. So we modify

85
Texas Tech University, Hao Zhang, August 2021

the bias operation in step (4). before the bias operation, the pattern solutions are sorted
from better to worse, then a set of new random values 𝑋 and random weight values 𝑊
are reassigned to the worst pattern solutions:

𝑋𝑖𝑤𝑜𝑟𝑠𝑡 (𝑡 + 1) = 𝐿𝐵 + (𝑈𝐵 − 𝐿𝐵) × 𝑟𝑎𝑛𝑑, 𝑖 = 1,2, ⋯ 𝑁𝑤𝑜𝑟𝑠𝑡


{ 𝑤𝑜𝑟𝑠𝑡 (4-30)
𝑊𝑖 (𝑡 + 1) = 𝑟𝑎𝑛𝑑, 𝑖 = 1,2, ⋯ 𝑁𝑤𝑜𝑟𝑠𝑡

where 𝐿𝐵 and 𝑈𝐵 represent the lower and upper searching boundaries; 𝑟𝑎𝑛𝑑 is
a random number within (0,1); 𝑁𝑤𝑜𝑟𝑠𝑡 is the number of the worst solutions to be
selected, by trial and error, 𝑁𝑤𝑜𝑟𝑠𝑡 = 0.3×𝑁𝑝𝑜𝑝 shows the best optimization effect.

Algorithm 1: Pseudocode of the modified neural network algorithm (M-NNA)


1: Initialization Randomly define 𝑁𝑝𝑜𝑝 × 𝐷 initial pattern solutions 𝑋 according to Eq. (4-23), and 𝑁𝑝𝑜𝑝 × 𝑁𝑝𝑜𝑝
initial weights 𝑊 according to Eq. (4-24), adjust the summation of weights according to Eq. (4-25)
2: While The stopping criterion is not met
3: Update the pattern solutions 𝑋𝑡+1 according to Eq. (4-26)
4: Select the target pattern solution and update the weights 𝑊𝑡+1 according to Eq. (4-27)
𝑊𝑜𝑟𝑠𝑡 𝑊𝑜𝑟𝑠𝑡
5: Pickup the 𝑁𝑤𝑜𝑟𝑠𝑡 worst pattern solutions 𝑋𝑡+1 and the corresponding weights 𝑊𝑡+1
6: For 𝑖 = 1 to 𝑁𝑤𝑜𝑟𝑠𝑡 %---------------The modification to NNA (shown in Section 4.3.3) ---------------
𝑤𝑜𝑟𝑠𝑡 (𝑖)
7: 𝑋𝑡+1 = 𝐿𝐵 + (𝑈𝐵 − 𝐿𝐵) × 𝑟𝑎𝑛𝑑 % 𝑈𝐵 and 𝐿𝐵 are the upper and lower bounds of pattern solutions
𝑤𝑜𝑟𝑠𝑡 (𝑖)
8: 𝑊𝑡+1 = 𝑟𝑎𝑛𝑑
9: End For
10: Define the bias factor 𝛽
11: For 𝑖 = 1 to 𝑁𝑝𝑜𝑝
12: Define a random value 𝑟𝑎𝑛𝑑
13: If 𝑟𝑎𝑛𝑑 ≤ 𝛽 %%---------------Bias operator for pattern solution & weight matrix ---------------
14: 𝑁𝑏 = 𝑅𝑜𝑢𝑛𝑑(𝐷 × 𝛽) % 𝑁𝑏 :Number of biased variables in a population of pattern solution
15: For 𝑗 = 1: 𝑁𝑏
16: 𝑋𝑡+1 (𝑗, 𝐼𝑛𝑡𝑒𝑔𝑒𝑟 𝑟𝑎𝑛𝑑[0, 𝐷]) = 𝐿𝐵 + (𝑈𝐵 − 𝐿𝐵) × 𝑟𝑎𝑛𝑑
17: End For
18: 𝑁𝑤𝑏 = 𝑅𝑜𝑢𝑛𝑑(𝑁𝑝𝑜𝑝 × 𝛽) % 𝑁𝑤𝑏 :Number of biased variables in the updated weight matrix
19: For 𝑗 = 1: 𝑁𝑤𝑏
20: 𝑊𝑡+1 (𝑗, 𝐼𝑛𝑡𝑒𝑔𝑒𝑟 𝑟𝑎𝑛𝑑[0, 𝑁𝑝𝑜𝑝 ]) = 𝑈(0,1)
21: End For
22: Else %%---------------Transfer function operator----------------
23: Transfer the pattern solution 𝑋𝑡+1 (𝑖) according to Eq. (4-29)
24: End IF
25: End For
26: 𝛽 = 𝛽 × 0.9
27: 𝑡 = 𝑡 + 1
28: End While
29: Output the optimum pattern solution

86
Texas Tech University, Hao Zhang, August 2021

The advantage of the modification is that it gives the worst pattern solutions a
chance to become better pattern solutions, or even become the “target solution”. Besides,
since random values are continually assigned to the pattern solutions throughout the
optimization, the risk of traping in local extremes caused by bad initial pattern solutions
or bad “target solution” significantly reduced. The reliability of the M-NNA algorithm
is testified by the above multi-extreme benchmark function in Eq.(3-21), The NNA and
M-NNA algorithms are testified, together the PSO and M-PSO introduced in Section
3.2, as a comparison. The optimization is performed 100 times independently with
random initial values, in each optimization, the pattern solution population 𝑁𝑝𝑜𝑝 is
defined as 10 and the total step for optimization iteration is set as 50. The statistic results
for the 100 times of optimization by different optimization algorithms are counted and
displayed in Figure 4.14.

(a) PSO (b) M-PSO

(c) NNA (d) M-NNA


Figure 4.14 Distribution of the 100 optimization results for different algorithms

87
Texas Tech University, Hao Zhang, August 2021

As shown in Figure 4.14, the PSO shows the worst optimization accuracy, for
only 28% of the optimizations reach the true optimal value, yet most optimizations are
trapped in the local extremes, which reveals a huge optimization risk. The M-PSO
performs better than the PSO with 54% optimizations reach the true optimal value. The
NNA presents far better performance with 80% accuracy, which reveals the superiority
of the machine learning algorithm over the traditional algorithms. Whereas the M-NNA
shows the best optimization ability, the accuracy of which surpasses the NNA by 9%
percent, reaches 89%; the rest of the optimization results are also located in the near-
optimal values. These statistic plots demonstrate the reliability and superiority of the M-
NNA.

4.3.4 The workflow of complex fracture network optimization

Based on the above CFN propagation and production model, together with the
NPV in Eq.(4-22) as the objective function, the CFN optimization framework is
constructed. The M-NNA algorithm is adopted as the optimization algorithm. Some
hydraulic fracturing operation parameters are selected as the optimization variables and
they form the pattern solution. The detailed workflow for CFN optimization is described
as follows and the flowchart can be seen in Figure 4.15.

(1) At the beginning of optimization, set a series of random particles 𝐗 =


(𝑋1 , 𝑋2 ⋯ 𝑋𝑁𝑝𝑜𝑝 ), which contains the parameters to be optimized.
(2) Based on the CFN propagation model in Section 4.1, perform the CFN propagation
simulation and obtain the CFN model for each pattern solution 𝑋𝑖𝑘 , where 𝑘
represents the current iteration step.
(3) Put the obtained CFN models in the EDFM in Section 4.2 for production simulation
and record the simulation result for every pattern solution 𝑋𝑖𝑘 .
𝑘
(4) According to the simulation results, calculate the NPV value 𝐽𝑜𝑏,𝑖 for each pattern
solution 𝑋𝑖𝑘 using Eq. (4-22), the pattern solution with the best NPV is regarded as
the target solution 𝑋 𝑡𝑎𝑟𝑔𝑒𝑡 (𝑘); then update the pattern solution 𝑋𝑖𝑘+1 according to
the M-PSO in Section 4.3.2.

88
Texas Tech University, Hao Zhang, August 2021

(5) Back to step 2, re-construct the CFN model based on the new pattern solution 𝑋𝑖𝑘+1 ,
then re-run the EDFM production simulation to calculate the corresponding new
𝑘+1
NPV value 𝐽𝑜𝑏,𝑖 .
(6) Repeatedly perform steps 2-5 until the designed iteration step is over or the target
solution converges.
Start

Initial
Pattern solutions

HF propagation
NF pattern
model
Crossing Criterion

CFN propagation model

EDFM production
Update pattern simulation
solutions using
M-PSO NPV Value

No NPV
Converged ?

Yes
Finish

Figure 4.15 Flow chart of the CFN optimization

4.4 Examples

4.4.1 Case 1: Complex fracture network propagation simulation and sensitivity


analysis

Based on the complex fracture network propagation numerical model proposed


in the above sections, a series of simulations are performed under different NF
distribution fields and hydraulic fracturing operations to investigate the impacts of
different NF distributions and HF operation parameters on the CFN propagation. The

89
Texas Tech University, Hao Zhang, August 2021

reservoir model is the same as that in Section 4.1.4 and the corresponding geological
parameters can be seen in Table 4.1. The detailed sensitivity analysis procedure and
results can be seen in the following subsections.

First, the impact of NF length on the fracture network is studied. A set of parallel
NFs, with an angle of 70 degrees to the y-axis, are generated with fracture length varying
from 10ft to 100ft by a step of 5ft. Considering the uncertainty of NF distribution, 50
stochastic NF fields for each NF length are generated according to the power-law
function in Eq.(3-13) with the same parameters, and some representative models are
shown in Figure 4.16. These CFN simulation results are compared based on the same
propagation length on the y-axis, and the total HF length in the CFN is used to evaluate
the scale of the CFN. The CFN propagation results are shown in Figure 4.17, where the
grey curves represent the CFN variation for 50 stochastic NF fields, and the red curve
represents the mean value. As can be seen that, when the NF is short, the impact of NF
length on the fracture network is not obvious. Because of the stress shadow effect, the
HF is hard to activate both sides of the NF and could only propagate along one side of
the NF. As the NF length increase, the stress shadow effect turns weak, the HF becomes
easier to open both sides of the NF and makes the CFN scale sharply increase. After the
NF is longer than 60 ft, the HF is more inclined to propagate along with encountering
NF, which reduces the chance for HF to encounter more NFs, so the increase of fracture
network tends to slow down.

(a) 10 ft (b) 40 ft (c) 90 ft


Figure 4.16 The fracture network propagation with different NF lengths

90
Texas Tech University, Hao Zhang, August 2021

Figure 4.17 The fracture network scale variation with NF length

Besides the above, the impact of the NF angle on the fracture network scale is
studied. A set of parallel NFs, with varying angles from 10 degrees to 90 degrees to the
y-axis with a step of 5 degrees, is generated. Same as the previous case, the NF
distribution uncertainty is considered, some representative models are shown in Figure
4.18 and the fracture network propagation result is summarized in Figure 4.19. It can be
seen from the result that, when the contact angle is small, the HF could only open one
side of the NFs, and the chance for HF to encounter NFs is also small, which both
prohibit the fracture network growth. As the contact angle surpasses 70 degrees, the HF
becomes easier to open both sides of NFs and create more branches, which also
increases the chance for HF to encounter more NFs. As a consequence, the scale fracture
network meets a sharp increase after 70 degrees.

(a) 20 degrees (b) 45 degrees (c) 80 degrees


Figure 4.18 The fracture network propagation with different NF angle

91
Texas Tech University, Hao Zhang, August 2021

Figure 4.19 The fracture network scale variation with NF angles

The spacing of HFs is also investigated for its great impact on fracture network
propagation. Examples for HFs with varying spacing from 10ft to 100ft with a step of
5ft are simulated in 50 stochastic NF fields (Figure 4.20), and the fracture network
propagation result is summarized in Figure 4.21. As is shown that, when HF spacing is
small, there will be a strong stress shadow to prevent the HF propagation. As HF spacing
increases, the stress shadow becomes weak and the induced stress could help open more
NFs between the two HFs, so more branches are generated. When the spacing continues
increasing, the effect of neighboring fractures becomes weak, the two HFs tend to
resemble isolated fractures, so the scale of the fracture network slightly reduces.

(a) 10 ft (b) 50 ft (c) 100 ft


Figure 4.20 The fracture network propagation with different HF spacing

92
Texas Tech University, Hao Zhang, August 2021

Figure 4.21 The fracture network scale variation with HF spacing

Finally, the effect of NF density is investigated, and a series of randomly


distributed NF fields are adopted to make the simulation more in line with the real
situation (Figure 4.22); the NF proportionality coefficient 𝛼 in the power-law function
Eq.(3-13) is used to control the NF density and the proportionality coefficient 𝛼
increases from 100 to 500 with a step of 50. The fracture network propagation result is
summarized in Figure 4.23. As can be seen that, as the NF density increase, the chance
for HF encounters the NF will also increase and create more fracture branches. However,
as the NF becomes denser, the stress shadow effect will become stronger, which
prevents the inner fracture branches from growing; only the branches located on the
outer sides of CFN are more likely to grow, which reduces the increasing tendency of
the fracture network.

(a) 𝛼 = 100 (b) 𝛼 = 200 (c) 𝛼 = 500


Figure 4.22 The fracture network propagation with different NF density

93
Texas Tech University, Hao Zhang, August 2021

Figure 4.23 The fracture network scale variation with NF density

4.4.2 Case 2: Complex fracture network productivity analysis

In this section, a case for the CFN propagation and production in a shale
reservoir is performed according to the methods mentioned in the above sections. First
of all, a square reservoir domain is defined with a dimension of 500ft×500ft in the
horizontal view and a dimension of 50ft in the vertical view. There is anisotropic in-situ
stress in this reservoir, for which the direction of minimum and maximum horizontal
stress 𝜎ℎ and 𝜎𝐻 along with the x-axis and y-axis, respectively. Other parameters for
geological and fluid properties related to hydraulic fracturing are listed in Table 4.1. An
NF pattern is generated in the domain based on the power-law probability function in
Eq.(3-13), with the parameters 𝛼 = 100, 𝑎 = 1.5, 𝑙 ∈ [20,100]𝑓𝑡 (Figure 4.24-a). It
can be seen from the NF pattern that the existence probability of shorter NFs is much
larger than that of longer NFs, it agrees with the observation results of real NFs in the
formation, which is that the shorter the NF is, the larger the number will be (Gale et al.,
2007; Wu and Olson, 2016). The Matlab software is used to construct the CFN
propagation and production numerical model.

First, a case for CFN propagation is performed in a reservoir domain. First of all,
An HF is designed at the bottom and propagation along the y-axis with the same
injection rate as the former case, the propagation stops when the HF reaches 300ft in

94
Texas Tech University, Hao Zhang, August 2021

the y-axis. The simulation result in Figure 4.24-b shows that the HF stimulates the
encountering NFs and forms a CFN, the side of NF along with the HF propagation
direction is more likely to be opened; besides, because of the induced stress, not all the
frac-branches in the CFN could continue propagating, only the frac-branches locate at
the top of CFN is easier to continue propagating because of the low induced stress there.
The CFN propagation shows good agreement with former scholars' researches. (Zhang
et al., 2015; Liu and Reynolds, 2019)

(a) NF pattern (b) CFN model


Figure 4.24 A sketch for complex fracture network propagation simulation

After propagation simulation, the CFN obtained is brought into the EDFM for
production simulation. The EDFM is constructed based on the HFM Modulus of the
Matlab toolkit MRST (Lie, 2019). Based on the EDFM generation method in Section
3.1, the reservoir is subdivided into 100×100 structured grids in the horizontal direction,
each of which with a dimension of 5ft×5ft×50ft; the CFN together with the NF pattern
is embedded into these grids. As shown in Figure 4.25, the CFN represented by the thick
blue segments, and NFs represented by the thin blue segments are perfectly embedded
into the grids, the shape of grids is not affected by the complexity of fractures.

95
Texas Tech University, Hao Zhang, August 2021

Figure 4.25 The CFN and NF pattern converts to EDFM

The initial pressure of the reservoir is homogenous with a value of 3000psi, and
the BHP is set as 500psi. Other geological parameters related to the production
simulation are displayed in Table 4.2, most of which derive from real field data (Yu and
Sepehrnoori, 2013; Wang, 2020). The shale gas flow model in Section 4.2.3 is adopted
to simulate the shale gas production, for which the adsorption mechanism and the
transport mechanisms in the matrix and fracture system are considered and shown in
Figure 4.26. The shale gas contains only one component CH4, the properties of which
are shown in Table 4.3 (Jiang and Younis, 2015; Wang, 2020). The total production
time is set as 10 years.

(a) Langmuir isotherm curve (b) Fracture permeability correction factor


Figure 4.26 Correction factors for adsorption and transport mechanisms

96
Texas Tech University, Hao Zhang, August 2021

Table 4.2 Geological parameters for production simulation

Parameters Value Parameters Value

Reservoir initial pressure 3000 psi Fracture porosity 0.9


Initial matrix permeability 100 nD Fracture compressibility 1.5-12 psi-1
Matrix porosity 0.1 Biot constant 0.5
Matrix compressibility 2.1-14 psi-1 Wellbore radious 0.3 ft
Initial HF permeability 100 mD The bottom hole pressure 500 psi
Initial NF permeability 10 mD Skin factor 10
HF aperture 0.1 in Formation thickness 50 ft
NF aperture 0.01 in

Table 4.3 Properties of shale and gas

Parameters Value Unit

Shale rock density 160 lb/ft3


Molecular weight, CH4 0.036 lb/mol
Critical pressure, CH4 670 psi
Critical temperature, CH4 190.6 K
Reservoir temperature 327.6 K
Langmuir pressure 0.055 psi
Langmuir volume 0.048 ft3/lb

To evaluate the effect of CFN on the production of fractured shale reservoirs,


two EDFM models are simulated, one is with CFN and the unstimulated NFs, another
with only the CFN. After 10 years of production simulation, the pressure field
distributions for both models are displayed in Figure 4.27 respectively, and the
corresponding cumulative shale gas productions are shown in Figure 4.28.

97
Texas Tech University, Hao Zhang, August 2021

(a) The CFN and the unstimulated NFs (b) The CFN without the unstimulated NFs
Figure 4.27 The pressure field (psi) after 10 years of production for the EDFM

Figure 4.28 The cumulative shale gas production comparison

From Figure 4.27, it can be seen that the CFN shows a significant effect in shale
gas production, the area covered by the CFN obtains a perfect production, the pressure
there is nearly the same as the BHP, which reveals that the shale gas there is almost
depletion. However, the production of the area away from the CFN is not very good,
the effect of CFN gradually weakens as the distance becomes longer. There is hardly
any pressure drop in the area 150ft away from the CFN, which means no gas is produced
beyond the area. Additionally, we can also see that the CFN makes a majority

98
Texas Tech University, Hao Zhang, August 2021

contribution to the gas production, yet the contribution of unstimulated NFs is very low;
for the pressure field in Figure 4.27 and the cumulative shale gas production in Figure
4.28. The simulation result reveals the unavoidable impact of CFN on shale reservoirs’
production.

4.4.3 Case 3: Complex fracture network optimization under M-NNA

As shown in the above case, the CFN possesses a significant effect on gas
production in fractured shale reservoirs, so optimizing the CFN configuration is
essential to the development of the fractured shale reservoir. Many scholars have
analyzed the geomechanical and fluid parameters’ effect on the CFN propagation and
estimated the optimum CFN under certain conditions (Zhang et al., 2015; Wu and Olson,
2016; Xu et al., 2018; Wu et al., 2019a), while the CFN optimization comprehensively
considering the propagation and production under geological and economical effects
haven’t gained enough study.

(a) Type 1: Aligning placement (b) Type 2: Alternating placement


Figure 4.29 Two types of hydraulic fracturing placement

In this section, a CFN optimization is performed based on a case of multi-well


hydraulic fracturing and production in a fractured shale reservoir. First of all, a
rectangular reservoir model is constructed, which is similar to that in the previous case,

99
Texas Tech University, Hao Zhang, August 2021

except for the horizontal scale. As shown in Figure 4.29, the reservoir’s horizontal
dimension is 500ft×1000ft to simulate the CFN propagation and interaction from
multiple wells. Two horizontal wells (represented by thick red segments) are drilled
parallelly along with the x-axis, and the HFs (represented by thin black segments) are
generated along with the y-axis. The HF placements are designed as two types, aligning
placements and alternating placements, as a comparison. As shown in Figure 4.30, the
power-law probability NF map is constructed within the reservoir domain according to
Eq.(3-13), under the relative parameters 𝛼 = 200, 𝑎 = 1.5, 𝑙 ∈ [20,150]𝑓𝑡.

Figure 4.30 The NF pattern within the reservoir domain

The NPV considering the cost of hydraulic fracturing and the revenue of shale
gas production in Eq.(4-22) is regarded as the optimization objective function, and the
relative economic parameters are shown in Table 4.4. Since geological parameters such
as in-situ stress, NF distribution, and formation fluid properties, are the unchangeable
parameters, to realize the CFN optimization, the fracturing operational parameters,
including the injection rate, the well spacing, as well as the fracturing spacing, are
chosen as the optimization variables. The whole optimization procedure is conducted
by the machine learning algorithm M-NNA introduced in Section 4.3.3, for which the
optimization variables mentioned above are used as the pattern solution 𝑋, and the

100
Texas Tech University, Hao Zhang, August 2021

pattern solution population 𝑁𝑝𝑜𝑝 is defined as 10. Table 4.5 shows the search space for
optimization variables, which is used to avoid the simulation collapsing due to extreme
values. The time for hydraulic fracturing is set as 30min, the timeline set as 10 years for
production, and the total step for optimization iteration is set as 50. All the other
geological and fluid parameters for CFN propagation and production simulation are the
same as those used in the previous case and are shown in Table 4.1, Table 4.2, and Table
4.3. Since revealed in the previous case that the unstimulated NFs don’t make too much
contribution to gas production, so only the CFN is considered in the EDFM. The entire
optimization framework is established using the Matlab software, based on the
workflow in Section 4.3.4.

Table 4.4 Economic parameters and their values

Parameters Value Unit

The price of shale gas 2 $/MSCF


The cost of well drilling 200 $/ft
The cost of fracturing fluid 5 $/ft3
Annual discount rate 0.1

Table 4.5 Optimization variables and their search space

Optimization variables Search space

The fracture spacing (ft) [10, 300]


The well spacing (ft) [200, 700]
The injection rate (bbl/min) [10, 50]

Based on the above models and workflow, the CFN optimization under the
aligning and alternating fracturing placements are performed respectively. After
optimization, the corresponding results are obtained as shown in the following figures.
Figure 4.31 and Figure 4.32 show the CFNs before and after optimization for the
aligning and alternating fracturing placements, respectively. Besides, the results of CFN
production simulation for both fracturing placements are also presented; Figure 4.33
and Figure 4.34 show the pressure field after 10 years of production under initial and

101
Texas Tech University, Hao Zhang, August 2021

optimal CFNs for the aligning and alternating fracturing placements, respectively.
Additionally, the cumulative gas production under initial and optimal CFNs for both
fracturing placements are compared in Figure 4.35, and the corresponding NPV
variations during the optimization iteration are displayed together in Figure 4.36. The
optimization variables variation during the optimization iterations for the aligning and
alternating fracturing placement are shown in Table 4.6 and Table 4.7, respectively.

(a) The initial CFN (b) The optimal CFN


Figure 4.31 The initial and optimal CFN for aligning fracturing placement

(a) The initial CFN (b) The optimal CFN


Figure 4.32 The initial and optimal CFN for alternating fracturing placement

102
Texas Tech University, Hao Zhang, August 2021

(a) The initial CFN (b) The optimal CFN


Figure 4.33 The pressure fields (psi) after 10 years of production under initial and
optimal CFNs for aligning fracturing placement

(a) The initial CFN (b) The optimal CFN


Figure 4.34 The pressure fields (psi) after 10 years of production under initial and
optimal CFNs for alternating fracturing placement

103
Texas Tech University, Hao Zhang, August 2021

(a) The aligning fracturing placement (b) The alternating fracturing placement
Figure 4.35 The accumulative gas production under initial and optimal CFNs

Figure 4.36 The objective function variation with the optimization iterations

Table 4.6 Optimization variables during optimization (Aligning fracturing placement)

Optimization Values in Values in Values in Values in Values in


Variables step 1 step 10 step 20 step 40 step 50

The fracture spacing (ft) 39.91 87.23 89.37 91.38 91.75


The well spacing (ft) 614.24 444.24 425.8 397.08 398.07
The injection rate (bbl/min) 29.68 46.05 51.81 53.21 53.18

104
Texas Tech University, Hao Zhang, August 2021

Table 4.7 Optimization variables during optimization (Alternating fracturing placement)

Optimization Values in Values in Values in Values in Values in


Variables step 1 step 10 step 20 step 40 step 50

The fracture spacing (ft) 33.85 54.81 69.92 72.23 72.97


The well spacing (ft) 600.12 426.01 417.12 415.64 415.54
The injection rate (bbl/min) 33.78 44.73 51.76 55.26 55.57

The results of CFN optimization in Figure 4.31 and Figure 4.32 show that the
optimal CFNs for both fracturing placements are improved compared with the
corresponding initial ones, larger areas are covered and more fracture branches are
generated by the optimal CFNs than the initial ones. The pressure fields after production
in Figure 4.33 and Figure 4.34, as well as the cumulative gas production in Figure 4.35,
can also show that after the CFN optimization, larger formation areas are stimulated and
more gas is produced. Then comparing the optimal CFN of these two fracturing
placements for analysis, it can be found that the alternating fracturing placement gives
the opposition CFNs a larger space for propagation; they are more likely to stagger with
each other rather than encounter, which can help to stimulate and connect more NFs
into networks. The cumulative gas production in Figure 4.35 and the NPV values in
Figure 4.36 also reveal that the alternating fracturing placement produces more gas and
leads to higher economic revenue. The comparison reveals that the alternating fracturing
placement shows a better performance for fractured shale reservoir development than
the aligning fracturing placement. To sum up, the above results and analysis have
enough persuasive power to demonstrate the reliability of the CFN simulation and
optimization method proposed in this study.

105
Texas Tech University, Hao Zhang, August 2021

CHAPTER V

UNSTRUCTURED QUAD-GRID BASED EMBEDDED DISCRETE


FRACTURE MODEL
The current EDFM subdivides the reservoir matrix with structured grids, which
significantly simplifies the gridding and effectively avoids the miscalculation caused by
arbitrarily oriented fractures, which simplifies the gridding and is efficient for the
simulation of reservoirs with arbitrarily oriented fractures. However, the structured grid
used in the EDFM still possesses drawbacks, for it has poor adaptivity to reservoirs with
irregular boundaries. Usually, the reservoir models used in EDFM are rectangular, so
the structured grid could adapt to these reservoirs very well; however, when the
reservoir is not rectangular and contains irregular boundaries, the structured grid will be
hard to adapt to its shape and cause inevitable deformation.

To conquer the drawbacks of the structured grid-based EDFM, an unstructured


quadrangular grid-based EDFM (UnQ-EDFM) is proposed in this chapter. In UnQ-
EDFM, a mature unstructured quadrangular grid is adopted that could perfectly adapt
to irregular boundaries without increasing the grid number; besides, grid refinement
could be achieved more flexibly in anywhere we want, without adding the gridding
complexity. Then the feasibility of the UnQ-EDFM is testified in an irregularly shaped
naturally fractured shale reservoir. The simulation results are analyzed and compared to
that of the structured grid-based EDFM, which proves the reliability and superiority of
the UnQ-EDFM. The paper is organized as follows. First, the drawbacks of the current
structured grid-based EDFM are illustrated in Section 5.1. Then the detailed statement
on the unstructured quadrangular grid generation is described in Section 5.2.
Furthermore, the optimization framework based on the UnQ-EDFM is constructed in
Section 5.3, A state-of-the-art neural network algorithm (NNA) is used as the
optimization algorithm, which is further improved into an M-NNA in this study and
shows higher accuracy and efficiency over the traditional algorithms. Lastly, several
numerical examples based on the structured grid-based EDFM and the UnQ-EDFM are
compared and discussed in Section 5.4.

106
Texas Tech University, Hao Zhang, August 2021

5.1 The drawback of the structured grid in EDFM

The EDFM subdivides the reservoir matrix with structured grids, which
significantly simplifies the gridding and effectively avoids the miscalculation caused by
arbitrarily oriented fractures. However, the structured grid used in the EDFM still
possesses drawbacks, for it has poor adaptivity to reservoirs with irregular boundaries.
Usually, the reservoir models used in EDFM are rectangular shape, the drawback for
the structured grid is not obvious; however, when the reservoir is not rectangular and
contains irregular boundaries, the structured grid will be hard to adapt to its shape and
cause inevitable deformations, which will reduce the simulation accuracy. To avoid this
drawback, Yan et al. (2016) applied the triangular grid into EDFM to match the shape
of irregular reservoirs, but the triangular grid increases the grid number comparing with
the structured grid and prolongs the simulation time. Another drawback for the
structured grid is that it causes a sharp pressure gradient between matrix and fractures
and weakens the simulation accuracy, especially with low fracture conductivity. Wang
(2020) coupled the LGR method with EDFM to alleviate the pressure gradient by
refining the grids around the fractures, but this will complicate the gridding process and
cannot perfectly adapt to complex fractures. Until now, the problems caused by the
structured grid are still not fully solved.

(a) Irregular reservoir domain

107
Texas Tech University, Hao Zhang, August 2021

(b) Large structured grids (c) Small structured grids


Figure 5.1 Structured grids generated in the reservoir with irregular boundary

As can be seen in Figure 5.1-a, there is an irregular-shaped reservoir, the length


for each boundary is marked in the axes, a type of structured grid with a 2m grid length
is generated in the reservoir (Figure 5.1-b). However, the structured grid cannot adapt
to the irregular boundary, which results in zig-zag boundaries in the grid map and causes
deformation to the original reservoir shape. This deformation caused by structured grids
will lead to errors in the numerical simulation. To reduce the deformation, we should
use a smaller-sized grid (0.5m) (Figure 5.1-c), but this will enlarge the number of grids
and prolong the simulation time, which is inefficient for the simulation. Facing the
drawbacks of the structured grid, scholars also adopted unstructured triangular grids into
the reservoir with irregular boundaries. The triangular grid shows flexible adaptivity to
irregular boundaries, so it has been widely used in complex-shaped reservoirs’
simulation (Yan et al, 2016; Liu and Reynolds, 2019); however, it is still not perfect. As
is shown in Figure 5.2, to cover the same reservoir area with the same grid length, the
triangular grid always needs a higher number of grids comparing with the structured
grid, which is still inefficient for the reservoir simulation.

108
Texas Tech University, Hao Zhang, August 2021

(a) The structured grid (b) The triangular grid

(c) The statistic of grid number


Figure 5.2 The comparison of the structured grid and the triangular grid in the same
irregular reservoir

5.2 The unstructured quadrangular grid

5.2.1 The principle of frontal Delaunay quadrangular grid

Facing the drawbacks of the currently used grids in EDFM, we propose an


unstructured quadrangular grid-based EDFM (UnQ-EDFM). In this method, a mature
unstructured quad-grid generation method is adopted, it could perfectly adapt to
irregular boundaries and the grid number is lower than the triangular grid, which
considers accuracy and efficiency simultaneously. The unstructured quadrangular grid
used in this study is the frontal Delaunay quadrangular grid (Remacle et al., 2013),

109
Texas Tech University, Hao Zhang, August 2021

which is a kind of indirect method that first generates triangular grids and then combines
the adjacent triangles to form quad grids. Different from other traditional indirect
methods, it creatively applies the 𝐿∞ -norm to directly generate right Delaunay triangular
grids, so the calculation efficiency is improved significantly. In a two-dimensional plane,
the distance between two points 𝐱1 (𝑥1 , 𝑦1 ) and 𝐱2 (𝑥2 , 𝑦2 ) is usually calculated by the
Euclidean norm (𝐿2 -norm). However, the distance of two points can also be defined by
other norms:

‖𝐱2 − 𝐱1 ‖1 = |𝑥2 − 𝑥1 | + |𝑦2 − 𝑦1 | 𝐿1 − norm


‖𝐱2 − 𝐱1 ‖2 = (|𝑥2 − 𝑥1 |2 + |𝑦2 − 𝑦1 |2 )1⁄2 𝐿2 − norm
(5-1)
‖𝐱2 − 𝐱1 ‖𝑝 = (|𝑥2 − 𝑥1 |𝑝 + |𝑦2 − 𝑦1 |𝑝 )1⁄𝑝 𝐿𝑝 − norm
{ ‖𝐱 2 − 𝐱1 ‖∞ = 𝑚𝑎𝑥( |𝑥2 − 𝑥1 |, |𝑦2 − 𝑦1 |) 𝐿∞ − norm

(a) Equilateral triangles (b) Right triangles


Figure 5.3 Comparison of different triangular grids (Remacle et al., 2013)

As shown in Figure 5.3, compared with the equilateral triangle in Figure 5.3-a,
the right triangle in Figure 5.3-b is with different side lengths in the 𝐿2 -norm, so it
cannot be generated by the ordinary Delaunay triangulation method (Watson et al., 1981;
Rebay, 1993). However, if the distance between two points is measured based on the
𝐿∞ -norm, side-lengths of the right triangle will be equal: ‖𝐲 − 𝐲1 ‖∞ = ‖𝐲 − 𝐲2 ‖∞ = 𝑎.
Based on this criterion, the right triangles could be directly generateded by the Delaunay
triangulation method in the 𝐿∞ -norm.

110
Texas Tech University, Hao Zhang, August 2021

(a) Grids in an annular domain (b) cross fields


Figure 5.4 The grids and the corresponding cross fields (Remacle et al., 2013)

To let the orientation of the generated grids aligned with the nearby boundary,
the parameter “cross-field” is used, which represents the preferred grid orientation at
each point 𝒖(x, y) of the domain and is controlled by 𝜃(𝒖). As shown in Figure 5.4, the
orientation of the grids generated around the point 𝒖(x, y) should be aligned to the
cross-field of it. Controlled by the 𝐿∞ -norm distance and the cross-fields, Delaunay right
triangles are generated aligned with the adjacent boundaries layer by layer. After the
gridding, the triangular grids are transformed into quad-grids. The entire procedure for
the frontal Delaunay quad-grid generation is listed below, and the schematic figure for
each step can also be seen in Figure 5.5:

(1) Set initial boundary points to define the shape of the domain (Figure 5.5-a).

(2) Connect the initial boundary points into a loop to constrain the domain for gridding,
and the segments connecting these points are treated as the boundaries (Figure 5.5-
b).

(3) Insert a series of new points on the boundaries to subdivide the boundaries into
several short segments, and then define the corresponding cross fields for these
points according to the boundary orientations (Figure 5.5-c).

(4) Subdivide the domain into Delaunay right triangles based on the 𝐿∞ -norm, the

111
Texas Tech University, Hao Zhang, August 2021

points on the boundaries are regarded as vertexes and the orientations of the
generated grids are controlled by the nearest cross fields (Figure 5.5-d).

(5) Combine two adjacent right triangular grids that share the same bevel edge into one
quad-grid and delete the shared bevel edge (Figure 5.5-e).

Figure 5.5 Example of Frontal Delaunay Quad Generation

According to the procedure above, a set of Frontal Delaunay quad-grid is


generated. Due to the boundary conditions' complexity, it is hard to make all of the
generated triangles be right triangles, so some little error is allowed. A detailed
description of the Frontal Delaunay quad-grid generation method can be seen in
Remacle et al. (2013).

Based on this unstructured quad-grid subdivision method, the irregularly shaped


reservoir domain in Figure 5.6-a is subdivided, and the grid map generated is compared
with the structured grid map in Figure 5.1, the number of grids for different these three
types are counted. As can be seen that, the frontal Delaunay quad-grid has a robust
adaptivity to the irregular boundary, and the original shape of the reservoir is not

112
Texas Tech University, Hao Zhang, August 2021

deformed comparing with the structured grid, so this grid will help to reduce the error
for numerical simulation. Besides, despite the irregular boundary, the shape of the most
quad-grid is pretty regular and distributes aligning to the orientation of the adjacent
boundaries, only a few grids are slightly deformed. The bar chart also reveals that the
grid number generated by the frontal Delaunay quad-grid is far less than the triangular
grid, this will reduce the calculation time and increase the simulation efficiency
significantly, compared with the triangular grid.

(a) The structured grid (b) The unstructured quad grid

(c) The statistic of grid number


Figure 5.6 The comparison of the structured grid and the triangular grid in the same
irregular reservoir

113
Texas Tech University, Hao Zhang, August 2021

5.2.2 Local grid refinement for unstructured quadrangular grid

The local grid refinement (LGR) is a necessary gridding measure that is


commonly used in some heterogeneous or anisotropic reservoirs, which could help to
increase the accuracy of the simulation. However, the LGR is not easy to be achieved
by the commonly used grids, which usually cause large amounts of calculation or
complex connection problems between refined grids and normal grids. While the
unstructured quad-grid could realize the LGR without causing any problems. To realize
the LGR, some points in the domain are defined as the key points, which possess values
that can control the side length of grids around them. The values for these key points
represent the lengths of grids generated around them, the larger the key point value is,
the larger the surrounding grids will be. When smaller values are assigned to the key
points in the region where we want to achieve the LGR, the varying-sized grid will be
generated.

Table 5.1 Values of key points

Key points Values (m)

B1 2
B2 2
B3 2
B4 0.5

Figure 5.7 shows an unstructured quad-grid LGR example, the vertex points in
the boundaries are defined as the key points, with the corresponding values listed in
Table 5.1. According to the key points’ values, a varying sized quad-grid map is
generated (Figure 5.7-b); as can be seen that the value for key point B4 is 0.5m in Table
5.1, which is smaller than the values of other key points, so the LGR is realized around
B4. Even with LGR, the grids generated are still strictly aligned to the orientation of the
adjacent boundaries and most of the grids are still in regular shape; besides, the size of
grids varies smoothly from the LGR region to the normal-sized region.

114
Texas Tech University, Hao Zhang, August 2021

(a) Reservoir domain with key points (b) Unstructured quad-grid with LGR
Figure 5.7 Unstructured quad grid with local grid refinement

All the above examples prove the feasibility of the unstructured quad-grid over
the currently used structured and triangular grids, for it can perfectly adapt to irregularly
shaped reservoirs without causing any deformation, cover the domain with fewer grids,
and achieve LGR flexibly in anywhere of the reservoir. Based on these advantages, the
unstructured quad-grid is implemented into the EDFM to construct the UnQ-EDFM.

5.3 The optimization method

The EDFM is mostly used for history matching or optimization problems, which
usually need hundreds or thousands of times of simulation and cause a large amount of
calculation (Yu et al., 2018; Wu et al., 2019b), so a small error in one simulation will
lead to a huge difference in the final results. To verify the reliability of the UnQ-EDFM
for these applications, an optimization framework is built.

5.3.1 Optimization objective function

For reservoir engineering problems, the ultimate destination of optimization is


usually achieving the maximum economical recovery with the lowest cost, so the net
present value (NPV) is chosen as the objective function of optimization, but the
formulation is a little bit different from that in the previous sections. The NPV function
for the shale gas production can be written as:

115
Texas Tech University, Hao Zhang, August 2021

𝑁 (𝑟𝑔 ⋅𝑄𝑔𝑛 −𝑟𝑤 ⋅𝑄𝑤


𝑛 )×𝛥𝑡 𝑛
,𝑗 𝑁 𝑁
𝑡𝑁 𝑝 ,𝑗 𝑝 𝑓
𝐽𝑜𝑏 = ∑𝑛=1 {∑𝑗=1 𝑛 } − ∑𝑗=1 (𝐿𝑤𝑗 ⋅ 𝐶𝑤 + ∑𝑖=1 𝐿𝑓𝑖𝑗 ⋅ 𝐶𝑓 ) (5-2)
(1+𝑏)𝑡 /365

where, 𝐽𝑜𝑏 represents the NPV value; 𝑁𝑡 is the total number of the shale gas
production simulation time step; 𝛥𝑡 is the length of the simulation time step and 𝑡 is the
total simulation time; 𝑟𝑔 is the gas price and 𝑟𝑤 is the cost of produced water disposal;
𝑛 𝑛
𝑄𝑔,𝑗 and 𝑄𝑤,𝑗 are the gas and water production rate of the 𝑗th well in the 𝑛th simulation
time step, respectively; 𝑁𝑝 and 𝑁𝑓 are the number of production wells and the number
of HFs for the 𝑗th well, respectively; 𝐶𝑤 is the well drilling cost per unit length, and 𝐿𝑤𝑗
is the length of the𝑗th well; 𝐶𝑓 is the hydraulic fracturing cost per unit length and 𝐿𝑓𝑖𝑗 is
the length of the 𝑖 th HF in the 𝑗th well; 𝑏 is the annual discount rate.

5.3.2 The workflow of optimization

The optimization algorithms used in this study is the M-NNA algorithm


proposed in above Section 4.3. The optimization is processed cyclically into several
steps. In each iteration step, build up the reservoir EDFM model according to
optimization variables firstly, then calculate the NPV value according to the simulation
results and update the optimization variables through the M-NNA. This procedure is
repeated until reaching the optimum value. The detailed optimization workflow is
illustrated as follows and the corresponding flowchart can be seen in Figure 5.8.

(1) At the beginning of optimization, convert the optimization variables into a pattern
solution 𝑋 in the M-NNA, set a series of random pattern solutions 𝐗 =
(𝑋1 , 𝑋2 ⋯ 𝑋𝑁𝑝𝑜𝑝 ), which contains the parameters to be optimized.

(2) Based on the geological parameters, build up the corresponding reservoir EDFM
model and grid for each pattern solution 𝑋𝑖𝑘 , where 𝑘 represents the current
iteration step. Perform the numerical simulation and record the simulation results.
𝑘
(3) According to the simulation results, calculate the NPV value 𝐽𝑜𝑏,𝑖 for each pattern
solution 𝑋𝑖𝑘 using Eq.(5-2), the pattern solution with the best NPV is regarded as
the target solution 𝑋𝑡𝑎𝑟𝑔𝑒𝑡 (𝑘); then update the pattern solution 𝑋𝑖𝑘+1 according to

116
Texas Tech University, Hao Zhang, August 2021

the M-NNA in Section 4.3.

(4) Back to step 2 and build up a new EDFM model based on the new pattern solution
𝑋𝑖𝑘+1 , and then calculate a new NPV value 𝐽𝑜𝑏,𝑖
𝑘+1
, accordingly.

𝑏𝑒𝑠𝑡
(5) Continue repeating the step 2-4 until the best-known NPV value 𝐽𝑜𝑏 converges.

Start

Initial values for


Optimization variables

Unstructured
EDFM model
quad-grid

Reservoir
Update variables simulation
using M-NNA
NPV Value

No NPV
Converged ?

Yes
Finish

Figure 5.8 Flowchart of M-NNA optimization procedure

5.4 Examples

In this section, the reliability and superiority of the unstructured quadrangular


grid-based EDFM (UnQ-EDFM) are verified using two cases of multi-stage fracturing
horizontal well production simulation in irregularly shaped shale reservoirs. In Case 1,
different sized unstructured quad-grids and structured grids are generated within the
same reservoir and their simulation results are compared; In case 2: a well placement
optimization is processed in the reservoir with different kinds of grids, the optimization
results are analyzed. The unstructured quad-grid generation method is compiled using
the C++ program and integrated into the HFM modulus of the Matlab toolkit MRST

117
Texas Tech University, Hao Zhang, August 2021

(Lie, 2019) to form the UnQ-EDFM, the optimization framework is constructed using
Matlab software.

5.4.1 Case 1: Verification of the unstructured quad-grid based EDFM

In this case, the simulation accuracy of the EDFM shale gas production model
mentioned is first validated by comparing it with a commercial simulator. A rectangular
reservoir with the dimension of 500m×500m×50m is constructed using EDFM and
discretized into 50×50×1 grids, each one is 10m×10m×50m. A horizontal well with 4
HFs is constructed in the reservoir, with the wellhead location (175m, 250m) and the
well length of 150m. The HFs are planar bi-wing fractures, with a half-length of 100m
and a spacing of 50m. As a comparison, the same reservoir model is constructed using
the Eclipse software, in which the HFs are presented by refined grids. The two models
are shown in Figure 5.9, for the EDFM in Figure 5.9-a, the blue segments represent the
HFs, and the block red segment represents the wellbore.

(a) EDFM model (b) Eclipse model


Figure 5.9 The reservoir model with fractured well in different simulators

The shale gas flow model in Section 4.2.3 is used to simulate the shale gas
production in EDFM, the geological properties of which are listed in Table 5.2, some of
which derive from the Barnett shale (Yu and Sepehrnoori, 2014; Cao et al., 2016; Wang,
2020). The shale gas contains only one component CH4, the property of which is shown
in Table 5.3 (Jiang and Younis, 2015; Wang, 2020), the correction factors of adsorption

118
Texas Tech University, Hao Zhang, August 2021

mechanism and transport mechanisms in the matrix and fracture system can be seen in
Figure 5.14. The maximum horizontal stress 𝜎𝐻 is 34MPa with the direction aligned to
the y-direction, while the minimum horizontal stress 𝜎ℎ is 30Mpa with the direction
aligned to the x-direction. The, The reservoir is with an initial pressure of 20MPa, and
the production time is set as 10 years with a constant bottom hole pressure (BHP) of
3.4MPa. The Eclipse model for comparison is simulated by the E300 simulator based
on the same geological and fluid properties.

(a) Langmuir isotherm curve (b) Fracture permeability correction factor


Figure 5.10 Correction factors for adsorption and transport mechanisms

Table 5.2 Geological parameters for the reservoir model

Parameters Value Parameters Value

Reservoir initial pressure 20 MPa Fracture porosity 1.0


Matrix permeability 200 nD Fracture compressibility 1-8 Pa-1
Matrix porosity 0.1 fracture aperture 3 mm
-10 -1
Matrix compressibility 1.45 Pa Biot constant 0.5
Minimum horizontal stress 30 MPa Wellbore radious 0.1 m
Maximum horizontal stress 34 MPa The bottom hole pressure 3.4 MPa
HF permeability 100 mD Skin factor 10
NF permeability 10 mD Formation thickness 50 m

119
Texas Tech University, Hao Zhang, August 2021

Table 5.3 Properties of shale and gas

Parameters Value Unit

Shale rock density 2500 kg/m3


Molecular weight, CH4 0.01604 kg/mol
Critical pressure, CH4 4.60 MPa
Critical temperature, CH4 190.6 K
Reservoir temperature 327.6 K
Langmuir pressure 7.89 MPa
Langmuir volume 0.003 m3/kg

After the simulation, the pressure fields of these two models are shown in Figure
5.11, and the accumulative gas productions are shown in Figure 5.12. As can be seen
that, the pressure fields obtained from the EDFM and Eclipse software are almost the
same; additionally, the accumulative shale gas production cure for the EDFM also
matches pretty well with that of the Eclipse model. These simulation results are
convincing enough to validate the accuracy of the EDFM.

(a) EDFM model (b) Eclipse model


Figure 5.11 Reservoir pressure (MPa) after 10 years of production for models in
different simulators

120
Texas Tech University, Hao Zhang, August 2021

Figure 5.12 The cumulative gas production for different models

Based on the validation of the EDFM shale gas production model, an irregularly
shaped reservoir is used to verify the reliability of the unstructured UnQ-EDFM. As
shown in Figure 5.13, the shape of the reservoir model derives from the Brugge reservoir,
which is commonly used in reservoir numerical simulations (Chen et al., 2010; Zhang
et al., 2019). The original Brugge reservoir is too large to be used in this case, so its
scale is adjusted to 1600m×800m, while its original shape is maintained. The color of
the Brugge reservoir represents its original permeability distribution, which is not used
in this simulation. Some boundary points are chosen as the key points and are connected
by red segments to constrain the boundaries of the reservoir during grid subdivision.

Figure 5.13 The horizontal view of the reservoir

121
Texas Tech University, Hao Zhang, August 2021

For the production simulation, a multi-stage hydraulic fractured horizontal well


is constructed in the reservoir, with a wellhead location (650m, 700m) and a well length
of 850m. The HFs are planar bi-wing fractures, with the fracture half-length of 100m
and fracture spacing of 50m. The HFs together with the NFs generated based on the
power-law probability function in Eq. (3-13) are shown in Figure 5.14, where the black
cure represents the reservoir boundary, the block blue segments represent the HFs and
the thin blue segments represent the NFs. The fracture aperture is set as 3mm; the initial
permeability for the HF and NF is 100md and 10md respectively. The reservoir’s stress
and pressure conditions, and other detailed geological and fluid properties the same as
the model mentioned above and can be seen in Table 5.2, Table 5.3, and Figure 5.10.
The production time is set as 20 years.

Figure 5.14 The fracture map of the reservoir

First, a sensitivity analysis of structured grid length’s effect on the EDFM


simulation accuracy is performed. The reservoir domain is subdivided by 1m, 2m, 5m,
10m, 20m, and 30m structured grids, which are shown in Figure 5.15, respectively (the
1m and 2m grids are too narrow to be shown). The simulation results for different grids
can be seen in Figure 5.16, the accumulative shale gas production for the different grids
are recorded and compared in Figure 5.17-a, and their errors comparing with 1m
structured grid is shown in Figure 5.17-b, the total number of grids and the
corresponding simulation times for different grids are shown in Figure 5.18.

122
Texas Tech University, Hao Zhang, August 2021

(a) 5m structured grid (b) 10m structured grid

(c) 20m structured grid (d) 30m structured grid


Figure 5.15 Reservoir models with fractures and different sized structured grids

(a) 1m structured grid (b) 2m structured grid

(c) 5m structured grid (d) 10m structured grid

123
Texas Tech University, Hao Zhang, August 2021

(e) 20m structured grid (f) 30m structured grid


Figure 5.16 The reservoir pressure (MPa) after 20 years of production under different
grid lengths

(a) Accumulative gas production


0.16
13.62%
0.14
0.12
0.10
Error

0.08
5.70%
0.06
0.04
0.02 1.02%
0.07% 0.15%
0.00
2m 5m 10m 20m 30m
Grid size

(b) The error comparing with 1m structured grid


Figure 5.17 The comparison of simulation results under different grids

124
Texas Tech University, Hao Zhang, August 2021

1.E+06 10000

1.E+05
1000

Simulation time (S)


1.E+04

Grid number 1.E+03 100

1.E+02
10
1.E+01

1.E+00 1
1m 2m 5m 10m 20m 30m 1m 2m 5m 10m 20m 30m
Grid size Grid size

(a) The number of grids (b) The simulation time


Figure 5.18 The statistics of grid number and simulation time for different grids

Analyze combining Figure 5.15, Figure 5.16, and Figure 5.17, it can be easily
seen that the structured grid shows bad adaptation to the irregular shaped when the grid
length is larger than 10m, a lot of jagged boundaries appear and the reservoir’s shape is
severely deformed; the accumulation gas production in Figure 5.17 also demonstrates
that as the grid length enlarges, the simulation error increases significantly. Whereas the
deformation is not obvious as the grid length is smaller than 10m, the simulation error
also reduces to a small extend. However, as the grid length decrease, the total grid
number increases exponentially (Figure 5.18-a), which dramatically prolongs the
simulation time to hundreds of times (Figure 5.18-b). So the structured grid is not
suitable for irregularly shaped reservoir simulation, too large grids will increase the
simulation error, yet too narrow grids will enlarge the simulation time.

As a comparison, unstructured quad-grids are generated in the reservoir domain.


Since is indicated in Figure 5.17 that the error of the 10m grid is only 1% compared to
the 1m grid, so the 10m grid is regarded as the standard grid in the following simulations.
The unstructured quad-grids are generated with grid lengths of 10m, 20m, and 30m
respectively, the grid maps together with the corresponding simulation results are shown
in Figure 5.19. We can see that the unstructured quadrangular grid strictly constrains
the original boundaries of the reservoir, there is no deformation whatever the size of the
grid. The cumulative shale gas production for the different sized structured and

125
Texas Tech University, Hao Zhang, August 2021

unstructured quad grids are compared in Figure 5.20, which shows that the simulation
errors between different sized unstructured quad grids are much smaller than the
corresponding structured grids. These all demonstrate the superiority of the unstructured
quad-grid over the structured grid.

(a) 10m unstructured quad grid map (b) 10m unstructured quad-grid pressure map

(c) 20m unstructured quad grid map (d) 20m unstructured quad-grid pressure map

(e) 30m unstructured quad grid map (f) 30m unstructured quad grid pressure map
Figure 5.19 Grid maps and the corresponding reservoir pressure (MPa) after 20 years of
production for different sized unstructured quad grids

126
Texas Tech University, Hao Zhang, August 2021

(a) Structured grid

(b) Unstructured quad grid

(c) Comparison of structured and unstructured grids

127
Texas Tech University, Hao Zhang, August 2021

0.14 Structured grid 12.73%


0.12 Unstructured quad-grid

0.10

Error
0.08

0.06
4.45%
3.77%
0.04
2.46%
0.02
0.36%
0.00%
0.00
10m 20m 30m
Grid size

(b) The errors comparing with 10m structured grid


Figure 5.20 Cumulative gas production under different grid types and sizes

However, even for the unstructured quad grid, when the grid size enlarges, it
will still cause errors for the simulation result comparing the small-sized grids. It is
because the large grid will cause a sharp pressure gradient between the matrix and
fractures and reduce the simulation accuracy. To alleviate the errors caused by the grid
size, a local grid refinement (LGR) is applied to the unstructured quadrangular grid.

Table 5.4 Values of key points

Key points Value (m) Key points Value (m)

B1 30 B7 10
B2 30 B8 10
B3 30 B9 10
B4 30 B10 10
B5 30 B11 30
B6 10 B12 30

As proposed in Section 5.2.2, the size of the unstructured quad grid is controlled
by the key points, so we can set different values to these key points in Figure 5.13 to
create varying-sized quadrangular grids. The values for key points are shown in Table
5.4, and the corresponding varying-sized grid map can be seen in Figure 5.21. As can
be seen that, controlled by the values of key points, the LGR to the unstructured

128
Texas Tech University, Hao Zhang, August 2021

quadrangular grid is generated, the region near the horizontal fracturing well is
subdivided by 10m grids, while the region away from the well is subdivided by 30m
grids. For the comparison, triangular grids are generated in the reservoir with the same
grid lengths (10m, 20m, 30m, and varying-sized), which are shown in Figure 5.22.

Figure 5.21 The reservoir models with varying-sized unstructured quad grid

(a) 10m triangular grid (b) 20m triangular grid

(c) 30m triangular grid (d) Varying-sized triangular grid


Figure 5.22 The reservoir models with different-sized triangular grids

129
Texas Tech University, Hao Zhang, August 2021

The corresponding pressure distributions after 20 years of production of the


varying-sized quadrangular and triangular grids are shown in Figure 5.23, and the
cumulative gas production results are displayed in Figure 5.24. Besides, the total grid
number, and the simulation time for different grids, are compared in Figure 5.25.

(a) Varying-sized unstructured quad grid (b) Varying-sized triangular grid

(c) 10m triangular grid (d) 20m triangular grid

(e) 30m triangular grid


Figure 5.23 Reservoir pressure (MPa) after 20 years of production for different grid-
based reservoirs

130
Texas Tech University, Hao Zhang, August 2021

(a) Unstructured quad grid

(b) Triangular grid


0.06
Triangular grid
0.05 Unstructured quad-grid 4.45%
3.90%
0.04
Error

0.03 2.46%

0.02
1.09%
0.01 0.47% 0.58%
0.27% 0.36%
0.00
10m 20m 30m Varying sized
Grid size

(c) The error comparing with 10m structured grid


Figure 5.24 The cumulative gas production under different grids

131
Texas Tech University, Hao Zhang, August 2021

120
10000 Structured grid Structured grids
100

Simulation time (S)


Unstruc-quad grid Unstruc-quad grid

Number of grids
8000
Triangular grid 80 Triangular grid
6000
60
4000
40

2000 20

0 0
10 m 20 m 30 m Varying 10 m 20 m 30 m Varying
sized sized
Grid size Grid size

(a) The number of grids (b) The simulation time


Figure 5.25 The statistics of grid number and simulation time for different grids

Comparing the simulation results of different sized unstructured quadrangular


grids in Figure 5.24-a, we can see that the cumulative gas production for the varying-
sized quadrangular grid is almost similar to the 10m grid, the simulation error is reduced
significantly; while the simulation time for the varying-sized quadrangular grid is only
half of the time spend for 10m grid (Figure 5.25-b). It means that the varying-sized
quadrangular grid leads to high simulation accuracy with shorter simulation time.
Besides, when comparing the unstructured quadrangular grid with the triangular grid,
we can also see that the triangular grid can also adaptive to the irregular boundaries and
achieves the same simulation accuracy as the unstructured quad grid (Figure 5.24).
However, the triangular grid generates larger numbers of grids than the unstructured
quadrangular grid at the same grid lengths (Figure 5.22, Figure 5.25-a); more grids will
prolong the simulation time and reduce the calculation efficiency. This is revealed
obviously in Figure 5.25-b, the unstructured quadrangular grid saves nearly half of the
simulation time compared to the triangular grid. To sum up all the analysis above, it is
proved that the unstructured quadrangular grid is the most excellent grid among the
three kinds of grids, which can adaptive to complex-shaped reservoirs, maintain high
simulation accuracy and reduce simulation time simultaneously.

132
Texas Tech University, Hao Zhang, August 2021

5.4.2 Case 2: Hydraulic fractured well optimization based on the unstructured


quad-grid based EDFM.

In this case, single and multiple hydraulic fractured wells are optimized in the
UnQ-EDFM to further verify the reliability and superiority of this model. The reservoir
model adopted in this case is the same as that used in case 1, except for the reservoir
permeability. In case 1, the reservoir permeability is defined as homogenous, but in this
case, a heterogeneous permeability field is used. The original permeability field of the
Brugge reservoir is converted into the nano-Darcy scale and matched to the
corresponding grid maps by the weighted average method, which is shown in Figure
5.26. Other geological and fluid parameters are the same as those used in Case 1 and
shown in Table 5.2, Table 5.3, and Figure 5.10, respectively.

(a) Varying-sized unstructured quad grid (b) 10m structured grid

(c) 20m structured grid (d) 30m structured grid


Figure 5.26 Heterogeneous permeability field with different types of grids (nD)

First, a multi-stage hydraulic fractured well is generated in this reservoir, the HF


aperture is the same as that used in Case 1. Since the reservoir is heterogeneous and

133
Texas Tech University, Hao Zhang, August 2021

with an irregular shape, different well placements will have different productions to a
large extend, so the optimization method based on the WOA in Section 4 is applied to
achieve optimum well placement and HF design. Fracturing-related parameters,
including the wellhead location, well length, HF spacing, as well as HF half-length, are
chosen as the optimization parameters. The varying-sized unstructured quadrangular
grid is constructed in the reservoir model based on the key point values shown in Table
5.4, to make sure the region with higher permeability is covered by refined grids. As a
comparison, structured grids with different sizes are also generated. All of these grids
are shown in Figure 5.26.

Table 5.5 Optimization variables and search space for single well optimization

Optimization variables Search space

The wellhead (m) Reservoir domain


The fracture spacing (m) [10, 100]
The fracture half-length (m) [0, 300]
The well length (m) [100, 1500]

Table 5.6 Economic parameters and their values

Parameters Value

The price of shale gas 2 $/MSCF


The cost of well drilling 750 $/m
The cost of fracturing 300 $/m
Annual discount rate 0.1

The optimization is processed according to the workflow mention in Section


5.3.4 and the flowchart in Figure 5.8. In the beginning, the initial values for optimization
variables are assigned randomly in the search space shown in Table 5.5. The hydraulic
fractured well is generated based on the optimization variables and then implemented
into the EDFM-based shale gas model for production simulation, working with a
constant BHP of 3.4MPa for 20 years. After the simulation, the cumulative gas
production and the parameters of the hydraulically fractured well are recorded and

134
Texas Tech University, Hao Zhang, August 2021

adopted to calculate the NPV in Eq.(5-2), the relative economic parameters are shown
in Table 5.6. Then the optimization variables are optimized according to the WOA
algorithm, the whale number for WOA is set as 20. The optimization processes
cyclically with the total optimization step of 50. The optimization framework is
constructed using Matlab software, and a parallel pool is employed to reduce the
calculation time. Based on these parameters, the well placement optimization is
performed under different grid maps, respectively. After optimization, the optimum well
placements together with the pressure fields after 20 years of production, under different
grid maps, are shown in Figure 5.27; the corresponding optimum accumulative shale
gas production, together with the NPV values’ variations during the optimization, for
grid types, are shown in Figure 5.28. Furthermore, the simulation times for all grid types
are displayed in Figure 5.29, the optimum values of the optimization variables for all
grid types are listed in Table 5.7.

(a) Varying-sized unstructured quad grid (b) 10m structured grid

(c) 20m structured grid (d) 30m structured grid


Figure 5.27 The optimum well placement and the corresponding pressure field (MPa)
for single well optimization

135
Texas Tech University, Hao Zhang, August 2021

(a) The accumulative shale gas production

(b) The NPV variation during the optimization


Figure 5.28 The comparison of optimization results under different grids for the single
well optimization

140
Toal optimization time (min)

120

100

80

60

40

20

0
10m structured 20m structured 30m structured Unstructured
grid grid grid quad grid

Grid types

Figure 5.29 Comparison of optimization times under different grids for single well
optimization

136
Texas Tech University, Hao Zhang, August 2021

Table 5.7 The optimum values after the optimization under different grids for single
well optimization

Varying-sized
Optimization 30m structured 20m structured 10m structured
unstructured
Variables grid grid grid
quad-grid

The wellhead location (m) (321.17, 852.58) (454.34, 843.00) (512.94, 791.77) (495.69, 818.51)
The fracture spacing (m) 29.27 27.77 25.31 25.02
The fracture half-length (m) 121.53 113.02 119.98 112.34
The well length (m) 878.24 888.72 885.88 896.01

As shown in Figure 5.27, the optimum well placements for all the grid types are
in the region with higher permeability after optimization, but different grid types lead
to different optimization results. As can be seen that the optimum well placement
obtained in the varying-sized unstructured quadrangular grid is similar to those obtained
in 10m and 20m structured grids, while there is a big difference for the well placement
obtained in the 30m structured grid. Figure 5.28 also reveals that the cumulative gas
productions obtained from the varying-sized unstructured quad grids and the 10m
structured grid are almost the same, but there are obvious differences for those from the
20m and 30m grids. It reveals that, for the structured grid, the deformation of the
reservoir becomes more severe as the grid size enlarges from 10m to 30m, which
reduces the accuracy of the optimization significantly; whereas, the varying-sized
unstructured quadrangular grid maintains the optimization accuracy within a high level.

Comparing the NPV values in Figure 5.28 and the total optimization times in
Figure 5.29 we can see that enlarging the grid size for the structured grids will help to
reduce the optimization time, but will also cause the reduction of the optimization
accuracy; the NPV from the 30m structured grids is far from the true optimum value.
However, the varying-sized unstructured quadrangular grid achieves a highly accurate
NPV similar to the 10m structured grid, while costs only half of the optimization time.
It is because that the unstructured quadrangular grid adapts to the original shape of the
reservoir with no deformation; besides, the refined grids alleviate the sharp pressure

137
Texas Tech University, Hao Zhang, August 2021

gradient between the reservoir matrix and fractures to increase the simulation accuracy,
and the coarse grids reduce the grid number to save the calculation time.

Table 5.8 Optimization variables and search space for multi-well optimization

Optimization variables Search space

The wellhead (m) Reservoir domain


The fracture spacing (m) [10, 100]
The fracture half-length (m) [0, 300]
The well spacing (m) [100, 500]
The well length (m) [100, 1500]

To further validate the feasibility of UnQ-EDFM, a multi-well optimization is


performed based on the same reservoir model. Two horizontal wells with staggering
placed HFs are constructed in the reservoir. The wellhead location, well length, HF
spacing, HF half-length, and well spacing, are chosen as the optimization parameters,
the search space of which is listed in Table 5.8. All the simulation and optimization
related parameters are the same as the above model. After optimization, the optimum
hydraulic fractured well placements together with the pressure fields after 20 years of
production under different grid maps, are shown in Figure 5.30; the corresponding
optimum accumulative shale gas production and the NPV values during the
optimization, for all grid types, are displayed in Figure 5.31. Furthermore, the
optimization times for single and multi-well optimization under all grid types are
displayed in Figure 5.32, and the optimum values of the optimization variables for all
grid types are listed in Table 5.9.

(a) Varying-sized unstructured quad grid (b) 10m structured grid

138
Texas Tech University, Hao Zhang, August 2021

(c) 20m structured grid (d) 30m structured grid


Figure 5.30 The optimum well placement and the corresponding pressure field (MPa)
for multi-well optimization

(a) Accumulative shale gas production (b) NPV variation during the optimization
Figure 5.31 The comparison of optimization results under different grids for multi-well
optimization
160

140 Single well


Multiple wells
Simulation time (S)

120

100

80

60

40

20

0
10m structured 20m structured 30m structured Unstructured
grid grid grid quad grid
Grid size
Figure 5.32 The comparison of optimization time for single and multi-well optimization
under different grids

139
Texas Tech University, Hao Zhang, August 2021

Table 5.9 The optimum values after the optimization under different grids for multi-
well optimization

Varying-sized
Optimization 30m structured 20m structured 10m structured
unstructured
Variables grid grid grid
quad-grid

The wellhead location (m) (541.27, 844.65) (471.81, 855.28) (474.59, 845.26) (475.38, 839.25)
The fracture spacing (m) 25.03 30.57 26.26 27.116
The fracture half-length (m) 63.44 91.34 103.47 98.36
The well spacing (m) 198.61 166.22 163.99 157.27
The well length (m) 925.46 921.85 917.14 906.53

As can be seen that the optimization results for multi-well are similar to those of
the single well. The optimum well placement obtained in the varying-sized unstructured
quad-grid is similar to those in 10m structured grids, while the difference gradually
becomes obvious as the grid enlarges to 30m. The accumulative shale gas production
and the NPV variation in Figure 5.31 also reveal that the varying-sized unstructured
quad-grid maintains the optimization accuracy to a high level, Furthermore, Figure 5.32
shows that the varying-sized unstructured quad-grid significantly reduce the
optimization times compared to the 10m grid; besides, the time for multi-well
optimization doesn’t increase too much compared to single well optimization. It reveals
that, even for the multi-well optimization, the advantage of reducing optimization time
while maintaining accuracy for the UnQ-EDFM is still achieved. The above simulation
results and analysis are convincing enough to prove that the UnQ-EDFM proposed in
this paper is more reliable and efficient than the traditional EDFM when dealing with
complex reservoirs and large amounts of numerical simulation.

140
Texas Tech University, Hao Zhang, August 2021

CHAPTER VI

SURROGATE-ASSISTED MULTI-OBJECTIVE HYDRAULIC


FRACTURE OPTIMIZATION CONSIDERING UNCERTAINTY
Since the reservoir description from well logging and core analysis is limited,
the reservoir models usually could not precisely describe the geological condition of the
entire reservoir, the uncertainty of geological parameters, becomes an unavoidable
factor and presents a significant impact on the reservoir development. However, the
research on the hydraulically fractured well optimization in unconventional reservoirs
is still limited. Few scholars have made a comprehensive study of hydraulically
fractured well optimization considering the geological uncertainty; besides, the novel
optimization strategies such as multi-objective optimization (MOO) as well as the
surrogate model methodology have not attracted much attention in the unconventional
reservoir development.

Facing these problems, a surrogate-assisted hydraulically fractured well multi-


objective optimization study in the unconventional reservoir with geological uncertainty
is proposed in this chapter. In Section 6.1, a state-of-the-art MOO algorithm named the
hybrid multi-objective particle swarm optimization algorithm (HMOPSO) is proposed.
It couples the advantages of MOPSO and NSGA-II and presents a higher accuracy and
efficiency over both of them. A Pareto-ranking scheme is adopted to search for non-
dominated solutions. Then, in Section 6.2, a Kriging model is employed as the surrogate
model, together with the HMOPSO to construct the framework for the surrogated-
assisted MOO. Several stochastic reservoir models are generated, and the expected
value and standard deviation of the NPVs for these stochastic models are regarded as
the objective functions. Finally, two examples are presented in Section 6.3, the
superiority of the HMOPSO against the traditional algorithms is validated using
benchmarks; and also, the proposed surrogated-assisted MOO framework is performed
into the hydraulic fracture well optimization in the shale reservoir, the results are
compared and analyzed.

141
Texas Tech University, Hao Zhang, August 2021

6.1 The multi-objective optimization algorithm

6.1.1 The multi-objective optimization problem and Pareto front

The general multi-objective optimization (MOO) problem aims to optimize two


or more conflicting objectives simultaneously, the multi-objective minimization
problem can be formulated as:

Minimize 𝐹(𝑥) = [𝑓1 (𝑥), 𝑓2 (𝑥), 𝑓3 (𝑥) ⋯ 𝑓𝑛 (𝑥)]𝑇


Subject to 𝑥 ∈ Ω (6-1)

where 𝑥 is the optimization vector and 𝛺 is the decision space; 𝐹(𝑥) is the
vector of objective functions, and 𝑓𝑛 (𝑥) is an objective function within this vector.

Different from the single-objective optimization (SOO), which aims to obtain a


unique optimum solution for the specific objective function, the MOO aims to obtain a
set of trade-off solutions to satisfy all the objectives. During the process of MOO, a
Pareto optimal ranking scheme is used to sort the solutions based on their domination
relationships. Considering two solutions 𝑥1 and 𝑥2 , 𝑥1 is regarded as dominating 𝑥2
when satisfying the following relationship (Zhao et al., 2020b):

∀𝑖 ∈ {1,2 ⋯ 𝑛}, 𝑓𝑖 (𝑥1 ) < 𝑓2 (𝑥2 )


𝑥1 ≺ 𝑥2 if and only if { (6-2)
∃𝑖 ∈ {1,2 ⋯ 𝑛}, 𝑓𝑖 (𝑥1 ) < 𝑓2 (𝑥2 )

where ≺ represents the domination relationship.

For a solution 𝑥 ∗ in the set, it can be regarded as the non-dominated solution if


no other solutions dominate this solution. All the non-dominated solutions make up a
Pareto set (PS), the objective function vector 𝐹(𝑥) obtained from the PS is named the
Pareto front (PF).

𝑃𝐹 = {[𝑓1 (𝑥), 𝑓2 (𝑥), 𝑓3 (𝑥) ⋯ 𝑓𝑚 (𝑥)]𝑇 | 𝑥 ∈ 𝑃𝑆} (6-3)

where 𝑚 means the number of non-dominated solutions in the PS. As shown in


Figure 6.1, the solutions on the PF are the optimal solutions that simultaneously satisfy
all the objectives.

142
Texas Tech University, Hao Zhang, August 2021

Figure 6.1 Dominated, non-dominated, and Pareto front (Mahesh et al., 2016)

6.1.2 The hybrid multi-objective PSO algorithm

Since the optimization searching scheme for MOO is different from the SOO,
the PF cannot be obtained by the traditional optimization algorithms. As illustrated in
the introduction section, the NSGA-II (Deb et al.,2002) and MOPSO (Coello et al., 2004)
are two popular MOO algorithms. However, both of them have their own drawbacks;
the NSGA-II can obtain a good PF, but the searching speed is pretty low; the MOPSO
has a fast searching speed, yet the solutions are easy to get stuck together and cannot
show a good PF. Besides, both of these algorithms do not have good performance with
low particle numbers. Facing these problems, we take synthetic advantages of these two
algorithms and proposed a hybrid multi-objective particle swarm optimization
algorithm (HMOPSO). In the HMOPSO, the optimization searching strategy of
MOPSO is adopted to maintain a high searching speed, and the crowding-distance and
elitist strategy of NSGA-II is employed to maintain a good PF; additionally, a Gaussian
distribution reinitializing strategy is used to increase the efficiency of optimization
searching. The detailed workflow for HMOPSO can be seen as follows, and the
pseudocode of HMOPSO is presented in Algorithm 1.

(1) At the beginning of the optimization, initialize a swarm of 𝑁-dimensional particles


in the searching space, the 𝑖 th particle is 𝑋𝑖 = (𝑥1 , 𝑥2 ⋯ 𝑥𝑁 ), To make the initial
particles cover the feasible region, a Latin hypercube sampling (LHS) method is

143
Texas Tech University, Hao Zhang, August 2021

adopted for the initialization (McKay et al., 2000). The corresponding initial
searching velocity is also randomly set as 𝑉𝑖 = (𝑣1 , 𝑣2 ⋯ 𝑣𝑁 ). Then define the
particle historical best position 𝑃𝑋𝑖 and the global best position 𝐺𝑋. Different from
the single objective PSO, the global best position 𝐺𝑋 is not a unique value but all
the non-dominated particles in the Pareto set 𝑃𝑆. At the initial iteration step, the
initial position of the particle is also its historical best position 𝑃𝑋𝑖 = 𝑋𝑖 , the global
best position and the PS is regarded as the particle swarm 𝐺𝑋 = 𝑃𝑆 =
{𝑋1 , 𝑋2 ⋯ 𝑋𝑁𝑝 }, where 𝑁𝑝 is the population of particles in the swarm.

(2) According to current values, update the particle’s searching velocities and positions
based on the following function:

𝑉𝑖𝑘+1 = 𝑐0 𝑉𝑖𝑘 + 𝑐1 𝑟1 (𝑃𝑋𝑖𝑘 − 𝑋𝑖𝑘 ) + 𝑐2 𝑟2 (𝐺𝑋𝑟𝑎𝑛𝑑


𝑘
− 𝑋𝑖𝑘 )
{ (6-4)
𝑋𝑖𝑘+1 = 𝑋𝑖𝑘 + 𝑉𝑖𝑘

where 𝑘 means the 𝑘th iteration step; 𝑐0 , 𝑐1 , and 𝑐2 are the weight coefficients for the
searching velocity, the particle’s historical best position, and the global best position;
𝑘
𝑟1 , 𝑟2 are randomly defined numbers within [0, 1]; 𝐺𝑋𝑟𝑎𝑛𝑑 is a particle randomly
selected from the global best position set 𝐺𝑋 𝑘 .

(3) Based on the new particles, calculate the corresponding objective solutions 𝑓(𝑋𝑖𝑘+1 ),
and update the particle historical best position 𝑃𝑋𝑖𝑘+1 . Sort the domination
relationship of the particles according to the Pareto ranking scheme in Eq.(6-2), pick
up all the non-dominated particles to obtain a new Pareto set 𝑃𝑆 𝑘+1 , and then
combine the new PS with the old one, and delete the duplicated particles.

𝑃𝑆 𝑘+1 = Unique{𝑃𝑆 𝑘+1 ∪ 𝑃𝑆 𝑘 } (6-5)

(4) To avoid the aggregation of particles in the PS, sort the domination relationship of
the 𝑃𝑆 𝑘+1 and then pick up 𝑁𝑝 best non-dominated particles based on the
crowding-distance and elitist strategies of NSGA-II (Deb et al.,2002). This set of
non-dominated particles are regarded as the new global best position set 𝐺𝑋 𝑘+1 . To
prevent premature convergence during the optimization process, the new global best

144
Texas Tech University, Hao Zhang, August 2021

position set 𝐺𝑋 𝑘+1 is mixed with the particle swarm 𝑋 𝑘+1 according to a bias factor
𝛽:

𝐺𝑋 𝑘+1 = {𝐺𝑋1𝑘+1 , 𝐺𝑋2𝑘+1 , ⋯ 𝐺𝑋𝛽𝑘+1 , 𝑋𝛽+1


𝑘+1 𝑘+1
, 𝑋𝛽+2 𝑘+1
, ⋯ 𝑋𝑁𝑝 } (6-6)

The bias factor 𝛽 is not a constant, it varies with the iteration step 𝑘:
𝑘
𝛽=⌈ × 𝑁𝑝⌉ (6-7)
𝐼𝑡𝑒𝑟_𝑚𝑎𝑥

where 𝐼𝑡𝑒𝑟_𝑚𝑎𝑥 is the maximum iteration step, ⌈𝑎⌉ means around up the non-
integer 𝑎 into an integer.

(5) Sort all the particles 𝑋 𝑘+1 from worst to best, pick up the 𝑁𝑝𝑤 worst particles and
reinitialize their positions, we define that 𝑁𝑝𝑤 = ⌈0.2 × 𝑁𝑝⌉ . A Gaussian
distribution reinitialization method is used to reinitialize the worst particles, the new
position of the 𝑗th worst particle can be calculated as (Das and Dennis, 1998):
𝑘+1
𝐺𝑋𝑟𝑎𝑛𝑑 −𝑃𝑋𝑗𝑘+1
𝑋𝑗𝑘+1 = 𝑁 ( 𝑘+1
, | 𝐺𝑋𝑟𝑎𝑛𝑑 − 𝑃𝑋𝑗𝑘+1 |) (6-8)
2

where 𝑁(𝜇, 𝜎) means a random number normally distributed with 𝜇 as the mean
value 𝜎 as the variance. As shown in Figure 6.2, the Gaussian distribution
reinitialization is a smart method, for it gives the worst particle a high probability to
find a better position or even the global best position. The reinitialized particles will
not be updated by Eq.(6-4) in the next iteration step.

Figure 6.2 Gaussian distribution reinitialization (Zapotecas and Coello, 2011)

145
Texas Tech University, Hao Zhang, August 2021

(6) After obtaining the new particles 𝑋𝑖𝑘+1 and their historical best positions 𝑃𝑋𝑖𝑘+1 , as
well as the new global best position set 𝐺𝑋 𝑘+1 , back to step 2 and start the next
iteration step.

(7) Repeat steps 2-6, until the designed step is over or the convergence is met.

Algorithm 1: Pseudocode of the hybrid multi-objective particle swarm optimization (HMOPSO)

1: Initialization Randomly define 𝑁𝑝 initial particles 𝑋, initialize the corresponding historical best solution 𝑃𝑋,
the Global best solution 𝐺𝑋, and the Pareto set 𝑃𝑆, 𝑘 = 1
2: While The stopping criterion is not met
3: For 𝑖 = 1: 𝑁𝑝
4: 𝑉𝑖𝑘+1 = 𝑐0 𝑉𝑖𝑘 + 𝑐1 𝑟1 (𝑃𝑋𝑖𝑘 − 𝑋𝑖𝑘 ) + 𝑐2 𝑟2 (𝐺𝑋𝑟𝑎𝑛𝑑
𝑘
− 𝑋𝑖𝑘 )
5: 𝑋𝑖𝑘+1 = 𝑋𝑖𝑘 + 𝑉𝑖𝑘
6: Calculated the objective solution 𝑓(𝑋𝑖𝑘+1 )
7: End For
8: Sort the domination relationships of particles according to the Pareto ranking scheme in Eq. (6-2), pick up the
nondominated particles to obtain a new Pareto set 𝑃𝑆 𝑘+1
9: 𝑃𝑆 𝑘+1 = unique(𝑃𝑆 𝑘+1 ∪ 𝑃𝑆 𝑘 )
10: Sort the domination relationship of the 𝑃𝑆 𝑘+1 and pick up 𝑁𝑝 best non-dominated particles as the new global
best position set 𝐺𝑋 𝑘+1 , according to the crowding-distance and elitist strategies of NSGA-II
𝑘
11: 𝛽 = 𝑟𝑜𝑢𝑛𝑑 (𝐼𝑡𝑒𝑟_𝑚𝑎𝑥 × 𝑁𝑝)

12: For 𝑖 = 1: 𝑁𝑝 %%---------------Bias operator for the global best position---------------


13: If 𝑖 ≤ 𝛽
14: 𝐺𝑋𝑖𝑘+1 = 𝐺𝑋𝑖𝑘+1
15: Else
16: 𝐺𝑋𝑖𝑘+1 = 𝑋𝑖𝑘+1
17: End If
18: End For
19: 𝑁𝑝𝑤 = 𝑟𝑜𝑢𝑛𝑑(0.2 × 𝑁𝑝)
20: Sort the particles 𝑋 𝑘+1 from worst to best, pick up the 𝑁𝑝𝑤 worst particles
21: For 𝑗 = 1 to 𝑁𝑝𝑤 %%---------------Reinitialization operator for the worst particles---------------
𝑘+1
𝐺𝑋𝑟𝑎𝑛𝑑 −𝑃𝑋𝑗𝑘+1
22: 𝑋𝑗𝑘+1 = 𝑁 ( 𝑘+1
, | 𝐺𝑋𝑟𝑎𝑛𝑑 − 𝑃𝑋𝑗𝑘+1 |)
2

23: End For


24: 𝑘 = 𝑘 + 1
25: End While
26: Output the final Pareto Front

146
Texas Tech University, Hao Zhang, August 2021

6.2 The Kriging surrogate model

6.2.1 The principle of Kriging model

For the reservoir optimization problem with uncertainty, a series of stochastic


models are established to describe the geological uncertainty, which enlarges the
calculation amount to hundreds of times than the optimization under certain conditions
and leads to pretty low optimization efficiency. The surrogate model is an effective
method to increase optimization efficiency. By training the mathematical parameters
with the real simulation data, the results obtained from the surrogate model will be more
and more converged to that from the numerical simulation, then the computationally
expensive numerical simulations can be replaced by the efficient surrogate model. For
this research, a Gaussian process Kriging model is adopted as the surrogate model.

For any individual 𝑋 ∈ 𝑅𝑛 , the Gaussian process model can be written as:

𝐹(𝑋) = 𝜇 + 𝑒(𝑋) (6-9)

where 𝐹(𝑋) is the unknown function to be estimated; 𝜇 is the known


approximated function, 𝑒(𝑋)~𝑁(0, 𝜎 2 ) is the stochastic component with Gaussian
distribution.

Consider individuals 𝑋, 𝑋′ ∈ 𝑅𝑛 , the correlation 𝑐(𝑋, 𝑋′) between the


components 𝑒(𝑋) and 𝑒(𝑋′) can be calculated by (Rasmussen and Williams, 2006):

𝑐(𝑋, 𝑋′) = 𝑒𝑥𝑝(− ∑𝑛𝑖=1 𝜃𝑖 |𝑥𝑖 − 𝑥𝑖 ′|2 ) (6-10)

where 𝜃𝑖 the parameter indicates the impact of 𝑥𝑖 on the function 𝐹(𝑋).

Consider 𝐾 points 𝑋1 , 𝑋2 , ⋯ 𝑋𝐾 ∈ 𝑅𝑛 and the corresponding real objective


functions 𝐹1 , 𝐹2 , ⋯ 𝐹𝐾 , the Kriging model can be built assuming that this is no error
between the approximated and real objective function, and the likelihood can be
formulated as (Rasmussen and Williams, 2006):

1 (𝐅−𝜇𝐪)𝑇 𝐂 −1 (𝐅−𝜇𝐪)
𝑃= 𝑒𝑥𝑝 (− ) (6-11)
(2𝜋𝜎 2 )𝐾/2 √𝑑𝑒𝑡(𝐂) 2𝜎 2

147
Texas Tech University, Hao Zhang, August 2021

where 𝐅 = (𝐹1 , 𝐹2 , ⋯ 𝐹𝐾 )𝑇 ; 𝐂 is a 𝐾 × 𝐾 dimensional matrix with the (𝑖, 𝑗) -


element of 𝑐(𝑋𝑖 , 𝑋𝑗 ); 𝐪 is a 𝐾-dimensional unit column vector.

When the likelihood 𝑃 is maximized, the expected value 𝜇 variance 𝜎 2 are


formulated as (Rasmussen and Williams, 2006):

𝐪𝑻 𝐂 −𝟏 𝐅
𝜇̂ =
𝐪𝑻 𝐂 −𝟏 𝐪
{ (6-12)
(𝐅−𝐪𝜇)𝑇 𝐂 −1 (𝐅−𝐪𝜇)
𝜎̂ 2 =
𝐾

Substitute the values of 𝜇̂ and 𝜎̂ 2 in Eq.(6-12) to Eq.(6-11), the likelihood


function will depend only on the 𝜃𝑖 . Eq.(6-11) can be maximized by the proper value of
𝜃𝑖 , and the proper 𝜃𝑖 can be obtained by training the Kriging model with the real
simulation result. The maximum likelihood function means the least error between the
surrogate model and the real numerical model.

6.2.2 A case for Kriging surrogate model validation

The feasibility of the Kriging surrogate model is testified based on a benchmark


function. The formulation of the function is listed as follows, and the image of the
function can be seen in Figure 6.3.

𝑓(𝑥, 𝑦) = 𝑠𝑖𝑛(𝑥 2 + 𝑦 2 )⁄√𝑥 2 + 𝑦 2 , 𝑥, 𝑦 ∈ [−20, 20] (6-13)

Figure 6.3 Image of the benchmark function

148
Texas Tech University, Hao Zhang, August 2021

The Kriging model generates different numbers of sample points using the LHS
within the range, the populations of sample points are 10, 50,100, and 200. The solutions
of these sample points are calculated by Eq. (6-13), and the approximation functions are
generated using the algorithms in Section 6.2.1. The images of the approximation
functions under different sample points are displayed in Figure 6.4, where the red dots
represent the locations of the sample points. As can be seen that, the shape of the
approximation function gets closer to the real function as the population of sample
points increases; when the sample points’ population is larger than 100, the shape of the
approximation function is almost the same as the real function.

(a) 10 sample points (b) 50 sample points

(a) 100 sample points (b) 200 sample points


Figure 6.4 Image of the approximation function with different sample points

149
Texas Tech University, Hao Zhang, August 2021

Besides, the comparison of function values between the real function and
approximate functions under different sample points is displayed in Figure 6.5, in which
the green lines represent the optimal condition that the approximate function value
equals the real function value. 50 test points are randomly picked up from different
approximated functions and compared with the corresponding real function values. As
can be seen that, for the approximate function built up with 10 sample points, the test
points are scattered and distribute away from the green line, which means a big
difference between the approximate function and the real function. As the sample points
increase, the test points distribute closer to the green line and even coincide with the
green line when the sample points are larger than 100, which means the approximate
function is the same as the real function.

(a) 10 sample points (b) 50 sample points

(a) 100 sample points (b) 200 sample points


Figure 6.5 Comparison of the approximate and real function values

150
Texas Tech University, Hao Zhang, August 2021

6.3 The surrogate-assisted multi-objective hydraulically fractured well


optimization considering uncertainty

6.3.1 The reservoir geological uncertainty description

For the fractured shale reservoir, the most essential uncertain parameter is the
natural fracture (NF) distribution, which shows a remarkable effect on the hydraulically
fractured well’s productively and brings about a significant risk for the development.
The distribution of NFs in the formation is complex, yet can still be described by two
main characteristics, the frac-number and the frac-length. Based on decades of micro-
seismic logging and core analysis, scholars have found that in unconventional reservoirs
the relation between the number and length of NFs could be summarized into a power-
law probability distribution trend (Segall and Pollard, 1983; Davy, 1993; Wu and Olson,
2016). In previous Section 3.1.3, a power-law probability function is illustrated in
Eq.(3-13) to describe the NF distribution. A real NF distribution map and the NF
distribution map generated based on the power-law probability function are shown in
Figure 6.6. As can be seen that the frac-length and frac-number of the real NF conform
pretty well to the power-law function, and the NF map generated based on the power-
law distribution function shows a similar distribution trend as the real NF map, the
number of NFs reduces as the length of NFs increases, and the entire generated NF
pattern is pretty closed to the real NF pattern.

(a) Real NF pattern (Davy, 1993) (b) Power-law probability NF pattern


Figure 6.6 Comparison of real NF map and power-law probability NF map

151
Texas Tech University, Hao Zhang, August 2021

By using the power-law distribution function, we can describe the relationship


between the frac-length and the corresponding frac-number, which presents the overall
distribution of the NF map; however, the exact distribution of every NF cannot be
precisely depicted, different NF patterns can be obtained even with the same parameter,
which brings unavoidable geological uncertainty. Scholars usually use a set of models
that contain as many as possible distribution maps to describe the geological uncertainty.
As shown in Figure 6.7, the reservoir geological uncertainty is described by a set of
stochastic NF patterns which are constructed based on the power-law probability
distribution function with the same parameter.

(a) NF pattern 1 (b) NF pattern 2

(c) NF pattern 3 (d) NF pattern N


Figure 6.7 Stochastic NF models generated using the same parameter

152
Texas Tech University, Hao Zhang, August 2021

6.3.2 The multi-objective functions for hydraulically fractured well optimization

As mentioned in the above chapters, the ultimate objective of reservoir


optimization problems is usually obtaining the highest economical benefit from the
recovery while spending the lowest cost for the investment, so the net present value
(NPV) is an effective economical parameter to adopted as the objective function for the
hydraulically fractured well optimization (Rammay et al., 2016). The formulation of
NPV can be seen in Eq.(5-2).

Considering the geological uncertainty, the stochastic reservoir models are all
simulated, so a set of NPVs will be obtained. To fully consider the balance between the
expected value and the possible development risk, the multi-objective optimization
method (MOO) is adopted, in which the expected value 𝐸𝑁𝑃𝑉 and standard deviation
𝜎𝑁𝑃𝑉 of these NPVs are regarded as the objective functions for the hydraulically
fractured well optimization, which are calculated as:
1
𝐸𝑁𝑃𝑉 = ∑𝑚
𝑖=1 𝐽𝑖 , (𝑖 = 1,2, … 𝑚) (6-14)
𝑚

1
𝜎𝑁𝑃𝑉 = √ ∑𝑚
𝑖=1(𝐸𝑁𝑃𝑉 − 𝐽𝑖 )
2 (6-15)
𝑚

where 𝐽𝑖 is the NPV of the 𝑖 th reservoir model and 𝑚 is the total number of
stochastic reservoir models.

Figure 6.8 The illustration of NPV distribution

153
Texas Tech University, Hao Zhang, August 2021

As is shown in Figure 6.8, all the NPVs obtained from the stochastic reservoir
models conform to a normal distribution, the expected value 𝐸𝑁𝑃𝑉 represents the
expected economical revenue from the development and the standard deviation 𝜎𝑁𝑃𝑉
reflect the possible development risk. The objective of the optimization is to obtain the
maximum 𝐸𝑁𝑃𝑉 while minimizing the 𝜎𝑁𝑃𝑉 .

6.3.3 Workflow of surrogate-assisted multi-objective hydraulically fractured well


optimization

Combining the Kriging surrogate model with the HMOPSO algorithm, and
make the expected value and standard deviation the objective functions, the surrogate-
assisted multi-objective hydraulically fractured well optimization is constructed. The
hydraulic fracturing operation parameters are regarded as the optimization variables.
The shale gas model together with the EDFM mentioned in Section 4.2 are used for the
production simulation. The optimization and surrogate model training are performed
simultaneously. First, the initial particle swarm is regarded as the initial data set to train
the surrogate model. Because of the limited particle number, the surrogate model will
not be accurate enough. Then in each of the following steps, randomly pick up some
representative particles from the swarm that is calculated by the surrogate model and
testify their accuracy with the corresponding numerical simulations. If not accurate
enough, the particle swarm will be calculated by the numerical simulation to obtain the
accurate solutions, and then add these particles into the data set to train the surrogate
model. So during the optimization steps, the data set will gradually enlarge, makes the
surrogate training getting closer to the numerical model. After the Kriging model is
totally matured and converged to the real numerical model, it will be used as the
surrogate model in the following iterations. The workflow is illustrated in detail as
follows and is displayed more intuitively in Figure 6.9.

(1) Firstly, integrate the optimization variables to form a particle 𝑋 = (𝑥1 , 𝑥2 ⋯ 𝑥𝑛 ),


where 𝑥𝑖 is the 𝑖 th optimization variable. Generate the initial particle swarm using
the LHS method. Then initialize the particle historical best solution 𝑃𝑋, the Global

154
Texas Tech University, Hao Zhang, August 2021

best solution set 𝐺𝑋, and the Pareto set 𝑃𝑆. Generate the stochastic reservoir models.

(2) Based on the stochastic reservoir models, establish the corresponding shale gas
EDFM production models according to Section 2. For each particle 𝑋𝑖𝑘 in the
swarm, construct the corresponding hydraulically fractured well according to the
value of 𝑋𝑖𝑘 , and bring it to every stochastic EDFM model and perform the
production simulation. After the simulation, calculate the NPV for each stochastic
EDFM model using Eq.(5-2) and then obtain the expected value 𝐸𝑁𝑃𝑉 𝑘𝑖 and
standard deviation 𝜎𝑁𝑃𝑉 𝑘𝑖 using Eq.(6-14) and Eq.(6-15). Repeat this process to all
the particles in the swarm.

(3) Add the particle swarm 𝑋 𝑘 into the current input data set to form a new set of input
data, add the set of expected value 𝐸𝑁𝑃𝑉 𝑘 and standard deviation 𝜎𝑁𝑃𝑉 𝑘 into the
current set of objective functions, then train the Kriging model using the updated
input data set and the objective functions. For the initial iteration step, the initial
particle swarm is regarded as the initial input data set, the same for the initial
objective functions.

(4) Update the particles 𝑋𝑖𝑘+1 , the historical best solution 𝑃𝑋𝑖𝑘+1 , the global best
solution set 𝐺𝑋 𝑘+1 , and the Pareto set 𝑃𝑆 𝑘+1 according to the HMOPSO algorithm
in Section 6.1.

(5) Randomly select three representative particles from the swarm, bring them into
EDFM simulation and obtain three real solutions according to the procedure in step
2, and then bring them into the Kriging model to obtain three approximate solutions.

(6) If the error between the real and approximate solutions for these three particles is
within the allowable range, calculate the solutions of the entire particle swarm using
the Kriging model and back to step 4 for the new iteration; if not, back to step 2 to
calculate the solutions of the entire particle swarm using EDFM simulation, then
add these particles into the current data set and retrain the Kriging model.

(7) Continue repeating the iterations to the designed step, then output the final PF.

155
Texas Tech University, Hao Zhang, August 2021

Start

Initialize particle
swarm using LHS

Stochastic EDFM models Kriging


Add
particle
surrogate model
Model 1 Model 2 Model n
swarm into
the input Calculate objective
NPV 1 NPV 2 NPV n data set functions using
surrogate model

Objective functions
Calculate the
ENPV σNPV Calculate the
solutions of solutions of
particle swarm particle swarm
using EDFM Update particle swarm using Kriging
simulation using HMOPSO surrogate model

Pick up 3 representative
particles to testify their
accuracy

No Particles are Yes


accurate?
If the iteration
step is over
Output the final
Pareto front

Figure 6.9 Flowchart of the surrogate-assisted multi-objective optimization

6.4 Examples

6.4.1 Case 1: Validation of the surrogate-assisted HMOPSO algorithm

In this case, the feasibility and superiority of the surrogate-assisted HMOPSO


algorithm proposed in the above sections are validated based on several benchmark
functions. First, the performance of HMOSP proposed in Section 6.1 is verified against
the classical MOO algorithms, including the NSGA-II and MOPSO. Several widely
used multi-objective benchmark functions (ZDT1, ZDT2, ZDT3, and ZDT6) are

156
Texas Tech University, Hao Zhang, August 2021

employed as the objective functions (Zitzler et al., 2000), the formulation of which can
be seen below:

(1) The formulation for benchmark function ZDT1 is:

𝑓1 (𝑥) = 𝑥1 𝑥 ∈ [0,1]
𝑥1
𝑓2 (𝑥) = 𝑔(𝑥) ⌊1 − √ ⌋ (6-16)
𝑔(𝑥)

{ 𝑔(𝑥) = 1 + 9 (∑𝑛𝑖=2 𝑥𝑖 )⁄( 𝑛 − 1)

(2) The formulation for benchmark function ZDT2 is:

𝑓1 (𝑥) = 𝑥1 𝑥 ∈ [0,1]
𝑥1 2
{𝑓2 (𝑥) = 𝑔(𝑥) ⌊1 − ( ) ⌋ (6-17)
𝑔(𝑥)
𝑔(𝑥) = 1 + 9 (∑𝑛𝑖=2 𝑥𝑖 )⁄( 𝑛 − 1)

(3) The formulation for benchmark function ZDT3 is:

𝑓1 (𝑥) = 𝑥1 𝑥 ∈ [0,1]
𝑥1 𝑥1
𝑓2 (𝑥) = 𝑔(𝑥) ⌊1 − √( )−( ) 𝑠𝑖𝑛(10𝜋𝑥1 )⌋ (6-18)
𝑔(𝑥) 𝑔(𝑥)

{𝑔(𝑥) = 1 + 9 (∑𝑛𝑖=2 𝑥𝑖 )⁄( 𝑛 − 1)

(4) The formulation for benchmark function ZDT6 is:

𝑓1 (𝑥) = 𝑥1 𝑥 ∈ [0,1]
𝑓 (𝑥) 2
𝑓2 (𝑥) = 𝑔(𝑥) ⌊1 − ( 1 ) ⌋
𝑔(𝑥) (6-19)
1
(∑𝑛
𝑖=2 𝑥𝑖 ) 4
{𝑔(𝑥) = 1 + 9 [ (𝑛−1)
]

where 𝑓1 (𝑥) and 𝑓2 (𝑥) are two objective functions of variation 𝑥.

The number of optimization objectives is 2. The dimension of variables is


designed as 100, and the range of the variables is 𝑥𝑖 ∈ [0,1], (𝑖 = 1,2 ⋯ 100). The
population of particles is 50 for all the algorithms and the iteration steps are designed
as 50. The corresponding optimization results for these benchmark functions under
different MOO algorithms are displayed in Figure 6.10.

157
Texas Tech University, Hao Zhang, August 2021

(a-1) ZDT1 with NSGA-II (a-2) ZDT2 with NSGA-II

(a-3) ZDT3 with NSGA-II (a-4) ZDT6 with NSGA-II

(b-1) ZDT1 with MOPSO (b-2) ZDT2 with MOPSO

158
Texas Tech University, Hao Zhang, August 2021

(b-3) ZDT3 with MOPSO (b-4) ZDT6 with MOPSO

(c-1) ZDT1 with HMOPSO (c-2) ZDT2 with HMOPSO

(c-3) ZDT3 with HMOPSO (c-4) ZDT6 with HMOPSO


Figure 6.10 The optimization results for different benchmark functions with different
MOO algorithms

159
Texas Tech University, Hao Zhang, August 2021

As is shown in Figure 6.10, the green curves represent the true PFs of these
benchmark functions, the back circles and red stars represent the dominated and non-
dominated solutions, respectively. The figures reveal that the optimization results of the
same benchmark function with different algorithms show a remarkable difference. The
NSGA-II maintains a good shape of PF but the accuracy and efficiency of optimization
searching are pretty low, after 50 steps the particles don’t research the true PF. The
searching speed for the MOPSO is remarkably faster than the NSGA-II, for the particles
can reach the true PF after the optimization; however, it is hard for MOPSO to maintain
the shape of PF, the particles are easy to attract each other and finally get stuck together.
This problem is even serious in the benchmark ZDT2, for all the particles get stuck into
one point. Meanwhile, the HMOPSO presents an excellent performance which
including both the optimization efficiency and the good PF shape, all the particles reach
the true PF and distribute evenly within the PF. These optimization results demonstrate
the reliability and superiority of the HMOPSO when dealing with complex multi-
objective problems. Additionally, to test the performance of HMOPSO with low particle
populations, it is adopted in the ZDT1 again with different particle numbers (50, 20,10,
and 5). As a comparison, the NSGA-II and the MOPSO are also adopted with the same
particle populations. The optimization results can be seen in Figure 6.11.

(a-1) NSGA-II with 50 particles (a-2) NSGA-II with 20 particles

160
Texas Tech University, Hao Zhang, August 2021

(a-3) NSGA-II with 10 particles (a-4) NSGA-II with 5 particles

(b-1) MOPSO with 50 particles (b-2) MOPSO with 20 particles

(b-3) MOPSO with 10 particles (b-4) MOPSO with 5 particles

161
Texas Tech University, Hao Zhang, August 2021

(c-1) HMOPSO with 50 particles (c-2) HMOPSO with 20 particles

(c-3) HMOPSO with 10 particles (c-4) HMOPSO with 5 particles


Figure 6.11 Optimization results for different MOO algorithms in ZDT1 with different
particle population

As is shown in Figure 6.11, the optimization accuracy of NSGA-II and MOPSO


all decrease dramatically with the particle population reduces. The particles either
cannot reach the real PF or the PF obtained cannot maintain a good shape. However, the
HMOPSO keeps an excellent optimization accuracy even with the particle population
reduces. Since the reinitialization operator in the HMOPSO increases the utilization of
the particles, it can still maintain a high accuracy even with 5 particles, and the particles
all distribute evenly to cover the whole PF. The results prove the superiority of the
HMOPSO with low particle populations.

162
Texas Tech University, Hao Zhang, August 2021

Secondly, based on the HMOPSO and the Kriging surrogate model, the
surrogate-assisted HMOPSO algorithm proposed in Section 6.3.3 is verified with the
benchmark function ZDT1. The designed optimization iteration step is 50, and the
particles’ population is defined as 20. During the optimization, the iteration times of
training the surrogate model are recorded. The final optimization result can be seen in
Figure 6.12-a, which shows that the PF obtained by the surrogate-assisted HMOPSO
matches the true PF pretty well. The bar chart in Figure 6.12-b indicates that the
algorithm takes 11 iteration steps to train the surrogate model, and 39 steps of
optimization searching are done by the well-trained surrogated model. These results
demonstrate that the surrogate-assisted HMOPSO could obtain the same optimization
accuracy and save the majority percentage calculation time.

(a) The optimization result


60
50
40
Iteration times

30
20
10
0
Train the Surrogate-assitied Total iteration
surrogate model optimization times
Operations
(b) The statistic of different iteration steps
Figure 6.12 The results of surrogate-assisted HMOPSO based on ZDT1

163
Texas Tech University, Hao Zhang, August 2021

6.4.2 Case 2: Multi-objective hydraulically fractured well optimization in the


uncertain reservoir using the surrogate-assisted HMOPSO

In this case, the surrogate-assisted HMOPSO is applied to a hydraulically


fractured well optimization in a fractured shale gas reservoir with geological uncertainty.
The reservoir model used for this optimization is also the Brugge reservoir, which is a
benchmark reservoir model widely applied in reservoir numerical simulations (Chen et
al., 2010; Zhang et al., 2019). As is shown in Figure 6.13, the scale of the Brugge
reservoir model is 1600m×800m, with a thickness of 50m. The background color
represents the Brugge reservoir’s original permeability field, which is not used in this
model. The UnQ-EDFM mentioned in Chapter 5 is adopted for the numerical
simulation. The boundaries of the reservoir are constrained by red segments and the
unstructured quad-grid mentioned in Section 2.3 is used for the gridding, with a grid
length of 20m. As is shown in Figure 6.13-b, better than the structured grid, the
unstructured quad grid presents excellent adaption to the irregularly shaped boundaries
and the reservoir’s original shape is not deformed in the grid map, which increases the
simulation accuracy (Zhang and Sheng, 2021).

(a) Horizontal view of the reservoir

164
Texas Tech University, Hao Zhang, August 2021

(b) The grid map of the reservoir


Figure 6.13 The Brugge reservoir model

The shale gas flow model illustrated in case Section 4.2.3 is integrated into the
EDFM to simulate the production of shale gas in fractured reservoirs, Similar to the
previous shale gas reservoir model in the previous case of Section 5.4, the geological
parameters for this reservoir model are shown in Table 5.2. There is only one
component of CH4 in the shale gas, with the properties listed in Table 5.3. For the shale
gas flowing in the rock matrix and fracture systems, the adsorption and transport
mechanisms are considered based on the corresponding correction factors in Figure 5.10.
The initial pressure of the reservoir is 20MPa, and during the production, the reservoir
is a closed boundary. The distribution of NFs is considered as the uncertain geological
parameter. As described in Section 6.3.1, the NF distribution map is simulated based
on the power-law probability function, with the corresponding parameters 𝛼 = 300,
𝑎 = 1.5, 𝑙 ∈ [50,200]𝑚, 𝑑𝑙 = 10𝑚, 50 stochastic models are generated to describe
the geological uncertainty, some of which are shown in Figure 6.14.

165
Texas Tech University, Hao Zhang, August 2021

(a) Stochastic NF map 1 (b) Stochastic NF map 5

(c) Stochastic NF map 10 (d) Stochastic NF map 20

(e) Stochastic NF map 30 (f) Stochastic NF map 50


Figure 6.14 The stochastic NF distribution maps

166
Texas Tech University, Hao Zhang, August 2021

After the establishment of the fractured shale gas reservoir with geological
uncertainty, a multi-stage hydraulically fractured horizontal well is generated in this
reservoir. The wellbore radius is 0.1m, the HFs are regarded as the planar bi-wing
fractures and could penetrate and connect the encountering NFs into fracture networks.
The fracture aperture is 3mm; the initial HF permeability is 100md and that for the NF
is 10md. The entire time for shale gas production is 20 years, with a constant BHP of
4MPa. The operational parameters related to the hydraulically fractured well
construction, involving the wellhead location, the well horizontal length, the HF half-
length, and the HF space, are selected to be the optimization variables, with the
corresponding search space shown in Table 6.1.

Table 6.1 Optimization variables and their search space

Optimization variables Search space

The wellhead location Reservoir domain


The well length (m) [100, 1500]
The fracture half-length (m) [10, 300]
The fracture spacing (m) [10, 100]

Table 6.2 Economic parameters and their values

Parameters Value Unit

The cost of well drilling 750 $/m


The cost of fracturing 300 $/m
The price of shale gas 2 $/MSCF
The annual discount rate 0.1

Since there are huge differences between the search spaces of different
optimization variables which will affect the accuracy of surrogate modeling and
optimization, the optimization variables are first normalized into the same search space
of [0, 1], based on the following equation:

(𝑋𝑖 −𝑋𝑖𝐹𝑙𝑜𝑜𝑟 )
𝑋𝑛𝑜𝑟𝑚,𝑖 = 𝐶𝑒𝑖𝑙𝑖𝑛𝑔 (6-20)
(𝑋𝑖 −𝑋𝑖𝐹𝑙𝑜𝑜𝑟 )

167
Texas Tech University, Hao Zhang, August 2021

𝐶𝑒𝑖𝑙𝑖𝑛𝑔
where 𝑋𝑖 is the 𝑖 th optimization variable; 𝑋𝑖 and 𝑋𝑖𝐹𝑙𝑜𝑜𝑟 are the upper and
lower limits for the 𝑖 th optimization variable; 𝑋𝑛𝑜𝑟𝑚,𝑖 is the 𝑖 th optimization variable
after normalization. After the optimization, the optimal values of optimization variables
are transformed into the values in the original search space.

Then based on the shale gas reservoir model and the uncertain NF pattern, the
hydraulically fractured well MOO is performed by the surrogate-assisted HMOPSO.
The NPVs from the production of stochastic models are calculated by the Eq.(5-2) in
Section 5.3.1, according to the economic parameter in Table 6.2, and the expected value
and standard deviation of these NPVs are regarded as the objective functions. The
particles’ population in the swarm is 20, and the entire step for optimization iteration is
set as 50. As a comparison, NSGA-II and MOPSO are also combined with the Kriging
surrogate model to construct the surrogate-assisted NSGA-II and surrogate-assisted
MOPSO, perform the optimization with the same reservoir models and parameters.
After the optimization, the PFs for different algorithms can be seen in Figure 6.16.

Figure 6.15 The optimization results of different surrogate-assisted algorithms

In Figure 6.16, the colored marks represent the non-dominated solutions on the
PF for different algorithms, and the gray circles represent the historical solutions during

168
Texas Tech University, Hao Zhang, August 2021

the optimization. As can be seen that after the optimization, the particles reach a PF that
satisfies both the increase of expected economical benefit and the reduction of
development risk. While comparing the PFs for these three optimization algorithms, we
can see that the PF obtained by HMOPSO is superior to the PF obtained by the other
two algorithms. The HMOPSO can push the particles toward a better front with a
smaller standard deviation and higher expectation of NPV, and also, the particles present
smooth distribution and reflect the best shape of the PF.
60

50
Iteration times

40

30

20

10

0
Train the Surrogate-assitied Total iteration
surrogate model optimization times
Operations

(a) The statistic of different iteration steps

(b) Variations of the accuracy of solutions on the PF


Figure 6.16 The analysis of the PF obtained by the surrogate-assisted HMOPSO

169
Texas Tech University, Hao Zhang, August 2021

The iteration steps for different operations of the surrogate assisted HMOPSO
are recorded and shown in Figure 6.16-a, it reveals that this algorithm takes 21 iteration
steps to train the surrogate model, which is then used to replace the time-consuming
numerical simulations for the rest 29 iterations. Since obtaining the solutions from the
surrogate model costs only a few tenths of a second, so the surrogate model helps to
save around 58% of the original optimization time. The non-dominated solutions on the
PF obtained by the Kriging model are compared with the solutions by real numerical
simulation and the accuracy is shown in Figure 6.16-b; as can be seen that the error for
all the solutions does not exceed 0.1%, which demonstrates that after training, the
surrogate model is accurate enough to replace the numerical simulation.

To better understand the difference between solutions on the PF, three


representative points are picked up from the PF, which stand for the maximal expected
values (point A), the minimal development risk (point B), and the medium point (point
C) (Figure 6.16). The values of the optimization variables of these three points are listed
in Table 6.3; additionally, the hydraulically fractured wells constructed according to
these three points and their corresponding production simulation results can be seen in
Figure 6.17 respectively. Comparing the results of these three points, we can see that
the well with a longer horizontal length and longer fracture half-length leads to better
shale gas production and a higher expectation of NPV with geological uncertainty, for
the HFs can cover a larger area and connect more NFs, yet it also needs higher economic
input and then cause larger development risk.

Table 6.3 The values of the optimization variables for different points

Optimization Variables Point A Point B Point C

The wellhead location (m, m) (523.63, 753.63) (771.62, 702.53) (648.35, 679.72)
The well length (m) 946.08 470.47 758.39
The fracture half-length (m) 127.16 39.87 80.91
The fracture spacing (m) 31.53 27.81 22.76

170
Texas Tech University, Hao Zhang, August 2021

(a) The hydraulic fractured well and the corresponding production result for point A

(b) The hydraulic fractured well and the corresponding production result for point B

(c) The hydraulic fractured well and the corresponding production result for point C
Figure 6.17 The hydraulically fractured wells and their corresponding reservoir pressure
map (MPa) after 20 years of production for different points

171
Texas Tech University, Hao Zhang, August 2021

(a) The statistics of the NPV distributions

(b) The box diagram of NPV distributions


Figure 6.18 The comparison of NPV distributions for different points

Furthermore, the NPV distributions for all the stochastic reservoir models under
these three solution points are compared in the histogram and box diagram of Figure

172
Texas Tech University, Hao Zhang, August 2021

6.18. As can be seen that, the NPVs of point A located towards the higher NPV region
with a mean value of 7.6×106 dollars, but they also present a more discrete distribution
comparing with the NPVs of the other two points, which means that if we design the
hydraulically fractured well according to the values of point A, we may obtain a higher
expectation of economic benefit yet have to undertake a higher risk of development. As
a comparison, the NPVs of point B distribute much closer and concentrate toward the
low-value region with a mean value of 0.9×106 dollars, which means if we design the
hydraulically fractured well according to point B, we don’t need to worry about the
development risk, yet we should also not expect a high economic benefit either.
Meanwhile, the NPVs of point C present a medium distribution trend and indicates a
medium economic benefit and development risk comparing with the other points. The
conditions of other points on the PF can also be derived from these three representative
points, the points closer to point A will bring higher expected economic benefits and
higher development risks, vice versa. The PF provides comprehensive trade-off
solutions for the engineers to design the hydraulically fractured well in unconventional
reservoirs considering the geological uncertainty.

173
Texas Tech University, Hao Zhang, August 2021

CHAPTER VII

CONCLUSION AND DISCUSSION


In this chapter, the key findings from the numerical model construction and
simulation results in this dissertation are summarized. Recommendations for future
work in this research area are also presented.

7.1 Summary

In view that the hydrocarbon resource produced from shale reservoirs is


occupying an ever-increasing percentage in today’s petroleum industry, and the
economic success of shale oil and gas development highly depends on the optimal
design of the hydraulically fractured well. This dissertation makes a comprehensive
numerical investigation on the hydraulically fractured well construction and production
optimization in the fractured shale reservoir considering the effect of the fracturing-
stimulated reservoir volume. The investigation is categorized into four chapters, and the
conclusions for each chapter are summarized and listed below.

7.1.1 Stimulated reservoir volume estimation and hydraulic fracturing


optimization

A simulation and optimization study for hydraulic fracturing well considering


the effect of SRV is proposed in this chapter. A DDM algorithm coupling the
mechanism of induced stress and formation pressure is used to estimate the SRV
stimulated by the hydraulic fractures, and the stress-dependent permeability (SDP)
within the SRV region is calculated according to the cubic law. Based on this SRV
estimation model, an automatic optimization process is established, in which a
Modified-PSO algorithm is proposed and employed to search for the optimal value of
NPV. Finally, the optimal hydraulic fracturing well design is achieved in a shale gas
reservoir. To sum up all the simulation and optimization results, the following
conclusions can be drawn from this chapter.

174
Texas Tech University, Hao Zhang, August 2021

 The SRV volume will be affected by hydraulic fracture half-length and fracture
spacing. As the hydraulic fracture half-length increases, the SRV volume will
increase, but the increasing extent or magnitude will decrease. As the hydraulic
fracture spacing increases, the SRV volume will first increase and then sharply
decrease, and finally reaches a constant. The best hydraulic fracture spacing for
SRV in this reservoir model is around 35ft.

 Comparing with the original PSO algorithm, the Modified-PSO algorithm shows a
better performance dealing with multi-extreme problems, it can maintain the
accuracy of the optimization result while reducing the calculation amount.

 The largest SRV volume doesn’t mean the best economic revenue. Comparing to
the SRV volume, the NPV is a better objective parameter for hydraulic fracturing
well optimization, which takes the investment cost for well drilling and SRV
stimulation into consideration, and avoids development risk.

 For a multiple hydraulic fracturing well design in the shale gas reservoir, the
alternating fracturing placement method presents better performance than the
aligning fracturing placement method.

7.1.2 Complex fracture network simulation and optimization

This chapter numerically investigates the complex fracture network (CFN)


propagation and production in the shale reservoir. First, an HF propagation model is
constructed by coupling the DDM and the fluid flow function; then combined with the
power-law probability NF pattern to construct a CFN propagation model. Secondly, the
CFN obtained is embedded into the EDFM to realize CFN production in the shale gas
reservoir. Furthermore, a CFN optimization considering both geological and economic
effects is performed, in which the CFN propagation and production model are combined
to achieve the optimal economic revenue under different fracturing strategies; an
advanced machine-learning algorithm M-NNA is proposed in this paper and adopted as
the optimization algorithm. From the results of the CFN simulation and optimization,
the following conclusions are summarized:

175
Texas Tech University, Hao Zhang, August 2021

 The NF distribution will affect the CFN propagation, a larger contact angle between
HF and NF helps to create more frac branches in CFN. Because of the induced
stress, the side of NF along with the HF propagation direction is more likely to be
opened, and frac branches located around the top of CFN are more likely to
continue propagating.

 The operational parameter, such as the hydraulic fracturing spacing, also affects the
CFN propagation. Too narrow spacing will bring a large stress shadow effect to
prevent the HF growth, while too large spacing will weaken the interaction of the
neighboring HFs.

 The CFN shows a significant impact on gas production in the fractured shale
reservoir, the area covered by the CFN obtained a perfect production, where the
shale gas nearly exhausted after 10 years of production; however, the effect of CFN
will get weaker as the distance away, the area 150ft away from the CFN can hardly
be affected.

 The machine learning algorithm processes a superiority over traditional algorithms


when dealing with complex optimization problems; the modification to the NNA
in this study reduces the risk for the solutions to trap in the local extreme and
increases the calculation accuracy.

 For a multi-well hydraulic fracturing, the alternating fracturing placement strategy


could make the HFs stimulate and connect more NFs to construct a better CFN
configuration, and to obtain higher economic revenue than the aligning fracturing
placement.

7.1.3 Unstructured quad-grid based embedded discrete fracture model

An unstructured quadrangular grid-based EDFM (UnQ-EDFM) is proposed in


this chapter. In this model, a frontal Delaunay quadrangular grid is adopted, which
overcomes the drawbacks of the traditional structured grid-based EDFM. Additionally,
an irregularly shaped naturally fractured shale reservoir model is constructed, based on
which the simulation of traditional EDFM and UnQ-EDFM are compared. Furthermore,

176
Texas Tech University, Hao Zhang, August 2021

a multi-stage fractured well is optimized in the above-mentioned shale reservoir model


and the reliability of UnQ-EDFM is testified based on thousands of times of simulation.
To sum up, the following conclusions are drawn from this study.

 The unstructured quadrangular grid presents better adaptivity than the structured
grid in the irregular shape reservoir, it perfectly constrains the irregular boundaries
and causes no deformation to the reservoir shape, regardless of the size of the grid.
Besides, when subdividing the same area with the same grid length, the grid number
for the unstructured quad grid is far less than that of the triangular grid.

 Compared to the structured grid-based EDFM, the UnQ-EDFM significantly


reduces the simulation error caused by reservoir deformation, especially with larger
grid size; additionally, it achieves the same simulation accuracy as the triangular
grid-based EDFM and costs only half of the calculation time.

 The UnQ-EDFM could achieve local grid refinement (LGR) flexibly anywhere in
the reservoir, without causing any additional gridding complexity; the LGR in the
UnQ-EDFM increases the simulation accuracy while not prolong the calculation
time.

 Testified by the fractured well optimization in a complex shale reservoir, the UnQ-
EDFM is proved to be more reliable and efficient than the structured grid-based
EDFM under large amounts of simulations.

7.1.4 Surrogate-assisted multi-objective hydraulic fracture optimization


considering uncertainty

In this chapter, a state-of-the-art surrogate-assisted multi-objective optimization


(MOO) framework is proposed for computationally expansive optimization problems
with geological uncertainty. Based on the EDFM, a shale gas reservoir model is
constructed, and a set of power-law distributed NF patterns are generated to represent
the geological uncertainty. Then, a hybrid multi-objective particle swarm optimization
algorithm (HMOPSO) is proposed, which couples the synthetic advantages of MOPSO
and NSGA-II; additionally, a Kriging surrogate model is integrated into the HMOPSO

177
Texas Tech University, Hao Zhang, August 2021

to construct a surrogated-assisted HMOPSO framework. Furthermore, a multi-objective


hydraulically fractured well optimization in shale reservoir under uncertainty is
performed by the surrogated-assisted HMOPSO, in which the expected value and
standard deviation of the NPV are regarded as the objective functions. Based on the
above simulation and optimization results, some essential conclusions are summarized
and listed as follows.

 Testified by the benchmark functions, the HMOPSO algorithm presents better


performance than the traditional NSGA-II and MOPSO in MOO, it can reach the
true PF with a faster speed than the NSGA-II and obtains a better PF shape than the
MOPSO; besides, the HMOPSO increases the utilization of particles and has higher
optimization accuracy in low particle populations.

 The Kriging model has pretty good matching quality with the real function; the
surrogate-assisted HMOPSO presents an excellent optimization performance in the
benchmark function, the Kriging surrogate model replaces the real function after
11 times of training and achieves the same accuracy.

 The surrogate-assisted HMOPSO is successfully applied in the hydraulically


fractured well optimization with geological uncertainty, the PF is obtained and
around 58% of the numerical simulation time is saved.

 The MOO of the hydraulically-fractured well provides a set of trade-off solutions


to the engineers to develop the shale gas reservoir under NF distribution uncertainty.
Higher economic recoveries are always accompanied by higher development risk,
and this research conducts a quantitative evaluation of the benefit and the risk.

7.2 Future works

There are many topics in which the work presented in the dissertation is far from
enough and could be extended. These topics are summarized as future work.

 In this dissertation, the investigations of hydraulic fracture propagation all focus on


the horizontal direction, and the models are all in two dimensions. However,

178
Texas Tech University, Hao Zhang, August 2021

hydraulic fracture vertical propagation is also of great practical significance,


especially for the multi-layered reservoir. For some situations, we don’t want
fractures to grow out of the target pay zone in the vertical direction, since it is not
only a waste of fracturing fluid but also causes a potential environmental hazard.
While in some other cases, the fracture is expected to break the barriers and
connected the thin layers to obtain the best recovery. So one of the future research
goals is the study of hydraulic fracture verticle propagation.

 This dissertation proposes robust hydraulic fractured well optimization considering


the reservoir geological uncertainty. However, with the fierce fluctuation in the oil
& gas prices, the economical uncertainty shows the same importance for the
hydraulic fractured well designing as the geological uncertainty, and should also be
taken into consideration. How to robustly design the hydraulic fractured well to
always maintaining the maximum economic revenue under the uncertain future oil
& gas prices is an essential topic. So the hydraulic fractured well optimization under
economic uncertainty should be investigated in the future.

 This dissertation numerically investigated the propagation and optimization of


complex fracture networks based on the geological condition and production of
shale gas reservoirs. Recently, geothermal production is becoming a hot topic,
which also highly depends on hydraulically fractured wells. The formation of
geothermal reservoirs also contains natural fractures that could be stimulated by
hydraulic fracturing to construct complex fracture networks, which will
significantly improve geothermal production. To construct and optimize the
hydraulically fractured wells considering the complex fracture network in the
geothermal reservoir has a huge application prospect and should be deeply studied
in the future.

179
Texas Tech University, Hao Zhang, August 2021

BIBLIOGRAPHY
AbouOmar, M. S., Zhang, H. J., & Su, Y. X. (2019). Fractional order fuzzy PID control
of automotive PEM fuel cell air feed system using neural network optimization
algorithm. Energies, 12(8), 1435. https://doi.org/10.3390/en12081435.

Abou-Sayed, A. S., Clifton, R. J., Dougherty, R. L., & Morales, R. H. (1984). Evaluation
of the influence of in-situ reservoir conditions on the geometry of hydraulic
fractures using a 3-d simulator: Part 2-case studies. In SPE Unconventional Gas
Recovery Symposium. Society of Petroleum Engineers.
https://doi.org/10.2118/10504-pa.

Adachi, J., Siebrits, E. M., Peirce, A., & Desroches, J. (2007). Computer simulation of
hydraulic fractures. International Journal of Rock Mechanics and Mining
Sciences, 44(5), 739-757. https://doi.org/10.1016/j.ijrmms.2006.11.006.

Alizadeh, R., Jia, L., Nellippallil, A. B., Wang, G., Hao, J., Allen, J. K., & Mistree, F.
(2019). Ensemble of surrogates and cross-validation for rapid and accurate
predictions using small data sets. Artificial Intelligence for Engineering Design,
Analysis and Manufacturing: AI EDAM, 33(4), 484-501.
https://doi.org/10.1017/s089006041900026x.

Alramahi, B., & Sundberg, M. I. (2012). Proppant embedment and conductivity of


hydraulic fractures in shales. In 46th US Rock Mechanics/Geomechanics
Symposium. American Rock Mechanics Association.

Arvind, H., Franz, D., & Martin, C. (2010). Volumetric Fracture Modeling Approach
(VFMA): incorporating microseismic data in the simulation of shale gas
reservoirs. In Proceedings of the SPE Annual Technical Conference and
Exhibition. https://doi.org/10.2118/134683-ms.

Bagheri, Mohammad, A. (2006). Modeling geomechanical effects on the flow


properties of fractured reservoirs. Doctoral Dissertation, University of Calgary.

Barenblatt, G. I., Zheltov, I. P., & Kochina, I. N. (1960). Basic concepts in the theory
of seepage of homogeneous liquids in fissured rocks [strata]. Journal of applied
mathematics and mechanics, 24(5), 1286-1303. https://doi.org/10.1016/0021-
8928(60)90107-6.

Barton, C. A., & Zoback, M. D. (2002). Discrimination of natural fractures from


drilling-induced wellbore failures in wellbore image data-implications for
reservoir permeability. SPE Reservoir Evaluation & Engineering, 5(03), 249-
254. https://doi.org/10.2118/78599-pa.

180
Texas Tech University, Hao Zhang, August 2021

Britt, L. K., Smith, M. B., Haddad, Z. A., Lawrence, J. P., Chipperfield, S. T., &
Hellman, T. J. (2006). Waterfracs: We do need proppant after all. In SPE annual
technical conference and exhibition. Society of Petroleum Engineers.
https://doi.org/10.2118/102227-MS.

Cao, P., Liu, J., & Leong, Y. K. (2016). A fully coupled multiscale shale deformation-
gas transport model for the evaluation of shale gas extraction. Fuel, 178, 103-
117. https://doi.org/10.1016/j.fuel.2016.03.055.

Carter, R. D. (1957). Appendix to Optimum fluid characteristics for fracture


extension. By GC Howard and CR Fast. Drill. And Prod. Prac, API, 267.

Chai, Z., Yan, B., Killough, J. E., & Wang, Y. (2016). Dynamic embedded discrete
fracture multi-continuum model for the simulation of fractured shale reservoirs.
In International Petroleum Technology Conference. International Petroleum
Technology Conference. https://doi.org/10.2523/IPTC-18887-MS.

Chang, Y., Petvipusit, K. R., & Devegowda, D. (2015). Multi-objective optimization


coupled with dimension-wise polynomial-based approach in smart well
placement under model uncertainty. In SPE Reservoir Simulation Symposium.
Society of Petroleum Engineers. Houston, Texas. SPE-173291-MS.
https://doi.org/10.2118/173291-MS.

Chen, Y., & Oliver, D. S. (2010). Ensemble-based closed-loop optimization applied to


Brugge field. SPE Reservoir Evaluation & Engineering, 13(01), 56-71..
https://doi.org/10.2118/118926-pa.

Cho, Y., Apaydin, O. G., & Ozkan, E. (2013). Pressure-dependent natural-fracture


permeability in shale and its effect on shale-gas well production. SPE Reservoir
Evaluation & Engineering, 16(02), 216-228. https://doi.org/10.2118/159801-PA.

Clifton, R. J., & Abou-Sayed, A. S. (1981). A variational approach to the prediction of


the three-dimensional geometry of hydraulic fractures. In SPE/DOE Low
Permeability Gas Reservoirs Symposium. Society of Petroleum Engineers.
https://doi.org/10.2118/9879-ms.

Coello, C. A. C., Pulido, G. T., & Lechuga, M. S. (2004). Handling multiple objectives
with particle swarm optimization. IEEE Transactions on evolutionary
computation, 8(3), 256-279. https://doi.org/10.1109/TEVC.2004.826067.

Cordero, J. A. R., Sanchez, E. C. M., & Roehl, D. (2019). Integrated discrete fracture
and dual porosity-Dual permeability models for fluid flow in deformable
fractured media. Journal of Petroleum Science and Engineering, 175, 644-653.
https://doi.org/10.1016/j.petrol.2018.12.053.

181
Texas Tech University, Hao Zhang, August 2021

Crouch, S. L. (1976). Solution of plane elasticity problems by the displacement


discontinuity method. I. Infinite body solution. International Journal for
Numerical Methods in Engineering, 10(2), 301-343.
https://doi.org/10.1002/nme.1620100206.

Cundall, P. A., & Strack, O. D. (1979). A discrete numerical model for granular
assemblies. geotechnique, 29(1), 47-65.
https://doi.org/10.1680/geot.1979.29.1.47.

Das, I., & Dennis, J. E. (1998). Normal-boundary intersection: A new method for
generating the Pareto surface in nonlinear multicriteria optimization
problems. SIAM journal on optimization, 8(3), 631-657.
https://doi.org/10.1137/s1052623496307510.

Davy, P. (1993). On the frequency‐length distribution of the San Andreas fault


system. Journal of Geophysical Research: Solid Earth, 98(B7), 12141-12151.
https://doi.org/10.1029/93jb00372.

De Pater, C. J., & Beugelsdijk, L. J. L. (2005). Experiments and numerical simulation


of hydraulic fracturing in naturally fractured rock. In Alaska Rocks 2005, The
40th US Symposium on Rock Mechanics (USRMS). American Rock Mechanics
Association.

Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. A. M. T. (2002). A fast and elitist
multiobjective genetic algorithm: NSGA-II. IEEE transactions on evolutionary
computation, 6(2), 182-197. https://doi.org/10.1109/4235.996017.

Durst, D. G., Harris, J. T., Contreras, J. D., & Watson, D. R. (2008). Improved Single-
Trip Multistage Completion Systems for Unconventional Gas Formations.
In SPE Tight Gas Completions Conference. Society of Petroleum Engineers.
https://doi.org/10.2118/115260-ms.

EIA. (2016a). Hydraulic fracturing accounts for about half of current U.S. crude oil
production. Energy Information Administration.
https://www.eia.gov/todayinenergy/detail.php?id=25372.

EIA. (2016b). Hydraulically fractured wells provide two-thirds of U.S. natural gas
production”. Energy Information Administration.
https://www.eia.gov/todayinenergy/detail.php?id=26112.

EIA. (2019). EIA adds new play production data to shale gas and tight oil reports.
Energy Information Administration.
https://www.eia.gov/todayinenergy/detail.php?id=38372.

182
Texas Tech University, Hao Zhang, August 2021

EIA. (2021). Annual energy outlook 2021. Energy Information Administration.


https://www.eia.gov/outlooks/aeo/.

Eshkalak, M. O., Al-shalabi, E. W., Sanaei, A., Aybar, U., & Sepehrnoori, K. (2014).
Enhanced gas recovery by CO2 sequestration versus re-fracturing treatment in
unconventional shale gas reservoirs. In Abu Dhabi International Petroleum
Exhibition and Conference. Society of Petroleum Engineers.
https://doi.org/10.2118/172083-ms.

Ewing, R. E. (Ed.). (1983). The mathematics of reservoir simulation. Society for


Industrial and Applied Mathematics. https://doi.org/10.1137/1.9781611971071.

Farah, N., Delorme, M., Ding, D. Y., Wu, Y. S., & Codreanu, D. B. (2019). Flow
modelling of unconventional shale reservoirs using a DFM-MINC proximity
function. Journal of Petroleum Science and Engineering, 173, 222-236.
https://doi.org/10.1016/j.petrol.2018.10.014.

Feiyu, W., Jing, G., Weiping, F., & Linyan, B. (2013). Evolution of overmature marine
shale porosity and implication to the free gas volume. Petroleum Exploration
and Development, 40(6), 819-824.https://doi.org/10.1016/s1876-
3804(13)60111-1.

Fisher, M. K., Heinze, J. R., Harris, C. D., Davidson, B. M., Wright, C. A., & Dunn, K.
P. (2004). Optimizing horizontal completion techniques in the Barnett shale
using microseismic fracture mapping. In SPE Annual Technical Conference and
Exhibition. Society of Petroleum Engineers. https://doi.org/10.2118/90051-MS.

Fisher, M. K., Wright, C. A., Davidson, B. M., Steinsberger, N. P., Buckler, W. S.,
Goodwin, A., & Fielder, E. O. (2005). Integrating fracture mapping technologies
to improve stimulations in the Barnett shale. SPE Production &
Facilities, 20(02), 85-93. https://doi.org/10.2118/77441-pa.

Fung, R. L., Vilayakumar, S., & Cormack, D. E. (1987). Calculation of vertical fracture
containment in layered formations. SPE formation evaluation, 2(04), 518-522.
https://doi.org/10.2118/14707-pa.

Gale, J. F., Reed, R. M., & Holder, J. (2007). Natural fractures in the Barnett Shale and
their importance for hydraulic fracture treatments. AAPG bulletin, 91(4), 603-
622. http://dx.doi.org/10.1306/11010606061.

Gale, J., & Holder, J. (2008). Natural fractures in shales and their importance for gas
production. In Tectonics studies group annual meeting (p. 71).
https://doi.org/10.1144/0071131.

183
Texas Tech University, Hao Zhang, August 2021

Garcia, X., Nagel, N., Zhang, F., Sanchez-Nagel, M., & Lee, B. (2013). Revisiting
vertical hydraulic fracture propagation through layered formations–A numerical
evaluation. In 47th US Rock Mechanics/Geomechanics Symposium. American
Rock Mechanics Association.

Geertsma, J., & De Klerk, F. (1969). A rapid method of predicting width and extent of
hydraulically induced fractures. Journal of petroleum technology, 21(12), 1-571.
https://doi.org/10.2118/2458-pa.

Gringarten, A. C., Ramey Jr, H. J., & Raghavan, R. (1974). Unsteady-state pressure
distributions created by a well with a single infinite-conductivity vertical
fracture. Society of Petroleum Engineers Journal, 14(04), 347-360.
https://doi.org/10.2118/4051-pa.

Gu, H., & Weng, X. (2010). Criterion for fractures crossing frictional interfaces at non-
orthogonal angles. In 44th US rock mechanics symposium and 5th US-Canada
rock mechanics symposium. American Rock Mechanics Association.

Guo, T., Zhang, S., Gao, J., Zhang, J., & Yu, H. (2013). Experimental study of fracture
permeability for stimulated reservoir volume (SRV) in shale
formation. Transport in porous media, 98(3), 525-542.
https://doi.org/10.1007/s11242-013-0157-7.

Heinemann, Z. E., Gerken, G., & von Hantelmann, G. (1983). Using local grid
refinement in a multiple-application reservoir simulator. In SPE Reservoir
Simulation Symposium. Society of Petroleum Engineers.
https://doi.org/10.2118/12255-ms.

Hinton, G. E., & Sejnowski, T. J. (Eds.). (1999). Unsupervised learning: foundations of


neural computation. MIT press.

Hossain, M. M., Rahman, M. K., & Rahman, S. S. (2002). A shear dilation stimulation
model for production enhancement from naturally fractured reservoirs. SPE
Journal, 7(02), 183-195. https://doi.org/10.2118/78355-PA.

Hoteit, H., & Firoozabadi, A. (2008). An efficient numerical model for incompressible
two-phase flow in fractured media. Advances in Water Resources, 31(6), 891-
905. https://doi.org/10.1016/j.advwatres.2008.02.004.

Hu, Y., Li, Z., Zhao, J., Tao, Z., & Gao, P. (2017). Prediction and analysis of the
stimulated reservoir volume for shale gas reservoirs based on rock failure
mechanism. Environmental Earth Sciences, 76(15), 1-13.
https://doi.org/10.1007/s12665-017-6830-3.

184
Texas Tech University, Hao Zhang, August 2021

Isebor, O. J., & Durlofsky, L. J. (2014). Biobjective optimization for general oil field
development. Journal of Petroleum Science and Engineering, 119, 123-138.
https://doi.org/10.1016/j.petrol.2014.04.021.

Jahandideh, A., & Jafarpour, B. (2016). Optimization of hydraulic fracturing design


under spatially variable shale fracability. Journal of Petroleum Science and
Engineering, 138, 174-188. https://doi.org/10.2118/169521-ms.

Jansen, J. D., Brouwer, R., & Douma, S. G. (2009). Closed loop reservoir management.
In SPE reservoir simulation symposium. OnePetro.
https://doi.org/10.2118/119098-ms.

Jiang, J., & Younis, R. M. (2015). Numerical study of complex fracture geometries for
unconventional gas reservoirs using a discrete fracture-matrix model. Journal of
Natural Gas Science and Engineering, 26, 1174-1186.
https://doi.org/10.1016/j.jngse.2015.08.013.

Jo, H., & Hurt, R. (2013). Testing and review of various displacement discontinuity
elements for LEFM crack problems. In ISRM International Conference for
Effective and Sustainable Hydraulic Fracturing. International Society for Rock
Mechanics and Rock Engineering. https://doi.org/10.5772/56445.

Kassis, S., & Sondergeld, C. H. (2010). Fracture permeability of gas shale: Effect of
roughness, fracture offset, proppant, and effective stress. In International oil and
gas conference and exhibition in China. Society of Petroleum Engineers.
https://doi.org/10.2118/131376-MS.

Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. In Proceedings of


ICNN'95-international conference on neural networks, 4, 1942-1948. IEEE.
https://doi.org/10.1109/icnn.1995.488968.

Keshavarzi, R., & Mohammadi, S. (2012). A new approach for numerical modeling of
hydraulic fracture propagation in naturally fractured reservoirs. In SPE/EAGE
European Unconventional Resources Conference & Exhibition-From Potential
to Production, cp-285-00039. European Association of Geoscientists &
Engineers. https://doi.org/10.2118/152509-ms.

Khristianovic, S. A., & Zheltov, Y. P. (1955). Formation of vertical fractures by means


of highly viscous liquid. In World Petroleum Congress Proceedings, 579-586.

King, G. E. (2010). Thirty years of gas shale fracturing: What have we learned?. In SPE
annual technical conference and exhibition. Society of Petroleum Engineers.
https://doi.org/10.2118/133456-ms.

185
Texas Tech University, Hao Zhang, August 2021

LaBrecque, D., Brigham, R., Denison, J., Murdoch, L., Slack, W., Liu, Q. H., ... &
Ahmadian, M. (2016). Remote imaging of proppants in hydraulic fracture
networks using electromagnetic methods: Results of small-scale field
experiments. In SPE Hydraulic Fracturing Technology Conference. OnePetro.
https://doi.org/10.2118/179170-MS.

Lecampion, B. (2009). An extended finite element method for hydraulic fracture


problems. Communications in Numerical Methods in Engineering, 25(2), 121-
133. https://doi.org/10.1002/cnm.1111.

Li, L., & Lee, S. H. (2008). Efficient field-scale simulation of black oil in a naturally
fractured reservoir through discrete fracture networks and homogenized
media. SPE Reservoir Evaluation & Engineering, 11(04), 750-758.
https://doi.org/10.2118/103901-PA.

Li, S., & Zhang, D. (2018). A fully coupled model for hydraulic-fracture growth during
multiwell-fracturing treatments: enhancing fracture complexity. SPE Production
& Operations, 33(02), 235-250. https://doi.org/10.2118/182674-PA.

Li, W., Dong, Z., & Lei, G. (2017). Integrating embedded discrete fracture and dual-
porosity, dual-permeability methods to simulate fluid flow in shale oil
reservoirs. Energies, 10(10), 1471. https://doi.org/10.3390/en10101471.

Liang, B., Jiang, H., Li, J., & Gong, C. (2016). A systematic study of fracture parameters
effect on fracture network permeability based on discrete-fracture model
employing Finite Element Analyses. Journal of Natural Gas Science and
Engineering, 28, 711-722. https://doi.org/10.1016/j.jngse.2015.12.011.

Lie, K. A. (2019). An introduction to reservoir simulation using MATLAB/GNU


Octave: User guide for the MATLAB Reservoir Simulation Toolbox (MRST).
Cambridge University Press. https://doi.org/10.1017/9781108591416.

Lin, R., Ren, L., Zhao, J., Wu, L., & Li, Y. (2017). Cluster spacing optimization of
multi-stage fracturing in horizontal shale gas wells based on stimulated reservoir
volume evaluation. Arabian Journal of Geosciences, 10(2), 38.
https://doi.org/10.1007/s12517-016-2823-x.

Liu, B., Grout, V., & Nikolaeva, A. (2017). Efficient global optimization of actuator
based on a surrogate model assisted hybrid algorithm. IEEE Transactions on
Industrial Electronics, 65(7), 5712-5721.
https://doi.org/10.1109/TIE.2017.2782203.

Liu, X., & Reynolds, A. C. (2016). Gradient-based multiobjective optimization for


maximizing expectation and minimizing uncertainty or risk with application to

186
Texas Tech University, Hao Zhang, August 2021

optimal well-control problem with only bound constraints. SPE Journal, 21(05),
1-813. https://doi.org/10.2118/173216-PA.

Liu, Z. (2020). Robust Life-Cycle Production Optimization with State Constraints.


Doctoral dissertation, The University of Tulsa.

Liu, Z., & Forouzanfar, F. (2018). Ensemble clustering for efficient robust optimization
of naturally fractured reservoirs. Computational Geosciences, 22(1), 283-296.
https://doi.org/10.1007/s10596-017-9689-1.

Liu, Z., & Reynolds, A. C. (2019). History matching an unconventional reservoir with
a complex fracture network. In SPE Reservoir Simulation Conference. Society
of Petroleum Engineers. https://doi.org/10.2118/193921-MS.

Lopez, B., Piedrahita, J., & Aguilera, R. (2015). Estimates of stress dependent properties
in tight reservoirs: Their use with drill cuttings data. In SPE Latin American and
Caribbean Petroleum Engineering Conference. Society of Petroleum Engineers.
https://doi.org/10.2118/177189-MS.

Mahesh, K., Nallagownden, P., & Elamvazuthi, I. (2016). Advanced Pareto front non-
dominated sorting multi-objective particle swarm optimization for optimal
placement and sizing of distributed generation. Energies, 9(12), 982.
https://doi.org/10.3390/en9120982.

Maxwell, S. C., Urbancic, T. I., Steinsberger, N., & Zinno, R. (2002). Microseismic
imaging of hydraulic fracture complexity in the Barnett shale. In SPE annual
technical conference and exhibition. Society of Petroleum Engineers.
https://doi.org/10.2118/77440-MS.

Mayerhofer, M. J., Lolon, E. P., Warpinski, N. R., Cipolla, C. L., Walser, D., &
Rightmire, C. M. (2010). What is stimulated reservoir volume?. SPE Production
& Operations, 25(01), 89-98. https://doi.org/10.2118/119890-PA.

Mayerhofer, M. J., Lolon, E. P., Youngblood, J. E., & Heinze, J. R. (2006). Integration
of microseismic-fracture-mapping results with numerical fracture network
production modeling in the Barnett Shale. In SPE annual technical conference
and exhibition. Society of Petroleum Engineers. https://doi.org/10.2118/102103-
ms.

McKay, M. D., Beckman, R. J., & Conover, W. J. (2000). A comparison of three


methods for selecting values of input variables in the analysis of output from a
computer code. Technometrics, 42(1), 55-61.
https://doi.org/10.1080/00401706.2000.10485979.

187
Texas Tech University, Hao Zhang, August 2021

Moës, N., Dolbow, J., & Belytschko, T. (1999). A finite element method for crack
growth without remeshing. International journal for numerical methods in
engineering, 46(1), 131-150. https://doi.org/10.1002/(sici)1097-
0207(19990910)46:1<131::aid-nme726>3.0.co;2-j.

Moinfar, A., Sepehrnoori, K., Johns, R. T., & Varavei, A. (2013). Coupled
geomechanics and flow simulation for an embedded discrete fracture model.
In SPE Reservoir Simulation Symposium. Society of Petroleum Engineers.
http://dx.doi.org/10.2118/163666-MS.

Mutalik, P. N., & Gibson, R. W. (2008). Case history of sequential and simultaneous
fracturing of the Barnett shale in Parker County. In SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers.
https://doi.org/10.2118/116124-ms.

Nassir, M., Settari, A., & Wan, R. G. (2014). Prediction of stimulated reservoir volume
and optimization of fracturing in tight gas and shale with a fully elasto-plastic
coupled geomechanical model. SPE Journal, 19(05), 771-785.
https://doi.org/10.2118/163814-PA.

Nordgren, R. P. (1972). Propagation of a vertical hydraulic fracture. Society of


Petroleum Engineers Journal, 12(04), 306-314. https://doi.org/10.2118/3009-
pa.

Olsen, S. R., & Watanabe, F. S. (1957). A method to determine a phosphorus adsorption


maximum of soils as measured by the Langmuir isotherm. Soil Science Society
of America Journal, 21(2), 144-149.
https://doi.org/10.2136/sssaj1957.03615995002100020004x.

Olson, J. E. (2004). Predicting fracture swarms-the influence of subcritical crack growth


and the crack-tip process zone on joint spacing in rock. Geological society,
London, Special Publications. 231 (1), 73-88.
https://doi.org/10.1144/gsl.sp.2004.231.01.05.

Olson, J. E. (2007). Fracture aperture, length and pattern geometry development under
biaxial loading: a numerical study with applications to natural, cross-jointed
systems. Geological Society, London, Special Publications. 289 (1), 123-142.
https://doi.org/10.1144/sp289.8.

Ong, Y. S., Nair, P. B., & Keane, A. J. (2003). Evolutionary optimization of


computationally expensive problems via surrogate modeling. AIAA
journal, 41(4), 687-696. https://arc.aiaa.org/doi/abs/10.2514/2.1999.

188
Texas Tech University, Hao Zhang, August 2021

Ozkan, E., Raghavan, R. S., & Apaydin, O. G. (2010). Modeling of fluid transfer from
shale matrix to fracture network. In SPE Annual Technical Conference and
Exhibition. Society of Petroleum Engineers. https://doi.org/10.2118/134830-MS.

Pan, L., He, C., Tian, Y., Wang, H., Zhang, X., & Jin, Y. (2018). A classification-based
surrogate-assisted evolutionary algorithm for expensive many-objective
optimization. IEEE Transactions on Evolutionary Computation, 23(1), 74-88.
https://doi.org/10.1109/tevc.2018.2802784.

Perkins, T. K., & Kern, L. R. (1961). Widths of hydraulic fractures. Journal of


Petroleum Technology, 13(09), 937-949. https://doi.org/10.2118/89-pa.

Pruess, K. (1983). GMINC: A mesh generator for flow simulations in fractured


reservoirs (No. LBL-15227). Lawrence Berkeley Lab., CA (USA).
https://doi.org/10.2172/6065621.

Rafiee, M., Soliman, M. Y., & Pirayesh, E. (2012). Hydraulic fracturing design and
optimization: a modification to zipper frac. In SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers.
https://doi.org/10.2118/159786-MS.

Rammay, M. H., & Awotunde, A. A. (2016). Stochastic optimization of hydraulic


fracture and horizontal well parameters in shale gas reservoirs. Journal of
Natural Gas Science and Engineering, 36, 71-78.
https://doi.org/10.1016/j.jngse.2016.10.002.

Rasmussen, C. E., & Williams, C. K. (2006). Gaussian Process for Machine Learning.
Adaptive Computation and Machine Learning.
https://doi:10.7551/mitpress/3206.001.0001.

Rebay, S. (1993). Efficient unstructured mesh generation by means of Delaunay


triangulation and Bowyer-Watson algorithm. Journal of computational
physics, 106(1), 125-138. https://doi.org/10.1006/jcph.1993.1097.

Remacle, J. F., Henrotte, F., Carrier‐Baudouin, T., Béchet, E., Marchandise, E.,
Geuzaine, C., & Mouton, T. (2013). A frontal Delaunay quad mesh generator
using the L∞ norm. International Journal for Numerical Methods in
Engineering, 94(5), 494-512. https://doi.org/10.1002/nme.4458.

Ren, L., Lin, R., Zhao, J. Z., Yang, K. W., Hu, Y. Q., & Wang, X. J. (2015).
Simultaneous hydraulic fracturing of ultra-low permeability sandstone
reservoirs in China: Mechanism and its field test. Journal of Central South
University, 22(4), 1427-1436. https://doi.org/10.1007/s11771-015-2660-1.

189
Texas Tech University, Hao Zhang, August 2021

Ren, L., Lin, R., Zhao, J., Rasouli, V., Zhao, J., & Yang, H. (2018). Stimulated reservoir
volume estimation for shale gas fracturing: Mechanism and modeling
approach. Journal of Petroleum Science and Engineering, 166, 290-304.
https://doi.org/10.1016/j.petrol.2018.03.041.

Ren, L., Su, Y., Zhan, S., Hao, Y., Meng, F., & Sheng, G. (2016). Modeling and
simulation of complex fracture network propagation with SRV fracturing in
unconventional shale reservoirs. Journal of Natural Gas Science and
Engineering, 28, 132-141. https://doi.org/10.1016/j.jngse.2015.11.042.

Ren, Q., Dong, Y., & Yu, T. (2009). Numerical modeling of concrete hydraulic
fracturing with extended finite element method. Science in China Series E:
Technological Sciences, 52(3), 559-565. https://doi.org/10.1007/s11431-009-
0058-8.

Renshaw, C. E., & Pollard, D. D. (1995). An experimentally verified criterion for


propagation across unbounded frictional interfaces in brittle, linear elastic
materials. In International journal of rock mechanics and mining sciences &
geomechanics abstracts, 32(3), 237-249. Pergamon.
https://doi.org/10.1016/0148-9062(94)00037-4.

Ross, S. A. (1995). Uses, abuses, and alternatives to the net-present-value


rule. Financial management, 24(3), 96-102. https://doi.org/10.2307/3665561.

Roussel, N. P., & Sharma, M. M. (2011). Optimizing fracture spacing and sequencing
in horizontal-well fracturing. SPE Production & Operations, 26(02), 173-184.
https://doi.org/10.2118/127986-ms.

Rubin, B. (2010). Accurate simulation of non Darcy flow in stimulated fractured shale
reservoirs. In SPE Western regional meeting. Society of Petroleum Engineers.
https://doi.org/10.2118/132093-ms.

Sadollah, A., Sayyaadi, H., & Yadav, A. (2018). A dynamic metaheuristic optimization
model inspired by biological nervous systems: Neural network
algorithm. Applied Soft Computing, 71, 747-782.
https://doi.org/10.1016/j.asoc.2018.07.039.

Schlichting H. (1968). Boundary-layer theory, 6th Edn. McGraw-Hill. New York.

Schweitzer, R., & Bilgesu, H. I. (2009). The role of economics on well and fracture
design completions of Marcellus Shale wells. In SPE Eastern Regional Meeting.
Society of Petroleum Engineers. https://doi.org/10.2118/125975-MS.

190
Texas Tech University, Hao Zhang, August 2021

Segall, P., & Pollard, D. D. (1983). Joint formation in granitic rock of the Sierra
Nevada. Geological Society of America Bulletin, 94(5), 563-575.
http://dx.doi.org/10.1130/00167606.

Sen, V., Min, K. S., Ji, L., & Sullivan, R. (2018). Completions and well spacing
optimization by dynamic SRV modeling for multi-stage hydraulic fracturing.
In SPE Annual Technical Conference and Exhibition. Society of Petroleum
Engineers. https://doi.org/10.2118/191571-MS.

Settari, A., & Cleary, M. P. (1984). Three-dimensional simulation of hydraulic


fracturing. Journal of petroleum technology, 36(07), 1-177.
https://doi.org/10.2118/10504-pa.

Shakiba, M., de Araujo Cavalcante Filho, J. S., & Sepehrnoori, K. (2018). Using
embedded discrete fracture model (EDFM) in numerical simulation of complex
hydraulic fracture networks calibrated by microseismic monitoring data. Journal
of Natural Gas Science and Engineering, 55, 495-507.
https://doi.org/10.2118/175142-ms.

Sheng, G., Su, Y., & Wang, W. (2019). A new fractal approach for describing induced-
fracture porosity/permeability/compressibility in stimulated unconventional
reservoirs. Journal of Petroleum Science and Engineering, 179, 855-866.
https://doi.org/10.1016/j.petrol.2019.04.104.

Shiqian, X., Yuyao, L., Yu, Z., Sen, W., & Qihong, F. (2020). A history matching
framework to characterize fracture network and reservoir properties in tight
oil. Journal of Energy Resources Technology, 142(4).
https://doi.org/10.1115/1.4044767.

Siddhamshetty, P., Yang, S., & Kwon, J. S. I. (2018). Modeling of hydraulic fracturing
and designing of online pumping schedules to achieve uniform proppant
concentration in conventional oil reservoirs. Computers & Chemical
Engineering, 114, 306-317.
https://doi.org/10.1016/j.compchemeng.2017.10.032.

Simonson, E. R., Abou-Sayed, A. S., & Clifton, R. J. (1978). Containment of massive


hydraulic fractures. Society of Petroleum Engineers Journal, 18(01), 27-32.
https://doi.org/10.2118/6089-pa.

Siraj, M. M., Van den Hof, P. M., & Jansen, J. D. (2015). Model and economic
uncertainties in balancing short-term and long-term objectives in water-flooding
optimization. In SPE Reservoir Simulation Symposium. Society of Petroleum
Engineers. https://doi.org/10.2118/173285-MS.

191
Texas Tech University, Hao Zhang, August 2021

Su, Y., Ren, L., Meng, F., Xu, C., & Wang, W. (2015). Theoretical analysis of the
mechanism of fracture network propagation with stimulated reservoir volume
(SRV) fracturing in tight oil reservoirs. PloS one, 10(5), e0125319.
https://doi.org/10.1371/journal.pone.0125319.

Sun, C., Jin, Y., Cheng, R., Ding, J., & Zeng, J. (2017). Surrogate-assisted cooperative
swarm optimization of high-dimensional expensive problems. IEEE
Transactions on Evolutionary Computation, 21(4), 644-660.
https://doi.org/10.1109/tevc.2017.2675628.

Tan, P., Jin, Y., Han, K., Hou, B., Chen, M., Guo, X., & Gao, J. (2017). Analysis of
hydraulic fracture initiation and vertical propagation behavior in laminated shale
formation. Fuel, 206, 482-493. https://doi.org/10.1016/j.fuel.2017.05.033.

Tao, Q., Ehlig-Economides, C. A., & Ghassemi, A. (2009). Investigation of stress-


dependent permeability in naturally fractured reservoirs using a fully coupled
poroelastic displacement discontinuity model. In SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers.
https://doi.org/10.2118/124745-MS.

Tian, Y., Cheng, R., Zhang, X., Su, Y., & Jin, Y. (2018). A strengthened dominance
relation considering convergence and diversity for evolutionary many-objective
optimization. IEEE Transactions on Evolutionary Computation, 23(2), 331-345.
https://doi.org/10.1109/TEVC.2018.2866854.

Tripoppoom, S., Ma, X., Yong, R., Wu, J., Yu, W., Sepehrnoori, K., ... & Li, N. (2020).
Assisted history matching in shale gas well using multiple-proxy-based Markov
chain Monte Carlo algorithm: The comparison of K-nearest neighbors and
neural networks as proxy model. Fuel, 262, 116563.
https://doi.org/10.1016/j.fuel.2019.116563.

Veatch Jr, R. W., & Moschovidis, Z. A. (1986). An overview of recent advances in


hydraulic fracturing technology. In International meeting on petroleum
engineering. Society of Petroleum Engineers. https://doi.org/10.2118/14085-ms.

Wang, B. (2020). MRST-Shale: An Open-Source Framework for Generic Numerical


Modeling of Unconventional Shale and Tight Gas Reservoirs. Preprints.
https://doi.org/10.20944/preprints202001.0080.v1.

Warpinski, N. R. (1991). Hydraulic fracturing in tight, fissured media. Journal of


Petroleum Technology, 43(02), 146-209. https://doi.org/10.2118/20154-pa.

Warpinski, N. R., & Teufel, L. W. (1987). Influence of geologic discontinuities on


hydraulic fracture propagation. Journal of Petroleum Technology, 39(02), 209-
220. https://doi.org/10.2118/13224-pa.

192
Texas Tech University, Hao Zhang, August 2021

Warren, J. E., & Root, P. J. (1963). The behavior of naturally fractured


reservoirs. Society of Petroleum Engineers Journal, 3(03), 245-255.
https://doi.org/10.2118/426-pa.

Waters, G. A., Dean, B. K., Downie, R. C., Kerrihard, K. J., Austbo, L., & McPherson,
B. (2009). Simultaneous hydraulic fracturing of adjacent horizontal wells in the
Woodford Shale. In SPE hydraulic fracturing technology conference. Society of
Petroleum Engineers. https://doi.org/10.2118/119635-ms.

Watson, D. F. (1981). Computing the n-dimensional Delaunay tessellation with


application to Voronoi polytopes. The computer journal, 24(2), 167-172.
https://doi.org/10.1093/comjnl/24.2.167.

Wiley, C., Barree, B., Eberhard, M., & Lantz, T. (2004). Improved Horizontal Well
Stimulations in the Bakken Formation, Williston Basin, Montana. In SPE
Annual Technical Conference and Exhibition. Society of Petroleum Engineers.
https://doi.org/10.2118/90697-ms.

Wu, K. (2014). Numerical modeling of complex hydraulic fracture development in


unconventional reservoirs. Doctoral Dissertation, University of Texas at Austin.

Wu, K., & Olson, J. E. (2016). Numerical investigation of complex hydraulic-fracture


development in naturally fractured reservoirs. SPE production &
operations, 31(04), 300-309. https://doi.org/10.2118/173326-MS.

Wu, K., Yu, W., & Miao, J. (2019a). Integrating complex fracture modeling and EDFM
to optimize well spacing in shale oil reservoirs. In 53rd US Rock
Mechanics/Geomechanics Symposium. American Rock Mechanics Association.

Wu, R., Kresse, O., Weng, X., Cohen, C. E., & Gu, H. (2012). Modeling of interaction
of hydraulic fractures in complex fracture networks. In SPE Hydraulic
Fracturing Technology Conference. Society of Petroleum Engineers.
https://doi.org/10.2118/152052-MS.

Wu, W., Zhou, J., Kakkar, P., Russell, R., & Sharma, M. M. (2019b). An experimental
study on conductivity of unpropped fractures in preserved shales. SPE
Production & Operations, 34(02), 280-296. https://doi.org/10.2118/184858-pa.

Wu, Y. S. and Pruess, K. (1988). A multiple-porosity method for simulation of naturally


fractured petroleum reservoirs. SPE Reservoir Engineering, 3 (01), 327-336.
https://doi.org/10.2118/15129-PA.

Wu, Y. S., Li, J., Ding, D. Y., Wang, C., & Di, Y. (2014). A generalized framework
model for the simulation of gas production in unconventional gas reservoirs. Spe
Journal, 19(05), 845-857. https://doi.org/10.2118/163609-pa.

193
Texas Tech University, Hao Zhang, August 2021

Xu, F., Yu, W., Li, X., Miao, J., Zhao, G., Sepehrnoori, K., ... & Wen, G. (2018). A fast
EDFM method for production simulation of complex fractures in naturally
fractured reservoirs. In SPE/AAPG Eastern Regional Meeting. Society of
Petroleum Engineers. https://doi.org/10.2118/191800-18ERM-MS.

Xu, Y. (2015). Implementation and application of the embedded discrete fracture model
(EDFM) for reservoir simulation in fractured reservoirs. Doctoral dissertation,
University of Texas at Austin.

Yan, X., Huang, Z., Yao, J., Li, Y., & Fan, D. (2016). An efficient embedded discrete
fracture model based on mimetic finite difference method. Journal of Petroleum
Science and Engineering, 145, 11-21.
https://doi.org/10.1016/j.petrol.2016.03.013.

Yang, C., Card, C., Nghiem, L. X., & Fedutenko, E. (2011). Robust optimization of
SAGD operations under geological uncertainties. In SPE reservoir simulation
symposium. Society of Petroleum Engineers. https://doi.org/10.2118/141676-
MS.

Yu, W., & Sepehrnoori, K. (2013). Optimization of multiple hydraulically fractured


horizontal wells in unconventional gas reservoirs. Journal of Petroleum
Engineering, 2013. https://doi.org/10.1155/2013/151898.

Yu, W., Wu, K., Liu, M., Sepehrnoori, K., & Miao, J. (2018). Production forecasting
for shale gas reservoirs with nanopores and complex fracture geometries using
an innovative non-intrusive EDFM method. In SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers.
https://doi.org/10.2118/191666-MS.

Zanganeh, B., Ahmadi, M., Hanks, C., & Awoleke, O. (2015). The role of hydraulic
fracture geometry and conductivity profile, unpropped zone conductivity and
fracturing fluid flowback on production performance of shale oil wells. Journal
of Unconventional Oil and Gas Resources, 9, 103-113.
https://doi.org/10.1016/j.juogr.2014.11.006.

Zapotecas Martí nez, S., & Coello Coello, C. A. (2011). A multi-objective particle
swarm optimizer based on decomposition. In Proceedings of the 13th annual
conference on Genetic and evolutionary computation, 69-76.
https://doi.org/10.1145/2001576.2001587.

Zeng, Q. D., Yao, J., & Shao, J. (2019). Study of hydraulic fracturing in an anisotropic
poroelastic medium via a hybrid EDFM-XFEM approach. Computers and
Geotechnics, 105, 51-68. https://doi.org/10.1016/j.compgeo.2018.09.010.

194
Texas Tech University, Hao Zhang, August 2021

Zeng, Q., & Yao, J. (2016). Numerical simulation of fracture network generation in
naturally fractured reservoirs. Journal of Natural Gas Science and
Engineering, 30, 430-443. https://doi.org/10.1016/j.jngse.2016.02.047.

Zhang, F., Zhu, H., Zhou, H., Guo, J., & Huang, B. (2017). Discrete-element-
method/computational-fluid-dynamics coupling simulation of proppant
embedment and fracture conductivity after hydraulic fracturing. SPE
Journal, 22(02), 632-644. https://doi.org/10.2118/185172-pa.

Zhang, H., & Sheng, J. J. (2021). An efficient embedded discrete fracture model based
on the unstructured quadrangular grid. Journal of Natural Gas Science and
Engineering, 85, 103710. https://doi.org/10.1016/j.jngse.2020.103710.

Zhang, H., Zhang, K., Zhang, L., Sheng, J., Yao, J., Wang, J., & Yang, Y. (2019).
Construction and optimization of adaptive well pattern based on reservoir
anisotropy and uncertainty. Journal of Petroleum Science and Engineering, 181,
106252. https://doi.org/10.1016/j.petrol.2019.106252.

Zhang, X., Zheng, X., Cheng, R., Qiu, J., & Jin, Y. (2018). A competitive mechanism
based multi-objective particle swarm optimizer with fast convergence.
Information Sciences, 427, 63-76. https://doi.org/10.1016/j.ins.2017.10.037.

Zhang, Z., Li, X., Yuan, W., He, J., Li, G., & Wu, Y. (2015). Numerical analysis on the
optimization of hydraulic fracture networks. Energies, 8(10), 12061-12079.
https://doi.org/10.3390/en81012061.

Zhao, M., Zhang, K., Chen, G., Zhao, X., Yao, C., Sun, H., ... & Yao, J. (2020a). A
surrogate-assisted multi-objective evolutionary algorithm with dimension-
reduction for production optimization. Journal of Petroleum Science and
Engineering, 192, 107192. https://doi.org/10.1016/j.petrol.2020.107192.

Zhao, M., Zhang, K., Chen, G., Zhao, X., Yao, J., Yao, C., ... & Yang, Y. (2020b). A
Classification-Based Surrogate-Assisted Multiobjective Evolutionary
Algorithm for Production Optimization under Geological Uncertainty. SPE
Journal. https://doi.org/10.2118/201229-pa.

Zhu, S. Y., Du, Z. M., Li, C. L., Salmachi, A., Peng, X. L., Wang, C. W., ... & Deng, P.
(2018). A semi-analytical model for pressure-dependent permeability of tight
sandstone reservoirs. Transport in Porous Media, 122(2), 235-252.
https://doi.org/10.1007/s11242-018-1001-x.

Zitzler, E., Deb, K., & Thiele, L. (2000). Comparison of multiobjective evolutionary
algorithms: Empirical results. Evolutionary computation. 8 (2), 173-195.
https://doi.org/10.1162/106365600568202.

195

You might also like