Professional Documents
Culture Documents
Liu Thesis 2020
Liu Thesis 2020
Liu Thesis 2020
by
Chuxi Liu
2020
The Thesis Committee for Chuxi Liu
Certifies that this is the approved version of the following Thesis:
APPROVED BY
SUPERVISING COMMITTEE:
Wei Yu
Automatic History Matching with Data Integration for Unconventional
Reservoirs
by
Chuxi Liu
Thesis
Presented to the Faculty of the Graduate School of
in Partial Fulfillment
of the Requirements
To my beloved parents, Ziyun Liu and Aimin Ma for their unconditional love and
encouragement.
Acknowledgements
Sepehrnoori, for his continuous supports and guidance for the entire period of my
master’s degree. His resourceful feedbacks and comments on the weekly meeting are
extremely helpful for me to envision the industry insights and also to build a more
rigorous thought process. It is my greatest pleasure and honor to work under this research
group, which is the leading research group in unconventional resource studies in the
Secondly, I am very thankful for Dr. Wei Yu. His professional expertise,
both in perspective of academia and industry. I am grateful for his utmost patience, trusts,
grateful for his review and writing advices given towards the completion of this thesis.
on automatic history matching algorithm. I also appreciate the help and supports from our
research group students for skill sharing, especially Esmail Eltathan for his tutorial on the
fracture modelling and reservoir modelling and his knowledge regarding the numerical
optimization.
Finally, I want to acknowledge the great assistances from my UT PGE friends for
their care and helps, so that I can experience a wonderful life in Austin. In addition, I
want to appreciate the emotional and financial supports from my parents. Without them,
v
Abstract
such as history matching can minimize the global error and find an optimal reservoir
model that can approximate the fracture geometry and petrophysical parameters in the
subsurface. For unconventional reservoirs, the idea behind the automatic history
matching is well developed and the workflow is also applied to statistically generate an
more uncertainties regarding fracture and reservoir properties could be further reduced by
using available information. Therefore, the objective of this study is to minimize
We utilized the developed automatic history matching (AHM) code and modified
the proxy engine, by substituting the neural network (NN) model with XGBoost
(XGBOOST) model. The XGBOOST is found to perform more efficiently and accurately
than NN, when the size of the available dataset for training is small. Furthermore, the
vi
AHM workflow is capable of modelling non-uniform half-length of hydraulic fractures in
the corner point gridding system and complex, realistic natural fracture distributions
using the fractal theory. Both of these functionalities partially fulfill some degrees of
software and naturally occurring patterns often found at cores. We applied this innovative
approach to actual shale gas and shale oil wells. We then found that by coupling
additional data into the AHM process, the fracture geometries and petrophysical
properties can be more accurately depicted. The obtained results are also highly
assimilating with the field experience from the engineers. In addition, by studying natural
fractures in the model, we found out that the connectivity between natural fractures and
potential.
This study is beneficial because more reliable and robust results based on
reservoir and fractures, can provide invaluable guidance towards well spacing planning,
vii
Table of Contents
List of Tables ..................................................................................................................... xi
1.1 Motivation.............................................................................................................1
4.8 Conclusions.........................................................................................................86
ix
5.2 Reservoir Model .................................................................................................87
5.8 Conclusions.......................................................................................................128
Glossary ...........................................................................................................................133
Acronyms ................................................................................................................133
Nomenclature ..........................................................................................................134
References ........................................................................................................................137
x
List of Tables
Table 3.1: Summary of significant parameters that is supported by AHM workflow. ......17
study. .............................................................................................................35
Table 4.2: Summary of basic reservoir, fracture and gas desorption parameters used
Table 4.3: Summary of uncertain parameters, their ranges, fixed parameters, and
Table 5.1: Summary of basic reservoir, fracture and gas desorption parameters used
Table 5.2: Summary of uncertain parameters, their ranges, fixed parameters, and
xi
List of Figures
Figure 2.1: Illustration of the concept for Embedded Discrete Fracture Model: (a)
Figure 3.2: Graphical illustration of the permeability decay and fracture closure
Figure 3.3: Schematic illustration of the Latin Hypercube sampling method. (Shams
2016) .............................................................................................................19
Figure 3.4: Summary of the workflow for inputting the hydraulic fractures in
Figure 3.5: Graphical illustration of the variables and definitions used for inputting
Figure 3.6: Graphical illustration of the fractal theory and the effects of different
dimension. .....................................................................................................24
Figure 3.9: Illustration of deviation of original best match (left) model’s fracture
ensemble. ......................................................................................................31
Figure 4.1: Visualizations of basic reservoir and well information in the corner point
gridding system: (a) Top view; (b) 3D view; (c) Regional 3D well
xii
Figure 4.2: Statistical distributions of spatial configuration properties of natural
Figure 4.3: Visualization of reservoir dip and depth variations for the bulk field
Figure 4.4: Relative permeability curves for modeling gas and water flow for this
Figure 4.5: Porosity field with heterogeneity in both horizontal directions and
Figure 4.6: Comparison between the conventional hydraulic fracture design and
Figure 4.7: Fracture modelling software outputs for the hydraulic fractures modelled
(light blue) from the perforation locations: (a) Overview; (b) Enlarged
view. ..............................................................................................................45
Figure 4.9: Visualization of tilting from the horizontally generated fractures (light
blue) to the fractures that aligns with the dip of the reservoir model
(green): (a) Overview; (b) Heel side view; (c) Toe side view. .....................46
Figure 4.10: Visualization of maximum and minimum permeability decline curve. ........49
Figure 4.11: Visualization of maximum and minimum half-lengths values for both
outer and inner fractures: (a) Minimum outer and inner fracture half-
Figure 4.14: Comparison between simulation results from initial iteration (Latin
flow rate; (b) Cumulative gas production; (c) Water flow rate; (d)
Cumulative water production; (e) Water gas ratio; (f) Flowing BHP. .........55
Figure 4.15: Number of history matching solutions versus iteration number. ..................57
results and actual production history: (a) Gas flow rate; (b) Cumulative
gas production; (c) Water flow rate; (d) Cumulative water production;
Figure 4.17: Comparison between best match’s simulation results and actual
production history: (a) Gas flow rate; (b) Cumulative gas production; (c)
Water flow rate; (d) Cumulative water production; (e) Water gas ratio;
Figure 4.18: Comparison between prior and posterior distributions of the studied
saturation; (d) Fracture height; (e) Outer fracture half-length; (f) Inner
Figure 4.19: Characterized fracture geometry from AHM workflow: (a) Non-uniform
xiv
Figure 4.20: Pressure-dependent fracture permeability decline curve of 55 history
history matching solution (light orange); and best match solution (red). .....68
Figure 4.22: Comparison between response parameters’ values and relative errors of
proxy model and simulation results for iteration 2, 4, 6, 8 and 10: (a)
Water flow rate at 325 days; (b) Flowing BHP pressure at 334 days; (c)
Water flow rate relative error; (d) Flowing BHP relative error; (e) Global
Figure 4.23: Scatter plots of values of different uncertain parameters and global error
height; (e) Outer fracture half-length; (f) Inner fracture half-length; (g)
Figure 4.24: Production forecast of 55 history matching solutions: (a) flowing BHP;
(b) Gas flow rate; (c) Cumulative gas production; (d) Water flow rate;
Figure 4.25: Connectivity analysis of the natural fractures presented in the history
Figure 4.26: Spatial connectivity analysis of the natural fractures magnified: (a) 1.5
times in both lengths and heights; (b) 1.8 times in both lengths and
heights. ..........................................................................................................80
xv
Figure 4.27: Visualizations of the effects of the natural fractures’ connectivity on
both length and height; (d) Natural fractures enlarged 1.8 times in both
natural fractures; (c) Natural fractures enlarged 1.5 times in both length
and height; (d) Natural fractures enlarged 1.8 times in both length and
height.............................................................................................................83
both length and height; (d) Natural fractures enlarged 1.8 times in both
Figure 4.30: Comparison of 20 year gas estimated ultimate recovery for four different
Figure 5.1: A 3D example model visualizations for the reservoir, hydraulic fractures
Figure 5.2: Fixed relative permeability used for this study: (a) Matrix; (b) Hydraulic
fractures.........................................................................................................90
Figure 5.3: Connectivity check and non-connected natural fractures (pink) removal
xvi
Figure 5.4: Illustration of negligible effects from removing non-connected fractures
distributions: (a) Df=2.6, Dl=2.45, α=4.5; (b) Df=2.8, Dl=2.45, α=4.5; (c)
Df=2.7, Dl=2.3, α=4.5; (d) Df=2.7, Dl=2.6, α=4.5; (e) Df=2.7, Dl=2.45,
Figure 5.7: Locations of proxy prediction points for calculating values of response
Figure 5.8: Comparison between simulation results from initial iteration (Latin
flow rate; (b) Cumulative gas production; (c) Water flow rate; (d)
Cumulative water production; (e) Water gas ratio; (f) Flowing BHP. .......103
Figure 5.9: Number of history matching solutions versus iteration number. ................105
results and actual production history: (a) Gas flow rate; (b) Cumulative
gas production; (c) Water flow rate; (d) Cumulative water production;
Figure 5.11: Comparison between best match’s simulation results and actual
production history: (a) Gas flow rate; (b) Cumulative gas production; (c)
Water flow rate; (d) Cumulative water production; (e) Water gas ratio;
saturation; (d) Matrix porosity; (e) Fracture height; (f) Outer fracture
Figure 5.13: Characterized fracture geometry from AHM workflow: (a) Overview;
history matching solution (light orange); and best match solution (red). ...117
Figure 5.16: Comparison between response parameters’ values and relative errors of
proxy model and simulation results for iteration 2, 4, 6, 8 and 10: (a)
BHP at 30 days; (b) WGR at 20 days; (c) BHP at 745 days; (d) WGR at
743 days; (e) BHP at 1626 days; (f) WGR at 1623 days; (g) BHP
relative error; (h) WGR relative error; (i) Global relative error. ................119
xviii
Figure 5.17: Scatter plots of values of different uncertain parameters and global error
porosity; (e) Fracture height; (f) Outer fracture half-length; (g) Inner
error. ............................................................................................................122
BHP; (b) Gas flow rate; (c) Cumulative gas production; (d) Water flow
Figure 5.19: Pressure distribution of matrix at different production time: (a) 365 days;
(b) End of history (1626 days); (c) 10 years; (d) 20 years. .........................127
time: (a) 365 days; (b) End of history (1626 days); (c) 10 years; (d) 20
years. ...........................................................................................................128
xix
Chapter 1: Introduction
1.1 MOTIVATION
to be only the source rock that provides hydrocarbon or prevents hydrocarbons from
escaping. With the unlock of technology such as hydraulic fracturing and horizontal
nowadays. Nevertheless, significant challenges remain for shale reservoirs because of the
reservoirs are mainly feasible because of hydraulic fractures, and without proper
fracture modelling techniques, properties of the shale petrophysics and hydraulic fracture
called history matching. Due to current technology limits, it is not possible to fully
visualize the actual geometries of hydraulic fractures subsurface, making the history
matching problem ill-posed. Myriad combinations of the physical properties can match
the production data, and it is nonviable to validate which of these combinations is true. In
addition, some of the physical processes are still unknown to engineers, and others are
not fully understood yet. The advent of computer aided (or automatic) history matching
1
From another aspects, the collaborations between geologist and engineers should
be even closer than before. The problem of how the work of field geologists can provide
insights to unconventional reservoir engineers is one of the keys to better solve the
knowledge would be the field-scale heterogeneity and natural fracture modelling, which
Bayesian method and advanced machine learning algorithm to better study and
geology model and fractal dominated natural fractures usually observed in outcrops, are
review on fracture modelling, history matching, corner point gridding of geology model,
and fractal theory behind natural fracture generation is provided.
Chapter 4 introduces a shale gas field example to demonstrate how geology model
2
forecast, natural fracture sensitivity studies, and pressure visualizations are also
performed.
In Chapter 5, another shale gas well is history matched automatically to show how
fractal theory could assist to generate a more realistic natural fracture network. Similarly,
In Chapter 6, all the studies and the corresponding findings are summarized, and
3
Chapter 2: Literature Review
developments, and potential solutions. Background in four main aspects, from fracture
modelling; history matching; complex gridding system and natural fractures, will be
thoroughly introduced.
With the advent of hydraulic fracturing and horizontal drilling, more hydrocarbon
though fracturing is vastly applied in nearly all unconventional reservoirs, the modelling
under developments, never minding the unfathomable myths about subsurface nature
significant because it greatly affects the quality of history matching results. Therefore, we
topics in fracture modelling, namely categories of the model; workflow of the model; and
related to the fracture modelling, some of which are not well understood. These includes
RTA (rate transient analysis) analysis; net pressure modelling; initial shut-in pressure
(ISIP) and geomechanical heterogeneity (McClure et al. 2020). Let alone the complexity
4
of hydraulic fracture modeling, the occurrences of natural fractures in the shale reservoirs
al. 2009; Yang et al. 2018), unstructured grid (Wang 2015; Bosma et al. 2017; Ding
2019), local grid refinement (Cipolla et al. 2009; Abdle Moneim et al. 2012; Xu et al.
2019), and embedded discrete fracture model (Xu et al. 2017; Yu et al. 2018b; Miao et al.
2018).
The first method, dual porosity dual permeability model (DPDK), has complex
Despite its considerable computational efficiency advantage, DPDK model cannot study
the complex fracture network system such as natural fractures and 3D configurations. In
addition, according to Rubin (2010), DPDK model fails to accurately model the matrix-
Furthermore, Kuchuk and Biryukov (2014) also points out that DPDK model’s limitation
and Sun et al. (2016) points out that DPDK model and its’ variants couldn’t properly
5
to be computed (Hassanpour et al. 2008). Therefore, for complex geometry of fractures,
unstructured grid struggles with the gridding arrangements and poor efficiency.
The third method, local grid refinement (LGR), has enhanced accuracy for
modelling fractured reservoirs. Nevertheless, Cipolla et al. (2009) proposed that LGR can
suffer from poor computational efficiency. Extra refinements around well blocks are also
required to accurately model flow scheme (Ghosh 1998). Tijink and Cottier (2019)
pointed out that extensive refinements must be performed to avoid invalidating results
due to vertically partial LGR if complex geological feature is presented. This would
corner point gridding model, which will be discussed later in this chapter.
The last method, embedded discrete fracture model (EDFM), has gained favorable
several researchers (Moinfar et al. 2014; Cavalcante Filho et al. 2015; Shakiba and
Sepehrnoori 2015; Xu et al. 2017). Extensive benchmarking of EDFM method with LGR
method is completed by Xu et al. (2017), and it is found out that EDFM method could
achieve similar accuracy with LGR but could reduce computational time by 20 times for
efficiency made it possible for properly model hydraulic or natural fractures in the history
matching process. Xu et al. (2017) constructed the EDFM method utilizing non-
blocks are added to original reservoir model to account for modified transmissibility
between fracture media, matrix media, and well media. One main drawback of the EDFM
method is that it assumes linear pressure distributions in matrix blocks that contain
6
(a) Physical domain
Figure 2.1: Illustration of the concept for Embedded Discrete Fracture Model: (a)
Physical domain; (b) Numerical domain. (Xu et al. 2017).
each method is not able to model hydraulic or natural fractures perfectly. Besides,
sensing (DTS/DAS), advances, more clarity will be brought to the discipline of fracture
modelling. For now, the EDFM method stands out to be the best fracture modelling
7
2.2 HISTORY MATCHING
the estimated ultimate recovery (EUR), the numerical optimization method such as
history matching is required. In this section, we review some of the current history
matching techniques. In general, history matching is divided into two main categories:
The manual history matching requires the engineer to manually adjust one
parameter at a time to potentially match the historical reservoir responses. Tavassoli et al.
(2004) pointed out that manual history matching is considered to be utterly time
inefficient due to its trial-and-error nature. In addition, Kabir et al. (2003) stressed that
manual history matching is highly subjective and depends exclusively on the engineer’s
field knowledge/experience. The manual history matching is also very difficult to keep
up with the pace of the updates of real-time data (Kabir et al. 2003). Moreover, due to the
ill-posed problem of history matching, it is often not enough to obtain single realization
of the reservoir properties. Results from manual history matching is deemed to be overly
The computer aided history matching (or automatic history matching, AHM) is
becoming the industry trend currently. The idea of AHM is to use mathematical
automatically, through machine languages. Many history matching work with various
optimization algorithms have been developed and applied. Arroyo Negrete et al. (2018)
test responses. However, this approach cannot handle larger number of uncertain
8
parameters, and the length of history data is too small to represent field production. Gao
et al. (2017) used Gauss Newton method to minimize objective function, a measure of the
history matching error, and obtained satisfactory results. Vazquez et al. (2015) applied
particle swarm algorithm in an oil reservoir, and Xie et al. (2015) implemented
ensemble of realizations. Nevertheless, the history matching methods discussed above all
fail to properly characterize the uncertainties associated with these reservoirs. In fact,
Goodwin (2015) proposed that these optimization-based algorithms for history matching
is statistically biased. Even though some of the algorithms could generate a solution
ensemble, they are not designed to analyze the probabilistic distributions of model
parameters.
With these challenges, the proxy-based Monte Carlo Markov Chain (MCMC)
Monte Carlo simulation with memory. Large sample size can well approximate the shape
proxy models must be adopted to substitute the role of reservoir simulator and quantify
the objective functions for the generated samples. Wantawin et al. (2017) used a
polynomial proxy model for shale gas history matching. Yu et al. (2018a) and
Tripoppoom et al. (2018) experimented with k-nearest neighbor (KNN) model and
artificial neural networks (ANN) for history matching of shale gas wells. Among these
approaches, the one with ANN proxy model has the highest accuracy and computational
efficiency. However, one problem with the ANN method is that its accuracy is sub-
optimal if the training data size is small (Sejdinovic 2015), which is common for history
9
matching problems due to the computation costs associated with reservoir simulations.
XGBoost, which has great accuracy when the training dataset size is small. This method
system must be adopted. In this section, we briefly review some of the work completed to
In corner point gridding system, combinations of coordinate lines and nodal depth
are given (Ponting 1989) to represent complex geological features such as faults, cross-
stratified beds, and sophisticated boundaries (Ding and Lemonnier 1995). The utilization
of geology model in corner point gridding system could thus greatly reduce the
al. (2013) employed the corner point geology model into the numerical simulation of oil
sand steam-assisted gravity drainage (SAGD) process, but the simulation time for the fine
gridded geology model is impractical. Therefore, upscaling of the geology model should
be performed to be suitable for history matching process, especially for the fractured
continuum and unstructured grid, they are extremely challenging to be adapted into the
sophisticated corner point gridding system (Du et al. 2017). However, numerous studies
10
of implementing EDFM in the corner point gridding system are conducted (Panfili and
Cominelli 2014; Panfili et al. 2015; Shah et al. 2016). Xu and Sepehrnoori (2019) further
extended the EDFM to numerically handle complex geometries encountered in the corner
point model. Therefore, a robust performance of generalized EDFM method in the corner
point gridding system is achieved and the formulations laid a solid foundation to couple
The effects of natural fractures on automatic history matching is not negligible, as natural
fractures both affect hydraulic fracture growth and also the probabilistic production
forecast from results of history matching. Wu and Olson (2016) extensively analyzed the
demonstrated the effects of natural fractures on long term productions. However, the
natural fractures could be highly biased and scale dependent. Darcel et al. (2003a) stated
that the natural fractures observed in many outcrops exhibit fractal distributions of their
center locations and power law distributions in their lengths. This observation is
supported by several other authors (Okubo and Aki 1987; Ouillon and Sornette 1996;
Main et al. 1990; Odling 1997). Kim and Schechter (2007) also stressed that the
unbiased.
11
Chapter 3: Methodology
3.1 INTRODUCTION
The objective of this chapter is to introduce the key steps in the automatic history
matching (AHM) workflow. The major components of the workflow contain Embedded
Discrete Fracture Model (EDFM), reservoir simulator, proxy model and Markov Chain
Monte Carlo sampling (MCMC). Flowcharts and schematic diagrams will be provided to
3.1. There are four major modules in this workflow: input module (light grey); iterative
optimization module (light yellow); solution module (light green); and post-processing
module (light blue). The input module checks data quality and assures that reservoir
model is correctly configured. It contains three components: field inputs; data treatments
and base model setup. The iterative optimization module generates hundreds of
realizations using artificial intelligence (AI) engine, advanced sampling methods, EDFM
sampling; hydraulic fractures (HFs) and natural fractures (NFs) modelled by EDFM;
reservoir simulator; AI; and Markov Chain Monte Carlo (MCMC) algorithm. Solution
module finds satisfactory results of AHM and obtains characterized parameters. This
module contains: solution filtering; best match; and posterior distribution. Post-
processing module provides potential guidance to the engineers with pressure drainage
and visualizations. In the next few sections, each module is introduced in detail.
12
Figure 3.1: General illustration of automatic history matching workflow.
13
3.3 INPUT MODULE
In this section, the input module of the automatic history matching workflow is
briefly introduced. The function of this module is to receive data, inspect and treat the
data, and prepare a base case model that is error free and also computationally efficient so
that the no problems will be encountered when the next module (iterative optimization
module) is executed.
available. However, not all of the available data could be used in the AHM workflow.
The most important input data for the AHM workflow is the production data. Such data
mainly includes production monitoring dates, gas flow rate, water flow rate, wellhead
pressure (WHP), bottomhole pressure (BHP), etc. for shale gas reservoir and an
additional oil flow rate for shale oil reservoir. Some of the other data might also be
The bottomhole pressure available in the field inputs are usually based on the
computation from rate transient analysis (RTA), which could have significant errors
wellbore model. However, these two methods are either unavailable or too costly in most
bottomhole pressure:
14
𝑊𝐺𝑅
𝐵𝐻𝑃 = 𝑊𝐻𝑃 + ∆𝑃 ∗ (3.1)
𝑊𝐺𝑅𝑚𝑎𝑥
where 𝐵𝐻𝑃 is bottomhole pressure, 𝑊𝐻𝑃 is wellhead pressure, ∆𝑃 is an empirical
constant pressure loss, 𝑊𝐺𝑅 is the water gas ratio, and 𝑊𝐺𝑅𝑚𝑎𝑥 is the maximum
the fluid flow rate data. It is acceptable to keep shut-in intervals if these intervals are
situated in the middle of the total history period. However, they should be removed if
they are situated near the end of the history period, because they increase the
After we completed the input and preparation of relevant data, it is crucial to build
a simulation model base case. Arbitrary values could be assigned to fracture geometrical
properties, petrophysical properties, etc., some of which will later become uncertain
parameters. The purpose of this component is not to get a match with the production data,
but rather to make sure the reservoir simulation runs error free and the total
the iterative optimization module will run smoothly, and hundreds of or even thousands
identifications. If the combination for the base model results in huge deviations of
15
3.4 ITERATIVE OPTIMIZATION MODULE
automatic history matching workflow is explained in detail. The function of this module
engine, advanced sampling methods, EDFM and reservoir simulator. The principles and
Based on the field experience and combinations of the base model, some of
for automatic history matching process to characterize. The rest of the parameters could
Parameters that the workflow current can assign to be either fixed/uncertain are listed in
Table 3.1. In this table, we also listed the sources of information that can provide some
unconventional reservoir type (gas or oil), depths intervals (deep, >3000 m or shallow,
<2000 m), geological structure (thin or thick net pay) and engineering designs (proppants,
pumping schedule, cluster/stage spacing). It is noteworthy to point out that for shale
This variable reduces the equivalent fracture permeability if the matrix pressure is
is the initial reservoir pressure, and 𝑃′ is a specific reservoir pressure after production
begins. Figure 3.2 graphically illustrates the concepts of fracture closure and permeability
decay factor. At deeper depth, the fracture is expected to close more easily, and thus a
General
Parameters Sources of information
classification
Matrix permeability DFIT/core data
Matrix properties Matrix porosity Well logging
Matrix water saturation Field experience
Fracture height DFIT/microseismic
Fracture half-length Microseismic
Fracture water saturation Empirical value
Fracture conductivity DFIT/experience
Fracture properties
Fracture aperture Usually fixed
Cluster efficiency Field experience
Fracture permeability decay
Depth interval
factor
Exponent of water relative
Empirical value
permeability curve
Endpoint of water relative
Relative Empirical value
permeability curve
permeability
Exponent of gas relative
properties Empirical value
permeability curve
Exponent of oil relative
Empirical value
permeability curve
Number of natural fractures Regional geology
Length of natural fractures Regional geology
Conductivity of natural
Natural fracture Regional geology
fractures
properties
3D fractal dimension Regional geology/Image Log
3D length dimension Regional geology/Image Log
3D alpha regulator Regional geology/Image Log
module and guarantee accurate results from solution module, we must perform an initial
sampling process. In addition, this sampling process must be smart enough to fully
consider the uncertainty space constructed by the possible combinations of the specified
uncertain parameters.
Shams (2016) has already discussed different types of sampling methods for the
initial sampling process. However, in our proposed workflow, the highly efficient Latin
Hypercube method is applied. The reason for this choice is because other methods, such
samples to be representative, but Latin Hypercube (LH) method only requires 2k number
greater number of samples, but 50 initial samples are deemed to be enough, based on the
18
balance between precision and time cost to run the reservoir simulation of these initial
samples.
The principle of the LH method is summarized next. First, the range for each of
the specified uncertain parameters is divided into N sub-intervals. Then, random samples
are drawn and only one sample is allowed to be drawn from each high dimension sub-
space. These samples are drawn in a way that no repeated samples are enclosed in the
same sub-interval for any uncertain parameters. Figure 3.3 schematically illustrates this
method. The black grids represent selected samples and notice two samples that share the
same column or row with each other is non-existent. The advantage of this method is that
a minimal number of generated samples can have a wide distribution across the entire
uncertainty space, so that the samples are not biased, or too stochastic.
Figure 3.3: Schematic illustration of the Latin Hypercube sampling method. (Shams
2016)
Based on the results from smart sampling, we need to input hydraulic fractures in
the model first. The main procedures for inputting hydraulic fractures are summarized in
Figure 3.4. First, we need to read in the well trajectory information and perforation
19
report. Then, according to geology-related information and also the sampling results of
the uncertain parameters, dip/strike/tilting angles are inputted as fixed variables while
half-length, height and heel bias information is given by one single realization of AHM
workflow. Next, the coordinates of all the perforation points are linearly interpolated
based on the X, Y, and Z, measured depth (MD) values of two adjacent well trajectory
ℎ 𝑡 ℎ)
𝑥𝑝 = 𝑥𝑤 + (𝑥𝑤 − 𝑥𝑤 ∗𝑟
ℎ 𝑡 ℎ
𝑀𝐷𝑝 − 𝑀𝐷ℎ (3.3)
{ 𝑦𝑝 = 𝑦𝑤 + (𝑦𝑤 − 𝑦𝑤 ) ∗ 𝑟 , 𝑟=
ℎ 𝑡 ℎ)
𝑀𝐷𝑡 − 𝑀𝐷ℎ
𝑧𝑝 = 𝑧𝑤 + (𝑧𝑤 − 𝑧𝑤 ∗𝑟
are the measured depths of perforation point, adjacent well trajectory point toward heel
Next, the coordinates of the two endpoints of the fracture’s middle axis are
where 𝑥𝑚𝑎 , 𝑦𝑚𝑎 , and 𝑧𝑚𝑎 are the coordinates of the two endpoints of the fracture’s
middle axis, 𝑥𝑓 is the half-length of the bi-wing planar hydraulic fracture, tilting is the
angle between x-y plane and the middle axis (refer to Figure 3.5b), and strike is the angle
20
Then, the coordinates of the four vertices of the planar hydraulic fracture is
computed based on the calculated 𝑥𝑚𝑎 , 𝑦𝑚𝑎 , and 𝑧𝑚𝑎 and geometrical projections of the
vertices pair on the same side of the wing on x-z and x-y planes, as shown in Figure 3.5c
where 𝑥𝑣 , 𝑦𝑣 , and 𝑧𝑣 are the coordinates of the four vertices of the hydraulic fracture,
ℎ𝑓 is the hydraulic fracture height, dip is the angle between hydraulic fracture plane and
x-y plane. Then, these coordinates are outputted and also natural fractures’ coordinates if
Figure 3.4: Summary of the workflow for inputting the hydraulic fractures in arbitrary
gridding system.
21
(a) 3D view (b) y-z plane view
Figure 3.5: Graphical illustration of the variables and definitions used for inputting the
hydraulic fractures in arbitrary gridding system.
Once the hydraulic fractures are successfully inputted into the reservoir model, we
can then add the background natural fractures in the model. There are generally three
ways to include natural fractures in our workflow. The first one involves a fixed or
characterized natural fracture network. This could come from a random distribution, or
22
from the field geologist working on the project. This method is deterministic and cannot
consider the associated uncertainty with natural fractures. The second method generate
random realizations of total number, length, and conductivity of the natural fractures
within the coordinate ranges of the model, with single/multiple sets of fixed dip and
strike angles. This method is probabilistic to properly consider the uncertainty associated
with natural fractures, but it is deemed to be too stochastic with high randomness. The
last method involves the use of 3D fractal theory. Darcel et al. (2003a) proposed that
natural fracture networks in nature, as observed in various outcrops, follow two unique
geometrical patterns: fractal distributions of fracture centers and power law distributions
of lengths. Hence, this method is beneficial because it is more realistic than the second
method, which lacks physical basis. The equation for fractal controlled natural fracture
3+𝐷
Γ ( 2 𝑙) (3.6)
𝑛(𝑙, 𝐿)𝑑𝑙 = 𝛼3𝐷 𝐿𝐷𝑐3𝐷 𝑙 −(𝐷𝑙3𝐷+1) , 𝛼3𝐷 = 𝛼√𝜋
2+𝐷
Γ ( 2 𝑙)
where 𝑛(𝑙, 𝐿) is the density of natural fractures given a minimum fracture length of 𝑙
and a model scale of 𝐿, 𝛼3𝐷 is the 3D alpha regulator are derived from 2D fractal
parameters, 𝐷𝑐3𝐷 is the 3D fractal dimension number, 𝐷𝑙3𝐷 is the 3D length dimension
number, 𝛼 is the 2D alpha regulator that controls natural fracture density, Γ is the
gamma function, and 𝐷𝑙 is the 2D length dimension number. The power law distribution
of the natural fracture can then be back-interpolated by this equation. Figure 3.6
schematically shows the effects of fractal and length dimensions on the spatial
23
(a) 2D fractal (b) 3D fractal
Figure 3.6: Graphical illustration of the fractal theory and the effects of different fractal
parameter on natural fracture network: (a) 2D fractal; (b) 3D fractal. (Darcel
et al. 2003b). The D is fractal dimension and a is length dimension.
After the preprocessing of both hydraulic and natural fractures in the reservoir
model using EDFM preprocessor, the reservoir simulations will be performed. The non-
intrusiveness of the EDFM preprocessor allows us to execute the simulations using any
of the third-party simulation software, including CMG, ECLIPSE, NEXUS, and etc. In
our proposed workflow, the black oil simulator – CMG’s IMEX is utilized. We did not
select the CMG’s compositional simulator – GEM because of its high computational
costs. One key aspect that needs to be considered is the Langmuir isotherm gas
desorption and absorption phenomenon that is unique for shale gas simulation. The
𝐾 ∙ 𝑃𝑚
𝛼𝑎𝑑𝑠 = 𝜌(1 − 𝜙)𝑉̅𝑚𝑎𝑥 ( ) (3.7)
1 + 𝐾 ∙ 𝑃𝑚
24
where 𝛼𝑎𝑑𝑠 is the unit fractional adsorption, 𝜌 is the rock density, 𝜙 is the matrix
porosity, 𝑉̅𝑚𝑎𝑥 is the maximum volume of absorbed gas per unit mass of rock, 𝐾 is the
establish the artificial intelligence engine using proxy models. The purpose of proxy
model is to establish a relationship between simulation input and output, so that it can
mimic a reservoir simulator and avoid actual simulation runs. There are many machine
nearest neighbor (KNN), deep neural network (DNN), polynomial, kriging, etc. However,
machine learning models that optimize the objective function collaboratively. The
ensemble learning models embedded in XGBoost are the regression decision trees, which
is predicts the outcome based on several inputs using the tree algorithms. A simplified
schematic structure for the XGBoost is provided in Figure 3.7. One unique feature about
XGBoost is the objective function is additive and customized, so that the loss function
where 𝐿𝑋𝐺𝐵 is the overall loss function, 𝑁 is the total training size, 𝐿(𝑦𝑖 , 𝐹(𝑥𝑖 )) is the
loss function value at data point i given input data of xi and output data of yi, 𝑀 is the
total number of sub-loss function, and Ω(ℎ𝑚 ) is the regularization function, which is
As outlined by Malik et al. (2020), there are several advantages of XGBoost. The
model has a build-in cross validation methods, which prevents excessive overfitting
issues common among machine learning models. It also has some embedded search
algorithm to partially aid the selection of hyperparameters. The code structure is
missing data. In general, the implementation of neural network is optimal when the size
of the available data is massive. For the ill-posed history matching problem, the data size
is limited because of the limited computational time we assign to the task. Therefore, the
26
XGBoost algorithm would relatively converge faster and have better accuracy than neural
The exact method we utilized the proxy model is to use the model to predict the
values of selected representative time locations along the history data. The input of the
proxy model would be combinations of the uncertain parameters’ combinations, and the
output of the proxy model would be the value of different response variable (water flow
rate, bottomhole pressure, etc.) at a single time location. Usually 10 proxy locations are
selected, so for each of the response variable, 10 proxy models are built. For shale gas
reservoir, we generally select BHP and water flow rate (or water gas ratio) to be the
response variables to be predicted by proxy models, while for shale oil reservoir, an
additional variable of oil gas rate is incorporated. To measure the error of the multi-proxy
models or actual simulation run for each response variable, we calculate the objective
function as
𝑦𝑖,𝑚𝑜𝑑𝑒𝑙 − 𝑦𝑖,ℎ𝑖𝑠𝑡
∑𝑝𝑖=1 | × 100| × 𝑤𝑖 (3.9)
𝑦𝑖,ℎ𝑖𝑠𝑡
𝐹𝑗 =
∑𝑝𝑖=1 𝑤𝑖
where 𝐹𝑗 is the objective function value for response variable j, p is the total proxy
locations selected for the response variable j, i is the index of the proxy location, 𝑦𝑖,𝑚𝑜𝑑𝑒𝑙
is the value of the response variable j at time i from either simulation run or proxy model,
𝑦𝑖,ℎ𝑖𝑠𝑡 is the value of the response variable j at time i from actual history data, and wi is
the weight assigned to the proxy point i, given a response variable j. Similarly, the global
objective function is calculated in a similar way (just sum up all response variables). The
global objective function value will be used in the next component (MCMC), and the
27
3.4.7 Markov Chain Monte Carlo (MCMC) Algorithm
Figure 3.8 highly summarizes the entire procedure of the MCMC algorithm used
in the iterative optimization module. First of all, we start at the step of initial 𝜃, which is
the combination of uncertain parameters from the smart sampling component with lowest
Next, we proceed into an iterative feature that repeats one hundred thousand steps
and a check is performed to see if the iterations reaches the maximum steps. If not, we
change called “walk distance” as 𝛿. Next a step acceptance ratio, a*, is calculated as
(𝜀 ∗ )2 −𝜀 2
∗
𝑎 = min {1, 𝑒 2𝜎2 } (3.10)
uncertain parameters, 𝜀 is the global objective function of the previous (in this case
function values for the smart sampling results. The multi-proxy model is embedded in
this step to predict the values of response variables for all of the locations designated
1. If the ratio is greater than this number, the proposed combination is accepted as the
new sample. Otherwise, the old combination is regarded as the new sample. The
procedure then repeats, and a new walk distance is assigned, until maximum steps is
reached.
After the maximum step is reached, we check with the overall acceptance ratio
and see if it is between the empirical range of 15% to 50%. If this criterion is not
28
satisfied, we redo the previous steps by adjusting the variance value used in Equation
3.10 using an adjusting ratio, which usually falls between 2 to 10. Otherwise, we can
proceed to the next step and remove the initial 20% of the generated samples, which is
called “burn-in” period. This is done because the initial samples have a larger bias and is
not representative of the true posterior distribution. The remaining 80% of the samples
are thus the outputs from the MCMC algorithm. However, we only filter out 25 samples
with minimum global objective function values to be validated with the reservoir
Figure 3.8: Simplified schematic structure of Markov Chain Monte Carlo algorithm.
This brief explanation about MCMC concludes the principles for the iterative
optimization. In general, the quality of the multi-proxy model built from the smart
29
sampling process is poor. However, as MCMC procedures continue and actual validation
cases accumulate, the multi-proxy model will gradually improve qualitatively. Based on
our history matching experience, a maximum iteration number of 10 for this module is
automatic history matching workflow. The function of this module is to filter out optimal
ensemble of history matching solutions, obtain the most representative match with the
lowest global error, and obtain the final posterior distributions for various uncertain
results are available. However, some of results will still have high values of global
objective function, and it would be much less flexible if we use global function values to
treat all the simulations. Therefore, we need to selectively filter the history matching
results by using different threshold for each individual objective function. Simulation
cases whose values of individual objective functions satisfy all of the thresholds will be
grouped as the final history matching solution ensembles. Based on experience, the
threshold for bottomhole pressure’s objective function is much stricter than the one for
water flow rate/ water gas ratio, because significant data quality issues are associated
30
with water production due to the water flow back phenomenon. Compared to overly
considers the uncertainties associated with the ill-posed history matching problem.
There is a one optimal history matching solution with the lowest global objective
function value from the ensemble of the history matching solutions. However, one
problem exists with this best match model. For some of the uncertain parameters, it is
possible that this optimal model is not lying close to the statistical P50 values of the
solution ensembles, as shown in Figure 3.9. This presents a statistical bias, since the best
match model cannot catch the variability associated with the uncertainty space, as P50
value is defined to be the most likely value in any arbitrary distributions. Therefore, in
our workflow, the best match model is generated around the P50 values of the filtered
Figure 3.9: Illustration of deviation of original best match (left) model’s fracture height
to the P50 value (right) of the history matching solution ensemble.
31
3.5.3 Posterior Distribution
Once the solution ensembled is filtered and the representative best match model is
obtained, the visualization for posterior distributions can be achieved, as shown in the
right plot of Figure 3.9. Given the specific uncertain range, the histograms of the
distributions of the solution ensemble is plotted as the orange bars, and the representative
best match is shown as the red vertical line. The prior distribution, which is assumed to
be uniform in our workflow, is also plotted and shown as grey bars. Lastly, we included
the cumulative probability function based on the histograms as dashed curve to help us
probabilistic analysis of ultimate estimated recovery (EUR) for the filtered solutions and
also visualize the pressure drawdown in the representative best match model, for both
matrix and fractures. This module’s function is intended for the work described above.
After the history matching period and depending on the length of history, we need
to assign BHP drawdown schedule for the production forecast. If the length of history is
very short (around or less than 3 months), we assign a gradual drawdown from the last
BHP value of the history period to a minimum BHP constraint (usually between 1 to 2
Mpa) over a period of time, which depends on the decline rate of BHP in the history
period. We could also assign a direct BHP drop to the minimum constraint if the length
of history is long. Then, the production will continue until a total simulation time of 20
years is achieved. The EUR for both gas phase and water phase will be computed from
32
the reservoir simulator. The process is probabilistic because the work is repeated for the
entire solution ensemble and the representative best match, instead of a deterministic
approach for single model EUR usually observed in conventional methods. This
component provides important appraisal of the productivity of the studied shale reservoir.
3.6.2 Visualizations
The last step in our workflow would be the visualizations of pressure drawdown
in the both the matrix and fracture media. For both structural gridding system and corner
point gridding system, our workflow is able to extract the pressure drawdown
connectivity of natural fractures, study the drainage extents of the hydraulic fractures.
These provides invaluable guidance towards well spacing optimization and future
33
Chapter 4: Application of Automatic History Matching to Corner-Point
Gridding System in Shale Reservoirs
4.1 INTRODUCTION
model from an actual shale gas field. The chapter consists of sequential explanations of
steps necessary to complete the study, including basic reservoir model description,
history matching and production forecast. We utilized the automatic history matching
module, which is composed of Markov chain Monte Carlo algorithm coupled with
XGBOOST proxy model, along with Embedded Discrete Fracture Model (EDFM) and
the corner-point gridding extension module to perform the study. The workflow is
applied to an actual shale gas well in Sichuan Basin in China with 334 days of production
history. The gas flow rate and water flow rate information are available, while the
This field case contains a single well with a horizontal section length of around
2000 m. Since the model is in corner point gridding system, the well is deviated and is
observed to pass through several geological layers. We assumed water saturation and
field geologists. The well was completed with 27 stages of hydraulic fractures with a total
34
For this study, we used a static model in corner point gridding system, and there
are 1139 natural fractures (only one characteristic set) inscribed in this model as
described by the geologists. Both a view of the reservoir model and 3D visualization of
well landing location is given in Figure 4.1. Statistical distributions regarding the spatial
configurations of these natural fractures are given in Figure 4.2. We also tabulated Table
4.1 to summarize the spatial information obtained from the geologists. Average length
and height of natural fractures are 66 m and 32 m, respectively, and the mean dip angle is
about 22 degrees. These information imply that the natural fractures have poor
connectivity with both the wellbore and hydraulic fractures and is not well communicated
between vertical layers. This will be discussed in detail in section 4.7. The model
dimension is approximately 1000 m long × 2200 m wide and the thickness varies across
locations and across layers. The shallowest point of the model is around 3194 m and the
deepest point is about 3489 m, as shown in Figure 4.3. The model in z-direction is
upscaled into 10 layers, with average thickness of 30 m, five layers of 6 m, 6.2 m, 4.5 m,
1.53 m, and 9.92 m. The number of grid blocks is 64 and 144 in x and y direction. Basic
reservoir, fractures, and Langmuir isotherm gas desorption properties are summarized in
Table 4.2. The relative permeability curves (Figure 4.4) for modeling gas and water flow
are assumed to be fixed because they affect the dynamic production insensitively. The
Table 4.1: Summary of statistical information regarding natural fractures in this study.
35
Model parameters Value Unit
Model dimension (x × y) 1000 × 2200 m
Number of grid blocks (x × y × z) 64 × 144 × 10 -
Initial reservoir pressure 70 MPa
Reservoir temperature 150 ℃
Residual water saturation 30% -
Total compressibility 4.35×10-7 kPa-1
Reservoir depth 3194 ~ 3489 m
Well length ~ 2000 m
Number of stages 27 -
Clusters per stage 3 -
Average stage spacing 21 m
Average cluster spacing 17.3 m
Langmuir pressure 7 MPa
Langmuir adsorption volume 3.5 m3/ton
Rock density 2.54 g/cm3
Table 4.2: Summary of basic reservoir, fracture and gas desorption parameters used in
this study.
(a) Top view of deviated well (red) and natural fractures (light blue)
37
(d) Global 3D well landing location
Figure 4.1: Visualizations of basic reservoir and well information in the corner point
gridding system: (a) Top view; (b) 3D view; (c) Regional 3D well landing
location; (d) Global 3D well landing location.
38
(c) Length (d) Surface area
Figure 4.3: Visualization of reservoir dip and depth variations for the bulk field model
(not upscaled).
39
Figure 4.4: Relative permeability curves for modeling gas and water flow for this
reservoir model.
Figure 4.5: Porosity field with heterogeneity in both horizontal directions and vertical
direction of the reservoir model used in this study.
40
4.3 HYDRAULIC FRACTURE GENERATOR IN THE CORNER-POINT GRID MODEL
example, the two hydraulic fractures closest to subsequent, previous stage are called outer
fractures. The one middle hydraulic fracture is called inner fractures. Usually, outer
fractures have larger half-length than the inner fracture(s) (Figure 4.6b). This design is
matching work, in which the fractures are assumed to have uniform half-length (Figure
4.6a). This is because the characterized non-uniform half-lengths both honors the fracture
growth modelled by fracture modelling software and provides useful guidance regarding
a more realistic drainage area and thus better understanding of optimal spacing. To
support this statement, we modelled fractures of this well on both one stage and the entire
well using fracture modelling software, as shown in Figure 4.7. We inputted maximum
and minimum horizontal stress of 98 Mpa and 86 Mpa, with a Young’s modulus of 46
GPa and a Poisson’s ratio of 0.24, which are representative of the geomechanical
information for this reservoir. The single stage model provides us with non-uniform
substantiated by the data provided by Wheaton et al. (2017). This implies that the non-
outer/inner fractures).
41
(a) Uniform half-lengths
Figure 4.6: Comparison between the conventional hydraulic fracture design and
simplified heel-biased hydraulic fracture design: (a) Uniform half-lengths;
(b) non-uniform half-lengths; (c) Non-uniform half-lengths with natural
fractures.
42
(a) Fracture modelling software output for a single stage
Figure 4.7: Fracture modelling software outputs for the hydraulic fractures modelled in
this study: (a) Single stage; (b) Entire well.
43
Since the actual perforation report for this well is available, the corner point
attentions must be paid regarding the dip of the reservoir. The region of the reservoir
model near the toe side of the well has a lower dip angle compared with the region near
the heel side. If hydraulic fractures are generated perfectly horizontal to the well bore,
they will grow outside the model and thus affects the results of the simulation, as shown
for the cluster at toe side and linearly increased the tilting angle until the first cluster of
14th stage, which has a tilting angle of 8 degrees. The clusters in the rest of the stages
toward heel side all have tilting angles of 8 degrees. This tilting scheme can be visualized
in Figure 4.9.
(a) Overview
44
(b) Enlarged view
Figure 4.8: Illustration of the problem of creating completely horizontal fractures (light
blue) from the perforation locations: (a) Overview; (b) Enlarged view.
(a) Overview
45
(b) Heel side view
Figure 4.9: Visualization of tilting from the horizontally generated fractures (light blue)
to the fractures that aligns with the dip of the reservoir model (green): (a)
Overview; (b) Heel side view; (c) Toe side view.
46
4.4 UNCERTAIN RESERVOIR AND FRACTURE PARAMETERS IDENTIFICATIONS
factor, fracture water saturation, fracture height, outer fracture half-length, inner fracture
half-length, fracture conductivity and natural fracture conductivity. These parameters are
example, since the depth of this reservoir is more than 3000 meters, the stress tends to be
high and therefore fractures are more prone to be closed. Thus, we used pressure
dependent permeability to model this effect. The equation for pressure dependent
parameters can be selected, but it would best to control the total number to be below 10 to
balance proxy efficiency and accuracy and to control non-linear degree of uncertainty
parameter. We also fixed the relative permeability curve (shown in Figure 4.4). The
matrix water saturation is assumed to be equal to residual water saturation, and the widths
of hydraulic and natural fractures are assumed to be 0.1 m to properly simulate water
flow back in early production period. Table 4.3 summarizes all the ranges used for the
eight uncertain parameters and deterministic values for the fixed parameters. The ranges
for the eight uncertain parameters are either based on history matching experience
engineers working on this field (fracture height, outer/inner fracture half-lengths). Figure
4.10 and Figure 4.11 contains the information regarding the maximum/minimum
47
pressure-dependent permeability decline and longest/shortest scenario for heel-biased
half-lengths, respectively.
Table 4.3: Summary of uncertain parameters, their ranges, fixed parameters, and their
deterministic values used for automatic history matching.
48
Figure 4.10: Visualization of maximum and minimum permeability decline curve.
49
(b) Maximum outer and inner fracture half-lengths
Figure 4.11: Visualization of maximum and minimum half-lengths values for both outer
and inner fractures: (a) Minimum outer and inner fracture half-lengths; (b)
Maximum outer and inner fracture half-lengths.
unconventional reservoir with two phase flow, gas flow rate is used constraint as well
instead of flowing bottomhole pressure (BHP). This is because the bottomhole pressure is
interpolated from wellhead pressure (WHP). In this study, the conversion from WHP to
BHP based on a pressure loss correlation is shown in Figure 4.12. A total of 20 proxy
points is assigned for the proxy model to make predictions on bottomhole pressure and
water flow rate, and the location assignments are shown in Figure. 4.13. These included
10 proxies for bottomhole pressure at day 3, 20, 50, 85, 121, 153, 200, 238, 297, and 334
50
(Figure. 4.13a), 10 proxies for water flow rate at day 3, 20, 50, 85, 121, 153, 200, 238,
297, and 325 (Figure. 4.13b). In theory, we can include infinite proxy points, but 10
Figure 4.13: Locations of proxy prediction points for calculating values of response
variables’ objective functions during history matching iterations: (a)
Flowing BHP; (b) Water flow rate.
After determining the ranges of uncertain parameters and the proxy locations, we
can proceed to the first iteration of the history matching by using the Latin Hypercube
sampling process. The 50 initial iteration samples are generated, and their simulation
results are shown in Figure 4.14. Again, we used gas flow rate as constraint for all
simulation runs (shown in Figure 4.14a) because this response is measured relatively
accurately. For one of the samples, the gas flow rate cannot be constrained, probably due
to its’ sub-optimal combinations of uncertain parameters. For water flow rate (Figure
4.14c) and flowing BHP (Figure 4.14f), the current ranges covers the production history,
meaning there are samples both over-estimating and under-estimating the history data.
52
(a) Gas flow rate
53
(c) Water flow rate
54
(e) Water gas ratio
Figure 4.14: Comparison between simulation results from initial iteration (Latin
Hypercube sampling process) and actual production history: (a) Gas flow
rate; (b) Cumulative gas production; (c) Water flow rate; (d) Cumulative
water production; (e) Water gas ratio; (f) Flowing BHP.
55
The automatic history matching process then continued iteratively. For each
iteration, we used single chain of MCMC and a proposal sample size of 20000. The
train:validation set ratio is 8:2, and the proxy model has an epoch size of 5 with a total
epoch number of 600. The maximum loop number for updating sampling variance is 8,
and the overall accept ratio should be between 0.15 and 0.51 for the posterior samples to
pass the filtering test. The variance update ratio is 2 for the chain if the overall accept
ratio is rejected. For the accepted posterior sample ensembles, the 25 samples with the
lowest global errors are selected to be inputted into the reservoir simulator. We also
defined different weight for each proxy location when calculating objective functions.
Proxies of BHP at days 3, 20, 297, and 334 and proxies of water flow rate at days 3, 20,
297, and 325 have weights of 1, while rest of the proxies have weights of 2. At early
period (days 3 and 20), the production is not stabilized, and the data does not necessarily
reflect the response from the reservoir. At late period (days 297+), there are some
fluctuations with the data as reflected in Figure 4.13. These two reasons caused the
For this study, a total of 10 extra iterations is run in addition to the initial iteration
(50 cases). For each of the extra iteration, 25 cases are included, thus giving a total of 300
simulation results. The entire history matching workflow took approximately 15 hours to
finish. We then filtered out the history matching solution out of the 300 simulations based
on one threshold for each of the response variables (flowing BHP and water flow rate).
Objective function threshold values of 10 and 33 are used for BHP and water flow rate,
respectively. The threshold value is stricter with BHP because the quality of BHP
response should be smoother, and issue of water flow back exists. A total of 55 solutions
satisfied the criteria. The number of solutions versus iteration run is visualized in Figure
56
4.15. As can be seen from this plot, the number of solutions first increases and then
decreases, implying that the maximum iteration number could be reduced to improve
efficiency. The simulation responses for these 55 history matching solutions can be
visualized in Figure 4.16. We can observe from this plot that all the history matching
solutions meet the well constraint of gas flow rate. In addition, the results for water flow
rate heavily concentrates near the actual production history, except for the fluctuations at
the late period. The results for pressure also cover the trend of the drawdown decline,
except for the time after the two shut-in periods. This can be explained by the operational
uncertainties associated with shut-in, and the quality of the match could be further
improved if shut-in time could be avoided. These general observations suggest that the
workflow successfully optimized the results compared to the initial iteration’s suboptimal
results. Figure 4.17 shows the simulation results for the best match model, and they
57
(a) Gas flow rate
58
(c) Water flow rate
59
(e) Water gas ratio
61
(c) Water flow rate
62
(e) Water gas ratio
Figure 4.17: Comparison between best match’s simulation results and actual production
history: (a) Gas flow rate; (b) Cumulative gas production; (c) Water flow
rate; (d) Cumulative water production; (e) Water gas ratio; (f) Flowing BHP.
63
The information regarding the best match model is given below. Matrix permeability is
38.1 nano-darcy (nd), fracture permeability decay factor is about 0.0249 and the decline
curve can be visualized in Figures 4.18 and 4.20. As can be implied from the
permeability decay curve, the degree of fracture closure is pronounced because the
pressure (the range for this parameters produced a maximum multiplier of 0.6 to a
minimum multiplier of 0.05). Fracture water saturation is about 0.853 and fracture height
is 8.62 m. The outer fracture and inner fracture half-lengths are 130 m and 54.3 m,
natural fractures are 90.8 md-m and 10.7 md-m. This best match model is generated
around the P50 values of all 55 history matching solutions to consider the effects of
uncertainties. From the plot, we can observe that the solution ensembles for matrix
permeability, fracture height, and fracture conductivity are more concentrated on the
lower end of their ranges, while solution ensembles for fracture water saturation and
outer fracture half-length are more concentrated on the higher end of their ranges. These
uncertain parameters more sensitively affect the dynamic production. However, other
parameters are less sensitive because any values within their ranges can compose the
solution ensembles. Usually, the half-length for hydraulic fractures are considered very
sensitive parameter. Nevertheless, in this study, the inner fractures with lower half-
lengths have less contributions to the production and are thus less sensitive. Parallel
coordinates plot is drawn in Figure 4.21 and it can also prove that the abovementioned
statements. More sensitive parameters should have solution ensemble lines passing
through focused part of the corresponding vertical coordinates, and less sensitive
parameters should have solution ensemble lines spanning the entire vertical coordinates.
64
(a) Matrix permeability (b) Fracture permeability decay factor
65
(g) Hydraulic fracture conductivity (h) Natural fracture conductivity
Figure 4.18: Comparison between prior and posterior distributions of the studied
uncertain parameters based on 55 history matching solutions: (a) Matrix
permeability; (b) Fracture permeability decay factor; (c) Fracture water
saturation; (d) Fracture height; (e) Outer fracture half-length; (f) Inner
fracture half-length; (g) Hydraulic fracture conductivity; (h) Natural fracture
conductivity.
Figure 4.19: Characterized fracture geometry from AHM workflow: (a) Non-uniform
fracture half-lengths; (b) Fracture height.
In order to gauge the accuracy and precision of the workflow, we plotted the
values of the two response variables (water flow rate and BHP) at the last proxy location
for both proxy model and the actual simulation runs (Figure 4.22a, b). In addition, the
objective function value for water flow rate and BHP, and global relative error are also
plotted, as shown in Figure 4.22c, d, e. Iteration 2, 4, 6, 8, and 10 are selected so that the
improvement of proxy model could be visualized along with progression of iterations.
From Figure 4.22, it is well perceived that the value of response variables tends to be
more consistent between proxy prediction and simulation runs, as training data for proxy
8, since the objective functions and global errors can be optimized at iteration 8.
68
(a) Water flow rate at 325 days
69
(c) Water flow rate relative error
70
(e) Global relative error
Figure 4.22: Comparison between response parameters’ values and relative errors of
proxy model and simulation results for iteration 2, 4, 6, 8 and 10: (a) Water
flow rate at 325 days; (b) Flowing BHP pressure at 334 days; (c) Water flow
rate relative error; (d) Flowing BHP relative error; (e) Global relative error.
uncertain parameters and global error versus simulation cases (In Figure 4.23, red circles
are AHM solutions). From Figure 4.23a, c, we can conclude that the matrix permeability
and fracture water saturations are very sensitive to the dynamic production. Figure 4.23d,
e, g substantiates that fracture height, outer fracture half-length and hydraulic fracture
conductivity are also sensitive, but not at same level with matrix permeability and
fracture water saturation. Figure 4.23b, f, h shows that fracture permeability decay factor,
inner fracture half-length and natural fracture conductivity are least sensitive parameters
to dynamic production. Finally, Figure 4.23i shows that the workflow can optimize the
72
(g) Hydraulic fracture conductivity (h) Natural fracture conductivity
Figure 4.23: Scatter plots of values of different uncertain parameters and global error
versus simulation index: (a) Matrix permeability; (b) Fracture permeability
decay factor; (c) Fracture water saturation; (d) Fracture height; (e) Outer
fracture half-length; (f) Inner fracture half-length; (g) Hydraulic fracture
conductivity; (h) Natural fracture conductivity; (i) global error.
the previous section. The BHP drawdown schedule after the history period can be
visualized in Figure 4.24a. This schedule follows a gradual (spanning 9 months) decline
from the BHP value of the last history date of each solution’s simulation result and
73
remains at the minimum BHP constraint of 800 kPa until a total simulation time of 20
years is achieved. The reason we adopted this strategy is to avoid sudden increases in
water/gas flow rate right after history period ends. The gas flow rates, cumulative gas
productions, water flow rates, cumulative water productions are provided in Figure 4.24b,
c, d, e, respectively. Although the gas flow rates and water flow rates shows peaks after
production history, the magnitude of this peak is well controlled. For 20-year gas
estimated ultimate recovery (EUR), the P10, P50, P90 and best match values are 114.5,
124.4, 131.3, and 119.0 million cubic meters. For 20-year water EUR, the P10, P50, and
P90 and best match values are 26.7, 28.7, 30.8, and 28.9 thousand cubic meters. Gas
EUR for best match model tends to be little bit lower than the P50 values, and water EUR
is perfectly matching with the P50 values. Indeed, the integration of geology model and
the automatic history matching workflow greatly reduced the uncertain range for both 20-
year gas and water EUR. Without the heterogeneity of porosity, the ranges for these EUR
75
(d) Water flow rate
Figure 4.24: Production forecast of 55 history matching solutions: (a) flowing BHP; (b)
Gas flow rate; (c) Cumulative gas production; (d) Water flow rate; (e)
Cumulative water production.
76
4.7 NATURAL FRACTURE SENSITIVITY STUDY AND VISUALIZATIONS
to analyze the connectivity of natural fractures and its effects on the pressure drawdown
in both matrix and fractures, drainage volumes and 20-year gas EUR. Figure 4.25 shows
the grouped natural fractures and the total number of natural fractures in each group.
Although the number of connected fractures is high, the communication between these
fractures is not negligible. In fact, only 2 natural fractures are connected to the wellbore.
Therefore, the impact of the poor connectivity on gas EUR is profound. To prove this
point, we compared four scenario with different connectivity of natural fractures, which
are no natural fractures; the original dimension of natural fractures; natural fractures’
location preserved but their length and heights are magnified by 1.5 times (Figure 4.26a);
and natural fractures’ location preserved but their length and heights are magnified by 1.8
times (Figure 4.26b). The number of non-connected fractures (grey fractures in Figure
4.26) for 1.5 times magnification and 1.8 times magnification are 237 and 135, meaning
the connectivity significantly improved. In Figure 4.27, we also plotted the matrix
pressure distributions of the eighth layer from the top, where the most proportion of the
well lands in, for these four scenarios. These four models are created using the best match
properties, and the only difference is the presence of natural fractures. It is very
noticeable that for the scenario without natural fractures, the matrix pressure drawdown is
only confined to the region near the wellbore. However, as the presence and connectivity
of natural fractures increase, the pressure drawdown area extends to region far away from
the wellbore. In addition, for the enlarged natural fractures (Figure 4.27c, d), the pressure
drawdown are connected from the 7th layer to 8th layer, as opposed to Figure 4.27a, b
77
where the toe side region of 8th layer do not display any drawdown because the well
trajectory penetrated the top layer. Figure 4.28 depicts the drainage volumes for the four
scenarios studied. For no natural fractures and original natural fractures scenarios (Figure
4.28a, b), the drainage volumes are enclosed only in bottom five layers, or only small
portions of top five layers. However, the enlarged natural fracture scenarios exhibit
prominent drainage in top five layers as well. Figure 4.29 describes the pressure
drawdown in all the fractures. The original natural fracture scenario (Figure 4.29b) shows
However, with the enlarged natural fractures, a significant network is developed and with
this improved connectivity, more gas could be unlocked from this low permeability shale
reservoir. The comparison between 20-year gas EUR of the four scenarios are shown in
Figure 4.30. The gas EUR for these four scenarios are 90.9, 123.8, 219.5 and 297.1
million cubic meters (or -26.6%, 0%, 77.3%, 140% comparatively if the original scenario
(a) Spatial distribution for different natural fracture groups (color coded)
Figure 4.25: Connectivity analysis of the natural fractures presented in the history
matched model: (a) Spatial distribution for different natural fracture groups;
(b) Total number of natural fractures in different groups.
Figure 4.26: Spatial connectivity analysis of the natural fractures magnified: (a) 1.5 times
in both lengths and heights; (b) 1.8 times in both lengths and heights.
80
(b) Original natural fractures
(c) Natural fractures enlarged 1.5 times in both length and height
(d) Natural fractures enlarged 1.8 times in both length and height
Figure 4.27: Visualizations of the effects of the natural fractures’ connectivity on matrix
pressure distributions after 20 years: (a) No natural fractures; (b) Original
natural fractures; (c) Natural fractures enlarged 1.5 times in both length and
height; (d) Natural fractures enlarged 1.8 times in both length and height.
81
(a) No natural fractures
82
(c) Natural fractures enlarged 1.5 times in both length and height
(d) Natural fractures enlarged 1.8 times in both length and height
(c) Natural fractures enlarged 1.5 times in both length and height
84
(d) Natural fractures enlarged 1.8 times in both length and height
Figure 4.29: Visualizations of the effects of the natural fractures’ connectivity on fracture
pressure distributions after 20 years: (a) No natural fractures; (b) Original
natural fractures; (c) Natural fractures enlarged 1.5 times in both length and
height; (d) Natural fractures enlarged 1.8 times in both length and height.
Figure 4.30: Comparison of 20 year gas estimated ultimate recovery for four different
natural fracture connectivity scenarios.
85
4.8 CONCLUSIONS
geology model with a modified proxy engine to achieve practicality, efficiency, and
proximity with reality. The workflow performed history match and statistical production
forecast for a real case of shale gas well. In this chapter, we concluded the followings:
BHP and water rate for a shale gas well. The workflow found 55 HM solutions
with 8 uncertain parameters from 11 iterations or 300 total simulation runs, which
2. The eight posterior distributions were obtained from the workflow. The matrix
factor, inner fracture half-length and natural fracture conductivity are least
sensitive.
3. The heterogeneous porosity field reduces the uncertainties associated with the
towards optimal well spacing designs and future fracking job designs.
workflow to further reduce uncertainties and supply field engineers with practical
guidance.
86
Chapter 5: Development of Realistic Natural Fracture Module in AHM
by Using Fractal Theory
5.1 INTRODUCTION
Due to the facts that extensive number of natural fractures and faults exists among
the shale gas reservoirs; it is paramount to incorporate natural fractures within the
numerical model. These geological features have huge implications regarding long-term
EUR, well spacing optimization, and fracturing/refracturing job design. Therefore, the
objective of this chapter is to study the effects of natural fractures on the results of history
lengths, our workflow utilizes the fractal theory to generate natural fractures in 3D space.
Darcel et al. (2003a) proposed that natural fracture networks in nature, as observed in
various outcrops, follow two unique geometrical patterns: fractal distributions of fracture
centers and power law distributions of lengths. Hence, the proposed workflow is
beneficial because it is more realistic than random natural fracture generation, which
All three main steps in the AHM workflow were performed including parameter
used another shale gas single well with 1626-day historical production data. To optimize
history matching efficiency, we included an extra step to analyze the fractal network’s
connectivity and remove the non-connected fractures from the network. Similar
constraints and BHP pressure conversion from last chapter is performed here.
The novel workflow is applied to a real shale gas well in Sichuan Basin in China.
We performed the case with both hydraulic fractures and natural fractures. The hydraulic
87
fractures are modelled considering the heel bias effect, which means the outer 2 fractures
in each stage have longer half-lengths than the fracture(s) in the middle. This design
actually honors the output characteristics of fracture modelling software, and is deemed
more realistic than simple, planar fractures with uniform half-lengths. For natural
fractures, we inputted the fractal dimension, fractal length dimension and alpha regulator
to be uncertain parameters, meaning that the total number of natural fractures will not be
constant. This field case contains a single well with a horizontal section length of around
1550 m. There is a total of 18 stages of hydraulic fractures with a total cluster number of
54, with a constant clusters per stage of 3. The cluster spacing is approximately 27.78
For this study, since the geology model for this reservoir is lacking, we created a
homogeneous, rectangular reservoir model with a dimension of 1700 m long, 800 m wide
respectively. An example model visualization for the non-uniform hydraulic fractures and
fractal natural fracture scenario are provided in Figure 5.1. The fractal based natural
fractures are fully 3D, meaning that complex dip/azimuth/height distributions are given
for natural fractures. The reservoir has a uniform depth of 2500 m and a residual water
saturation of 0.30, which is equal to matrix water saturation. The initial reservoir pressure
is 44.79 MPa with a temperature of 100 ℃. A long history of 1626 days is available for
history matching. In normal situations, the available history is usually no longer than 1
year. Therefore, the availability of this long production history can truly test the
workflow’s robustness to capture long term production trends. Basic reservoir, fractures,
and Langmuir isotherm gas desorption properties are summarized in Table 5.1. The
relative permeability curves for modeling gas and water flow are assumed to be fixed
88
because they affect the dynamic production less sensitively. We adopted non-linear
relative permeability curves for matrix and straight-line relative permeability curves for
Table 5.1: Summary of basic reservoir, fracture and gas desorption parameters used in
this study.
Figure 5.1: A 3D example model visualizations for the reservoir, hydraulic fractures
(blue) and fractal natural fractures (green) in this study.
89
(a) Matrix
Figure 5.2: Fixed relative permeability used for this study: (a) Matrix; (b) Hydraulic
fractures.
90
5.3 CONNECTIVITY ANALYSIS AND EFFICIENCY OPTIMIZATION
To facilitate the efficiency of the workflow, an extra procedure is built into the
workflow. This procedure detects the connectivity between the fractal generated natural
fractures, hydraulic fractures, and wellbore. Only the natural fractures that are connected
to wellbore or hydraulic fractures are preserved for history matching purpose. The rest of
natural fractures that are isolated are removed from the reservoir model to save
computational time. For example, in Figure 5.3, pink fractures are isolated and will be
removed from the model. This procedure should not affect the results of history
matching/ production forecast since negligible contributions from these fractures are
observed. To validate this statement, we used the field production data of actual gas flow
rate as the well constraint for the history period, and then reduced bottomhole pressure
from the last history day stepwise to a minimum of 1000 kPa until 20 years (Figure 5.4a).
discrepancies can be found for long term cumulative gas and water production.
Figure 5.3: Connectivity check and non-connected natural fractures (pink) removal for
faster history matching.
91
(a) Bottomhole pressure
92
(c) Cumulative water production
be inputted into the history matching workflow. These parameters include matrix
first eight parameters are important parameters related to reservoir and engineering, as
length ratio of the natural fractures, and alpha regulator controls the overall natural
fracture density. Therefore, the fractal dimension really provides information regarding
93
what degree of randomness is associated with the fractal network, while the length
dimension reflects what ratio of faults and smaller micro-fractures/fissures exists. Darcel
et al. (2003a) states that the fractal dimension ranges between 1.5 to 2 for 2D (2.5 to 3 for
3D), but we focused the range more on lower end because the lower fractal dimension is,
the fractures tend to be less randomly distributed and rather exist in regional clusters. In
addition, the lower the length dimension, the larger is the difference between maximum
and minimum natural fracture lengths. The length dimension range here is designed also
to emphasis the combined effects of both faults and fractures on a smaller scale.
Because the permeability and aperture of natural fracture are much insensitive to
dynamic production, we assume constant values of 50 md and 0.01 m for them. Table 5.2
summarizes all the ranges used for the eleven uncertain parameters and the values for the
constant parameters. The ranges for the first eight uncertain parameters are based on the
experience from the last chapter, since the well in this study is situated in the region very
close to the shale gas well, we studied in the last chapter. The ranges for the fractal
parameters are determined so that the total number of natural fractures (both connected
and isolated) are controlled below 5000 to optimize computational speed. For better
94
Uncertain Parameters Distribution Minimum Maximum Unit
Matrix permeability Uniform 0.00001 0.001 md
Fracture permeability decay
Uniform 0.01 0.052 -
factor
Fracture water saturation Uniform 0.5 0.9 -
Matrix porosity Uniform 5% 9% -
Fracture height Uniform 5 20 m
Outer fracture half-length Uniform 50 180 m
Inner fracture half-length Uniform 20 80 m
Fracture conductivity Uniform 10 500 md-m
3D fractal dimension (𝐷𝑓 ) Uniform 2.6 2.8 -
3D length dimension (𝐷𝑙 ) Uniform 2.3 2.6 -
3D alpha regulator (𝛼) Uniform 4 5 -
Exponent of water relative
Deterministic 1.95 -
permeability
Exponent of gas relative
Deterministic 1.8 -
permeability
Endpoint of water relative
Deterministic 0.842 -
permeability
Table 5.2: Summary of uncertain parameters, their ranges, fixed parameters, and their
deterministic values used for automatic history matching.
96
(d) Df=2.7, Dl=2.6, α=4.5
97
(f) Df=2.7, Dl=2.45, α=5
converting the WHP to BHP, as shown in Figure 5.6. Next, we selected a total of 20
proxy points for the proxy model to make predictions on bottomhole pressure and water
gas ratio (WGR), and the location assignments are shown in Figure. 5.7. In this study,
WGR was used as the response variable instead of water flow rate because it is less
noisy. The selected proxy locations included 10 proxies for bottomhole pressure at day
30, 90, 226, 374, 500, 745, 1000, 1200, 1400, and 1626 (Figure. 5.7a), 10 proxies for
water gas ratio at day 20, 140, 226, 374, 500, 743, 983, 1200, 1400, and 1623 (Figure.
5.7b). Although the historical data length is much longer than the one in the last chapter,
98
we did not increase the proxy points because this would jeopardize computational
99
(b) Water gas ratio
Figure 5.7: Locations of proxy prediction points for calculating values of response
variables’ objective functions during history matching iterations: (a)
Flowing BHP; (b) Water gas ratio.
Latin Hypercube samples, run the simulations for these samples and obtained the
corresponding results, as shown in Figure 5.8. Similarly, gas flow rate is used as
constraint for all simulation. Most initial samples can constrain the gas flow rate very
well, as can be seen from the Figure 5.8a. For water flow rate (Figure 5.8c) and flowing
BHP (Figure 5.8f), the current ranges covers the production history, meaning there are
samples both over-estimating and under-estimating the history data. Even though the
initial samples tend to overestimate flowing BHP, our algorithm should be able to find
100
(a) Gas flow rate
101
(c) Water flow rate
102
(e) Water gas ratio
Figure 5.8: Comparison between simulation results from initial iteration (Latin
Hypercube sampling process) and actual production history: (a) Gas flow
rate; (b) Cumulative gas production; (c) Water flow rate; (d) Cumulative
water production; (e) Water gas ratio; (f) Flowing BHP.
103
The automatic history matching process then continued iteratively. For each
iteration, we used single chain of MCMC and a proposal size of 30000. The other
parameters for MCMC are kept same as the last chapter. Similarly, differing weights for
each proxy location when calculating objective functions are defined. The first three
proxies of BHP and WGR at have weights of 1, while rest of the proxies have weights of
2. This is because there are huge uncertainties associated with water flow rate data at
early period, due to water flow back. There is also a rapid drop in BHP, meaning the well
Likewise, a total of 11 iterations, or 300 simulations are designed for this study.
The entire history matching workflow took approximately 13 hours to finish. Next,
history matching solutions are filtered out based on objective function threshold values of
35 and 70 for BHP and WGR. Each criterion is relaxed compared to the last chapter,
because more data uncertainties are introduced as production history increases. A total of
52 solutions are filtered out. The number of solutions versus iteration run is given in
Figure 5.9. This figure implies that the quality of proxy model improves progressively
with the iteration number, and the last 2 iterations can be removed to optimize efficiency.
The resulting production curves of response variables for these 52 history matching
solutions can be visualized in Figure 5.10. The figure shows that despite the early
period’s data fluctuations, our workflow is capable to match the late period production
data excellently even though the initial iteration overestimates actual BHP. Figure 5.11
shows the simulation results for the best match model, which proves that our proposed
workflow is robust in history matching production data that spans much longer time
period.
104
Figure 5.9: Number of history matching solutions versus iteration number.
105
(b) Cumulative gas production
106
(d) Cumulative water production
107
(f) Flowing bottomhole pressure
109
(d) Cumulative water production
110
(f) Flowing bottomhole pressure
Figure 5.11: Comparison between best match’s simulation results and actual production
history: (a) Gas flow rate; (b) Cumulative gas production; (c) Water flow
rate; (d) Cumulative water production; (e) Water gas ratio; (f) Flowing BHP.
The following information contains the best match model parameters’ values.
Matrix permeability is 36.2 nano-darcy (nd), fracture permeability decay factor is about
0.0307 and the decline curve can be visualized in Figures 5.12 and 5.14. These results the
equivalent matrix permeability is low and thus the cluster spacing could be further
reduced to maximize recovery. The fracture permeability multiplier reaches 0.25 when
pressure drops to atmospheric pressure, meaning the fracture closure phenomenon is less
severe than the study in Chapter 4. Because the depth of the reservoir in this study is
shallower, and assuming similar fracture job design is applied for both study, the
compaction should be less severe for this study and thus less conductivity loss occurs,
which is corroborated by our results. Matrix porosity is 6.8 %. Compared to the previous
111
chapter, in which the model has a greater depth and a lower average porosity between 4-
5%, the obtained porosity is again backed up by physics. The fracture height, outer
(visualized in Figure 5.13). The fracture height is able to grow higher due to shallower
depth, but half-lengths tend to be limited possibly because of the effect of natural
fractures 256 md-m, implying a less severe water flow back problem. Finally, the 3D
fractal dimension, 3d length dimension and 3d alpha regulator characterized are 2.65,
2.58, and 4.29. These results imply that the natural fractures tend to be less randomly
distributed, but the length discrepancy between different scales of natural fractures is low.
There is also a relatively lower density of natural fractures. Parallel coordinates plot is
shown in Figure 5.15, and we see that matrix permeability, fracture height, outer/inner
112
(c) Fracture water saturation (d) Matrix porosity
113
(i) 3D fractal dimension (j) 3D length dimension
Figure 5.12: Comparison between prior and posterior distributions of the studied
uncertain parameters based on 52 history matching solutions: (a) Matrix
permeability; (b) Fracture permeability decay factor; (c) Fracture water
saturation; (d) Matrix porosity; (e) Fracture height; (f) Outer fracture half-
length; (g) Inner fracture half-length; (h) Hydraulic fracture conductivity; (i)
3D fractal dimension; (j) 3D length dimension; (k) 3D alpha regulator.
114
(a) Overview
115
(c) Fracture height
Figure 5.13: Characterized fracture geometry from AHM workflow: (a) Overview; (b)
Non-uniform fracture half-lengths; (c) Fracture height.
In order to gauge the robustness of the workflow to catch long term production
trend, we plotted the values of the BHP and WGR at the early period (Figure 5.16a, b),
mid-period (Figure 5.16c, d), and late period (Figure 5.16e, f). In addition, the objective
function value for water flow rate and BHP, and global relative error are also plotted, as
For WGR at 1623 days, the values for water flow rate is almost 0, making the values for
proxy/simulation very low. From Figure 5.16, the proxy quality, especially for BHP
117
(a) BHP at 30 days (b) WGR at 20 days
Figure 5.16: Comparison between response parameters’ values and relative errors of
proxy model and simulation results for iteration 2, 4, 6, 8 and 10: (a) BHP at
30 days; (b) WGR at 20 days; (c) BHP at 745 days; (d) WGR at 743 days;
(e) BHP at 1626 days; (f) WGR at 1623 days; (g) BHP relative error; (h)
WGR relative error; (i) Global relative error.
error versus simulation cases (In Figure 5.17, red circles are AHM solutions). This figure
proves our previous statements about sensitivities of the various uncertain parameters.
119
For global error (Figure 5.17l), the average values for history matching solutions are
higher than the one obtained in last chapter, because of the longer data history.
120
(e) Fracture height (f) Outer fracture half-length
121
(k) 3D alpha regulator (l) Global error
Figure 5.17: Scatter plots of values of different uncertain parameters and global error
versus simulation index: (a) Matrix permeability; (b) Fracture permeability
decay factor; (c) Fracture water saturation; (d) Matrix porosity; (e) Fracture
height; (f) Outer fracture half-length; (g) Inner fracture half-length; (h)
Hydraulic fracture conductivity; (i) 3D fractal dimension; (j) 3D length
dimension; (k) 3D alpha regulator; (l) Global error.
history matching solutions. The BHP drawdown schedule after the history matching
period is visualized in Figure 5.18a. To maintain the BHP drawdown rate from the late
Because the history data spans 4 years, the reservoir has lost a large amount of energy
and there is no more flow rate peaks after history period. For 20-year gas estimated
ultimate recovery (EUR), the P10, P50, P90 and best match values are 112.0, 121.2,
127.7, and 115.1 million cubic meters. For 20-year water EUR, the P10, P50, P90 and
best match values are 5.24, 6.50, 7.91, and 7.06 thousand cubic meters. Compared to last
122
chapter, the ranges for the probabilistic production forecasting in this study, without
porosity heterogeneity and inclusion of natural fractures, span more widely, further
validating that the characterization of natural fracture distributions and geology model is
123
(c) Cumulative gas production
124
(e) Cumulative water production
Figure 5.18: Production forecasting of 52 history matching solutions: (a) Flowing BHP;
(b) Gas flow rate; (c) Cumulative gas production; (d) Water flow rate; (e)
Cumulative water production.
At last, the pressure in matrix and complex fracture network are visualized
separately for both the short term and long term. In Figure 5.19, we plotted the matrix
pressure at 365 days, end of history period (1626 days), 10 years and 20 years. From this
figure, we can see that the drainage area gradually increases as time progresses. However,
this area is only confined to the region near wellbore, implying that the characterized
fractal network has very deficient connectivity. Figure 20 contains the pressure in
fractures at the same time. This figure again shows that the natural fracture network in
125
(a) 365 days
(c) 10 years
Figure 5.19: Pressure distribution of matrix at different production time: (a) 365 days; (b)
End of history (1626 days); (c) 10 years; (d) 20 years.
127
(c) 10 years
(d) 20 years
5.8 CONCLUSIONS
We applied the AHM workflow to another shale gas well in southwest china and
investigated the effects of fractal based natural fracture networks on history matching
results. It is found out that the presence of natural fractures reduced possibility of
existence of longer and taller fractures. In other words, our workflow is robust enough to
substantiate the fact that the existence of natural fractures has profound impacts on the
stress shadowing effect of hydraulic fractures propagation. From another perspective, the
history matching results suggest that the natural fractures in this shale reservoir,
according to fractal theory, are less randomly distributed, have a smaller number of faults
and have a rather low density. Compared to the study performed in the last chapter and
128
field experience from geologists, a much more developed natural fracture network is
usually seen in the deeper shale gas reservoirs, and our results indeed agree with the
statement. To close the loop, more subsurface data (such as good-quality microseismic
data, well image log, etc.) should be introduced to further validate the findings if
available.
129
Chapter 6: Summary, Conclusions, and Recommendations for Future
Work
This chapter aims to summarize all the conclusions and findings from this study,
and to deliver some of the potential suggestions for future improvements and extensions
shale reservoirs with natural fractures characterized from both a geology model and based
simple structural grids and complex corner point grids. In addition, the coupling of
XGBoost with MCMC algorithm enables us to optimize the workflow efficiency and
obtain accurate matching results. These results are then applied to probabilistically
gridding system and geologist-characterized natural fracture network into the AHM
process for a shale gas well. We introduced the complex characteristics of the reservoir,
discussed the difficulties of inputting fractures in this model, and explained how this
problem is solved. Then, we employed the AHM workflow to this reservoir model and
connectivity analysis are lastly studied. We found out that the inclusion of porosity
heterogeneity reduced the uncertainties associated with the probabilistic EUR forecasts,
and that the geologist-characterized natural fracture network has very limited
connectivity with both the wellbore and hydraulic fractures, creating a bottleneck for the
results of automatic history matching. Again, each of the module in our proposed AHM
workflow is implemented and acceptable results are achieved. In this chapter, we found
out that the presence of natural fracture does affect the growth of hydraulic fractures, and
based on the characterizations of fractal parameters, we know that the natural fractures in
this reservoir are less randomly distributed and more fractal-distributed. Furthermore,
there is a small number of large-scale fractures and a low natural fracture density.
This study is very beneficial, since the integration of the geology with engineering
any shale reservoirs. By using a physical geology model, a realistic natural fracture
To further improve/extend some aspects of this study and make the numerical
optimization process more realistic, the following bullet points are enumerated:
sampling methods, potential suitable proxy models and their results comparison
would bring a more wholesome picture for the iterative optimization module.
131
• Microseismic data could be brought into the AHM process. For example, plane
fittings to seismic events and perform automatic cut off of the originally fitted
algorithm could not control the total number of natural fractures. Moreover,
unconventional reservoirs.
data could be utilized in the AHM workflow to further reduce the uncertainties
could be integrated with our AHM workflow, and automatic cutoff features can
132
Glossary
ACRONYMS
AI = Artificial intelligence
MD = Measured depth
MH = Metropolis-Hasting
NN = Neural network
POR = Porosity
NOMENCLATURE
𝑀𝐷ℎ = Measured depth of the adjacent well trajectory point toward heel side
𝑀𝐷𝑝 = Measured depth of an arbitrary perforation point
134
𝑀𝐷𝑡 = Measured depth of the adjacent well trajectory point toward toe side
𝑃𝑚 = Matrix pressure
𝑟 = Interpolation ratio
135
𝑦𝑝 = Y coordinate of an arbitrary perforation point
𝑦𝑤ℎ = Y coordinate of the adjacent well trajectory point toward heel side
𝑦𝑤𝑡 = Y coordinate of the adjacent well trajectory point toward toe side
𝛼 = 2D alpha regulator
Γ = Gamma function
𝜙 = Matrix porosity
𝜌 = Rock density
136
References
Abdle Moneim, S. S., Rabee, R., Shehata, A. M., and Aly, A. M., 2012. Modeling
Hydraulic Fractures in Finite Difference Simulators Using Amalgam LGR (Local
Grid Refinement). Paper SPE 148864, presented at the North Africa Technical
Conference and Exhibition, Cairo, Egypt, 20-22 February.
Arroyo Negrete, E., Rodiguez, J., Goryachev, S., Belova, N., Al Blooshi, A., and Basioni,
M., 2018. Automatic History Matching Theory, Implementation, and Field
Applications. Paper SPe 193018, presented at the Abu Dhabi International
Petroleum Exhibition & Conference, Abu Dhabi, UAE, 12-15 November.
Al-Shaalan, T. M., Fung, L. S. K., and Dogru, A. H., 2003. A Scalable Massively Parallel
Dual-Porosity Dual-Permeability Simulator for Fractured Reservoirs with Super-
K Permeability. Paper SPE 84371, presented at the SPE Annual Technical
Conference and Exhibition, Denver, Colorado, 5-8 October.
Bentéjac, C., Csörgő, A., and Martínez-Muñoz, G., 2019. A Comparative Analysis of
XGBoost. doi: 10.1007/s10462-020-09896-5.
Bosma, S. B. M., Hajibeygi, H., Tene, M., and Tchelepi, H. A., 2017. Multiscale Finite
Volume Method for Discrete Fracture Modeling with Unstructured Grids. Paper
SPE 182654, presented at the SPE Reservoir Simulation Conference,
Montgomery, Texas, 20-22 February.
Cavalcante Filho, J.S.A., Shakiba, M., Moinfar, A., and Sepehrnoori, K., 2015.
Implementation of a Preprocessor for Embedded Discrete Fracture Modeling in an
IMPEC Compositional Reservoir Simulator. Paper SPE 173289, presented at the
SPE Reservoir Simulation Symposium, Houston, Texas, 23-25 February.
Cipolla, C. L., Lolon, E., Erdle, J. C., and Rubin, B., 2009. Reservoir Modeling in
ShaleGas Reservoirs. Paper SPE 125530, presented at the SPE Eastern Regional
Meeting, Charleston, West Virginia, 23-25 September.
Cipolla, C. L., Weng, X., Mack, M. G., Ganguly, U., Gu, H., Kresse, O., and Cohen, C.
E., 2011. Integrating Microseismic Mapping and Complex Fracture Modeling to
Characterize Hydraulic Fracture Complexity. Paper SPE 140185, presented at the
SPE Hydraulic Fracturing Technology and Conference, the Woodlands, Texas,
24-26 January.
CMG-IMEX., 2017. IMEX User’s Guide, Computer Modeling Group Ltd.
Darcel, C., Bour, O., Davy, P., and de Dreuzy, J.R., 2003a. Connectivity Properties of
Two-Dimensional Fracture Networks with Stochastic Fractal Correlation. Water
Resources Research 39 (10): SBH1-1 – SBH1-13.
137
Darcel, C., Bour, O., Davy, P., 2003b. Stereological Analysis of Fractal Fracture
Networks. Journal of Geophysical Research 108 (9): ETG13-1 – ETG13-14.
Ding, X., 2019. Using Unstructured Grids for Modeling Complex Discrete Fracture
Network in Unconventional Reservoir Simulation. Paper SPE 195051, presented
at the SPE Middle East Oil and Gas Show and Conference, Manama, Bahrain, 18-
21 March.
Ding, Y., & Lemonnier, P., 1995. Use of Corner Point Geometry in Reservoir
Simulation. Paper SPE 29933, presented at the International Meeting on
Petroleum Engineering, Beijing, China, 14-17 November.
Du, S., Liang, B., and Yuanbo, L., 2017. Field Study: Embedded Discrete Fracture
Modeling with Artificial Intelligence in Permian Basin for Shale Formation.
Paper SPE 187202, presented at the SPE Annual Technical Conference and
Exhibition, San Antonio, Texas, 9-11 October.
Fumagalli, A., Pasquale, L., Zonca, S., and Micheletti S., 2016. An Upscaling Procedure
for Fractured Reservoirs with Embedded Grids. Water Resources Research (52)
8: 6506-6525.
Gao G., Jiang H., Hagen P., Vink J., and Wells T., 2017. A Gauss-Newton Trust-Region
Solver for Large-Scale History-Matching Problems. SPE Journal 22 (6): 1999-
2011.
Ghosh, S., 1998. Curvilinear Local Grid Refinement. Paper SPE 50633, presented at the
European Petroleum Conference, The Hague, Netherlands, 20-22 October.
Glover, K., Naser, G., and Mohammadi, H., 2015. Creep Deformation of Fracture
Surfaces Analysis in a Hydraulically Fracture Reservoir Using the Finite Element
Method. Journal of Petroleum and Gas Engineering 6 (6): 62-73.
Goodwin, N., 2015. Bridging the Gap Between Deterministic and Probabilistic
Uncertainty Quantification Using Advanced Proxy Based Methods. Paper SPE
173301, presented at the SPE Reservoir Simulation Symposium, Houston, Texas,
23-25 February.
Hassanpour, R. M., Leuangthong, O., and Deutsch, C. V., 2008. Calculation of
Permeability Tensors for Unstructured Grid Blocks. Paper PETSOC-2008-187,
presented at the Canadian International Petroleum Conference, Calgary, Alberta,
Canada, 17-19 June.
Kabir, C. S., Chien, M. C. H., and Landa, J. L., 2003. Experiences With Automated
History Matching. Paper SPE 79670, presented at the SPE Reservoir Simulation
Symposium, Houston, Texas, 3-5 February.
Kim, T. H., and Schechter, D. S., 2007. Estimation of Fracture Porosity of Naturally
Fractured Reservoirs with No Matrix Porosity Using Fractal Discrete Fracture
138
Networks. Paper SPE 110720, presented at the SPE Annual Technical Conference
and Exhibition, Anaheim, California, 11-14 November.
Kuchuk, F. J., and Biryukov, D., 2012. Transient Pressure Test Interpretation from
Continuously and Discretely Fractured Reservoirs. Paper SPE 158096, presented
at the SPE Annual Technical Conference and Exhibition, San Antonio, Texas, 8-
10 October.
Kumar, A., Close, J. C., Picone, M. M., Agustsson, H., Card, C. C., and Kjosavik, A.,
2013. A New and Practical Workflow for Large Multi-pad SAGD Simulation- A
Corner Oil Sands Case Study. Paper SPE 165511, presented at the SPE Heavy Oil
Conference-Canada, Calgary, Alberta, Canada, 11-13 June.
Lu, J., Zhu, T., and Tiab, D., 2009. Pressure Behavior of Horizontal Wells in Dual-
Porosity, Dual-Permeability Naturally Fractured Reservoirs. Paper SPE 120103,
presented at the SPE Middle East Oil and Gas Show and Conference, Manama,
Bahrain, 15-18 March.
Main, I. G., Meredith, P. G., Sammonds, P. R., and Jones, C., 1990. Influence of Fractal
Flaw Distributions on Rock Deformation in the Brittle Field. Geological Society
Special Publications 54: 71-79.
Malik, S., Harode, R., and Kunwar, A., 2020. XGBoost: A Deep Dive into Boosting
(technical report). doi: 10.13140/RG.2.2.15243.64803.
McClure, M., Picone, M., Fowler, G., Ratcliff, D., Kang, C., Medam, S., and Frantz, J.,
2020. Nuances and Frequently Asked Questions in Field-Scale Hydraulic Fracture
Modeling. Paper SPE 199726, presented at the SPE Hydraulic Fracturing
Technology and Conference, the Woodlands, Texas, 4-6 February.
Miao, J., Yu, W., Xia, Z., Zhao, W., Xu, Y., and Sepehrnoori, K., 2018. An Easy and Fast
EDFM Method for Production Simulation in Shale Reservoirs with Complex
Fracture Geometry. Paper ARMA-DFNE-18-0926, presented at the 2nd
International Discrete Fracture Network Engineering Conference, Seattle,
Washington, 20-22 June.
Moinfar, A., Varavei, A., Sepehrnoori, K., and Johns R.T., 2014. Development of an
Efficient Embedded Discrete Fracture Model for 3D Compositional Reservoir
Simulation in Fractured Reservoirs. SPE Journal 19 (2): 289-303.
Odling, N. E., 1997. Scaling and Connectivity of Joint Systems in Sandstones from
Western Norway. Journal of Structural Geology 19 (10): 1257-1271.
Okubo, P. G., and Aki, K., 1987. Fractal Geometry in the San Andreas Fault System.
Journal of Geophysical Research 92 (B1): 345-355.
Ouillon, G., and Sornette, D., 1996. Unbiased Multifractal Analysis: Application to Fault
Patterns. Geophysical Research Letters 23 (23): 3409-3412.
139
Panfili, P., and Cominelli, A., 2014. Simulation of Miscible Gas Injection in a Fractured
Carbonate Reservoir Using an Embedded Discrete Fracture Model. Paper SPE
171830, presented at the Abu Dhabi International Petroleum Exhibition and
Conference, Abu Dhabi, UAE, 10-13 November.
Panfili, P., Colin, R., Cominelli, A., Giamminonni, D., and Guerra, L., 2015. Efficient
and Effective Field Scale Simulation of Hydraulic Fractured Wells: Methodology
and Application. Paper SPE 175542, presented at the SPE Reservoir
Characterization and Simulation Conference and Exhibition, Abu Dhabi, UAE,
14-16 September.
Peaceman, D. W., 1983. Interpretation of Well-Block Pressures in Numerical Reservoir
Simulation with Nonsquare Grid Blocks and Anisotropic Permeability. SPE
Journal 23 (03): 531-543.
Ponting, D. K., 1989. Corner Point Geometry in Reservoir Simulation. Presented at the
1st European Conference on the Mathematics of Oil Recovery, Cambridge, UK,
14-16 July.
Rao, X., Cheng, L., Cao, R., Jia, P., Dong, P., and Du, X., 2019. A Modified Embedded
Discrete Fracture Model to Improve the Simulation Accuracy During Early-Time
Production of Multi-Stage Fractured Horizontal Well. Paper SPE 196263,
presented at the SPE/IATMI Asia Pacific Oil & Gas Conference and Exhibition,
Bali, Indonesia, 29-31 October.
Rubin, B., 2010. Accurate Simulation of Non Darcy Flow in Stimulated Fractured Shale
Reservoirs. Society of Petroleum Engineers. Paper SPE 132093, presented at the
SPE Western Regional Meeting, Anaheim, California, 27-29 May.
Rwechungura, R. W., Dadashpour, M., and Kleppe, J., 2011. Advanced History Matching
Techniques Reviewed. Paper 142497, presented at the SPE Middle East Oil and
Gas Show and Conference, Manama, Bahrain, 25-28 September.
Sejdinovic, D., 2015. Model complexity and generalization. Lecture, Oxford University,
Oxford, UK.
Sepehrnoori, K., Xu, Y., and Yu, W., 2020. Embedded Discrete Fracture Modeling and
Application in Reservoir Simulation. 1st Ed.; Publisher: Elsevier, Cambridge,
USA. ISBN: 978-0-12-819688-5.
Shah, S., Møyner, O., Tene, M., Lie, K.A., and Hajibeygi, H., 2016. The Multiscale
Restriction Smoothed Basis Method for Fractured Porous Media (F-MsRSB).
Journal of Computational Physics 318 (C): 36-57.
Shakiba, M., and Sepehrnoori, K., 2015. Using Embedded Discrete Fracture Model
(EDFM) and Microseismic Monitoring Data to Characterize the Complex
Hydraulic Fracture Networks. Paper SPE 175142, presented at the SPE Annual
Technical Conference and Exhibition, Houston, Texas, 28-30 September.
140
Shams, M., 2016. Reservoir Simulation Assisted History Matching: From Theory to
Design. Paper SPE 182808, presented at the SPE Kingdom of Saudi Arabia
Annual Technical Symposium and Exhibition, Dammam, Saudi Arabia, 25-28
April.
Sun, J., Schechter, D., and Huang, C.-K., 2016. Grid-Sensitivity Analysis and
Comparison Between Unstructured Perpendicular Bisector and Structured
Tartan/Local-Grid-Refinement Grids for Hydraulically Fractured Horizontal
Wells in Eagle Ford Formation with Complicated Natural Fractures. SPE Journal
21 (6): 2260-2275.
Tavassoli, Z., Carter, J. N., and King, P. R., 2004. Errors in History Matching. SPE
Journal 9 (3): 352-361.
Tijink, P., & Cottier, J., 2019. The Description and Quantification of the Truncation
Errors Produced by Local-Grid Refinement in Reservoir Simulation. SPE
Reservoir Evaluation & Engineering 22 (2): 660-672
Tripoppoom, S., Yu, W., Sepehrnoori, K., and Miao, J., 2019. Application of Assisted
History Matching Workflow to Shale Gas Well Using EDFM and Neural
Network-Markov Chain Monte Carlo Algorithm. Paper URTEC-2019-659,
presented at the SPE/AAPG/SEG Unconventional Resources Technology
Conference, Denver, Colorado, 22-24 July.
Tripoppoom, S., 2019. Assisted History Matching Workflow for Unconventional
Reservoirs. Master thesis, The University of Texas at Austin, Austin, Texas (May
2019).
Vazquez, O., Young, C., Demyanov, V., Arnold, D., Fisher, A., MacMillan, A., and
Christie, M., 2015. Produced-Water-Chemistry History Matching in the Janice
Field. SPE Reservoir Evaluation & Engineering 18 (4): 564-576.
Wang, Y., 2015. A Hybrid Dual-Continuum Discrete Fracture Modeling Approach for
Numerical Simulation of Production from Unconventional Plays. Paper SPE
178749, presented at the SPE Annual Technical Conference and Exhibition,
Houston, Texas, 28-30 September.
Wantawin, M., Yu, W., and Sepehrnoori, K., 2017. An Iterative Response-Surface
Methodology by Use of High-Degree-Polynomial Proxy Models for Integrated
History Matching and Probabilistic Forecasting Applied to Shale-Gas Reservoirs.
SPE Journal 22 (6): 2012-2031.
Wheaton, B., Haustveit, K., Deeg, W., Miskimins, J., and Barree, R., 2016. A Case Study
of Completion Effectiveness in the Eagle Ford Shale Using DAS/DTS
Observations and Hydraulic Fracture Modeling. Paper SPE 179149, presented at
the SPE Hydraulic Fracturing Technology Conference and Exhibition, The
Woodlands, Texas, 9-11 February.
141
Wu, K., and Olson, J.E., 2016. Numerical Investigation of Complex Fracture Networks in
Naturally Fractured Reservoirs. SPE Production & Operations 31 (4): 300-309.
Xie, J., Yang, C., Gupta, N., King, M. J., and Datta-Gupta, A., 2015. Integration of Shale
Gas-Production Data and Microseismic for Fracture and Reservoir Properties
With the Fast Marching Method. SPE Journal 20 (2): 347-359.
Xue, X., Yang, C., Onishi, T., King, M. J., and Datta-Gupta, A., 2019. Modeling
Hydraulically Fractured Shale Wells Using the Fast Marching Method with Local
Grid Refinements LGRs and Embedded Discrete Fracture Model EDFM. Paper
SPE 193822, presented at the SPE Reservoir Simulation Conference, Galveston,
Texas, 10-11 April.
Xu, Y., Cavalcante Filho, J. S. A., Yu, W., and Sepehrnoori, K., 2017. Discrete-Fracture
Modeling of Complex Hydraulic-Fracture Geometries in Reservoir Simulators.
SPE Reservoir Evaluation & Engineering 20 (2): 403-422.
Xu, Y., and Sepehrnoori, K., 2019. Development of an Embedded Discrete Fracture
Model for Field-Scale Reservoir Simulation with Complex Corner-Point Grids.
SPE Journal 24 (4): 1552-1575.
Yang, C., King, M. J., and Datta-Gupta, A., 2017. Rapid Simulation of Naturally
Fractured Unconventional Reservoirs with Unstructured Grids Using the Fast
Marching Method. Paper SPE 182612, presented at the SPE Reservoir Simulation
Conference, Montgomery, Texas, 20-22 February.
Yang, D., Xue, X., and Chen, J., 2018. High Resolution Hydraulic Fracture Network
Modeling Using Flexible Dual Porosity Dual Permeability Framework. Paper SPE
190096, presented at the SPE Western Regional Meeting, Garden Grove,
California, 22-26 April.
Yu, W., Tripoppoom, S., Sepehrnoori, K., and Miao, J., 2018a. An Automatic History-
Matching Workflow for Unconventional Reservoirs Coupling MCMC and Non-
Intrusive EDFM Methods. Paper SPE-191473-MS, presented at the SPE Annual
Technical Conference and Exhibition, 24-26 September, Dallas, Texas.
Yu, W., Miao, J., and Sepehrnoori, K., 2018b. A Revolutionary EDFM Method for
Modeling Dynamic Behaviors of Complex Fractures in Naturally Fractured
Reservoirs. Paper ARMA-DFNE-18-0928, presented at the 2nd International
Discrete Fracture Network Engineering Conference, Seattle, Washington, 20-22
June.
Yu, W., and Sepehrnoori, K., 2018. Shale Gas and Tight Oil Reservoir Simulation, 1st
Ed.; Publisher: Elsevier, Cambridge, USA. ISBN: 978-0-12-813868-7.
Yu, W., Hu X., Liu, M. and Wang, W., 2019. Investigation of the Effect of Natural
Fractures on Multiple Shale-Gas Well Performance Using Non-Intrusive EDFM
Technology. Energies 12 (5): 932.
142