SPE-195928-MS

Integrated Characterization of the Fracture Network in Fractured Shale Gas

Reservoirs—Stochastic Fracture Modeling, Simulation and Assisted History
Matching

Yonghui Wu, China University of Petroleum-Beijing, and Texas A&M University; Linsong Cheng, China University
of Petroleum-Beijing; John E. Killough, Texas A&M University; Shijun Huang, China University of Petroleum-
Beijing; Sidong Fang, Sinopec Exploration and Production Research Institute; Pin Jia, Renyi Cao, and Yongchao
Xue, China University of Petroleum-Beijing

Copyright 2019, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Annual Technical Conference and Exhibition held in Calgary, Alberta, Canada, 30 Sep - 2 October 2019.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
The large uncertainty in fracture characterization for shale gas reservoirs seriously affects the confidence
in making forecasts, fracturing design, and taking recovery enhancement measures. This paper presents a
workflow to characterize the complex fracture networks (CFNs) and reduce the uncertainty by integrating
stochastic CFNs modeling constrained by core and microseismic data, reservoir simulation using a novel
edge-based Green element method (eGEM), and assisted history matching based on Ensemble Kalman
Filter (EnKF).
In this paper, the geometry of CFNs is generated stochastically constrained by the measurements of
hydraulic fracturing treatment, core, and microseismic data. A stochastic parameterization model is used
to generate an ensemble of initial realizations of the stress-dependent fracture conductivities of CFNs. To
make the eGEM practicable for reservoir simulation, a steady-state fundamental solution is applied to the
integral equation, and the technique of local grid refinement (LGR) is applied to refine the domain grids
near the fractures. Finally, assisted-history-matching based on EnKF is implemented to calibrate the DFN
models and further quantify the uncertainties in the fracture characterization.
The proposed technique is tested using a multi-stage fractured horizontal well from a shale gas field.
After analyzing the history matching results, the proposed integrated workflow is shown to be efficient in
characterizing fracture networks and reducing the uncertainties. The advantages are exhibited in several
aspects. First, the eGEM-based Discrete-Fracture Model (DFM) is shown to be quite efficient in assisted
history matching of large field applications because of eGEM’s high precision with coarse grids. This
enables simulations of CFNs without upscaling the fractures using continuum approaches. In addition, CFNs
geometry can be generated with the constraints of core and microseismic data, and a primary conductivity of
CFNs can be generated using the hydraulic fracturing treatment data. Moreover, the uncertainties for CFNs
characterization and EUR predictions can be further reduced with the application of EnKF in assimilating
the production data.
2 SPE-195928-MS

This paper provides an efficient integrated workflow to characterize the fracture networks in fractured
unconventional reservoirs. This workflow, which incorporated several efficient techniques including
fracture network modeling, simulation and calibration, can be readily used in field applications. In
addition, various data sources could be assimilated in this workflow to reduce the uncertainty in fracture
characterization, including hydraulic fracturing treatment, core, microseismic and production data.

Introduction
Hydraulic fracturing has been the major technology for the commercial development of unconventional
resources. Because unconventional reservoirs are rich in natural fractures, which can be reactivated after
hydraulic fracturing, and CFNs will be generated around the horizontal wellbore. The production of
unconventional wells is dependent on the complexity of the fracture networks to a great extent. Therefore,
the simulation and characterization techniques of CFNs in unconventional reservoirs have attracted the
word’s interest.
There are mainly 3 types of models, including analytical, semi-analytical and numerical models, for the
simulation of the production performance of fractured wells with CFNs.
Analytical models are widely used for simulating fracture networks, which are handled using the concept
of stimulated reservoir volume (SRV). Multiple porosity models are often used for the SRV, and the solution
of the model are obtained analytically based on the linear flow assumption in each subdomain (Brown et
al. 2009; Bello and Wattenbarger 2010; Al-Ahmadi and Wattenbarger 2011; Stalgorova and Mattar 2013;
Fuentes-Cruz and Valko 2015; Wu et al. 2019a). This model is quite practical and widely used in production
data analysis. However, many complexities cannot be considered using analytical models, such as handling
CFNs explicitly, and nonlinearities and heterogeneities in the model, etc.
Many semi-analytical models for simulation of fractures are based on Green’s functions and Boundary
Element Method (BEM) in petroleum engineering (Gringarten and Ramey 1972; Cinco-Ley et al. 1978;
Larsen and Hegre 1991; Kikani and Horne 1992, 1993; Ozkan and Raghavan 1991; Chen and Raghavan
1997). With the development of unconventional resources, many semi-analytical models are proposed to
simulate CFNs, which can be represented explicitly in the semi-analytical models. Biryukov and Kuchuk
(2012) used an analytical element free method for modeling the transient behavior of CFNs. Zhou et al.
(2014) are the first to use Green’s functions in modeling the transient behavior of CFNs in unconventional
reservoirs. After that, this method is used by many researchers to handle CFNs in different models (Jia et
al. 2016; Yu et al. 2016; Zhao et al. 2016; Chen et al. 2016; Yang et al. 2017; Cheng et al. 2017; Bao et
al. 2017a). To handle arbitrary boundaries and simple heterogeneities, BEM is used by several researchers
to simulate CFNs in unconventional reservoirs (Jia et al. 2017; Cao et al. 2017; Idorenyin and Shirif 2018;
Wu et al. 2019b). The way CFNs handled in BEM is quite like Green’s functions, and CFNs are discretized
into many elements and each element is treated as a source or sink. However, these models are limited to
linear and simple heterogeneous problems. In addition, the computation is rather expensive when numbers
fractures are handled.
Numerical methods are widely used in reservoir stimulation because many complexities can be
incorporated into numerical models. There are mainly two numerical models to handle fluid flow in
fractures, including continuum approaches and explicit fracture models. The continuum approaches handle
fractures using classic dual-porosity/dual-permeability (DP/DP) model (Warren and Root 1963; Kazemi,
H. et al. 1976), and multi-porosity model, MINC (Pruess and Narasimhan 1985; Wu and Pruess 1988;
Wu et al. 2011). However, this approach can only handle well-connected fractures, and complex fracture
geometry cannot be captured. Recently, it has been popular to use Discrete-Fracture Model (DFM) and
Embedded-Discrete-Fracture Model (EDFM) to simulate CFNs explicitly. DFM uses unstructured grid
systems to represent CFNs more realistically (Kim and Deo 2000; Karimi-Fard et al. 2004; Hoteit and
Firoozabadi 2005; Geiger et al. 2009; Cipolla 2010; Mayerhofer et al. 2010). Owing to the severe contrast
SPE-195928-MS 3

in permeability between the matrix and fractures, DFM models typically come with the significant cost of
complex unstructured gridding and computation. EDFM does not require the domain grid to conform to
fracture geometry, so this approach can largely reduce expenses in gridding and computation (Lee et al.
2001; Li and Lee 2008; Moinfar et al. 2014; Jiang et al. 2016; Yan et al. 2016; Ding et al. 2017; Chai et
al. 2018; Tang et al. 2018). Cartesian grids are often used in EDFM, the fracture segments that intersect
grid cells are treated as discrete fracture computational domain, and the concept of the transport index
is introduced to compute fluid exchange between fractures and the connected matrix elements. Another
approach for reservoir simulate is by using GEM, which is first proposed by Taigbenu (1995). The initial
GEM cannot handle discrete fracture, and Wu et al. (2019c) made two important modifications to the
classical GEM to make it practical to simulate CFNs. Except for these approaches for reservoir simulations,
data-driven models are also very interesting techniques for reservoir engineering (Bao and Gildin 2017b;
Tian and Horne 2019; Pan et al. 2019).
The modeling of the fractures in unconventional reservoirs is a major step before reservoir simulation.
Although obtaining the actual geometries and conductivities of fractures is impossible, it is very important
to use the accessible techniques to roughly infer the fractures generated in the formation. This step is not
only important for making predictions but also critical for taking further measures to enhance the recovery
of unconventional oil and gas. At present, several approaches have been reported in the literature to assess
fracture properties. Cipolla and Wright (2000) shew that there are mainly two types of techniques: far-field
and near-wellbore methods. The near-wellbore methods include borehole imaging, acoustic and temperature
logs. Tang et al. (2017) proposed a technique to interpret nuclear magnetic resonance and electromagnetic
logs and assess micro-fracture density. Li and Zhu (2018) investigated the temperature behavior during
multiple stage fracturing in horizontal wells, and Zhang and Zhu (2019) proposed an inversion approach
for assessing the length and conductivity of fractures. These approaches are mainly based on planar fracture
and orthogonal fracture network assumptions. With the wide application of microseismic tools, many
researchers used the microseismic data to roughly obtain the geometry of CFNs. Cipolla et al. (2011),
and Weng et al. (2011) proposed an approach to use microseismic events locations to calibrate the CFNs
generated by a geomechanical model. Yu et al. (2016) proposed a workflow to combine the multi-event
moment-tensor inversion method to extract the geometry of CFNs. Gamboa et al. (2016), and Sun et al.
(2016) proposed a stochastic approach to generate CFNs constrained by microseismic events and cores
data. In addition, some conceptual ideas of hydraulic-natural fracture interaction are used in the approach,
although hydraulic propagation mechanics is not considered.
History matching is a very important step to calibrate the generated CFNs models. Ensemble Kalman
Filter (EnKF) is a very popular and promising method for history matching because it is independent of the
simulators, and convenient for reservoir monitoring, performance prediction and uncertainty quantification
(Evensen 2003; Zafari and Reynolds 2007; Aanonsen et al. 2009; Fan et al. 2016). EnKF is also widely used
by many researchers for history matching of fracture parameters in unconventional reservoirs. Although
the DP/DP model and the concept of SRV cannot characterize fracture geometry, these models are used by
many researchers in history matching, because they are quite practical and the same model can be used in
history matching. Nejadi et al. (2015) used the DP model to upscale the CFNs in shale gas reservoirs and the
EnKF is applied to calibrate the model. Chang and Zhang (2018) used the concept of SRV to characterize
the CFNs and EnKF is used to match the properties of the SRV. In the model, only the main fractures are
modeled explicitly, and other fractures are characterized using a continuum approach. Wantawin et al. (2017)
proposed a proxy-based history matching approach by using the design of experiments, response surface
methodology, and Markov Chain Monte Carlo Algorithm (MCMC). However, CFNs are not considered in
the model. Elahi and Jafarpour (2018) calibrated the dynamic fracture parameters for both planar fracture
and orthogonal fracture networks models using tracer and production data. Chai et al. (2018) proposed
a novel compartmental EDFM (cEDFM) to address problems with two-phase flow and flow barriers in
reservoir simulations, and EnKFis also used to calibrate two cases with several fractures. Liu and Reynolds
4 SPE-195928-MS

(2019) used a stochastic fractal model to generate CFNs, and EnKF is used to estimate the geometric
configuration and fracture properties. In these works, CFNs are explicitly represented, and EnKF is used
for a synthetic case. In conclusion, the characterization of CFNs in unconventional reservoirs is promising,
but very few studies have considered applying EnKF to field cases with CFNs handled explicitly.
In this paper, we proposed an integrated workflow to characterize the CFNs in multiple fractured shale
gas reservoirs. The stochastic CFNs generation approach proposed by Gamboa et al. (2016) is applied to
obtain the primary fracture configuration. To decrease the uncertainties of the generated CFNs, multiple
data, including microseismic, well log, core, and hydraulic fracturing treatment data, are used as constraints.
The eGEM proposed in our previous work (Wu et al. 2019c) is used as the forward model. In addition,
LGR and a steady-state fundamental solution are used to modify the eGEM to make it more convenient for
reservoir simulation. Furthermore, we validated the modified eGEM and applied it in a proposed workflow
for history matching of the CFNs parameters. Finally, the proposed workflow is implemented in a field case
from Southwestern China.

Integrated Workflow for CFNs Characterization

This section provides an integrated workflow for CFNs generation and calibration. This workflow begins
with CFNs generation with multiple constraint data, including microseismic, well log, core, and hydraulic
fracturing treatment data. After stochastically generating an ensemble of CFNs model parameters, a novel
eGEM is used to predict the production performance of each realization in the ensemble. Finally, the models
are calibrated by using the EnKF to assimilate the production data. The main steps of the workflow are
shown in Fig. 1.

Figure 1—Workflow for CFNs characterization by adsorbing multiple data.

SPE-195928-MS 5

CFNs Generation and Fracture Parameters

CFNs modeling for fractured unconventional reservoirs is with large uncertainty, so it is important to make
the most of the data we have in CFNs generation. CFNs models consist of two parts, including fracture
geometries and properties. In this paper, the CFNs geometry is generated using the method proposed
by Gamboa et al. (2016), and Sun et al. (2016). The fracture properties, including width, porosity, and
permeability, are modeled stochastically constrained by hydraulic fracturing treatment data.
CFNs Geometry. As shown by Mayerhofer et al. (2010), microseismic events are caused by the
reactivations of pre-existing natural fractures, which results in the release of energy in the form of seismic
waves. If microseismic monitoring information is accessible, the microseismic event locations can be used
as the locations of natural fractures. The procedure proposed by Gamboa et al. (2016) is used to generate
the CFNs geometry, and there are mainly two procedures in generating the CFNs geometries, including
stochastic natural fracture generation and hydraulic fracture generation.
In natural fracture generation, three parameters, including fracture location, length, and orientation,
should be generated. The fracture locations are given by the microseismic event locations. Here, we assume
that only one fracture can pass through an event location. Geostatistical results can be used in the generation
of fracture length and orientation. From core and image logging results, we can obtain the orientation, length,
and number of each natural fracture set, so the distributions of length and orientation of each fracture set can
be obtained. The number of a fracture set can be used as the probability in the generation of the fracture set.
In the generation of hydraulic fractures, a simple approach is used to model the fracture propagation
process with the presence of natural fractures. The approach incorporates some common ideas from physical
modeling of hydraulic-natural fracture interaction and hydraulic fracture propagation (Dahi Taleghani and
Olson, 2014). The hydraulic fractures are assumed to propagate along the horizontal maximum stress if there
are no natural fractures connected. When a hydraulic fracture intersects with a natural fracture, we assume
that the fracturing fluid will fill in the natural fracture rapidly and the hydraulic fracture will propagate along
the natural fracture until it arriving the end of the natural fracture or intersecting the next natural fracture.
This propagation process for a fracture stage will stop until the total hydraulic fracture length reaches the
reference length of this stage. Here, the values of reference length for different stages are assumed to be
proportional to the volumes of fracturing fluid used in hydraulic fracturing. Because the 2D problem is
assumed in this paper, fractures are assumed to fully penetrate the target formation.
CFNs Properties. After obtaining the CFNs geometry, the properties should be given, and the
properties mainly include fracture length, permeability, porosity, and width. Zhang et al. (2014) tested
the conductivities of planar hydraulic fractures under different proppant content and closure pressure
conditions. The results show that a propped fracture can have satisfied conductivity even at high closure
stress. Unpropped, aligned fractures can provide a conductive path, but the conductivity is on the order of
0.02 mD.ft at 4,000 psi closure stress. On the other hand, the conductivity of both propped and unpropped
fractures are closure stress-dependent, and the unpropped fractures are more stress-dependent. Gamboa et
al. (2016) used a nonlinear function to fit the experimental data, and a good agreement is obtained. This
nonlinear expression is also used in this paper to characterize the fracture properties, and it is given by:
(1)
In which, σn is closure stress; Cd fracture conductivity; and Cd0, Cdt and κ are fitted parameters.
The fracture conductivity in the initial state is
(2)
Taking Eq. (1)/ Eq. (2), we can obtain the following form
(3)
6 SPE-195928-MS

Where, σni is the closure stress at the initial state, and Δσn = σn – σni.
Stress change can be expressed with pressure changes in the fractures using a uniaxial strain state
(Bachman et al. 2011):

(4)

Where, α is Biot constant, v is Poisson’s ratio, p is pressure, and Δp= p − pi.

Therefore, Eq. (1) can be expressed with
(5)
The conductivity modulus γ, conductivity parameter Cd0, and initial conductivity Cdi are uncertain
parameters that should be matched. In addition, for different fractures, the proppant concentrations are
different, and these parameters are different, especially for propped and unpropped fractures.
Another important issue is to determine which fractures are propped. Some researchers studied proppants
transport in fractures with intersections, and they found that part of the proppants can be carried into the
intersected natural fractures at the intersections. However, more proppant is observed to remain in the
hydraulic fracture beyond the junction than in the natural fractures (Han et al. 2016; Tong and Mohanty
2016). We assume that only the main fractures can be propped, and the interconnected induced fractures
are not propped. Because of the dramatic difference in the properties of propped and unpropped fractures,
their properties should be given respectively.
For each stage, the total length of the propped fractures is assumed to relate to the total amount of
proppants used in the stage.
(6)
In which, Xp,i is the total length of propped fractures in the ith stage; βp is a constant; Vp,i is the total
amount of proppant used in the stage.
These formulations are used to generate the initial ensemble of CFNs parameters. The final fracture
properties should be updated using history matching.

The EGEM-based Forward Model

Review of the Previous EGEM. In our previous work (Wu et al. 2019c), we propose an eGEM for forward
simulation of fluid flow in the unconventional formation with CFNs handled explicitly. The eGEM is a
good extension of BEM for handling nonlinear and heterogeneous problems in fractured reservoirs. In
addition, compared with BEM, much computation can be saved because there are fewer integrals, which
need to be evaluated numerically. Comparing to FDM-based numerical simulator, coarse grids can be used
to simulate the transient behavior of fracture networks, especially in early time. In our previous work, we
mainly focused on presenting the mathematical fundamentals of eGEM and model validation, and transient
Green’s fundamental solution is used in derivation to obtain high precisions for pressure transient behavior
in early flow regimes. Here, we just present several main steps in the model derivation.
For fluid flow in fractured media, wells and fractures are treated as sources or sinks, and the governing
equation is given by

(7)

One should notice that the equations in this section are in the Laplace domain and dimensionless form,
and the definition of the dimensionless variables can be found in our previous work (Wu et al. 2019c).
SPE-195928-MS 7

The transient state fundamental solution proposed by Zauderer (1983) is used in the previous model in
deriving the integral equation, and it is given by
(8)
Using Eq. (8), the integral equation for flow in the reservoir with CFNs is given by

(9)

A schematic of the discretization of the previous eGEM model is shown in Fig. 2(a). Cartesian grids are
used to discretize the domain, and the fractures are embedded into the background grids and discretized
accordingly. We can find that the mesh is quite like the mesh of EDFM, but there are different numerical
models. The solution points of the domain grids are located on the midpoints of the edges. The major
advantage is that the flux terms between any two connected domain blocks are equal in magnitude and
opposite in direction, so they can be eliminated in computation. For blocks on the boundary of the reservoir,
the boundary conditions can be used in computation for edges on the boundary. For closed boundary
condition, the flux term for an edge on the boundary equals zero. The solution points for fracture elements
are also located on the midpoints, and this can avoid singularity on the edges. Matrix properties, such as
permeability and compressibility, are given based on each domain block. Therefore, ▽ln kmD = 0 if the
properties of a block are constant.

Figure 2—Schematics of the eGEM for cases with and without LGR.

Using this discretize, and applying Eq. (9) to each block, we obtain

(10)

Eq. (10) can be written in a more compact form as

(11)
8 SPE-195928-MS

Where, , , , ,

.
FDM is used for fluid flow in the fracture networks, and the flow equations for the matrix and fracture
systems are coupled to obtain the final solution of the model. As we presented in our previous work, eGEM
has a good early time precision when the transient state fundamental solution is used, so it is quite suitable
for pressure transient analysis. However, we can find that the parameters in Eq. (11) should be updated step
by step because is time-dependent. This will take much time to calculate these parameters in reservoir
simulation, especially for history matching.
Modified EGEM for Reservoir Simulation. In this paper, we are focused on making production predictions
for history matching, so two modifications are made to the eGEM to make the program run faster. First, we
used a steady-state flow fundamental solution in the integral equation to avoid parameters calculation step
by step. Second, LGR is used to refine the background grids connected with the fractures.
The fundamental steady-state fundamental solution for Eq. (7) in Laplace domain is given by
(12)
Using Eq. (12), the integral equation for flow in the reservoir with DFN is given by

(13)

Comparing Eqs. (13) and (9), we can find that another integral term over the domain, the fourth term of
Eq. (13), is added to the integral equation. This is because the steady-state fundamental solution is used.
The discretization is quite like our previous work, excepting that LGR is used in this paper. A schematic
is shown in Fig.2(b). First, Cartesian grids are used for background grids to discretize the domain, and
fractures are embedded into the background grids. Then, for domain grids connected with fractures, these
grids are refined. In this way, we can use coarse grids for domain grids far away from the fractures, and
fine grids for the grids connected to the fracture elements. The midpoints of the block edges and fracture
elements are also used as solution points. For different blocks, the number of edges and fracture elements
may not be the same, and this is different from the previous eGEM. For the case shown in Fig. 2 for an
example, there are only 4 edges for all domain blocks, but the number may be larger than 4 if the grids are
refined. There are only 4 edges for blocks ⑥ and ⑦, but there are 6 edges for blocks ① and ②, 8 edges
for block ③, and 10 edges for block ④. Therefore, the discretized integral equation can is given by

(14)

For simplicity, the integral in the fourth term is approximated using average pressure in the block. The
average pressure can be approximated by the weighted average of the pressures on the solution points. It
SPE-195928-MS 9

should be noted that Δln kmD = 0 if the matrix properties are given based on each domain block. Thus, Eq.
(14) can be written in a more compact form as

(15)

Where, , and Cl is given as the length of the edge. In this derivation, the well

is assumed to be controlled by pressure, so the flux term caused by the wellbore is presented on the side
of Eq. (15), which is different from Eq. (11).
We can apply Eq. (15) to each domain block, and let i go through all the solution points of the block to
generate the solution system. In each block, the midpoints of all edges, fracture elements, and wellbores are
given as solution points, so we will present three scenarios in the following to generate the solution system.
Scenario 1: Locating point i on the midpoints of all edges of the eth grid, we can obtain
(16)
Where

Because the solution points are located on the midpoints of the edges, θi = π.
The rank of is Nb, we could rewrite Eq. (16) as
(17)
If we define , Eq. (16) can be written as
(18)
Thus, the equation for one of the Nb midpoints p is
(19)
Where, , and are respectively the pth row of , and .
For each edge within the domain, Eq. (19) can be used to generate two equations, which are respectively
from two connected domain block. As presented in our previous work (Wu et al. 2019c), the summation
10 SPE-195928-MS

of the two obtained from the two connected blocks equals zero. Therefore, we can eliminate
by adding the two equations up.

(20)

Therefore, coupling the equations obtained using all edges, we can obtain the following compact form
(21)
Scenario 2:In each block, locating i on the midpoints of the fracture elements, we can obtain
(22)
Where

Because the fracture elements are within the block, so θi = 2π for this case.
Taking Eq. (18) into Eq. (22), we obtain
(23)
Where , , .
Coupling all the equations obtained using the fracture elements, we obtain
(24)
Scenario 3:In each block, locating i on the wellbore, we obtain
(25)
Where
SPE-195928-MS 11

Taking Eq. (18) into Eq. (25), we obtain

(26)
Where , , .
Coupling the equations obtained using all wellbores, we obtain
(27)
We could obtain the eventual equations for fluid flow in the tight matrix system by coupling Eqs. (21),
(24) and (27).

(28)

In Eq. (28), we already have Nb + NF + Nw equations. However, we have Nb + 2NF + Nw unknowns, so

other NF equations are needed to obtain the ultimate solution of the system. These equations can be obtained
using the equations for fluid flow within the fracture networks. FDM can be used for this problem, and more
details are presented in our previous work (Wu et al. 2019c). The solution of the model is obtained in the
Laplace domain, and the numerical algorithm proposed by Stehfest (1976) is used to inverse the solution to
the time domain. In addition, Picard iteration is used to handle nonlinear parameters in the model.

The EnKF Approach for History Matching

The EnKF is widely used in history matching of production data to calibrate the reservoir model. EnKF-
based history matching is independent of reservoir simulator, and uncertainties in reservoir characterization
can be quantified, so this method is used to update fracture parameters and quantify the uncertainties in this
paper by combining the stochastic fracture generation method and eGEM-based reservoir simulator. The
main steps including initial ensemble generation, prediction step, and assimilation step, and some details
are presented below.
The history matching with EnKF starts with initial ensemble generation. In this step, an ensemble of Ne
model realizations is generated by sampling from the prior PDF of the parameter space.
For each realization, the state variables contain the unknown parameters in the model. In this paper, we are
mainly focused on interpreting the fracture parameters, so the tight matrix is supposed to be homogeneous.
As shown before, each fracture includes three parameters, Cd0, Cdi and γ. Therefore, each state vector is
given by
12 SPE-195928-MS

(29)

The state vector can be defined as follows

(30)

Where m is the vector of a model realization with the length of Nm, g(m) is the vector of production
performance parameters with the length of Nd, and u(m) is the vector of state variables, which include
fracture pressure profile in this paper. u(m) is a vector with the length of Nu.
The initial ensemble can be written as
(31)

Defining , g(m) can be written as

(32)
In the forecasting step, the Ne models in the ensemble are simulated, and the state vectors u(m) and
production performances g(m) are collected.
In the assimilation step, production data in this step is used to update the models in the ensemble.
(33)
Where dobs,n is the observation production performance data at the time step; Ke,n is the Kalman gain at
the time step, and it is calculated by
(34)
Where

(35)

The Validity of the Forward Model

The objective of the following parts is to validate the modified eGEM, and apply the proposed integral
workflow to a field case. In this section, gas flow in the formation is considered. Picard iteration is used to
update the pressure dependent parameters in the model. At present, single-phase gas flow in 2D reservoirs
is addressed for all the cases.
In our previous work, the transient fundamental solution is used in the forward model, and the precision
of the eGEM has been detailed verified with multiple cases. However, the steady-state fundamental solution
and LGR are used in this paper to reduce the computation of the forward model, so the modified eGEM
should be tested.
This case is from a deep shale gas reservoir, and the formation depth is about 3500m. The matrix
and fracture properties are presented in Table 1, and gas properties are shown in Fig. 3. Although many
researchers pointed that complex mechanisms can be significant in the nanopores of shale, such as complex
gas adsorption, non-Darcy flow, and phase behavior (Wu et al. 2016; Zhang et al. 2017, 2018; Cui et al.
2018; Sun et al. 2018, 2019; Xiong et al. 2017, 2019), these effects are not considered in this paper because
we are focused on the simulation and calibration of CFNs.
SPE-195928-MS 13

Table 1—Parameters for the validation case.

Parameter Value

Initial pressure (MPa) 69.8

Formation temperature (K) 375
Formation thickness (m) 47
Matrix permeability (mD) 5×10−5
Matrix porosity 0.045
Rock compressibility (MPa ) −1 1×10−4
Initial water saturation (Irreducible) 0.45
Gas relative density 0.55
Hydraulic fracture permeability (mD) 1×104
Hydraulic fracture width (m) 0.01
Hydraulic fracture porosity 0.3
Bottom hole pressure, (MPa) 5

Figure 3—Gas PVT properties.

To save the computation, only a stage of the fractured shale gas well is simulated. The dimension of the
simulated domain is 400m×300m, and the configuration of the CFNs simulated in this section is shown in
Fig. 4. In the following, we first analyzed the effects of grid sizes on the simulation results, as shown in Fig.
4 (a)∼(d). Then, the technique of LGR is tested using 5 schemes, as shown in Fig. 4 (e)∼(i).
14 SPE-195928-MS

Figure 4—Schematics of the validation model of the modified eGEM.

To study the effects of grid sizes on the simulation results, we compared five scenarios, including
100m×100m, 50m×50m, 25m×25m, 12.5m×12.5m and 6.25m×6.25m. The simulation results are shown
in Fig. 5, in which the gas production rates are compared. We can find that the results will converge to a
single line when we gradually refine the background grids. Particularly, the results are almost the same for
the scenarios with the grid sizes of 25m×25m and smaller. In addition, when the grid size is 100m×100m or
larger, the simulation results deviate significantly from the accurate results. When the grid size is 50m×50m,
a good precision can be obtained if the time is larger than 400 days, but the results are with much error
within 400 days. This is because the early time precision is closely related to the fluid exchange between
fracture segments and the connected domain grids. In addition, the pressure drop around the fractures is
much larger than far away from the fractures. The overall computation precision will be affected when
coarse grids are used for the whole domain. Therefore, it will be helpful to refine the 50m×50m background
grids around the fractures.
Five LGR schemes are compared to find a suitable mesh in simulation. For each scenario, a coarse grids
system is first used for the background grids, and then we refine the grids that are connected with discrete
fractures. Taking the scenario shown in Fig. 4 (e) for an example, the size of the fundamental background
grids is 100m×100m, then the domain grids are refined with 50m×50m if the local grids are connected with
fractures, and the grids for the fractures are regenerated using the refined background grids. The simulation
results for the five LGR schemes are shown in Fig. 6. Two primary background grids are compared, and
they are respectively 100m×100m and 50m×50m. The gas production rate curve with fine grids is used as a
SPE-195928-MS 15

reference solution. As shown in Fig. 6, we can find that the early time productions are in accordance with the
reference solution when the size of the domain grids near the fractures is refined with 25m×25m or smaller.
However, the late productions are dependent on the size of the primary background grids. Therefore, it
is better to use 50m×50m primary domain grids and the size of the refined grids should be 25m×25m or
smaller.

Figure 5—Comparison results of the production rate with different grid sizes.

Figure 6—Comparison results of the production rate with different LGR schemes.
16 SPE-195928-MS

Application to a Multiple Stage Fractured Shale Gas Well

Description of the Field Case
A multiple fractured shale gas well from Southwestern China is used in the analysis. The buried depth of the
shale formation is about 3500m, and the effective thickness of the formation is 30m. The initial pressure and
temperature in the formation are 67.3MPa and 375K, respectively. Initial water saturation in the formation
is about 0.45. Laboratory tests show that the water is almost reducible, and the molar composition of CH4
is more than 98%, so single-phase gas flow is assumed in this case. Gas PVT properties are shown in Fig.
3. Core analysis shows that the porosity of the tight matrix is about 0.04. The tested permeability value has
a wide range, which is from 10−5mD to 10−2mD. This wide range is caused by micro-fractures generated
in the tested cores. Another important factor is matrix compressibility, which is often unknown. Therefore,
the permeability, porosity, and compressibility of the matrix should be calibrated in history matching.
The horizontal well was hydraulically fractured with 17 stages, and there are 3 clusters in each stage.
The induced microseismicity was recorded by 10 receivers placed on the surface, and the microseismic data
is shown in Fig. 7(a). After collecting the results of image log and core data, the azimuth distributions of
natural fractures are obtained and shown in Fig. 7(b). The azimuths of the horizontal well and maximum
horizontal stress are also shown in Fig. 7(b). In hydraulic fracturing, the amount of fracturing fluid and
proppants used in each stage are shown in Fig. 8.

Figure 7—The microseismic data and the azimuth of the natural fracture sets for the field case.

Figure 8—The amount of fracturing fluid and proppants used in each stage.
SPE-195928-MS 17

Stochastic CFNs Modeling

There are mainly 2 steps in modeling the CFNs, generating the natural fractures and modeling hydraulic
fracture propagation. In the generation of natural fractures, we should determine four parameters, including
position, ratio, azimuth, and length. The points of the microseismic events are used as the location of natural
fractures. Each point divides the relevant fracture into two segments, and the ratio of the two segments'
length is assumed to obey a uniform distribution. The length and azimuth of each natural fracture are also
randomly generated. According to Fig. 7(b), we can divide natural fractures into 3 sets, which respectively
strike EW, N65°E, and N10°E. We assume the azimuth of each fracture set obeys a normal distribution,
and the deviations are respectively 3°, 3°, and 10°. According to the fracture numbers in each set, the ratio
of fracture number of the three sets is given as 0.28/0.2/0.52. In practice, we can generate a uniformly
distributed random number. If the number is smaller than 0.28, we assumed the fracture strike EW. If the
number is between 0.28 and 0.48, we assumed the fracture strike N65°E. Otherwise, the fracture is assumed
to strike N10°E. The fracture length is assumed to obey a truncated normal distribution. For example, we
assume the mean of the fracture length is 80m, and the deviation is 20m, and the lower and upper limits
are 100m and 50m, respectively.
The second step is to model hydraulic fracture propagation in the formation with the presence of natural
fractures. In each fracture stage, there are three clusters, and we assume that there will be a hydraulic fracture
in each cluster after hydraulic fracturing. In addition, the amount of fracturing fluid and proppants used in
the three clusters are the same, and they are respectively 1/3 of the total amount for the stage. Hydraulic
fractures will propagate along the maximum horizontal stress until they interact with natural fractures. If
a hydraulic fracture interacts with a natural fracture, it is assumed to propagate along the natural fracture.
In addition, two constraints are added in the fracture propagation. First, the speed of hydraulic fracture
propagation in both sides of a cluster is assumed to be same, excepting that hydraulic fracture propagates
in natural fractures. This is because fracture propagating along a natural fracture is much easier than in
the tight matrix. The second constraint is that the total length of hydraulic fractures (including reactivated
natural fractures) for a fracture stage is proportional to the volume of fracture fluid used in this stage.
Using the procedures presented before, we can obtain the generated CFNs. In Fig. 9, we present 3 CFNs
results with 3 different distributions for natural fracture length. One should note that Fig. 9(a), (c), and
(e) show the CFNs generation results with all the fractures. The points are microseismic events, the black
and red lines are natural fractures and hydraulic fractures, respectively. Fig. 9(b), (d), and (f) only present
the fractures that are connected to the wellbore both directly and indirectly. The blue lines are reactivated
natural fractures.
18 SPE-195928-MS

Figure 9—The CFNs generation results for the field case with different distributions for natural fracture length.

For the length of natural fractures, the case shown in Fig. 9(a) and (b) is with the mean fracture length of
40m, the deviation of 20m, and the lower and upper limits of 10m and 60m, respectively. Fig. 9(c)∼(f) show
the CFNs generated with longer natural fractures. The mean of the fracture length is given as 80m, and the
deviation is 20m. The lower and upper limits are respectively 50m and 100m. For the length of hydraulic
fractures, the largest fracture length is given as 400m in the cases shown in Fig. 9(a)∼(d), and it is 1.3
SPE-195928-MS 19

times of the half-length of the well-space. The largest fracture length in Fig. 9(e) and (f) is given as 200m.
According to the results of the first case, shown in Fig. 9(a) and (b), the generated fractures are not complex,
and many hydraulic fractures are planar fractures. Comparing the first two case, respectively shown in Fig.
9(a) and (c), the results show that the generated hydraulic fractures are more complex if natural fractures
are longer. Comparing the last two case, respectively shown in Fig. 9(c) and (e), the results show that larger
SRVs will be generated if a larger hydraulic fracture length is given.

History Matching of CFNs Parameters

After generating the CFNs, we applied the proposed workflow to calibrate the model using the production
data of the shale gas well. The daily report of the well is presented in Fig. 10, which shows the BHP of 200
days. Here, the CFNs generated in Fig. 9 (c) is used in history matching. In total 387 fractures are modeled,
and each fracture is assumed to fully penetrate the target formation. We should first generate the mesh.
The dimension of the reservoir is given as 1700m×1600m. As the validation case is shown in Fig. 6, the
primary size of the domain grids is given as 50m×50m, and the grids connected to the fractures are refined
to 25m×25m. The results of the mesh used in history matching are shown in Fig. 11.

Figure 10—The daily reports of the BHP for the field case study.

Figure 11—The mesh used for the field case study.

20 SPE-195928-MS

Table 2—Parameters for the validation case.

Parameter Range

log(km), logarithm of matrix permeability −6 ∼ −4

Φm, matrix porosity 0.04 ∼ 0.06

log(cm), logarithm of rock compressibility −5 ∼ −3

log(Cd0), for reactivated natural fractures −4 ∼ −2

log(Cdi), for reactivated natural fractures −2 ∼ 0

γ, for reactivated natural fractures 0.15 ∼ 0.35

log(Cd0), for hydraulic fractures −1 ∼ 0

log(Cdi), for hydraulic fractures 0∼2

γ, for hydraulic fractures 0.05 ∼ 0.15

The unknown parameters are permeability, porosity, and compressibility of the matrix, and conductivity
of the fractures. In history matching, the logarithms of matrix permeability and compressibility are used,
and the formulation shown in Eq. (5) are used to characterize the stress-dependent fracture properties. Table
1 shows the basic formation and fracture parameters for the field case. The distribution of the uncertain
parameters is shown in Fig. 11 and should be determined using history matching. Because very few geology
statistic data are available, so the actual distributions of these parameters are not available. For this case,
large ranges are used for all the parameters, which are assumed to be evenly distributed, as shown in Table 2.
The initial ensemble is generated by sampling from the uniform distributions of the parameters shown in
Table 2. The matrix properties are assumed to be heterogeneous, so only 3 matrix parameters are updated
in history matching. For fractures, there are three parameters for each fracture, so 3NF (NF = 387 in this
case) fracture parameters are updated in each step. The ensemble size is given as 50. In forecasting steps,
the BHP is given day-by-day to make sure that the control mode of the model is the same as the actual
well. The cumulative gas production is used in history matching because the daily production of the well
varies significantly and the values can be erroneous. In addition, the production data for every ten days
is assimilated and used to update the model in the EnKF. The 200 days’ production history is gradually
assimilated using 20-time steps.

Figure 12—The results of history matching and production predictions after updating 5, 10, 15, 20 steps using EnKF.
SPE-195928-MS 21

The history matching results are shown in Fig. 12. One can notice that the uncertainties in history
matching hatching are significantly reduced after 5 steps assimilation. In addition, good matches for the
production data can be obtained after 10 steps assimilation. After history matching, we predicted the gas
production for 10000 days using the models in the updated ensemble at the BHP of 5MPa. The cumulative
production profiles are shown in Fig. 12 (e)∼(h). The predicted eventual shale gas production using the
models in the initial and final ensembles are shown in Fig. 13 (a) and (b). we can also find that the
uncertainties in the EUR of the well are significantly reduced after history matching. The EUR of the well
is estimated to be 1.17×108m3.
The posterior distribution of the model parameters is shown in Fig. 13. The blue bars are the posterior
distributions of the parameters. Although the prior distribution ranges of the parameters are very wide, we
can find that the posterior distributions of the parameters are significantly narrowed down. This shows that
the uncertainties of the parameters are significantly reduced after history matching. The results show that the
matrix permeability and porosity are almost determined after history matching because the distributions are
very narrow, as shown in Fig. 13 (c) and (d). However, as shown in Fig. 13 (e), the matrix compressibility
is hard to be determined because the rock compressibility is very small comparing to gas compressibility.
In addition, gas production is not sensitive to matrix compressibility. Fig. 13 (f)∼(g) show the distribution
of the initial fracture conductivity and conductivity modulus of an unpropped natural fracture, and Fig. 13
(h)∼(i) show those distributions of a propped fracture. To understand the heterogeneities of the fracture
properties, we presented the profiles of initial conductivities and conductivity modulus of the CFNs in
Fig. 14. The results show that the conductivities are much larger for fractures directly connected with the
wellbore. In addition, the conductivity moduli are much smaller for fractures directly connected with the
wellbore. This shows that the fractures near the wellbore are propped, while there may be no proppants or
very few proppant for many reactivated natural fractures far away from the wellbore. These phenomena are
in accordance with our basic knowledge about hydraulic fracturing and the proppant transport simulation
results (Han et al. 2016; Tong and Mohanty 2016).
22 SPE-195928-MS

Figure 13—The posterior distributions of gas production and model parameters.

Figure 14—The initial conductivity and conductivity modulus of CFNs for one of the realizations in the updated ensemble.

Conclusions
In this paper, we introduced an integrated workflow for characterizing the CFNs properties in shale gas
reservoirs. A stochastic CFNs modeling approach, eGEM-based discrete fracture model, and EnKF-based
history matching are incorporated in this workflow. The eGEM is modified by using LGR and steady-state
SPE-195928-MS 23

fundamental solutions in order to make the eGEM more convenient in reservoir simulation. We applied
the workflow to analyze a field case from Southwestern China. The following conclusions are guaranteed
from this study:

• Although the steady-state fundamental solution is used in deriving the integral equation, the eGEM
can still remain a quite high precision with coarse grids when handling CFNs. The grid size near the
fractures has a significant influence on the precision of early time production, so LGR should be
used to this kind of DFMs to reduce the computation. The simulation cases show that proposed LGR
technique is quite efficient for in practice. Therefore, coarse grids can be used for the background
grids far away from the fractures, and we just need to slightly refine the background grids connected
with the fractures to have high precisions in early production time.
• The stochastic CFNs generation approach provides an efficient method to utilize microseismic,
core and hydraulic fracturing treatment data in modeling. The generated CFNs geometries are
very sensitive to the length of both natural and hydraulic fractures, which are difficult to obtain in
practice. However, the unknowns are totally different for different realizations of CFNs geometry,
updating the CFNs with different unknowns may be impossible. More future work should be
conducted to characterize the CFNs for this problem.
• The proposed workflow is shown to be efficient in the characterization of CFNs in unconventional
reservoirs. The application of EnKF enables history matching and uncertainty quantification in
the same process. The results of the field case show that the uncertainties of the CFNs parameters
and EUR are significantly reduced after history matching. Reliable EUR predictions and CFNs
properties can be obtained using this workflow.

Acknowledgments
The authors acknowledge that this study was partially funded by the National Major Science and Technology
Projects of China (No. 2017ZX05037001 and 2016ZX05013004). We also thank the National Natural
Science Foundation of China (No. U1762210, 51574258 and 41672132) for financial support.

Nomenclature
cm = compressibility of the matrix, MPa−1
Cd = fracture conductivity, mD.m
G = fundamental Green’s solution
k = permeability, mD
kF = discrete fracture permeability, mD
kFD = dimensionless discrete fracture permeability
km = matrix permeability, mD
kr = reference permeability, mD
l = the length along a fracture, m
lD = dimensionless length along a fracture
Lr = reference length, m
n = outer normal on the boundary
p = pressure, MPa
pFD = dimensionless discrete fracture pressure
pi = initial pressure, MPa
pmD = dimensionless matrix pressure
pwD = dimensionless bottom hole pressure
qFD = Dimensionless flux strength along the fracture
24 SPE-195928-MS

qr = reference flow rate, m3/d

qwD = dimensionless wellbore production rate
rD = dimensionless radius
s = Laplace constant
t = time, day
wF = fracture width, m
x = x- direction, m
xD = dimensionless x-direction
y = y- direction
yD = dimensionless y-direction

Greeks symbols
Ф = porosity, m3/m3
Фm = the porosity of matrix, m3/m3
ФF = the porosity of discrete fracture, m3/m3
Γ = boundary
Ф = pressure transmit coefficient, mD.MPa/mPa.s
ηmD = dimensionless diffusivity ratio of matrix
ηFD = dimensionless diffusivity ratio of discrete fracture
∂ = differential operator
μ = fluid viscosity, mPa.s
δ = Dirac constant
θi = the angle enclosed by the boundary
Ω = domain

Superscripts
b = boundary
e = the eth block
F = fracture
j = the jth boundary element
k = the kth fracture element
nf = the nfth fracture element

Subscripts
D = dimensionless
F = discrete fracture
m = matrix
i = the ith element
j = the jth element
k = the kth element
sc = surface condition
p the pth perforation
w wellbore
SPE-195928-MS 25

References
Al-Ahmadi, H. A., Wattenbarger, R. A. 2011. Triple-porosity Models: One Further Step towards Capturing Fractured
Reservoirs Heterogeneity. Presented at the SPE/DGS Saudi Arabia Section Technical Symposium and Exhibition,
15-18 May, Al-Khobar, Saudi Arabia. SPE-149054-MS.
Aanonsen, S. I., Nævdal, G., Oliver, D. S., et al 2009. The ensemble Kalman filter in reservoir engineering--a review.
SPE J, 14(03): 393–412. SPE-117274-PA.
Bachman, R. C., Sen, V., Khalmanova, D. et al 2011. Examining the Effects of Stress Dependent Reservoir Permeability on
Stimulated Horizontal Montney Gas Wells. Presented at the Canadian Unconventional Resources Conference, Calgary,
15–17 November. SPE-149331-MS.
Bao, A., Hazlett, R. D., & Babu, D. K. 2017a. A Discrete, Arbitrarily Oriented 3D Plane-Source Analytical Solution to
the Diffusivity Equation for Modeling Reservoir Fluid Flow. SPE J, 22(05), 1–609. SPE-185180-PA.
Bao, A., & Gildin, E. 2017b. Data-Driven Model Reduction Based on Sparsity-Promoting Methods for Multiphase Flow
in Porous Media. In SPE Latin America and Caribbean Petroleum Engineering Conference. Presented at the SPE Latin
America and Caribbean Petroleum Engineering Conference, 17-19 May, Buenos Aires, Argentina. SPE-185514-MS.
Bello, R. O., & Wattenbarger, R. A. 2010. Modeling and Analysis of Shale Gas Production with a Skin Effect. J Can
Petrol Technol. 49(12): 37–48. SPE-143229-PA.
Biryukov, D., & Kuchuk. F. J. 2012. Transient Pressure Behavior of Reservoirs with Discrete Conductive Faults and
Fractures. Transport in Porous Media, 95(1):239–268.
Brown, M. L., Ozkan, E., Raghavan, R. S., et al 2009. Practical Solutions for Pressure-Transient Responses of Fractured
Horizontal Wells in Unconventional Shale Reservoirs. SPE Res Eval & Eng, 14(6): 663–676. SPE-125043-PA.
Cao, Y., & Killough, J. E. 2017. An Improved Boundary Element Method for Modeling Fluid Flow through Fractured
Porous Medium. Presented at the SPE Reservoir Simulation Conference, 20-22 February, Montgomery, Texas, USA.
SPE-182658-MS.
Chai, Z., Tang, H., He, Y., et al 2018. Uncertainty quantification of the fracture network with a novel fractured reservoir
forward model. Presented at the SPE Annual Technical Conference and Exhibition, 24-26 September, Dallas, Texas,
USA. SPE-191395-MS.
Chang, H., & Zhang, D. 2018. History matching of stimulated reservoir volume of shale-gas reservoirs using an iterative
ensemble smoother. SPE J, 23(02): 346–366. SPE-189436-PA.
Chen, C. C., & Raghavan, R. 1997. A Multiply-Fractured Horizontal Well in a Rectangular Drainage Region. SPE J, 2(4):
455–465. SPE-37072-PA.
Chen, Z., Liao, X., Zhao, X., Lv, S., Zhu, L. 2016. A semi-analytical approach for obtaining type curves of multiple-
fractured horizontal wells with secondary-fracture networks. SPE J. 21(2): 538–549. SPE-178913-PA.
Cheng, L., Fang, S., Wu, Y., et al 2017. A hybrid semi-analytical model for production from heterogeneous tight oil
reservoirs with fractured horizontal well. J Petrol Sci Eng, 157: 588–603.
Cinco-Ley, H., Samaniego-V, F., and Dominguez-A, N. 1978. Transient pressure behavior for a well with a finite-
conductivity vertical fracture. SPE J, 18:253–264.
Cipolla, C. L., & Wright, C. A. 2000. Diagnostic techniques to understand hydraulic fracturing: what? why? and how?
Presented at the SPE/CERI Gas Technology Symposium, 3-5 April, Calgary, Alberta, Canada.
Cipolla, C. L., Lolon, E. P., Erdle, J. C., et al 2010. Reservoir Modeling in Shale-Gas Reservoirs. SPE Res Eval & Eng,
13 (4): 638–653. SPE-125530-PA.
Cipolla, C., Weng, X., Mack, M., et ale 2012. Integrating microseismic mapping and complex fracture modeling to
characterize fracture complexity. Presented at the SPE Hydraulic Fracturing Technology Conference, 24-26 January,
the Woodlands, Texas, USA. SPE-140185-MS.
Cui, X., Yang, E., Song, K., et al 2018. Phase equilibrium of hydrocarbons confined in nanopores from a modified Peng-
Robinson equation of state. Presented at the SPE Annual Technical Conference and Exhibition, 24-26 September,
Dallas, Texas, USA.
Dahi Taleghani, A., Olson, J. E. 2014. How Natural Fractures Could Affect Hydraulic-Fracture Geometry. SPE J, 19
(01):161–171. SPE-167608-PA.
Ding, D. Y., Farah, N., Bourbiaux, B., et al 2018. Simulation of Matrix/Fracture Interaction in Low-Permeability Fractured
Unconventional Reservoirs. SPE J, 23(04): 1389–1411. SPE-182608-PA.
Elahi, S. H., & Jafarpour, B. 2018. Dynamic Fracture Characterization from Tracer-Test and Flow-Rate Data with
Ensemble Kalman Filter. SPE J, 23(02): 449–466. SPE-189449-PA.
Evensen, G. 2003. The ensemble Kalman filter: Theoretical formulation and practical implementation. Ocean dynamics,
53(4): 343–367.
Fan, Z., Zhang, Y., & Yang, D. T. 2016. Estimation of three-phase relative permeabilities for a water-alternating-gas
process by use of an improved ensemble randomized maximum-likelihood algorithm. SPE Res Eval & Eng, 19(04):
683–693. SPE-180931-PA.
26 SPE-195928-MS

Fuentes-Cruz, G. and Valko, P. P. 2015. Revisiting the Dual-Porosity/Dual-Permeability Modeling of Unconventional

Reservoirs: the Induced-Interporosity Flow Field. SPE J, 20(1): 124–141. SPE-173895-PA.
Gamboa, E. S., Sun, J., & Schechter, D. 2016. Reducing Uncertainties of Fracture Characterization on Production
Performance by Incorporating Microseismic and Core Analysis Data. Presented at the SPE Asia Pacific Hydraulic
Fracturing Conference, 24-26 August, Beijing, China. SPE-181785-MS.
Geiger, S., Matthai, S., Niessner, J., et al 2009. Black-Oil Simulations for Three-Component, Three-Phase Flow in
Fractured Porous Media. SPE J, 14 (2): 338–354. SPE-107485-PA.
Gringarten, A. C., Ramey, H. J. and Raghavan, R. 1972. Unsteady-State Pressure Distribution Created by a Well with a
Single Infinite-Conductivity Vertical Fracture. SPE J, 14 (4): 347–360. SPE-4051-PA.
Han, J., Yuan, P., Huang, X., et al 2016. Numerical study of proppant transport in complex fracture geometry. Presented
at the SPE Low Perm Symposium, 5-6 May, Denver, Colorado, USA. SPE-180243-MS.
Hoteit, H., & Firoozabadi, A. 2005. Multicomponent Fluid Flow by Discontinuous Galerkin and Mixed Methods in
Unfractured and Fractured Media. Water Resour Res, 41 (11): 274–282.
Idorenyin, E. H., & Shirif, E. 2018. Semianalytical Solution for Modeling the Performance of Complex Multifractured
Horizontal Wells in Unconventional Reservoirs. SPE Res Eval & Eng, 21(04): 961–980. SPE-194019-PA.
Jia, P., Cheng. L., Huang S., et al 2016. A Semi-analytical Model for the Flow Behavior of Naturally Fractured Formations
with Multi-scale Fracture Networks. J Hydrol, 537: 208–220.
Jia, P., Cheng, L., Huang, S., et al (2017). A comprehensive model combining Laplace-transform finite-difference and
boundary-element method for the flow behavior of a two-zone system with discrete fracture network. J Hydrol, 551:
453–469.
Jiang, J., & Younis, R. M. 2016. Hybrid Coupled Discrete Fracture-Matrix and Multicontinuum Models for
Unconventional Reservoir Simulation. SPE J, 21(3): 1009–1027. SPE-178430-PA
Karimi-Fard, M., Durlofsky, L. J., & Aziz, K. 2004. An efficient discrete fracture model applicable for general purpose
reservoir simulators. SPE J, 9 (2): 227–236. SPE-88812-PA.
Kazemi, H., Merrill Jr, L. S., Porterfield, K. L., et al 1976. Numerical simulation of water-oil flow in naturally fractured
reservoirs. SPE J, 16(06): 317–326. SPE-5719-PA.
Kikani, J., & Horne, R. N. 1992. Pressure-Transient Snalysis of Arbitrarily Shaped Reservoirs with the Boundary-Element
Method. SPE Form Eval, 7(1): 53–60. SPE-18159-PA.
Kikani, J., & Horne, R. N. 1993. Modeling Pressure-Transient Behavior of Sectionally Homogeneous Reservoirs by the
Boundary-Element Method. SPE Form Eval, 8(2): 145–152. SPE-19778-PA.
Kim, J. G., & Deo, M. D. 2000. Finite-Element, Discrete-Fracture Model for Multiphase Flow in Porous Media. AIChE
J. 46 (6): 1120–1130.
Larsen, L., & Hegre, T. M. 1991. Pressure-Transient Behavior of Horizontal Wells with Finite-Conductivity Vertical
Fractures. Presented at the International Arctic Technology Conference, Anchorage, USA, 29–31 May. SPE-22076-
MS.
Lee, S. H., Lough, M. F. and Jensen, C. L. 2001. Hierarchical Modeling of Flow in Naturally Fractured Formations with
Multiple Length Scales. Water Resour Res, 37 (3): 443–455.
Li, L., & Lee, S. H. 2008. Efficient Field-Scale Simulation of Black Oil in a Naturally Fractured Reservoir through Discrete
Fracture Networks and Homogenized Media. SPE Res Eval & Eng, 11 (4):750–758. SPE-103901-PA.
Li, X., & Zhu, D. 2018. Temperature Behavior during Multistage Fracture Treatments in Horizontal Wells. SPE Prod
Oper, 33(03), 522–538. SPE-181876-PA.
Liu, Z., & Reynolds, A. C. 2019. History Matching an Unconventional Reservoir with a Complex Fracture Network.
Presented at the SPE Reservoir Simulation Conference, 10-11 April, Galveston, Texas, USA. SPE-193921-MS.
Mayerhofer, M. J., Lolon, E., Warpinski, N. R., et al 2010. What Is Stimulated Reservoir Volume? SPE Prod Oper, 25(1):
89–98. SPE-119890-PA.
Moinfar, A., Varavei, A., Sepehrnoori, K., Johns, R. T. 2014. Development of an efficient embedded discrete fracture
model for 3D compositional reservoir simulation in fractured reservoirs. SPE J, 19(2): 289–303. SPE-154246-PA.
Nejadi, S., Leung, J. Y., Trivedi, J. J., et al 2015. Integrated Characterization of Hydraulically Fractured Shale-Gas
Reservoirs—Production History Matching. SPE Res Eval & Eng, 18(04): 481–494. SPE-171664-PA.
Ozkan, E., & Raghavan, R. 1991. New Solutions for Well-Test Analysis Problems. Part 1: Analytical Considerations. SPE
Form Eval, 6(3): 359–368. SPE-18615-PA.
Ozkan, E., & Raghavan, R. 1991. New Solutions for Well-Test-Analysis Problems. Part 2: Computational Considerations
and Applications. SPE Form Eval, 6(3): 369–378. SPE-18616-PA.
Pan, Y., Zhou, P., Deng, L., & Lee, J. 2019. Production Analysis and Forecasting for Unconventional Reservoirs Using
Laplacian Echo-State Networks. Presented at the SPE Western Regional Meeting, 23-26 April, San Jose, California,
USA. SPE-195243-MS.
SPE-195928-MS 27

Pruess, K., & Narasimhan, T. N. 1982. A practical method for modeling fluid and heat flow in fractured porous media.
SPE J, 25(1): 14–26.SPE-10509-PA.
Stalgorova, K., Mattar, L. 2013. Analytical Model for Unconventional Multifractured Composite Systems. SPE Res Eval
& Eng, 16(3): 246–256. SPE-162516-PA.
Stehfest, H. 1970. Numerical Inversion of Laplace Transforms. ACM Commun, 13 (1): 47–49.
Sun, J., Niu, G., & Schechter, D. 2016. Numerical simulation of stochastically-generated complex fracture networks by
utilizing core and microseismic data for hydraulically fractured horizontal wells in unconventional reservoirs–a field
case study. Presented at the SPE Eastern Regional Meeting, 13-15 September, Canton, Ohio, USA.
Sun, Z., Shi, J., Wu, K., et al 2018. Transport capacity of gas confined in nanoporous ultra-tight gas reservoirs with real
gas effect and water storage mechanisms coupling. Int J Heat Mass Tran, 126: 1007–1018.
Sun, Z., Wu, K., Shi, J., et al 2019. Effect of pore geometry on nanoconfined water transport behavior. AIChE J, e16613.
Tang, H., Killough, J. E., Heidari, Z., & Sun, Z. 2017. A new technique to characterize fracture density by use of neutron
porosity logs enhanced by electrically transported contrast agents. SPE J, 22(04): 1034–1045. SPE-181509-PA.
Tang, H., Hasan, A. R., & Killough, J. 2018. Development and application of a fully implicitly coupled wellbore/reservoir
simulator to characterize the transient liquid loading in horizontal gas wells. SPE J, 23(05): 1615–1629. SPE-187354-
PA.
Tian, C., & Horne, R. 2019. Applying Machine-Learning Techniques To Interpret Flow-Rate, Pressure, and Temperature
Data From Permanent Downhole Gauges. SPE Res Eval & Eng, 22(2): 386–401.
Tong, S., & Mohanty, K. K. 2016. Proppant transport study in fractures with intersections. Fuel, 181: 463–477.
Taigbenu, A. E. 1995. The Green Element Method. Int J Numer Meth Eng, 38(13): 2241–2263.
Wantawin, M., Yu, W., Dachanuwattana, S., & 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 J, 22(06): 2012–2031. SPE-187938-PA.
Warren, J. E., & Root, P. J. 1963. The Behavior of Naturally Fractured Reservoirs. SPE J, 3 (3): 245–255. SPE-426-PA.
Weng, X., Kresse, O., Cohen, C.-E. et al 2011. Modeling of Hydraulic-Fracture-Network Propagation in a Naturally
Fractured Formation. SPE Prod Oper, 26 (4):368–380. SPE-140253-PA.
Wu, K., Li, X., Guo, C., Wang, C., & Chen, Z. 2016. A unified model for gas transfer in nanopores of shale-gas reservoirs:
coupling pore diffusion and surface diffusion. SPE J, 21(05): 1583–1611. SPE-2014-1921039-PA.
Wu, Y. S., & Pruess, K. 1988. A multiple-porosity method for simulation of naturally fractured petroleum reservoirs. SPE
Res Eval & Eng, 3(01): 327–336. SPE-15129-PA.
Wu, Y. S., Di, Y., Kang, Z., et al 2011. A multiple-continuum model for simulating single-phase and multiphase flow in
naturally fractured vuggy reservoirs. J Petrol Sci & Eng, 78(1): 13–22.
Wu, Y., Cheng, L., Huang, S., et al 2019a. An analytical model for analyzing the impact of fracturing fluid-induced
formation damage on rate transient behavior in tight formations. J Petrol Sci Eng, 179: 513–525.
Wu, Y., Cheng, L., Huang, S., et al 2019b. A Semi-analytical Model for Simulating Fluid Flow in Naturally Fractured
Reservoirs With Nonhomogeneous Vugs and Fractures. SPE J, 24(01): 334–348. SPE-194023-PA.
Wu, Y., Cheng, L., Fang, S., et al 2019c. A Novel Edge-Based Green Element Method for Simulating Fluid Flow in
Unconventional Reservoirs with Discrete Fractures. Presented at the SPE Western Regional Meeting, 23-26 April, San
Jose, California, USA. SPE-195342-MS.
Xiong, H., Huang S., Liu H., et al 2017. A novel model to investigate the effects of injector-producer pressure difference
on SAGD for bitumen recovery. Int J Oil Gas Coal T, 16(3): 217–235.
Xiong, H., Devegowda, D., Huang, L. 2019. EOR Solvent-Oil Interaction in Clay-Hosted Pores: Insights from Molecular
Dynamics Simulations. Fuel, 249:233–251.
Yan, X., Huang, Z., Yao, J., Li, Y., et al 2016. An Efficient Embedded Discrete Fracture Model Based on Mimetic Finite
Difference Method. J Petrol Sci Eng. 145: 11–21.
Yang, R., Huang, Z., Li, G., et al 2017. A Semianalytical Approach to Model Two-Phase Flowback of Shale-Gas Wells
with Complex-Fracture-Network Geometries. SPE J, 22(6): 1808–183. SPE-181766-PA.
Yu, W., Wu, K., Sepehrnoori, K., 2016. A Semianalytical Model for Production Simulation from Nonplanar Hydraulic-
Fracture Geometry in Tight Oil Reservoirs. SPE J, 21(3): 924–939. SPE-178440-PA.
Zafari, M. & Reynolds, A.C. 2007. Assessing the Uncertainty in Reservoir Description and Performance Predictions with
the Ensemble Kalman Filter. SPE J, 12 (3): 382–391. SPE-95750-PA.
Zauderer, E., 1983. Partial Differential Equations in Applied Mathematics. Pub. Wiley-Interscience.
Zhang, J., Kamenov, A., Hill, A. D. et al 2014. Laboratory Measurement of Hydraulic-Fracture Conductivities in the
Barnett Shale.SPE Production & Operations, 29(03): 216–227. SPE-163839-PA.
Zhang, S., & Zhu, D. 2017. Inversion of downhole temperature measurements in multistage fracture stimulation in
horizontal wells. Presented at the SPE Annual Technical Conference and Exhibition, 9-11 October, San Antonio, Texas,
USA. SPE-187322-MS.
28 SPE-195928-MS

Zhang, T, Li, X, Sun, Z, et al 2017. An analytical model for relative permeability in water-wet nanoporous media. Chem
Eng Sci, 174:1–12.
Zhang, T, Li, X, Wang, X, et al 2018. Modeling the water transport behavior in organic-rich nanoporous shale with
generalized lattice Boltzmann method. Int J Heat Mass Tran, 127:123–134.
Zhao, Y. L., Xie, S. C., Peng, X. L., et al 2016. Transient pressure response of fractured horizontal wells in tight gas
reservoirs with arbitrary shapes by the boundary element method. Environ Earth Sci, 75(17): 1220.
Zhou, W., Banerjee, R., Poe, B. D., et al 2012. Semi-Analytical Production Simulation of Complex Hydraulic Fracture
Network. SPE J, 19(1): 6–18. SPE-157367-PA.