Accepted Manuscript

Insights to fracture stimulation design in unconventional reservoirs based on machine

learning modeling

Shuhua Wang, Shengnan Chen

PII: S0920-4105(18)31072-6
DOI: https://doi.org/10.1016/j.petrol.2018.11.076
Reference: PETROL 5548

To appear in: Journal of Petroleum Science and Engineering

Received Date: 26 July 2018

Revised Date: 26 November 2018
Accepted Date: 28 November 2018

Please cite this article as: Wang, S., Chen, S., Insights to fracture stimulation design in unconventional
reservoirs based on machine learning modeling, Journal of Petroleum Science and Engineering (2018),
doi: https://doi.org/10.1016/j.petrol.2018.11.076.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

1 Insights to fracture stimulation design in unconventional reservoirs based on

2 machine learning modelling
3
4 Shuhua Wang, Shengnan Chen*

PT
5 Department of Chemical and Petroleum Engineering, Schulich School of Engineering,
6 University of Calgary, AB, T2N 1N4, Canada
7 E-mail address: snchen@ucalgary.ca (S. Chen), shuhwang@ucalgary.ca (S. Wang).

RI
8

SC
9 ABSTRACT

10 With rapid development of unconventional tight and shale reservoirs, considerable amounts of
data sets are increasing rapidly. Data mining techniques are becoming attractive alternatives for

U
11
12 well performance evaluation and optimization. This paper develops a comprehensive data mining
AN
13 process to evaluate well production performance in Montney Formations in western Canadian
14 sedimentary basin. The general data visualization and statistical data evaluation are used to
M

15 qualitatively and quantitatively evaluate the relationships between the stimulation parameters and
16 first-year oil production. Then, the recursive feature elimination with cross validation (RFECV)
D

17 is used to identify the most important factors on the first-year oil production in unconventional
18 reservoirs. In addition, four commonly used supervised learning approaches including random
TE

19 forest (RF), adaptive boosting (AdaBoost), support vector machine (SVM), and neural network
20 (NN) are compared to estimate the first-year well production. The results show that 6 features are
EP

21 the most important variables for constructing an accurate prediction model: well latitude,
22 longitude, well true vertical depth (TVD), proppant pumped per well, well lateral length, and
fluid injected per well. Compared to other algorithms, RF has the best prediction performance for
C

23
24 the first-year oil production. Furthermore, such data-driven models are found to be very useful
AC

25 for reservoir engineers when designing hydraulic fracture treatments in Montney tight reservoirs.
26
27 Keywords: Fracture stimulation design, Data mining, Feature selection, Machine learning,
28 Montney tight reservoirs
29

1
 ACCEPTED MANUSCRIPT

1 1. Introduction

2 Unconventional tight or shale reservoirs are becoming important hydrocarbon resources

3 with the successful application of advanced horizontal well drilling and multi-stage hydraulic
4 fracturing techniques (Kerr, 2010; Hughes, 2013). Optimizing completion and stimulation

PT
5 strategies is critical to maximize the well productivity and/or oil and gas recovery of
6 unconventional formations (Barree et al., 2015; Bhattacharya and Nikolaou, 2016). Using

RI
7 reservoir simulation to evaluate and optimize the performance of completion and stimulation
8 strategies is challenging due to extensive complex processes of hydraulic fracturing. The

SC
9 numerical modeling process itself is also quite time-consuming in solving fluid flow in a very
10 low permeability reservoir with multi-stage hydraulic fractures. With the rapid development of
11 unconventional reservoirs, the amount of data related to petroleum exploitation and production is

U
12 increasing rapidly and continuously (Tan et al., 2017). It is of practical interest for reservoir
AN
13 engineers to develop a reliable and fast well performance analysis and prediction model based on
14 the knowledge extracted from the available data.
15 Data-driven modeling techniques, which integrate the statistical data analysis, data
M

16 clustering and classification, and machine learning algorithms, provide a promising and
17 alternative approach to evaluate the performance of fractured wells in unconventional formations
D

18 (Shaheen et al., 2011). Several attempts have been made to assess the well completion and
TE

19 stimulation strategies through various machine learning methods (Shelley et al., 2008; Shelley
20 and Harris, 2009; Montgomery and Sullivan, 2017; Shelley and Stephenson, 2000; Awoleke and
Lane, 2011; LaFollette et al., 2012; Gullickson et al., 2014; Lolon et al., 2016; Centurion et al.,
EP

21
22 2012; Zhou et al., 2014; Schuetter et al., 2015). With the widespread use of machine learning
23 algorithms in oil and gas industry, more and more challenges arising from the complex nature of
C

24 petroleum exploitation need to be overcome. The most fundamental problem of production

AC

25 forecasting is to identify the importance and influence of various factors on the well productivity.
26 It is well known that the well performance in tight or shale reservoirs is largely affected not only
27 by the fracture stimulation design including lateral length of horizontal well, number of fracture
28 stages, mass of proppant and volume of water used for hydraulic fracturing, but also by the well
29 locations or geological properties. Previous studies either simply ignored geological features in
30 the models (Shelley and Stephenson, 2000; Awoleke and Lane, 2011; LaFollette et al., 2012;
31 Gullickson et al., 2014; Lolon et al., 2016) or assumed geological homogeneity within a small
2
 ACCEPTED MANUSCRIPT

1 sample of wells (Centurion et al., 2012; Zhou et al., 2014). One of the major reasons is that it is
2 difficult to collect all geological properties, such as permeability, porosity, reservoir pressure,
3 and initial water saturation for each horizontal well. However, geological properties in tight or
4 shale formations have been found to vary enormously over even relatively short distances
5 (Clarkson et al., 2012). Thus, incorporating these geological features into machine learning

PT
6 models is necessary. Another alternative approach is to use well locations (e.g., coordinates) to
7 represent the geological differences between wells (Schuetter et al., 2015). In addition, machine

RI
8 learning models trained with more features are expected to provide a better prediction
9 performance, but in practice, excessive features may not only largely slow down the learning

SC
10 process, but also decrease the accuracy of prediction models (Yu and Liu, 2004). Feature
11 selection has been an effective approach to improve model accuracy by eliminating the irrelevant

U
12 or redundant features from the original dataset. Therefore, it is essential to select the features that
13 are most useful and relevant for the problem before applying the machine learning algorithms on
AN
14 the dataset.
15 The neural network method is most commonly used in oil and gas industry for either
M

16 classification or regression problems (Shelley and Stephenson, 2000; Awoleke and Lane, 2011;
17 Ashenayi et al., 1994; Al-Fattah and Startzman, 2003; Hegeman et al., 2009). More recently,
D

18 other data mining approaches such as decision tree (DT), random forest, adaptive boosting, and
19 support vector machine have been gradually applied in various application domains (Tan et al.,
TE

20 2017; Lolon et al., 2016; Schuetter et al., 2015). However, there remains a major challenge in
21 choosing the right machine learning algorithms for production prediction problem in
EP

22 unconventional reservoirs. Inappropriate selection of methods may lead to slow training process,
23 low prediction accuracy, and overfitting problems. In general, SVM has a better performance for
C

24 relatively small data sets with fewer outliers, while RF may require more data sets, but they can
25 come up with a robust model. Compared to the SVM and RF, NN requires a large amount of data
AC

26 to work well, whereas, setting up a neural network is tedious and time-consuming. Nevertheless,
27 comprehensive evaluation of various supervised learning methods is essential to achieve a fast
28 and robust prediction model.
29 In this study, a comprehensive data mining process is performed on the Montney field data.
30 The dataset not only contains well information such as well locations, true vertical depth, well
31 lateral length, and wellbore direction, but also provides detailed stimulation information

3
 ACCEPTED MANUSCRIPT

1 including number of fracture stages, total volumes of fluid and proppant, and fracturing fluid
2 type. General data visualization and statistical data analysis are firstly carried out to qualitatively
3 interpret the relationship between the stimulation design and first-year well production. Then,
4 recursive feature elimination method is used to evaluate the importance of features in
5 constructing the prediction models. After removing redundant and uncorrelated features, four

PT
6 commonly used supervised learning approaches including random forest, adaptive boosting,
7 support vector machine, and neural network are evaluated to predict the first-year well

RI
8 production in Montney tight reservoirs. Finally, the newly developed prediction model is used to
9 optimize the fracture stimulation strategies (e.g., the mass of proppant and volume of fracturing

SC
10 fluid) for field cases.
11

U
12 2. Data preparation
AN
13 The data used in this study comes from horizontal wells drilled in Montney Formation.
14 Montney Formation is the primary oil and gas producing layers in western Canadian sedimentary
15 basin (WCSB) located in southwest Alberta, as shown in Fig. 1. The lower Triassic Montney
M

16 Formation was deposited in an open shelf marine environment during Jurassic-Cretaceous

17 (Ghanizadeh et al., 2015). Formation thickness increases from less than 1 meter in the east up to
D

18 300-400 meters in the west. The depth to the top of the Montney also increases in the same
TE

19 direction from approximately 500 meters in the east to over 4000 meters in the west (Rivard et
20 al., 2014). This variation in burial results in the development of a continuous hydrocarbon
generation including oil, volatile oil, gas condensate, and dry gas zones from east to west
EP

21
22 (Ghanizadeh et al., 2015; Rivard et al., 2014; Kuppe et al., 2012). From the perspective of
23 lithostratigraphy, the Montney Formation can be subdivided into two main members: the lower
C

24 siltstone-sandstone and the upper shale members (Ghanizadeh et al., 2015; Dixon, 2000). The
AC

25 upper Montney is about 100 meters in thickness dominated by fine-to-coarse-grained silts and
26 fine-grained sands with minor inter-bedded shales (Kuppe et al., 2012). In addition to the lower
27 Montney, it is dominantly composed of inter-bedded shales and fine-grained siltstones with
28 approximately 100 meters in thickness as well. Because the upper Montney is more organic
29 richer than the lower member, most of horizontal wells were drilled within the upper Montney
30 Formation (Chalmers and Bustin, 2012). Although the matrix permeability and porosity are
31 extremely low, which are typically in the ranges of 1.0×10-3 to 6.5×10-7 mD and 3% to 12%,
4
 ACCEPTED MANUSCRIPT

1 respectively, the Montney Formation is becoming a well-established producing zone for both oil
2 and gas throughout western Alberta and northeastern British Columbia due to the development
3 of advanced multi-stage hydraulic fracturing and horizontal drilling technologies. It is estimated
4 that more than 4000 multi-stage fractured horizontal wells were drilled or licensed in Montney
5 Formation since 2005.

PT
6

RI
U SC
AN
M

Horizontal wells
D

7
TE

8 Fig. 1. The study area of Montney Formation and wells’ distribution.

9
EP

10 In this work, a total of 4790 horizontal wells were obtained from the well completion and
11 fracture database (i.e., Well Completion and Frac DataBase by GeoLogic Systems). Two kinds
12 of features were considered the most important for oil and gas production: the geological
C

13 properties of formation and well completion/stimulation parameters. However, due to the

AC

14 limitation of accessible open database, the geological properties of formation such as

15 permeability, porosity, reservoir pressure, initial water saturation, and geothermal gradient were
16 difficult to be obtained. Thus, well locations (i.e., latitude, longitude, and well true vertical depth)
17 were used to indirectly represent geological differences between wells. Other features including
18 well lateral length, well completion type, wellbore direction, and the usage of fracturing fluid
19 and proppant were applied to represent the influence of well completions/stimulation on the

5
 ACCEPTED MANUSCRIPT

1 production. It should be noted that the fracture half-length was not included into input features.
2 This is because the fracture half-length cannot be directly obtained from any existing database. It
3 needs to be estimated from an accurate fracture propagation modeling. Although many fracture
4 propagation models are proposed recently, it still cannot guarantee the calculated fracture half-
5 length is accurate enough to develop machine learning models. If input features have much noise

PT
6 and uncertainty, it can significantly affect the accuracy of machine learning models. Therefore,
7 the fracture half-length was not considered an input feature in this study.

RI
8 For each horizontal well, 11 features were extracted from the database. These features can
9 be subdivided into two categories: well information and fracture stimulation design. Table 1

SC
10 presents the range and statistical properties of selected features and target employed for this
11 study. To keep the data points as more as possible for data-driven modelling, only horizontal

U
12 wells with missing features are removed from the database. Then, the wells with all features are
13 assumed to be good quality for developing data-driven models. Thus, a total of 3610 horizontal
AN
14 wells are remained to the data mining process. With regard to the well features, the well
15 locations (i.e., latitude, longitude, and well true vertical depth) are used to identify the geological
M

16 differences between wells due to the lack of the geological properties of horizontal wells. Other
17 well features contain the well lateral length, well completion type, and wellbore direction. The
D

18 commonly used well completion types in hydraulic fractured horizontal wells are perforated
19 cased-hole or open-hole completions. For cased-hole well completion, the cement is pumped
TE

20 down the inside of casing and circulated to the surface through the annulus between the casing
21 and wellbore. This allows to cement the casing tube in place and provides isolation between
EP

22 various production zones. On the contrary, no casing or liner is cemented in place across the
23 whole production zone in the open-hole completion process. It is found that such completion
C

24 types can largely affect the connections between horizontal wellbores and formations. Thus, it is
25 selected as the important feature for the data-driven modeling.
AC

26

6
 ACCEPTED MANUSCRIPT

1 Table 1
2 Statistical properties of input and output variables used in this study.
Standard
Type Parameters Units Range Mean
deviation
Well Latitude [°] 53.78 ~ 56.68 55.36 0.67

PT
information Longitude [°] -121.96 ~ -116.50 -119.26 1.50
Well TVD [m] 805.90 ~ 3863.90 2264.57 527.29
Lateral length [m] 147.00 ~ 3996.00 1702.13 567.81

RI
Well completion type / cased-hole, open-hole
Wellbore direction / E/W, NE/SW, N/S, NW/SE
Fracture Total proppant [ton] 0.70 ~ 10433.26 1309.12 1259.94

SC
stimulation Total fracturing fluid [m3] 45.00~ 59724.01 6247.55 6749.86
design Fracturing fluid type / oil, slickwater, surfactant, water
Fracture stages / 1 ~ 87 18 9

U
Use of interval clusters / yes, no
Production 12-month BOE [mbo] 0.55 ~ 740.59 138.17 96.36
AN
3
4 In addition, when a horizontal well is hydraulically fractured, a transverse or longitudinal
M

5 fracture is created. The propagation pattern of induced fractures is mainly determined by the
6 direction of minimum principal stress (Daneshy, 1971; Daneshy, 1973; Salah et al., 2016). In
D

7 general, the induced fracture has propagated in a direction perpendicular to the minimum
8 principal stress. If the well is drilled along the minimum stress, transverse fractures would be
TE

9 created. However, if the well is drilled perpendicular to the minimum stress, it induces a
10 longitudinal fracture, as shown in Fig. 2. Therefore, it is essential to take the wellbore direction
EP

11 into consideration in the data-driven modelling. In this work, the wellbore directions are
12 classified into four groups on basis of four cardinal directions and four intercardinal directions:
13 east/west (E/W), northeast/southwest (NE/SW), north/south (N/S), northwest/southeast (NW/SE).
C

14
AC

7
 ACCEPTED MANUSCRIPT

PT
RI
SC
1

U
2 Fig. 2. Two fracture propagation patterns in a horizontal well (Salah et al., 2016).
3
AN
4 The hydraulic fracturing process involves injecting a large volume of fluid at a pressure that
5 is sufficiently high to break down the rock and initiate the fractures. Then, a large amount of
M

6 proppant is pumped down to keep fractures open and provide a long-term and high-conductive
7 fractures. The most important fracture stimulation parameters include the total mass of proppant,
D

8 volume of fracturing fluid, number of fracture stages, fracturing fluid type, and use of interval
9 clusters. It is important to note that the average fracture spacing is not included in these features.
TE

10 This is because it can be calculated by the ratio of well lateral length to number of fracture stages.
11 The target variable considered in this work is the cumulative barrels of oil equivalent (BOE)
EP

12 produced by a horizontal well during the first year. The unit of BOE used in this work is mbo,
13 which denotes the thousand barrels of oil. The BOE provides a comparable indicator, which
C

14 combines the oil and gas productions for wells in different types of reservoirs (e.g., oil-producing
15 reservoir, gas-condensate reservoir). Moreover, the first-year BOE production is a good metric
AC

16 not only for the estimated ultimate recovery in unconventional reservoirs, but also makes it
17 possible to neglect the impact of well interference on the production which often occurs at the
18 later production period.
19
20
21

8
 ACCEPTED MANUSCRIPT

1 3. Methodology

2 3.1. Feature selection

3 Feature selection is a process which automatically select the features that contribute most to
4 the prediction variable or target. It is an indispensable step in data mining process because it can

PT
5 largely reduce the training time, improve the accuracy of prediction models, and avoid the
6 overfitting problems. In this work, the recursive feature elimination with cross validation

RI
7 (RFECV) is used to identify which features are most important to predict the first-year oil
8 production in unconventional reservoirs. Although a set of features are assembled in the database,

SC
9 it is not clear which features are more important for developing the prediction model. RFECV
10 performs the recursive feature elimination process in a cross-validation loop to identify and
11 remove the features of low importance. By recursively employing the subsets of features into

U
12 prediction models, the training data set can be pruned to only the features responsible for
AN
13 prediction. In recursive feature elimination process, the random forest is used as the estimator.
14 When the data set is fed into the model, the importance of each feature can be calculated based
15 on Mean Decrease Impurity (MDI). More information on MDI and feature importance
M

16 evaluation with random forests can be found elsewhere (Louppe, 2014). The major concept of
17 feature importance evaluation in random forests is the relative rank of a feature used as a
D

18 decision node in a decision tree. Features used at the top of the tree contribute to the final
TE

19 prediction decision of a larger fraction of the input samples. The expected fraction of the samples
20 they contribute to can thus be used as an estimation of the relative importance of features. For
each recursion step within an iteration, the feature with the lowest importance is eliminated and
EP

21
22 the random forest model is re-trained by left features, until a single feature remains.
23 Fig. 3 shows the data partitioning scheme and RFECV process. It should be noted that the
C

24 original data set was randomly split to training set (90%) and testing data set (10%). 9-fold cross
AC

25 validation was used for the RFECV process. This data partition strategy guarantees the original
26 samples are partitioned into ten equal size subsamples. In other words, the testing data set (10%)
27 is equal to each subset (10% of original data set) of 9-fold cross validation. To avoid the
28 confusion with the testing set in cross validation process, the testing data set was called the
29 holdout set. No training was performed on the holdout set. Overall, 1000 model runs are
30 performed for the feature selection. Following steps are carried out for each model run:

9
 ACCEPTED MANUSCRIPT

1 (1) The original database including 11 features and 3610 observations is randomly partitioned
2 into the training (90%) and holdout (10%) data set. No training is carried out on the holdout
3 set.
4 (2) Recursive feature elimination with 9-fold cross validation is performed.
5 Firstly, the training data set is subdivided into 9 cross-validation folds. Each fold consists of

PT
6 10% of original data set (i.e., 11.11% of the training data set). Eight folds are aggregated to train
7 the model, and the remaining fold is used to test the performance of prediction model. Random

RI
8 forest is used as the prediction model in RFECV process. The performance of prediction model
9 is evaluated by mean squared error (MSE) and coefficient of determination ( ):

SC
10

1
MSE = − (1)

=1− U ∑ −
AN
∑ −
(2)

11
M

12 where denotes the number of data points; is the th target value; denotes th predicted
13 value; is the mean value of observations.
D

14 Then, the prediction model is recursively re-trained and evaluated by systematically

TE

15 removing features with the lowest importance from the training data set. It should be noted that
16 the database partition process is randomness, the trimmed database varies in features between
17 model runs. Nine recursive feature elimination iterations would be carried out for each model run
EP

18 because of 9-fold cross-validation process. For each of the nine RFECV iterations, 11 random
19 forests are generated with the feature set size decreasing from 11 to 1. Each random forest is
C

20 trained and evaluated by the cross-validation score (i.e., MSE). The feature set size
AC

21 corresponding to the highest cross-validation score is reported.

22 (3) The features of training data set are reduced to the optimum features identified by RFECV in
23 step 2. Thus, a new prediction model is developed and trained by the trimmed training data
24 set (i.e., feature-reduced training data set). Finally, the prediction performance is evaluated
25 using the holdout set.
26

10
 ACCEPTED MANUSCRIPT

PT
RI
U SC
1
AN
2 Fig. 3. Data partitioning scheme for recursive feature elimination with cross validation.
3
4 Model uncertainty assessment is essential for most of machine learning algorithms. The
M

5 learning process is an inverse problem with non-unique solutions for a given training dataset (Ma
6 et al., 2015). The model uncertainty of random forest may stem from the random database
D

7 partition process and bootstrapping (i.e., random sampling with replacement). In this work, 1000
TE

8 model runs are performed, and the optimal feature size is determined by weighting the number of
9 features, mean value and variance of performance score (i.e., ). The model with a higher
10 performance score, fewer features, and lower variance is preferred. A trade-off equation is
EP

11 proposed for optimum feature size determination:

1
optimum result = max % ∙ 3
(.'+(/'
01 2
C

&'()*+',
(3)

1
6
AC

012 =4 −
5−1 (.'+(/'
(4)

where &'()*+', denotes the number of features; (.'+(/' is the mean value of coefficient of
012 represents the unbiased
12

13 determination for the model runs with feature size of &'()*+', ;

14 sample standard deviation of coefficient of determination for the model runs with feature size of

11
 ACCEPTED MANUSCRIPT

1 &'()*+', ; is the coefficient of determination of th model run; 5 is the number of prediction

2 models.
3

4 3.2. Data preprocessing

PT
5 Data normalization is a common requirement for many machine learning algorithms. To
6 improve the training efficiency of the models, the standardization method is applied to normalize

RI
8 −9
7 the inputs and target:

7 =
:
(5)

1
6

SC
9= 8 (6)

:=4
1
U 8 −9
AN
(7)

where 7 is the standard score of the th sample; 8 denotes the th sample; 9 is the mean value of
the samples; : is the standard deviation of the samples;
8
M

9 denotes the number of samples. After

10 standardization, the samples are rescaled to a standard normal distribution with a mean of 0 and a
D

11 standard deviation of 1.
TE

12

13 3.3. Random forest

EP

14 Random forest (RF) is an ensemble learning method proposed by Breiman (2001). It

15 constructs a large number of randomized decision trees on bootstrapped training samples and
C

16 aggregates their predictions by averaging the results (James et al., 2015). It has become a major
17 data mining tool for both regression and classification problems. More recently, the consistency
AC

18 of RF has been proved by Scornet et al. (2015). Compared to other machine learning algorithms
19 such as neural networks, RF can achieve a relatively high prediction performance with only few
20 parameters to tune (Genuer et al., 2017). In addition, the independent decision trees can be built
21 in parallel in the RF model, which significantly reduces the training time. Furthermore, because
22 the RF model is developed by aggregating a number of decision trees, it is able to reduce the
23 variance of prediction models and improve the robustness of models. The detailed

12
 ACCEPTED MANUSCRIPT

1 implementations of RF algorithms can be found elsewhere (Breiman, 2001; Scornet et al., 2015).
2 In this work, each RF consists of an ensemble of 500 decision trees. Bootstrap sampling method
3 (i.e., random sampling with replacement) is used to develop the RF models. The bootstrapped
4 training data sets are obtained by repeatedly sampling observations from the original data set.
5 The performance of the RF models is evaluated by the metrics defined by equation 1 and 2.

PT
6

7 3.4. Adaptive boosting

RI
8 Adaptive boosting (AdaBoost) method is one of most successful ensemble learning

SC
9 algorithm proposed by Freund and Schapire (1991). Unlike RF and bagging method, AdaBoost
10 sequentially generates a set of weak learners (e.g. decision trees) and combines their results using
11 a weighted average approach. In this technique, all data is applied to train each model, but

U
12 samples that are mis-predicted by the previous models are given more weights so that subsequent
AN
13 models give more training on them. Thus, boosting method can largely reduce the bias between
14 the prediction and target. However, in some cases, this training strategy results in overfitting
15 problems especially for the noisy training samples in comparison with the bagging method (Sun
M

16 et al., 2016). A detailed description of AdaBoost method can be found elsewhere (Freund and
17 Schapire, 1991; Drucker, 1997). In this work, the decision tree is used as the base predictor in
D

18 AdaBoost algorithm. The number of predictors is set to be 500. MSE and defined by equation
TE

19 1 and 2 are used to assess the performance of AdaBoost models.

20
EP

21 3.5. Support vector machine

22 Support vector machine (SVM) invented by Cortes and Vapnik (1995) is widely used for
C

23 regression and classification problems. The basic conception of SVM is to construct a

24 hyperplane or a set of hyperplanes in a high-dimension feature space to classify the data points.
AC

map the input vectors into a high-dimension space. The ; -support vector regression (; -SVR) is
25 The kernel functions such as linear, polynomial, and radial basis function (RBF) are employed to

applied in this study. The detailed mathematical formulations of ;-SVR can be found elsewhere
26

(Vapnik, 1998; Smola and Scholkopf, 2004). RBF is used as the kernel function in ; -SVR. There
27

are two important parameters (i.e., penalty parameter < and kernel parameter =) that should be
28
29
30 chosen carefully to avoid the over fitting problems. The larger the penalty parameter is used; the

13
 ACCEPTED MANUSCRIPT

more error is penalized. In this work, prior to the models’ comparison, the penalty parameter (< )
and kernel parameter (=) are fine-tuned to improve the performance of ; -SVR. The optimum
1

values of < and = are 500 and 0.1, respectively.

2
3
4

PT
5 3.6. Neural network

6 Neural Network (NN) is powerful machine learning method that is capable of modeling the

RI
7 complex and nonlinear relationship between inputs and outputs through mimicking the biological
8 neural network behavior (LeCun et al., 2015). It is widely used for data classification, pattern

SC
9 recognition, and nonlinear regression. The most commonly used neural network is the multi-
10 layer perception (MLP) network that is composed of an input layer, an output layer, and any
11 number of hidden layers, as shown in Fig. 4 (a). A neural network contains a number of highly

U
12 interconnected and parallel computational elements called neurons or nodes. Fig. 4 (b) shows
AN
13 that each neuron consists of a bias, a transfer function, and weights that connect it to other
14 neurons. The mathematical model of the th neuron can be written as:
E

> = ?@ A B 8B + D F
M

(8)
B

where A B is the weight connected between neuron G of preceding layer and neuron
D

current layer; D is bias or threshold value at neuron of current layer; 8B denotes the output
15 of the
TE

value of node G in preceding layer; > is the output value of node in current layer; ? ∙ is the
16

transfer function or activation function; H is the number of neurons in the preceding layer.
17
18
EP

19
C
AC

14
 ACCEPTED MANUSCRIPT

PT
RI
SC
1
2 Fig. 4. Illustration of a neural network model. (a) a multi-layer perception; (b) computational
3
U
process of a neuron.
AN
4
5 The process of training a neural network involves tuning values of weights and biases in
M

6 each step to correlate output with input variables. There are two phases namely feed-forward
7 pass and backward pass in the training process (Rumelhart et al., 1985). In feed-forward pass, the
D

8 function signal is propagated from the input layer to the output layer. For backward pass, the
9 error signal between predicted targets and actual values is calculated and propagated back to the
TE

10 input layer to update weights and biases. One of the major problems of developing a neural
11 network is the overfitting problem. The early stopping technique is applied in this work to
EP

12 overcome this problem. In the early stopping technique, the data set is divided into three subsets.
13 This first subset is used to train the neural network model. The network weights and biases are
14 iteratively updated to minimize the loss function (i.e., mean squared error) of the training set.
C

15 The second subset is validation set. The loss function on the validation set is monitored during
AC

16 the training process. The validation error normally decreases during initial epochs of training
17 process, as shown in Fig. 5. However, when the training data set is overfitted by the neural
18 network model, the validation error starts to increase although the training error keeps decreasing.
19 Once the validation error increases for a certain number of iterations, the training process is
20 stopped, and the weights and biases with minimum validation error are returned.
21

15
 ACCEPTED MANUSCRIPT

PT
RI
1
2 Fig. 5. Illustration of early stopping technique.

SC
3 The activation functions used in the hidden layers and output layer are rectified linear unit
4 (ReLU) and linear operator, respectively. The stochastic gradient descent is used as the optimizer
for the training process. The learning rate, number of hidden layers, and number of neurons are

U
5
6 evaluated before the comparison of different machine learning algorithms. The optimal learning
AN
7 rate is 0.01 with an adaptive strategy. This adaptive learning process keeps the learning rate a
8 constant value of 0.01 as long as the objective function (i.e., the mean squared error) keeps
M

9 decreasing. Otherwise, the learning rate is divided by 5. The neural network with three hidden
10 layers provides an optimum prediction result. Each hidden layer has 200 neurons. Table 2 shows
D

11 the parameters for different machine learning methods used in this study. In this work, Scikit-
12 learn package proposed by Pedregosa et al. (2011) was used to develop the random forest,
TE

13 AdaBoost, and SVM models. Python-based Keras (Chollet, 2015) was used to develop ANN
14 models.
EP

15
16 Table 2
17 Parameters for different machine learning methods.
C

Methods Parameters Values

AC

Random forest number of decision trees 500

penalty parameter <

AdaBoost number of decision trees 500

kernel parameter =
SVM 500
0.1
kernel function radial basis function
Neural network activation function Rectified (ReLU)
learning rate 0.01
number hidden layers 3

16
 ACCEPTED MANUSCRIPT

number of neurons 200

1

2 4. Results and discussion

3 4.1. General data visualization

PT
4 Fig. 6 summaries the history of Montney Formation in terms of fracture stimulation
5 parameters (i.e., average mass of proppant pumped per well, average fracturing fluid injected per

RI
6 well, and number of fracture stages per well) and first year cumulative oil and gas productions.
7 Very few horizontal wells were drilled before 2007 in Montney Formation, as shown in Fig. 6 (b)

SC
8 (gray bars, primary y-axis), and the number of fracture stages of these wells were less than 10
9 (blue line, secondary y-axis). The number of new wells released per month was moderately
10 increased between 2008 and 2009. However, the hydraulic fracturing process was not intensive
11

U
during this period. A small quantity of proppant and fracturing fluid were used in a single
AN
12 horizontal well (blue bars and red line in Fig. 6 (a)). Moreover, the number of fracture stages was
13 still less than 10. Therefore, the improvement of 12-month oil and gas productions was not
impressive around 2008 and 2009. It is interesting to note that the oil and gas productions were
M

14
15 dramatically increased around 2014. This is not only on account of the increase of new wells, but
16 also because of the significant improvement of hydraulic fracturing technologies. For example,
D

17 Fig. 6 (b) shows that the number of fracture stages was increased to 20 in 2014. Meanwhile, the
TE

18 mass of proppant and volume of fluid used in a single horizontal well were also considerably
19 increased. The mass of proppant was increased to 2000 tons per well, and the volume of
fracturing fluid was increased to 10000 m3 per well. However, the increment of cumulative oil
EP

20
21 and gas production before the 2014 was much less than that after 2014, although the number of
22 wells and fracture stages were increased continuously. This is can be attributed to the fact that
C

23 the increases of fracturing fluid and proppant per well were relatively small before 2014. After
AC

24 2014, it is found that the usage of fracturing fluid and proppant per well were increased
25 significantly. Furthermore, it should be noted that both the number of new wells and productions
26 were decreased because of the downturn of oil prices around 2015 and 2016.
27

17
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE

1
2 Fig. 6. History of Montney Formation. (a) average fluid injection per well and average proppant
EP

3 pumped per well; (b) number of new release wells per month and average fracture stages per
4 well; (c) 12-month gas and oil productions.
C

5
AC

18
 ACCEPTED MANUSCRIPT

PT
RI
1
2 Fig. 7. Normalized 12-month oil production distribution with and without TVD.

SC
3

U
AN
M
D

4
5 Fig. 8. Normalized 12-month gas production distribution with and without TVD.
TE

6
C EP
AC

7
8 Fig. 9. Normalized 12-month water production distribution with and without TVD.
9
10 Figs. 7 through 9 illustrate the distributions of 12-month cumulative oil, gas, and water
11 production. The figure without TVD means that all data points are projected into the longitude-
19
 ACCEPTED MANUSCRIPT

1 latitude plane. It can be seen that oil, gas, and water production in Montney Formation is
2 unevenly distributed, implying that the reservoir quality varies with well locations across the
3 study area. The cumulative oil production decreases from northeast to southwest, on the contrary,
4 both the cumulative gas production and well true vertical depth increase from northeast to
5 southwest. This can be attributed to the mixed hydrocarbon system in Montney Formation

PT
6 because of the variation of formation depth. The deeper Montney reservoirs in the southwest area
7 tends to be less rich in liquids, while shallower Montney reservoirs in the northeast are liquid-

RI
8 rich and may be oil saturated (Kuppe et al., 2012). As a result, more gas is produced from the
9 southwest area of Montney Formation, and more oil or condensate is produced from the

SC
10 northeast area of Montney formation. In addition, these hydrocarbon production distributions are
11 useful to identify the hot zones in the Montney Formation with the potential to yield a high

U
12 production.
13
AN
14 4.2. Statistical data evaluation

15 In this work, the traditional data analysis such as the plotting of factors versus the target
M

16 variable is found to be difficult to quantify the relationships between the stimulation parameters
17 and the cumulative production. However, the general trends can be qualitatively identified using
D

18 the boxplot. Fig. 10 shows the general relationships between various stimulation parameters and
TE

19 the cumulative 12-month BOE. In the boxplot, the box ranges from the first quartile (25%) to the
20 third quartile (75%) of the grouped data set. The number above or below the box represents the
percentage of number of wells that belong to each range. For example, 3.9% in Fig. 10 (a)
EP

21
22 represents 3.9% of horizontal wells has a true vertical depth less than 1000 m. Furthermore, the
23 Pearson correlation is used to quantitatively evaluate the linear relationship between various
C

24 parameters. Fig. 11 depicts the Pearson correlation matrix. The Pearson coefficient has a range
AC

25 from +1 to -1. A value of 0 indicates that there is no linear association between the two variables.
26 A value of +1 represents a perfect positive (i.e., increasing) linear relationship, while -1 indicates
27 a negative (i.e., decreasing) linear relationship between the two variables.
28

20
 ACCEPTED MANUSCRIPT

PT
1

RI
U SC
2
AN
M
D

3
4 Fig. 10. Boxplots of effects of fracture stimulation parameters on 12-month cumulative BOE. (a)
TE

5 well true vertical depth; (b) well lateral length; (c) well completion type; (d) wellbore direction;
6 (e) proppant pumped per well; (f) fracturing fluid injected per well; (g) fracturing fluid type; (h)
EP

7 number of fracture stages; (i) use of interval clusters.

8
C

9 Several interesting findings are noted. First, Fig. 10 (a), (b), (e), (f), and (h) illustrate that the
10 12-month cumulative production increases with the increase of well true vertical depth, well
AC

11 lateral length, proppant pumped per well, fluid injected per well, and number of fracture stages.
12 These positive correlations are also indicated by the Pearson coefficients, as shown in Fig. 11. It
13 is found that the proppant has the largest Pearson coefficient (i.e., =0.53), indicating the
14 greatest positive impact on the cumulative production in comparison with other four factors. It is
15 also interesting to noted that the number of fractures stages has a minimal impact on the
16 cumulative production (i.e., =0.24). The effects of fracturing parameters on the 12-month

21
 ACCEPTED MANUSCRIPT

1 cumulative production are in the following order in terms of their positive impact: proppant
2 pumped per well, well lateral length, fluid injected per well, well true vertical depth, and number
3 of fracture stages. Fig. 10 (a) shows that well true vertical depth plays important roles in oil and
4 gas production in Montney Formation. The 12-month BOE production increases with the
5 increase of well true vertical depth. This can be attributed to several reasons. One of major

PT
6 reasons is that the reservoir pressure increases with the increase of well true vertical depth. The
7 pressure gradient of Montney Formation is estimated to a range of 8 to 12 kPa/m by Seifert et al.

RI
8 (2015). If the average value of 10 kPa/m is used, the pressure difference between wells with the
9 true vertical depth of 1000 and 3000 m is calculated to be 20 MPa. This means the reservoir

SC
10 pressure of horizontal well with 3000 m TVD is 20 MPa larger than that of well with 1000 m
11 TVD. The main production mechanism of oil and gas in unconventional Montney Formation is

U
12 primary drive mechanism (i.e., pressure-depletion mechanism). The larger initial reservoir
13 pressure implies the larger reservoir energy available for production. Another reason is that the
AN
14 formation thickness increases with the increase of well true vertical depth. Kuppe et al. (2012)
15 demonstrated that the depth from the surface to the top of Montney increases from northeast to
M

16 southwest. Similarly, the thickness of Montney Formation increases from less than 1 m at the
17 northeast edge to more than 350 m along the west of Alberta border. Thus, horizontal wells with
D

18 a larger true vertical depth were drilled in a thicker Montney Formation. These wells have a
19 larger drainage or production area compared to the wells with a small true vertical depth. As a
TE

20 result, horizontal wells with larger true vertical depth have better production performance in
21 Montney tight reservoirs. Fig. 10 (b) shows that the BOE production increases with the increase
EP

22 of well lateral length. This is because the well drainage area increases as the horizontal well
23 lateral length increases. It is also found that 12-month BOE increases with the increase of
C

24 proppant pump per well, as shown in Fig. 10 (e). In the hydraulic fracturing process, the
25 proppant is used to keep fractures opening and provide a high conductivity flow path for oil and
AC

26 gas. The more proppant is used, the larger stimulated reservoir volume is created. Thus, the BOE
27 production is significantly increased by pumping more proppant. In addition, Fig. 10 (f) also
28 depicts that the BOE production increases with the increase of fluid injected per well. This can
29 be attributed to the fact that more fractures are created when a larger amount of fracturing fluid is
30 injected into the formation.

22
 ACCEPTED MANUSCRIPT

1 Second, as mentioned previously, the horizontal wells with open-hole system have a larger
2 flow area and lower skin factors in comparison with the wells used cased-hole completion. Thus,
3 a higher cumulative production can be expected with the open-hole completion. However, Fig.
4 10 (c) demonstrates that, although more than half of wells (i.e., 65.1%) were completed with
5 open-hole system in Montney Formation, the open-hole system does not outperform the cased-

PT
6 hole system in terms of 12-month cumulative production. Meanwhile, the Pearson coefficient
7 between well completion type and the production is equal to zero, which indicates that there is

RI
8 no apparent linear correlation between these two variables.
9 Third, it is worthwhile noting that about half of horizontal wells (i.e., 50.1%) were drilled

SC
10 along the NW/SE direction in Montney Formation, as shown in Fig. 10 (d). The boxplot also
11 shows that these wells have a better production performance. This is because the regional stress

U
12 regime of the Montney Formation is compressional with the regional maximum horizontal stress
13 direction approximately N40°E (Steinhoff, 2013), as shown in Fig. 12. Thus, the horizontal wells
AN
14 with NW/SE direction are likely to create the transverse fractures (as shown in Fig. 2). In
15 addition, the transverse fractures outperform the longitudinal fractures in cases where reservoir
M

16 permeability is extremely low (Economides et al., 2010; Salah et al., 2016). Therefore, as for
17 Montney Formation with permeability typically less than 0.001 mD, a horizontal well with
D

18 NW/SE wellbore direction and transverse fractures can achieve a better production performance.
19 Finally, Fig. 10 (g) presents that 49.8% of wells were fractured with slickwater. This is
TE

20 because the slickwater treatment significantly reduces gel damage within the fractures compared
21 with a typical water-based crosslinked fracturing fluid. Other benefits of slickwater treatment
EP

22 include lower costs, high fracture-network complexity, potential for improved height
23 containment (Palisch et al., 2010). It is also found that the wells with slickwater treatment have
C

24 the highest average cumulative production of 165.75 mbo. This indicates that the slickwater
25 treatment is the more preferred fracturing fluid in Montney Formation. Furthermore, Fig. 10 (i)
AC

26 compares the production performance for the wells with and without multiple perforation
27 clusters. It can be seen from this figure that, although only 23% of wells were fractured with
28 multiple perforation clusters, these wells still have a higher average cumulative production of
29 166.73 mbo, which is 28.6% higher than those without clusters. This indicates that the hydraulic
30 fracturing process with multiple perforation clusters increases the hydrocarbon production in
31 tight formations.

23
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
2
M

3 Fig. 11. Pearson correlation matrix of various variables.

4
D
TE
EP
C
AC

24
 ACCEPTED MANUSCRIPT

1 Fig. 12. Maximum horizontal stress orientations (modified from Steinhoff, 2013).
2

3 4.3. Feature selection with RFECV

4 It should be noted that Pearson correlation only can be used to evaluate the linear

PT
5 relationship between two variables. However, a meaningful relationship (e.g., a curvilinear
6 relationship) can exist even if the Pearson correlation coefficient is equal to 0. Thus, the Pearson

RI
7 correlation cannot be used as an indicator of feature importance or feature selection. It is
8 necessary to perform the RFECV process for feature selection in this work. Fig. 13 plots the

SC
9 prediction performance of the model versus the number of features employed to train the model
10 during cross validation. It should be noted that each box represents aggregated model scores for
11 nine cross-validation iterations for all 1000 model runs. It can be seen that if the number of

U
12 features is less than 4, the prediction ability of the model is significantly reduced as the decrease
AN
13 of number of features. On the other hand, when the number of features is larger than 4, the
14 increment of performance score is relatively small (<<0.01) as the increase of number of features.
15 Fig. 14 shows the importance of 11 features according to their frequency of occurrence in all
M

16 1000 model runs. It is interesting to note that, for all 1000 model runs, four features are
17 consistently selected as the input variables: latitude, longitude, well true vertical depth, and
D

18 proppant pumped per well. The first three variables determine the location of a well, which, in
TE

19 practice, defines the geological properties of the well. This indicates that the geological property
20 of the reservoir or well location is the most important factor when developing the well
performance prediction models. Moreover, our results also suggest that the mass of proppant
EP

21
22 pumped per well is one of the most important factors in predicting the first year well productions.
23 This is not surprise, because many studies also claimed that the mass of proppant is crucial to
C

24 achieve a successful hydraulic fracturing treatment. Compared to the proppant, other fracturing
AC

25 parameters such as the volume of fracturing fluid, and fracture stages are less important. This is
26 attributed to the fact that there are moderate linear relationships among these variables. For
27 example, the Pearson coefficient between the proppant and fracturing fluid is calculated to be
28 0.82, as shown in Fig. 11. This means the more proppant is pumped, the more fracturing fluid is
29 used in hydraulic fracturing process. Therefore, the effects of fracturing fluid on the production
30 have been properly considered when employing the mass of proppant to train the prediction
31 models.
25
 ACCEPTED MANUSCRIPT

PT
RI
SC
2
3 Fig. 13. Cross validation prediction performance as a function of number of features.
4

U
AN
M
D
TE

5
EP

6 Fig. 14. The frequency of features through RFECV (in percentage of number of model runs).
7
C

8 Fig. 15 depicts the holdout set prediction performance as a function of number of features.
As indicated in Fig. 3, the training data set is trimmed to match the feature set selected by
AC

9
10 RFECV. Then, the model is re-trained on the trimmed data set and the model prediction accuracy
11 is evaluated using the holdout set. Note that the feature set varies in size and composition for
12 each of the 1000 model runs. It is found that feature sets selected by RFECV are not less than 4
13 features and only 1 of the 1000 RFECV-selected feature sets composes of 5 features. The min-
14 max range of 9-fold cross validation prediction performance is also shown in this figure. In terms
15 of the mean value of , the holdout set has a slightly better prediction performance in

26
 ACCEPTED MANUSCRIPT

1 comparison with the cross-validation set. This can be expected as only 80% of the total data set
2 is employed to train the model during the cross-validation process, while the prediction models
3 are trained by 90% of the total data set for each holdout iteration. On the other hand, the
4 predictions for the holdout set have a relatively larger variance. To obtain the optimum feature
5 set for prediction models, it is necessary to select the feature set that not only maximizes the

PT
6 mean value of , but also minimizes the prediction variance. Table 3 gives the average value
7 and unbiased sample standard deviation (i.e., equation 4) for each feature set. It can be seen that

RI
8 the feature set with 6 features has the largest mean value and a relatively lower standard
9 deviation. Therefore, 6 features are selected as the input variables for the following studies.

SC
10 According to Fig. 14, this feature set consists of well latitude, longitude, well TVD, proppant
11 pumped per well, well lateral length, and fluid injected per well.

U
12 AN
M
D
TE
EP

13
14 Fig. 15. Holdout set prediction performance as a function of number of features.
15
C

16 Table 3
AC

17 Average values and standard deviation of for the holdout set.

Number of features Average value, (.'+(/' Standard deviation, 012
4 0.6366 0.0400
5 0.6784 /
6 0.6606 0.0354
7 0.6313 0.0519
8 0.6386 0.0329

27
 ACCEPTED MANUSCRIPT

9 0.6297 0.0396
10 0.6335 0.0360
11 0.6328 0.0379
1

2 4.4. Comparison of RF performance with other approaches

PT
3 Fig. 16 compares the prediction performance of the RF, AdaBoost, SVM, and NN
4 algorithms using the 6 features selected by RFECV. For each algorithm, a total of 1000 model

RI
5 runs are performed. For each model run, 90% of the total data set is used to train the prediction
6 model, and 10% of the total data set is used to test the model. The prediction performance for the

SC
7 training and testing set (i.e., holdout data set) is compared. The average mean-squared-errors
8 (MSE) of the holdout set are found to be 0.3659 for RF, 0.3851 for AdaBoost, 0.6260 for SVM,
and 0.4910 for NN (Table 4). It is found that the RF performs the best in terms of prediction

U
9
10 accuracy, which suggests that this approach is a useful and powerful tool for forecasting the first-
AN
11 year oil production in unconventional reservoirs. It should be noted that the AdaBoost method
12 also gives a comparable prediction ability compared to the RF. However, the differences of MSE
M

13 between the training set and holdout set is very large, indicating that the AdaBoost method
14 undergoes overfitting problems. Thus, it is recommended to use the RF rather than AdaBoost
D

15 algorithm to develop the production forecasting models.

16
TE
C EP
AC

17
18 Fig. 16. Comparison of prediction performance of the RF, AdaBoost, SVM, and NN algorithms.
19

28
 ACCEPTED MANUSCRIPT

1
2
3 Table 4
4 Comparison of prediction performance of the RF, AdaBoost, SVM, and NN algorithms.

Machine learning Training set (90%) Holdout set (10%)

PT
algorithms Average MSE Average R2 Average MSE Average R2
RF 0.0532 0.9468 0.3659 0.6310

RI
AdaBoost 0.0002 0.9998 0.3851 0.6131
SVM 0.4375 0.5621 0.6260 0.3736
NN 0.4242 0.5758 0.4910 0.5064

SC
5
6 Fig. 17 compares the target and predicted BOE for training set, holdout set, and ordered
7 holdout set with 90% prediction intervals by using the RF model. It should be noted that this RF
8
U
model is the optimum one of 1000 runs. Good agreement between the target and predicted BOE
AN
9 is obtained for both training and testing data set. The values of for training and holdout set are
10 calculated to be 0.95 and 0.75, respectively. As for the holdout set, 88.6% of observations (320
M

11 out of 361 data points) fall into the 90% prediction intervals, indicating that the developed
12 production forecasting model is accurate and robust for further fracture stimulation design.
D

13
TE
C EP
AC

14

29
 ACCEPTED MANUSCRIPT

1 Fig. 17. Comparison of target and predicted BOE for (a) training set; (b) holdout set; and (c)
2 ordered holdout set with 90% prediction intervals.
3

4 4.5. Case study: optimal proppant and fracturing fluid

PT
5 An important application of the production forecasting model is to optimize the fracture
6 stimulation design in field cases. In this work, four horizontal wells are chosen to be examples

RI
7 that demonstrate the successful application of the production forecasting models developed by
8 RF algorithm. Table 5 shows the information of selected horizontal wells. All wells were drilled

SC
9 along the northwest direction and completed with open-hole system. The slickwater was
10 employed to create the hydraulic fractures. Interval clusters were not used during the fracturing
11 treatment. Because the selected wells are close to each other, the true vertical depths are around

U
12 ~ 3000 meters, indicating that these wells are producing in the same reservoir layers. However,
AN
13 these wells have significantly different production performance in terms of the first-year
14 production. This can be attributed to the different hydraulic fracturing treatments. Previous study
15 indicates that the proppant and fracturing fluid pumped per well are the most important
M

16 fracturing parameters. Thus, the RF prediction model is used to optimize the proppant and
17 fracturing fluid for these four wells.
D

18
TE

19 Table 5
20 Information of four horizontal wells used in this study.
EP

Parameters Well 1 Well 2 Well 3 Well 4

Latitude 54.5626 54.5410 54.6259 54.4982
Longitude -118.5933 -118.7025 -118.8794 -118.6407
C

Well TVD (m) 2908.1 3175.7 3153.7 3196.1

Lateral length (m) 3119.5 2793.9 1698.0 2548.82
AC

Well completion type open-hole open-hole open-hole open-hole

Wellbore direction NW NW NW NW
Total proppant (ton) 4688.7 4471.8 2104.2 4117.3
Total fluid (m3) 30550.3 30130.4 16394.8 27550.1
Fracturing fluid type slickwater slickwater slickwater slickwater
Number of stages 30 28 21 24
Use of interval clusters No No No No
12-month BOE (mbo) 412.2 131.1 216.3 229.6

30
 ACCEPTED MANUSCRIPT

1
2 The production forecasting model developed by RF method can be employed to optimize the
3 mass of proppant and volume of fluid used in the hydraulic fracturing process. As mention
4 earlier, there is a strong linear relationship between the mass of proppant and the volume of
5 fracturing fluid, prior to the optimization process, such relationship is investigated by plotting all

PT
6 field cases. Fig. 18 shows the relationship between the proppant and fracturing fluid per stage. It
7 should be noted that there is a practical engineering design for the usage of proppant and

RI
8 fracturing fluid. The upper and lower bounds can be determined from these field cases. It is of
9 practical importance to use these bounds to check the optimum design obtained by the data-

SC
10 driven models.
11

U
AN
M
D
TE

12
EP

13 Fig. 18. The relationship between proppant and fracturing fluid for field cases.
14
15 Fig. 19 describes the effects of mass of proppant and volume of fluid injected per stage on
C

16 the first-year oil production for four horizontal wells. The blue circle represents the original field
AC

17 design for the corresponding well. Some interesting findings are noted. First, the mass of
18 proppant pumped per stage has the greatest influences on the production. If the mass of proppant
19 is less than 40 tons per stage, the first-year oil production is extremely low. When the mass of
20 proppant pumped per stage is increased from 40 to 200 tons, the production increases
21 significantly.

31
 ACCEPTED MANUSCRIPT

1 In addition, when the mass of proppant pumped per stage is less than 200 tonnages, the
2 volume of fluid injected per stage has less effects on the production. However, when the
3 proppant is larger than 200 tonnages, the production performance can be simply divided into
4 three different regions due to the variation of the volume of fracturing fluid. Two non-optimum
5 regions and one optimum region can be easily identified based on the volume of fracturing fluid.

PT
6 In the optimum fracture design region, the fracturing fluid ranges from 300 to 1000 m3 per stage.
7 It should be noted that, when the fracturing fluid injected per stage is larger than 1000 m3, the

RI
8 fluid leakage and water blocking around the hydraulic fractures could be the major problem in
9 tight formations, which leads to a significant reduction in the well productivity. On the contrary,

SC
10 when the fracturing fluid is less than 300 m3 per stage, proppant suspension in the fracturing
11 fluid could be a problem. If the injection volume of fracturing fluid is not enough to keep the

U
12 proppant in suspension, the proppant settles faster and cannot be transported deeper into the
13 induced fractures, which leads to more shorter and low-conductive fractures. As a result, the well
AN
14 productivity is relatively lower.
15
M
D
TE
EP

16
C
AC

17
18 Fig. 19. Optimal design of proppant and fracturing fluid per stage for four field cases.
32
 ACCEPTED MANUSCRIPT

1
2 Furthermore, it is interesting to note that the optimum design regions are located between
3 the upper and lower bounds for four horizontal wells. This indicates that the optimized the mass
4 of proppant and volume of fracturing fluid are practical and feasible to the field applications. It
5 also should be noted that the original fracturing treatment designs of four horizontal wells are

PT
6 non-optimum. This data-driven model suggests that the first-year oil production of four wells can
7 be further increased by injecting more proppant during the stimulation treatment. Given this

RI
8 data-driven model, the hydraulic fracturing design can be optimized for unconventional
9 reservoirs. It is very useful to guide the reservoir engineers when designing the hydraulic

SC
10 fracturing treatments for horizontal wells in tight formations. These data-driven models can
11 further be integrated into existing reservoir management and decision-making routines.

U
12 AN
13 5. Conclusions

14 In this paper, we have shown how to comprehensively analyze the fracture stimulation
15 design in unconventional reservoirs by using general data visualization, statistical data evaluation,
M

16 and advanced machine learning algorithms. The focus of this paper is to develop an accurate and
17 robust production prediction model for hydraulically fractured horizontal wells in
D

18 unconventional reservoirs. First, the general data visualization and statistical data evaluation are
TE

19 performed to qualitatively and quantitatively evaluate the relationships between the stimulation
20 parameters and the first-year production. It is found that open-hole system does not outperform
the cased-hole system in terms of 12-month oil production. The horizontal wells with NW/SE
EP

21
22 wellbore direction and transverse fractures are more preferred in Montney Formation. The
23 average cumulative production of wells with multiple perforation clusters is 28.6% higher than
C

24 that without clusters. Then, the recursive feature elimination with cross validation is used to
AC

25 evaluate the importance of the features in constructing the prediction models. 6 features are
26 identified as the most important variables: well latitude, longitude, well TVD, proppant pumped
27 per well, well lateral length, and fluid injected per well. After removing the redundant and
28 uncorrelated features, four commonly used supervised learning approaches including RF,
29 AdaBoost, SVM, and NN are evaluated in this study. It is found that the RF performs the best in
30 terms of prediction accuracy in comparison with other three methods. Finally, the newly
31 developed prediction models are successfully used to optimize the fracture stimulation strategies
33
 ACCEPTED MANUSCRIPT

1 for field cases. These data-driven models are found to be very useful for reservoir engineers
2 when designing the hydraulic fracture treatments in unconventional tight formations. In this
3 work, the study area only focuses on the Montney Formation. Thus, the observations and insights
4 may only be valid for Montney Formations. However, the data mining process and machine
5 learning models proposed in this work are portable and easy-to-use. It can be easily applied to

PT
6 other Formations as long as the corresponding database are available. Future studies will be
7 conducted to develop the data mining models for other unconventional formations.

RI
8

9 Acknowledgement

SC
10 The authors gratefully acknowledge the financial support from Natural Sciences and Engineering
11 Research Council of Canada (NSERC), Alberta innovates technology futures scholarship, and

U
12 Canada First Research Excellence Fund: Global Research Initiative in Sustainable Low Carbon
AN
13 Unconventional Resources. The authors also wish to express their appreciation for Well
14 Completions and Frac Database provided by geoLOGIC systems Ltd.
15
M

16 Abbreviations
D

AdaBoost Adaptive Boosting

TE

BOE Barrels of Oil Equivalent, mbo

DT Decision Tree
E East
EP

mbo Thousand Barrels of Oil

MDI Mean Decrease Impurity
C

MLP Multi-Layer Perception

MSE Mean Squared Error
AC

N North
NE Northeast
NN Neural Network
NW Northwest
RBF Radial Basis Function
ReLU Rectified Linear Unit

34
 ACCEPTED MANUSCRIPT

RF Random Forest
RFECV Recursive Feature Elimination with Cross Validation
S South
SE Southeast
SVM Support Vector Machine

PT
SW Southwest
TVD True Vertical Depth, m

RI
W West
WCSB Western Canadian Sedimentary basin

SC
1

2 Nomenclature

D
<
U
Bias value at neuron of current layer

? ∙
AN
Penalty parameter of SVM models

H
Transfer function or activation function
Number of neurons in the preceding layer
M

Number of data points

&'()*+',

5
Number of features
D

Number of prediction models

Pearson correlation coefficient
TE

Coefficient of determination
Coefficient of determination of th model run
(.'+(/'
EP

01 2
Mean value of coefficient of determination

AB Weight connected between neuron G of preceding layer and neuron

Unbiased sample standard deviation of coefficient of determination
C

8
8B Output value of node G in preceding layer
th sample
AC

> Output value of node

th target value
th predicted value

7
Mean value of observations
Standard score of the th sample

35
 ACCEPTED MANUSCRIPT

9
:
Mean value of samples

=
Standard deviation of samples
Kernel parameter of SVM models
1
2

PT
3 References

RI
4 Al-Fattah, S.M., Startzman, R.A., 2003. Neural network approach predicts U.S. natural gas
5 production. SPE Production & Facilities 18 (2), 84-91.

SC
6 Ashenayi, K., Nazi, G.A., Lea, J.F., Kemp, F., 1994. Application of an artificial neural network
7 to pump card diagnosis. SPE Computer Applications 6, 9-14.
8 Awoleke, O.O., Lane, R.H., 2011. Analysis of data from the Barnett shale using conventional

U
9 statistical and virtual intelligence techniques. SPE Reservoir Evaluation & Engineering 14
AN
10 (5), 544-556.
11 Barree, R.D., Cox, S.A., Miskimins, J.L., Gilbert, J.V., Conway, M.W., 2015. Economic
12 optimization of horizontal well completions in unconventional reservoirs. SPE Production &
M

13 Operations 30 (4), 293-311.

14 Bhattacharya, S., Nikolaou, M., 2016. Comprehensive optimization methodology for stimulation
D

15 design of low-permeability unconventional gas reservoirs. SPE Journal 21 (3), 947-964.

TE

16 Breiman, L., 2001. Random Forests. Machine learning 45 (1), 5-32.

17 Centurion, S., Cade, R., Luo, X., 2012. Eagle Ford shale: hydraulic fracturing, completion, and
18 production trends: part II. In: SPE Annual Technical Conference and Exhibition, San
EP

19 Antonio, Texas.
20 Chalmers, G.R.L., Bustin, R.M., 2012. Geological evaluation of halfway-Diog-Montney hybrid
C

21 gas shale-tight gas reservoir, northeastern British Columbia. Marine and Petroleum Geology
AC

22 38, 53-72.
23 Chollet, F., 2015. Keras, from https://github.com/fchollet/keras.
24 Clarkson, C.R., Jensen, J.L., Chipperfield, S., 2012. Unconventional gas reservoir evaluation:
25 what do we have to consider? Journal of Natural Gas Science and Engineering 8, 9-33.
26 Cortes, C., Vapnik, V., 1995. Support-vector networks. Machine Learning 20, 273-297.
27 Daneshy, A.A. 1971. True and apparent direction of hydraulic fractures. In: Drilling and Rock
28 Mechanics Conference, Austin, Texas.

36
 ACCEPTED MANUSCRIPT

1 Daneshy, A.A. 1973. A study of inclined hydraulic fractures. SPE Journal, 13 (2), 61-68.
2 Dixon, J., 2000. Regional lithostratigraphic units in the Triassic Montney formation of western
3 Canada. Bulletin of Canadian Petroleum Geology 48 (1), 80-83.
4 Drucker, H., 1997. Improving regressors using boosting techniques. In: Fourteenth International
5 Conference on Machine Learning, San Francisco, CA.

PT
6 Economides, M.J., Marongiu-Porcu, M., Yang, M., Martin, A.N., 2010. Fracturing horizontal
7 transverse, horizontal longitudinal and vertical wells: criteria for decision. In: Canadian

RI
8 Unconventional Resources and International Petroleum Conference, Calgary, Alberta,
9 Canada.

SC
10 Freund, Y., Schapire, E., 1991. A decision-theoretic generalization of on-line learning and an
11 application to boosting. Journal of Computer and System Sciences 55, 119-139.

U
12 Genuer, R., Poggi, J., Tuleau-Malot, C., Villa-Vialaneix, N., 2017. Random forests for big data.
13 Big Data Research 9, 28-46.
AN
14 Ghanizadeh, A., Clarkson, C.R., Aquino, S., Ardakani, O.H., Sanei, H., 2015. Petrophysical and
15 geomechanical characteristics of Canadian tight oil and liquid-rich gas reservoir: I. pore
M

16 network and permeability characterization. Fuel 153, 664-681.

17 Gullickson, G., Fiscus, K., Cook, P., 2014. Completion influence on production decline in the
D

18 Bakken/Three Forks play. In: SPE Western North American and Rocky Mountain Joint
19 Regional Meeting, Denver, Colorado.
TE

20 Hegeman, P., Dong, C., Varotsis, N., Gaganis, V., 2009. Application of artificial neural networks
21 to downhole fluid analysis. SPE Reservoir Evaluation & Engineering 12 (1), 8-13.
EP

22 Hughes, J.D., 2013. Energy: a reality check on the shale revolution. Nature 494, 307-308.
23 James, G., Witten, D., Hastie, T., Tibshirani, R., 2015. An introduction to statistical learning.
C

24 Springer, NY, USA.

25 Kerr, R.A., 2010. Natural gas from shale bursts onto the scene. Science 328, 1624-1626.
AC

26 Kuppe, F., Haysom, S., Nevokshonoff, G., 2012. Liquid rich unconventional Montney: the
27 geology and the forecast. In: SPE Canadian Unconventional Resources Conference, Calgary,
28 Alberta.
29 LaFollette, R.R., Holcomb, W.D., Aragon, J., 2012. Impact of completion system, staging, and
30 hydraulic fracturing trends in the Bakken formation of the eastern Willison Basin. In: SPE
31 Hydraulic Fracturing Technology Conference, Woodlands, Texas.

37
 ACCEPTED MANUSCRIPT

1 LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521, 436-444.
2 Lolon, E., Hamidieh, K., Weijers, L., Mayerhofer, M., Melcher, H., Oduba, O., 2016. Evaluating
3 the relationship between well parameters and production using multivariate statistical models:
4 a middle Bakken and Three Forks case history. In: SPE Hydraulic Fracturing Technology
5 Conference, Woodlands, Texas.

PT
6 Louppe, G. 2014. Understanding random forests: from theory to practice. PhD Thesis, University
7 of Liege, Belgium.

RI
8 Ma, Z., Leung, J.Y., Zanon, S., Dzurman, P., 2015. Practical implementation of knowledge-
9 based approaches for steam-assisted gravity drainage production analysis. Expert Systems

SC
10 with Applications 42, 7326-7343.
11 Montgomery, J.B., O’Sullivan, F.M., 2017. Spatial variability of tight oil well productivity and

U
12 the impact of technology. Appl. Energy 195, 344-355.
13 Palisch, T., Vincent, M.C., Handren, P.J., 2010. Slickwater fracturing: food for thought. SPE
AN
14 Production & Operations 25 (3), 327-344.
15 Pedregosa, F., Varoquax, G., Gramfort, A., et al. 2011. Scikit-learn: machine learning in Python.
M

16 Journal of Machine Learning Research, 12, 2825-2830.

17 Rivard, C., Lavoie, D., Lefebvre, R., Sejourne, S., Lamontagne, C., Duchesne, M., 2014. An
D

18 overview of Canadian shale gas production and environmental concerns. International

19 Journal of Coal Geology 126, 64-76.
TE

20 Rumelhart, D., Hinton, G., Williams, R., 1985. Learning internal representations by error
21 propagation. University of California, San Diego, La Jolla, Institute for Cognitive Science,
EP

22 ISC-8506.
23 Salah, M., Orr, D., Meguid, A., Crane, B., Squires, S., 2016. Multistage horizontal hydraulic
C

24 fracture optimization through an integrated design and workflow in Apollonia tight chalk,
25 Egypt from the laboratory to the field. In: Abu Dhabi International Petroleum Exhibition &
AC

26 Conference, Abu Dhabi, UAE.

27 Schuetter, J., Mishra, S., Zhong, M., LaFollette, R., 2015. Data analytics for production
28 optimization in unconventional reservoirs. In: Unconventional Resources Technology
29 Conference, San Antonio, Texas.
30 Scornet, E., Biau, G., Vert, J., 2015. Consistency of random forests. The Annals of Statistics 43
31 (4), 1716-1741.

38
 ACCEPTED MANUSCRIPT

1 Seifert, M., Geol, P., Lenko, M., and Lee, J. 2015. Optimizing completions within the Montney
2 resource play. In: Unconventional Resources Technology Conference, San Antonio, Texas,
3 USA.
4 Shaheen, M., Shahbaz, M., ur Rehman, Z., Guergachi, A., 2011. Data mining applications in
5 hydrocarbon exploration. Artificial Intelligence Review 35, 1-18.

PT
6 Shelley, B., Grieser, B., Johnson, B.J., Fielder, E.O., Heinze, J.R., Werline, J.R., 2008. Data
7 analysis of Barnett shale completions. SPE Journal 13 (3), 366-374.

RI
8 Shelley, B., Harris, P.C., 2009. Data mining identifies production drivers in a complex high-
9 temperature gas reservoir. SPE Production & Operations 24, 74-80.

SC
10 Shelley, B., Stephenson, S., 2000. The use of artificial neural networks in completion stimulation
11 and design. Computers & Geosciences 26, 941-951.

U
12 Smola, A.J., Scholkopf, B., 2004. A tutorial on support vector regression. Statistics and
13 Computing 14, 199-222.
AN
14 Steinhoff, C., 2013. Multicomponent seismic monitoring of the effective stimulated volume
15 associated with hydraulic fracture stimulations in a shale reservoir, pouce coupe field,
M

16 Alberta, Canada. MSc thesis, Colorado School of Mines, Golden, Colorado.

17 Sun, B., Chen, S., Wang, J., Chen, H., 2016. A robust multi-class AdaBoost algorithm for
D

18 mislabeled noisy data. Knowledge-Based Systems 102, 87-102.

19 Tan, F., Luo, G., Wang, D., Chen, Y., 2017. Evaluation of complex petroleum reservoirs based
TE

20 on data mining methods. Computational Geosciences 21 (1), 151-165.

21 Vapnik, V., 1998. Statistical learning theory. John Wiley & Sons, NY, USA.
EP

22 Yu, L., Liu, H., 2004. Efficient feature selection via analysis of relevance and redundancy.
23 Journal of Machine Learning Research 5, 1205-1224.
C

24 Zhou, Q., Dilmore, R., Kleit, A., Wang, J.Y., 2014. Evaluating gas production performances in
25 Marcellus using data mining technologies. Journal of Natural Gas Science and Engineering
AC

26 20, 109-120.
27

39
 ACCEPTED MANUSCRIPT

Insights to fracture stimulation design in unconventional reservoirs based on

machine learning modelling

PT
Shuhua Wang, Shengnan Chen*

Department of Chemical and Petroleum Engineering, Schulich School of Engineering,

RI
University of Calgary, AB, T2N 1N4, Canada

SC
* Corresponding author.
E-mail address: snchen@ucalgary.ca (S. Chen), shuhwang@ucalgary.ca (S. Wang).

U
AN
HIGHLIGHTS

• The comprehensive data mining process is developed to evaluate well performance in

M

unconventional tight reservoirs.

• Feature selection is performed using recursive feature elimination with cross validation.
D

• Various machine learning algorithms are compared to predict the first-year oil production.
• The usage of proppant and hydraulic fracturing fluid are optimized by data-drivel models.
TE
C EP
AC