SPE-217297-MS

Comparative Analysis Between Empirical Correlations and Time Series

Models for the Prediction and Forecasting of Unconventional Bakken Wells
Production

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

A. Laalam, Department of Petroleum Engineering, University of North Dakota, Grand Forks, ND, USA; O. S.
Tomomewo, College of Engineering and Mines Energy Studies, University of North Dakota, Grand Forks, ND,
USA; H. Khalifa and N. Bouabdallah, Department of Petroleum Engineering, University of North Dakota, Grand
Forks, ND, USA; H. Ouadi, College of Engineering and Mines Energy Studies, University of North Dakota, Grand
Forks, ND, USA; T. H. Tran and M. E. Perdomo, School of Chemical Engineering, University of Adelaide, Adelaide,
Australia

Copyright 2023, Society of Petroleum Engineers DOI 10.2118/217297-MS

This paper was prepared for presentation at the Asia Pacific Unconventional Resources Symposium held in Brisbane, Australia on 14 – 15 November 2023.

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
Accurately forecasting oil and gas well production, especially in complex unconventional reservoirs, is vital.
Leveraging advanced techniques like machine learning and deep learning is becoming more common due to
ample historical data availability. While traditional methods work for conventional reservoirs, they struggle
in unconventional scenarios. Modern machine and deep learning models excel in such challenges, offering
insights while bypassing temporary disruptions or pressure issues. This study compares ten empirical
production forecast models with state-of-the-art deep learning and time series models (ARIMA, LSTM,
GRU) in the Bakken shale play of the Williston Basin. After thorough calibration using extensive data,
model efficacy is assessed using R2-score and MSE. Results highlight well-specific performance, with no
single model consistently outperforming across all wells. Notably, optimally adjusted ARIMA produced
commendable results for many wells. This research aids reservoir engineers by simplifying production
decline trend identification, reducing reliance on intricate decline curve analyses. It ushers in a streamlined
and dependable paradigm for production forecasting.

Introduction
Oil production forecast has always stood as a critical challenge for the energy sector. Accurate forecasting
is imperative, not only for individual oil producers but also for governments, policymakers, and other
stakeholders involved in energy management (Smith et al., 2018). Among the many oil-producing regions in
the US, the Bakken Formation, extending across parts of North Dakota, Montana, and Canada, has emerged
as a significant area of interest due to its vast reserves and unconventional production methods (Boualam,
2022). The extraction and production of oil in unconventional reservoirs like the Bakken Formation present
unique challenges that demand complex geological and engineering understanding (Ouadi et al., 2022).
2 SPE-217297-MS

Unlike conventional reservoirs, where oil flows relatively freely through porous rock, unconventional
reservoirs consist of rock formations with low permeability (Wang et al., 2016). Traditional methods like
DCA have been widely used but can lead to significant inaccuracies in unconventional reservoirs (Ilk et al.,
2008). Material balance models and reservoir simulations, although more precise, may not account for the
complex fluid flow behavior and may require substantial data that may not always be accessible (Clarkson
& Pedersen, 2011).
These limitations result in discrepancies in production forecasts for unconventional reservoirs like
the Bakken Formation, with far-reaching implications. Inaccurate forecasts can lead to poor investment

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

strategies, such as over or under-allocation of resources, influencing not only individual companies but
the wider economy (Lee et al., 2015). Governments and regulatory bodies may implement policies based
on flawed data, affecting the broader energy strategy (Smith et al., 2018). Misguided predictions can
also influence market prices and create unnecessary volatility, impacting global energy markets (Wang
et al., 2016). Additionally, inefficient resource extraction due to imprecise forecasts can lead to more
extensive environmental impacts, including increased water usage and greenhouse gas emissions (Clarkson
& Pedersen, 2011).
In the wake of Industry 4.0, the utilization of machine learning (ML) and time series analysis in various
sectors has unlocked new avenues to address complex challenges (Chen et al., 2017). The petroleum
engineering domain has witnessed a surge in the application of machine learning techniques to tackle
intricate problems. Laoufi et al. (2022) delved into sand production, a prevalent issue in global oil
and gas assets, and proposed a classification approach using ML algorithms to determine the optimal
sand control method. Boualam et al. (2020) highlighted the complexities of the Three Forks Formation
and employed both traditional petrophysical models and machine learning algorithms to estimate water
saturation, emphasizing the potential of ML in capturing the intricacies of such formations. Kiran et al.
(2022) shifted the focus to geothermal exploration, underscoring the need for specialized drilling systems.
They applied machine learning and deep learning to analyze drilling operational parameters, aiming to
enhance real-time monitoring systems. Hamadi et al. (2023) presented a robust ML framework to predict
key parameters critical for CO2-enhanced oil recovery projects, showcasing the capability of ML models to
outperform traditional correlations. Leveraging these advanced techniques offers a promising approach to
enhancing the accuracy and reliability of unconventional wells’ production forecasts (Al-Anazi & Gates,
2010). ML models can digest vast quantities of data, discern hidden patterns, and adapt to changes that
conventional models might overlook (Chen et al., 2017). When coupled with time series analysis, this
synergy has the potential to revolutionize how we predict oil production (Wang et al., 2016).
The objective of this paper is to explore the integration of machine learning and time series analysis as a
novel method to improve production forecasting in Bakken oil wells. This research addresses the limitations
of the existing models in the literature and proposes a framework that can significantly contribute to the
precision of forecasts. In the following sections, we will delve into the existing methodologies, elucidate
the unique characteristics of Bakken oil wells that pose challenges for traditional forecasting models, and
present our innovative approach. Through rigorous evaluation, we aim to demonstrate the effectiveness of
our models, contributing to the wider body of knowledge and influencing future work in the realm of oil
production forecasting.

Literature Review
Oil production estimation has always been a complex task, tackled with different methodologies.
Traditional methods like decline curve analysis (DCA), well-test analysis, and numerical/analytical
reservoir stimulation have been prominent in estimating oil production rates for both conventional and
unconventional reservoirs (Chahar et al. 2022). However, these techniques face challenges in dealing with
uncertainties such as noise, missing data, and anomalies (Bao et al. 2020). Despite an R2 value exceeding
SPE-217297-MS 3

0.9 indicating high accuracy, no single model can be identified as universally superior (Chahar et al. 2022).
DCA, popular for estimating EUR (Estimated Ultimate Recovery), has its roots in the works of Arps in
1945 and further developments by Brons and Fetkovitch (Fetkovitch 1983). However, it falls short in long-
term precision, prompting Guillen Falcon (2021) to devise an algorithm utilizing production and flowing
pressure data, enhancing insights into parameters like OGIP (Original Gas in Place) and k*h (permeability-
thickness product).
Machine learning (ML) has emerged as a powerful tool for addressing these challenges. For Falcon
(2021), ensemble ML methods like Support Vector Machines (SVM) and Multilayer Perceptron (MLP)

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

tackled high-dimensional time-series problems with an impressive overall R2 of 98%. Kubota and Souto
(2019) utilized linear regression models for oil production forecasting without the need for geological
models or numerical simulators, while Temizel et al. (2022) demonstrated the LSTM model's efficacy
against a physics-based reservoir simulator. Sagheer and Kotb, (2019) used a deep long-short-term memory
(DLSTM) architecture, an advanced LSTM, showed superior predictive capabilities in actual oilfield
production data. The DLSTM model outperformed standard methods, including deep RNN, deep GRU, and
the ARIMA model. In contrast, the Deep Gated Recurrent Unit (GRU) model effectively handled complex
temporal datasets with a low-complexity structure, outperforming standard methods and demonstrating
future potential in the domain (Al-Shabandar et al., 2021).
Several studies have focused on unconventional reservoirs, where traditional methods often fall short.
Ning et al. (2022) presented a machine learning-based time series forecasting method using ARIMA,
LSTM, and Prophet, showing robust results in predicting oil production rate in unconventional reservoirs.
Abdullayeva and Imamverdiyev (2019) developed a hybrid model based on CNN and LSTM for forecasting
oil production time series. Zhang and Jia (2021) proposed a forecasting method based on multivariate time
series (MTS) and vector autoregressive (VAR) machine learning model for waterflooding reservoir, showing
more accurate results than numerical reservoir simulation. Other innovative approaches include Zhan et al.
(2019) using LSTM to tackle production forecasting, showing dynamic and non-linear event capturing. Liu
et al. (2019) used LSTM to predict oilfield production using static and dynamic developing parameters,
showing high consistency with actual production. Alimohammadi et al. (2020) developed a deep learning
approach for long-term well performance prediction in unconventional resources, using different recurrent
neural networks (RNNs) including LSTM, GRU, and Bidirectional Recurrent Neural Networks.
In summary, we can observe a growing trend towards the integration of machine learning and deep
learning techniques in oil production estimation. These methods offer promising results in handling complex
and nonlinear data, providing more accurate and robust predictions. The application of these techniques
in unconventional reservoirs and the development of hybrid models further demonstrate the versatility
and potential of machine learning in enhancing oil production forecasting. Future research may focus on
optimizing these models, integrating them with traditional methods, and exploring their applicability across
various oil and gas scenarios.

Study area
The Bakken Formation, situated within the vast expanse of northwestern North Dakota, northeastern
Montana, southern Saskatchewan, and southwestern Manitoba, represents one of the most significant
contiguous oil and natural gas reservoirs in the United States (LeFever, 1991). An intricate sequence of
black shale, siltstone, and sandstone, this formation, which dates back to the Late Devonian to Early
Mississippian era, is nested within the Williston Basin (Pitman et al., 2001). Boasting three distinct layers
- a lower shale member, a middle sandstone, and an upper shale member - each is rich in organic content,
indicative of its marine origins, and a potential source of oil and natural gas. While historically regarded as
a marginal resource due to its low permeability rock structure, technological advances in horizontal drilling
and hydrofracturing have revolutionized its potential, turning the Bakken Formation into a powerhouse
4 SPE-217297-MS

of oil and natural gas production (Clark, 2011). By 2012, the formation catapulted North Dakota to the
second-highest oil producer in the US, surpassed only by Texas (Helms, 2012). The transformation of
the Bakken Formation has not only reshaped the energy landscape but also invigorated North Dakota's
economy, ushering in unprecedented employment levels. As a testament to its abundant resources, the USGS
Assessment projected that the Bakken Formation in the U.S. holds an average undiscovered volume of 3.65
billion barrels of oil, 1.85 trillion cubic feet of natural gas, and 148 million barrels of natural gas liquids
(Gaswirth et al., 2013). Furthermore, its Canadian counterpart is hailed as one of Canada's largest oil fields.
What sets the Bakken apart from other prolific oil and gas reservoirs is not mere discovery, but the innovative

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

methods employed to unlock its potential, marking it as a cornerstone in North America's energy history.
Figure 1 presents the geographical and components of the Bakken Petroleum System.

Figure 1—Geographical localization of the Bakken Formation (Clark, 2009) and Localization
of the Middle Bakken Member within the Bakken Petroleum System (Wells, 2023)

The majority of oil and natural gas production in the State of North Dakota is derived from horizontal
fractured wells that penetrate the middle Bakken member, a layer known for holding the most proven
reserves within the formation (Chellal et al., 2022). As of June 2023, the state's active well count,
encompassing those in production, reached an impressive tally of nearly 18,000 wells. This remarkable
growth has been facilitated by the North Dakota Industrial Commission, whose commitment to transparency
and collaboration has resulted in the availability of an extensive repository of historical data. This wealth
of information has become an invaluable asset for researchers and industry experts alike. For the purpose
of this study, we meticulously collected the production history data from 15 strategically selected wells
drilled in the Bakken pool within McKenzie County. A comprehensive and detailed description of the data,
including the methodology employed in its collection and the specific attributes examined, will be presented
in the subsequent section.

Methodology and Methods

This section is dedicated to outlining the methods adopted for data collection, analysis, and interpretation. By
providing a clear roadmap of the research process, it reinforces the credibility of the study's outcomes. The
subsequent discussions will offer a comprehensive insight into the specific methodologies and techniques
that have been utilized, ensuring a transparent and thorough understanding of the research mechanics.

General Workflow
The workflow begins with data collection and integration, advancing to extraction and analytical processing.
After preprocessing the data, its integrity is verified. Any discrepancies lead to refining preprocessing steps.
Concurrently, empirical models are developed, fitted, and evaluated. If unsuccessful, these models undergo
SPE-217297-MS 5

comparison and re-evaluation. Alongside, machine learning (ML) models are constructed, preprocessed,
and trained. Their performance is gauged through a blind test on three new wells. If the ML models do
not surpass the empirical ones, alternative ML strategies are explored. The entire process concludes with
deriving insights and key takeaways, as visualized in Figure 2.

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 2—Proposed workflow of the study

6 SPE-217297-MS

Empirical Models
Empirical models for oil production forecasting utilize historical production data to predict future outputs.
Unlike models grounded in physical principles, these are based on observed production trends. They offer
simplicity and quick estimates, especially when detailed reservoir data is unavailable (Coutry et al., 2023).
In this study, 10 different empirical correlations are evaluated against the production data of the 12 Bakken
wells. Table 1 provides the mathematical equation of each model.

Table 1—Mathematical formulation of the ten empirical models used in this study

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Model Equations Source

Exponential

(1)

Hyperbolic

Arps models (2) Arps (1945)

Harmonic

(3)

Weibull's model (4) Weibull (1951)

Matthews-Lefkovits Model (MLM) (5) Matthews and Lefkovits (1956)

Hsieh model (6) Hsieh et al. (2001)

Logistic Growth Model (LGM) (7) Clark et al. (2011)

Extended Exponential Decline Curve Model

(8) Zhang et al. (2015)
(EEDCM)

Fractional Decline-Curve (FDC) (9) Zuo et al. (2016)

Variable Decline Modified Arps (VDMA)

(10) Gupta et al. (2018)
model

Time Series Models

Time series models for oil production forecasting analyze sequential data points collected over intervals
to predict future outputs. Leveraging patterns and trends observed over time, these models are particularly
adept at capturing temporal dependencies and seasonality in production data. While they can provide
nuanced insights into production behavior, their efficacy is contingent on the consistency and granularity
of the historical data. In this research, a thorough evaluation was conducted on three distinct time series
models, culminating in the introduction of a fourth hybrid model that synergistically combines the strengths
of LSTM and ARIMA.
ARIMA (AutoRegressive Integrated Moving Average). ARIMA is a foundational method for time series
forecasting, combining autoregressive (AR), moving average (MA), and differencing (I) components. The
model is denoted by ARIMA (p, d, q), where:
SPE-217297-MS 7

• p: Order of the AR part.

• d: Differencing order to make data stationary.

• q: Order of the MA part.

This combination allows ARIMA to handle various time series structures, capturing linear relationships
and seasonality. As Ayeni & Pilat (1992) found this method to be more accurate than the traditional decline
curve method in estimating crude oil reserves and predicting future behavior of oil and gas wells.

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

LSTM (Long Short-Term Memory). LSTM, a variant of recurrent neural networks (RNNs), excels in
recognizing patterns in time series. It was utilized to predict oil field production at ultra-high water cut
stage, and the prediction results were good, which verifies the versatility of the method (Wang et al., 2020).
It uses gates to control the flow of information:
(11)

(12)

(13)

(14)
Where it, ft, gt, and ot are the input, forget, cell, and output gates respectively. W and b are the weights
and biases for each gate.
GRU (Gated Recurrent Unit). GRU, an RNN variant, captures long-term dependencies using two gates:
(15)

(16)
Where rt and zt are the reset and update gates respectively, and W and b are gate-specific weights and
biases. Cheng & Yang (2021) show that LSTM and GRU have their respective advantages under different
input parameters, providing an effective method for dynamic prediction of oil well production.

Evaluation Metrics
The R2-Score, or Coefficient of Determination, quantifies the proportion of variance in the dependent
variable that's explained by the independent variables in the model. It offers insight into how well the
observed outcomes are replicated by the model, with a higher R2-score indicating that a larger proportion
of the variance is accounted for. Ideally, an R2-score of 1 suggests perfect prediction, while a score of 0
implies the model merely predicts the mean of the dependent variable.
(17)
Where:

• SSR = Sum of Squares of Residuals

• SST = Total Sum of Squares

Mean Squared Error (MSE) is a metric that gauges the average squared difference between the actual
observed outcomes and the predictions made by the model. A low MSE indicates a model that closely fits
the observed data, with a value of 0 indicating absolute fit.
8 SPE-217297-MS

(18)

Where:

• yi = Actual value

• = Predicted value
• n = Number of observations

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Data collection and description
The data consist of the oil production history of 15 unconventional oil wells drilling in the middle Bakken
formation in the North Dakota part of the Williston Basin. The data span over at least 12 years of production.
The wells were divided between training and testing wells (12 wells), and 3 wells were saved for blind testing
the machine learning models. Table 2 represents a statistical description of the testing and validation wells.

Table 2—Statistical Description of the training and testing dataset (12 wells)

Well Count Mean Std Min 25% 50% 75% Max

1 116.0 60.2178 90.8315 0.0000 19.2052 32.5949 62.4979 614.3333

2 116.0 79.7365 156.1646 0.0000 19.5968 32.3988 63.3468 1248.6667

3 116.0 91.3758 141.6748 0.0000 31.5747 43.1667 84.4627 948.0000

4 120.0 51.4712 59.4389 11.2500 20.1161 28.5548 51.2097 375.8387

5 104.0 54.5742 45.9494 0.0000 24.3214 33.0308 65.6056 199.6774

6 120.0 84.1749 126.0911 0.0000 25.5441 40.1290 73.4734 732.2333

7 134.0 77.3712 72.9922 0.0000 39.9191 56.0468 85.2486 534.3667

8 139.0 46.5576 55.5095 0.0000 20.4082 31.2903 54.8667 528.7273

9 141.0 27.2582 33.0421 1.3333 11.6000 15.7667 26.2727 250.3333

10 127.0 47.7623 86.1496 0.0000 18.8709 28.0968 45.8498 884.5000

11 144.0 50.5467 64.2723 0.0000 20.8917 29.8083 41.4774 468.2500

12 137.0 67.2278 86.6062 0.0000 22.7667 39.6333 67.7333 598.3548

The graph of Figure 3 represents the oil rate of 12 different wells over a span of 12 years, from 2011 to
2023. As is common with many oil wells, there's an observable decline in production rates over time. This
decline can be attributed to the depletion of reservoir pressure and the drop in the driving forces responsible
for pushing oil to the surface. The observed decline is rapid and sharpens just after few months of production
for some wells, which is typical for unconventional reservoirs.
SPE-217297-MS 9

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 3—Oil Rate (bbl/day) versus Time (in years) for the 12 wells within the training and
testing dataset. Distinct colors delineate the decline curve associated with each individual well.

Most of the wells exhibit a consistent decline, suggesting that they might be tapping into similar
geological formations or reservoirs with similar pressure and saturation characteristics. The depletion of
reservoir pressure over time is a natural phenomenon in the lifecycle of an oil well, and it's evident
from the general trend of the graph. However, some wells exhibit a unique behavior. They start at a
moderate production rate, experience a sharp increase, followed by a decline, and then another noticeable
increase. Such fluctuations can be attributed to several factors. The well might have tapped into multiple
reservoirs with different pressure regimes, or perhaps certain interventions or stimulations like refracturing
were applied to enhance production (Merzoug et al., 2022). Another possibility is that the well might
have faced operational issues initially, which were later rectified, leading to a spike in production. The
subsequent decline and rise might indicate changing reservoir conditions or again, operational interventions.
Additionally, we can observe that toward the end of the production period, it's noticeable that the production
rates of the wells seem to converge. This could indicate that the wells are reaching their economic limit,
where the production rate is too low to justify the operational costs. Alternatively, the reservoirs might be
nearing their ultimate recovery, leading to similar decline rates for all wells.

Modeling, Results and Discussion

Testing of Empirical Correlations
The ten empirical models outlined in the preceding section were evaluated using the production data from
12 wells. The performance of each model varied across the wells. To maintain brevity and clarity, we
will highlight the results from the top-performing model for each well in Figure 4. For a comprehensive
breakdown of the results for all models, please consult Appendix A.
10 SPE-217297-MS

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 4—The fitting results of the best performing model for each well

For a comparative analysis of the models, we employed a ranking system grounded on the R2-score and
MSE for each well. Figure 5 presents a heatmap derived from this evaluation, offering a visual guide for
selecting the most fitting model for the production data across the majority of Bakken oil wells.
SPE-217297-MS 11

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 5—The heatmap showing the ranking of the ten tested model for each well based on the R2-score and the MSE

The performance of various models reveals distinct patterns. Notably, the Arps hyperbolic, LGM, Hsieh,
and MLM models consistently rank among the top five across most wells, underscoring their robustness.
However, the Hsieh model presents an intriguing behavior, securing the top rank in two wells while
underperforming in several others. This variability suggests that while certain models may excel in specific
scenarios, their efficacy can vary across different well conditions.
12 SPE-217297-MS

Time Series Models Development

ARIMA Model. As described previously, the ARIMA model is a powerful tool when dealing with
sequential data. Before any modeling, the ADF (Augmented Dickey-Fuller) test is employed to assess the
data's stability. The ADF test is a common statistical procedure used to determine whether a time series is
stationary or contains a unit root, which can influence its predictability. If the data proves unstable, necessary
adjustments are undertaken.
For each well, an ARIMA model is constructed and meticulously fine-tuned to ensure optimal
performance. The effectiveness of each model is gauged using metrics such as R² and MSE, with the

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

predicted results being visually compared to actual data. Any deviations between the model's forecasts and
the real outcomes are closely examined. Following this, predictions from all individual models are averaged
to yield a consolidated outcome. This combined prediction's accuracy is then contrasted with individual
model outputs and the original data.
To wrap up the methodology, models are harnessed to predict the next 24 months of production, which
are visually presented and systematically tabulated.
Building and Evaluating the Individual Well's Models. To ensure accurate forecasting, the data's
stationarity was first verified. The ADF test was applied to each of the 12 wells’ production data. The
findings revealed a notably low p-value, almost zero, and an ADF statistic more negative than the benchmark
critical values at 1%, 5%, and 10% confidence levels. These metrics suggest that the time series data for
all wells is stationary.
The model was the run using default hyperparameters (1,0,1). The order (1,0,1) implies the following:

• AR (1): AutoRegressive term. The model uses one previous value (lag of 1) in the series to predict
the current value.
• I (0): Integrated. The data does not require differencing to achieve stationarity, implying that it's
already stationary.
• MA (1): Moving Average term. The model uses one lagged forecast error term to improve the
accuracy of the predictions.
Table 3 represents the metrics of the ARIMA model with the default hyperparameters.

Table 3—The evaluation metrics for the default ARIMA model (1,0,1) applied to all wells

Well R2-Score (%) MSE

1 35.55 95.0079

2 77.68 135.1705

3 76.57 84.4156

4 75.21 24.5100

5 84.25 20.0026

6 45.93 106.1245

7 77.59 27.6164

8 80.95 14.7150

9 87.13 4.3076

10 80.00 16.3718

11 90.47 11.5657

12 87.35 24.6229
SPE-217297-MS 13

Figure 6 showcases the efficacy of the default ARIMA model applied to Well 7. This particular well was
chosen as a representative example to demonstrate the model's capability to accurately track production
trends, even amidst pronounced and unpredictable fluctuations.

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 6—Fitting curves of the default ARIMA Model applied on Well 07

The table and accompanying graph highlight that the initial model results could be improved. To enhance
accuracy, hyperparameter tuning was employed to identify the optimal AR, I, and MA parameters for each
well. The results are shown in Table 4 and Figure 7.

Table 4—The results of the hyperparameters tuned ARIMA model for each of the 12 wells

Well Order R2-Score (%) MSE

1 (1, 0, 3) 41.95 85.5788

2 (1, 0, 1) 77.68 135.1705

3 (2, 0, 0) 76.69 83.947

4 (1, 0, 0) 75.21 23.5798

5 (3, 0, 3) 85.55 18.3502

6 (3, 0, 3) 54.61 92.0396

7 (2, 0, 3) 78.14 26.9399

8 (2, 0, 2) 87.90 9.3467

9 (2, 0, 3) 88.85 3.7317

10 (3, 0, 0) 81.62 15.0423

11 (2, 0, 1) 91.59 10.2007

12 (2, 0, 3) 88.03 23.2909

Figure 7 shows the prediction results for Well 11, who witnessed the best match among all wells.
14 SPE-217297-MS

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 7—Prediction results for Well 11 using the hyperparameters tuned ARIMA Model

Individual models were tailored for each well to capture its unique production characteristics. However,
to enhance the forecasting robustness and be in line with the inherent methodology of ARIMA, an approach
of averaging the predictions across all individual models was adopted. This strategy, rooted in the principle
of mitigating anomalies and noise from individual datasets, harnesses the collective predictive power of
all models.
Building and Evaluating the Combined ARIMA Model. To create a cohesive prediction model for oil
production across various wells, we'll first extract forecasts from each well's ARIMA model, optimized
using the best hyperparameters identified earlier. Ensuring these predictions are date-aligned and span a
consistent time frame is crucial. We then compute an average across all well predictions to produce a
comprehensive oil production forecast. This consolidated model will be assessed using key evaluation
metrics such as R2-score and MSE. Finally, through graphical representation, we'll juxtapose the actual data
against both individual and averaged model forecasts to gauge their comparative efficacy (see Figure 8
below).
The plot provides a clear visual representation of the model's predictions in relation to the actual oil
production values. The red line, representing predicted values, closely mirrors the blue line of original
values, capturing the overarching trend of oil production across all wells. However, noticeable deviations
between the two lines suggest that while the model is adept at grasping the general production pattern,
it occasionally struggles with the nuances of individual well behaviors or external factors influencing
production. This observation is further substantiated by the evaluation metrics, where an R² score of 0.923
underscores the model's proficiency in fitting the data, even though there's room for refinement in capturing
specific well dynamics.
SPE-217297-MS 15

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 8—The plot shows the original (blue) and predicted (red) values
for the combined model, which averages the predictions across all wells.

Building the Forecast. Utilizing the general ARIMA Model, a 24-month forecast of the oil rate was
conducted. The results for Well 11 can be observed in Figure 9. For detailed outcomes from all wells,
reference is made to Figure B1 in Appendix B.

Figure 9—Predicted and Forecasted production using the averaged ARIMA model across all wells

LSTM Model. The workflow forecasts oil production using an LSTM (Long Short-Term Memory) neural
network. Data from multiple wells is preprocessed, scaling the production values and creating sequences
suitable for LSTM training. The data is then divided into training and testing sets (80 % and 20 %
respectively). An LSTM model, consisting of a single LSTM layer followed by a dense layer, is trained
16 SPE-217297-MS

on these sequences (Table 5). After training, the model predicts and forecasts oil production for each well.
The results, including original production data, LSTM predictions, and forecasts, are visualized in a plot
for each well.

Table 5—Architecture parameters of the LSTM model

Layer (type) Output Shape Param #

lstm_2 (LSTM) (None, 50) 10400

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

dense_2 (Dense) (None, 1) 51

The evaluation metrics of the LSTM model for each well, namely R2-score and MSE, are listed in Table 6.

Table 6—Evaluation metrics of the LSTM model for the 12 wells

Well R2-Score (%) MSE

1 74.14 206.5858

2 80.16 214.3468

3 82.62 272.2508

4 84.86 89.5891

5 86.02 135.9011

6 42.88 352.8993

7 81.43 196.6190

8 79.68 79.1423

9 93.47 8.6498

10 78.18 68.6394

11 95.92 19.4773

12 87.14 128.1885

Average 80.54167 147.6908

While many wells display commendable R2-score values, the elevated MSE values for most are
concerning. This suggests that while the model adeptly captures the overall trend, it struggles to precisely
predict specific values with high accuracy.
Figure 10 showcases the prediction and forecast outcomes derived from the LSTM model. Well 11
stands out with the most commendable performance, though several other wells also demonstrate impressive
results, as detailed in Table 06. Comprehensive results for the remaining wells can be found in Figure B-2
in Appendix B.
SPE-217297-MS 17

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 10—Example of the performance of the LSTM model (well 11)

GRU Model. The presented workflow conducts a correlation analysis and predicts oil production using a
GRU (Gated Recurrent Unit) neural network. Initially, the oil rate values are scaled between 0 and 1 using
the MinMaxScaler. To structure the data for the GRU model, sequences or windows of data are generated
from the scaled oil production values. Each sequence is defined by a window size of 12 consecutive data
points, with the subsequent data point serving as the target.
The sequences are then reshaped to meet the GRU input specifications and divided into training and
testing datasets, adhering to an 80-20 split. A GRU model is constructed with the following specifications:

• A single GRU layer with 50 units, using the ReLU activation function.

• An input shape determined by the window size, which is 12 in this case.

• A dense output layer with a single unit.

The model is compiled using the Adam optimizer and Mean Squared Error (MSE) as the loss function.
Finally, the model undergoes training on the training data for 200 epochs, with a batch size of 16, and
utilizes 20% of the training data for validation to monitor performance and mitigate overfitting. The results
are reported in Table 7.

Table 7—Evaluation metrics of the GRU model for the 12 wells

Well R2 Score MSE

1 70.27% 62.9429

2 87.63% 8.5724

3 77.57% 39.0952

4 76.04% 106.6373

5 45.68% 132.4569

6 45.10% 63.6655

7 60.69% 145.7908

8 71.19% 40.1555
18 SPE-217297-MS

Well R2 Score MSE

9 83.28% 35.7439

10 77.15% 9.1912

11 86.59% 29.1996

12 82.98% 47.3963

Average 72.01% 60.0706

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 11 displays the prediction and forecast results obtained using the GRU model. Well 02 emerges
as the standout, exhibiting the most exemplary performance. However, numerous other wells also present
notable outcomes, as delineated in Table 06. In-depth results for the other wells are provided in Figure B3
in Appendix B.

Figure 11—Example of the performance of the GRU model (well 02)

Blind Testing. The blind test workflow systematically applied a suite of pre-trained ARIMA models to new
well data, emphasizing ensemble techniques for improved prediction accuracy. Through meticulous data
preparation, individual model application, and ensemble averaging, the methodology offered both current
predictions and 24-month forecasts for each well. This comprehensive approach, complemented by visual
assessments and in-depth analyses, showcased the potential of ensemble modeling in forecasting unseen
well behaviors, setting the stage for the subsequent blind testing discussion. Please refer to Figure 12 and
Table 8 for the results of the blind test for the three models.
SPE-217297-MS 19

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 12—Results of the blind test on three unseen wells using ARIMA, LSTM and GRU models

Table 8—Evaluation metrics of the blind test on the three unseen wells for ARIMA, LSTM and GRU models

ARIMA R2-Score LSTM R2-Score GRU R2-Score

Well ARIMA MSE LSTM MSE GRU MSE
(%) (%) (%)

Well 1 91.33 17.1512 15.21 188.1743 24.92 179.9759

Well 2 69.06 53.5619 43.61 72.4420 55.49 76.8644

Well 3 83.92 28.3193 34.47 117.0963 54.13 88.1785

The results of the blind testing confirm the superiority of the ARIMA model compared to the LSTM and
GRU models. With a higher R2-score and a lower MSE, we can observe a pretty good fit of the decline
curve by this model. The LSTM model did not deliver the expected results, with two observable gaps at
the beginning and ending of the production period, low R2-score and high MSE, explaining the necessity
for a better tuning of its hyperparameters.

Comparison between the ARIMA Model and ARPS Hyperbolic Model

In forecasting oil production, the ARIMA model was utilized, with individual well predictions optimized
through hyperparameters. These predictions were synchronized by date and subsequently averaged to offer
20 SPE-217297-MS

a holistic view across all fields. Evaluation metrics were extracted from this consolidated data (Table 9).
The accompanying plots (refer to Figure 13) present a visual juxtaposition of the original values with the
averaged predictions from both the top-performing empirical model and the leading time series model.

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure 13—Original vs Predicted values for the averaged data across all wells. (a) ARIMA model (b) Arps hyperbolic model

Table 9—Comparison of the evaluation metrics of the Arps hyperbolic

model vs the ARIMA model on the averaged dataset across all wells

Metrics ARPS Hyperbolic Model ARIMA Model

R -Score
2 83.29% 92.30%

Mean Squared Error (MSE) 330.93 171.61

As seen in the above table, the superiority of the ARIMA model over the ARPS Hyperbolic model is
evident with an impressive R2- score of 92.30%. This high score, combined with reduced MSE values,
highlights the ARIMA model's enhanced precision in capturing data variability and offering reliable
predictions. Although the ARIMA's data-driven nature ensures adaptability, the empirical basis of the ARPS
Hyperbolic model remains relevant in specific scenarios. These results are further corroborated by the graphs
presented in Figure 13.

Discussion
• One of the salient observations was the variability in performance across different wells. This
suggests that while empirical models have been traditionally reliable for production forecasting,
their efficacy might be contingent on well-specific characteristics in unconventional reservoirs.
This also applies to the time series model, where the performance also varies from one well
to another. For example, the ARIMA model, the best performing model, had relatively low
performance in well 1 compared to other empirical and time series models
• The production data from the Bakken shows clear complexity. Well 6 consistently had lower
performance across all models, with an R2-score always below 50%. This is due to the highly
variable production data for this well.
• The emphasis on highlighting the top-performing model for each well is crucial. It accentuates the
idea that no one-size-fits-all model exists and that each well might require its tailor-made predictive
approach. The comparison heatmap offers a visually intuitive and expedient means to ascertain
the best-suited model for any given well. The results further reiterate that model performance is
SPE-217297-MS 21

intricately tied to well-specific attributes. This underscores the need for reservoir engineers to
approach forecasting with a flexible toolkit, ready to deploy different models based on the unique
characteristics of each well.
• The incorporation of the Augmented Dickey-Fuller (ADF) test before modeling reflects the
importance of data stationarity in time series analysis. Non-stationary data can lead to unreliable
and spurious results, and thus, ensuring stationarity is paramount.
• The ARIMA model, renowned for its adeptness with sequential data, showcased its versatility
in our study. When optimally calibrated, the ARIMA model outperformed several empirical

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

models, underscoring its potential as a robust tool for forecasting in unconventional reservoirs. By
averaging predictions from individual ARIMA models, we aimed to achieve a more generalized
forecasting tool. The comparative assessment between this consolidated prediction, individual
model outputs, and actual data provides a holistic understanding of the model's performance
spectrum.
• Blind testing results unequivocally highlight the ARIMA model's dominance over the LSTM
and GRU models. The ARIMA model, with its superior R2-score and reduced MSE, adeptly fits
the decline curve. In contrast, the LSTM model's performance was suboptimal, evidenced by
noticeable discrepancies at both the onset and conclusion of the production period. Its diminished
R2-score and elevated MSE underscore the need for refined tuning.
• In the production prediction model of the LSTM model (and in less degree, the GRU model), two
distinct gaps without predicted production values are observed. The first gap, at the beginning of
the prediction curve, is a direct result of the model's architecture. LSTM models require a sequence
of data as input to predict subsequent values. As a result, the first 12 months of data were used as
the input window to predict the production from the 13th month onwards. Therefore, no predictions
were made for these initial months. The second gap, found at the end of the original data before
the forecasting starts, is associated with the length of the data used for predictions.

Conclusion
In this comprehensive exploration into unconventional Bakken wells production forecasting, a rigorous
comparative analysis was undertaken, setting ten empirical production forecast models against a selection of
state-of-the-art time series models. A notable observation was the distinct variability in the performance of
empirical models across the 15 Bakken wells studied. Historically, while these empirical models, especially
the Arps hyperbolic model, have been instrumental in providing valuable insights and proving most suitable
for a majority of the Bakken wells, this variability underscores the necessity for a more nuanced approach
when selecting models, tailoring them to specific well characteristics. Among the time series models
assessed, the ARIMA model stood out. When fine-tuned optimally, the ARIMA model demonstrated
superior predictive capabilities for a significant majority of the wells, often surpassing the performances of
several empirical models. This study also highlighted the significance of a robust, adaptive framework. The
implementation of the Augmented Dickey-Fuller (ADF) test to verify data stability prior to the ARIMA
modeling process showcases a forward-thinking, data-driven methodology. This strategy not only bolsters
the model's robustness but also its reliability, positioning it as a powerful instrument in the toolkit of reservoir
engineers.
It's important to note that incorporating well-specific conditions can significantly enhance the accuracy
of results for each well. Observations have shown that many time series models struggle to capture certain
events in the production history that deviate from the general data trends. To address this issue, adding more
detailed information about each well, including factors like lateral length, perforation details, and hydraulic
fracturing specifications, can enable the models to become more adept at detecting and predicting these
unique events, ultimately improving their overall performance. However, it's essential to acknowledge that
22 SPE-217297-MS

these additions may introduce greater complexity into the modeling process, and a more intricate and time-
consuming tuning process may be required to ensure that the models continue to deliver accurate results.
In conclusion, while it's evident that no single model has emerged as the unequivocal best for forecasting
unconventional Bakken wells, this research firmly advocates for the synergistic use of both empirical and
time series models. By harnessing the unique strengths of these two model types in combination, industry
stakeholders stand to benefit from more precise and consistent production forecasts. This achievement
symbolizes a noteworthy stride toward a data-centric future in the oil and gas sector. This research and its
workflow lay a solid foundation for future studies aimed at enhancing production forecasts not only for

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Bakken wells but also for other shale plays worldwide. The insights and methodologies developed here
can serve as a valuable reference point for researchers and practitioners seeking to optimize production
forecasting in the dynamic field of shale oil and gas exploration.

Acknowledgment
The authors wish to express their gratitude to the North Dakota Industrial Commission (NDIC) for
financially supporting this work and for providing access to real field data.

References
1. Abdullayeva, F. J., & Imamverdiyev, Y. (2019). Development of Oil Production Forecasting
Method based on Deep Learning. Statistics, Optimization and Information Computing, 7(4).
https://doi.org/10.19139/soic-2310-5070-651
2. Al-Anazi, A., & Gates, I. D. (2010). A support vector machine algorithm to classify lithofacies
and model permeability in heterogeneous reservoirs. Engineering Geology, 114 (3-4), 267–277.
https://doi.org/10.1016/j.enggeo.2010.05.004
3. Alimohammadi, H., Rahmanifard, H., & Chen, N. (2020). Multivariate Time Series Modelling
Approach for Production Forecasting in Unconventional Resources. Paper Presented at the
SPE Annual Technical Conference and Exhibition, Virtual, October 2020. Paper Number:
SPE-201571-MS. https://doi.org/10.2118/201571-ms
4. Al-Shabandar, R., Jaddoa, A., Liatsis, P., & Hussain, A. J. (2021). A deep gated recurrent neural
network for petroleum production forecasting. Machine Learning With Applications, 3, 100013.
https://doi.org/10.1016/j.mlwa.2020.100013
5. Ayeni, B., & Pilat, R. (1992). Crude oil reserve estimation: An application of the autoregressive
integrated moving average (ARIMA) model. Journal of Petroleum Science and Engineering, 8,
13–28. https://doi.org/10.1016/0920(92)90041-X.
6. Bao, Anqi, Eduardo Gildin, Jianhua Huang, and Emilio Jose Rocha Coutinho. 2020. "Data-
Driven End-To-End Production Prediction of Oil Reservoirs by EnKF-Enhanced Recurrent
Neural Networks." SPE Latin American and Caribbean Petroleum Engineering Conference
Proceedings 2020-July. doi: 10.2118/199005-MS
7. Boualam, A. (n.d.). Impact Of Stress On The Characterization Of The Flow Units In The
Complex Three Forks Reservoir, Williston Basin. UND Scholarly Commons. https://
commons.und.edu/theses/2839
8. Boualam, A., Rasouli, V., Dalkhaa, C., & Djezzar, S. (2020). ADVANCED PETROPHYSICAL
ANALYSIS AND WATER SATURATION PREDICTION IN THREE FORKS, WILLISTON
BASIN. Paper Presented at the SPWLA 61st Annual Logging Symposium, Virtual Online
Webinar, June 2020. Paper Number: SPWLA-5104. https://doi.org/10.30632/spwla-5104
9. Chahar, Jaiyesh, Jayant Verma, Divyanshu Vyas, and Mukul Goyal. 2022. "Data-
Driven Approach for Hydrocarbon Production Forecasting Using Machine Learning
SPE-217297-MS 23

Techniques." Journal of Petroleum Science and Engineering 217:110757. doi: 10.1016/

J.PETROL.2022.110757.
10. Chellal, H. A., Merzoug, A., Rasouli, V., & Brinkerhoff, R. (2022). Effect of rock elastic
anisotropy on hydraulic fracture containment in the bakken formation. All Days. https://
doi.org/10.56952/arma-2022-0777
11. Chen, Y., Zhang, D., & Zhang, J. (2017). Data-driven hybrid method for tight oil well production
performance prediction and optimization. Applied Energy, 195, 497–506. https://doi.org/10.1016/
j.apenergy.2017.03.051

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

12. Cheng, Y., & Yang, Y. (2021). Prediction of oil well production based on the time series model
of optimized recursive neural network. Petroleum Science and Technology, 39, 303 - 312. https://
doi.org/10.1080/10916466.2021.1877303.
13. Clark, A. J. (2009). Determination of Recovery Factor in the Bakken Formation, Mountrail
County, ND. In SPE Annual Technical Conference and Exhibition. Society of Petroleum
Engineers. https://doi.org/10.2118/133719-STU
14. Clark AJ, Lake LW, Patzek TW (2011) Production forecasting with logistic growth models.
OnePetro. https://doi.org/10.2118/144790-MS
15. Clark, A. (2011). The Bakken Boom: A Modern-Day Gold Rush. The Oil Drum.
16. Clarkson, C. R., & Pedersen, P. K. (2011). Production data analysis of unconventional gas wells:
Review of theory and best practices. International Journal of Coal Geology, 87 (2), 66–81.
https://doi.org/10.1016/j.coal.2011.05.009
17. Coutry, S., Tantawy, M. A., & Fadel, S. (2023). Assessing the accuracy of empirical decline
curve techniques for forecasting production in unconventional reservoirs: a case study of
Haynesville, Marcellus, and Marcellus Upper Shale. Journal of Engineering and Applied Science,
70 (1). https://doi.org/10.1186/s44147-023-00233-5
18. Duong, Anh N. 2011. "Rate-Decline Analysis for Fracture-Dominated Shale Reservoirs." SPE
Reservoir Evaluation & Engineering 14 (03):377–87 doi: 10.2118/137748-PA.
19. Gaswirth, S.B., Marra, K.R., Cook, T. A, Charpentier, R.R., Gautier, D.L., Higley, D.K., Klett,
T.R., Lewan, M.D., Lillis, P.G., Schenk, C.J., Tennyson, M.E., and Whidden, K.J., 2013,
Assessment of undiscovered oil resources in the Bakken and Three Forks Formations, Williston
Basin Province, Montana, North Dakota, and South Dakota, 2013: U.S. Geological Survey Fact
Sheet 2013–3013, 4 p., https://pubs.usgs.gov/fs/2013/3013.
20. Guillen Falcon, Gerardo E. 2021. "Flowing Material Balance and Decline Curve Analysis: A
Hybrid Approach for Better EUR and Decline Profile Predictions for CSG Wells." SPE/AAPG/
SEG Asia Pacific Unconventional Resources Technology Conference, Asia Pacific URTeC 2021
610–36. doi: 10.15530/AP-URTEC-2021-208381.
21. Gupta I, Rai C, Sondergeld C, Devegowda D (2018) Variable exponential decline: modified
arps to characterize unconventional- shale production performance. SPE Reservoir Eval Eng
21:1045–1057. https://doi.org/10.2118/194005-PA
22. Hamadi, M., Mehadji, T. E., Laalam, A., Zeraibi, N., Tomomewo, O. S., Ouadi, H., & Dehdouh,
A. (2023). Prediction of key parameters in the design of CO2 miscible injection via the
application of machine learning algorithms. Eng, 4 (3), 1905–1932 https://doi.org/10.3390/
eng4030108
23. Helms, L. (2012). North Dakota Department of Mineral Resources Monthly Report.
24. Hsieh, Frank S., Cecilia Vega, and Lucy Vega. 2001. "Applying a Time-Dependent Darcy
Equation for Decline Analysis for Wells of Varying Reservoir Type." Proceedings - SPE Rocky
Mountain Petroleum Technology Conference, Petroleum Technology - Leads the Way 139–48.
doi: 10.2118/71036-MS.
24 SPE-217297-MS

25. Ilk, D., Rushing, J. A., Perego, A. D., & Blasingame, T. A. (2008). Exponential vs. Hyperbolic
Decline in Tight Gas Sands—Understanding the Origin and Implications for Reserve Estimates
Using Arps’ Decline Curves. SPE Annual Technical Conference and Exhibition. https://
doi.org/10.2118/115258-MS
26. J ARPs, By J. and Member Aime. 1945. "Analysis of Decline Curves." Transactions of the AIME
160 (01):228–47 doi: 10.2118/945228-G.
27. Johnson, R., & Bustin, R. M. (2006). Coalbed methane potential of the Canadian
Cordillera coalfields. Bulletin of Canadian Petroleum Geology, 54 (2), 100–115. https://

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

doi.org/10.2113/54.2.100
28. Kiran, R., Dansena, P., Salehi, S., & Rajak, V. K. (2022). Application of machine learning and
well log attributes in geothermal drilling. Geothermics, 101, 102355. https://doi.org/10.1016/
j.geothermics.2022.102355
29. Kubota, Leonardo, and Fabio Souto. 2019. "Application of Machine Learning to Oil Production
Forecast under Uncertainties-The Linear Model." Offshore Technology Conference Brasil 2019,
OTCB 2019. doi: 10.4043/29883-MS.
30. Laoufi, H., Megherbi, Z., Zeraibi, N., Merzoug, A., & Ladmia, A. (2022). Selection of sand
control completion techniques using machine learning. All Days. https://doi.org/10.56952/
igs-2022-118
31. Lee, T. K., Dennis, S. A., & Simlai, P. (2015). The effect of oil production on bank deposits in
North Dakota's Bakken formation. In Research in finance (pp. 1–16). https://doi.org/10.1108/
s0196-382120150000031001
32. LeFever, J. A. (1991). History of oil production from the Bakken Formation, North Dakota.
Williston Basin Symposium.
33. Liu, W., Liu, W. D., & Gu, J. (2019). Petroleum Production Forecasting Based on Machine
Learning. ICAIP ’19: Proceedings of the 2019 3rd International Conference on Advances in
Image Processing, November 2019, Pages 124–128. https://doi.org/10.1145/3373419.3373421
34. Matthews CS, Lefkovits HC (1956) Gravity drainage performance of depletion-type reservoirs in
the stripper stage. Transact AIME 207:265–274. https://doi.org/10.2118/665-G-P
35. Merzoug, A., Chellal, H. A., Brinkerhoff, R., Rasouli, V., & Olaoye, O. (2022). Parent-Child
Well Interaction in Multi-Stage Hydraulic Fracturing: A Bakken case study. All Days. https://
doi.org/10.56952/arma-2022-0188
36. Ning, Y., Kazemi, H., & Tahmasebi, P. (2022). A comparative machine learning study for time
series oil production forecasting: ARIMA, LSTM, and Prophet. Computers & Geosciences, 164,
105126. https://doi.org/10.1016/j.cageo.2022.105126
37. Ouadi, H., Mellal, I., Chemmakh, A., Djezzar, S., Boualam, A., Merzoug, A., Laalam, A.,
Mouedden, N., Khetib, Y., Rasouli, V., & Tomomewo, O. (2022). New approach for Stress-
Dependent permeability and porosity response in the Bakken formation. Day 3 Wed, October 05,
2022. https://doi.org/10.2118/210104-ms
38. Pitman, J. K., Price, L. C., & LeFever, J. A. (2001). Diagenesis and fracture development in the
Bakken Formation, Williston Basin: Implications for reservoir quality in the middle member.
U.S. Geological Survey Professional Paper. https://doi.org/10.3133/pp1653
39. Sagheer, Alaa, and Mostafa Kotb. 2019. "Time Series Forecasting of Petroleum Production
Using Deep LSTM Recurrent Networks." Neurocomputing 323:203–13 doi: 10.1016/
j.neucom.2018.09.082.
40. Smith, L., Boschee, P., LeFever, J., & Helms, L. (2018). North Dakota's Bakken Oil and Gas
Reservoirs: A Review. North Dakota Geological Survey, 45(1), 1–8.
SPE-217297-MS 25

41. Temizel, Cenk Temizel, Uchenna Odi, Nouf Al-Sulaiman, Karri Reddy, Dike Putra, Mesut
Yurukcu, Hakki Aydin, and Cengiz Yegin. 2022. "Production Forecasting in Conventional Oil
Reservoirs Using Deep Learning." SPE Western Regional Meeting Proceedings 2022-April:26–
28. doi: 10.2118/209277-MS.
42. Valkó, Peter P. 2009. "Assigning Value to Stimulation in the Barnett Shale: A Simultaneous
Analysis of 7000 plus Production Hystories and Well Completion Records." Society of
Petroleum Engineers - SPE Hydraulic Fracturing Technology Conference 2009 386–404. doi:
10.2118/119369-MS.

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

43. Wang, G., Su, Y., & Li, X. (2016). A new model for the prediction of liquid production from
shale gas wells. Fuel, 181, 478–491. https://doi.org/10.1016/j.fuel.2016.05.016
44. Wang, H., Mu, L., Shi, F., & Dou, H. (2020). Production prediction at ultra-high water cut stage
via Recurrent Neural Network. Petroleum Exploration and Development, 47, 1084–1090. https://
doi.org/10.1016/s1876(20)60119-7.
45. Wells, B. (2023). First North Dakota Oil Well. American Oil & Gas Historical Society. https://
aoghs.org/petroleum-pioneers/north-dakota-williston-basin/
46. Zhan, C., Sankaran, S., LeMoine, V., Graybill, J., & Mey, D. S. (2019). Application of machine
learning for production forecasting for unconventional resources. Proceedings of the 7th
Unconventional Resources Technology Conference. https://doi.org/10.15530/urtec-2019-47
47. Zhang, He, Martin Cocco, Dean Rietz, Adam Cagle, and John Lee. 2015. "An Empirical
Extended Exponential Decline Curve for Shale Reservoirs." Proceedings - SPE Annual Technical
Conference and Exhibition 2015-January:4118–40. doi: 10.2118/175016-MS.
48. Zhang, R., & Jia, H. (2021). Production performance forecasting method based on multivariate
time series and vector autoregressive machine learning model for waterflooding reservoirs.
Petroleum Exploration and Development, 48(1), 201–211. https://doi.org/10.1016/
s1876(21)60016-2
26 SPE-217297-MS

Appendix A
Empirical Models Results

Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023

Figure A1—Average R2-score and overall ranking of empirical models

Table A1—Empirical Models Ranking for Each Well

Well Rank 1 Rank 2 Rank 3 Rank 4 Rank 5 Rank 6 Rank 7 Rank 8 Rank 9 Rank 10

Ar-Hyp MLM Ar-Har Weibull's LGM Hsieh FDC Ar-Exp EEDCM VDMA
1
(85.54%) (85.54%) (85.14%) (85.08%) (84.77%) (82.61%) (74.79%) (71.65%) (−44.33%) (−44.33%)

LGM MLM Ar-Hyp Ar-Har VDMA Weibull's Hsieh Ar-Exp FDC EEDCM
2
(97.88%) (97.86%) (97.85%) (97.82%) (97.76%) (97.76%) (95.93%) (88.93%) (86.80%) (00.00%)

Ar-Hyp MLM Ar-Har LGM Weibull's Hsieh Ar-Exp FDC EEDCM VDMA
3
(92.13%) (92.13%) (92.06%) (91.75%) (91.49%) (88.69%) (83.73%) (76.23%) (00.00%) (−41.96%)

LGM Ar-Hyp MLM Weibull's VDMA Ar-Har Ar-Exp Hsieh FDC EEDCM
4
(86.07%) (80.04%) (80.04%) (77.77%) (77.78%) (77.44%) (76.80%) (72.34%) (40.82%) (00.00%)

LGM Ar-Hyp Ar-Exp Weibull's MLM Hsieh Ar-Har FDC EEDCM VDMA
5
(89.83%) (77.60%) (77.50%) (77.49%) (75.06%) (70.05%) (69.83%) (33.32%) (00.00%) (−142.43%)

Ar-Hyp Hsieh Weibull's VDMA LGM MLM Ar-Exp Ar-Har FDC EEDCM
6
(51.15%) (51.82%) (51.19%) (51.19%) (51.18%) (51.05%) (50.45%) (49.37%) (45.68%) (−44.93%)

Ar-Hyp MLM Ar-Har Weibull's VDMA LGM Hsieh Ar-Exp FDC EEDCM
7
(88.68%) (88.68%) (87.55%) (85.50%) (85.49%) (83.58%) (77.94%) (71.24%) (63.78%) (18.71%)

Hsieh LGM Ar-Hyp MLM VDMA FDC Weibull's Ar-Har Ar-Exp EEDCM
8
(96.04%) (95.73%) (95.02%) (95.01%) (95.47%) (95.19%) (91.78%) (86.48%) (64.16%) (−70.86%)

Ar-Hyp MLM Ar-Har Weibull's Ar-Exp FDC EEDCM VDMA LGM Hsieh
9
(71.23%) (71.23%) (70.18%) (67.99%) (66.42%) (25.10%) (00.00%) (00.29%) (−12.18%) (−67.73%)

Hsieh LGM FDC Ar-Hyp MLM Weibull's Ar-Har Ar-Exp EEDCM VDMA
10
(96.56%) (96.37%) (96.04%) (95.81%) (95.81%) (89.24%) (87.50%) (66.18%) (00.00%) (−28.08%)

LGM Weibull's VDMA Ar-Hyp MLM Ar-Har Hsieh Ar-Exp FDC EEDCM
11
(96.87%) (96.68%) (96.68%) (96.61%) (96.61%) (96.58%) (93.62%) (85.32%) (83.27%) (00.00%)

Ar-Hyp MLM Ar-Exp Ar-Har FDC EEDCM LGM VDMA Weibull's Hsieh
12
(67.85%) (67.76%) (66.17%) (64.44%) (20.04%) (00.00%) (−29.62%) (−60.00%) (−60.00%) (−60.00%)
 Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023
27

Figure B1—Original, Predicted and Forecasted values by the ARIMA models for the 12 wells
Times Series Models Results
Appendix B
SPE-217297-MS
 Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023
SPE-217297-MS

Figure B2—Original, Predicted and Forecasted values by the LSTM model for the 12 wells
28
 Downloaded from http://onepetro.org/SPEURCE/proceedings-pdf/24SA02/2-24SA02/D021S007R001/3306562/spe-217297-ms.pdf by Colin Jordan on 12 November 2023
29

Figure 14B3—Original, Predicted and Forecasted values by the GRU model for the 12 wells
SPE-217297-MS