See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/343206534

Production Forecasting in Shale Reservoirs Using LSTM Method in Deep

Learning

Conference Paper · July 2020

DOI: 10.15530/urtec-2020-2878

CITATIONS READS
22 2,375

8 authors, including:

Cenk Temizel Celal Hakan Canbaz

Saudi Aramco Climpach
180 PUBLICATIONS 727 CITATIONS 52 PUBLICATIONS 559 CITATIONS

SEE PROFILE SEE PROFILE

Onder Saracoglu Dike Putra

Flux Energy Solutions RBC - PLC - UIR
22 PUBLICATIONS 109 CITATIONS 30 PUBLICATIONS 186 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Celal Hakan Canbaz on 27 July 2020.

The user has requested enhancement of the downloaded file.

URTeC: 2878

Production Forecasting in Shale Reservoirs Using LSTM Method in

Deep Learning

Cenk Temizel*1, Celal Hakan Canbaz2, Onder Saracoglu3, Dike Putra4, Ali Baser3, Tomi
Erfando5, Shanker Krishna6, Luigi Saputelli7, 1. Aramco, 2. Ege University, 3. METU, 4.
Rafflesia Energy, 5. Universitas Islam Riau, 6. Pandit Deendayal Petroleum University,
7. Frontender Corporation
Copyright 2020, Unconventional Resources Technology Conference (URTeC) DOI 10.15530/urtec-2020-2878

This paper was prepared for presentation at the Unconventional Resources Technology Conference held in Denver, Colorado, USA,
20-22 July 2020.

The URTeC Technical Program Committee accepted this presentation on the basis of information contained in an abstract
submitted by the author(s). The contents of this paper have not been reviewed by URTeC and URTeC does not warrant the
accuracy, reliability, or timeliness of any information herein. All information is the responsibility of, and, is subject to corrections by
the author(s). Any person or entity that relies on any information obtained from this paper does so at their own risk. The information
herein does not necessarily reflect any position of URTeC. Any reproduction, distribution, or storage of any part of this paper by
anyone other than the author without the written consent of URTeC is prohibited.

Abstract
Predicting EUR in unconventional tight-shale reservoirs with prolonged transient behavior is a
challenging task. Most methods used in predicting such long-term behavior have shown certain
limitations. However, long short-term memory (LSTM) – an artificial recurrent neural network (RNN)
architecture used in deep learning – has proven to be well-suited to classifying, processing, and making
predictions based on time series data with lags of unknown duration between important events. This study
compares LSTM and reservoir simulation forecasts.
Available unconventional tight-shale reservoir data is analyzed by LSTM and predictions obtained. A
reservoir simulation model based on the same data is used to compare the LSTM forecast with results
from a physics-based model. In the LSTM forecasting, any operational interferences to the well are taken
into account to make sure the machine learning model is not impacted by interferences that do not reflect
the actual physics of the production mechanism on the behavior of the well.
The forecasts from the LSTM machine learning model and the physics-based reservoir simulation model
are compared. The LSTM model shows a good level of accuracy in predicting long-term unconventional
tight-shale reservoir behavior using the physics-based reservoir simulation model as a benchmark. An
analysis of the comparison shows that the LSTM machine learning model provides robust predictions
with its long-term forecasting capability. This allows for better data-driven forecasting of EUR in
unconventional tight-shale reservoirs. A detailed analysis is done using the forecast results from LSTM
and the reservoir simulation model, and the key drivers of the EUR response are evaluated and outlined.
Deep learning applications are limited in the oil and gas industry. However, it has key advantages over
other conventional machine learning methods; especially where relationships are in time and space and
not very clear to the modeler. This study provides a detailed insight into deep learning applications in the
oil and gas industry by using LSTM for long-term behavior prediction in unconventional shale reservoirs.
URTeC 2878

Introduction
Over the years, unconventional hydrocarbon resources have essentially contributed to the global supply of
oil and gas. These resources include shale oil and gas, tight oil and gas, coal bed methane and gas
hydrates. Though they are bound to be the most important sources to meet the rising energy demand,
unconventional hydrocarbon reservoirs demand ground-breaking approaches towards exploration and
completion techniques that are capable of achieving economically viable production of hydrocarbons.
This is primarily due to their characteristic features such as heterogeneity and low permeability along
with anisotropic petrophysical and geomechanical characteristics (Ghanizadeh et al., 2015; Figueiredo et
al., 2017; Gu et al., 2008; Hanson et al., 1981). Shale and tight reserves were once thought of as
unreachable formations but have now been unlocked effectively thanks to advanced techniques like
directional/horizontal drilling, multistage hydraulic fracturing or a combinative approach. The rate of
production of hydrocarbons from these reservoirs have increased significantly over the years (Aloulou
and Wagener, 2018; Hughes, 2013a, 2013b).
Throughout the whole stages involved in the process of exploration and development, an assessment of
future production is critical while making decisions. During the initial phase of exploration, an appraisal
of hydrocarbons in place and the resources that are recoverable is very crucial in determining whether to
advance to field development. During the initial phase of development, surface facilities and sales
contracts are quantified on the basis of production rate estimations. Estimates of reserves and production
rates are very necessary while the fields are being passed on from one developer to another. Production
estimate has substantial impact on planning the development of field and evaluating the economic
aspects.
Numerical reservoir simulations have been used to assess the performance in terms of production from
horizontal wells that are hydraulically fractured (Wilson and Durlofsky, 2013; Yu et al., 2014). However,
it is a tiresome and time-consuming process when one tries to estimate the multi-well pad performance in
the field condition. In order to correctly model the flow of fluid in the neighbourhood of wells that are
fractured hydraulically, higher stages of grid refinement are employed.
Numerical reservoir simulation process to establish oil rate forecast encompasses construction of
geological and numerical models along with history matching curves (Clarkson et al., 2015; Kalra et al.,
2018; Temizel et al., 2018) which further depends on formation data, rock-fluid data and many more.
Moreover, this process of building accurate numerical models is wearisome and consumes a lot of time
(Nwaobi and Anandarajah, 2018). The flow characteristics of hydrocarbons beneath can also be described
using analytical models, that may at times depend on time-consuming and expensive physical
investigatory studies. The representation of the reservoir using physical and mathematical models bring in
a whole lot of assumptions that pertain to boundary states and other physical parameters such
compressibility factors, capillary forces, thermal effects (Zhang et al., 2019; Clarkson and Qanbari, 2016;
Du et al., 2017) to name a few. The scope of applicability of analytical approaches that are bounded by
assumptions gets reduced.
Decline curve analysis (DCA) is an extensively used method in the field to forecast the future production
owing to its fastness and simplicity as compared to numerical reservoir simulation. This is an empirical
regression method applied on production data and does not depend on core and fluid data (including
experimental results), seismic and logging data. It makes predictions on the production in future (of the
well that has drawn interest) on the basis of databases that are available on the public domain. The
preliminary concepts of DCA was initially put forward by Arnold and Anderson (1908) and Arnold
(1923) while the same was later on modified additionally by Arps (1945). With respect to shale reservoir,
several techniques of DCA have been put forth (Robertson, 1998; Ilk et al., 2008; Valko 2009; Seshadri
and Mattar, 2010; Duong, 2011; Clark et al., 2011; Can and Kabir, 2012; Zhang et al., 2016; Zuo et al.,
2016; Hu et al., 2018). The shortcomings of these techniques have also been discussed (Paryani et al.,
2016; Zuo et al., 2016; Tan et al., 2018). The prevailing techniques of DCA take the production history
URTeC 2878

alone into consideration and therefore cannot describe the alterations in the production due to field
activities such as well shut-in (Lee et al., 2019). While conducting the regression, irregularities in
production rates (unforeseen hike or drop) are left out. Significant variations in the production history
would require the initial point of the analysis to be determined again. The process of fitting the curve for
each individual well while conducting the analysis on an unconventional reservoir (that includes several
numbers of wells due to high-density drilling operations) is time-consuming and tiresome.
Machine Learning (ML) techniques have advanced (LeCun et al., 2015; Littman, 2015; Ghahramani,
2015) and progressively reached out to well completion datasets, production datasets and well stimulation
datasets at ease thereby revolutionizing the approaches towards production forecasting and analysis in the
oil and gas industry (Shelley et al., 2008; Awoleke and Lane, 2011; Akbilgic et al., 2015; Ma et al., 2015;
Wang and Chen, 2019; Temizel et al., 2019). In the process of reservoir description by way of estimating
petrophysical features like porosity, permeability, fluid saturation etc. using seismic attributes and well
log data, ML techniques have drawn substantial focus (Lim and Kim, 2004; Ahmadi, 2015; Zhao et al.,
2015). Researchers have employed algorithms that concentrate on extracting features (Canchumuni et al.,
2017; Lee et al., 2018) and algorithms that effectively target reservoir uncertainty through clustering
analysis (Sheidt and Caers, 2007; Lee et al., 2017; 2019).
The applicability of ML or deep learning (DL) techniques in various domains of hydrocarbon exploration
and production sector have become quite popular drawing a massive deal of attention as well as interest
(Alkinani et al., 2019). These techniques have been employed by researchers in activities such as to
predict PVT (pressure-volume-temperature) properties (Gharbi et al., 1999; El-Sebakhy et al., 2007),
optimize the process of production and placement of wells (Hassani et al., 2011; Shi 2013; Nwachukwu et
al., 2018; Ahmadi and Bahadori, 2015), conduct sensitivity analysis and perform history matching (Aifa,
2014; Montgomery and Sullivan, 2017; Guo et al., 2018a; 2018b) and forecast production (He et al.,
2001; Shaheen et al., 2011; Ahmadi et al., 2017; Ahmadi et al., 2015; Velasco et al., 2016; Chakra et al.,
2013; Li et al., 2013; Aizenberg et al., 2016; Panja et al., 2018; Wang et al., 2019; Sagheer and Kotb,
2019).
ML techniques have become a key point of interest in the oil and gas industry owing to its efficiency in
enhancing the decision-making process and operational performances. In unconventional hydrocarbon
plays, ML tools can be employed to predict the productivity of the well and to improve the process of
hydraulic fracturing operation design, for instance, in terms of fracturing fluid volume estimation and
proppant tonnage estimation. The extremities involved in interpreting the huge datasets that has been
acquired is too bulky and the process is overwhelmingly time-consuming and practically impossible for
humans in terms of direct interpretability. The ML algorithms are capable enough to scientifically unknot
the mutual associations between the independent and dependent variables.
To enhance the success of production forecasting and analysis outcomes, it is very much necessary to
select and adopt the appropriate data science methods and practices. The aim of this study is to present a
review of various ML techniques adopted by the industry to reveal the estimates of future production
rates.

Theory and Methods

Machine Learning (ML) Technique
Machine Learning (ML), often considered as a subset of artificial intelligence, is a technique in which
analytical models are built through automated procedures. With minimal intervention from humans,
systems comprehend data, classify patterns and form decisions. The systems acquire knowledge from past
experience to avoid explicit programming requirements. ML has travelled through a long distance from
merely being showcased in science fantasies to being a dependable tool in the real world that is capable of
URTeC 2878

strengthening various features of a business or an operation. Characteristic approach of an ML technique

(Figure 1.) can be visualized as follows.
1. Raw datasets are extracted from various sources and integrated to the database that has been
identified as the target database.
2. The target data undergoes a cleaning process during which the noises are removed and issues
pertaining to duplication and inconsistency are checked for and resolved.
3. Further, the transformation of data (or cleaned data) into suitable forms takes place by way of
normalization or dimension reduction.
4. Best candidate model is chosen after subjecting the dataset to several ML algorithms.
5. The decision-making process relies on these ML models that have been developed
To successfully implement a data science project, expertise in three major areas is very much necessary,
such as a) Programming, b) Mathematics and Statistics and c) Business theme associated with the scope
of the field of application. ML algorithms are generally classified as a) Supervised, b) Unsupervised and
d) Reinforcement.

Figure 1. Flow-chart representation of the approach of Machine Learning Algorithms

Supervised Learning algorithm comprises of the outcome variable or the dependent variable that needs to
be predicted from a given set of predictor variables or independent variable. Based on these variables, a
function is generated that carries out the task of mapping the inputs to the desired outcomes. The training
process lasts until the expected level of accuracy is attained by the model on the training data. Random
Forest, Decision Tree and Gradient Boost are some of the techniques that fall under this category. In this
technique, a model is built that makes forecasts on the basis of evidence in the occurrence of uncertainty.
The algorithm accepts a given set of input data and its responses to train a model that is capable of
predicting the responses to a new set of data. Classification and Regression methods are employed by
supervised learning to build predictive models. Unsupervised Learning algorithm does not include any
outcome variable to make any predictions. This technique is used to group a population in different
clusters. K-means is an example of unsupervised learning technique. Hidden patterns or inherent
trends/structures within the data are identified by this technique. Datasets that include input data without
any characterized response or outcome are inferred by this technique. Clustering is the most popular
technique that is categorized as unsupervised learning method. This technique can be employed to train a
model to determine a better internal representation by splitting the data into different clusters. In
URTeC 2878

Reinforcement Learning technique, the machine is subjected to a specific environment that is employed
by the machine to get trained continually with the help of trial and error method. The past experience
teaches the machine to acquire the best possible information that helps in making appropriate business
decisions. Markov-Decision process is an example of reinforcement learning technique.
Deep Learning (DL) Technique
Deep Learning (DL), often well thought out to be a subset of ML, has exhibited advanced performance in
extracting massive wealth of information from vast data that brings in remarkable potential to the
industry. It has already come to prominence owing to its effectiveness in executing tasks such as speech
recognition (Graves et al., 2013a; 2013b), image recognition (Krizhevsky et al., 2012; LeCun et al.,
2015), natural language comprehension (Sarikaya et al., 2014; Blunsom et al., 2017), playing games
(Silver et al., 2016) and biomedical applications (Cao et al., 2018). Production data is basically
categorized as time-series data. Time Series Forecasting (TSF) is a technique that includes prediction of
the upcoming behaviour of a system on the basis of the data pertaining to the system’s past and present.
Particularly in reference to shale gas, acquiring enough data points ordered in time would be easy on
account of high-density drilling operations associated with its production. Data in such quantity would be
necessary to apply DL techniques. As stated earlier, traditional methods of making predictions consume a
huge amount of time and are complex. Thus, researchers have been focusing on the use of huge and
widely available quantities of historical data to carry out production forecasting accurately with the help
of techniques that deduce the stochastic relationship between the past values and future behaviour.
DL is an artificial neural network (ANN) based learning technique that is influenced by the arrangement
of neurons situated in the human brain (Bengio, 2009). This technique utilizes a multilayer arrangement
and the model that possess multiple layers (Figure 2) for processing tend to acquire knowledge on the
data representation with multiple levels of thought process.

Figure 2. Schematic representation of DL model (Xing and Du, 2018)

Recurrent Neural Network (RNN)

Recurrent Neural Network (Rumelhart et al. 1986) or RNN is an extensively used DL technique to
conduct time-series forecasting owing to its capability to effectively operate input and output
arrangements belonging to different ranges of lengths as opposed to multilayer perceptron (MLP, a
feedforward artificial neural network scheme) or convolutional neural network (CNN). The process of
choosing a particular artificial neural network (ANN) scheme, a widely established forecasting tool, is
primarily influenced by three major factors viz. the intricacy of the solution, the expected accuracy in the
forecast and the features of the data. When we take intricacy and accuracy into consideration, feed
forward neural network (FFNN) scheme dominates. In this scheme, the information travels through the
network in the forward course alone. But, as we take the data characteristics into consideration, RNN
gains the upper hand.
As in the case of human beings, it becomes necessary to comprehend a word being read from an essay on
the basis of the comprehension of the previous words. Everything is not lost and the thought process
begin from scratch. Human thoughts possess the quality of persistence. Traditional neural network
URTeC 2878

schemes do not possess this attribute and this is a major deficiency. RNN deals with this problem as they
are networks that have loops (Figure 3) that enables the information to stay or be preserved. RNN can be
considered as several replicas of the same network wherein information is passed on from one network to
the succeeding network.

Figure 3. Schematic representation of RNN (a) in rolled form (b) in unrolled form (from https://colah.github.io/posts/2015-08-Understanding-
LSTMs/)

ARIMA forecasting technique for unconventional hydrocarbon plays

Auto-Regressive Integrated Moving Average (ARIMA) method is a statistical modeling practice that is
employed to carry out time-series forecast (Box and Pierce, 1970). It is a univariate time-series prediction
technique that is based on utilizing the former values of the time-series to forecast the upcoming trend or
values. It has been in use for years, specifically to make predictions from time-series data, owing to its
predictive accuracy and generalization. Being a conventional statistical technique, ARIMA has an upper
hand over DL algorithms with respect to computational easiness and confirmed efficacy in applications to
several domains. The strong fitting capabilities of this technique make it useful in various sectors such as
electric power, transport, agriculture, economy and energy. Studies have shown that this method falls short
of sensitivity to variations in the data and is not appropriate for non-linear forecasting. Wang and Jiang
(2019) brought together linear as well as non-linear prediction techniques, ARIMA and ANN, to develop
hybrid prediction models in order to predict the monthly production of shale in Pennsylvania, Texas.
ARIMA, that has excellent linear fitting capabilities, was combined with ANN, that has capabilities to pull
out non-linear dependencies, to enhance the precision in the prediction outcomes. The technique of
hybridizing ARIMA and ANN to construct a combinatorial model for conducting time-series forecast with
better accuracy and effectiveness was proposed by Wang et al. (2013). It is worth observing the difference
between traditional techniques based on statistical analysis and the DL algorithms to optimize the method
that can be adopted to predict the production of shale gas. Interestingly, the size and type of data, when
taken into consideration, have a huge impact on the performance of forecasting method and these may give
an edge to ML and DL techniques.
Random Forest (RF)
Random Forest (Ho, 1995) or RF is an algorithm that comprises of a large quantity of individual decision
trees that function as an ensemble (collectively to achieve better performance in prediction). This ensemble
learning algorithm (Figure 4) constructs multiple number of binary decision trees, which are also termed as
Classification and Regression Trees (Breiman et al. 1984) or CART, from the training data and the
descriptive variables that are pulled out randomly from the whole training dataset. The model of the
constructed tree and the predicted values obtained from it are diverse because of the difference in the learned
dataset possessed by each tree. The dataset that is not used to build the decision trees (out-of-bag selection)
will be utilized to validate the model. Each individual unit has a different number of out-of-bag selections
URTeC 2878

in the whole decision tree forming the random forest and the classified values during selection are foretold
differently for each tree. The probability of prediction is estimated for an individual unit. The decision trees’
tendency to overfit to the training dataset is rectified by random decision trees and the accuracy of prediction
gets boosted. When the size of the forest becomes large with large number of trees, the misclassification
rate (generalized errors) reaches a limit. Randomly built data from the whole learning dataset is utilized
while recovering individual decision trees which remain unaffected by the occurrence of noise as well as
outliers.

Figure 4. Schematic representation of Random Forest Algorithm (Han et al., 2020)

Support Vector Machine (SVM)

Support Vector Machine (Cortes and Vapnik, 1995) or SVM is a widely used supervised ML algorithm that
finds its applications in problems pertaining to classification and regression. For a given set of training data,
this algorithm determines a single or a group of hyperplanes in a space having infinite dimensions that are
capable of classifying the data. Out of several hyperplanes, the best one is chosen in such a way that it
denotes the farthest distance of separation from the nearest point in the training dataset of any class. This
separation is termed as functional margin and generally, if the margin becomes large, the generalization
error (an assessment of the predictive accuracy of an algorithm when applied to formerly unseen data) of
the classifier gets reduced. This technique is efficacious in high-dimensional spaces where the data is
modelled on the basis of several attributes. The decision function is composed of a subgroup of training
data points, termed as support vectors, that stays nearer to the hyperplane and have a major impact on its
position as well as orientation. With the help of these support vectors, the classifier’s margin gets
maximized. Apart from carrying out linear classification, SVM algorithms (Figure 5) are capable of
executing non-linear classification with the help of kernel method as the inputs are indirectly mapped into
high-dimensional distinctive attribute spaces. When features outnumber samples to a great extent, kernel
functions must be chosen appropriately to evade overfitting issues and the regularization term becomes
critical.

Figure 5. Schematic representation of Support Vector Machine Algorithm (Han et al., 2020)
URTeC 2878

Gradient Boosting (GB)

Gradient Boosting (Brieman, 1997) or GB is an algorithm that generates a model for prediction as an
ensemble of forecasting models that are weak, usually decision trees. This ensemble learning method works
on the basis of the ideology that weak learners can be trained to convert them into robust learners. The
predictors that follow, or decision trees in specific, learn from the shortfalls of the previous ones. This
ensures that the observations do not possess equal probability of occurring in succeeding models.
Fascinatingly, it is guaranteed that those with the highest value of error occur most. Since, the new
predictors acquire knowledge from the faults of preceding predictors, time consumed (or the number of
iterations required) to reach the actual estimates gets reduced. However, it is necessary to indicate the
stopping conditions with utmost care to avoid issues associated with overfitting on the data employed for
training purposes. The key intention of this supervised learning algorithm (Figure 6) is to describe and
minimize a cost function or loss function on the basis of gradient descent, i.e. minimizing the function by
progressing iteratively in the direction of most abrupt descent as described by the negative slope, for
modifying the forecasts. The forecasts are updated in such a way that the residuals sum up to 0 or reaches
minimum while the estimates are satisfactorily nearest to the actual values.

Figure 6. Schematic representation of Gradient Boosting Algorithm (Han et al., 2020)

Production forecasting techniques applied on unconventional hydrocarbon resources

Liao et al. (2020) employed Random Forest (RF) technique to assess the dependency of estimated ultimate
recovery (EUR) on sensitive and crucial independent variables based on Geography, Petrophysics and
Engineering (GPE) aspects. The entire data (that includes seismic, well logging, laboratory-scale
experiments, production, well testing to name a few) was obtained from 1286 wells positioned at Cardium
tight oil formations in Canada and had more than 50 variables. The key independent variables that were
identified, on the basis of sensitivity analysis, to forecast production in Cardium formations (Canada)
bearing tight oil were well location (WL), true vertical depth (TVD), resource density (RD), stimulated
length (SL), total stage count (TSC), pumped fluid per length (PFL), pumped proppant per length (PPL),
sand concentration (SC) and injection rate (IR). Models were developed on account of various ML
algorithms and the outcomes of the prediction process were compared. Their study pointed out that the RF
algorithm made predictions with an accuracy of about 90% which is far ahead when compared with other
ML algorithms like Gradient Boosting (GB) and Support Vector Machine (SVM). The predictive model
built on the basis of RF algorithm was identified as an economically viable tool to make predictions. Their
study highlighted the feasibility and accuracy of the technique to carry out production forecast and assess
the economic viability of the project in Cardium reservoir.
Han et al. (2020) made use of ML algorithms like RF, GB and SVM to develop a robust model that can be
used to predict productivity during the early phase of production (within 6 months). The required datasets
were obtained from 150 wells, targeting shale gas, stationed at Eagle Ford shale formations. Key parameters
associated with reservoir properties, well stimulation and completion were identified as the input variables
URTeC 2878

whilst the cumulative production of gas during a span of 3 years was identified as the target variable. They
pointed out that the model based on RF algorithm performed better achieving high accuracy.
Bhattacharya et al. (2019) developed models to forecast daily production of gas accurately on the basis of
RF, ANN and SVM and compared the efficacy of each model. Apart from the geoscientific, completion
and surface-based data, the required information was extracted from the data recorded by fibre-optic
monitoring facilities like distributed acoustic sensing (DAS) and distributed temperature sensing (DTS).
They aimed at assessing the performance of a 28-stage hydraulically fractured horizontal well positioned
at Marcellus shale formations within Appalachian basin. The sensitivity analysis conducted by them during
the study highlighted the key parameters required to forecast the production of gas (because they largely
influence the production of gas) as Poisson’s ratio, casing pressure, minimum horizontal stress, gamma log
and distributed acoustic sensing energy. It was observed that RF algorithm performed better than the other
algorithms by pushing up the accuracy towards a value of 96% and pulling down the cost as well as time
required to perform the computation (Temizel et al., 2019).
Luo et al. (2019) built non-linear models using RF and Deep Neural Network (DNN) algorithms to forecast
the cumulative production of oil during a span of 6 months (Balaji et al., 2018). The whole dataset was
obtained from around 3600 wells positioned at Eagle Ford formations. Geological parameters such as
structural depth, thickness of the formation, total organic carbon (TOC), number of calcite layers and
average thickness of the layer (thickness of the formation divided by the total count of the layers) were
considered as key input parameters that influenced the productivity of wells in Eagle Ford. Additionally,
completion parameters such as proppant loading, volume of fracturing fluid and its type, TVD, measured
depth (MD), length of perforation, completion and production dates along with oil gravity represented the
key input parameters that described the productivity of wells in Eagle Ford. The performance of non-linear
models in terms of predictive capabilities were found to be better owing to the complexities and nonlinearity
involved in the dependencies between input and output variables. The R2 accuracy of the forecast made by
the best model constructed during their study was found to be 62% only and the value dropped further for
wells whose production lasted for a shorter duration. Their study highlighted the occurrence of other input
variables that could enhance the forecast of output variable (which is nothing but 6-month cumulative
BOE).
The predictive accuracy of RF algorithm was examined by Biswas (2019) on large datasets from wells
positioned in Wolfcamp shale formations within Delaware basin. The entire data was obtained from 5716
horizontal wells and had 131 predictor variables while EUR was taken as the target variable. RF algorithm
was observed to put forth better performance than optimized DL algorithm. However, DL algorithm was
expected to perform well with increased size of database. Zhong et al. (2015) had previously compared the
applicability of different machine learning techniques including RF, GB and SVM algorithms to predict the
quality of production from 476 horizontal wells positioned in Wolfcamp shale formations within Permian
basin. The target variables were cumulative production of oil in barrel during the first 12 producing months,
maximum monthly production of oil during the first 12 producing months and the obtained production
efficiency by the maximum monthly production of oil during the first 12 producing months over the entire
lateral length in bbl./ft. Parameters associated with various aspects covering the geographic location of the
wells, architecture of wells, well completion and well stimulation were considered as the input variables.
Long Short-Term Memory (LSTM)
The engineering and geosciences data pertaining to the subsurface are really huge and have been classified
as big data that is described by the three Vs: velocity which represents the real-time data obtained as in the
case of logging-while-drilling (LWD), variety which indicates the fact that the data being dealt with belongs
to different scales and come from various domains and volume which specifies the quantity of the data (its
size) that involves complexities. Building more efficient and accurate models to predict hydrocarbon
production is a complex problem that has drawn the focus of several researchers.
URTeC 2878

Long short-term memory (Hochreiter and Schmidhuber, 1997) or LSTM networks are a particular type of
RNN that is being widely used by researchers in the recent years to carry out time-series prediction studies.
There are instances when the gap between the pertinent information and the location where it is required
becomes huge making RNN incapable of understanding the process of connecting the information. LSTMs
tend to avoid issues associated with long-term dependency such that information is remembered for
expanded periods of time. This technique avoids vanishing gradient problem (Grosse, 2017), an issue
observed in training ANN with gradient based learning techniques as well as backpropagation algorithms.
Weights of each network get updated in a way that is proportional to the error function’s partial derivative
with respect to the existing weight during the training’s each iteration. At times, the gradient becomes
negligibly small and avoiding an update in the weight’s value thus stopping the training. This technique
avoids exploding gradient problem (Grosse, 2017) as well, a situation in which huge error gradients get
accumulated leading to huge modifications to the model weights at the time of training. This can cause
instabilities to the model making it uncapable to acquire knowledge from the training data. As opposed to
FFNN, LSTM (Figure 7) includes connections for feedback. It consists of three gates (input, output and
forget) to control the flow of data through the cell that recollects values over random intervals of time.
Studies conducted over the years have led to a gradual and very significant progress (Sak et al., 2014; Tai
et al., 2015) in this technique making it a widely used algorithm.

Figure 7. Schematic representation of the repeating module in an LSTM network (from https://colah.github.io/posts/2015-08-Understanding-
LSTMs/)

Review of the application of LSTM method in production forecasting

Lee et al. (2019) employed LSTM algorithm to construct a model for predicting the shale-gas production
of the future. In order to deduce the input features, shut-in period and the gas production of the past were
considered. The training dataset were obtained from 300 wells positioned at Duverney formations in
Alberta, Canada. The model was tested for 15 wells stationed in the same area. The model, that was trained,
exhibited capabilities to forecast the production rates for a longer period (55 months). They pointed out that
the method can be used to predict the production rates of the future much faster and examine the influence
of added attributes like shut-in period. It was highlighted that the method can yield more effectual and
consistent forecast of shale-gas production and that this technique can be used in both conventional and
unconventional scenario. It is worth noting that further tuning and refinement of the feature selection
process can yield a method with enhanced predictive capabilities in terms of reliability and utilization. It
was proposed to include stimulation parameters, attributes extracted from geological information and
refracturing as potential features that have a crucial effect on the production of shale-gas and that can boost
the performance and predictive capabilities of the method. The underperformance of DCA and the high-
density drilling associated with unconventional hydrocarbon resources make this method more effective in
those circumstances.
URTeC 2878

Zhan et al. (2019) tested the applicability of LSTM technique to predict the rate of oil production for a span
of 2 years or even further by utilizing very fewer past data acquired during the initial stages of production.
The required dataset was obtained from over 300 wells stationed at onshore unconventional formations.
During the initial few years of production, approx. 70% of the total EUR can be recovered from shale wells
after which a rapid decline is observed. The steepness of the decline makes it difficult to completely analyse
the trend causing over-estimation. They highlighted the importance of this technique to forecast production
and evaluate the reservoir in such complex scenario. They observed that the average difference between the
estimated and observed accumulated production stayed within 0.2% while the variance did not cross 5%.
Calvette et al. (2019) evaluated the capabilities of LSTM based model to predict the production. It was
observed that the error stayed low, thus justifying the applicability of this method instead of reservoir
simulator (Temizel et al., 2020). They adopted two case studies, a synthetic simple box reservoir and a
modified version of PUNQ S3 reservoir, to evaluate the predictive efficiency of the model. Sagheer and
Kotb (2019) tested the predictive efficacy of deep LSTM (DLSTM) network wherein more LSTM layers
were stacked to overcome the shortfalls of shallow arrangement while working with long interval time-
series data. They pointed out that the proposed method performed much better than other models used in
the study, such as those based on deep RNN, deep gated recurrent unit (GRU) and ARIMA. The models
were evaluated and verified on the basis of their applicability in two real-field case studies i.e. Huabei oil
field in China and Cambay basin oil field in India. An assessment of the model performance was carried
out on account of root mean square error (RMSE) and root mean square percentage error (RMSPE) and
DLSTM was fond to perform well with a considerable difference in the error values. It is worth noting that
the model performance was evaluated using testing data (unseen data) instead of training data.
Recently, Song et al. (2020) compared the performance of a hybrid model constructed on the basis of LSTM
subjected to particle swarm optimization (PSO) with traditional neural network and ARIMA models to
forecast the daily oil production from volcanic reservoir of Xinjiang oilfield (Canbaz et al., 2019) They
highlighted the capability of the novel approach to perceive the complexities in the pattern of variation in
the oil production rate under the influence of several factors. It was observed that, in the case of LSTM, the
predictive accuracy improved at the expense of computational time with the increase in the number of
hidden layers. The importance of setting the parameters to ensure an effective performance of the LSTM
model was pointed out during the study. The mean absolute percentage error (MAPE), mean absolute error
(MAE) and RMSE corresponding to LSTM model were found to be the least (9.88, 1.60 and 2.02,
respectively) in comparison with other models, thus proving the predictive efficiency of the model. Liu et
al. (2020) proposed an ensemble empirical mode decomposition (EEMD) based LSTM to enhance the speed
and accuracy in the oil production forecasting task. In order to assess and verify the efficacy of the model,
it was tested on two real-field cases, SJ and JD oilfields based in China. EEMD-LSTM model was compared
with EEMD-ANN and EEMD-SVM models. It was found that the EEMD-LSTM model performed much
better than the other models by providing the forecast perfectly and with great quality. Interestingly, the
values of RMSE, MAPE, MAE and R2 corresponding to the proposed model were found to be the least as
compared with the values corresponding to the other models.

Type of Decline Analysis in Production Forecasting

Currently there are some things that a drag associated with the predictions of the rate of production within
the oil and gas industry began less efficient, some of them are incorrect production estimates and poor
economic analysis, the feasibility of each project must be analyzed with the right data to avoid
miscalculation because it can lead to the calculation of income and investment from the well also be
wrong.(Iyke & Princewill, 2018). Arps (1945) developed a basis for analyzing the decline in production
in conventional reservoirs in the oil industry with using decline curve that can evaluate the estimated
ultimate recovery of a well. This model can be used efficiently in conventional formations if it has the
following assumptions:
• There were no significant changes during operation
URTeC 2878

• The well produces under BHP

• Unchanged drainage area (Boundary Dominated Flow)
• Produced in homogeneous formation and stable flow. (Sharma & Guttery, 2019)

However, the above conditions cannot be applied in an unconventional reservoir. The decline curve
model can be semianalytic or empirical in the broader DCA category. The most common DCA methods
that have historically been applied to shale reservoirs are the Arps (1944), Fetkovich (1980), Hsieh
(2001), and Blasingame and Rushing (2005) methods originally developed for conventional reservoirs
used in unconventional reservoirs, although with certain limitations. On the other hand, Ilk et al., (2008),
Valko and Lee (2010), Duong (2010), and Fulford and Blasingame (2013) were specifically developed for
unconventional shale reservoirs.(Gupta, Rai, Sondergeld, & Devegowda, 2018). So with the DCA method
if the reservoir conventional or unconventional that can forecast production and the reservoir performance
and certainly the rate of production can be estimated with different methods and models of decline as well
adjust to the condition of the reservoir.
Arps Model
The Arps method is one of the classic Decline Curve Analysis (DCA) models proposed by Arps, in which
Arps simulates a decrease in this empirical flow rate by proposing three parameters based on its exponent
hyperbolic function (Tan, Zuo, & Wang, 2018) which has been used a long time ago, which is about more
than 60 years, so this method is also often called the oldest or most traditional method in the Decline
Curve Analysis equation. Arps model which has the following equation: (Hong, Bratvold, Lake, & Ruiz
Maraggi, 2018).

𝑞0 (1 + 𝑏𝐷𝑖 𝑡)−1⁄𝑏 0<𝑏≤1

𝑞𝑡 = { } (1)
𝑞0 𝑒 −𝐷𝑖𝑡 𝑏=0

In applying the Arps model, the reservoir flow form, Boundary Dominated Flow (BDF) with bottomhole
pressure tends to be constant, has a fixed skin factor. This model applies the concept of loss ratio To
reduce the decline curve analysis model, where the two assumptions proposed by Arps, the first
assumption that the loss ratio is constant, and the second assumption is the first derivative loss ratio is
estimated to be constant. In the loss-ratio will be made a ratio among the decline in the rate of production
at the present time, with the previous time.
The Arps model is considered less accurate in adjusting the decline curve that occurs in unconventional
reservoirs and also predicts Estimated Ultimated Recovery (EUR) even though this model is known to be
quite simple and fast. This model will tend to cause an overestimation of EUR shale gas wells by
assuming that the flow form is controlled by the Boundary Laminated Flow (BDF). Whereas the shale gas
well itself, very rarely has the form of BDF flow, so that the arps model cannot be applied directly
without modification before adjusting to the unconventional reservoir characteristics. (Tan et al., 2018)
This model is divided into three based on the value of hyperbolic exponent (b), namely on decreasing
exponential hyperbolic exponent, it will be 0 (b = 0), on the hyperbolic decrease (0 <b <1), and decreasing
harmonic (b = 1), So it can be concluded that this traditional Arps model applies a method with
hyperbolic exponent (b) values, ie between hyperbolic exponent values (b) may exceed 1 in some
applications, but will produce unrealistic estimates. Then the Arps is further divided up, 0 ≤ 𝑏 ≤ 1.
In an exponential decline condition with the exponential decline curve (b=0), this straight-line relationship
can occur in 2 states, the first is when the flow rate (q) versus time (t) is plotted onto a graph on a semi-
log scale, and the second is when the flow rate (q) versus cumulative production is plotted into a
Cartesian scale chart.(T. Ahmed, 2019)
URTeC 2878

In an exponential decrease, the value of hyperbolic exponent (b) will be equal to zero. This decrease is
referred to as a constant rate decrease which has two conditions, namely the initial production and the
condition of the rate of decline. This model shows a decrease in the rate of change in flow or production
over time, which will tend to be constant (Belyadi, Fathi, & Belyadi, 2019). Exponential decline which
can also be referred to as geometric decline, semilog, or constant percentage has a tendency that the
decline in the rate of production each time will be proportional to the rate of production. (Arps, 1945).
Main Exponential decline equations can be given as below:
The equation for Rate;
𝑞 = 𝑞𝑖 𝑒 −𝐷𝑡 (2)
Equation of cumulative oil production:
𝑁𝑝 = (𝑞𝑖 − 𝑞)⁄𝐷 (3)
Nominal Decline Rate equations are given in (4) and (5) as:
𝐷 = −ln (1 − 𝐷𝑒 ) (4)
𝐷𝑒 = (𝑞𝑖 − 𝑞)/𝑞𝑖 (5)
Effective Decline Rate:
𝐷𝑒 = 1 − 𝑒 −𝐷 (6)
Life of the well is given as;
𝑡 = 𝑙𝑛(𝑞𝑖 /𝑞)/𝐷 (7)

Classification of these decline curves are given in Figure 8.

Figure 8. Decline Curve Classification

URTeC 2878

In an harmonic decline curve (b=1), in relation to the scale of semi-log curve is shown in a straight line
between the flow rate versus cumulative production, and has a curved shape when in a state of semi-log,
log-log, or on a Cartesian graph of flow rate versus time. If the hyperbolic exponent is 1, a harmonic
decline condition will occur, with a constant decrease. Harmonic curve decline has an equation that can
be seen in the following table,

Table 1. Harmonic Decline Equations, B=1 (PETROBJECTS, 2004)

Stretched Exponential Decline

SEPD is a decline curve model that is empirically similar to the Arps decrease curve model, and the
SEPD method was proposed by Valko (2009). The problem of uniqueness in the SEPD becomes a
problem that is a little difficult to solve so that Valko proposes another procedure by means of non-linear
equations to obtain parameters. Some SEPD model parameters are quite similar to the Arps model such
as, qo, n, and τ, which are empirical constants, where n in the SEPD is similar to the value of b or
hyperbolic exponent in the Arps model.SEPD produces behavioral inheritance similar to hyperbolic at the
start but eventually forms its natural boundary. (Can, Texas, Kabir, & Corporation, 2011). Which can be
seen in the following picture.
Stretched Exponential: adding bound components to impose limits on EUR developed by Valko, (2009)
which is a new variation on the continued development of the Arps model. The extended exponential
relationship between production rate and time is defined as follows:

Figure 9. Comparison of SEPD and Arps Decline Relations

URTeC 2878

The SEPD model tends to be superior to the PLE model, and there are also similarities with the PLE
model, but there will be differences in behavior at the end of time. And in equation above n is a time
exponent whereas t is a characteristic time used in the Stretched Exponential Production Decline
equation. And for the equation of the cumulative relationships in SEPD are described as follows:
(Belyadi, Fathi, & Belyadi, 2019).

(13)

(14)
Because of its bound nature, this method is more suitable for use in unconventional reservoirs. And the
model will provide EUR values to the end if n, t, and qi have positive values.
Table 2. The Stretched Exponential Production Decline Model

Each volume has an exponential decay rate, with a characteristic time constant distribution (𝜏). The
decrease in actual production is determined by how much the volume contributes to which the natural
interpretation of this model is advanced by Valko and Lee (2010). SEPD takes about 36 months to be able
to estimate parameters accurately and tend to be more suitable for transient flow than BDF. The SEPD
model also requires a complex solution in determining the unknown parameters shown by Zuo et al
(2016)The concave and convex portion of the decline curve will join one another, this is one of the
advantages shown by the model of Stretched Exponential Production Decline (SEPD) and without the
addition of the parameters needed in estimating the value of EUR proposed by (Nesheli, 2012). Zuo et
al.,(2016) show SEPD model tends to lead to transient flow compared to boundary dominated flow,
requiring a production time of more than 36 months to accurately estimate parameters. This SEPD model
also requires complex solutions in determining unknown parameters. (Ahmed, 2019)
So the conclusion obtained from Stretched Decline is
• Stretched equations cannot be relied on to predict EUR without the presence of Boundary Flow
Dominates (BDF) data.

• Although SEPD can always solve this problem, the ability to predict EUR is no better than
Power Law Decline.

• SEPD has the same level of complexity as Power Law, except if there is a solution for
cumulative production. If the decreasing coefficient is obtained manipulation of the equation
may be acceptable.

• Observing a nominal decrease with time can lead to the formulation of new equations based on
URTeC 2878

functions that provide more data curvature. (Freeborn & Russell, 2012)

Terminal / Limiting Decline Rate

Changes in the rate of hyperbolic decreases that turn to exponential decline are also known as Terminal /
Limiting Decline Rate. If the beginning of the decline effective (De) in the well is around 65%, then the
decrease can switch if the decline effective value changes, ie from a hyperbolic decline to an exponential
decline. EUR predictions get lower along with the higher terminal decline and can result so that the
transition from hyperbolic to exponential will be faster.(Belyadi et al., 2019).

Figure 10. Hyperbolic versus modified hyperbolic decline

Terminal decline that occurred between the curves in the image above is assumed at 5% where after the
effective reduction (de) reached 5%, then the remaining life of the well hyperbolic decline will turn into
exponential decline. The concept of a terminal/limiting decline rate is preceded by a hyperbolic decline
curve that undergoes a transition to an exponential decline curve at a certain limit that limits the effective
decline rate (De). A decline curve can be calculated at any time as:

(19)

The time at which change changes from hyperbolic to the exponential rate of decline can be calculated as:

(20)

This approach can be effective in predicting production performance if used with caution. However, this
is a non-unique approach that can produce reserves estimates that differ greatly with time, and / or on
stable estimates. (Alsaadoun, 2018)
Power Law Exponential (PLE)
Power Law Exponential is a decline curve used for observing the inverse loss ratio against time, which
occurs when the inverse loss ratio forms a straight line curve on a graph plot at the log-log scale. Power
law Exponential is used when the flow rate time data shows the change in power law in the inverse loss
ratio. (Odi, Bacho, & Daal, 2019). The power law exponential decline model was developed by Ilk et al,
URTeC 2878

(2008) which is a modification of the exponential decline in Arps. Powe Law exponential decline model
was developed specifically for gas-tight wells and defined in the following equation as:

(21)

Or it can be diminished to the following equation:

(22)

n is the exponent of time, D1 is the decreasing constant at a particular time, and the constant decreasing at
infinite time is D∞,, and Di is the initial rate of decline of % per year. This PLE model is more suitable to
be applied to production data in regions with transient or BDF flow without EUR overestimated.
Ilk et al., propose another approach for formulating parameter b:
(23)

(24)

(25)
The model defined by Equation 2 is named the facility law exponential decline model. this is often
supported the Arps decline curve and uses Stevens' law to approach production levels. This model was
developed specifically for shale gas wells. The exponential decline power-law model is defined in
Equation 2, where the equation model is obtained based on the Arps decline curve to achieve a production
rate that is close to, and the shale gas resources are very suitable to use this exponential decline power-
law model. There are several additional variables in the PLE model in calculating transient flow and
Boundary-Dominated flow. Parameters or variables such as qˆi, Dˆ i, D∞, and nˆ are four unknowns that
can cause differences or discrepancies in field data. The PLE model can be solved by a number of
methods such as the interpolation algorithm like the least least square method. McNeil et al. showing that
this Exponential Decline Power Law is suitable for use in production in transient or boundary-dominated
areas. Without causing an overestimation of the remaining reserves, so the resulting estimate will still be
in fair condition. Then the PLE model can also be used in linear flow, bilinear flow followed by linear
flow, and linear flow followed by Boundary Dominated Flow, or bilinear flow followed by linear flow and
ending with BDF flow, this statement was stated by Kanfar and Wattenbarger. Whereas if historical
production data 67% of the time are very suitable when using the PLE method, Paryani et al., also stated
that the PLE model gives a low estimateamong all DCA models in the study, with the most conservative
estimate (Tan, Zuo, & Wang, 2018). Many other researchers develop special correlations to predict shale-
well production performance. Ilk et al. (2008) propose a power-law loss-ratio model, and Table 4 shows
the proposed equation. D1 in the equation refers to the rate of decline after one unit of time, and D1 refers
to the rate of decline when the well has entered SRV-dominated flow. In this period, the rate of decline
becomes constant, and the PLE equation decreases to be an exponential decrease equation. The authors
propose that their model is flexible enough to model transient, transition, and Boundary-Dominated flow.
This model, like the composite-decline curve model, follows the assumption that the well enters the SRV-
dominated boundary flow at the end of time, and the nature of the reservoir is constant. In the PLE model,
there is no easy way to calculate constants like D∞, D1, and n using production data. If the initial data is
interrupted, the D1 calculation might have a sizeable error. D∞ and n can only be estimated through trial
URTeC 2878

and error. (Gupta, Rai, Sondergeld, & Devegowda, 2018).

Table 3. Decline-curve equations proposed by Ilk et. al. (2008).

Modified Power Law

The modification of the power law equation model is based on the PLE model which is an empirical
equation so that it can be used also in analytical reservoir models. This modification is divided into two
parts, namely when the flow is transient and when the Boundary dominated Flow condition. So the
Modified Power Law equation in transient conditions can be calculated with the formula as follows:
The value of n is in the range between 0.3 to -0.13. while the Boundary Dominated Flow itself begins after
the transient condition is over, and is obtained through the investigation finger equation. By following the
assumption of hyperbolic reduction with b = 0.5 under the BDF conditions for gases.

𝑞 = 𝑞̂𝑖 𝑒𝑥𝑝(−𝐷̂𝑖 𝑡 𝑛 ) (28)

Decline curve of the variable Power Law Exponential for observing the legal behavior of the inverse loss
ratio power variable against time. This inverse loss ratio versus time plot creates a curved line relationship
on the graph plot with a log-scale scale.
Table 4. Variable Power Law Exponential Formulas.

The variable Power Law Exponential is used when time production rate data shows the behavior of the
Power Law variable by inverse loss ratio to time. (Odi, Bacho, & Daal, 2019)
Extended Arps
A decline curve for observing the initial linear change in time in a derivative loss ratio followed by a
period of derivatives constant loss ratio until the next time. Used when time production rate data shows
evidence of linear changes in the derivative loss ratio followed by the constant loss ratio derivative
behavior.
URTeC 2878

Table 5. Extended Arps Model Formulas.

Duong Decline
Duong decline model is a model of DCA that empirically derived the previous model found on shale gas
wells and long-term linear flow in reservoirs. Duong decline is developed into analysis and estimation of
production from non-conventional reservoirs. Duong decline uses the reserve probability distribution in
the source of play estimation, which can represent inaccuracies in estimations and calculate reserve
estimates (Duong, 2010). Duong decline was specifically improved in analyzing non-conventional
reservoirs that have low permeability reservoirs. Shapes of a proper curve are for wells this shows the
transient flow in the long-term. This method will show the limited EUR value and tend also conservative
than the decline in the Arps method with a value of b smaller than 1 (Duong, 2010). A preventive
approach can be to estimate the EUR value of wells where the flow from the fracture will be dominant
and the contribution of the matrix can be ignored. In this method, the density of fractures connected to the
area where the fracture has occurred must increase from time to time which is useful to support the
performance of this method (Duong, 2010). Duong proposes an empirical model developed from the
observation that the log-plot of the top level of cumulative versus time production is an observation to
correspond to the straight line. The assumption behind the model is that well production is dominated by
linear or bilinear fracture flow and that wells rarely reach boundary dominated flow (Gupta et al, 2018).
The cumulative gas equation becomes:
𝑡
𝐺𝑝 = ∫0 𝑞𝑑𝑡 (35)

𝑡 1−𝑛
𝐺𝑝 = 𝑞1 (1−𝑛) (36)

The equation q and Gp can be derived as follows (Duong, 2011):

𝑎 1−𝑚 −1)
𝑞
𝑞1
= (𝑡 −𝑚 ) 𝑒 1−𝑚(𝑡 (37)

And;
𝑎 1−𝑚 −1)
𝑞1
𝐺𝑝 = 𝑎
𝑒 1−𝑚(𝑡 (38)
URTeC 2878

Q1 are the rate at 1 day, a and m are the constant value generated from log q / GP vs day when t is in one
day. Estimated a and m are originated from intercepts in y-axis and slope. Plot of t versus q and the slope
of a line shows the level at day 1 production. Intercept (q∞) equals to flow rate at time = ∞ (infinite).
These Model adapts the concept of developing Stimulated Reservoir Volume (SRV) which leads to
infinite linear flow. This concept refers to the activation of faults and fractures where changes in local
stress during depletion are the cause (Warpinski & Branagan, 1989). A constant SRV model can also
show long linear flow and is ignored if the permeability matrix is very low. The Duong equation for
production and cumulative levels is as follows:

𝑞 = 𝑞1 𝑡(𝑎, 𝑚) + 𝑞∞ (39)
And;
𝑞1 𝑡(𝑎, 𝑚)
𝐺𝑝 = (40)
𝑎𝑡 −𝑚

Where;
𝑎
𝑡(𝑎, 𝑚) = 𝑡 −𝑚 𝑒𝑥𝑝 (1−𝑚 (𝑡1−𝑚 − 1)) (41)

Some tight gas zones and shales for the production of dry gas and high liquid gas are applied to test
Duong's decline. All show predictable straight-line trends, by the intercept and slope associated with the
kind of fracture flow-regime. The performance of individual wells or EUR is reduced based on the range
of productions in the best three-month average or the level of production that can be known or estimated.
These results indicate this method is easy to apply, estimates a credible EUR, and this method is used to
replace the role of the previous method for non-conventional reservoirs. This new method can more
present methods of statistical for analyzing production estimates from resource estimates and for
establishing various forecast results, including the distribution of reserve probabilities lower plane (P90)
to a higher plane (P10) (Duong, 2010). The advantage of the Duong method is that it is simple to predict
future rates and EUR values developed to tight gas wells where dominated by unconventional reservoir
fractures, Duong method can also provide statistical methodologies to analyze source play production
forecasting and set the range of yields for this forecast, include the distribution of reserve probabilities
form P90 to P10 and EUR for this method not based on the theory of dominant boundary flow but in the
other hand limitations of a newest trend (m and q∞) with times limit and industrial levels (Duong, 2011).
The disadvantage of the Duong decline method is that the initial pressure applied in this modeling of
numerical if be based gradients of fluid perhaps inappropriate. The result of statistical modeling maybe
doesn't represent the characteristics of flow in gas shale (Duong, 2011).
Modified Duong Model
Modified duong model, to serve a preferable fit for data in BDF overall regime. Analysis of the actual
production well and the results of analytic simulations performed on the elected shale gas wells to
juxtapose the many decline curve model beforehand for ensure that an model provided the low difference
in the estimated reserves trace. Well and grouped well analyzes were performed to examine the success of
various models. The previous Duong method and these modifications present a further accurate estimate
of the reserves for data sets rather than the Arps method. This method can help to estimate production gas
quickly and also perfectly in the shale reservoir (Joshi & Lee,2013).
(Duong,2010) suggested the plot rate q versus times must form a straight-line that was revised in the
beginning but current operating conditions which might not prove so. For wells with a production history
URTeC 2878

of fewer than 18 months, forcing a line that is regressed in origin (q∞ equal to 0) refers to a minimum
error in estimate the production reserve (Joshi & Lee, 2013). Modifications made to the Duong method
are switched to the Arps BDF models after a specific speed or time. By using a value of b 0,4 is applied
for the model of Arps after the switch. (Fetkovich et al, 1996), stated that b = 0,5 must be used for gas
wells by bottom hole pressure flowing equally to nil while b value 0,4 is suitable if bottom hole pressure
is around 10% in reservoir pressure. A controversial issue for Duong modified by transfer to Arps is the
rate of decline at the time of switch (Joshi & Lee, 2013).
The figure below shows the comparison between the original Duong (with q ∞ ≠ 0), and the modified
Duong (forcing q ∞ = 0).

Figure 11. The comparison of Original Duong and the Modified Duong (Joshi & Lee, 2013)

The advantages of the Modified Duong gave more accurate the results than SEDM and Modified Arps
Model when higher than 12 months of production history data is avail, although some uncertainties are
still related to production forecasts and the modified Duong method for the reservoir approach
specifically will produce more accurate estimates (Joshi & Lee, 2013). The weakness of the modified
Duong method is that none of the models studied yields accurate estimates with historical production data
of six months or less and In the Modified Duong Model, the straight line used to determine q ∞ must be
forced through zero intercepts, mainly when production history data are avail for less than 24 months
(Joshi & Lee, 2013).
Multi-segment decline
In many companies, forecasting production using multi-segment gets its attractive power. Theoretically,
the multi-segment method is applied to observe the various flow regime in a reservoir for a certain term,
produces decrease in Arps method of three segments that allow any segment to observe different flow-
regime, include transient-flow, When value b > 1, boundary dominated flow When value b is smaller than
one and large than zero, and an decline of exponential when value b is zero. Selecting segments and
parameter b is usually based on guesses about reservoir knowledge and visual observations on slope
changes in decline (Jeyachandra et al., 2016).
URTeC 2878

Figure 12. Flow regimes and corresponding "b" values (Xiong, Gao, & Li, 2017).

Multi-segment production forecasting methods use Rate Transient Analysis (RTA), RTA is positing the
theory that pressure data and production data are signals originated in the reservoir. RTA would be used
to identifying differences in flow-regime observed in the reservoirs, and transition times between flow
regimes can be found using superposition plots. This transition time can then be reduced to a multi-
segment plot of estimated production to divide the production history accurately. In this segment, we can
use a decreasing curve that is best suited for that segment. For example, if we have accurately identified
the long-term transient flow, we can use the Duong method to represent that segment, and if we find a
strong boundary effect, we can step back in that segment and find a better value of Arps. Several case
studies are presented for conventional and unconventional reservoirs including wells in vertical and
horizontal paths, and the proposed technique is compared to Single Arps and Duong methods
(Jeyachandra et al., 2016). Flow regimes, such as radial, bilinear, linear formation, Stimulated Reservoir
Volume (SRV), and flow dominated by boundaries, and flow regime transitions could be the
identification in the RTA diagnostic chart. Transient linear flow is general flow behavior in Tight gas
wells or unconventional wells. The flow regime here is considered as a half slope in the log plot of the
rates of gas vs. time (Jeyachandra et al., 2016). (Bello & Wattenberger, 2008) conduct RTAs on wells that
show temporary linear behavior. They claim that drainage from the block of the matrix to the surrounding
fractures is a possible cause of a transient linear regime. Because of this, fractures have negligible
pressure drops which result in transient linear flow, which is consistent with the concept of dual-porosity.
(Clarkson & Pederson, 2010) investigated classical transient flow rates in multi fractures horizontal wells
with single porosity and multiple reservoirs. They concluded that the flow regime must be identified first
before choosing to apply different models and a combination of techniques is needed to reach a solution.
(Belyadi et al., 2015) also showed the importance of complexity in the interpretation of linear flow in
shale wells.
In principle, by following the flow regime to divide well performance into segments. Then we can know
the decline character for each segment (flow regime). And then find "B" and Di for each segment to
represent their decline character (Gao & Li, 2017).
URTeC 2878

Identify Flow Regime: Log-Log and Normalized Pressure vs. Sqrt time plots are used to identify the flow
regime.

Figure 13. Log-Log and Normalized pressure vs. Sqrt time plots.

Calculating B with a time: b-parameters are calculated and Log-Log graph for b vs. time is plotted.

Figure 14. Log-Log plot of b-parameters vs. time.

Divide into 3 segments (b and di): Strengths This method is suitable for unconventional reservoirs that
show a variety of flow regimes and Multi-segment is a simple, yet powerful method, to utilize RTA, to
identify changes in flow regime, and to apply Arps or other DCA techniques and will produce better
estimates (Jeyachandra et al., 2016). Arp`s hyperbolic DCA model is given in below equation.
1
𝑞 = 𝑞𝑖
(1 + 𝑏𝐷𝑖 𝑡)(1/𝑏)
(42)
URTeC 2878

Comparing Results:

Figure 15. Log-Log plot for q & Gp vs. Time (Xiong, Gao, & Li, 2017).

The multi-segment DCA methodology and workflow can help in estimating the behavior of the well and
EUR rate correctly (Xiong, Gao, & Li, 2017). Deficiencies in the multi-segment models presuppose
reservoir conduct that might don't occur in the fields, and various operation conditions could tilt the slope
of the productions rate and could cover the flow regime observation in the reservoir, this cause to
inaccurate estimates. Transient rate analysis can erase inaccurate estimates of flow transitions, assume
flow regimes in reservoirs, and remove masking effects on pressure (Jeyachandra et al., 2016).
Logistic Growth Model
Traditional empirical Arps method uses to determine reserve in conventional reservoirs often provides
more than estimated reserves when used in this very low or unconventional permeability formation.
Generally, the value of b obtained is higher than 1, The results in a un pragmatic level of that production
are nevermore close to zero. (Ilk et al., 2008) has explained the problems faced with traditional models
such as the Arps empirical method which provides inaccurate estimates in estimate reserves in reservoirs
with low permeability. A new reduction model has been suggested to improve accuracy in determining
reserves; Nevertheless, no one has obtained broad industry approval. Thus a new decline model for
forecasting productions has been made to based on very low reservoir permeability both from reservoir
oil and gas (Clark et al., 2011).
URTeC 2878

Logistic-Growth model is based on the concept that growths are only possible for certain sizes. The
maximum size of growth might be called "carrying capacity". Logistic-Growth is useful for providing
population growth models and their first adaptability in oil management (Hubbert, 1956). (Clark et al.,
2011) be expanded the logistic-growth model that models productions for a single-well instead. This
model adjusts to the basic principles of a growing population that determine that growing is probable up
to a certain size. This model includes volumetric quantities of known oil or gas toward the calculation and
limits the estimated reserves to a rational amount. Logistic-Growth has been used in various fields for
various things. In addition to population modeling, this model has modeled yeast growth, organ
regeneration, and,intrusion a recent product in marketplace (Tsoularis & Wallace,2001). The Logistics-
Growth Model used in various fields has taken a different form; However, it has been merged with
(Tsoularis & Wallace, 2001) toward a general "Logistic-Growth Model" with the following this equation:
𝛾
𝑑𝑁 𝑁 𝛽
𝑑𝑡
= 𝑟 𝑁 𝛼 [1 − (𝐾 ) ] (43)

Where;
N = Population
r = Constant
𝛼, 𝛽, 𝛾 = Exponents
K = Carrying Capacity
To estimate production in well of oil and gas, this had held changed after an analysis of empirical into
following this equation (Clark et al., 2011).

𝐾𝑡 𝑛
𝑄(𝑡) = 𝑎+𝑡 𝑛
(44)

Where;
Q = Cumulative Population
K = Carrying Capacity
a = Constant
n = Hyperbolic Exponent
t = Time

𝑑𝑄 𝐾𝑛𝑏𝑡 𝑛−1
𝑞(𝑡) = =
𝑑𝑡 (𝑎 + 𝑡 𝑛 )2
(45)
Where;
q = Production Rate
URTeC 2878

According to (Kanfar & Wattenbarger, 2012) Rate and cumulative equations are as follows:
𝐾𝑛𝑎𝑡 𝑛−1
𝑞=
(𝑎 + 𝑡 𝑛 )2
(46)
And,

𝐾𝑡 𝑛
𝑄=
𝑎 + 𝑡𝑛
(47)

Herein, “K” is a carrying capacity or recoverable resource. The advantages of this model are simple to use
and very able to being a positive data trend of existing productions and giving a make sense estimate of
future productions. This model can very accurately assess the downward trend in behavior in a well. The
ultimate recovery value can be predicted with the logistical growth model and is more a bit simpler than
the Arps model. This model gives a new alternative method to forecast productions in non-conventional
reservoirs. The drawback is that the Logistic-Growth Model cannot predict non-physical values (Clark et
al., 2011).
Type Curve Analysis
Type curves are graphical representations of theory solution of flow rates equation. This analysis includes
finding theoretical type curves which "match" the response of the reservoir and test wells when there is a
change in production or pressure. Matches are obtained graphically with the principle of superposition
from the well test data graph, in the same chart of the type of curve, and search for type of curve will
provide the best matching. Because the type of curve is a plot of a theoretical solution for the equations of
pseudo steady state and transient flow, its normally present as the dimensionless variable (Ahmed, 2018).
The purpose of Type curve analysis is to find the type of curve which "matches" with the response well
and reservoir when the test. The reservoir parameters, as permeability & skin, could be determined for
dimensionless parameters to describe the desired type of curve (Gringarten, 1987). There is some the
method of type curve : Gringarten, Agarwal-Gardner, Blasingame, Linear compound, Normalized
pressure integral, Wattenbarger, Fetkovich and Transient method where this will be discussed in another
section
Model Dari Type Curve:
Radial, a condition where the reservoir is a cylinder with a vertical well in the middle. Dimensionless
parameters are given in the following equation (Illman and Neuman, 2000):
𝜋𝑘𝐿 𝑇𝑠𝑐
𝑤𝐷 = ( )𝑤
𝑄𝑠𝑐 𝑇𝑝𝑠𝑐 𝑤
(48)
𝑘𝑡𝑝̅
𝑡𝐷 =
∅𝜇𝑟𝑤2
(49)
URTeC 2878

Figure 16. Model Dari type curve analysis (Illman, 2005)

The Figure above that explained initial time data down in straight-line follows slope, which shown
compressive water storage in spherical flow-rate model interval test. Pressure peaks are not seen in the
logarithmic scale in the figure above due to the two-phase flow condition (Illman, 2005). The behavior
slope negative unit is likewise related to the harmonic decrease. The time and time derivatives in graph
logs are very useful because of the usability features that differ during periods of transient and pss flow.
These graph likewise beneficial for determining the value of gas-in-place. Besides this type is useful for
estimate parameters of reservoir as permeability & skin effects (Agarwal et al., 1998). Fracture, Cylinder
or rectangular reservoir with infinite conductivity fracture in the middle. Formations that experience
fractures naturally are actually from heterogeneous systems of vug, fracture, and random matrix. The
fracture in parallel with the main axis is permeability. Liquids in this state can be assumed to be a slightly
compressive single-phase, pressure in the wellbore in an infinite reservoir (Engler & Tiab, 1996).
Analysis of well fractures is related to the identification of variables wells and reservoirs that have an
impact on future well performance (Ahmed, 2018).

Figure 17. Model Dari type curve analysis (Illman, 2005)

URTeC 2878

Figure 18. Infinite acting condition (Zareenejad et al., 2012).

𝑞 1 1 𝑁𝑝
∆𝑝
=𝑏 − 𝑁𝑐 (50)
𝑝𝑠𝑠 𝑓 𝑝𝑠𝑠 ∆𝑝
𝑏

Where;
∆𝑝 = (𝑝𝑖 − 𝑝𝑤𝑓 ) (51)
And,
141.2𝐵𝑜 𝜇𝑜 1 4 𝐴
𝑏𝑝𝑠𝑠 = 𝐿𝑛 ( − 2 ) (52)
𝑘ℎ 2 𝑒𝐴 𝐶𝐴 𝑟𝑤𝑎

Water Drive, Cylinder reservoir with concentric cylinder aquifers. The water drive model is based on a
radial composite model, with an outside aquifer, which is defined by different levels of mobility from
different aquifers and relative to the reservoir. The mobility ratio ranges from 0 to 10 which means from
the absence of an aquifer to an aquifer in a constant state.

Figure 19. Dimensionless Rate vs. Dimensionless Production plot (Chambers, Karra & Mortimer,1980).
URTeC 2878

The aquifer model assumes that the water zone is infinite.. Curve type analysis to predict the performance
of wells with water drive was first proposed by Blades and Stright. They found that the Gas oil ratio was
relatively insensitive to changes in relative permeability, anisotropy, and skin effects, but was very
sensitive to oil viscosity and capillary pressure. Judging from the results of several runs from a computer,
the type curves is developed with the thickness of the oil column, the total rate of liquid production,
viscosity, and capillary pressure as parameters. The typical type curves of this study are reproduced in the
figure below. This particular type of curve shows the effect of the total rate of liquid production on oil cut
and oil recovery. The fixed parameters are oil viscosity, net pay, and capillary pressure (Chambers, Karra
& Mortimer, 1980).
Horizontal, a square-shaped reservoir in the middle is found horizontal well. This analytical model was
first developed in the 1980s. This analysis horizontal well test be based on the approach on a source line
to the well with a vertical fracture that is penetrated in part from the total of the entire reservoir, assuming
that the horizontal well can be seen production well from the source in infinite acting condition in the
reservoirs. Reservoirs in this model can be horizontal, homogeneous, anisotropic with Kh and Kv is
constant, constant thickness, and constant porosity. Reservoirs can be the situation becomes infinite or
finite in lateral size. Reservoirs are limited by closed upper and lower bounds. Reservoir pressure is
constant at the start. Single-phase fluids and production rates are considered constant (Guo et al., 2012).
The mathematical model of horizontal well in the reservoir tight gas makes the pressure curve decrease
and the type reflects the infinite horizontal well pressure performance in reservoir gas. The results are
shown in the figure below (Guo et al., 2012).

Figure 20. Horizontal well pressure performance in reservoir gas. (Guo et al.,2012).

Elliptical, The elliptical reservoir has a finite conductivity fracture in the center of the well. The elliptical
focus on a dual porosity is the inner & outer regions that are clarified based on microseismic observations.

Figure 21. Pressure transient curves of elliptical flow.

URTeC 2878

Elliptical flow model uses a modified Mathieu function, Laplace transformation method, and a Stehfest
algorithm comprehensive. Wellbore pressure dimensionless curve (pwD) and derivative pressure curve
(p wD x tD / CD) associated with tD / CD are plot to understand the various flow regime presented in
following figure (Xu et al., 2015).

Figure 22. Rate transient curves of elliptical flow.

Finite conductivity fracture, This model is a highly long formed fractures formed by “massive hydraulic
fracture”. This model requires a big-hearted amount of binding material to continue open and a result the
permeability of the Kf fracture, is weaker than a infinite conductivity fracture. This fracture has limited
conductivity characterized by a pressure drop that can be measured on the fracture and hence shows a
different pressure response during well testing on hydraulic fractures (Ahmed, 2018).
Gringarten type curve
Gringarten serves to analyze a well that has wellbore storage & skin in it with a system multi porosity.
Gringarten curve applied to fractures, acidizing, and well that have fractures on the fissured reservoir.
Reservoir and well parameters, produce characteristics parameter that can present quantitative data about
the fissured volume and size of blocks porous in reservoir. Gringarten type curve is be based on the
solutions of the diffusivity radial equations & with the resulting assumption: vertically wells in steady
production rate; reservoir in infinite-acting, homogeneity type; single phase, fluid which has a slightly
compress flow; not much influenced by skin factors, and the coefficient of wellbore storage is constant.
This assumption shows that this type of curve did develop specially for drawdown tests on undersaturated
reservoirs oil. Gringarten curve is likewise beneficial for analyzing for gas well and pressure buildup
tests, These curves are the most complete and practical to use. They are also the foremost widely utilized
in refining industry literature (Bourdet, & Gringarten, 1980).
The several effective ways to classify interpretations of the Gringarten model is to use a pressure
derivative associated beside the natural log from several elapse time functions. Plot logs of pressure vs.
pressure drops elapse time produces some characteristics for various interpretation models that are easily
recognized. This feature is illustrated in the Figure below (Gringarten, 1987). This type of curve describes
the variation of P vs t for the reservoir determined from the well configuration. Determined used an
analytical model and displayed in variables without dimensionless, the analytical model used by
Gringarten to describe vertical wells in an infinite homogeneous reservoir (Daviau, 1986).
URTeC 2878

Figure 23. Dimensionless Pressure vs. Dimensionless Curve plot (Gringarten et al., 1979).

The above figure shows type of gringarten curve given from the PD vs. tDe equation. The ordinate of the
measured match point in the axis type curve system is given as “pD” and in the field data axis system:
“Δp” (Gringarten et al., 1979).
0.000264 𝑘𝑡
𝑡𝐷𝑒 = 2
∅𝜇𝑐𝑡 𝑟𝑤𝑒
(53)

And,
𝑘ℎ
𝑃𝐷 = 141.2 𝑞𝐵𝜇 ∆𝑝 (54)

Proportionality factor between pD and Δp. can be used to determine kh (horizontal permeability) in a
reservoir (Gringarten et al., 1979).
𝐷 𝑀 (𝑃 )
𝑘ℎ = 141.2 𝑞𝐵𝜇 (∆𝑃) (55)
𝑀

In the same way, the absence of a match point, M, is measured in the curve type axis system, tD / CD, and
in the field data axis system: Δt.

0.000295 𝑘ℎ
𝑡𝐷 ⁄𝐶𝐷 = 𝜇𝐶
∆𝑡 (56)

The proportionality factor between tD / CD and Δt can be used to calculate C, wellbore storage by using
below formula;
0.000295 𝑘ℎ (∆𝑡)𝑀
𝐶= 𝜇 (𝑡𝐷 /𝐶𝐷 )𝑀
(57)

Then the CD value is calculated by using the Equation 58 (Gringarten et al., 1979).
0.89 𝐶
𝐶𝑑 = ℎ∅𝑐 𝑟2 (58)
𝑡 𝑤
URTeC 2878

The values above are used to determine the skin:

1 (𝐶𝐷 𝑒𝑥𝑝(2𝑆))
𝑆 = 2 ln 𝐶𝐷
(59)

Type curves are made from the production of a steady flow rate (drawdown). These curve could be used
directly to analyze buildup test if [Δt << tp; <t << tp (n - 1) and tp (n-1)]. If this condition doesn't exist,
the use of types can effect to incorrect results (Bourdarot & Balvet, 1998). If a match with the buildup
data point and the drawdown curve is forced it will produce a type curve that is too high in the curve set
and therefore will produce inaccurate data. The buildup measurement point plotted vs Δte lie on the
drawdown curve and then the flattening effect of the buildup will disappear (Bourdarot & Balvet, 1998).
Agarwal-Gardner type curve
These type of curves are a process reducing production to be analyzed in the production of radial and
vertical oil wells and gas wells. 1970 Agarwal present a model for infinite-acting reservoirs (sans outer
boundary). Matching for this type of curve is used to estimate reservoir parameters and determine an
initial location of a semi-log for a straight line (Rosa & Horne, 1983). This curve has been developed to
produce a model that is easier to estimate gas or oil in the reservoir and to predict the value of
permeability, skin, fracture length, conductivity, etc. Agarwal analyzes production data by using a double
porosity type curve to gas and oil wells with vertical fractures condition (Alom et al.,2017). These types
of production curves represent Progress from previous methods because an obvious difference can be
created between flow matches associated with transients and boundaries. this model also provides a more
direct way to determine reserves. This production decline curve was first created specifically for
hydraulic fracturing wells, both infinite and finite conductivity (Agarwal et al.,1998).
These types can be used to wells by hydraulic fractures because they can estimate half-length/fracture
conductivity fractures. Agarwal-Gardner has arranged & presented a type of curve for investigating data
production. These method give type of curves by dimensionless variable be base in conventional wells
tests descriptions, with a match to the definition of dimensionless Fetkovich using by Blasingame et al.
This method likewise includes plots derivatives of primary and semi-log pressure. Besides, this method
presents a decline curve in an additional format to a standard normal level versus plot times (Ahmed,
2019). This type of curve has the best settlement model for calculating permeability in reservoirs with
natural fractures. Agarwal-Gardner adapted the model from Palacio and Blasingame which showed that
the rate of constant production and bottom-hole pressure can be changed to an equivalent steady liquid
solution rate (Zareenejad, Kalantari & Nasriani, 2012).

Figure 24. Forecast Agarwal-Gardner Type Curve (Ahmed, 1979).

URTeC 2878

Agarwal and Gardner type curves produce plot vs. time production rates in the log plot log. Data has been
normalized to make dimensionless units. Stems of this type of curve try to match the external drainage
radius well. For better matching (Ahmed, 2019).

Figure 25. Agarwal-Gardner Type Curve Analysis (Oil Example) (Ahmed, 1979).

Figure 26. Agarwal-Gardner Type Curve Analysis (Gas Example) (Ahmed, 1979).

In this particular case, the well shifts to boundary-dominated flow in the first few months of production
(Ahmed, 2019).
URTeC 2878

Rate-Cumulative Production Analysis

Dimensionless Type curves for Cumulative Rate production is given in below graph.

Figure 27. Rate-cumulative production decline-type curves (Agarwal et al., 1998).

Correlations fo oil well and gas well is given in Equations 60 and 61, respectively.
𝑁𝑝
𝑄𝑚 = 𝑐 (60)
𝑡 (𝑝𝑖 −𝑝𝑤𝑓 )

And,

𝐺(𝑃𝑝𝑖 −𝑃̅𝑝 ) 2𝑞𝑡𝑐𝑎 𝑃𝑖

𝑄𝑚 = = (𝑐 (61)
(𝑃𝑝𝑖 −𝑃𝑤𝑓 ) 𝑡 𝜇𝑍)𝑖 (𝑃𝑝𝑖 −𝑃𝑤𝑓 )

Rate-Time Production Analysis

Dimensionless Type curves for time dependant rate production is given in below graph.

Figure 28. Time dependant rate production decline-type curves (Agarwal et al., 1998).
 URTeC 2878

Dimensionless Type curves

Equations for dimensionless type curve analysis is given below (Agarwal et al., 1998).
For Oil Well;
𝑁𝑝
𝑡𝑐 = 𝑞
(62)

𝑞 𝑞
∆𝑝
= 𝑃 −𝑃 (63)
𝑖 𝑤𝑓

1 1
𝜕
= ∆𝑝
(64)
𝜕
𝑞
(𝜕𝑙𝑛𝑡 )
𝑐

For Gas Well;

𝜇𝑖𝐶𝑡𝑖 𝑡 𝑞(𝑡)
𝑡𝑐𝑎 = ∫ 𝑑𝑡 (65)
𝑞(𝑡) 0 𝜇(𝑃̅)𝑐𝑡 (𝑃̅)

𝑞 𝑞
=
∆𝑝𝑝 𝑃𝑝𝑖 − 𝑃𝑝𝑤𝑓
(66)

1 1
=
𝜕 ∆𝑝𝑝
𝜕
𝑞
(𝜕𝑙𝑛𝑡 )
𝑐𝑎

(67)

Agarwal and Gardner type curve analysis method uses the following models:
i. Radial, The cylindrical reservoir that has a well-located vertically in the middle. Dimensional
parameters are given in the equation (Agarwal et al., 1970).
ii. Fraktur, Cylinder or rectangular reservoir with infinite conductivity fracture in the middle (Agarwal et
al., 1970).
iii. Water Drive, Cylinder reservoir with concentric cylinder aquifers. The water drive model is based on a
radial composite model, with an outside aquifer, which is defined by different levels of mobility from
different aquifers and relative to the reservoir (Chambers, Karra & Mortimer, 1980). Boundary
dominated match function is to get information about drainage areas and reserve, which focuses on the
stems (depletion) which are dominated by boundaries from type curve. Analysis of this method
explains that this type of curve doesn't need a hyperbolic exponent. Alternatively, the data is matching
the stem with a single depletion. When the plot data is relocated, OGIP / OOIP continues to experience
updates. For gas reservoirs, the fluid is very function as pressure, the data points will "stretch" and
"contract" when moved. This is due to the pseudo-time effect. When a match is modified, OGIP,
average reservoir pressure, and material balance time are calculated automatically and data points are
repositioned according to known data. The graph is prepared into a type curve for flow with the given
URTeC 2878

Outer and Inner boundary conditions. Depending on the type of drawdown and buildup curve solution
that is generated (Agarwal, 1980). Transient match to get data about skin and permeability, it is
recommended to focus on the transient stems of this type curve. Agarwal-Gardner type curve,
appearing on the left of the plot with different reD values for a radial and water drive models, while
the re / xf value for the fracture model. Choose the curve type that best matches the data; this gives an
accurate reD (or re / xf) value. Pressure build-up assumes that the production period before closing is
long enough that causing transients when the flow period doesn't influence the next pressure buildup
data (Agarwal, 1980).
Blasingame Type Curve Analysis
This type curve analysis method provide pressure normalized and the material balance time rates to
expand rate type curve of analytical constant which merits the stem of a single depletion that not
dependent on reservoirs drive, including the shape and size. The meaning of time of material balance is a
concept of time which brings history of variable rate and modify that into the equivalent history of
constant rate. The material balance time improves the solution curve of constant pressure to become the
curve with a steady rate or known as harmonic curve. And as the result in each condition of flowing either
rate nor pressure is constant, the stem of harmonic the model of Fetkovich may be good enough in
analyzing all types of production data in consistent way (Mattar & Andeerson, 2003). The results of
Blasingame analysis among others skin factor, formation permeability, the drainage area of the reservoir,
and capacity of hydrocarbons in place. Moreover, to make sure a more precise in the matching of type
curve, function of flow rate integral derivative and flow rate integral was used rather than just make use
the data of flow rate. These integral functions also be able to get rid of issues related to data of production
with bottom-hole pressure behavior and the rate is not consistent. Blasingame type curve method can be
used on elliptical and radial flow wells, fractured wells or known as cylindrical, horizontal well, and well
with water drive mechanism. (Wilson, 2017). Moreover, if radius of wellbore, total compressibility and
thickness of reservoir are known, we could get information such as fracture half length, skin, permeability
of formation, area of drainage, dimensionless conductivity of fracture, and then amount of oil or gas
initial (Fekete, 2014a).
More precisely, type-curve plot the Palacio-Blasingame presents the following (Houzé, Viturat, & Fjaere,
1988):
Material Balance time:
𝑄
𝑡𝑒 = 𝑞𝑡 (68)
𝑡

Normalized rate:
𝑞
𝑃𝐼(𝑡) = 𝑃 −𝑃𝑡 (69)
𝑖 𝑤(𝑡)

Normalized rate integral:

1 𝑡 1 𝑡 𝑞
𝑃𝐼 𝐼𝑛𝑡. = 𝑡 ∫0 𝑒 𝑃𝐼(𝜏)𝑑𝜏 = 𝑡 ∫0 𝑒 𝑃 −𝑃𝑡 𝑑𝜏 (70)
𝑒 𝑒 𝑖 𝑤(𝑡)

Normalized rate integral derivative:

𝜕(𝑃𝐼𝑖𝑛𝑡 )
𝑃𝐼 𝐼𝑛𝑡. 𝐷𝑒𝑟𝑖𝑣𝑎𝑡𝑖𝑣𝑒 = 𝜕 ln(𝑡𝑒 )
(71)
URTeC 2878

Compound Linier Type Curve Analysis

Based on history, type curve matching such as Fetkovic and Blasingame was proved as a rightful method
for interpretation of data of production. The existence of conventional models are no longer helpful, as the
emergence of settlement for complex stratified in reservoirs with tight fractured. Then type curves of
Coumpound Linear Flow that more proper to unconventional was introduced (Mahmoud & Anderson,
2015).
The Compound Linear Flow type curves do not include end effects or boundary dominated flow, to
decrease confusion and complexity. As a replacement, Liang (2012) focus concern on the primary flow
regimes of interest for rate-transient analysis- early linear flow, transition and compound linear flow.
The type curves represent the behavior of only one single fracture, shown in the Figure below, for multi-
fractured well, in an infinitely large reservoir, several assumptions are used as following (Mahmoud &
Anderson, 2015) :
• Fractures along the horizontal well have the same spaced
• Conductivity and half-length of fracture are equally
• The distancing inter wells of horizontal is very large
• Reservoir known to have characterized such as the homogenous, uniformity in thickness and
single porosity of isotropic system, and single-layer.
• The flow of horizontal well is negligible

Figure 29. Reservoir Model used to generate the Compound Linear Type Curves

Type curves (constant rate) developed for a range of aspect ratios, xe/xf = 0.1, 0.2, 0.5, 1, 2 and 5. The
xe/xf ratios were not considered if obtained less than 0.1 and greater than 5, due to these geometries did
not indicate reasonable fracture spacings. The type curves could be presented in two forms among others
pressure time and rate time. The pressure-time solves for constant rate solution, meanwhile the rate time
is just the reverse of that solution.The using CLF type curves give outcome for determining the efficiency
URTeC 2878

of stimulation through the length of lateral. Which the length of the effective stimulated horizontal well
divided by the length of the actual horizontal well was the definition of stimulation efficiency. Moreover,
it can be used to guess the width of the stimulated reservoir that is objectivecally and consistently
reasonable. And these outputs can be used to be the source of the model of analytical for the next history
matching and forecasting.
Fetkovich Type Curve Analysis
This method found by Fetkovich after he identified where the method explained by J. J. Arps only can be
applied during the period of depletion with the result that Arps method did not give concern in the early
period of well namely transient flow. Fetkovich uses the dimensionless equation for the flow of analytical
and its use not only to boundary-dominated analysis but also transient flow analysis (Fekete, 2014a). The
advantages of this method among other methods are the interpretation doesn't use superposition time
functions to plot data, represent the production performance of the unconventional well. But this method
has a shortcoming, in the transient type curves was restricted to the systems of radial flow (Wilson, 2017).
In this method there are 2 assumptions used. These assumptions used in this method are constant flowing
pressure and fluid of slightly compressible. In consequence of the interlude in the data of production due
to well under compression or shut-in is negligible. However, this method can't carry out for production
with a restricted rate. (Mattar & Andeerson, 2003). Values of re and kh, Di and qi can be obtained if
reached a good match. Moreover, the decline exponent (b) can be obtained by pick type-curve the right
one. The pore volume of the reservoir could be forecasted, in case the boundary of external distance had
known. The NPI method not only gives output Np for any certain abandonment rate but also the
prediction of performance in future (Houzé et al., 1988).
Normalized Pressure Integral Type Curve Analysis
Normalized pressure integral (NPI) type curve is the reverse of type curve of Agarwal-Gardner. By
changing the time with a time that is equivalent, can be established as the comparison between the
cumulative and the rate of flow. The analogous in steady rate completion can lead one action farther if,
this NPI works under rate-normalized pressure instead of working with a pressure-normalized rate.
𝑝𝑖 −𝑝𝑤(𝑡)
The plot for the liquid case conduct between 𝑞𝑡
vs. te on the scale of log-log, the unit slop can be
obtained on boundary dominated flow. In addition, part of transient going show the stabilization at stage
related to mobility. Nonetheless, this connection only applies to oil. Except done rate normalization with
concern to ∆m(p) and normalized pseudo time. The resemblance with pressure transient analysis is thus
perfect. Besides the derivation have the level of noise is generally overly high. One of the solutions is run
with normalized integral pressure that can be seen in the figure below, in an analogous manner to what
had done on the type curve of Palacio-Blasingame (Odagme, 2016). Normalized pressure Integral
(Equation 72) and the Bourdet derivation of normalized pressure integral (Equation 73) is given below.
1 𝑡 𝑃𝑖 −𝑃𝑤(𝑡)
𝐼(𝑡𝑒 ) = 𝑡 ∫0 𝑒 𝑞(𝑡)
𝑑𝜏 (72)
𝑒

𝜕𝐼(𝑡𝑒 )
𝐼 ′ (𝑡𝑒 ) = (73)
𝜕 ln(𝑡𝑒 )
URTeC 2878

Figure 30. Integral of Normalized Pressure and Derivative vs. 𝑡𝑒 .

Transient Type Curve Analysis

The transient type curve shows an alternative ideal point of view for very short periods of production
analysis, also analysis for reservoir with the permeability of very low. Based on the data set of transient
production, comparing conducted has been taken between type curve match using type curve Agarwal-
Gardner and Transient can be seen below (IHS Markit, 2018).

Figure 31. Comparison of Agarwal-Gardner Type Curve and Transient Type Curve analyses.

The example above shows that the transient one has a unique curve than agarwal-gardner type curve. This
is because transient data is more suitable for use in transient format (qD vs. tD). Therefore for the analysis
of boundary-dominated flow types it is not recommended to use this method. The time function that is
suitable for use in this method is time of material balance, because it more precise for boundary-
dominated flow. However, when faced with transient data, material balance time data is not the best
choice. The time of radial-superposition is the basic of time-superposition for transient flow. When using
this type curve, both of the radial and linear superposition time options must available.
URTeC 2878

The calculation of parameters transient type curve analysis can be used (Fekete, 2014b) on neither vertical
nor horizontal well, fracture with has infinite conductivity, and then fracture with has finite conductivity.
Wattenbarger Type Curve Analysis
Expanded periods of linear flow that had monitored in lots of wells of gas. These gas wells generally in
the reservoir of gas with characterized very tight. Besides, hydraulic fracturing planning to installed to
lengthen the unto drainage boundary of well (Fekete, 2014c). Wattenbarger (1998) explained the latest
type of curve to data production analysis of these wells. In addition, this type curve using the assumption
presence of hydraulic fractured well that located on the center of a reservoir with a rectangular shape
(Ghedan et al., 2009). This fractured reached the boundaries of reservoir.

Figure 32. Location of a Hydraulic Fractured well which explained by Wattenbarger Type Curve Analysis.

And the equation that using for closed-reservoir and production with constant rate can be expressed as;

𝑥 2 2
𝜋 𝑌 1 2 𝑌𝑒 1 2 2 𝑥𝑓
𝑃𝐷 = ( 𝑒 ) [ + ( 𝑓) 𝑡𝐷𝑥𝑓 ] − ( ) ∑∞
𝑛−1 𝑒 [−𝑛 𝜋 ( ) 𝑡𝐷𝑥𝑓 ] (74)
2 𝑥𝑓 3 𝑌𝑒 𝜋2 𝑥𝑓 𝑛2 𝑌𝑒

Where dimensionless variables can be expressed as:

𝑘ℎ(𝑝𝑖 −𝑝𝑤𝑓 )
𝑝𝐷 = (75)
141.2 𝑞𝐵𝜇

0.00633𝑘𝑡
𝑡𝐷𝑥𝑓 = ∅𝜇𝑐𝑡 𝑥𝑓2
(76)

Wattenbarger (1998) exhibited different types of curve types based on ratios xf/Ye as follow:
URTeC 2878

Figure 33. Different Curve types based on the ratios xf/Ye.

PSS Waterdrive
The PSS waterdrive method depends on the invention by Fetkovich which is executed in the Agarwal-
Gardner, Blasingame, and Normalized Pressured integral models (Fekete, 2014c). The assumption that
used in this model is given below:
• Finite aquifer
• A geometry not depend on productivity index
• The reservoir has the mobility range medium-high for ideal limited aquifers
• Planned to analyze early time period, prior to water breakthrough occurrence
• Estimation material balance of reservoir. Besides determine of water influx, including the effect.
A scheme of PSS Waterdrive method can be illustrated in figure below. The aim of this model is to decipher
the equation of material balance with the equation of aquifer influx in simultaneously.

Figure 34. PSS Waterdrive method Scheme.

The implementation of this method implicates a process which known as reverse modelling. This method
makes the type curve of standard as the basic model, also to reduce the effect of water influx. In order to
the procedure of reverse modeling become effective, the data of aquifer effects must appear clearly. The
major parameters in this pseudo-steady state matching model namely the capacity of water initial and the
index of productivity of aquifer.
URTeC 2878

Results
4 different hydrocarbon types such as; volatile oil (Volatile_15), light oil (LightOil_21), condensate
(Condensate_9) and dry gas (Dry_Gas_3) were used in models and predictions.Feature sets created to be
used in XGBOOST model as “dayofweek”, “quarter”, “month”, “year”, “dayofyear”, “dayofmonth”,
“weekofyear”. Xgboost model trained by using the created features. Hyperparameter tuning is performed
manually. Daily oil rate data compared with the prediction data as well as test statistics were compared
with the prediction statistics. MSE and RMSE Errors also calculated.

Volatile Oil Production (Volatile_15)

The field oil rate data which recorded in standart conditions were compared with the prediction results
(Figure 35). Herein, the sudden increase of production at the beginnning of the dataset is removed by
filtering the data points earlier than the date of “13.02.2020”. Train and the test datasets created by
splitting the dataset into two at split date of “31.12.2021”.

Figure 35. Daily Oil Rates (SC) comparison with the prediction data for volatile oil.

Besides, the statistics of the test results were compared with the prediction results (Table 6). It can be
clearly seen that the field rates show differencies with the XGBOOST prediction rates, especially in
earlier times. It can be seen from the differencies of mean, and standard deviation results as well.
Table 6. Comparison of Field Data (Test) and XGBOOST Prediction Statistics.
URTeC 2878

Following the XGBOOST prediction, Holt-Winters Triple Exponential Smoothing (TES) model trained
by using the train and test data which were created in the XGBOOST section. Herein, trend and seasonal
variables set as multiplicative with dampened trends. Seasonal periods set as 2 months. The field oil rate
data (TEST) which recorded in standart conditions were compared with the ALL prediction results
(Figure 36).

Figure 36. Daily oil rates (SC) comparison with all prediction models for volatile oil case.

Volatile oil field data (TEST) statistics were compared with the TES prediction statistics in Table 7.
Results showed that the maximum values were closer than the XGBOOST Prediction.
Table 7. Comparison of field data (Test) and TES prediction statistics for volatile oil case.
URTeC 2878

Light Oil Production (LightOil_21)

As it has been done in Volatile oil example, the date data resampled into the start of the month format and
the sudden increase on the dataset in earlier times is removed by filtering the data points earlier than the
date of “13.02.2020”. XGBOOST model set with the same format of volatile oil (as “dayofweek”,
“quarter”, “month”, “year”, “dayofyear”, “dayofmonth”, “weekofyear”) and the XGBOOST model
trained by using the created features. Again, hyperparameter tuning is performed by hand (grid search
option can be used in future studies). The field oil rate data which recorded in standart conditions were
compared with the prediction results (Figure 37).

Figure 37. Daily Oil Rates (SC) comparison with the prediction data for light oil case.

Light oil rate data statistics were compared with the XGBOOST prediction statistics and some differences
were observed in maximum-minimum rate values as well as in 25%, 50% and 75% values of the rates. It
also affected the mean and standard deviation results (Table 8).
Table 8. Comparison of field data (Test) and TES prediction statistics for light oil case.

Holt-Winters Triple Exponential Smoothing model is also trained for the Light oil case by using the train
and the test date which were created in XGBOOST section. Trend and the seasonal variables set as
URTeC 2878

multiplicative with dampened treneds. It worth to pointing out that the seasonal periods set as 3 months in
this case. The field oil rate data (TEST) which recorded in standart conditions were compared with the
ALL prediction results (Figure 38).

Figure 38. Daily oil rates (SC) comparison with all prediction models for light oil case.

Light oil field data (TEST) statistics were compared with the TES prediction statistics in Table 9. Results
showed that the maximum values were matching compared with the test data.

Table 9. Comparison of field data (Test) and TES prediction statistics for light oil case.
URTeC 2878

Condensate Production (Condensate_9)

Similar to the previous cases, the sudden increase of the production at the beginning of the dataset is
removed by filtering option. It resulted by the removal of the data points earlier than the date of
“13.02.2020”. Train and test datasets created by splitting the dataset into two parts at the split date of
“31.12.2021”. XGBOOST model set with the same format of volatile oil (as “dayofweek”, “quarter”,
“month”, “year”, “dayofyear”, “dayofmonth”, “weekofyear”) and the XGBOOST model trained by using
the created features. Again, hyperparameter tuning is performed by hand (grid search option can be used
in future studies). The field oil rate data which recorded in standart conditions were compared with the
prediction results (Figure 39).

Figure 39. Daily Oil Rates (SC) comparison with the prediction data for condensate case.

Condensate rate data statistics were compared with the XGBOOST prediction statistics and some
differences were observed in maximum-minimum rate values as well as in 25%, 50% and 75% values of
the rates. It also affected the mean and standard deviation results (Table 10).
Table 10. Comparison of field data (Test) and TES prediction statistics for condensate case.

Similar to the volatile and light oil cases, Holt-Winters Triple Exponential Smoothing model is also trained
for the Light oil case by using the train and the test date which were created in XGBOOST section. Trend
and the seasonal variables set as multiplicative with dampened treneds. It worth to pointing out that the
URTeC 2878

seasonal periods set as 5 days in this case. The field oil rate data (TEST) which recorded in standart
conditions were compared with the ALL prediction results (Figure 40).

Figure 40. Daily oil rates (SC) comparison with all prediction models for condensate case.

Condensate field data (TEST) statistics were compared with the TES prediction statistics in Table 11.
Results showed that the maximum values as well as the mean data are perfectly matching compared with
the Test data.
Table 11. Comparison of field data (Test) and TES prediction statistics for condensate case.
URTeC 2878

Dry Gas Production (Dry Gas_3)

Similar to the previous cases, the sudden increase of the production at the beginning of the dataset is
removed by filtering option. It resulted by the removal of the data points earlier than the date of
“13.02.2020”. Train and test datasets created by splitting the dataset into two parts at the split date of
“31.12.2021”. XGBOOST model set with the same format of volatile oil (as “dayofweek”, “quarter”,
“month”, “year”, “dayofyear”, “dayofmonth”, “weekofyear”) and the XGBOOST model trained by using
the created features. Again, hyperparameter tuning is performed by hand (grid search option can be used
in future studies). The field oil rate data which recorded in standart conditions were compared with the
prediction results (Figure 41).

Figure 41. Daily oil rates (SC) comparison with all prediction models for dry gas case.

Dry Gas rate data statistics were compared with the XGBOOST prediction statistics and its observed that
there is slight matching between maximum-minimum rate values as well as in 25%, 50% and 75% values
of the rates. It also affected the mean and standard deviation results (Table 12).
Table 12. Comparison of field data (Test) and TES prediction statistics for dry gas case.
URTeC 2878

Similar to the previous 3 cases, Holt-Winters Triple Exponential Smoothing model is also trained for the
dry gas case by using the train and the test date which were created in XGBOOST section. Trend and the
seasonal variables set as multiplicative with dampened treneds. It worth to pointing out that the seasonal
periods set as 5 days in this case. The field oil rate data (TEST) which recorded in standart conditions were
compared with the ALL prediction results (Figure 42).

Figure 42. Daily oil rates (SC) comparison with all prediction models for dry gas case.

Dry Gas field data (TEST) statistics were compared with the TES prediction statistics in Table 13. Results
showed that all values are perfectly matching compared with the test data.
Table 13. Comparison of field data (Test) and TES prediction statistics for dry gas case.
URTeC 2878

Conclusions
Volatile oil, light oil, condensate, and dry gas cases were used in models and predictions. XGBOOST,
and Holt-Winter TES results were compared with the test data. Especially, in TES results, the accuracy
and data matching ratio is increased when the error precentage and the standard deviation decreases in
condensate and dry gas cases compared with volatile oil and light oil cases.
The results and analysis indicate that the data-driven Machine Learning models need a good amount of
history in order to perform well as they compare to that of the DCA methods which involve more
engineering judgement, which obviously may introduce some degree of bias as well.
When the results of different fluid types and history are used, the outcome differs a lot as expected where
the simulation results serve as the benchmark/reference forecast.
Even though a simulation-based approach has been used, the model outcome has been arranged so that
the results are realistic not only physically but also operationally. The operationally more realistic results
help the machine learning models to perform better as compared to smooth curves/data.

References
Ahmadi, M. A. (2015). Connectionist approach estimates gas–oil relative permeability in petroleum
reservoirs: application to reservoir simulation. Fuel, 140, 429-439.
Ahmadi, M. A., & Bahadori, A. (2015). A LSSVM approach for determining well placement and conning
phenomena in horizontal wells. Fuel, 153, 276-283.
Ahmadi, M. A., Hasanvand, M. Z., & Bahadori, A. (2017). A least-squares support vector machine
approach to predict temperature drop accompanying a given pressure drop for the natural gas production
and processing systems. International Journal of Ambient Energy, 38(2), 122-129.
Ahmadi, M. A., Soleimani, R., Lee, M., Kashiwao, T., & Bahadori, A. (2015). Determination of oil well
production performance using artificial neural network (ANN) linked to the particle swarm optimization
(PSO) tool. Petroleum, 1(2), 118-132.
Aifa, T. (2014). Neural network applications to reservoirs: Physics-based models and data models.
Journal of Petroleum Science and Engineering, 123, 1-6.
Aizenberg, I., Sheremetov, L., Villa-Vargas, L., & Martinez-Muñoz, J. (2016). Multilayer neural network
with multi-valued neurons in time series forecasting of oil production. Neurocomputing, 175, 980-989.
Akbilgic, O., Zhu, D., Gates, I. D., & Bergerson, J. A. (2015). Prediction of steam-assisted gravity
drainage steam to oil ratio from reservoir characteristics. Energy, 93, 1663-1670.
Alkinani, H. H., Al-Hameedi, A. T. T., Dunn-Norman, S., Flori, R. E., Alsaba, M. T., & Amer, A. S.
(2019, March). Applications of Artificial Neural Networks in the Petroleum Industry: A Review. SPE
Middle East Oil and Gas Show and Conference. Society of Petroleum Engineers.
Aloulou, F., & Wagener, D. (2018). Tight Oil Remains the Leading Source of Future US Crude Oil
Production.
Arnold, R., & Anderson, R. (1908). Preliminary report on the Coalinga oil district, Fresno and Kings
Counties, California (Vol. 357). US Government Printing Office.
Arnold, R. (1923). Two decades of petroleum geology, 1903-22. AAPG Bulletin, 7(6), 603-624.
Arps, J. J. (1945). Analysis of decline curves. Transactions of the AIME, 160(01), 228-247.
Awoleke, O., & Lane, R. (2011). Analysis of data from the Barnett shale using conventional statistical
and virtual intelligence techniques. SPE Reservoir Evaluation & Engineering, 14(05), 544-556.
URTeC 2878

Bengio, Y. (2009). Learning deep architectures for AI. Foundations and Trends® in Machine Learning,
2(1), 1-127.
Bhattacharya, S., Ghahfarokhi, P. K., Carr, T. R., & Pantaleone, S. (2019). Application of predictive data
analytics to model daily hydrocarbon production using petrophysical, geomechanical, fiber-optic,
completions, and surface data: A case study from the Marcellus Shale, North America. Journal of
Petroleum Science and Engineering, 176, 702-715.
Biswas, D. (2019, October). Adapting Shallow and Deep Learning Algorithms to Examine Production
Performance–Data Analytics and Forecasting. SPE/IATMI Asia Pacific Oil & Gas Conference and
Exhibition. Society of Petroleum Engineers.
Blunsom, P., Cho, K., Dyer, C., & Schütze, H. (2017). From characters to understanding natural language
(C2NLU): Robust end-to-end deep learning for NLP (Dagstuhl Seminar 17042). Dagstuhl Reports (Vol.
7, No. 1). Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik.
Box, G. E., & Pierce, D. A. (1970). Distribution of residual autocorrelations in autoregressive-integrated
moving average time series models. Journal of the American Statistical Association, 65(332), 1509-1526.
Breiman, L. (1997). Arcing the edge. Technical Report 486, Statistics Department, University of
California at Berkeley.
Breiman, L. (2001). Random forests. Machine learning, 45(1), 5-32.
Breiman, L., Friedman, J., Stone, C. J., & Olshen, R. A. (1984). Classification and regression trees. CRC
press.
Calvette, T., Gurwicz, A., Abreu, A. C., & Cavalcanti Pacheco, M. A. (2019, October). Forecasting Smart
Well Production via Deep Learning and Data Driven Optimization. Offshore Technology Conference,
Brazil.
Temizel, C., Zhiyenkulov, M., Ussenova, K., Kazhym, T., Canbaz, C. H., & Saputelli, L. A. (2018,
September). Optimization of well placement in waterfloods with optimal control theory under geological
uncertainty. In SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers.
Temizel, C., Irani, M., Ghannadi, S., Canbaz, C. H., Moreno, R., & Bashtani, F. (2019, September).
Optimization of Steam-Additive Processes with DTS/DAS Applications in Heavy Oil Reservoirs. In SPE
Annual Technical Conference and Exhibition. Society of Petroleum Engineers.
Temizel, C., Balaji, K., Canbaz, C. H., Palabiyik, Y., Moreno, R., Rabiei, M., ... & Ranjith, R. (2019,
April). Data-Driven Analysis of Natural Gas EOR in Unconventional Shale Oils. In SPE Oklahoma City
Oil and Gas Symposium. Society of Petroleum Engineers.
Temizel, C., Canbaz, C. H., Palabiyik, Y., Putra, D., Asena, A., Ranjith, R., & Jongkittinarukorn, K.
(2019, March). A Comprehensive Review of Smart/Intelligent Oilfield Technologies and Applications in
the Oil and Gas Industry. In SPE Middle East Oil and Gas Show and Conference. Society of Petroleum
Engineers.
Temizel, C., Canbaz, C. H., Alsaheib, H., & Monfared, H. (2020, January). Optimization of Smart Well
Placement in Waterfloods Under Geological Uncertainty in Intelligent Fields. In International Petroleum
Technology Conference. International Petroleum Technology Conference.
Can, B., & Kabir, S. (2012). Probabilistic production forecasting for unconventional reservoirs with
stretched exponential production decline model. SPE Reservoir Evaluation & Engineering, 15(01), 41-50.
Canbaz, C. H., Deniz-Paker, M., Hosgor, F. B., Putra, D., Moreno, R., Temizel, C., & Alkouh, A. (2019,
October). Optimization of Development of Heavy Oil Reservoirs through Geochemical Characterization.
In SPE Kuwait Oil & Gas Show and Conference. Society of Petroleum Engineers.
URTeC 2878

Canchumuni, S. A., Emerick, A. A., & Pacheco, M. A. (2017, October). Integration of ensemble data
assimilation and deep learning for history matching facies models. Offshore Technology Conference,
Brazil.
Cao, C., Liu, F., Tan, H., Song, D., Shu, W., Li, W., Zhou, Y., Bo, X. & Xie, Z. (2018). Deep learning
and its applications in biomedicine. Genomics, Proteomics & Bioinformatics, 16(1), 17-32.
Chakra, N. C., Song, K. Y., Gupta, M. M., & Saraf, D. N. (2013). An innovative neural forecast of
cumulative oil production from a petroleum reservoir employing higher-order neural networks (HONNs).
Journal of Petroleum Science and Engineering, 106, 18-33.
Clark, A. J., Lake, L. W., & Patzek, T. W. (2011, January). Production forecasting with logistic growth
models. SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers.
Clarkson, C. R., & Qanbari, F. (2016). A semi-analytical method for forecasting wells completed in low
permeability, undersaturated CBM reservoirs. Journal of Natural Gas Science and Engineering, 30, 19-27.
Clarkson, C. R., Williams-Kovacs, J. D., Qanbari, F., Behmanesh, H., & Sureshjani, M. H. (2015).
History-matching and forecasting tight/shale gas condensate wells using combined analytical, semi-
analytical, and empirical methods. Journal of Natural Gas Science and Engineering, 26, 1620-1647.
Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine learning, 20(3), 273-297.
Du, D. F., Wang, Y. Y., Zhao, Y. W., Sui, P. S., & Xia, X. (2017). A new mathematical model for
horizontal wells with variable density perforation completion in bottom water reservoirs. Petroleum
Science, 14(2), 383-394.
Duong, A. N. (2011). Rate-decline analysis for fracture-dominated shale reservoirs. SPE Reservoir
Evaluation & Engineering, 14(03), 377-387.
El-Sebakhy, E. A., Sheltami, T., Al-Bokhitan, S. Y., Shaaban, Y., Raharja, P. D., & Khaeruzzaman, Y.
(2007, January). Support vector machines framework for predicting the PVT properties of crude oil
systems. SPE Middle East Oil and Gas Show and Conference. Society of Petroleum Engineers.
Figueiredo, B., Tsang, C. F., Rutqvist, J., & Niemi, A. (2017). Study of hydraulic fracturing processes in
shale formations with complex geological settings. Journal of Petroleum Science and Engineering, 152,
361-374.
Ghahramani, Z. (2015). Probabilistic machine learning and artificial intelligence. Nature, 521(7553), 452-
459.
Ghanizadeh, A., Clarkson, C. R., Deglint, H., Vahedian, A., Aquino, S., & Wood, J. M. (2016, August).
Unpropped/propped fracture permeability and proppant embedment evaluation: A rigorous core-
analysis/imaging methodology. Unconventional Resources Technology Conference (URTEC).
Gharbi, R. B., Elsharkawy, A. M., & Karkoub, M. (1999). Universal neural-network-based model for
estimating the PVT properties of crude oil systems. Energy & Fuels, 13(2), 454-458.
Ghedan, S. G., Canbaz, C. H., & Mtawaa, B. (2009, January). Effect of Shape Factor, IFT and Amott
Method Derived Wettability on the Imbibition Process. In International Petroleum Technology
Conference. International Petroleum Technology Conference.
Graves, A., Mohamed, A. R., & Hinton, G. (2013, May). Speech recognition with deep recurrent neural
networks. IEEE International Conference on Acoustics, Speech and Signal Processing, 6645-6649. IEEE.
Graves, A., Jaitly, N., & Mohamed, A. R. (2013, December). Hybrid speech recognition with deep
bidirectional LSTM. IEEE Workshop on Automatic Speech Recognition and Understanding, 273-278.
IEEE.
URTeC 2878

Grosse, R. (2017). Lecture 15: Exploding and vanishing gradients. University of Toronto Computer
Science.
Gu, H., Siebrits, E., & Sabourov, A. (2008, January). Hydraulic fracture modeling with bedding plane
interfacial slip. SPE Eastern Regional/AAPG Eastern Section Joint Meeting. Society of Petroleum
Engineers.
Guo, Z., Chen, C., Gao, G., & Vink, J. (2018). Enhancing the performance of the distributed Gauss-
Newton optimization method by reducing the effect of numerical noise and truncation error with support-
vector regression. SPE Journal, 23(06), 2-428.
Guo, Z., Chen, C., Gao, G., Cao, R., Li, R., & Liu, H. (2018). Integration of support vector regression
with distributed Gauss-Newton optimization method and its applications to the uncertainty assessment of
unconventional assets. SPE Reservoir Evaluation & Engineering, 21(04), 1-007.
Han, D., Jung, J., & Kwon, S. (2020). Comparative Study on Supervised Learning Models for
Productivity Forecasting of Shale Reservoirs Based on a Data-Driven Approach. Applied Sciences, 10(4),
1267.
Hanson, M. E., Shaffer, R. J., & Anderson, G. D. (1981). Effects of various parameters on hydraulic
fracturing geometry. SPE Journal, 21(04), 435-443.
Hassani, H., Sarkheil, H., Foroud, T., & Karimpooli, S. (2011, January). A proxy modeling approach to
optimization horizontal well placement. 45th US Rock Mechanics/Geomechanics Symposium. American
Rock Mechanics Association.
He, Z., Yang, L., Yen, J., & Wu, C. (2001, January). Neural-network approach to predict well
performance using available field data. SPE Western Regional Meeting. Society of Petroleum Engineers.
Ho, T. K. (1995, August). Random decision forests. Proceedings of 3rd International Conference on
Document Analysis and Recognition, 1, 278-282. IEEE.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), 1735-
1780.
Hu, Y., Weijermars, R., Zuo, L., & Yu, W. (2018). Benchmarking EUR estimates for hydraulically
fractured wells with and without fracture hits using various DCA methods. Journal of Petroleum Science
and Engineering, 162, 617-632.
Hughes, J. D. (2013). Energy: A reality check on the shale revolution. Nature, 494(7437), 307.
Hughes, J. D. (2013). Drill, baby, drill: can unconventional fuels usher in a new era of energy abundance?
http://www.postcarbon.org/publications/drill-baby-drill/.
Ilk, D., Rushing, J. A., Perego, A. D., & Blasingame, T. A. (2008, January). 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. Society of Petroleum Engineers.
Kalra, S., Tian, W., & Wu, X. (2018). A numerical simulation study of CO 2 injection for enhancing
hydrocarbon recovery and sequestration in liquid-rich shales. Petroleum Science, 15(1), 103-115.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional
neural networks. Advances in Neural Information Processing Systems. 1097-1105.
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436-444.
Lee, K., Kim, S., Choe, J., Min, B., & Lee, H. S. (2019). Iterative static modeling of channelized
reservoirs using history-matched facies probability data and rejection of training image. Petroleum
Science, 16(1), 127-147.
URTeC 2878

Lee, K., Lim, J., Ahn, S., & Kim, J. (2018). Feature extraction using a deep learning algorithm for
uncertainty quantification of channelized reservoirs. Journal of Petroleum Science and Engineering, 171,
1007-1022.
Lee, K., Lim, J., Choe, J., & Lee, H. S. (2017). Regeneration of channelized reservoirs using history-
matched facies-probability map without inverse scheme. Journal of Petroleum Science and Engineering,
149, 340-350.
Lee, K., Lim, J., Yoon, D., & Jung, H. (2019). Prediction of Shale-Gas Production at Duvernay
Formation Using Deep-Learning Algorithm. SPE Journal.
Li, X., Chan, C. W., & Nguyen, H. H. (2013). Application of the Neural Decision Tree approach for
prediction of petroleum production. Journal of Petroleum science and engineering, 104, 11-16.
Liao, L., Zeng, Y., Liang, Y., & Zhang, H. (2020, January). Data Mining: A Novel Strategy for
Production Forecast in Tight Hydrocarbon Resource in Canada by Random Forest Analysis. International
Petroleum Technology Conference (IPTC).
Lim, J. S., & Kim, J. (2004, January). Reservoir porosity and permeability estimation from well logs
using fuzzy logic and neural networks. SPE Asia Pacific Oil and Gas Conference and Exhibition. Society
of Petroleum Engineers.
Littman, M. L. (2015). Reinforcement learning improves behaviour from evaluative feedback. Nature,
521(7553), 445-451.
Liu, W., Liu, W. D., & Gu, J. (2020). Forecasting oil production using ensemble empirical model
decomposition based Long Short-Term Memory neural network. Journal of Petroleum Science and
Engineering, 189, 107013.
Luo, G., Tian, Y., Sharma, A., & Ehlig-Economides, C. (2019, March). Eagle Ford Well Insights Using
Data-Driven Approaches. International Petroleum Technology Conference (IPTC).
Balaji, K., Rabiei, M., Suicmez, V., Canbaz, C. H., Agharzeyva, Z., Tek, S., ... & Temizel, C. (2018,
June). Status of data-driven methods and their applications in oil and gas industry. In SPE Europec
featured at 80th EAGE Conference and Exhibition. Society of Petroleum Engineers.
Ma, Z., Leung, J. Y., Zanon, S., & Dzurman, P. (2015). Practical implementation of knowledge-based
approaches for steam-assisted gravity drainage production analysis. Expert Systems with Applications,
42(21), 7326-7343.
Montgomery, J. B., & O’Sullivan, F. M. (2017). Spatial variability of tight oil well productivity and the
impact of technology. Applied Energy, 195, 344-355.
Nwachukwu, A., Jeong, H., Pyrcz, M., & Lake, L. W. (2018). Fast evaluation of well placements in
heterogeneous reservoir models using machine learning. Journal of Petroleum Science and Engineering,
163, 463-475.
Nwaobi, U., & Anandarajah, G. (2018). Parameter determination for a numerical approach to
undeveloped shale gas production estimation: The UK Bowland shale region application. Journal of
Natural Gas Science and Engineering, 58, 80-91.
Olah, C. (2015). https://colah.github.io/posts/2015-08-Understanding-LSTMs/
Panja, P., Velasco, R., Pathak, M., & Deo, M. (2018). Application of artificial intelligence to forecast
hydrocarbon production from shales. Petroleum, 4(1), 75-89.
Paryani, M., Ahmadi, M., Awoleke, O., & Hanks, C. (2016, September). Using improved decline curve
models for production forecasts in unconventional reservoirs. SPE Eastern Regional Meeting. Society of
Petroleum Engineers.
URTeC 2878

Robertson, S. (1988). Generalized hyperbolic equation. Society of Petroleum Engineers.

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating
errors. Nature, 323(6088), 533-536.
Sagheer, A., & Kotb, M. (2019). Time series forecasting of petroleum production using deep LSTM
recurrent networks. Neurocomputing, 323, 203-213.
Sak, H., Senior, A. W., & Beaufays, F. (2014). Long short-term memory recurrent neural network
architectures for large scale acoustic modeling.
Sarikaya, R., Hinton, G. E., & Deoras, A. (2014). Application of deep belief networks for natural
language understanding. IEEE/ACM Transactions on Audio, Speech and Language Processing, 22(4),
778-784.
Scheidt, C., & Caers, J. (2009). Uncertainty quantification in reservoir performance using distances and
kernel methods--application to a West Africa deep-water turbidite reservoir. SPE Journal, 14(04), 680-
692.
Seshadri, J. N., & Mattar, L. (2010, January). Comparison of power law and modified hyperbolic decline
methods. Canadian Unconventional Resources and International Petroleum Conference. Society of
Petroleum Engineers.
Shaheen, M., Shahbaz, M., ur Rehman, Z., & Guergachi, A. (2011). Data mining applications in
hydrocarbon exploration. Artificial Intelligence Review, 35(1), 1-18.
Shelley, B., Johnson, B. J., Fielder, E. O., Heinze, J. R., & Werline, J. R. (2008). Data Analysis of Barnett
Shale Completions. SPE Journal, 13 (3), 366-374.
Shi, G. (2013). Data mining and knowledge discovery for geoscientists. Elsevier.
Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J.,
Antonoglou, I., Panneershelvam, V., Lanctot, M. & Dieleman, S. (2016). Mastering the game of Go with
deep neural networks and tree search. Nature, 529(7587), 484.
Song, X., Liu, Y., Xue, L., Wang, J., Zhang, J., Wang, J., Jiang, L. and Cheng, Z. (2020). Time-series
well performance prediction based on Long Short-Term Memory (LSTM) neural network model. Journal
of Petroleum Science and Engineering, 186, 106682.
Tai, K. S., Socher, R., & Manning, C. D. (2015). Improved semantic representations from tree-structured
long short-term memory networks. arXiv preprint arXiv:1503.00075.
Tan, L., Zuo, L., & Wang, B. (2018). Methods of decline curve analysis for shale gas reservoirs. Energies,
11(3), 552.
Valko, P. P. (2009, January). Assigning value to stimulation in the Barnett Shale: a simultaneous analysis
of 7000 plus production histories and well completion records. SPE Hydraulic Fracturing Technology
Conference. Society of Petroleum Engineers.
Velasco, R., Panja, P., & Deo, M. (2016). New production performance and prediction tool for
unconventional reservoirs. Unconventional Resources Technology Conference (URTEC).
Wang, L., Zou, H., Su, J., Li, L., & Chaudhry, S. (2013). An ARIMA‐ANN hybrid model for time series
forecasting. Systems Research and Behavioral Science, 30(3), 244-259.
Wang, S., & Chen, S. (2019). Insights to fracture stimulation design in unconventional reservoirs based
on machine learning modeling. Journal of Petroleum Science and Engineering, 174, 682-695.
Wang, S., Chen, Z., & Chen, S. (2019). Applicability of deep neural networks on production forecasting
in Bakken shale reservoirs. Journal of Petroleum Science and Engineering, 179, 112-125.
URTeC 2878

Wang, Q., & Jiang, F. (2019). Integrating linear and nonlinear forecasting techniques based on grey
theory and artificial intelligence to forecast shale gas monthly production in Pennsylvania and Texas of
the United States. Energy, 178, 781-803.
Wilson, K. C., & Durlofsky, L. J. (2013). Optimization of shale gas field development using direct search
techniques and reduced-physics models. Journal of Petroleum Science and Engineering, 108, 304-315.
Xing, W., & Du, D. (2019). Dropout prediction in MOOCs: Using deep learning for personalized
intervention. Journal of Educational Computing Research, 57(3), 547-570.
Yu, W., Luo, Z., Javadpour, F., Varavei, A., & Sepehrnoori, K. (2014). Sensitivity analysis of hydraulic
fracture geometry in shale gas reservoirs. Journal of Petroleum Science and Engineering, 113, 1-7.
Zhan, C., Sankaran, S., LeMoine, V., Graybill, J., & Mey, D. O. S. (2019). Application of Machine
Learning for Production Forecasting for Unconventional Resources. Unconventional Resources
Technology Conference (URTEC).
Zhang, H., Rietz, D., Cagle, A., Cocco, M., & Lee, J. (2016). Extended exponential decline curve
analysis. Journal of Natural Gas Science and Engineering, 36, 402-413.
Zhang, R. H., Zhang, L. H., Tang, H. Y., Chen, S. N., Zhao, Y. L., Wu, J. F., & Wang, K. R. (2019). A
simulator for production prediction of multistage fractured horizontal well in shale gas reservoir
considering complex fracture geometry. Journal of Natural Gas Science and Engineering, 67, 14-29.
Zhong, M., Schuetter, J., Mishra, S., & Lafollette, R. F. (2015, February). Do data mining methods
matter? : A Wolfcamp Shale case study. SPE Hydraulic Fracturing Technology Conference. Society of
Petroleum Engineers.
Zuo, L., Yu, W., & Wu, K. (2016). A fractional decline curve analysis model for shale gas reservoirs.
International Journal of Coal Geology, 163, 140-148.
Agarwal, R. G. (1980, January). A new method to account for producing time effects when drawdown
type curves are used to analyze pressure buildup and other test data. In SPE Annual Technical Conference
and Exhibition. Society of Petroleum Engineers.
Agarwal, R. G., Gardner, D. C., Kleinsteiber, S. W., & Fussell, D. D. (1998, January). Analyzing well
production data using combined type curve and decline curve analysis concepts. In SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers.
Agarwal, R. G., Al-Hussainy, R., & Ramey Jr, H. J. (1970). An investigation of wellbore storage and skin
effect in unsteady liquid flow: I. Analytical treatment. Society of Petroleum Engineers Journal, 10(03),
279-290.
Ahmed, H. (2019). History-matching and Forecasting of Three Unconventional Oil and Gas Reservoirs
Using Decline Analyses and Type Curves (Doctoral dissertation).
Ahmed, T. (2018). Reservoir engineering handbook. Gulf Professional Publishing.
Ahmed, T. (2019). Reservoir Engineering Handbook. In Reservoir Engineering Handbook.
https://doi.org/10.1016/b978-0-12-813649-2.00016-5
Alom, M. S., Tamim, M., & Rahman, M. M. (2017). Decline curve analysis using rate normalized
pseudo-cumulative function in a boundary dominated gas reservoir. Journal of Petroleum Science and
Engineering, 150, 30-42.
Alsaadoun, D. (2018). Analysis of the Gas Production from Marcellus Shale Horizontal Wells Using
Decline Curves.
Arps, J. J. (1945). Analysis of Decline Curves. SPE, 160(1).
URTeC 2878

Bello, R. O., & Wattenbarger, R. A. (2008, January). Rate transient analysis in naturally fractured shale
gas reservoirs. In CIPC/SPE Gas Technology Symposium 2008 Joint Conference. Society of Petroleum
Engineers.
Belyadi, H., Fathi, E., & Belyadi, F. (2019). Hydraulic Fracturing in Unconventional Reservoirs. 311–
340. https://doi.org/10.1016/B978-0-12-817665-8.00017-5
Belyadi, H., Yuyi, S., & Junca-Laplace, J. P. (2015, October). Production analysis using rate transient
analysis. In SPE Eastern Regional Meeting. Society of Petroleum Engineers.
Bourdarot, G., & Balvet, B. B. (1998). Well testing: Interpretation methods. Paris, France: Editions
Technip.
Bourdet, D., & Gringarten, A. C. (1980, January). Determination of fissure volume and block size in
fractured reservoirs by type-curve analysis. In SPE annual technical conference and exhibition. Society of
Petroleum Engineers.
Can, B., Texas, A., Kabir, C. S., & Corporation, H. (2011). Probabilistic Performance Forecasting for
Unconventional Reservoirs with Stretched-Exponential Model.
Chambers, B., Karra, S., & Mortimer, L. (1980). Use of Type Curve Analysis In Predicting the Behaviour
of a Water-drive Reservoir. Journal of Canadian Petroleum Technology, 19(01).
Clark, A. J., Lake, L. W., & Patzek, T. W. (2011, January). Production forecasting with logistic growth
models. In SPE annual technical conference and exhibition. Society of Petroleum Engineers.
Clarkson, C. R., & Pedersen, P. K. (2011, January). Production analysis of Western Canadian
unconventional light oil plays. In Canadian Unconventional Resources Conference. Society of Petroleum
Engineers.
Daviau, F. (1986). Interprétation des essais de puits: les méthodes nouvelles. Éditions Technip.
Duong, A. N. (2010, January). An unconventional rate decline approach for tight and fracture-dominated
gas wells. In Canadian unconventional resources and international petroleum conference. Society of
Petroleum Engineers.
Duong, A. N. (2011). Rate-decline analysis for fracture-dominated shale reservoirs. SPE Reservoir
Evaluation & Engineering, 14(03), 377-387.
Engler, T., & Tiab, D. (1996). Analysis of pressure and pressure derivative without type curve matching,
4. Naturally fractured reservoirs. Journal of Petroleum Science and Engineering, 15(2-4), 127-138.
Fekete. (2014a). Fetkovich Typecurve Analysis Theory. Retrieved from
http://www.fekete.com/SAN/WebHelp/FeketeHarmony/Harmony_WebHelp/Content/HTML_Files/Refer
ence_Material/Analysis_Method_Theory/Fetkovich_Theory.htm.
Fekete. (2014b). Transient Typecurve Matching Equations. Retrieved from
http://www.fekete.com/SAN/WebHelp/FeketeHarmony/Harmony_WebHelp/Content/HTML_Files/Refer
ence_Material/Analysis_Method_Theory/Blasingame_Theory.htm#Transient_Typecurve_Matching_Equ
ations
Fekete. (2014c). Wattenbarger Typecurve Analysis Theory. Retrieved from
http://www.fekete.com/SAN/WebHelp/FeketeHarmony/Harmony_WebHelp/Content/HTML_Files/Refer
ence_Material/Analysis_Method_Theory/Wattenbarger_Theory.htm
Fetkovich, M. J., Fetkovich, E. J., & Fetkovich, M. D. (1996). Useful concepts for decline curve
forecasting, reserve estimation, and analysis. SPE Reservoir Engineering, 11(01), 13-22.
Freeborn, R., & Russell, B. (2012). How to Apply Stretched Exponential Equations to Reserve
Evaluation. (September), 24–25.
URTeC 2878

Gringarten, A. C. (1987). Type-curve analysis: what it can and cannot do. Journal of petroleum
technology, 39(01), 11-13.
Gringarten, A. C., Bourdet, D. P., Landel, P. A., & Kniazeff, V. J. (1979, January). A comparison
between different skin and wellbore storage type-curves for early-time transient analysis. In SPE Annual
Technical Conference and Exhibition. Society of Petroleum Engineers.
Gupta, I., Rai, C., Sondergeld, C., & Devegowda, D. (2018, September). Variable Exponential Decline-
Modified Arps to Characterize Unconventional Shale Production Performance. In Unconventional
Resources Technology Conference, Houston, Texas, 23-25 July 2018 (pp. 3552-3568). Society of
Exploration Geophysicists, American Association of Petroleum Geologists, Society of Petroleum
Engineers.
Guo, J. J., Zhang, S., Zhang, L. H., Qing, H., & Liu, Q. G. (2012). Well testing analysis for horizontal
well with consideration of threshold pressure gradient in tight gas reservoirs. Journal of Hydrodynamics,
24(4), 561-568.
Hong, A., Bratvold, R. B., Lake, L. W., & Ruiz Maraggi, L. M. (2018). Integrating model uncertainty in
probabilistic decline curve analysis for unconventional oil production forecasting. SPE/AAPG/SEG
Unconventional Resources Technology Conference 2018, URTC 2018, 1–21.
https://doi.org/10.15530/urtec-2018-2900625
Houzé, O., Viturat, D., & Fjaere, O. S. (1988). Data Analysis; The theory and practice of Pressure
Transient, Production Analysis, Well Performance Analysis, Production Logging and the use of
Permanent Downhole Gauge Data, KAPPA.
Hubbert, M. K. (1956, January). Nuclear energy and the fossil fuel. In Drilling and production practice.
American Petroleum Institute.
Ilk, D., Rushing, J. A., Perego, A. D., & Blasingame, T. A. (2008, January). Exponential vs. hyperbolic
decline in tight gas sands: understanding the origin and implications for reserve estimates using Arps'
decline curves. In SPE annual technical conference and exhibition. Society of Petroleum Engineers.
Illman, W. A. (2005). Type curve analyses of pneumatic single‐hole tests in unsaturated fractured tuff:
Direct evidence for a porosity scale effect. Water Resources Research, 41(4).
Illman, W. A., & Neuman, S. P. (2000). Type‐curve interpretation of multirate single‐hole pneumatic
injection tests in unsaturated fractured rock. Groundwater, 38(6), 899-911.
IHS Markit. (2018). Transient Typecurve Theory. Retrieved from
https://www.ihsenergy.ca/support/documentation_ca/Harmony_Enterprise/2018_3/content/html_files/ref_
materials/analysis_method_theory/typecurve/transient_typecurve_theory.htm
Iyke, A. C., & Princewill, O. N. (2018). Comparative Study of Oil Production Forecast by Decline Curve
Analysis and Material Balance. (April). https://doi.org/10.24018/ejers.2018.3.4.472
Jeyachandra, B., Sharma, A., Dwivedi, P., & Gupta, S. (2016, June). Reservoir Flow Regime Assisted
Multi-Segment Production Forecasting. In SPE Argentina Exploration and Production of Unconventional
Resources Symposium. Society of Petroleum Engineers.
Joshi, K., & Lee, W. J. (2013, February). Comparison of various deterministic forecasting techniques in
shale gas reservoirs. In SPE Hydraulic Fracturing Technology Conference. Society of Petroleum
Engineers.
Kanfar, M., & Wattenbarger, R. (2012, January). Comparison of empirical decline curve methods for
shale wells. In SPE Canadian Unconventional Resources Conference. Society of Petroleum Engineers.
Mahmoud, I. S., & Anderson, D. M. (2015). Practical Application of Type Curve Analysis to the
Interpretation of Multi-Stage Fractured Horizontal Well Data. SPE.
 URTeC 2878

Marapira, T. (n.d.). Production Decline Analysis in Unconventional Reservoirs using Modified

Hyperbolic Decline Model.
Mattar, L., & Andeerson, D. M. (2003). A Systematic and Comprehensive Methodology for Advanced
Analysis of Production Data. SPE Annual Technical Conference and Exhibition.
Mattar, L., & Moghadam, S. (2009). Modified Power Law Exponential Decline for Tight Gas. 1–11.
https://doi.org/10.2118/2009-198
Odagme, B. (2016). Evaluation of the Methodologies of Analyzing Production and Pressure Data of Tight
Gas Reservoir. (December).
Odi, U., Bacho, S., & Daal, J. (2019). Decline curve analysis in unconventional reservoirs using a
variable power law model: A barnett shale example. SPE/AAPG/SEG Unconventional Resources
Technology Conference 2019, URTC 2019. https://doi.org/10.15530/urtec-2019-39
Rosa, A. J., & Horne, R. N. (1983, January). Automated type-curve matching in well test analysis using
Laplace space determination of parameter gradients. In SPE Annual Technical Conference and
Exhibition. Society of Petroleum Engineers.
Sah, S. (2013). Development of New Decline Model for Shale Oil Reserves.
Seshadri, J., & Mattar, L. (2010). Comparison of power law and modified hyperbolic decline methods.
Society of Petroleum Engineers - Canadian Unconventional Resources and International Petroleum
Conference 2010, 2(October), 984–1000. https://doi.org/10.2118/137320-ms
Sharma, A., & Guttery, B. (2019). Modeling terminal decline rate in flow regime transition using
alternating conditional expectation non-linear regression methods. Society of Petroleum Engineers - SPE
Liquids-Rich Basins Conference - North America 2019, LRBC 2019.
Shin, H. J., Lim, J. S., & Shin, S. H. (2014). Estimated ultimate recovery prediction using oil and gas
production decline curve analysis and cash flow analysis for resource play. Geosystem Engineering,
17(1), 78–87. https://doi.org/10.1080/12269328.2014.886970
Tan, L., Zuo, L., & Wang, B. (2018). Methods of decline curve analysis for shale gas reservoirs. Energies,
11(3). https://doi.org/10.3390/en11030552
Tsoularis, A., & Wallace, J. (2002). Analysis of logistic growth models. Mathematical biosciences,
179(1), 21-55.
Warpinski, N. R., & Branagan, P. T. (1989). Altered Stress Fracturing. JPT 41 (9): 990–997.
Wattenbarger, R. A. (1998). Production Analysis of Linear Flow Into Fractured Tight Gas Wells. SPE.
Wilson, S. (2017). Production Decline Analysis In Forecasting Performance Of Producing Wells.
Politecnico Di Torino.
Xiong, H., Gao, S., & Li, H. (2017). Generate type well performance curves by combining multi-segment
decline models and calibrated numerical simulation models for UR wells in permian basin.
Unconventional Resources Technology Conference (URTEC).
Xu, J., Guo, C., Teng, W., Wei, M., & Jiang, R. (2015). Production performance analysis of tight oil/gas
reservoirs considering stimulated reservoir volume using elliptical flow. Journal of Natural Gas Science
and Engineering, 26, 827-839.
Zareenejad, M. H., Kalantari Asl, A., & Nasriani, H. R. (2012). Application of decline analysis in
fractured reservoirs, field case studies. Journal of Chemical and Petroleum Engineering, 46(1), 53-62.

View publication stats