Abstract
This study proposes the use of different machine learning techniques to predict the estimated ultimate recovery (EUR) as a function of the hydraulic fracturing design. A set of data includes 200 well production data, and completion designs were collected from oil production wells in the Niobrara shale formation. The completion design parameters include the lateral length, the number of stages, the total injected proppant and slurry volumes, and the maximum treating pressure measured during the fracturing operations. The data set was randomly split into training and testing with a ratio of 75:25. Different machine learning methods were to predict EUR from the completion design including linear regression, random forest (RF), and decision tree (DT) in addition to gradient boosting regression (GBR). EUR prediction from the completion data showed a low accuracy. As result, an intermediate step of estimating the well IP30 (the initial well production rate for the first month) from the completion data was carried out; then, the IP30 and the completion design were used as input parameters to predict the EUR. The linear regression showed some linear relationship between the output and the inputs, where the EUR can be predicted with a linear relationship with an R-value of 0.84. In addition, a linear correlation was developed based on the linear regression model. Moreover, the other ML tools including RF, DT, and GBR presented high accuracy of EUR prediction with correlation coefficient (R) values between actual and predicted EUR from the ML model higher than 0.9. This study provides ML application with an empirical correlation to predict the EUR from the completion design parameters at an early time without the need for complex numerical simulation analysis. Unlike the available empirical DCA models that require several months of production to build a sound prediction of EUR, the main advantage of the developed models in this study is that it requires only an initial flow rate along with the completion design to predict EUR with high certainty.
Similar content being viewed by others
Explore related subjects
Discover the latest articles, news and stories from top researchers in related subjects.Avoid common mistakes on your manuscript.
Introduction
In shale-gas reservoirs, the ultralow-permeability matrix is not capable to flow fluid at a feasible rate and to deliver an acceptable drainage volume. Horizontal drilling with multistage fracture completion has turned out to be the key stimulation technology for the development of shale plays (Beckwith 2011; King 2010; Wiley et al. 2004). Wells are to accomplish a series of fracturing processes, with a high injection rate, large fracturing slurry volume, and low proppant concentration, during the multistage design (Seale et al. 2006).
The economic feasibility and production improvement of an oil and gas well largely depend on the efficiency of hydraulically generated fractures. Reserve estimation is the key step for economic and investment calculations. Reserve, estimated ultimate recovery (EUR), is the estimation of oil and gas volume that can be economically extracted under the current technological constrictions. Estimating EUR in very tight reservoirs and shale plays has long been challenging. Reserves assessment is a process that is regularly updated the age of the reservoir production history. The data availability and the forecasting method are the two vital elements that determine the estimated EUR accuracy (Sidle and Lee 2010). Therefore, the development strategies are highly dependent on the accuracy of the forecast. The high-confidence EUR estimate is the total hydrocarbon recovery when the well reaches the abandonment conditions. However, no development action is required. Hence earlier and more accurate estimation of EUR is required for the field development.
EUR estimation methods
Several methods can be used to evaluate EUR in oil and gas reservoirs. These methods may be purely based on physics and other methods are built on empirical and analogy methods. Each technique has its limitations and inaccuracies (Alarifi and Miskimins 2021). These techniques include analogy theory, volumetric calculations, material balance, rate transient analysis (RTA), decline curve analysis (DCA), type curves, and numerical simulation.
Volumetric calculations and analogy methods depend on knowing the reservoir dimensions and petrophysical properties, which are difficult to be estimated in tight formations and early time development (Shanley et al. 2004; Sidle and Lee 2010). Similarly, material balance (MBE) techniques are based on PVT data to calculate hydrocarbon in-place and estimate EUR. However, MBE requires extended production data in addition to accurate reservoir pressure measurements. Moreover, MBE has many assumptions, which makes its applicability in unconventional reservoirs unsatisfactory. RTA is an analytical or numerical method widely used for the production forecast of unconventional reservoirs (E1-Banbi and Wattenbarger 1998; Ibrahim et al. 2020; Ibrahim and Wattenbarger 2005). However, RTA methods face drawbacks due to a lack of precise measurements of reservoir rock and fluid properties and appropriate awareness of the physics controlling multiphase flow conditions.
Numerical simulation and history-matching techniques can be used to estimate future well production and EUR. However, an accurate formation of petrophysical properties and fracture parameters is required, in addition to extended production data to be matched (Freeman et al. 2013; Miller et al. 2010; Olorode et al. 2013; Shanley et al. 2004; Sun et al. 2015).
The decline curve is the most commonly used technique for estimating the hydrocarbon EUR for its simplicity and low input parameters (Mahmoud et al. 2018). Arps’ decline method (Arps 1945) is the famous empirical correlation that has been generally used as the main industry model. However, Arp's method over-forecasts reserves when applied in tight permeability reservoirs (Agarwal et al. 1999; Sharma and Lee 2016). Several DCA models have been introduced to forecast production tailored for unconventional wells. A common method is the stretched exponential production decline (SEPD) (Valko 2009), in addition to the Arps hyperbolic decline with a “best-fit” hyperbolic decline exponent “b” value. But all are based on empirical observations of a particular scenario and have their own limitations, which can yield unreasonable EUR estimation (Ilk et al. 2008; Mahmoud et al. 2018; Miao et al. 2018; Valko 2009; Zhang et al. 2015; Alarifi 2021).
Machine learning
Machine learning (ML) has been used for diverse applications in oil and gas and environmental problems. Different ML tools such as artificial neural network (ANN), random forest (RF), decision tree (DT), gradient boosting regression (GBR), function networks (FN), adaptive neuro-fuzzy inference system (ANFIS), and support vector machine (SVM) can be applied to predict specific parameters from readily accessible data (Abdelgawad et al. 2019; Al Dhaif et al. 2022; Alarifi and Miskimins 2021; Ibrahim et al. 2022; Moussa et al. 2018).
ANN machine learning method includes three main layers types. These can be defined as input and output layers, in addition to hidden layers. The input layer includes the input parameters that are handled by neurons within the hidden layers to finally calculate the output layer. The layers are linked with a set of weights and biases that are boosted by many training processes. These weights and biases can be modified to ultimately accomplish the lowest achievable error in the objective (Hinton et al. 2006). RF technique includes many decision trees (DT) that attain great performance in a low-dimensional dataset. In RF, many trees are structured together to overcome the overfitting problem in the single DT technique by adjusting different hyperparameters to improve the model accuracy (Yarveicy et al. 2019). GBR combines multiple simple models into a single composite model. This is also why boosting is known as an additive model, since simple models (weak learners such as DT or LR) are added one at a time while keeping existing trees in the model. As more simple models are added, the complete final model becomes a stronger predictor. The term "gradient" in "gradient boosting" reflects the fact that the GBR uses gradient descent to minimize the loss.
Multistage hydraulic fracturing is the main stimulation technique in developing shale formations. EUR is the key parameter to evaluate the well profitability and different methods can be used to estimate the formation reserve, including volumetric calculations, material balance equation, numerical simulation, and decline curve analysis. These techniques are associated with different inputs and cannot be used in the early life of the well. Therefore, this study emphasizes developing a new methodology to estimate EUR for multistage fractured horizontal shale wells in the Niobrara formation using different machine learning tools. The well-completion data were used as input to forecast future production.
Methodology
Data analysis
A data set of production data and completion design were collected points were collected from around 200 horizontal wells completed in Niobrara shale formations, where each well represents one realization. These wells reached the abandonment conditions. Hence, EUR for each well is basically the total production until the abandonment conditions. The data set includes the completion design parameters which include the lateral length (L), the number of stages(N), the total injected proppant (Mprop)and slurry volumes (Vslurry), and the maximum treating pressure (Pmax) measured during the fracturing operations. Table 1 reviews the different statistical parameters for the data set to define the data position and range, in addition to the distribution shape. The ranges of the parameters are as follows: the lateral length 1500–11,200 ft, the number of stages 5–62, and the corresponding EUR varied from 1.2E4 to 3.9E5 BOE. Figure 1 shows a matrix plot for the collected dataset to present the connections between the input completion parameters and the EUR. The diagonal displays the spreading of the data and their ranges.
Pearson’s correlation coefficients (R) were used to describe the relationship between the input and the output parameters. The two coefficients were calculated using the following equations.
where R is Pearson’s correlation coefficient between the output parameter and the input parameters, \({x}_{i}\) is the independent feature which includes the lateral length, the number of stages, the total injected proppant and slurry volumes, and the maximum treating pressure measured during the fracturing operations. \({y}_{i}\) is the dependent parameter. \({\upmu }_{\mathrm{x}}, {\upmu }_{\mathrm{y}}, and {\upsigma }_{\mathrm{Rx}}, {\upsigma }_{\mathrm{y}}\) are the mean and the standard deviations.
Figure 2 presents a heat map for Pearson’s correlation coefficients between all the input features with each other and with the EUR. Generally, R varies from − 1 to 1. At R of − 1, the EUR is a strong inverse related to the completion parameter. For R of 1, a strong direct relationship is found between the EUR and the completion parameter. A strong relationship was found between the EUR and the initial flow rate with an R-value of 0.81. In addition, strong relation was found between the EUR and the well length and the number of stages with R-values higher than 0.5. Most of the parameters' R-values are at least 0.5 between the different completion parameters and EUR except the maximum treating pressure showed an R of 0.23.
Model development
Different machine learning models (ML) including linear regression (LR), random forest (RF), and decision tree (DT) in addition to gradient boosting regression (GBR) were applied to the dataset to forecast the EUR. The dataset (200 data points) was used to build the model after optimizing the splitting ratio of the training to testing datasets. The quality of the model was measured using absolute average error (AAPE) which represents the error between the actual EUR from the well production data and the estimated values of EUR from the ML model.
where \({y}_{i\mathrm{ actual}}\) and \({y}_{i\mathrm{ predicted}}\) are the actual and the estimated output value (EUR\()\), respectively, and N is the number of points in the dataset. The correlation coefficient (R) was used as the goodness of fit indicator between the actual EUR from the well production data and the estimated EUR value from the model, and it was calculated using Eq. 2, where \({\mathrm{x}}_{\mathrm{i}}\) and \({\mathrm{y}}_{\mathrm{i}}\) are the actual and the estimated EUR values, respectively.
Figure 3 presents a schematic for the different model-developing processes. After data collection and transformation, the data set were applied to train and test the ML models. Different training/testing ratios were testing to training set to be from 70 to 90%.
Results
Linear regression model results
LR module was implemented on the well-completion input parameters to calculate the EUR and to examine the linearity between the output and the inputs. The completion data were splatted to an optimized training /testing ratio of 70/30%. Figure 4 shows the actual versus the predicted EUR values. The results showed low accuracy during the training and testing data sets with R-values of 0.64 and 0.62, respectively. These results reflect the underfitting problem and there is a need for adding other input features for the model development. Hence, the initial flow rate was required to be added to the input features similar to most of the DCA tools where the IP30 is an essential parameter for the EUR prediction.
Therefore, an intermediate step of predicting the IP30 from the completion data was introduced. The predicted IP30 was then used as an input for the EUR prediction model. Figure 5 presents the LR cross-plot in both training and testing sets for the EUR model with perfect alignment with the 45-degree line. The LR model was able to predict the EUR with R-value of 0.88 and 0.80 for training and testing data sets, respectively.
These results prove that there are some linearity between the output and the inputs, where the EUR can be predicted with a linear relationship with an average R-value of 0.84. Moreover, a linear correlation was developed based on the linear regression model weights and biases as shown in Eq. 3. The linear correlation can be used to directly predict the EUR as a function of the completion design parameters with a fair accuracy
where X1-7 are the normalized input parameters including the number of stages, the lateral length, the total vertical depth, the maximum treating pressure measured during the fracturing operations the total injected proppant and slurry volumes, and the initial production rate. The normalization was done by subtracting the minimum value from the parameter value then dividing the result by the difference between the maximum and the minimum of each input parameter. The maximum and minimum values for each input parameter are presented in Table 1.
RF model results
Similar to LR model, RF technique was implemented on the well-completion input parameters to calculate the EUR. Figure 6 shows the actual versus the predicted EUR values. The model shows low accuracy, especially in the testing set with R-value of 0.79. These results reflect the underfitting problem similar to LR, and there is a need for adding the initial flow rate.
Therefore, an intermediate step of predicting the IP30 from the completion data was introduced. The predicted IP30 was then used as an input for the EUR prediction model. For the IP30 model, the optimum parameters for RF models were maximum features = “SQRT”, Maximum depth = 20, and the number of estimators = 150. Figure 7a presents the RF cross-plot in both training and testing sets for the IP30 model with perfect alignment with the 45-degree line. The RF model was able to predict the IP30 with R-value of 0.98 and 0.95 for both training and testing data sets, respectively. And the AAPE was estimated to be 7%.
The predicted IP30 was used as an input in addition to the completion data to the RF model to predict the EUR. Similar to the IP30 model, the hyperparameters were updated to reach the best model performance with maximum features = ‘Auto”, Maximum depth = 30, and the number of estimators = 100. Figure 7b presents the RF cross-plot in both training and testing sets for the EUR model with perfect alignment with the 45-degree line. The RF model was able to predict the EUR with R-value of 0.99 and 0.93 for training and testing data sets, respectively.
Figure 8 shows the distribution of the residuals, where the residual is randomly scattered around zero. The residuals follow the normal distribution, which displays that the scattering degree of the points is similar for all fitted EURs with no biases towered at the high or the low end of the EUR range.
Decision tree model results
DT technique was implemented on the well-completion input parameters to calculate the EUR. DT is a special case of RF where only one decision tree is used. DT resolves the ML problem by transforming the data into tree representation. Each internal node of the tree representation represents an attribute and each leaf node denotes a class label. The DT hyperparameters were optimized to reach the best model performance with maximum features = ‘Auto”, Maximum depth = 4, and minimum samples split = 2. Figure 9 presents the DT cross-plot in both training and testing sets for the EUR model with perfect alignment with the 45-degree line with R-value of 0.93 and 0.85 for training and testing data sets, respectively.
Gradient boosting regressor model results
GBR combines multiple simple models into a single composite model. This is also why boosting is known as an additive model, since simple models (weak learners such as DT or LR) are added one at a time, while keeping existing trees in the model. As more simple models are added, the complete final model becomes a stronger predictor. The term "gradient" in "gradient boosting" reflects the fact that the GBR uses gradient descent to minimize the loss. The GBR hyperparameters were optimized to reach the best model performance with loss function of = 'squared error', learning rate = 0.1, number of estimators = 100, subsample = 1.0, minimum samples split = 2, minimum samples leaf = 1, minimum weight fraction leaf = 0.0, and maximum depth = 3. GBR showed the highest EUR prediction performance. As shown in Fig. 10, the GBR cross-plot in both training and testing sets for the EUR model with perfect alignment with the 45-degree line with R-value of 0.99 and 0.94 for training and testing data sets, respectively and AAPE less than 5%.
Figure 11 shows the bar chart for the distribution of the residuals, which are randomly scattered around zero. This trend displays that the scattering degree of the points is similar for all fitted EURs with no biases towered at the high or the low end of the EUR range.
ANN model results
Similar to the RF model, the ANN technique was implemented on the collected data points to develop the ANN model. The optimized hyperparameters were designated based on the best model performance indicators. The ANN model was built with one hidden layer with 8 neurons. The “trainlm” function was selected to be the training function with 'logsig' as the data transfer function. Figure 12 presents the actual EUR values versus the estimated EUR cross-plot from the ANN model in training and testing datasets. Figure 7 shows the capability of the ANN model to calculate the EUR as a function of completion data and initial flowrate with a good alignment with the 45-degree line. The R-value for the training set with found to be 0.96 with an AAPE error of 13%. The testing data results AAPE of 12% with R-values of 0.12.
Moreover, an equation was built using the weights and the biases from the developed ANN model including one hidden layer, 8 neurons, and the transformation function of 'logsig'. Equation 4 presents the developed correlation using the weights and biases;
where \(W{2}_{i}\) are the weights for the neurons from the hidden layer to the output layer and their bias is \({b}_{2}\). \(W{1}_{i,1-7}\) is the weights for the neurons from the input layer to the hidden layer for the input parameters (x1-7); the number of stages, the lateral length, the total injected proppant and slurry volumes, the maximum treating pressure measured during the fracturing operations, and the initial production rate and \(b{1}_{i}\) is the improved biases for the hidden layer related with each neuron (i) from 1 to (neurons number (n) = 8). This equation was established based on the modified weights and biases of the optimized ANN model. The tuned weights and biases of the EUR model are recorded in Table 2 to be a substitute in Eq. 4.
Multistage hydraulic fracturing is the main stimulation technique in developing shale formations. EUR is the key parameter to evaluate the well profitability, and different methods can be used to estimate the formation reserve, including volumetric calculations, material balance equation, numerical simulation and decline curve analysis. These techniques are associated with different input and cannot be used in the early life of the well. Hence, the current study provides an application of different machine learning tools to estimate the well EUR as a function of the completion design at the early time of well life.
Conclusions
The estimated ultimate recovery and reserve is the vital parameter to evaluate the shale well profitability. Volumetric calculations, material balance equations, numerical simulations, and decline curve analysis are the typical methods to estimate hydrocarbon reserves. These techniques are associated with different inputs and cannot be used in the early life of the well. Therefore, this study emphasizes developing a new methodology to estimate EUR for multistage fractured horizontal shale wells in the Niobrara formation using different machine learning tools. The applied ML tools include weaker learners, such as linear regression and decision trees, and stronger learners, including gradient boosting regression, random forest, and neural networks. The input parameters were the different completion parameters such as lateral length, the number of stages, total injected proppant and slurry volumes, and the maximum treating pressure measured during the fracturing operations. The following are the main findings:
-
1.
The estimated ultimate recovery prediction requires well IP30 (initial well production rate) for accurate estimation.
-
2.
The linear regression showed some linearity between the output and the inputs, where the estimated ultimate recovery can be predicted with a linear relationship with an R-value of 0.84.
-
3.
Gradient boosting regression model showed the highest performance of predicting EUR from the completion data with an R-value of 0.99 and 0.94 in the case of the training and testing data sets, respectively.
-
4.
The estimated ultimate recovery of the multistage fractured horizontal wells is highly dependent on the lateral length and the number of stages.
-
5.
The developed ML models can be applied to accurately estimate the estimated ultimate recovery at the early stage of the well without the need of conducting expensive numerical simulations or wait until a very late stage of the well’s life.
The proposed models are expected to have the same performance if the data set used falls in the same range of the data used for training the predictive models. It is recommended to update the parameters of the models when used for another field having different inputs' ranges and geologic properties to ensure a robust prediction with reasonable accuracy.
Change history
19 June 2023
This article has been retracted. Please see the Retraction Notice for more detail: https://doi.org/10.1007/s13202-023-01666-7
Abbreviations
- b:
-
Arp’s hyperbolic decline exponent
- \(b{1}_{i}\) :
-
Improved biases for the hidden layer related with each neuron i
- \({b}_{2}\) :
-
Bias the hidden layer and the output layer
- IP30:
-
Initial production rate in the first month
- L:
-
Well lateral length
- Mprop :
-
Total injected proppant, Ib
- N:
-
Number of stages
- n:
-
The number of data points
- Pmax :
-
Maximum treating pressure measured during the fracturing operations.
- R:
-
Correlation coefficient
- Vslurry :
-
Slurry volume, bbl
- \(W{1}_{i,1-7}\) :
-
Weights for the neurons from the input layer to the hidden layer for the input parameters
- \(W{2}_{i}\) :
-
Weights for the neurons from the hidden layer to the output layer and
- \({x}_{i}\) and:
-
The independent parameter
- \({y}_{i}\) :
-
The dependent parameters
- \({\sigma }_{x} and {\sigma }_{y}\) :
-
The standard deviation for the independent and dependent parameters
- \({\mu }_{x} and {\mu }_{y}\) :
-
Mean for the independent and dependent parameters
- AAPE:
-
Average absolute percentage error
- ANFIS:
-
Adaptive neuro-fuzzy inference system
- ANN:
-
Artificial neural network
- BOE:
-
Barrel of oil equivalent
- DT:
-
Decision tree
- DCA:
-
Decline curve analysis
- EUR:
-
Estimated ultimate recovery
- GBR:
-
Gradient boosting regression
- LR:
-
Linear regression
- MBE:
-
Material balance equations
- ML:
-
Machine learning
- RF:
-
Random forest
- RTA:
-
Rate transient analysis
- SEPD:
-
Stretched exponential production decline
- SVM:
-
Support vector machine
References
Abdelgawad K, Elkatatny S, Moussa T, Mahmoud M, Patil S (2019) Real-time determination of rheological properties of spud drilling fluids using a hybrid artificial intelligence technique. J Energy Resour Technol 141(3):032908
Agarwal RG, Gardner DC, Kleinsteiber SW, Fussell DD (1999) Analyzing well production data using combined-type-curve and decline-curve analysis concepts. SPE Reserv Eval Eng 2:478–486
Al Dhaif R, Ibrahim AF, Elkatatny S (2022) Prediction of surface oil rates for volatile oil and gas condensate reservoirs using artificial intelligence techniques. J Energy Resour Technol Trans, ASME. https://doi.org/10.1115/1.4051298/1109768
Alarifi SA (2021) Production data analysis of hydraulically fractured horizontal wells from different shale formations. Appl Sci 11:2165. https://doi.org/10.3390/app11052165
Alarifi SA, Miskimins J (2021) A new approachpl to estimating ultimate recovery for multistage hydraulically fractured horizontal wells by utilizing completion parameters using machine learning. SPE Prod Oper 36:468–483
Arps JJ (1945) Analysis of decline curves. Trans AIME 160:228–247
Beckwith R (2011) Shale gas: promising prospects worldwide. J Pet Technol 63:37–40. https://doi.org/10.2118/0711-0037-JPT
E1-Banbi AH, Wattenbarger RA (1998) Analysis of linear flow in gas well production, in: spe gas technology Symposium, Calgary, Alberta, Canada, March 1998. Paper Number: SPE-39972-MS. OnePetro, pp 1–18. https://doi.org/10.2118/39972-MS
Freeman CM, Moridis G, Ilk D, Blasingame TA (2013) A numerical study of performance for tight gas and shale gas reservoir systems. J Pet Sci Eng 108:22–39
Hinton GE, Osindero S, Teh Y-W (2006) A fast learning algorithm for deep belief nets. Neural Comput 18:1527–1554. https://doi.org/10.1162/neco.2006.18.7.1527
Ibrahim AF, Assem A, Ibrahim M (2020) A novel workflow for water flowback RTA analysis to rank the shale quality and estimate fracture geometry. J Nat Gas Sci Eng 81:103387
Ibrahim AF, Elkatatny S, Abdelraouf Y, Al Ramadan M (2022) Application of various machine learning techniques in predicting water saturation in tight gas sandstone formation. J Energy Resour Technol. 144:83009
Ibrahim M, Wattenbarger RA (2005) Analysis of rate dependence in transient linear flow in tight gas wells, In: Canadian International Petroleum Conference 2005, CIPC 2005,, Calgary, Alberta, June (2005). Paper Number: PETSOC-2005–057. OnePetro. https://doi.org/10.2118/2005-057
Ilk D, Rushing JA, Perego AD, Blasingame, TA (2008) Exponential vs. hyperbolic decline in tight gas sands: understanding the origin and implications for reserve estimates using Arps’ decline curves, presented at the SPE Annual Technical Conference and Exhibition, Denver, Colorado, USA, September 2008. Paper Number: SPE-116731-MS
King GE (2010) Thirty years of gas-shale fracturing: what have we learned? J Pet Technol 62:88–90. https://doi.org/10.2118/1110-0088-JPT
Mahmoud O, Ibrahim M, Pieprzica C, Larsen S (2018) EUR prediction for unconventional reservoirs: state of the art and field case. presented at the SPE trinidad and tobago section energy resources conference, Port of Spain, Trinidad and Tobago, June 2018. Paper Number: SPE-191160-MS. https://doi.org/10.2118/191160-MS
Miao Y, Li X, Lee J, Zhou Y, Wu K, Sun Z, Liu S (2018) A new rate-decline analysis of shale gas reservoirs: coupling the self-diffusion and surface diffusion characteristics. J Pet Sci Eng 163:166–176
Miller MA, Jenkins C, Rai R (2010) Applying innovative production modeling techniques to quantify fracture characteristics, reservoir properties, and well performance in shale gas reservoirs, Paper presented at the SPE Eastern Regional Meeting, Morgantown, West Virginia, USA, October 2010. Paper Number: SPE-139097-MS
Moussa T, Elkatatny S, Mahmoud M, Abdulraheem A (2018) Development of new permeability formulation from well log data using artificial intelligence approaches. J Energy Resour Technol 140(7):072903
Olorode OM, Freeman CM, Moridis GJ, Blasingame TA (2013) High-resolution numerical modeling of complex and irregular fracture patterns in shale-gas reservoirs and tight gas reservoirs. SPE Reserv Eval Eng 16:443–455
Seale R, Donaldson J, Athans J (2006) Multistage fracturing system: improving operational efficiency and production, in: spe eastern regional meeting, Canton, Ohio, USA, October 2006. Paper Number: SPE-104557-MS. OnePetro, pp 218–225. https://doi.org/10.2118/104557-MS
Shanley KW, Cluff RM, Robinson JW (2004) Factors controlling prolific gas production from low-permeability sandstone reservoirs: Implications for resource assessment, prospect development, and risk analysis. Am Assoc Pet Geol Bull 88:1083–1121
Sharma, A., Lee, W.J.,: Improved workflow for EUR prediction in unconventional reservoirs, Paper presented at the SPE/AAPG/SEG Unconventional Resources Technology Conference, San Antonio, Texas, USA, August 2016. Paper Number: URTEC-2444280-MS (2016).
Sidle RE, Lee WJ (2010) An update on the use of reservoir analogs for the estimation of oil and gas reserves, Paper presented at the SPE Hydrocarbon Economics and Evaluation Symposium, March 8–9, 2010. Paper Number: SPE-129688-MS
Sun H, Chawathe A, Hoteit H, Shi X, Li L (2015) Understanding shale gas flow behavior using numerical simulation. Spe J 20:142–154
Valko PP (2009) Assigning value to stimulation in the Barnett Shale: a simultaneous analysis of 7000 plus production hystories and well completion records, Paper presented at the SPE Hydraulic Fracturing Technology Conference, January 19–21, 2009. Paper Number: SPE-119369-MS
Wiley C, Barree B, Eberhard M, Lantz T (2004) Improved horizontal well stimulations in the bakken formation, Williston Basin, Montana, in: SPE annual technical conference and exhibition , Houston, Texas, September 2004. Paper Number: SPE-90697-MS. OnePetro, pp 3559–3568. https://doi.org/10.2118/90697-MS
Yarveicy H, Saghafi H, Ghiasi MM, Mohammadi AH (2019) Decision tree-based modeling of CO2 equilibrium absorption in different aqueous solutions of absorbents. Environ Prog Sustain Energy 38:S441–S448
Zhang HE, Cocco M, Rietz D, Cagle A, Lee J (2015) An empirical extended exponential decline curve for shale reservoirs, Paper presented at the SPE Annual Technical Conference and Exhibition, September 28–30, 2015. Paper Number: SPE-175016-MS
Funding
No external fund for this research and the authors would like to thank KFUPM for giving permission to publish this work.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that there is no conflict of interest regarding the publication of this paper.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This article has been retracted. Please see the retraction notice for more detail: https://doi.org/10.1007/s13202-023-01666-7
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Ibrahim, A.F., Alarifi, S.A. & Elkatatny, S. RETRACTED ARTICLE: Data-driven EUR for multistage hydraulically fractured wells in shale formation using different machine learning methods. J Petrol Explor Prod Technol 13, 1123–1134 (2023). https://doi.org/10.1007/s13202-022-01602-1
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13202-022-01602-1