Applying Machine Learning to Oil Rate Fall Prediction for Bakken Shale Oil Wells | Scientific Reports – Nature.com

Literature review

The main objective of this study is to develop an ML-based model that can be used for the prediction of production rate decline for a large number of Bakken Shale wells over a very short period. This method will be much faster than commercial tank simulators because it does not require solving a large number of finite difference equations. Production from unconventional shale oil and gas began many years ago in the United States. Since then, many exploration companies have collected data from a significant number of oil and gas wells drilled and produced from these reservoirs, resulting in a large amount of horizontal well data. This information is available in several publicly accessible website databases¹. Various data analysis methods can be used to evaluate publicly available data to uncover underlying patterns and equilibrium points in these reservoirs that could be beneficial for future horizontal well development.^2,3,4. The most widely used method for projecting future production from shale oil wells is the projection of production decline curves.⁵. Decline curve models are mathematical equations used to model existing well production data and predict future well decline¹. Developing an empirical model of the production rate decline from the initial well performance and extrapolating this model into the future can predict future production potential and EUR. The most commonly used production decline curve model is the hyperbolic Arps model. However, fitting the Arps hyperbolic model to shale oil well production data has often resulted in physically unrealistic values ​​of the hyperbolic decline coefficient.¹. SEDM has been used to predict production from unconventional wells to solve this challenge⁵. SEDM is more suitable for shale oil wells than the hyperbolic Arps model, because they are in a transient flow regime for most of their lifetime. For the positive \({q}_{i}\)n and SEDM, SEDM returns a finite EUR value¹. Accordingly, SEDM was used in the study to predict the decline in production rate and EUR for the test wells.

In a similar study, an alternative approach for flow/pressure deconvolution was presented. Trained physics-based parameters and algorithms play a key role in the effective implementation of the recommended strategy by preserving the physics of superposition transient flows⁶. The main drawback of this study is that this method does not give satisfactory results when highly variable and limited data are available. The main drawback of this study is that it is highly dependent on the availability of a sufficient amount of data. Another study proposed a model to predict permeability of technically challenging (extremely heterogeneous) carbonate rock based on Random Forest regression, which can effectively acquire dependent physical parameters and provide assured permeability prediction compared to models. conventional empirics.^seven. The main drawback of this study is that it is highly dependent on the availability of good quality noise-free data. In a similar study, the authors used data-driven modeling to predict the rate of decline of Eagle Ford Shale oil wells.⁸. Another study proposed an ANN-based model to predict the rate of decline of Eagle Ford Shale oil wells⁹. The main drawback of these studies was that their applicability was limited only to Eagle for shale oil wells.

In a similar study, fuzzy logic, ANN (artificial neural network) and imperialist competitive algorithms were compiled to build an oil flow prediction model.^ten. The main drawback of this study is the determination of the optimized ANN architecture. Another study compiled several machine learning algorithms to predict porosity and permeability through the inclusion of petro-physical logs¹¹. The main drawback of this study is the involvement of complicated machine learning algorithms that take too long. In another study, the authors presented a deep belief network (DBN) model to predict the production of unconventional wells reliably and accurately. The authors ran 815 numerical simulation cases to develop a database for model training and to optimize hyper parameters using the Bayesian optimization algorithm. The proposed modeling framework was able to predict production from unconventional wells more reliably and accurately than compared to traditional machine learning techniques. The main limitation of this study is that training the model requires many simulations to be performed.¹².

Search problem

Commercial reservoir simulators can take hours or even days to predict the rate drop of a single well^13,14,15,16. Commercial reservoir simulators solve the discretized form of the mass balance equations. The number of grid blocks used in a reservoir model can be in the millions, which requires solving matrix equations million by million. As the reservoir becomes more and more heterogeneous and complicated, a finer resolution model (with a higher number of grid blocks) must be used. Additionally, accurate and comprehensive reservoir parameters including porosity, permeability, saturation, and other variables are essential to run one or even multiple reservoir simulations for the wells considered in the study, which are not not always available in the field.

Purpose and novelty

An alternative method based on machine learning was presented in this study which is very fast and accurate as it does not require solving matrix equations. It makes predictions based on previously collected field data. Machine learning can be used as an effective tool to predict oil rate decline in the type of data presented in this study. This study took less than a minute to estimate the rate of decline for all the wells used for the predictions. In machine learning based predictions, it has been observed that using the full data set to develop a machine learning model can lead to considerable errors due to data variability. To overcome this limitation, an alternative approach used in this work included cross-validation using k-fold validation and model averaging using the ensemble technique (Polyak–Ruppert averaging). This method divides the training data into several folds (k-folds) or subsets of data points, and a model was evaluated in one of the folds while the other folds were used for training. Therefore, by applying different subsets of training data to minimize the overfitting problem, we will have multiple machine learning models derived from a single training dataset at the end of training. The final prediction for test data/new data is based on a weighted average of the predictions made by all these models.

In this study, the ranking of variables was used to show which variables/parameters have a significant impact on the prediction of the decline in the rate and to rank them in order of priority. This data analysis was performed to understand the dataset before using it to make predictions. This study also used exploratory analysis to incorporate human judgment for more accurate conclusions.

Field of study

North Dakota, South Dakota, Montana, Manitoba and Saskatchewan are all part of the Williston Basin, which includes the Bakken Shale and its Three Forks. The Bakken Shale can be seen in Fig. 1 with oil and gas wells (Natural Gas Intelligence). All oil wells in this study were selected from the Bakken shales in Richland County (included in the green rectangle). SEDM was employed in this work to predict the decline in production.