Applicability of deep neural networks on production forecasting in Bakken shale reservoirs

Abstract

The unconventional shale and tight formations, such as the Bakken in Williston Basin, are becoming an important hydrocarbon sources since the development of advanced horizontal drilling and multi-stage stimulation techniques. The appropriate completion and stimulation designs are crucial to maximize the well productivity and oil recovery. Although reservoir simulations are commonly used for production analysis, the modeling of hydraulic fractures and intrinsic complexity of shale have restricted an understanding of the fluid flow behavior. Owing to the available of a large amount of data in unconventional tight and shale reservoirs, data-driven approaches, such as deep neural networks (DNNs), provide an alternative approach for production analysis in these formations.In this study, a total of 2919 wells including 2780 multi-stage hydraulic fractured horizontal wells and 139 vertical wells in Bakken Formation were collected and analyzed using a deep learning method and Sobol's global sensitivity analysis. In the deep learning model, one-hot encoding method was used to deal with the categorical data. Xavier initialization, dropout technique, and batch normalization were evaluated to develop a reliable deep neural network model. Moreover, k-fold cross validation was used to evaluate the prediction ability and robustness of models. Prior to the performance evaluation of DNN algorithms, an optimum DNN model was achieved by optimizing the dropout layer, activation function, and hyperparameters of DNN models. It is found that the number of hidden layers and neurons in each layer have significant effects on the performance of models. The optimized DNN model has 3 hidden layers and 200 neurons for each layer. The prediction performance of DNN models is acceptable with the low mean squared errors for both 6- and 18-months cumulative oil productions. Furthermore, the importance of each parameter was studied using the Sobol's analysis. Results show that the average proppant placed per stage is the most important parameter, which accounts for more than 35% of variability in both 6- and 18-months cumulative oil productions. These data-driven models can be readily updated when the new well information becomes available. It is worth mentioning that the proposed DNN models can be directly integrated into existing hydraulic fracturing design routines.