Employing Neural Networks in Reservoir Simulation for Hydraulic-Fractured Unconventional Wells and a Case Study in STACK Area

Abstract This study aims to develop an integrated workflow for a Deep Neural Network (DNN)-based reservoir simulator to model hydraulic fracturing and production forecasts. Given the complex and nonlinear characteristics of reservoir behaviors, our goal is to implement a scalable, automated simulator capable of addressing these challenges. This workflow consolidates data from varied sources, offering a detailed perspective on reservoir dynamics which aids in real-time decision-making and boosts operational efficiency. The workflow utilizes an extensive dataset comprising both actual and synthetic data encompassing geological, geophysical, petrophysical, drilling, completion, and production details to generate customized reservoir model. The development involved several crucial phases: data preparation, feature selection and physical model integration, model selection, training, and validation. We employed deep neural network to enable the model to learn and predict reservoir responses during both the hydraulic fracturing and production stages. The approach utilizes a unique algorithm tailored to address the high-dimensional, nonlinear, and temporal complexities associated with hydraulic fracturing and subsequent production. The case study has chosen 100+ wells of the STACK area. Our results show that the DNN-based reservoir simulation substantially improves both the accuracy and speed of reservoir performance predictions when compared to conventional simulators. More importantly, the workflow was able to integrate fracture treatment and EUR optimization. This demonstrates the model's superior computational efficiency, which is vital for real-time decision-making. The application of customized model can adeptly handle the intricate dynamics of hydraulic fracturing and production for operational efficacy.