A Surrogate Model for Predicting Dynamic Reservoir Pressure Profiles in Complex Hydraulic Fracture Systems

Dynamic intelligent prediction of the production pressure field in unconventional reservoirs aids in optimizing hydraulic fracturing designs and improving decision-making. However, traditional numerical simulation methods struggle to balance accuracy and efficiency, while most current deep-learning-based surrogate models either consider limited factors or oversimplify fracture geometries. To address these challenges, this study proposes a novel dynamic surrogate model (Dy-Pre-Net) for predicting pressure field variations during postfracturing production. The model incorporates complex hydraulic fracture geometries along with spatial data such as permeability, porosity, and the initial pressure field. It combines a channel attention mechanism with a U-Net framework to predict pressure changes across the reservoir grid. During model development, three data sets were used, and transfer learning was applied to enhance training efficiency and performance. Case studies indicate that transfer learning significantly accelerates model training. Compared with numerical simulation results, the surrogate model exhibits relatively higher absolute errors in the fracture zones and near the reservoir boundaries during the early production phase, with a maximum absolute error of up to 1.2 MPa. However, these errors gradually decrease over time. Furthermore, the surrogate model requires less than 10 s for prediction. Compared to traditional reservoir simulators, the surrogate model achieves comparable accuracy in predicting reservoir pressure fields while significantly reducing computational time.