Automated Reservoir History Matching Framework: Integrating Graph Neural Networks, Transformer, and Optimization for Enhanced Interwell Connectivity Inversion

Abstract

Understanding interwell connectivity during water-flooding reservoir development is crucial for analyzing the characteristics of remaining oil and optimizing technical measures. The key lies in establishing an inversion method to identify interwell connectivity. However, traditional history matching methods based on numerical simulation suffer from high computational costs and limited adaptability to complex spatiotemporal dependencies in production data. To address these challenges, this study combines a surrogate model trained using a graph neural network (GNN) and Transformer encoder with a differential evolution particle swarm optimization (DEPSO) algorithm for automated reservoir history matching. The surrogate model is constructed by embedding the capacitance–resistance model (CRM) into a graph structure, where wells are represented as nodes and interwell connectivity parameters as edge features. When applied to the conceptual model, the coefficient of determination (R²) was found to be approximately 0.95 during the training phase by comparing the production data predicted by the surrogate model with the actual observed data. The DEPSO algorithm aimed to minimize the differences between surrogate predictions and observed data, achieving good fitting results. When applied to a complex case study, the average water-cut fitting rate for each production well in its well group reached 87.8%. The results indicate that this method significantly improves fitting accuracy and has substantial practical value.