Multi-graph interactive oil production prediction network integrating physics priors and data-driven approaches

Abstract

In the digital transformation of smart oilfields, accurate oil production prediction is essential for optimizing production strategies. However, the production process is a complex nonlinear system involving coupled electrical, fluid, and thermodynamic domains where sensor data is high-dimensional and noisy. Traditional mechanistic models struggle with parameter calibration, while standard data-driven models like LSTM often overlook spatial topological relationships and physical constraints. To address these challenges, this paper proposes the multi-graph interaction learning network (MGILN), a “dual-track, three-stage” architecture integrating petroleum engineering knowledge with deep learning. Following feature embedding, the model constructs parallel “static physical prior” and “dynamic feature similarity” graphs to characterize rigid constraints and implicit statistical correlations. An innovative cross-graph interaction mechanism then aligns mechanistic logic with measured data, while gated spatio-temporal units capture long-term evolution. Experimental results on 10-dimensional features show that MGILN significantly outperforms benchmarks in accuracy and physical consistency. By utilizing a compound optimization function with differential physical loss and graph Laplacian regularization, this study balances interpretability and robustness, providing a technical foundation for intelligent decision support in smart oilfields.