A method for analyzing interwell connectivity based on gated recurrent network with knowledge interaction

Traditional interwell connectivity analysis methods for water-flooding reservoirs suffer from two major limitations: insufficient integration of seepage physics, leading to poor interpretability, and inadequate temporal modeling, which fails to capture the dynamic evolution of injection–production relationships. To overcome these issues, this study proposes a Knowledge-Interactive Gated Recurrent Unit (KIGRU) model that integrates physical constraints with temporal deep learning. The model adopts a dual-subnet architecture: Net-INJ encodes injection rates and interwell connectivity through gate functions and connection matrices, while Net-VOL characterizes reservoir volume changes. By embedding material balance equations into the network design, the model ensures physical consistency, while GRU modules effectively capture long-term temporal dependencies. Numerical experiments on synthetic reservoir cases demonstrate that KIGRU outperforms conventional neural networks and the Capacitance-Resistance Model (CRM) in both history matching and production forecasting. The model accurately identifies high-permeability channels, quantifies non-equilibrium flow, and yields more reliable predictions of liquid production rates. These results confirm that KIGRU achieves a balance between physical interpretability and predictive accuracy, offering a practical and theoretically sound tool for interwell connectivity analysis.