A physics-informed deep learning model for inter-well connectivity analysis of waterflooding reservoirs

Abstract

Accurate identification of inter-well connectivity is crucial for the efficient development of waterflooding reservoirs. Traditional methods, such as tracer tests, well testing, and numerical simulations, are often costly and computationally intensive, while purely data-driven approaches suffer from limited physical interpretability and poor adaptability. This study proposes a physics-informed deep learning model for predicting inter-well connectivity, integrating actual injection-production performance data and static geological properties of the reservoir. The model employs a Long Short-Term Memory (LSTM) network to extract time-series features from injector-producer pairs, combines static geological attributes including permeability, porosity, and layer thickness, and computes an inter-well connectivity matrix via a cross-well attention mechanism. By incorporating the material balance equation as a physical constraint into the loss function, the model achieves end-to-end training. Additionally, a learnable distance-decay weighting mechanism is introduced to enhance the spatial consistency of the inferred connectivity. Evaluated on a multi-layer injection-production dataset from an actual oilfield, the model yields connectivity matrices that reasonably reflect the injection-production correlations, accompanied by low errors in liquid production history matching. Experimental results demonstrate that the integration of physical constraints significantly improves both the physical consistency and generalization capability of connectivity predictions.