Physics-Guided Spatiotemporal Injection-Production Integrated Deep Learning Model for Water-Cut Prediction in Waterflooding Reservoirs

Abstract

Accurate water-cut prediction in mature waterflooding reservoirs is essential for identifying remaining oil distribution, warning of water breakthrough, and optimizing injection--production strategies. However, conventional numerical simulation is computationally intensive, while purely data-driven models often lack physical interpretability and struggle to capture the complex spatiotemporal dynamics of high-water-cut reservoirs. To address this issue, this study proposes a physics-guided spatiotemporal injection--production integrated network (PSTI-Net) for multi-step water-cut forecasting. The network distills dynamic physical knowledge from a simulator into a lightweight student model. Three branches are built: a three-dimensional convolutional neural network (3DCNN) extracts spatial heterogeneity from saturation fields; time-of-flight constraints capture fluid arrival dynamics near the wellbore; an injection-production module computes correlation-weighted injection rates using geometric distance and time-lagged correlation. The fused features enter a cascaded Gated Recurrent Unit and Long Short-Term Memory (GRU-LSTM) for multi-step forecasting. Tested on 40 years of production data from 402 wells in a Xinjiang oilfield, the model achieves an R2 of 0.865, outperforming numerical simulation (R2=0.754) and baselines like LSTM and Informer, with an approximately sixfold speed advantage. Ablation studies show temporal features contribute most, while spatial and injection-production modules enhance stability. The method captures sudden water breakthroughs and high-permeability channeling reliably. By distilling simulator physics, PSTI-Net provides fast, physically consistent water-cut forecasts without solving flow equations, bridging reliability and efficiency for early warning, optimization, and rapid field-scale decisions.