A hybrid physics-informed machine learning framework for water cut prediction in waterflooding reservoirs

• Unified mathematical model comprehensively describes convex, S-shaped, and concave water cut trends in diverse reservoirs. • LightGBM-based feature optimization identifies key drivers including cumulative injection and well productivity patterns. • Novel hybrid model integrates machine learning with a generalized water cut prediction framework, achieving R²=0.994 accuracy. • Field validation demonstrates 0.75% slower water cut rise and 262 million CNY economic gains through development optimization. • Physics-ML synergy establishes a transferable paradigm for data-physics integration in reservoir engineering challenges. The variation of water cut plays a crucial role in optimizing waterflooding development strategies and delaying production decline. This study proposes a novel approach that combines machine learning with a general water cut prediction model to improve forecasting accuracy and applicability. Based on the data from 140 waterflooding blocks in Daqing Oilfield, a general mathematical model is established to uniformly describe various water cut trends, including convex, sigmoidal, and concave shapes, to meet the development needs of different reservoir types. Furthermore,key features were selected through feature importance analysis and recursive feature elimination. Machine learning models, including Random Forest and LightGBM, were constructed and optimized, with a 70%/15%/15% training/validation/test split and bootstrapping for rigorous evaluation. Finally, a hybrid model was developed by integrating the general water cut prediction model with the best-performing LightGBM model. The predictive performance of the hybrid model was evaluated on the DBGDD block over a historical period (1963-2021) and for a future forecast horizon (2021-2024). It significantly outperforms that of the individual models, achieving R² = 0.994 (95% CI: 0.984-0.996) and RMSE = 2.46 (95% CI: 2.25-2.68). The results indicate that the hybrid model not only improves prediction accuracy and stability but also demonstrates strong generalization ability and practical applicability, as further validated in two additional blocks (N2DG and X7DSP). This method provides a precise water cut prediction tool for waterflooding development, enhancing waterflooding efficiency and economic benefits, while offering valuable insights for the integration of physical models with machine learning in other engineering fields.