Physics-constrained deep learning for reservoir thermal structure prediction: Enhanced interpretability and...

Rapid prediction of reservoir vertical thermal structure is crucial for implementation flexible reservoir optimization strategies aimed at ecological protection. Data-driven models can efficiently forecast water temperature dynamics but have limitations: limited measured data reduces the prediction accuracy, weak physical interpretability of stratification, and extrapolation to future scenarios remains unreliable. To overcome these challenges, this study proposes a physically constrained deep learning framework (P-DL). Mechanism-driven process models are used to augment training data and identify key influencing key factors. Vertical temperature profiles are transformed into physically interpretable parameters to describe stratification intensity and improve extrapolation through weak physical constraints. The Xiangjiaba (XJB) reservoir was selected as a case study, to compare the proposed framework was compared with Random Forest (RF), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) model. All models captured historical temperature variations, whereas P-DL provided more accurate representation of short-term local fluctuations. Among the outputs, parameter D reflects the temporal evolution of stratification, and the difference between A and B indicates overall intensity and peak timing. Under the SSP5-8.5 scenario, the framework outperformed others in predicting surface temperatures during strong stratification (RMSE: 0.83-1.1 °C; R²: 0.88-0.9) and showed superior consistency at both local and overall consistency (KLD: 2.85-5.71; KSS: 0.2-0.4). Overall, the framework improves prediction accuracy, physical interpretability, and extrapolation stability, providing a reference for intelligent thermal management of reservoirs. The hybrid model and weak physical constraints concept involved in the framework can also guide improving data-driven predictions for other environmental factors.