Multi-Task Learning Framework for Reservoir Water Temperature Field Prediction Using Modal Decomposition and Federated Learning

Reservoir thermal stratification affects withdrawal temperatures and downstream aquatic ecosystems, making water temperature prediction important for sustainable reservoir operation. However, predicting two-dimensional water temperature fields across different operating conditions remains challenging. This study proposes a framework integrating proper orthogonal decomposition (POD), mutual information, dual-channel BiLSTM, bidirectional residual cross-attention, multi-task learning, and personalized federated learning. POD reduces field dimensionality, while multi-task regression heads directly predict modal coefficients for different forecast days. To account for differences among operating conditions, feature-extraction networks remain local, and only regression-network parameters are aggregated. Experiments were conducted using EFDC-simulated temperature fields from 75 operating conditions of Wudongde Reservoir. The model achieved MAE and RMSE values of 0.120 °C and 0.183 °C for 2-day forecasts and 0.314 °C and 0.449 °C for 10-day forecasts, respectively. Compared with full-model aggregation, theoretical parameter communication was reduced by 96.29%, with slightly higher prediction errors. The predicted temperature fields may support the assessment of selective-withdrawal conditions and downstream ecological temperature requirements, contributing to sustainable reservoir management.