Explainable Prediction of Power Generation for Cascaded Hydropower Systems Under Complex Spatiotemporal Dependencies

Hydropower plays a key regulating role in new-type power systems, and both forecasting accuracy and interpretability are critical for power dispatch. However, cascade hydropower forecasting is constrained by strong spatiotemporal coupling among multi-dimensional features, flow propagation delays, as well as the limited transparency of deep learning models. To tackle these issues, this paper develops a hybrid framework integrating Maximal Information Coefficient (MIC), the Long- and Short-term Time-series Network (LSTNet), and the SHapley Additive exPlanations (SHAP) interpretability method. First, an MIC-based nonlinear screening mechanism is employed to remove redundant noise and construct a high-quality input space. Second, an LSTNet model is developed to deeply extract spatiotemporal coupling features among cascade stations and flow evolution patterns, achieving high-accuracy forecasting of both system-level and station-level outputs. Finally, SHAP is used for global and local interpretability analysis to perform physics-consistency verification with respect to the model’s decision-making rationale. Experimental results indicate that the proposed approach achieves low errors in total output forecasting, reducing error levels by approximately 57–88% compared with Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), and Informer. Moreover, SHAP feature-dependence analysis reveals a nonlinear response change of station D around 7.8 MW, providing evidence for the physical consistency of the model outputs and improving model interpretability.