Whole-process interpretable wind speed combination forecasting with multi-objective optimization and reinforcement learning

Wind speed forecasting is challenged by strong volatility and non-stationarity, which often causes not only large errors but also pronounced error fluctuations over time. This study proposes a whole-process interpretable ensemble framework that explicitly treats ensemble weighting as a multi-objective optimization problem, jointly considering prediction error and error fluctuation. Meteorological features are constructed and then compacted via two-stage selection, while an optimized variational mode decomposition is used to characterize the wind speed series. Three deep forecasting models are integrated (Informer–BiGRU, Crossformer–LSTM, and TCN–BiLSTM), and the ensemble weights are learned by a chaotic evolutionary reinforcement-learning-based multi-objective optimizer. To provide verifiable interpretability, we build a closed-loop explanation pipeline combining feature-level analysis (TreeSHAP/RollingSHAP), deep-model attribution (LRP), and ensemble-level complementarity analysis (IMI), further validated by model-agnostic checks (Kernel SHAP, AOPC, and counterfactual perturbations). Experimental results show competitive forecasting performance and a consistent, verifiable interpretability process, offering practical insights for wind power operation.