• • On rainy days, the proposed model reduces ERMSE and EMAE by 49.3% and 51.9% on average compared to a single BiLSTM model, directly improving forecast reliability under volatile weather—critical for grid frequency regulation and reducing reserve capacity requirements.
• • The entropy-based causal feature extraction method quantifies nonlinear coupling among meteorological factors, eliminating redundant inputs that plague linear correlation methods like Pearson coefficients, thereby enhancing feature selection quality and model generalization.
• • The parallel BiLSTM and TCN-Transformer architecture captures multi-scale temporal dependencies: BiLSTM handles short-term fluctuations, while TCN-Transformer extracts local patterns and long-range dependencies, achieving superior performance over single models in all three weather scenarios.
• • The improved Q-learning algorithm with a cosine-modulated dynamic reward-penalty mechanism adaptively adjusts sub-model weights, overcoming the static weight limitations of error inverse and mean fusion methods, and preventing premature convergence or local optima during training.