Short-Term Electricity Price Forecasting via a Reinforcement Learning-Based Dynamic Soft Ensemble Strategy

To address the high volatility of spot market prices and the feature extraction limitations of single models, a short-term electricity price forecasting method based on a reinforcement learning dynamic soft ensemble strategy is proposed. First, a complementary dual-branch architecture is constructed: the CNN-LSTM-Attention branch mines local temporal features, while the Transformer branch captures long-range global dependencies. Second, the Q-learning algorithm is introduced to model weight optimization as a Markov Decision Process. An intelligent agent perceives fluctuation states to adaptively allocate weights, overcoming the rigidity of traditional ensembles. Case studies on PJM market data demonstrate that the proposed model outperforms advanced benchmarks in MAE and RMSE metrics. Notably, prediction accuracy is significantly improved during price spikes and negative price periods. The results verify that the strategy effectively copes with market concept drift, supporting reliable bidding and risk mitigation.