Deep Reinforcement Learning-Based Dual-Agent Framework for Wind Power Forecasting

Accurate ultra-short-term wind power forecasting is essential for reliable wind power integration and operational dispatch, yet the non-stationary and regime-dependent nature of wind power signals makes it difficult for any single forecasting model to remain robust under diverse operating conditions. To address this issue, we propose a dual-agent deep reinforcement learning framework for wind power forecasting. A TD3-based fusion agent first assigns adaptive ensemble weights to four heterogeneous base models according to the meteorological context and historical errors. A SAC-based compensation agent then further refines the fused prediction by correcting systematic residual bias through temporal error patterns. Experiments on the public SDWPF real-world wind power dataset show that the proposed method consistently outperforms individual models, static ensemble methods, and a dynamic fusion baseline across all forecast horizons. Compared with the best individual model, the proposed method reduces MAE from 163.74kW to 96.81kW, corresponding to a 40.9% improvement, while increasing R² from 0.7255 to 0.8686. Moreover, compared with Agent₁, the second-stage compensation further reduces MAE by 30.2%. These results demonstrate the effectiveness of the proposed dual-agent framework for ultra-short-term wind power forecasting.