Soft Regime-Aware Residual Learning for Robust Wind Power Forecasting

Abstract

Accurate wind power forecasting is essential for reliable power system operation and renewable energy integration. However, strong non-stationarity, nonlinear dynamics, and rapidly changing operating conditions continue to challenge existing forecasting models. To address these issues, this paper proposes a Soft Regime-Aware Residual Learning framework based on a hybrid forecasting architecture. An OLinear model is adopted to capture global temporal patterns. At the same time, a Gaussian Mixture Model (GMM) is employed to learn probabilistic operating regimes, providing a soft representation of heterogeneous wind power states. A regime-aware residual expert is introduced to adaptively correct forecasting deviations, where LightGBM is adopted as the residual learner. In addition, physics-guided features, including ramp energy, turbulence indicators, wind shear, and wind speed statistics, are introduced to enhance robustness and physical interpretability. Experimental results on a real-world wind power dataset demonstrate that the proposed framework achieves a 4.45% reduction in MAE and a 7.69% reduction in MSE compared with the OLinear baseline, while consistently improving forecasting performance under high-ramp operating conditions. These results demonstrate the effectiveness of integrating global forecasting, probabilistic regime learning, and adaptive residual refinement for complex wind power prediction.