Short-Term Wind Power Forecasting Based on Sigmoid-Weighted Soft-Switching Strategy Between XGBoost and BP Neural Networks

Abstract

Wind power generation is characterized by strong intermittency and volatility, making accurate short-term forecasting essential for reliable power system operation and stable dispatch of power grids. Conventional single-model approaches often struggle to simultaneously maintain overall forecasting accuracy and reliably capture power peaks. To address this issue, this paper proposes a hybrid forecasting model with a sigmoid-weighted soft-switching strategy between Extreme Gradient Boosting (XGBoost) and Back Propagation (BP) neural networks. A peak-weighted Mean Square Error (MSE) loss function is introduced to improve the capability of XGBoost in learning extreme power samples, while the standard deviation of historical wind power is employed as a fluctuation indicator to determine the fluctuation regime, and a Sigmoid function maps this indicator to a continuous BP weight. The final prediction is therefore obtained by smooth weighted fusion rather than hard model selection: XGB Wtd dominates under stable conditions, while the contribution of BP Std increases as power fluctuation rises. Experimental results show that compared with the peak-weighted XGBoost (XGB Wtd), the proposed model increases the overall MAE by only 1.02% while reducing the peak MAE by 26.80%. Overall, the proposed model maintains a lightweight structure and provides an empirical interval for descriptive error-dispersion analysis, demonstrating practical potential for short-term wind power forecasting.