Knowledge-enhanced XGBoost for Wind Power Forecasting under Extreme Meteorological Events

Abstract

To address deteriorated generalization of conventional data-driven wind power forecasting under extreme meteorological conditions, this study develops a knowledge-enhanced XGBoost prediction framework. Physical thresholds are used to divide meteorological and turbine control variables into state labels for wind speed, temperature, yaw angle, and pitch angle. The labels are compressed into two-dimensional numerical features and combined with sliding window temporal features for model training. Experiments based on the SDWPF dataset verify that embedded physical prior information mitigates out-of-distribution prediction drift. Compared with vanilla XGBoost, the proposed method reduces 1-hour MAE by 23.77% with statistical significance and achieves evident accuracy improvement across diverse extreme weather subsets.