An EGWO-XGB-RF Feature Engineering Approach for Short-Term Photovoltaic Power Forecasting

Abstract

Accurate photovoltaic (PV) power prediction is essential for efficient PV plant operation and energy management; however, it remains challenging due to the high volatility of meteorological factors. To address this limitation of existing feature engineering methods in extracting effective features from complex nonlinear data, this paper proposes an EGWO-XGB-RF feature engineering method. The approach first employs Pearson correlation analysis to preliminarily screen meteorological features. Subsequently, XGB-RF is applied to generate new features through decision tree structures. Finally, EGWO is utilized to select a high-quality feature subset, which is subsequently used for PV power prediction with LightGBM. Experimental results show that the proposed EGWO-XGB-RF based model achieves an RMSE of 0.1875, an MAE of 0.0962, and R² of 0.9854 on the held-out test set, outperforming the eight compared baseline models under the same experimental conditions. The proposed method provides a reliable technical pathway for improving the accuracy of renewable energy forecasting.