Wind Power Prediction Correction and Climbing Characteristics Based on the Combination of XGBoost and LSTM
Year: 2025
Type: article
Abstract: In wind power forecasting, Existing methods face dual challenges, specifically manifested as insufficient modeling capabilities for complex time series features, low prediction accuracy of climbing events, and limited model generalization capabilities. Facing the problem of insufficient accuracy of traditional prediction methods when facing nonlinear and abrupt power changes, a hybrid prediction correction model combining XGBoost and LSTM is proposed, considering the capabilities of XGBoost in static feature extraction and LSTM in dynamic time series modeling. Construct a multi-modal feature fusion architecture, quantitatively evaluate the importance of input features through XGBoost, assist LSTM in capturing key inputs, and improve the model’s ability to perceive wind power changing trends. A power sudden change point detection algorithm based on phase space reconstruction is designed to identify wind power climbing and downclimbing events accurately. By establishing a coupling model between power change rate and grid frequency response, the system’s response prediction ability to power sudden change caused by grid disturbance is strengthened. Combining feature importance-driven error pattern recognition with LSTM memory cell state compensation strategy, the model output is dynamically adjusted, and the cumulative error is significantly reduced. The traditional LSTM model does not capture the long-term dependence characteristics of wind power series enough, resulting in the root mean square error (RMSE) of ultra-short-term prediction of 13.5 kW⋅h and the false report rate of climbing events of 42.7%, Although the single XGBoost model can deal with nonlinear features, there is a lag in time series dynamic modeling, the prediction accuracy rate is only 76.2%, and the recognition rate of extreme climbing events with slope exceeding 24.8 kW⋅h/10 min is less than 59.3%. In contrast, the measured data in a certain area shows that the prediction error rate of the model without the correction strategy soared to 22.3% in strong turbulent weather, while the model initially combining XGBoost and LSTM reduced the RMSE to 7.2 kW.h. The climbing event capture rate increased to 88%, and the extreme scene prediction accuracy reached 91.6%, showing the potential for collaborative processing of complex features.
Authors: Tingting Zheng, Hongquan Yin, Da Wang, Xu Feng, Jin Zhu, Jiangcheng Li
Institution: State Grid Corporation of China (China)
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4416995575