关键词: 风电功率预测, DTHSA, cWGAN-GP, B-sLSTM, 爬坡事件

Abstract: Wind power plays a critical role in renewable energy, and accurate wind power prediction is essential for enhancing grid stability and economic efficiency. Ramp events, characterized by sharp increases or decreases in wind power over short periods, represent a minority in wind power datasets. Consequently, prediction models often fail to adequately learn the features of these events, leading to reduced prediction accuracy during ramp periods. To address this issue, this study proposes a wind power prediction method based on virtual sample generation for ramp events. The approach first employs DTHSA (Dual-Threshold Hysteretic Slope Algorithm) to accurately identify ramp event segments. Subsequently, cWGAN-GP (Conditional Wasserstein Generative Adversarial Network with Gradient Penalty) is utilized to generate virtual ramp samples, thereby augmenting the number of ramp instances in the original dataset. On this basis, the proposed B-sLSTM algorithm is applied to perform wind power prediction on the enhanced dataset. Experimental validation using real-world data from a wind farm demonstrates that, for 4-hour-ahead wind power prediction, the proposed B-sLSTM model reduces RMSE and MAE by 0.265 MW and 0.22 MW, respectively, and increases R2 by 3.33% compared to the baseline sLSTM model. These results confirm that the proposed method effectively improves short-term wind power prediction accuracy.

Key words: wind power forecasting, dual-threshold hysteretic slope algorithm, data augmentation, B-sLSTM, ramp event