Short-Term Wind Power Forecasting under Extreme Weather Based on Improved TimeGAN and GAT-Crossformer
Abstract
To overcome the combined issues of intricate weather feature dynamics and insufficient data under extreme conditions, a short-term wind power prediction framework is presented, leveraging a modified time-series generative adversarial network (TimeGAN) alongside a GAT-Crossformer network. First, raw data corresponding to extreme weather events are extracted. To mitigate the problem of data scarcity, an improved TimeGAN, capable of capturing temporal dynamics, is employed to augment the extreme weather data. Second, in order to address the significant fluctuations in wind power output during extreme weather, the dataset is decomposed and reassembled using a combined approach of Variational Mode Decomposition and Sample Entropy. Finally, a hybrid model based on GAT-Crossformer is constructed. By incorporating multidimensional attention mechanisms, the model effectively captures the latent correlations and temporal dependencies in extreme weather scenarios, thereby enhancing forecasting accuracy. Experiments confirm that the proposed method significantly improves short-term forecasting performance under extreme conditions and exhibits strong generalization capability.