Ultra-short-term wind power forecast based on multi-feature information fusion of GAT-Crossformer

To address the issue that relationships among different wind clusters are rarely considered adequately in ultra-short-term wind power union forecasting, a multi-site wind power joint forecasting method based on graph attention network (GAT) and Crossformer is proposed. The spatial correlations among different wind farm sites as well as the cross-dimensional and cross-temporal dependencies between different feature variables at the same site are taken fully into consideration. Initially, the maximum information coefficient (MIC) is used for feature importance analysis and to select the factors that most influence wind power. Next, the spatial relationship of wind power sequences at neighboring sites is explored using a multi-site topological graph, and correlations are obtained by a designed GAT. Subsequently, a Crossformer model is constructed with a novel double attention mechanism to extract temporal dependencies within a single variable and spatial dependencies between multiple variables across different features at the same site. Validation experiments were performed on two real-world wind power clusters. The testing results indicate that compared to classical prediction models such as Crossformer, GAT-TCN, GAT-LSTM, GCN-LSTM, GAT, GCN, TCN, and LSTM, the proposed model demonstrates better generalization and excellent prediction performance in ultra-short-term joint wind power forecasting of multi-sites.