Short-term wind speed forecast based on dynamic spatio-temporal directed graph attention network

The accurate prediction of wind speed is crucial for the advancement of the wind power industry. This study introduces a new wind speed forecasting model called the Dynamic Spatio-temporal Directed Graph Attention Network (DSTDGAT), which aims to enhance prediction accuracy by capturing the time-varying and asymmetric spatio-temporal correlations among turbines. To address changing weather conditions, the model decomposes association patterns into bidirectional long-term and unidirectional short-term patterns. A dynamic directed graph learning module is designed to optimize adjacency matrices, followed by graph attention layers and gated temporal convolution layers to extract hidden spatio-temporal features for wind speed predictions. A two-stage training strategy is proposed for long- and short-term patterns, with results integrated through an adaptive fusion module to enhance model performance. Experimental results using actual wind speed data demonstrate that the proposed approach improves wind speed prediction accuracy and effectively leverages directed correlations within the complex graph structures of real-world wind farms.