A novel dynamic spatio-temporal graph convolutional network with multi-scale transformer for high-resolution wind vector forecasting

• Dynamic graph learning captures latent, time-varying spatial dependencies adaptively. • The designed gated deep GCN prevents over-smoothing during message passing. • Using multi-scale Transformer encoders to capture long-range temporal dependencies. • The recursive framework with dynamic updates improves long-term forecasting. • Multivariate joint modeling learns inter-variable coupling for better generalization. Accurate wind speed forecasting is crucial for the grid integration of wind farms and the dispatch of power systems. However, due to the inherent randomness and variability of wind speed, achieving high-precision wind speed forecasting (WSF) remains a significant challenge. To address this issue, a novel dynamic spatio-temporal graph convolutional network with multi-scale transformer (DSTGformer) is proposed in this study for wind speed vector forecasting. This model integrates the powerful local feature extraction of graph convolutional networks (GCNs) with the superior global information fusion of Transformers. Firstly, a dynamic spatio-temporal graph convolution module is designed, in which a self-attention mechanism is incorporated to generate a dynamic adjacency matrix. A gated residual deep graph convolutional network is then employed to capture the dynamic spatiotemporal dependencies of wind speed data by performing temporal graph convolution operations on the dynamic matrix. Furthermore, a multi-scale Transformer encoder is integrated to model long-range dependencies within the input sequences and extract effective features from multi-scale temporal fluctuations. A recursive prediction framework is adopted, enabling multi-step-ahead forecasting while maintaining the collaborative prediction performance of multiple meteorological variables. Compared to eight existing baseline WSF models, the proposed DSTGformer model achieves the highest prediction accuracy in wind speed forecasting.