Spatio-Temporal Feature Fusion via U-shaped Architecture for Accurate Wind Speed Prediction

Wind speed forecasting is a crucial support for ensuring the safe and stable operation of wind farms. However, due to the significant spatiotemporal variability of wind speed and the influence of complex meteorological factors, achieving highaccuracy forecasting remains a core challenge in the field of renewable energy generation. This study proposes a spatiotemporal feature fusion network based on a U-shaped architecture (U-STNet), which fully exploits the spatiotemporal dependencies within wind speed data by integrating spatial information from multiple wind turbines and long-range temporal dependencies across different periods. The model employs an embedding mechanism to project wind speed time series into a high-dimensional feature space, and utilizes an encoder-decoder Ushaped structure to perform encoding, reconstruction and multi-scale extract of high-dimensional features, effectively capturing the complex periodicity and seasonal variation patterns inherent in wind speed sequences. Experimental results on the SDWPF dataset demonstrate that the proposed method consistently outperforms existing mainstream forecasting models across multiple real-world wind speed datasets, significantly improving prediction accuracy. This approach provides an efficient and reliable spatiotemporal feature fusion modeling scheme for wind speed forecasting, with promising potential for broader application.