Multi-step prediction of spatio-temporal wind speed based on the multimodal coupled ST-DFNet model

Wind energy, as an efficient renewable energy source, has become a key component of the global energy mix transition. To improve short-term wind speed prediction accuracy and optimize wind farm resource allocation, a novel spatio-temporal wind speed prediction model is proposed. By combining altitude and meteorological factors, the paper establishes an effective coupling relationship utilizing the coupled mapping lattice model (CML) through multimodal equations, which overcomes the shortcomings of the traditional spatio-temporal model in dealing with the fusion of multimodal data. The improved Gated Spatial Attention Transformer model (GSAT) effectively extracts spatial data features. Based on the idea of "splitting and integrating" to target the data in the temporal and spatial domains, the Spatiotemporal-Disentangle Fusion Network (ST-DFNet) shows the highest prediction accuracy in different time scales and has been experimentally validated to outperform all comparative models. In addition, owing to breaking through the limitation of traditional interval prediction, this study proposes the adaptive interval estimation method ASKDE and constructs a dynamically adjustable interval prediction framework by revealing the correlation between heights and utilizing the super-Lorenz system, which successfully realizes the accurate coverage of wind speeds at high altitudes from low-altitude regions. The findings indicate that the model has strong forecasting capability and wide application prospects, especially in the fields of wind farm siting, turbine installation height optimization, and grid scheduling.