STGATN: a wind speed forecasting method based on geospatial dependency

Accurate wind speed forecasting is crucial for power systems, but wind speed as a spatially continuous field presents high randomness, fluctuation, and spatial heterogeneity under complex geographical environments, leading to challenges for predictive modeling. This paper proposes a wind speed forecasting method STGATN based on spatial dependency, which effectively reduces uncertainties caused by randomness, fluctuation, and spatial heterogeneity when capturing spatial dependencies between areas, thereby improving prediction accuracy. First, the geospatial continuous field is appropriately converted into discrete computational units that preserve topological relations. Second, a self-attention mechanism captures complex spatial dependency patterns between areas caused by factors such as terrain and surface pressure, breaking through the limitations of traditional methods that rely solely on spatial distance. Finally, a spatiotemporal graph attention neural network is designed to effectively extract and fuse spatiotemporal features, balancing time series trends and cycles with spatial dependency. Experiments demonstrate that our method significantly outperforms existing single and hybrid models on four datasets from Xinjiang, China. This superior performance is achieved by precisely modeling the real-world data spatial dependencies in wind speed evolution between different areas. Our approach provides new insights for wind speed forecasting in complex geographical environments.