Short-term Ground Wind Field Forecasting Based on DGCN-2TS

Wind forecasting plays a crucial role in renewable energy dispatch and extreme weather warnings. However, due to the highly dynamic spatiotemporal characteristics of wind and its rapid variability, accurate wind forecasting has always been a challenging task. In recent years, spatiotemporal graph neural networks have shown potential in integrating temporal information and spatial correlations among stations, providing a new solution for wind forecasting. Nevertheless, existing methods still face two main challenges: a) Predefining the adjacency relationships between stations using a static graph structure fails to effectively capture the dynamic meteorological features that change over time, resulting in the forecasting model's lack of adaptability to complex weather variations. b) The rapid and frequent changes in wind speed and direction make it difficult for traditional models to fit these nonlinear and high-frequency variations, affecting the prediction accuracy. To address these issues, we propose a new spatiotemporal forecasting framework that dynamically combines the spatial correlation structure among stations with the temporal correlations, adaptively generating the adjacency matrix based on meteorological features at each time step to capture the spatiotemporal dynamics of weather conditions. Furthermore, we design a multi-level forecasting model that first predicts meteorological variables with smaller changes and stronger trends, providing foresight information for subsequent wind forecasting and enhancing the model's adaptability to complex meteorological conditions. Experiments conducted on a real dataset from Southwest China demonstrate that the proposed method outperforms existing models in the wind forecasting task in this region.