Abstract: Accurately forecasting wind speed can enhance the accuracy of wind power forecasting. However, existing wind speed forecasting methods mostly ignore the spatial correlation between neighboring wind farms. There is significant potential for improving wind speed forecasting accuracy when abundant data from multiple wind farms and their strong interdependencies are available. To fully exploit the spatial correlation information, we propose a novel wind speed forecasting model based on GCN and frequency-enhanced decomposed transformer (FEDformer), i.e., GFformer. The GCN is utilized for extracting spatial features of wind speed, while the FEDformer is employed for learning temporal features. Moreover, this paper constructs a complex adjacency matrix that characterizes the correlation relationship from two dimensions: intensity and temporal lag. This enables GFformer to capture the spatiotemporal correlations of wind speed between neighboring wind farms more comprehensively, thereby further improving the accuracy of wind speed forecasting. In a case study with a dataset consisting of 25 wind farms, GFformer outperforms other benchmark models.