Wind power forecasting is a crucial topic in the field of wind energy generation. With the increasing penetration of renewable energy in power systems, wind energy, as a rapidly developing renewable resource, necessitates accurate short-term forecasting for the energy industry. In this paper, a Gaussian Mixture Model-gated Recurrent Unit （GMM-GRU） based on Gaussian graph convolution is proposed. Firstly, the XGBoost algorithm optimized by Particle Swarm Optimization （PSO） is utilized to construct a feature selection network for identifying important features. Secondly, using the graph auto-encoder and cosine correlation fusion method, a network graph to effectively capture the potential long-distance correlations among sites and accurately describe the spatial correlation of multi-site features are built. Finally, the Gaussian Mixture Model （GMM） is employed to deeply extract the intrinsic relationships within the network graph, and the Gated Recurrent Unit （GRU） integrates its spatio-temporal correlations to address the power forecasting problem. The experiments on real wind farm data are carried out. The comparisons with other models demonstrate that the proposed model significantly enhances wind power forecasting performance.

wind power forecasting  /  graph autoencoder  /  Gaussian mixture models  /  recurrent neural networks  /  PSO-XGBoost