Wind power prediction method based on graph-structured spatio-temporal feature decomposition and spatio-temporal attention

Against the background of global low-carbon energy transition, wind energy has developed rapidly, but its randomness, intermittency and volatility bring challenges to power system operation, making accurate wind power prediction crucial. For sparse wind farms, traditional graph neural networks have insufficient capability to extract temporal dependencies, and inconsistent wind speed variation trends of adjacent turbines degrade spatiotemporal feature extraction; meanwhile, the nonlinearity of wind speed further reduces prediction accuracy. This paper proposes an ultra-short-term wind speed prediction method based on graph-structured spatiotemporal feature decomposition and spatiotemporal attention graph neural network (TSAGNN). The method uses variational mode decomposition (VMD) to reduce data nonlinearity, and integrates temporal and spatial attention modules to enhance spatiotemporal feature capture. Experiments show that TSAGNN outperforms mainstream comparison models on multiple metrics, with high reliability under wind speed fluctuations, verifying its superiority for sparse wind farms.