A robust spatio-temporal prediction approach for wind power generation based on spectral temporal graph neural network

At present, the penetration of wind power generation is increasing remarkably worldwide, and the accurate wind power forecasting (WPF) is essential to ensure the reliability and economy of the power system. Most of the current work of WPF only capture temporal correlation in the time domain but ignore the spatial correlation. In this study, a spectral time graph neural network based on the maximum correlation criterion (MCC-Stem-GNN) is proposed to improve the accuracy of WPF for multiple sites and horizons. The self-attentive mechanism in the MCC-Stem-GNN automatically learns the correlations between the multivariate sequences. Besides, this model combines the Graph Fourier Transform (GFT) to model spatial correlation and the Discrete Fourier Transform (DFT) to model temporal correlation. The effectiveness of the proposed robust deep learning framework is verified on the simulated wind energy dataset over 16 locations in Ohio, US through considering different sample contamination types and levels, comprehensive case study is carried out to show the superiority of the MCC-Stem-GNN over the benchmarks.