Multi-feature extraction spatio-temporal interaction graph network for wind speed forecasting in windfarm

Accurate wind speed prediction is crucial for grid stability and energy management. However, the inherent instability and stochastic nature of wind speed present significant challenges for achieving precise forecasts. As deep learning advances, graph neural networks (GNNs) have shown great potential for time series modeling. However, current models often overlook the spatiotemporal synchronization and uniformity of wind speed data, instead modeling temporal and spatial components independently. This study proposes the Multi-Feature Extraction Spatio-Temporal Interaction Graph Network (MFSGN), which utilizes a de-stationary attention mechanism and frequency enhancement technique to capture the non-stationary and frequency-domain features of wind speed data. The model employs a spatial graph learning module to model both long-term static correlations and short-term dynamic relationships between wind turbines, ultimately fusing these to capture the spatial dependencies of wind speed. By incorporating both spatio-temporal separation and unification, MFSGN learns the spatio-temporal characteristics of wind speed, improving model efficiency and predictive performance. The parallel training of gated temporal convolution, dynamic graph convolution, and Fourier map neural networks further enhances the approach. Extensive experiments on real wind speed data from Shanxi Province, China, demonstrate that MFSGN outperforms existing benchmark methods in prediction accuracy and generalization, contributing to improved wind energy utilization and the advancement of the wind power industry.