Wind Power Forecasting Based on Adaptive Spatio-Temporal Graph Neural Network

The rapid increase in wind power consumption presents new challenges for power system operation. Accurate forecasting of wind power is recognized as a major contribution for promoting its consumption and improving the stability of grid operation. This study focuses on forecasting wind power for an onshore wind farm in the United States by incorporating spatiotemporal relationships among 200 wind turbines. We proposed a novel encoder-decoder structure model. In the encoder module, a graph neural network structure is used to capture the intricate geographical distance features and sequence similarity features of the wind turbines as spatial features. The extracted spatial features are then integrated with temporal features into the model by serving as inputs to the gated recurrent unit (GRU) module. In the decoder module, the last state of the top layer of the encoder serves as the initial hidden parameters of the GRU blocks. Finally, the predicted wind powers are given by the decoder. Our model exhibits superior accuracy and enhanced generalization capabilities.