A wind speed forecasting framework for multiple turbines based on adaptive gate mechanism enhanced multi-graph attention networks

Accurately forecasting wind speed is crucial for efficiently utilizing wind energy and scheduling power grids. Recently, Graph Neural Network (GNN) models have been widely utilized to forecast wind speed, which explicitly utilizes the correlation between turbine sites in a wind farm. However, it is challenging to appropriately construct graphs to characterize multiple latent but unknown interdependencies among turbines. This paper proposes a novel multi-site wind speed forecasting framework AG-MGAT, based on an adaptive gate mechanism enhanced multi-graph attention networks. In detail, the contributions of our work are threefold. Firstly, multiple graphs are explicitly constructed, which respectively measure the wind behavioral similarity and the directional causality between turbine sites. Secondly, to calibrate the potential misalignment of spatial-temporal GNNs using these task-agnostic graphs, an adaptive gate mechanism enhanced Graph Attention Network (GAT), AG-GAT, is innovatively designed, which uses a learnable adjacency matrix as gate to adaptively weight the sites' embeddings from the current GAT layer and these directly from the previous AG-GAT layer. At each timestep, the proposed AG-GATs working on the constructed graphs are used to extract the turbine sites' representations that embed the multiple correlations among sites, which are then sent to recurrent neural network for further processing the temporal interdependency. Finally, thorough experiments on real wind speed dataset are conducted and the experimental results show the superiority of our schemes over other state-of-the-art GNN-based forecasting schemes.