Applicability analysis of transformer to wind speed forecasting by a novel deep learning framework with multiple atmospheric variables

Accurate wind speed forecasting plays a crucial role in the efficient and economical management of power supply systems. In this study, a novel framework combining variational mode decomposition (VMD), graph neural network (GNN) and temporal forecasting component is proposed for wind speed forecasting using multiple atmospheric variables. VMD is employed to decompose atmospheric variables into distinct subsequences at various frequencies, and GNN is utilized to effectively pass, aggregate, and update variable features, thus enabling the extraction of pairwise dependencies among the different variables. Subsequently, the transformer model is used as the temporal forecasting component in the proposed framework. Compared with several state-of-the-art transformer-based models and baseline models in AI field, the superior performance of our hybrid framework is observed. Lastly, several other deep learning models, including multi-layer perceptrons (MLP), long short-term memory (LSTM), and gated recurrent unit (GRU), are selected as forecasting component for assessing the applicability of the transformer. Interestingly, the transformer exhibits the lowest performance, while the GRU demonstrates the most promising results, with the mean absolute error (MAE) of 0.1356 m/s and 0.1085 m/s for one-step ahead forecasting, respectively. Overall, this research provides valuable insights into a novel framework and the applicability of the transformer for wind speed forecasting.