A Multi-Modal Transformer for Short-Term Wind Power Forecasting with Spatio-Temporal Meteorological Data

Wind power forecasting (WPF) is a significant and effective technique to cope with the negative impact of its intermittency on grid stability. Although several deep learning (DL) methods have been applied to WPF as time series analysis approaches, there are still fewer studies considering multi-modal features for short-term WPF. In this paper, a novel transformer-based DL architecture is proposed to extract temporal and spatial dependencies from regional meteorology and time series using different encoders customized for each modality. Finally, a hierarchical multi-head cross-attention module is designed based on the logical dependencies between different modalities. The experimental results show that, due to the effectiveness of spatiotemporal context and cross-attention, the proposed model reduces root mean square error by 0.495-0.950 MW over other baselines, and the forecasting accuracy is improved by 0.56%-1.15%.