DeepWind: a heterogeneous spatio-temporal model for wind forecasting

Deep learning (DL) has shown great potential in enhancing the performance of traditional numerical weather prediction (NWP) methods in weather forecasting. Certain applications such as wind power generation desire more accurate wind predictions which is challenging due to limited observations and complex dynamics. To this end, this paper introduces a DL-based heterogeneous network named DeepWind for NWP correction, which can simultaneously correct the NWP of diverse wind variables across multiple weather stations. In particular, it first exerts the meteorological domain knowledge to achieve an effective transformation of target variables and then develops a heterogeneous neural network to learn spatio-temporal representations. A novel difference loss function is further designed for stable temporal learning. Moreover, this study might be the first to expose an underlying evaluation problem in deep forecasting, which we call evaluation inconsistency, thereby necessitating the assessment of model performance across diverse evaluation metrics. Experimental results demonstrate the superiority of the proposed approach over strong DL baselines, which makes it positioned for deployment in the real-world production environment. Source code is released at https://github.com/Rittersss/DeepWind.