Recently, there has been a dramatic rise in the demand for accurate temperature forecasts. However, challenges arise from modeling and fusing complex spatial and temporal features in temperature data. In this study, we propose a physics-informed directed-graph-based temperature prediction model to mitigate the challenges of purely data-driven prediction algorithms. Firstly, a directed graph design module was designed and then used to construct an asymmetric adjacency matrix based on the locations of temperature-monitoring stations. This module can capture the asymmetric relations between temperature data at different stations. Then, the directed adjacency matrix was incorporated into the graph attention module and the graph-gating module to extract the spatial and temporal features of the temperature data, and a fusion module was designed to integrate the spatial–temporal features and the directed graph adjacency matrix to provide better temperature prediction performance. Numerical simulations based on a real-world dataset collected in southern China demonstrate that our proposed physics-informed temperature prediction model can deliver superior prediction performance with a mean absolute error of less than 0.75 °C.
Cai et al. (Tue,) studied this question.