DGFormer: a physics-guided station level weather forecasting model with dynamic spatial-temporal graph neural network

Abstract

In recent years, there has been an increased interest in understanding and predicting the weather using weather station data with Spatial-Temporal Graph Neural Networks (STGNN). However, it has large prediction errors as a result of the inherent non-linearities and the influence of dynamic spatio-temporal auto-correlation. Using a continuously-varying graph topology chronologically, while embedding domain knowledge to enforce validity, can effectively resolve the issue, but the implementation of such concept constitutes an interdisciplinary challenge for researchers. A Dynamic Graph Former (DGFormer) model is proposed to address this challenge. It combines a topology learner through a deep generative layer with domain knowledge enhancement inserted into the STGNN structure, where the derived physics-guided method allows for an efficient integration with the earth system. For capture of the optimal topology, we merge a node-embedding-based similarity metric learning and the superposition principle as physical assistants into the dynamic graph module. We evaluate our model with a real-world weather dataset on short-term (12 hours) and medium-range (360 hours) prediction tasks. DGFormer achieves outstanding performance with obvious improvements by up to 34.84% at short-term prediction and by up to 23.25% at medium-range prediction compared with the state-of-the-art methods. We also conducted detailed analyses for cities in three regions and visualized the dynamic graph, revealing the characteristics, advantages, and graph visualization of our model.