DCPGRN: A Dynamic Climate Pattern Graph Recurrent Network for Weather Forecasting

Currently, to understand and predict the weather still faces a significant issue due to the inherent non-linearities and the involvement of dynamical spatio-temporal auto-correlation over time, causing an interdisciplinary challenge for researchers. We propose a Dynamic Climate Pattern Graph Recurrent Network (DCPGRN) for meteorological forecasting to address such challenge. It combines a topology learner through a deep generative layer with domain knowledge enhancement into the spatio-temporal network. We merge a node-embedding-based similarity metric learning and the superposition principle as physical assistant into the dynamic graph module to capture the optimal topology. Besides, we evaluate the DCPGRN with some real-world weather datasets and achieve outstanding performance with obvious improvements compared with the representative algorithms in this field. Moreover, we conduct a city insights in 7 typical city stations, revealing the interpretability, advantages, and insidious discussions of our model.