Dynamic Spatial-temporal Graph Prediction for Short-term Precipitation

Abstract

Precipitation is a crucial component of the natural water cycle, and inadequate timeliness and precision in precipitation prediction can result in agricultural losses, traffic disruptions, flood catastrophes, and even threats to human life. Consequently, precipitation prediction is a key problem in the domain of meteorology. However, the current methodologies pay close attention to the explicit spatial connections of precipitation regions while neglecting the implicit spatial connections over time. There are often challenging for traditional convolutional neural networks and graph neural networks to capture, leading to inaccurate spatial regions and poor timeliness of model predictions. To resolve this problem, we propose a Dynamic spatial-temporal graph prediction model for short-term precipitation (Dst-pred), which dynamically explores implicit connections among meteorological stations in the target region through graph neural networks and constructs dynamic spatial-temporal graphs to predict precipitation in the region. We have verified our Dst-pred model on our proprietary precipitation dataset from Guangxi Province, China, and the ERA5-Land dataset, and it can extract the implicit spatial connections between individual stations from the precipitation data of meteorological stations. The precipitation process capture of our model enhances the timeliness and accuracy of nowcasting precipitation prediction with the best performance.