Precipitation Nowcasting with Graph Neural Network and Gated Recurrent Units

Soobin Park, Yoonhyuk Choi, Chong-Kwon Kim

Abstract: Precipitation is one of the most unpredictable events in meteorology. While long-term forecasting remains in the domain of conventional NWP (Numerical Weather Prediction), recent advances in deep learning-based forecasting techniques provide competitive performances in nowcasting, forecasting over zero to two hours into the future. This paper introduces a novel nowcasting method that combines Graph Convolutional Network and Gated Recurrent Unit techniques. The proposed method utilizes optimized edge weights that capture the nonlinear relations between the end nodes via an MLP. The proposed model first applies a GCN (Graph Convolutional Network) to uncover latent features of weather stations. Then node feature vector is input to a Gated Recurrent Unit that captures temporal relations and predicts precipitation in the future. We performed extensive experiments to evaluate the performance of the proposed method using a real-world dataset obtained from KMA (Korea Meteorological Administration). The experiments show that the proposed method predicts precipitation nowcasting comparable to the NWP system.