A Study on Short-Term Nowcasting Precipitation Forecasting Methods Based on Radar Echo Maps

Short-To-medium-term precipitation nowcasting aims to predict changes in mesoscale and microscale weather systems, particularly severe convective systems, within 0 to 2 hours. Compared to large-scale weather systems, mesoscale and microscale systems are characterized by their abruptness, complex mechanisms, and short life cycles, making short-term precipitation forecasting a highly challenging field. This study focuses on the Yinchuan region as a case study, processing radar echo data through dimensionality reduction, noise reduction, frame interpolation, and data augmentation to construct datasets suitable for training and testing. The IAM4VPx5 model is employed for forecasting and is compared against traditional spatiotemporal prediction models, including SimVP, ConvLSTM, and ConvGRU. Experimental results demonstrate that the IAM4VPx5 model outperforms SimVP, ConvLSTM, and ConvGRU across multiple metrics, such as Critical Success Index (CSI), Heidke Skill Score (HSS), and Probability of Detection (POD). Additionally, IAM4VPx5 exhibits superior performance in Mean Squared Error (MSE) and Root Mean Squared Error (RMSE). Overall, the IAM4VPx5 model demonstrates exceptional performance compared to SimVP, ConvLSTM, and ConvGRU in various aspects. These results validate the effectiveness and advantages of the IAM4VPx5 model in processing radar echo data, offering an efficient solution for weather radar forecasting.