PredAt-GAN: a new spatiotemporal forecast model for precipitation nowcasting with weather radar echo images

Short-term heavy rainfall prediction is a critical and practical research domain with direct implications for property and life safety. However, existing RNN-based models utilizing Weather Radar Echo Images for prediction tasks lack error calibration mechanisms at each time step, leading to a significant issue of cumulative error that results in a rapid decline in prediction accuracy. Moreover, these models have not fully explored and assimilated knowledge about heavy rainfall events present within the dataset, indicating further progress is required before their deployment to actual prediction tasks. In this paper, PredAt-GAN model is proposed, which integrates the ConvRNN with the GAN to adjust the ConvRNN predicted output at each time step, effectively alleviated the problem of rapidly accumulating errors. Additionally, by incorporating attention units within the model network and introducing weight mechanisms in the training loss function, it enhances the model's focus on regions characterized by intense rainfall. To validate model's performance, comparative and ablation experiments using real reflectivity fields acquired from the Shenzhen weather radar station were conducted. The results demonstrate significant improvements in SSIM, MSE, HSS, CIS, and image visualization.