Integrating Spatio-Temporal and Generative Adversarial Networks for Enhanced Nowcasting Performance

Abstract

Nowcasting has emerged as a critical foundation for services including heavy rain alerts and public transportation management. Although widely used for short-term forecasting, models such as TrajGRU and PredRNN exhibit limitations in predicting low-intensity rainfall and low temporal resolution, resulting in suboptimal performance during infrequent heavy rainfall events. To tackle these challenges, we introduce a spatio-temporal sequence and generative adversarial network model for short-term precipitation forecasting based on radar data. By enhancing the ConvLSTM model with a pre-trained TransGAN generator, we improve feature resolution. We first assessed the model’s performance on the Moving MNIST dataset and subsequently validated it on the HKO-7 dataset. Employing metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), Structural Similarity Index Measure (SSIM), Critical Success Index (CSI), Probability of Detection (POD), and False Alarm Ratio (FAR), we compare our model’s performance to existing models. Experimental results reveal that our proposed ConvLSTM-TransGAN model effectively captures weather system evolution and surpasses the performance of other traditional models.