TriPhysGAN-Attn: A Physics-Informed Generative Model for Radar Echo Forecasting via Triple Mechanism Decomposition and Attention Fusion

Nowcasting of precipitation plays a critical role in meteorological operations and disaster prevention. However, existing deep learning-based radar echo prediction models often suffer from unclear physical mechanisms and insufficient capability to represent the multi-scale dynamic processes of precipitation systems, making them less effective in handling severe convective weather. In this study, we propose TriPhysGAN-Attn, a novel radar echo prediction model guided by meteorological physical mechanisms. We divide the evolution of precipitation into three fundamental processes: advection, initiation and dissipation, and deformation. Accordingly, we design an advection branch to model multi-scale motion fields, an initiation and dissipation branch driven by state differences, and a deformation branch guided by motion field gradients. To enhance physical consistency, we incorporate a physically informed loss function along with meteorological priors and supervision signals. A multi-head cross-attention mechanism is employed to dynamically fuse information across the three processes, facilitating collaborative modeling. Additionally, a generative adversarial framework and a coordinate attention module are integrated to better capture local discontinuities and spatial structures. Experimental results show that TriPhysGAN-Attn outperforms existing methods on real radar datasets, achieving superior performance in structural preservation, prediction accuracy, and robustness under extreme weather conditions.