This dissertation advances the field of precipitation nowcasting through three novel deep learning frameworks that progressively address the fundamental challenges of forecast accuracy, physical consistency, and image clarity. The research first introduces the Synthetic-data Task-segmented Generative Model (STGM), which combines synthetic data generation with task segmentation to extend reliable forecasting capabilities up to six hours while significantly reducing forecast blurriness. Building upon this foundation, the Physical-Driven Diffusion Network (PDDN) is developed, integrating latent diffusion models with numerical weather prediction data to enhance the physical consistency and reliability of precipitation forecasts. The research culminates in the development of RainCast-GPT, an...[ Read more ]

This dissertation advances the field of precipitation nowcasting through three novel deep learning frameworks that progressively address the fundamental challenges of forecast accuracy, physical consistency, and image clarity. The research first introduces the Synthetic-data Task-segmented Generative Model (STGM), which combines synthetic data generation with task segmentation to extend reliable forecasting capabilities up to six hours while significantly reducing forecast blurriness. Building upon this foundation, the Physical-Driven Diffusion Network (PDDN) is developed, integrating latent diffusion models with numerical weather prediction data to enhance the physical consistency and reliability of precipitation forecasts. The research culminates in the development of RainCast-GPT, an innovative framework that leverages large language models within an Ensemble-Meritocratic Task-segmented Generative Framework, further improving forecast clarity and accuracy. Each model demonstrates superior performance in six-hour precipitation forecasts compared to traditional methods, while addressing different aspects of the nowcasting challenge. Together, these three approaches represent a significant advancement in precipitation nowcasting, offering a comprehensive solution that combines deep learning techniques with physical understanding of meteorological systems. This work establishes new benchmarks in rainfall forecasting methodology and provides practical tools for emergency services, infrastructure management, and public safety applications.