Skillful Precipitation Nowcasting Using Physical-Driven Diffusion Networks

Abstract

Accurate and timely precipitation nowcasting is essential for numerous applications including emergency services, infrastructure management, and agriculture. Recently, deep learning (DL) techniques have shown promise in enhancing nowcasting capabilities. This study introduces a novel Physical-Driven Diffusion Network (PDDN) model that leverages both radar and numerical weather prediction (NWP) data to improve the accuracy and physical consistency of precipitation nowcasts. Our approach integrates the strengths of data-driven DL techniques with physics-based NWP models. The PDDN model utilizes latent diffusion models and autoencoders within a two-stage architecture to predict future radar images, incorporating the Weather Research and Forecasting (WRF) model data to enhance understanding of atmospheric dynamics. Our results demonstrate significant improvements over traditional models, particularly in short-term forecasting up to 6 hr. This research highlights the potential of combining advanced machine learning techniques with conventional meteorological data, offering new directions for enhancing the accuracy and reliability of weather forecasting.