Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification

Sudden onset of precipitation frequently endangers human lives and causes damage and disruption to infrastructure through flooding and landslides, and is often accompanied by other hazardous weather phenomena such as hail, lightning and windstorms. Precipitation is also a fundamental driver of agriculture and hydroelectric power generation. Consequently, short-term precipitation forecasts are important tools that can benefit infrastructure managers, emergency services and the general public if provided in a timely manner. Numerical weather prediction (NWP) models can typically forecast the probability and general intensity of precipitation occurring in a wider area, but they struggle at short spatial and temporal scales [1] because of the long running time and the time needed to assimilate data, i.e. to incorporate observational data used as the initial conditions. This problem is particularly severe with convective precipitation, which is associated with the highest rainfall rates, and originates from cells with a spatial scale on the order of a few tens of kilometers, making the exact location of the precipitation difficult to predict with NWP [2]. Experience over decades has shown that at lead times of minutes to a few hours, statistical and data-driven models that make optimal use of the latest available observations are useful tools for the short term prediction, or nowcasting, of precipitation. Such models have been widely deployed by meteorological agencies. A common way to implement precipitation nowcasting is Lagrangian extrapolation: using motion-detection algorithms to derive motion vectors from consecutive measurements of rainfall by weather radar, then advecting the precipitation field using these vectors to predict its future movement [3, 4].