Skilful precipitation nowcasting using deep generative models of radar

Precipitation nowcasting, crucial for weather-dependent decision-making, is typically done using radar-based wind estimates and advection of precipitation fields. However, this struggles with nonlinear events. Deep learning methods directly predict rain rates but lack constraints, leading to blurry long-term nowcasts. This paper introduces a deep generative model for probabilistic nowcasting, improving forecast quality, consistency, and value. The model produces realistic predictions across large regions and lead times, outperforming competitive methods in expert meteorologist evaluations (89% ranked first for accuracy and usefulness). This generative nowcasting approach provides probabilistic predictions that enhance forecast value and operational utility, especially at resolutions and lead times where other methods falter.

Suman Ravuri, Karel Lenc, Matthew Willson, Dmitry Kangin, Remi Lam, Piotr Mirowski, Megan Fitzsimons, Maria Athanassiadou, Sheleem Kashem, Sam Madge, Rachel Prudden, Amol Mandhane, Aidan Clark, Andrew Brock, Karen Simonyan, Raia Hadsell, Niall Robinson, Ellen Clancy, Alberto Arribas, Shakir Mohamed