Abstract
Precipitation nowcasting is crucial across various industries and plays asignificant role in mitigating and adapting to climate change. We introduce anefficient deep learning model for precipitation nowcasting, capable ofpredicting rainfall up to 8 hours in advance with greater accuracy thanexisting operational physics-based and extrapolation-based models. Our modelleverages multi-source meteorological data and physics-based forecasts todeliver high-resolution predictions in both time and space. It captures complexspatio-temporal dynamics through temporal attention networks and is optimizedusing data quality maps and dynamic thresholds. Experiments demonstrate thatour model outperforms state-of-the-art, and highlight its potential for fastreliable responses to evolving weather conditions.