Physics-Guided Deep Learning for Quantitative Precipitation Nowcasting
Abstract
Accurate short-term quantitative precipitation forecasting (QPF) or 'nowcasting' is vital for hydrological and water resource management, such as developing urban flooding warning systems. The state-of-the-art methods in precipitation nowcasting include extrapolation-based methods applied to radar echoes, numerical weather prediction (NWP) models, and hybrid methods that blend disparate approaches for improved forecasting. Despite the diversity of techniques used, QPF remains a challenging problem - further exacerbated by the increasing intensity and frequency of extreme precipitation events under a changing climate. Recent advancements in deep learning (DL) methods combined with ever-increasing computational power have demonstrated remarkable performance across various problems in image and language processing. Emerging research at the intersection of QPF and DL suggests the latter may also improve upon existing methods for QPF by taking advantage of growing data from in-situ sensors, radar stations, and satellites. In this work, we test the hypothesis that DL-based methods can outperform existing state-of-the-art methods for QPF with reduced latency. Further, we propose integrating process-based and data-driven paradigms to develop a physically consistent DL model (PCDL) that forecasts precipitation 6-8 hours in the future at a kilometer-scale spatial resolution and a temporal resolution of the order of 10 minutes. The aspired objectives of the PCDL model are to go beyond point predictions into uncertainty quantification and hydrometeorological explainability, as well as potentially generalizing to newer and possibly even data-poor regions. For a robust analysis, we evaluate our model against various methods ranging from naive Persistence to the best operational NWP model available from the National Oceanic and Atmospheric Administration (NOAA) in the United States.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMGC45F0870Y