Hybrid physics-AI outperforms numerical weather prediction for extreme precipitation nowcasting

Area of Science:

• Hydrology and Meteorology
• Artificial Intelligence and Machine Learning

Background:

• Precipitation nowcasting is crucial for flood and river management but remains a significant challenge.
• Existing Numerical Weather Prediction (NWP) models struggle with accurate detection, particularly for extreme events.
• Recent advancements in deep generative modeling (DGM) and physics-embedded machine learning (ML) offer potential improvements.

Purpose of the Study:

• To evaluate the performance of a state-of-the-art physics-embedded DGM, NowcastNet, against NWP models for precipitation nowcasting.
• To conduct a co-evaluation with water resources and river managers to assess practical applicability.
• To identify the strengths and limitations of advanced ML models in hydrological forecasting.

Main Methods:

• Comparison of NowcastNet with the High Resolution Rapid Refresh (HRRR) NWP model, advection, and persistence.
• Evaluation using metrics such as the Critical Success Index (CSI) for heavy precipitation events (extremes over 16 mm/h).
• Co-evaluation involving ML developers, hydrologists, and river managers.

Main Results:

• NowcastNet significantly outperformed HRRR, advection, and persistence, achieving a median CSI of 0.30 versus 0.04 for HRRR for extreme precipitation.
• The model demonstrated hydrologically-relevant improvements in point-by-point forecasts.
• Caveats include a tendency for overestimation of spatially aggregate precipitation at longer lead times.

Conclusions:

• Physics-embedded DGMs like NowcastNet represent a significant advancement over current NWP models for precipitation nowcasting, especially for extreme events.
• These findings suggest potential for enhanced flood emergency response and hydropower management.
• Further research and validation are needed to address limitations in aggregate precipitation forecasting.