Kong et al. (2026) Combining AI and numerical models for improving short-medium range precipitation prediction

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2026
• Date: 2026-09-15
• Authors: Hanbing Kong, Guoxing Chen, Ziyin Zhang, Wei-Chyung Wang, Jie Feng, Guihua Wang, Yijun Zhang, Yuejian Zhu
• DOI: 10.1038/s41612-026-01547-w

Research Groups

• Department of Atmospheric and Oceanic Sciences, Institute of Atmospheric Sciences, Fudan University, Shanghai, China.
• Shanghai Frontier Science Center of Atmosphere-Ocean Interaction and Research Center of AI for Earth Sciences, Fudan University, Shanghai, China.
• Institute of Urban Meteorology, China Meteorological Administration, Beijing, China.
• Atmospheric Sciences Research Center, University at Albany, State University of New York, Albany, NY, USA.

Short Summary

This study presents a hybrid approach to enhance short-to-medium-range precipitation forecasts by combining deep learning and numerical weather prediction (NWP) models. The Fast Precipitation Prediction (FPP) model outperforms the NCEP Global Forecast System (GFS) in predicting surface precipitation.

Objective

• Develop a data-driven FPP model that predicts 24-hour total precipitation using meteorological fields, bypassing uncertainties in NWP cloud parameterizations.
• Enhance short-to-medium-range precipitation forecasts by driving the FPP model with NWP-predicted meteorological fields.

Study Configuration

• Spatial Scale: The study focuses on the Chinese mainland, covering an area of approximately 9.6 million km².
• Temporal Scale: The study covers a period of 44 years (1979–2022) and evaluates forecasts for lead times ranging from 1 to 16 days.

Methodology and Data

• Models used: The FPP model uses a deep neural network architecture, specifically the Attention U-Net, to predict precipitation.
• Data sources: ERA5 reanalysis data (1979–2022) and CHM precipitation dataset (1979–2022) were used for training and validation.

Main Results

• The FPP model outperforms the GFS in predicting surface precipitation, with a lower root-mean-square error (RMSE) and higher pattern correlation coefficient (PCC).
• The FPP model demonstrates superior performance in predicting moderate to heavy precipitation at lead times beyond 48 hours.
• The study evaluates the hybrid approach using various metrics, including RMSE, PCC, equitable threat score (ETS), and false alarm rate (FAR).

Contributions

• This study presents a novel hybrid approach that combines deep learning and NWP models to enhance short-to-medium-range precipitation forecasts.
• The FPP model outperforms traditional NWP models in predicting surface precipitation, demonstrating its potential for operational implementation.

Funding

• This research was supported by the National Natural Science Foundation of China (Grant No. 41975044) and the Shanghai Municipal Education Commission (Grant No. 2021QH0010).

Citation

@article{Kong2026Combining,
  author = {Kong, Hanbing and Chen, Guoxing and Zhang, Ziyin and Wang, Wei-Chyung and Feng, Jie and Wang, Guihua and Zhang, Yijun and Zhu, Yuejian},
  title = {Combining AI and numerical models for improving short-medium range precipitation prediction},
  journal = {npj Climate and Atmospheric Science},
  year = {2026},
  doi = {10.1038/s41612-026-01547-w},
  url = {https://doi.org/10.1038/s41612-026-01547-w}
}