Improving the Forecast of Extreme Rainfall Event “21.7” in Henan Province Base on a Physical Guided RainNet

Extreme precipitation is one of the most destructive and influential natural disasters, and the accurate prediction of its occurrence and evolution is very challenging due to its rarity and uniqueness.

From 17-22 July 2021, Henan suffered unprecedented extreme heavy precipitation, with significant extremes on cumulative rainfall, daily rainfall, hourly rainfall, and the coverage of heavy rain. The maximum process rainfall amounted to 1,122.6 mm, the maximum daily cumulative precipitation of 777.5 mm occurred in Hebi city, and the maximum hourly precipitation amounted to 201.9 mm, which broke the historical record of hourly rainfall intensity in China's inland. The maximum hourly rainfall occurred in Zhengzhou, a large city with a population of over ten million and a GDP of over one hundred million, the flash floods and urban flooding triggered by the extreme heavy rain caused significant casualties and hazards to people's lives and socio-economic properties.

In this study, we took the "21.7" extreme precipitation event in Henan Province in 2021 as the research object, explored the potential of using physical guided machine learning method to improve the forecast of extreme intensity and location of the precipitation. Three ways of physical guiding were used in this paper: (1) physical analysis of the anomalous circulation and thermo-dynamical factors were made and selected as the input feature; (2)understanding of the multi-model forecast bias and its physical reasons were used to ensemble; (3) professional knowledge of forecasters were considered to design the loss function, the optimization index and constraints are more suitable for the physics and distribution of the precipitation，and the RainNet was built. The study showed that by learning the relationship between anomalous physical features and heavy precipitation, the intensity of precipitation forecast can be significantly improved, but the location was rarely adjusted, meanwhile larger false alarms occurs. It may because the anomalous features used here mainly contained information of large-scale systems and factors, and it was consistent with the model precipitation deviation, on the other hand, the samples of extreme precipitation were sparse, and then the algorithms adopted was simple accordingly. However, by ensemble "good and different" multi model with machine learning, the advantages of each model were extracted and then the location of precipitation center was significantly improved. Hence, combining the appropriate anomalous features with multi-model fusion, a integrated improvement of intensity and location were achieved. The study provides an innovative exploration to improve the refined forecasting of heavy precipitation with high extremes and high variability, it also provides a reference for the deep fusion of physics and the artificial intelligence methods in the future to improve intense rain forecast.