Improved forecasting via physics-guided machine learning as exemplified using “21·7” extreme rainfall event in Henan

As a natural disaster, extreme precipitation is among the most destructive and influential, but predicting its occurrence and evolution accurately is very challenging because of its rarity and uniqueness. Taking the example of the “21·7” extreme precipitation event (17–21 July 2021) in Henan Province, this study explores the potential of using physics-guided machine learning to improve the accuracy of forecasting the intensity and location of extreme precipitation. Three physics-guided ways of embedding physical features, fusing physical model forecasts and revised loss function are used, i.e., (1) analyzing the anomalous circulation and thermodynamical factors, (2) analyzing the multi-model forecast bias and the associated underlying reasons for it, and (3) using professional forecasting knowledge to design the loss function, and the corresponding results are used as input for machine learning to improve the forecasting accuracy. The results indicate that by learning the relationship between anomalous physical features and heavy precipitation, the forecasting of precipitation intensity is improved significantly, but the location is rarely adjusted and more false alarms appear. Possible reasons for this are as follows. The anomalous features used here mainly contain information about large-scale systems and factors which are consistent with the model precipitation deviation; moreover, the samples of extreme precipitation are sparse and so the algorithm used here is simple. However, by combining “good and different” multi models with machine learning, the advantages of each model are extracted and then the location of the precipitation center in the forecast is improved significantly. Therefore, by combining the appropriate anomalous features with multi-model fusion, an integrated improvement of the forecast of the rainfall intensity and location is achieved. Overall, this study is a novel exploration to improve the refined forecasting of heavy precipitation with extreme intensity and high variability, and provides a reference for the deep fusion of physics and artificial intelligence methods to improve intense rain forecast.