An Experimental Study of the Short-time Heavy Rainfall Event Forecast Based on Ensemble Learning and Sounding Data

Sounding analysis is one important method for short-term heavy rainfall event forecasting. By using sounding data of 119 stations at 0800 BT of 1 June -30 September during 2015-2019, based on XGBoost integrated learning framework, a prediction model for short-term heavy rainfall events (not less than 20 mm·h^-1) is proposed. Sounding data and derivative physical elements are used as characteristics parameters. The model can forecast whether short-term heavy rainfall occurs around the sounding station in following 12 h. Then an optimization method of high-risk weather is proposed. Using piecewise cost function as a loss function, different weighting factors are used to make the model more sensitive. This will ensure the total number of false prediction samples do not increase, but more false alarms rather than missing ones, leading to a slight increase on threat score (TS), a great improvement on probability of detection (POD) and the false alarm rate (FAR) will not exceed the threshold such as 0.5. After that, two tests are designed including a weighted sensitivity test for the piecewise loss function and a comparison test of the loss function using 12 datasets of 7 regional center sounding stations. The efficiency of the model optimization method is verified and the prediction ability before and after the improvement are compared. At last, a test of national short-term heavy precipitation forecast is designed, by using sounding data from 1 June to 30 September in 2019 as independent test set. Results show that reducing w_TP will decrease the number of hits and false alarm of the model's forecast; reducing w_FN will increase the number of hits and false alarms; w_TN and w_FP have little influence on the prediction. Compare with other commonly used cost function, the model with piecewise weight cost function has better forecasting skills, in which the TS is improved by 0.05-0.1, the POD is increased by more than 0.15, and the FAR is improved by 0.05-0.1. The model shows a clear tendency of forecasting positive instead of missing. In addition, the model shows similar results in all independent experiments, indicating that the optimization method has consistent effects on the results. The independent test of the national short-term heavy rainfall forecast experiment shows that the improved model has a certain short-term heavy rainfall forecast ability, with POD of 0.66, FAR of 0.37, and TS of 0.47. Above all, a short-term heavy rain prediction model is constructed based on the integrated decision tree and sounding data. The optimization method which could enhance the forecast skill of model is also proposed and verified.