Search

To improve the accuracy of hail forecasting, this study applies the random forest (RF) algorithm in hail identification and prediction in Shandong Peninsula. Hail observation data of 41 meteorological stations in Shandong Peninsula from 1998 to 2013 are used. The hail forecasting model with a 0–6 h range based on the RF algorithm is constructed using the convection index and related physical quantities calculated by the reanalysis data of the European Centre for Medium-Range Weather Forecasts during the same period. The model is built by undersampling within the RF algorithm (balanced RF), and the cross-validation is adopted to select the optimal forecast probability. The cross-validation exhibits high simulation accuracy, stable fitting effect, and small average generalization error. The performance of the balanced RF is tested by the independent data samples from 2014 to 2018, which shows excellent results. A trial report on the weather process on 13 June 2018 shows that the model is effective in identifying hail-fall areas and capable of forecasting all hail stations and the occurrence time of hail disasters. The RF algorithm focuses on thermal factors. The physical meaning of the selected factors is clear and consistent with the subjective prediction. The thresholds of the thermal factors, such as the lifted index, Showalter stability index, and total index, can be utilized as a reference for hailstorm prediction over Shandong Peninsula. Comparison of forecast (BFR model, the probability threshold is 0.4) and observation for hail weather over Shandong Peninsula on 13 June 2018. Colour shade is the probability of hail occurrence of the model output. Unlabelled Image • On the test model of random forest algorithm, the simulated probability for forecasting hail weather of detection, false alarm rate and critical success index are 100%, 20% and 80%, respectively. • The physical significance of key factors selected by the random forest model is clear and the physical factors are consistent with the experience of subjective forecasting. [ABSTRACT FROM AUTHOR]