Abstract:In the complex terrain of Sichuan Region, although the frequency of thunderstorm gale is relatively low, its impact is significant. There are few objective forecast products of the thunderstorm gale, and even if there are, the time resolution is lower. In order to further improve the accuracy of thunderstorm gale forecasting under complex terrain in Sichuan, in this article we comprehensively consider terrain factors, model physical quantity factors and time factors. According to the altitude, Sichuan is divided into highaltitude and low altitude areas. Based on the data from 2018 to 2021 and three machine learning methods of random forest, adaptive boosting and extreme random tree, we construct a thunderstorm gale prediction model and make a forecast for the 2022 thunderstorm gales obtaining a 3 h thunderstorm gale potential forecast. Then, based on the climate background, we downscale the 3 h forecast time to 1 h, and make a 0-12 h hourly thunderstorm gale forecast. At the same time, the forecasting effect is tested. The results show that, the adaptive boosting method of 3 h thunderstorm gale forecast has the best effect. The longtime and individual case tests show that the 0-12 h hourly thunderstorm gale forecast product obtained by the adaptive boosting method is superior to the forecasts of National Meteorological Centre with the TS score increased from 0.0104 to 0.0595, and the false alarm rate decreased from 0.988 to 0.808. This indicates that the adaptive boosting method has a higher application value in forecasting operation application value.