Abstract: Severe convective weather is a type of weather disaster with extremely destructive power, but due to its suddenness and small scale, it is still difficult to accurately forecast in meteorological business operations. Therefore, there is an urgent need for a model based on Fully Convolutional Networks to improve the above shortcomings.This article uses ground observations of severe convective weather from 2017 to 2021 and ECMWF (the European Center of Medium range weather forecasts) numerical model data as the training set, and 2022 as the test set. The FCN (Fully Convolutional Network) algorithm was adopted to construct three types of severe convective weather forecast models for hail, convective gust, and short-term heavy rainfall within 0-12 hours in Gansu Province. The model was applied in forecast business operations in 2023. The results are as follows. (1) In the training set, the FCN model performed well, with a total false negative rate (FNR) of only 16.6% for severe convective weather and non-severe convective weather. In the test set of 2022, the average critical success index (CSI) of the three types of severe convective weather was 25.8%, and the average probability of detection (POD) remained above 65.2%. Short-term heavy rainfall had the best forecasting effect. (2) In the validation set of forecast business operations in 2023, the average CSI of the three types of severe convective weather was 24.3%, the average POD was 62.6%, and the average false alarm ratio (FAR) was 71.2%. For all data sets, the CSI of short-term heavy rainfall was the highest, reaching 45.5%. At the same time, the POD of short-term heavy rainfall and convective gust exceeded 70%, while the FAR results of hail and convective gust were similar, both exceeding 77%, and the average FAR of short-term heavy rainfall was the lowest. (3) For hourly data, hail performed best at hours 4, 8, and 10, convective gust performed best at hour 6, and short-term heavy rainfall performed best at hours 2 and 4. Therefore, the FCN model constructed in this study performs well in forecasting severe weather, providing a broad application prospect for future automation of meteorological business.