Fangchenggang City, Guangxi, China (CNFAN), is a typical region for severe thunderstorms, with the annual thunderstorm days exceeding 85. Frequent thunderstorms have led to high rate of lightning strike trip-out on transmission and distribution lines, thereby threatening the operational stability of power grid. A great engineering value has become increasingly apparent, for developing an early warning system of severe thunderstorm potential (12 ~ 24 h in advance). This study established a two-stage artificial neural network (ANN) based method for predicting severe thunderstorm potential and intensity, using six-year continuous daily ground flash frequency data and meteorological parameters across 35 geographic grids covering the CNFAN area. In the first stage, a probabilistic classification model was developed to predict the occurrence potential of severe thunderstorms. In the second stage, a regression model was constructed to further predict ground flash frequencies and geometric mean (GM) of lightning current amplitudes for severe thunderstorms. Validation results indicate that this method achieved a probability of detection (POD) of 94.48% for severe thunderstorms. Over 30% of samples exhibited relative prediction errors below 50% for ground flash frequencies and over 80% of samples exhibited relative errors under 50% for the GM of lightning current amplitudes. Through the two-stage modelling strategy, this study not only achieves high-accuracy potential prediction of severe thunderstorms, but also reveals their intensity characteristics. The results can provide multiple risk assessment bases for the early warning of lightning hazard to power grids.