The atmospheric electric field (AEF) has significant importance in thunderstorm warning. This paper first analyzed AEF variations during thunderstorms based on 386 samples and then divided the period into four stages, in which a certain time stage provided valuable information for lightning prediction. It was defined as the “Stage T.” Fast Fourier transform was employed to study the frequency-domain patterns of the AEF in this stage. The result showed that there were notable differences in the frequency spectrum distribution between thunderstorm and non-thunderstorm weather conditions. During thunderstorms, the AEF frequency spectrum amplitude was higher and the waveform exhibited a wide range of variations compared to non-thunderstorm conditions. Then we used the Euclidean distance classifier to discriminate between thunderstorm events and non-thunderstorm events in the modeling samples for lightning forecasting. The remaining 256 samples were used to validate and evaluate the effectiveness of the algorithm. For prediction time of 5min, 15min and 25min, probability of detection(POD) is greater than 77.78%, false alarm rate (FAR) is less than 16.67%, critical success index (CSI) is greater than 0.57. These results indicated that this method was effective for short-term thunderstorm forecasting.