A lightning warning method based on atmospheric electric field features and deep learning algorithms is proposed to address the problem that the existing lightning warning method based on atmospheric electric field is ineffective. Firstly, the frequency domain features of the atmospheric electric field are extracted using ensemble empirical mode decomposition （EEMD）, and the above features are further dimensionally compressed by sparse auto encoder （SAE）, and combined with LSTM neural network to distinguish whether it is a thunderstorm weather or not. Finally, the effectiveness of the method is tested by using the lightning location data. The results show that the lightning warning method can predict lightning more accurately, and the highest accuracy of 81% can be achieved by using 45-minute time series of electric field. Compared with the traditional method, the proposed algorithm can be adapted for training and debugging for different regional conditions and has universal applicability. As the model runs longer and the amount of data increases, it can be updated and iterated to ensure the forecast performance is continuously improved. Thus, the method proposed in this paper has reference value and indicative significance for further improving the lightning protection level in Guangzhou.