Measuring the atmospheric electric field is of crucial importance for studying the discharge phenomena of thunderstorm clouds. If one is used to indicate the occurrence of a lightning event and zero to indicate the non-occurrence of the event, then a binary classification problem needs to be solved. Based on the established database of weather samples, we designed a lightning prediction system using deep learning techniques. First, the features of time-series data from multiple electric field measurement sites are extracted by a sparse auto encoder (SAE) to construct a visual picture, and a binary prediction of whether lightning occurs at a specific time interval is obtained based on the improved ResNet50. Then, the central location of lightning flashes is located based on the extracted features using a multilayer perceptron (MLP) model. The performance of the method yields satisfactory results with 88.2% accuracy, 92.2% precision rate, 81.5% recall rate, and 86.4% F1-score for weather samples, which is a significant improvement over traditional methods. Multiple spatial localization results for several minutes before and after can be used to know the specific area where lightning is likely to occur. All the above methods passed the reliability and robustness tests, and the experimental results demonstrate the effectiveness and superiority of the model in lightning short-time proximity warning.