Lightning prediction plays an important role in lightning protection of power systems. For the problems of low accuracy and complex modeling of lightning prediction methods in power systems today, a new lightning spatio-temporal proximity prediction model, LG-ResUnet, was proposed with the SEVIR dataset as the background. The model, based on the Unet architecture incorporates the residual structure and the spatio-temporal attention mechanism, which can adaptively capture the spatio-temporal dynamic change characteristics of lightning. Experiments were conducted to test the performance comparison with several existing models under the relevant evaluation indexes such as probability of detection (POD), critical success index (CSI), and mean squared error (MSE). The results show that the hit rate and critical success index of the LG-ResUnet model are improved by 44.3% and 38%, respectively, compared with other models, and the mean squared error is reduced by 4% on average at a blitz threshold of 5 times. The ablation experiment verifies that the overall model performance is improved by more than 10% on average under the joint effect of residual structure and attention mechanism. Meanwhile, the extrapolation results under different thunderstorm scenarios show that the model has better generalization ability and robustness advantage, which has certain application and reference value.