Abstract
In order to meet practical business application requirements, this study considered the physical characteristics of lightning, combined with the risk of lightning disasters and disaster responses, to develop a precise classification model for lightning safety risk warnings at target points based on the the multi-scale attention residual network. By analyzing 1404 sets of thunderstorm data from four warning levels in Guangdong region, we trained convolutional neural networks, support vector machines, random forests, extreme gradient enhancement networks, as well as the multi-scale attention residual network (MSA) constructed by our research institute. The results showed that the multi-scale attention residual network has interpretability for lightning safety risk warning, (1) This model proposes an attention mechanism to fuse different features, obtain the importance distribution of different features, and increase the number of neural network layers to extract deeper features. (2) This lightning safety risk warning model is the most reliable among the five models, with an accuracy rate of 93%. (3) For the four-level classification model, it was found that the accuracy of the lightning safety risk warning models based on the MSA remained above 70% (77%), and also achieved the highest recall, lowest standard deviation, and lowest log loss.