After the PV power supply is connected to the distribution network, the uncertainty, intermittency and fluctuation of PV power generation increase the difficulty of identifying faults in the distribution network. To address this problem, this paper proposes a method combining the entropy-variance modal component with a neural network to improve the ResNet model. Firstly, a PSCAD simulation model of the distribution network containing PV power is built to obtain batch data under different complex fault scenarios. Secondly, the entropy-variance modal (E-VMD) method is used to reconstruct the feature matrix of the samples, and then the improved residual network is used to further explore the implied features of the fault samples, and then the model is trained and tested. In comparison with the classification results of other models in the literature, the improved ResNet model achieves an average accuracy of 99.95% for fault type identification and 99.75% for fault feeder identification, and has good robustness, which can effectively achieve fast fault identification in distribution networks containing PV power supplies.