Abstract: The improvement of the accuracy of multiple classification of faults in smart meters is of great significance in the operation of power systems. In the traditional machine learning, the oversampling methods, such as the smote algorithm and its variants, seldom consider the global distribution of the data. And the subsequent classification algorithms are unable to obtain the deeper feature information from the data. We propose an imbalanced multi-classification model of CVAE-CNN. It takes labels as constraints and builds a CVAE network composed of fully connected layers to generate minority class samples. The network models the hidden variables that obey the multi-dimensional and independent Gaussian distribution in each dimension according to the lower bound of variation, and learns various distribution characteristics and the global characteristics of the data set to improve the quality of the generated data. The balanced data is classified using a CNN, in which the hidden complex features are extracted through the one-dimensional convolution operation. The maximum pooling method is constructed to improve the fault tolerance rate of the model, the recognition rate of a few categories increased. Taking 15 KEEL public data sets and the smart meter fault data collected in recent years as actual calculation examples, the proposed model is verified to have higher classification accuracy compared with the typical oversampling methods and classification methods.