Abstract：Electricity theft by users is the main cause of non-technical loss of electric energy in power grids，

which causes huge economic losses and resource wastage to the power system. Compared with a large number of

users' normal electricity samples，electricity theft users belong to a minority class of samples，and the traditional

electricity theft classification methods perform poorly in the case of sparse or imbalanced samples. As a result，a

user electricity theft classification method based on the data enhancement of improved time series generative

adversarial network （TimeGAN）was proposed，TimeGAN was used to enhance the original small-sample

electricity theft data，generating the augmented samples similar to the distribution of the original data，and

considering that the augmented samples are noisy or untrustworthy，the quality of augmented samples was

evaluated using the Mahalanobis distance to complete the untrustworthy sample rejection. Convolutional neural

network（CNN）was used to extract features from the augmented electricity load data，and long-short time memory network（LSTM）was used to extract the temporal correlation of the feature quantities and complete the feature classification，and furthermore，the sparrow search algorithm（SSA）was used to optimize the parameters of the CNN-LSTM network，so as to improve the accuracy of the model detection. The experimental results show that the proposed method can effectively solve the binary classification problem of sample imbalance in the identification of user's electricity theft behavior.

      Key words：TimeGAN model；Marginal distance；sparrow search algorithm（SSA）；electricity theft identification