Abstract: In order to solve the problems of low detection efficiency and low accuracy caused by the complex features of power data and unequal data samples, a power data analysis model based on deep learning is proposed. First of all, considered the imbalance of power theft data samples and the limited number of samples, a condition generation countermeasure network based on Wasserstein criterion is proposed to balance the power theft data and improve the diversity of data. Secondly, an automatic coder based on superposition convolution noise reduction is proposed to extract the user's power behavior characteristics, so as to improve the user's power feature extraction ability and enhance the model training efficiency. Finally, a power data classification model based on gradient lifting decision tree is proposed, which can effectively reduce the over fitting problem and improve the classification accuracy of the model. In the experimental stage, the proposed model is analyzed and verified by taking the power consumption data set released by the State Grid Corporation of China as an example. Compared with random oversampling（ROS）, synthetic minority over-sampling technique（SMOTE）, generative adversarial network（GAN）and other data enhancement methods, the proposed data enhancement method can effectively improve the training performance of the model. In addition, compared with logistic regression (LR), support vector machine (SVM), long-term and short-term memory network (LSTM) and other models, the proposed model has better performance in the test set, with accuracy and recall of 89.3% and 69%, respectively. The simulation results further verify the effectiveness and accuracy of the proposed model.