Abstract: Traditional electricity theft detection methods mostly build models based on one-dimensional power consumption data sequence, and a single classification model usually limits the deep mining of the law of electricity users’ behavior. In order to further improve the detection rate of electricity theft behavior, a novel electricity theft detection method based on the graph transformation and hybrid convolutional neural network is proposed. First, in order to better capture the difference in the periodic characteristics of electricity consumption before and after electricity theft behavior of users, the Gramian angular summation field image transformation method is introduced to realize the two-dimensional transformation of the electricity consumption data. Then,according to different dimensional forms of input, the hybrid convolutional neural network is applied to synchronously extract the global features of the original one-dimensional sequence data and the temporal correlation features retained by the two-dimensional image data. Actual case results show that the three performance indicators of the proposed model, i. e. area under the receiver operation characteristic curve, recall, and F1-score, have been effectively improved compared with the models such as random forest and wide ＆ deep convolutional neural network.