The losses of the power grid can be divided into two main categories, which are technical and non-technical losses. Non-technical losses, also called commercial losses, are mainly caused by electrical theft. To help utility companies discover the doubts of electric burglars, we proposed a hybrid Convolutional Neural Network - Support Vector Machine model for electricity theft detection. In this model, first, we used the convolutional neural network to learn features between different days from massive and varying smart meter data by the operations of the convolution and the sub-sampling layers. In addition, a dropout layer was added to avoid over-fitting. After that, the Support Vector Machine was used as a classifier layer to replace the soft-max classifier layer to detect dishonest customers. Finally, we used the actual data set collected by State Grid Corporation of China, in which the electricity theft will be flagged (Flag = 1) to test the accuracy of the model, the evaluations were recognized through the confusion matrix and ROC curve. A comparison with previous popular models was carried out to verify the benefit of the proposed hybrid model.