Aiming at the high cost and difficulty of obtaining labeled data for power grid companies， and the difficulty of training an effective electricity theft detection model with unlabeled data， this paper proposes a method based on CT-GAN （Co-training Generative Adversarial Networks） semi-supervised electricity theft detection method. Firstly， the principles and structures of generative adversarial networks and semi-supervised generative adversarial networks are explored. Secondly， it is proposed to replace the JS （Jensen-Shannon） divergence and KL （Kullback-Leibler） divergence distance with the Wasserstein distance to solve the problem of unstable model training and low quality of generated data caused by the gradient disappearance and mode collapse of the generative confrontation network problem， and built a multi-discriminator Co-training model to avoid the problem of high distribution error of a single discriminator. At the same time， it enhanced the ability of GAN to generate label sample data. By expanding the label sample data set， the model detection accuracy and generalization ability were improved. Finally， the accuracy and effectiveness of the method are verified using the Irish power grid dataset.