Due to the difficulty of the inspection of distributed photovoltaic (PV) electricity theft, the number of electricity theft samples collected by relevant departments is limited, which cannot meet the needs of data-driven electricity theft detection. This paper proposes a data augmentation method for distributed PV electricity theft using Wasserstein generative adversarial network (WGAN). First, WGAN can explicitly learn the time correlation that is difficult to model in the PV electricity theft data sequence. Furthermore, it can generate new electricity theft samples with similar distributions to the real ones through the confrontation training of the generator and discriminator networks. Then, according to the typical PV electricity theft model and data characteristics, the convolutional neural network (CNN) is selected for electricity theft detection. Finally, through the case analysis, it is shown that the electricity theft samples generated by WGAN can conform to the fluctuation law of authentic samples and the probability distribution characteristics of historical data, thereby effectively improving the detection performance.