Non-intrusive Load Monitoring (NILM) seeks to monitor the energy consumption and the usage of individual appliances in real-time through the total power reading of the whole unit. To improve the generalization capability and accuracy of the NILM model, the training dataset needs to be expanded accordingly. However, the collection of large amounts of power data is challenging and becomes a common conundrum in NILM. To solve the data shortage problem, TimeGAN, which takes advantages of unsupervised GAN and supervised training, is applied in the NILM field to generate realistic and high-quality power data. The results demonstrate that the data generated by TimeGAN is superior in quality to the baseline both from a qualitative and quantitative perspective.