Existing power anomaly detection is mainly based on analyzing static offline data. But this method takes a long time and has low identification accuracy when detecting timing and frequency anomalies, since this method requires offline screening, classification and preprocessing of the collected data, which is very laborious. Anomaly detection with supervised learning requires a large amount of abnormal data and cannot detect unknown anomalies. So, this paper innovatively proposes the idea of applying Time-series Generative Adversarial Networks (Time-GAN) in a dispatching automation system for the identification, diagnosis and prediction of massive data flow anomalies. First of all, regarding the problem of insufficient abnormal data, we use Time-GAN to generate a large number of reliable datasets for training fault diagnosis models. In addition, Time-GAN can ameliorate the imbalance between various types of data. Secondly, unsupervised learning methods such as Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and K-means are used to detect unknown anomalies that may exist in the power grid. Finally, some supervised learning methods are selected to compare with unsupervised learning methods. Experimental results show that the proposed algorithm has a higher recognition rate of known anomalies than other benchmark algorithms and it can find new unknown anomalies. It lays a good foundation for the safe, stable, high-quality and economical operation of the power grid.