School of Electrical Engineering, Beijing Jiaotong University, Beijing 100044, China

The massive measurement data lays the foundation for the data-driven transient stability prediction method. However, the inherent unbalance nature of unstable samples restricts the performance of this type of methods. In order to solve the problem of sample imbalance in transient stability prediction, a data augment method is proposed based on the improved deep convolutional generative adversarial network (DCGAN). New and effective unstable samples are generated by adversarial training of the generator and discriminator, which are used as a supplement to the original training set. In offline training, a deep belief network is used as the classifier, and the extended sample set is used for training, which effectively improves the recognition rate of the unstable samples. In online application, once the system changes unexpectedly, the offline model is updated by samples transferring and model fine-tuning technology, and after that, the unstable samples are further enhanced, which can greatly improve the transfer speed of transient stability adaptive assessment and the recognition rate of unstable samples in the new scenarios, making the evaluation results more reliable. The simulation results on IEEE 39-bus system and IEEE 140-bus system verify the effectiveness of the proposed method.