Data-driven transient stability assessment method has become the focus of research in the field of power network security. However

transient unstable situation in the actual power system is very rare

which brings great difficulties to the data acquisition method for judging the instability. This paper proposes a data augment method for the synthesis of unstable samples in the transient stability assessment. It enhances the adaptability of training methods for conditional generative adversarial network(CGAN)to improve their learning stability and uses the improved CGAN training generators and discriminators during offline training to learn the distribution characteristics of raw data. Then

the extreme learning machine(ELM)classifier is used to filter out the generated samples with the highest G-mean value among the multiple sets of samples generated by the improved CGAN. The unstable samples are used to augment the original unstable samples

and the augmented original samples are used to train the classifier to achieve online transient stability assessment. The simulation results show that the proposed method can effectively learn the distribution characteristics of the original data by the improved CGANs. The method has the advantages of strong anti-noise interference and good robustness to high-dimensional data

and it can effectively balance the unstable data of power system.