Power Grid Safety Assessment Based on an Improved Generative Adversarial Network

Abstract

Critical clearing time (CCT) is one of the important indices for transient stability evaluation. From the data-driven perspective, the research on CCT prediction is troubled by changing models and insufficient samples, that is, the small sample characteristics of the target system may be submerged by the large sample characteristics of the source system. In this paper, a prediction method of CCT based on improved generative adversarial network (WGAN) is put forward to solve this problem, and the network structure of WGAN is redesigned. This method is suitable for the transition phase of power grid when the operation mode changes. Through the unsupervised training of WGAN, the neural network will automatically learn the complex relationship between the small sample data of the target system, so that the generator can produce small sample data with high precision. Then the CCT prediction model was built on the expanded balanced dataset. Results validate that the method can effectively learn the distribution law of samples, improve the CCT prediction ability, and has the advantage of precision in the transition phase of power grid with insufficient samples.