Transient Stability Self-Adaptive Assessment Based on Sample Enhancement of DCGAN

Abstract

Since there are very few transient instabilities in the actual power system, the stable and unstable samples are extremely unbalanced, which brings great difficulties to the data-driven transient stability prediction. And due to the time-varying nature of the power system, the rapidity of the generation of new training samples in the online application of temporary stability prediction faces challenges. Aiming at the problem of decreased prediction accuracy when the AI network is applied to the transient stability prediction of the power system when the operation mode or topology of the power grid changes greatly, a method of transient stability self-adaptive assessment based on sample enhancement of improved DCGAN is proposed in this paper. While offline, perform cluster analysis on the historical topology of the power grid, generate a historical sample database, and train the temporary stability prediction model. When the topology of power grid changes, the new scene is matched with the historical scene and the samples are transfered to the new scene through a certain transfered principle. Then, the instability sample set is screened out, and the instability sample is enhanced by the improved Deep Convolutional Generative Adversarial Network(DCGAN). Finally, the offline model is fine-tuned with the new training samples after transfer and enhancement to realize online transient stability self-adaptive assessment. The experimental results of the Central China Power Grid show that the method can generate effective training samples for changing systems, balance power system instability data, and greatly reduce the sample generation time when the model is updated online, and improve the accuracy and Gmean value of transient stability prediction in new scenarios, which considering both the speed and accuracy of self-adaptive assessment.