Generative adversarial network-enabled learning scheme for power grid vulnerability analysis

Abstract

Real measurements of power grids are usually limited for research and modelling of extreme events such as the impact of typhoons due to confidentiality concerns. To overcome the dearth of valuable, trustworthy data, this paper proposes an adaptive learning method based on the generative adversarial network. To obtain informative examples, the falsely classified examples together with examples that are correctly classified with low confidence are used to train a GAN for producing synthetic examples to reinforce the learning. The new power grid examples are selected according to the likelihood of the true data distribution. An evaluation was conducted with data acquired by the China Southern Power Grid in Hainan. Most significantly, the performance of detecting the occurrence of a power grid fault under the impact of typhoons is greatly improved. It was demonstrated that the proposed method improved the performance of predicting power grid fault in extreme events by 8.9%. Using the modulated GAN network, the synthetic data closely follows the distribution of the real data as indicated by large p-values. Our method takes minutes to complete training a model, which enables an efficient response to disasters with modern computing facilities such as edge computing.