Unbalanced Voltage Sag Dataset Enhancement Based on Improved Balancing Generative Adversarial Network

Abstract

In the actual power grid environment, there are severe class imbalances in the data sets of various monitoring systems. The existing deep learning methods are highly dependent on the training data sets, and it is difficult to mine the minority sample features in the imbalanced data sets, resulting in Classification recognition accuracy is severely limited. This paper takes the voltage sag as the research object, and proposes a dataset enhancement method based on an improved balanced generative adversarial network. This method combines the advantages of autoencoder and generative adversarial network, learns the public knowledge in the data set through the initialization training of autoencoder, and uses the majority class sample features to assist in generating minority class samples. In addition, an embedding layer is introduced into the self‐encoder architecture to add label information to the eigenvectors, which solves the problem of strong coupling between the distribution of voltage sag eigenvectors and improves the generation effect of samples. At the same time, the Wasserstein distance optimized by the gradient penalty term is used to guide the model adversarial training, which overcomes the problems of vanishing gradient and mode collapse of the original model. Through experimental analysis, the model can fully extract the features of minority samples under unbalanced data sets, ensure the distribution consistency between generated samples and original samples, and take into account the authenticity and diversity of generated samples, which can effectively improve the recognition effect of deep learning model on minority classes and improve the overall recognition accuracy. © 2023 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.