Research on LeakGAN-Based Text Augmentation Technology for Transformer State Assessment

With the rapid increase in the proportion of renewable energy generation in power grids, the operational environment of electrical equipment has become increasingly complex, making accurate state assessment of critical devices such as transformers crucial for ensuring power system stability. However, the scarcity of transformer fault samples leads to prominent data imbalance issues, significantly constraining the diagnostic accuracy of data-driven models. To address this challenge, this study focuses on fault sample generation technology based on Generative Adversarial Networks (GANs), proposing an enhanced Information Leakage GAN (LeakGAN) model. By integrating key state parameters including transformer load coefficients, temperature data, and harmonic characteristics, the model strengthens its capability to characterize transformer operational behaviors and optimizes the quality of fault sample generation. Comparative experiments with four mainstream GAN variants were conducted, employing BLEU and ROUGE automated evaluation metrics alongside manual assessments to validate the superiority of the proposed model in terms of sample authenticity, diversity, and state relevance. The research establishes a comprehensive dataset for transformer state assessment, providing reliable support for data-driven fault diagnosis and condition evaluation, while offering novel optimization pathways for power equipment sample generation technologies.