Abstract
Vacuum surface flashover remains a major constraint for a variety of high-voltage vacuum insulation systems. Because flashover is governed by numerous coupled effects of material, geometry, voltage waveform, surface condition, and discharge history, reliable flashover threshold prediction using statistical and physics-based models remains challenging. To address this issue, a first-of-its-kind machine-learning (ML) based framework for vacuum surface flashover is proposed. First, a comprehensive vacuum flashover database is constructed. Six ML models were evaluated on the original dataset, and random forest (RF) and Histogram Gradient Boosting (HGB) demonstrated superior performance to the others. To further increase the prediction accuracy, several data augmentation methods were employed to increase the size of the database. It was found that threefold PureBootstrap performed the best, achieving a column shapes score of 96.10% and a column pair trends score of 97.78%. After augmentation, the dataset containing 902 samples and 19 features was used to further train RF and HGB for flashover threshold prediction. Compared with the J. C. Martin empirical formula and physical model based on the secondary electron emission avalanche, the ML models achieve substantially better prediction performance, with R 2 values of 0.92 for RF and 0.93 for HGB, higher than 0.24 and 0.22 for the J. C. Martin formula and the physical model, respectively. The employed ML models are interpretable, and illustrated that insulation system geometry and applied voltage waveform are among the most critical factors determining the flashover threshold. This work offers novel insights into flashover risk assessment in vacuum insulation design.
Funding Information
  • National Natural Science Foundation of China (52537007)

This publication has 81 references indexed in Scilit: