Rethinking Shallow and Deep Learnings for Transformer Dissolved Gas Analysis: A Review

Abstract

Dissolved gas analysis (DGA) of power transformers has attracted attention for years. Extensive machine learning techniques have been adopted to DGA for fault classification. Recently, deep learning (DL) techniques have been brought to deal with DGA issues, while their performances are not significantly improved compared to shallow learning (SL) algorithms. For a comprehensive investigation, this article tests popular SL algorithms and reports DL algorithms on four different DGA datasets. The results show that SL algorithms have efficient capacity for DGA analysis, while DL algorithms may not as great as they expect. In addition of complex structure and numerous parameters to tune, DL algorithms may even perform worse than SL algorithms. This work can be a reference for future DGA algorithm development.