Machine learning model for dissolved gas analysis: methodological review with a case in Hong Kong

Abstract

Dissolved gas analysis is a valuable diagnostic tool used to monitor transformer health by analysing the gases dissolved in insulation oil. However, its practical application is hindered by the absence of a universal standard, leading to varied interpretations and implementations across different contexts. Scholars have turned to machine learning to advance DGA anomaly detection, but the existing literature prioritises model development over methodological rigour; issues such as dataset imbalance, appropriate evaluation metrics, and testing and validation procedures are often overlooked. This study addresses the existing gaps by critically reviewing the methodological steps involved in machine learning modelling with DGA datasets. The proposed design considerations are justified, and the analysis of the IEC TC 10 dataset and a dataset from a Hong Kong company is enhanced. Using Python’s scikit-learn, over 18 combinations of datasets, data preprocessing techniques, and model architectures were assessed, and the random forest model (F1 = 0.88) and the support vector classifier with an RBF kernel (F1 = 0.87) were identified as the top-performing models, after applying Yeo-Johnson transformation. The models’ source code is available on the author’s GitHub repository.