Transformer fault diagnosis using dissolved gas analysis: a hybrid ensemble model with data preprocessing

Abstract

Failures and guaranteed dependability of the electrical grid, early fault diagnosis in power transformers is essential. By examining gas ratios suggestive of faults, dissolved gas analysis (DGA) continues to be a vital component for transformer health monitoring. Using four preprocessing techniques raw data, min-max normalization, logarithmic transformation, and square root transformation; this study suggests a machine learning method for fault detection using DGA gas ratios (such as CH₄/H₂, C₂H₂/C₂H₄). Random forest (RF), support vector machines (SVM), gradient boosted trees (GBT), and a hybrid RF-GBT model that uses prediction fusion were the four supervised classifiers assessed. Performance was assessed using accuracy, precision, recall, f1-score, and Cohen's kappa. Experimental results show that the hybrid RF-GBT model with logarithmic transformation achieves the highest performance, with 94.93% accuracy and 92.37% Cohen's kappa, significantly outperforming individual classifiers. Data preprocessing, particularly logarithmic and square root transformations, enhances diagnostic robustness by mitigating feature skewness. This study underscores the importance of tailored preprocessing and ensemble methods for reliable transformer fault diagnosis.