Fault diagnosis in power transformers based on dissolved gas analysis using hybrid stacking and voting ensembles

While machine learning techniques can successfully diagnose power transformer faults, improper application of Synthetic Minority Over-sampling Technique (SMOTE) results in data leakage and deceptive overfitting. In this study, a hybrid machine learning architecture with high generalization capability, free from the problem of overfitting for power transformers, has been proposed. To ensure methodological reliability, the dataset was strictly divided into a synthetically balanced training set (1120 samples) and a completely independent, pure test set (280 samples) that the model had never seen during the training phase. Basic algorithms (Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), k-Nearest Neighbors (k-NN), Support Vector Machine (SVM) and Decision Tree (DT) were optimized with hyperparameter constraints and regularization techniques to prevent memorization; then these models were used with hybrid voting and hybrid stacking methods. Experimental results have shown that the proposed Hybrid Stacking (DT + k-NN + SVM) model achieved a high accuracy of 94.29% on completely independent test data. Moreover, the difference between the model’s training and test performances remained at only 5.68%, provides evidence of satisfactory generalization that the system did not memorize synthetic data and can reliably diagnose faults in real field data.