ransformers are a crucial part of electrical power transmission and distribution systems. Fault prediction with high accuracy reduces downtime, lowers maintenance costs, and increases the lifespan of transformers. This study presents a novel multi-featured machine learning framework for transformer fault prediction, evaluated over dual-scale datasets. Machine learning techniques Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and K Nearest Neighbour (KNN) are used to detect transformer faults using two different datasets that include historical information such as temperature, vibration, oil level, humidity, load current, oil viscosity, and power transformer age. The performance of machine learning methods is compared in terms of accuracy, precision, F1 score and recall using two test case studies. The first case considered a simple dataset of 50 thousand, and the second case study involved complex data of more than one lakh transformers. The Decision Tree and Support Vector Machine models attained the highest classification accuracy across both datasets, with SVM showing near-perfect accuracy but failing to give meaningful precision, recall, and F1-score values—indicating potential overfitting or class imbalance in the present experimental evaluation. The Random Forest algorithm has obtained the best results in recall and F1 score, which indicates that it is better to recognize the fault pattern in complex operating conditions. While logistic regression has a moderate level of recall, it suffers from low precision and predictive power. The K-Nearest Neighbour method shows moderate performance in all the metrics and performance is quite consistent as the size of the dataset’s changes. The decision tree method produced a trade-off between interpretability and predictive ability that was a compromise.
These findings proved the feasibility of machine learning for transformer fault diagnostics, with notable implications for real-time health monitoring in advanced power systems.