Power transformer failures represent a significant risk to electrical grid reliability and can cause costly outages. Accurate fault diagnosis is essential for preventive maintenance and operational safety. Traditional dissolved gas analysis—the industry standard for detecting transformer faults—relies on chemical sampling and interpretation, which can be time-consuming and operator-dependent.
Researchers have now demonstrated that modern machine-learning techniques substantially improve fault detection performance. The study evaluated multiple algorithms, including Light Gradient Boosting Machine (LGBM), Categorical Boosting, and Multilayer Perceptron neural networks, on a dataset of 700 transformer samples compiled from the International Technical Committee 10 and Egyptian Electricity Holding Company databases.
LGBM was selected for its computational efficiency and ability to process large datasets through histogram-based learning and leafwise growth strategies. Categorical Boosting was chosen for its superior handling of categorical variables while reducing bias and overfitting risk. The Multilayer Perceptron captured nonlinear fault patterns that tree-based models alone could miss.
The hybrid LGBM + Extra Trees Classifier ensemble achieved the highest performance, with accuracy ranging from 93% to 98%—a substantial improvement over both conventional dissolved gas analysis methods and individual machine-learning approaches. This combination leverages the strengths of both boosting and ensemble methods, enhancing both generalization and reliability across diverse fault scenarios.
The implications for utilities are significant. Machine-learning models can process transformer operational data in real time, enabling rapid fault identification before catastrophic failures occur. Implementation could reduce diagnostic turnaround times, lower maintenance costs, and improve grid resilience. As utilities digitalize their asset management systems, integrating these algorithms into condition-monitoring platforms offers a path toward more predictive, data-driven transformer management strategies across power networks worldwide.
