Optimization-Driven Intelligent Transformer Fault Diagnosis Using Machine Learning: A Comparative Benchmark of Genetic Algorithm, Particle Swarm Optimization, AutoML, and Conventional DGA Techniques

Abstract

Power transformers are critical assets in electrical power systems, and their reliable operation depends on accurate and timely fault diagnosis. Although Dissolved Gas Analysis (DGA) is widely used for transformer condition monitoring, conventional diagnostic techniques such as the Rogers Ratio and Duval Triangle often exhibit limited accuracy due to predefined decision rules and their inability to effectively handle complex fault patterns. This study proposes an optimization-driven machine learning framework for intelligent transformer fault diagnosis by integrating Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Tree-based Pipeline Optimization Tool (TPOT) AutoML with supervised machine learning classifiers. Unlike existing studies that primarily investigate a single optimization strategy, the proposed framework provides a unified comparative evaluation of three optimization approaches and benchmarks them against conventional DGA techniques under identical experimental conditions. Experimental evaluation using five-fold cross-validation demonstrates that the GA-optimized Logistic Regression model achieved the highest testing accuracy of 96.22%, followed by PSO-optimized Logistic Regression (95.67%) and TPOT AutoML (95.00%). In comparison, the Rogers Ratio and Duval Triangle methods achieved accuracies of only 44.31% and 60.70%, respectively. Overall, the proposed framework improved diagnostic accuracy by 51.91 and 35.52 percentage points over the Rogers Ratio and Duval Triangle methods, respectively. These findings demonstrate that optimization-driven machine learning provides a robust, accurate, and practical solution for intelligent transformer health monitoring and predictive maintenance in modern power systems.