PREDICTIVE MAINTENANCE OF POWER TRANSFORMERS USING DISSOLVED GAS ANALYSIS AND MACHINE LEARNING

Abstract

Power transformers are critical assets in transmission and distribution networks, and unexpected failures can cause extensive equipment damage, service interruptions, maintenance costs, and reduced grid reliability. Predictive maintenance provides an effective alternative to purely time-based maintenance by continuously evaluating transformer condition and identifying developing faults before they progress into severe failures. Dissolved Gas Analysis is one of the most widely used diagnostic techniques for oil-filled transformers because abnormal thermal and electrical activity produces characteristic gases within the insulating oil. However, conventional DGA interpretation can become difficult when gas concentrations overlap across different fault types or when several abnormal conditions occur simultaneously. This study presents a machine-learning-based predictive maintenance framework that uses dissolved gas concentrations, gas ratios, preprocessing, feature selection, and fault classification to assess transformer condition. A Random Forest classification model is considered for identifying normal operation, thermal faults, electrical faults, and discharge-related conditions. Representative results indicate a classification accuracy of 96.4%, precision of 95.8%, recall of 95.2%, and F1-score of 95.5%. The proposed framework demonstrates that combining DGA with machine learning can improve fault classification, support early-warning decisions, and provide a practical foundation for condition-based transformer maintenance