Comparative Study on the Accuracy of the Conventional DGA Techniques and Artificial Neural Network in Classifying Faults Inside Oil Filled Power Transformers

Abstract

Power transformers are expensive yet crucial for power system reliability. As the installed base ages and failure rates rise, there is growing interest in advanced methods for monitoring and diagnosing faults to mitigate risks. Power transformer failures are often due to insulation breakdown from harsh conditions like overloading, that leads to prolonged outages, economic losses and safety hazards. Dissolved Gases Analysis (DGA) is a common diagnostic tool for detecting faults in oil-filled power transformers. However, it heavily relies on expert interpretation and can yield conflicting results, complicating decision-making. Researchers have explored Artificial Intelligence (AI) to address these challenges and improve diagnostic accuracy. This study investigates using Machine Learning (ML) techniques to enhance DGA for diagnosing power transformers. It employs an Artificial Neural Network (ANN) with Feed Forward Back Propagation, a Bayesian Regularizer for predictions, Principal Component Analysis (PCA) for feature selection and Adaptive Synthesizer (ADASYN) for data balancing. While traditional DGA methods are known for their accuracy and non- intrusiveness, they have limitations, particularly with undefined diagnostic areas. This research focuses on these limitations, to demonstrate that ANN provides more accurate predictions compared to conventional methods, with an average accuracy of 76.8% versus lower accuracies of 55% for Dornenburg, 40% for Duval, 38.4% for Roger and 31.8% for IEC (International Electrotechnical Commission) Methods. The study findings prove that ANN can effectively operate independently to improve diagnostic performance.