Document Type : Research Articles

Authors

• Hamid Reza Sezavar ¹
• Hamid Karimi ¹
• Navid Fahimi ²

¹ Qom University of Technology, Qom, Iran

² Iran University of Science and Technology, Tehran, Iran

Abstract

A comparative approach is pretend in this paper that evaluates different Artificial Intelligence (AI) methods for diagnosing power transformer faults using Dissolved Gas Analysis (DGA). Traditional approaches like the Rogers Ratio Method and Duval Triangle have been used for many years, but offer unreliable results for complex cases. Although, newer AI methods present better results, but still vary in how well they work. In this paper, several AI approaches are evaluated including Support Vector Machines (SVMs), Random Forest (RF), Gradient Boosting Machines (GBMs), Deep Neural Networks (DNNs) and a new combinational model is proposed based on comparing results. A real DGA dataset is used for covering six different fault types for the proposed testing. The results show that while all AI methods do better than traditional approaches, the combinational approach performs the best with 92.3% accuracy. This is found 20.2% better than traditional methods and 4.8% better than the best single AI model. Rational explanation is provided for how each method works and practical recommendation is presented for choosing the right approach based on particular requirements and available resources in real practices.

Keywords

• Transformer Diagnosis
• Dissolved Gas Analysis
• Machine Learning Comparison
• Hybrid AI Models
• Fault Classification

Main Subjects

• Power systems