Enhanced Power Transformer Fault Diagnosis Using Key Chemical Gases with DGA, Integrating Machine Learning and Traditional Methods

Abstract

The reliability of fault diagnosis and the stability of electrical grids demand accurate analysis of key chemical gases in power transformer oil. DGA, which quantitatively measures concentrations of critical chemical gases, is still a cornerstone technique in the field, but is infamous for its interpretative complexity: the correlation between gas levels and specific fault types is too complex. The proposed methodology in this study attempts to integrate the strengths of traditional methods of DGA interpretation along with the power of a machine learning model, specifically a Random Forest algorithm. The process comprises the preprocessing of DGA data to extract meaningful chemical features from them further developing the model using machine learning to classify the different kinds of faults based upon those chemical features. This approach has been validated on multiple scenarios for the data coming from DGA transformer faults after a lot of testing. Results show that this method delivered an average accuracy of 95.86% for three types of faults and 93.67% for the same types of faults with varying conditions. For six types of faults, the delivery was placed at an average accuracy and consistency of 88.85% and 87.47%, respectively. This approach significantly shows improved performance in the traditional methods of diagnostics while promising much more accurate fault detection. In addition to enhanced diagnostic accuracy, it supports proactive, hence preventive, maintenance strategies, resulting in improved system efficiency and reduced downtime. The paper details a technique that combines chemical data analysis with machine learning, from which distinct solutions can be conceived to address the complex challenges facing industries.