Performance Analysis of Multiple Machine Learning Models Applied to Dissolved Gas Analysis Under Thermal Fault Classification Considering IEEE C57.104-2019TM Standard for Substation Transformer

The use of Dissolved Gas Analysis (DGA) as a non-intrusive test for assessing transformer health is widespread in the field. However, traditional DGA-based diagnostics have several limitations. The Rogers ratio method is confined to gases involved in the computation, the interpretation of the IEC Ratio method can be ambiguous, and the Duval triangle method is unable to differentiate between healthy and faulty oil degradation. Furthermore, these traditional methods rely on subjective judgment from experts. To address these shortcomings, a two-layer framework has been proposed in this paper. The framework utilizes the random forest algorithm with the IEEE C57.104 - 2019TM guide as a watchdog in the first layer to distinguish between unhealthy and normal oil degradation. The second layer employs a neural network to classify different thermal fault occurences in the transformer. Traditional methods like Doernenburg, Duval, Rogers, and IEC ratio methods have been used for years, with Duval Triangle showing the best accuracy at 87.18%. Machine learning methods outperform traditional ones, with Support Vector Machine achieving 95.70% and Neural Networks reaching the highest accuracy of 97.90%. Despite high accuracy, the machine learning framework faced challenges due to a limited and imbalanced dataset.