New Dissolved Gas Analysis Diagnostics Framework for Substation Transformers using Random Forest Algorithm and IEEE C57.104 – 2019TM Guide

Abstract

Dissolved Gas Analysis (DGA) is a practical, non-intrusive test to check transformer
health status, as it is widely used in the field. However, the traditional methods of DGA
based diagnostics have intrinsic weaknesses. For example, the Rogers ratio method is
limited only to gases involved in the computation. The interpretation of the IEC Ratio
method can be unknown at some point. The Duval triangle method cannot diagnose
healthy degradation of oil from faulty ones. All traditional methods were subject to
expert subjective judgment. To fill these gaps, this paper introduces the two-layer
framework using a random forest algorithm with the IEEE C57.104 – 2019TM guide
as a watchdog (layer 1) for unhealthy oil degradation versus normal ones. The
prediction model (layer 2) used the random forest algorithm. Using the 277 DGA
datasets from Distribution Utilities from different parts of the Philippines, the
framework surpassed the accuracy of traditional methods (Duval triangle method, IEC
ratio, Doernunberg method) with an accuracy of 100%. The Duval triangle got 98.92%
accuracy, the IEC ratio had 28.32% accuracy, and the Doernunberg method had an
accuracy of 27.50%. Other ML algorithms, such as ANN (MLP), K-nearest neighbors,
SVM (linear), and J48, were also used for comparison. The ANN (MLP), K-Nearest
neighbor, and SVM (linear) got 78.6%, 85.7%, and 78.6% accuracy, respectively. The
random forest got the highest cross-validation score (89.14% ave.) among all ML
methods. Further evaluations were used for J48, DT, and Random Forest since all got
100% accuracy. RF algorithm still got the highest PR-AUC (94%, 89%) and ROC
AUC (95%, 97%) scores among the J48 and DT in the 70/30 and 80/20 data split.