Abstract
Dissolved Gas Analysis (DGA) is an essential online diagnostic for assessing the integrity of power transformer liquid and solid insulation. To improve the consistency and automation of DGA interpretation, this study uses IEEE C57.104‑2019 key gases as features (𝐻2, 𝐶𝐻4,𝐶2𝐻2, 𝐶2𝐻4, 𝐶2𝐻6, 𝐶𝑂, 𝐶𝑂2) and evaluates two supervised machine‑learning classifiers—K‑Nearest Neighbors (KNN) and Support Vector Machines (SVM)—for automated fault classification (Normal, Partial Discharge, Thermal, Arcing). The dataset comprises online DGA measurements from four unit transformers installed at Kusile Power Station; each is a three‑phase, three‑winding, core‑type unit transformer rated at 110 MVA with a 22/15.75/15.75 kV configuration and an off‑load tap changer. These units operate under comparable loading and environmental conditions, enabling consistent benchmarking of algorithms.
Online data were cleaned, standardized, and split using a stratified 70/30 train–test partition, with SMOTE applied to the training set to address class imbalance. KNN is employed as a non‑parametric, instance‑based learner using Euclidean distance in the standardized feature space, while SVM is a margin‑based learner using an RBF kernel to capture non‑linear boundaries. Model performance is quantified via accuracy, precision, recall, F1‑score, and confusion matrices. Across the four transformers, both models achieved high accuracy, with KNN slightly outperforming SVM on three datasets: Transformer 1, 96.54% (KNN) vs 96.66% (SVM); Transformer 2, 95.06% vs 91.67%; Transformer 3, 91.74% vs 90.86%; and Transformer 4, 93.26% vs 92.45%. Class‑wise reports and confusion matrices show Arcing as the most separable class, while most misclassifications occur between partial discharge and thermal, reflecting overlapping gas profiles. Overall, the comparison indicates that on four comparable in-service transformers using the same key-gas feature set and preprocessing workflow, both a distance-based KNN and a kernel-based SVM approach can deliver similarly strong fault-classification performance, and that the remaining classification uncertainty is concentrated primarily in separating partial discharge from thermal conditions.