A New Application of Machine Learning Based Methods for Disk Space Variation Fault Diagnosis in Transformer Windings

Frequency response analysis (FRA) is a highly effective methodology for discerning mechanical anomalies within the windings of power transformers. In this research, disk space variation (DSV), one of the most prevalent transformer winding faults, in various positions with different severities is considered on a 1.6 MVA distribution transformer with 20 kV windings, and the corresponding FRA results are measured and computed. All four frequency response components (amplitude, phase, real, and imaginary) are explored in this study's dataset. The constructed dataset is utilized to predict DSV faults using a trained Decision Tree, Random Forest, SVM Linear, SVM Polynomial, and XGBoost. The applied new machine learning (ML) approaches for interpreting FRA results are employed to obtain and extract crucial characteristics from frequency response test results so as to distinguish the location and intensity of DSV within the transformer windings. The outcomes of the experiments have validated the efficacy of the suggested intelligent techniques, which can predict fault location and extent with 93.5% and 100% accuracy in validation and test datasets, respectively.