Abstract:
Power system reliability and stability depends considerably on proper operation of its transformers. Power transformer health monitoring and fault detection is very important for uninterrupted power flow. Moreover, power transformer faults may endanger the life of power system protection workers. Due to thermal, mechanical, chemical and electrical stresses, and aging, power transformer insulation materials such as mineral oil and cellulose paper condition deteriorate, and give off different kinds of gases, e.g. hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2). These dissolved gases are valuable parameters for the detection of power transformer faults. Conventional methods used in the industry suffer from problems such as low accuracy and ‘no decision’ in or ‘overlapping regions’ in various cases. To overcome the shortcomings, various machine-learning-based approaches are reported in the literature that provide very good accuracy. Empirical mode decomposition (EMD) and intrinsic time-scale decomposition (ITD) have been successfully used in providing effective features in classifying nonlinear signals with machine-learning yielding promising results. However, the application of these, especially ITD in transformer fault diagnosis is rather very limited. In addition, compared to EMD, ITD can generate better proper rotation components. Thus, there is ample scope for developing effective and novel machine-learning techniques in the EMD and ITD domains for power transformer fault classification and diagnosis.
In this thesis, empirical mode decomposition (EMD) and intrinsic time-scale decomposition (ITD)-based features are extracted from a large set of DGA parameters obtained from publicly available DGA data of 376 transformers. The parameters are ranked from lower to higher skewness. Single-level EMD and ITD are performed on the ranked DGA parameters. The resulting intrinsic mode functions (IMF) and proper rotation components (PRC) are used as features in XGBoost-based classifiers. The discriminative ability (for six classes of faults) of the features is demonstrated using box-whisker plots. The hierarchical ensemble classifier with EMD as well as the ITD- based XGBoost classifier provides high accuracy (more than 90%), sensitivity (average 90% or more) and F1-score (mostly near 1 or1), the latter being better. Overall, the proposed methods yield a superior performance as compared to conventional and recently reported machine-learning based techniques.
