A Data Mining Based Protection and Classification of Transients for Two-Core Symmetric Phase Angle Regulators

Abstract

Several conventional and non-conventional transient conditions cause differential relays associated with Phase Angle Regulators to malfunction. For Two-core Symmetric Phase Angle Regulators, this article investigates the suitability of time and time-frequency feature-based estimators to differentiate internal faults from other transient conditions such as overexcitation, external faults with current transformer (CT) saturation, and magnetizing inrush. Subsequently, the faulty core unit (series or exciting) is located, and the transients are identified. Six well-known classifiers are trained on features extracted from one-cycle of post transient 3-phase differential currents filtered by an event detector. Maximum Relevance Minimum Redundancy, Random Forest, and exhaustive search with Decision Trees are used to select the relevant wavelet energy, time-domain, and wavelet coefficient features respectively. The fault detection scheme trained on XGBoost classifier with hyperparameters obtained from Bayesian Optimization gives an accuracy of 99.8%. The reliability of the proposed scheme is verified with varying tap positions, noise levels, and transformer ratings; and under different conditions like CT saturation, fault during magnetizing inrush, series core saturation, low current faults, and integration of wind energy. As a potential application, the methodology can be deployed to supervise microprocessor-based differential relays to improve the security and dependability of the protection system.

Keywords

• Bayesian optimization
• fault detection
• feature selection
• machine learning
• phase shift transformer
• transient classification