Ensemble-Based Fault Detection Using Single-Phase and Positive-Sequence Phasors of the IEEE 14-Bus System in a Digital Twin

To advance real-time transmission-line protection, this paper constructs a HYPERSIM-based cyber-physical testbed of the IEEE 14-bus grid, injects seven fault types at 60 km intervals, and streams PMU phasors through SEL-351S relays and an SEL-3530 RTAC for software-in-the-loop validation. A two-stage framework first detects anomalies and then classifies the fault, benchmarking deep networks (CNN, LSTM, CNN-LSTM) against ensemble trees (Random Forest, XGBoost, LightGBM) on both positive-sequence and individual-phase phasors. With positive-sequence inputs, accuracy peaks near 55 % for seven-class discrimination but exceeds 94 % when faults are coarsened into four classes, whereas individual-phase features push every model beyond 97 % and allow ensembles to hit 0.994 accuracy and 0.99 F1. These findings indicate that phase-resolved PMU data, paired with lightweight, interpretable ensembles, can enable near-perfect, low-latency fault discrimination on existing substation hardware.