Preliminary Voltage Analysis and Explainable Hybrid LSTM-XGBoost Framework for Kualanamu Airport 20 kV Distribution Network

Abstract

Voltage reliability in a 20 kV distribution network is important for maintaining airport operational continuity. This study uses ten MVDP 2 monitoring screenshots as preliminary observations and proposes an Explainable Hybrid LSTM-XGBoost framework for multi-horizon voltage-state prediction. The screenshots provide phase-to-phase voltage, three-phase current, frequency, and, in several observations, active power and power factor. The average line-to-line voltage across the ten observations ranges from 19.770 to 20.840 kV, with an overall mean of 20.355 kV, while frequency ranges from 49.93 to 50.22 Hz. Using the internal preferred operating band of 20.2-20.5 kV only as a preliminary descriptive label, 5 observations are classified as Low and 5 as High, with no Normal observation. The photographs are therefore useful as a pilot dataset and variable-feasibility evidence, but they are not sufficient for LSTM training or t+5, t+15, and t+30 prediction because continuous timestamps are unavailable. The main dataset should be obtained from timestamped SCADA/power-meter historical exports. Random Forest, XGBoost, LSTM, and Hybrid LSTM-XGBoost will subsequently be evaluated using walk-forward validation, macro F1-score, MCC, abnormal-class recall, and SHAP.