Power Quality Disturbance Classification Using Wavelet-Transform Features and Ensemble Machine Learning

PQ disturbances, voltage sags, swells, interruptions, harmonics, flicker, and transients, degrade grid reliability and damage sensitive equipment. Automating their classification is essential for modern distribution-network monitoring, particularly in Indonesia's PLN grid, where renewable-energy integration accelerates PQ degradation. This study proposes a two-stage pipeline: DWT feature extraction followed by ensemble machine learning. Level-5 wavelet multiresolution analysis decomposes each voltage waveform using three wavelet families (db4, sym4, coif4), yielding 24 features per signal (energy, entropy, RMS, and standard deviation across six bands). Four classifiers, SVM (baseline), Random Forest, XGBoost, and CatBoost, are evaluated on 3,550 synthetic signals from IEEE 1159 parameters across 13 disturbance classes with realistic imbalance. SMOTE oversampling is applied to training folds only to prevent leakage. The best configuration (coif4 + CatBoost) achieves 94.2% accuracy and 0.932 macro-F1. On the public Mendeley PQ Dataset (699 samples, 12 classes), ensembles exceed 98% accuracy. A signal-to-noise-ratio sweep (20-50 dB) confirms graceful degradation: accuracy remains above 73% at 20 dB. Wavelet-domain features paired with gradient-boosted ensembles offer a reproducible, high-performance approach to PQ disturbance classification. The headless Python pipeline is lightweight and edge-deployable for IoT-ready smart-grid PQ monitoring.