Sampling-Frequency-Aware Wavelet-Statistical Features for Fault Diagnostics in Inverter-Integrated Medium-Voltage Networks

Abstract

Inverter-based resources (IBRs) alter fault behaviour in medium-voltage (MV) networks, while heterogeneous monitoring devices operating at different sampling rates provide different bandwidths and representations of fault transients, complicating consistent feature formulation for data-driven fault diagnostics. This study develops a sampling-frequency-aware wavelet–statistical framework for fault detection, classification and location (FDCL). Voltage and current signals from 1800 simulated cases spanning six fault categories on a 33 kV radial feeder were sampled at 120 kHz, with anti-aliased decimation producing datasets at 12 and 1.2 kHz. Six Nyquist-valid feature families were extracted at each sampling rate, from which compact subsets were formulated using minimum-redundancy maximum-relevance (MRMR) and ReliefF feature selection. Four supervised machine-learning (ML) models—decision tree, k-nearest neighbour (KNN), bagged-tree ensemble and neural network—were evaluated for classification and regression using an 80/20 train–test split. Classification remained comparable across sampling tiers, with best test accuracies of 94.9%, 95.1% and 94.9% at 1.2, 12 and 120 kHz, respectively. Regression was more configuration-dependent: the 12 kHz full-feature neural network achieved the lowest unseen-location root-mean-square error (RMSE) of 2.603 km, with mean absolute error (MAE) = 1.845 km and coefficient of determination (R2) = 0.992, while MRMR-20/KNN retained a comparable RMSE of 2.659 km. Selected features shifted from predominantly wavelet descriptors at lower rates towards high-frequency band-energy descriptors at 120 kHz without improved prediction. The results show that increasing sampling bandwidth expands the observable frequency content but does not guarantee improved diagnostic accuracy; predictive performance depends on sampling frequency, physically valid feature formulation, feature selection and learner choice.