Compact wavelet-packet energy–entropy feature selection for overhead-line fault classification
Reliable classification of overhead transmission-line faults becomes more difficult when increased fault resistance weakens transient current components and redundant features reduce class separability. This study develops a compact wavelet-packet energy–entropy feature-selection framework for limited-sample fault-family classification. Three-level wavelet packet decomposition is first applied to the three-phase fault currents. Eight three-phase-averaged terminal-sub-band energy proportions are then combined with three phase-wise energy-entropy values to construct an 11-dimensional intermediate feature vector. Candidate feature subsets are evaluated by the mean five-fold cross-validation error of a one-versus-one multiclass linear support vector machine, while the Pied Kingfisher Optimizer is used to search the feature-subset space. In the investigated 230 kV, 200km transmission-line simulation system, the selected five-feature subset achieved an accuracy of 95.4% (103/108), compared with 92.6% (100/108) for the complete 11-dimensional baseline. The corresponding Wilson 95% confidence interval for the selected subset was 89.6%–98.0%. Among the three tested search methods, PKO achieved the highest classification accuracy, retained the smallest feature subset, and required the shortest recorded optimization time. The resistance-binned results showed larger accuracy gains in the 10–50 Ω and 150–300 Ω intervals, while stronger noise reduced classification performance. These results demonstrate that classifier-oriented selection of a compact energy–entropy representation provides a favorable balance among classification accuracy, feature dimensionality, and interpretability for transmission-line fault-family classification.