High impedance fault detection using combination of multi-layer perceptron neural networks based on multi-resolution morphological gradient features of current waveform

In this study a new pattern recognition-based algorithm is presented for detecting high impedance faults (HIFs) in distribution networks with broken or unbroken conductors and distinguishing them from other similar phenomena such as capacitor bank switching, load switching, no-load transformer switching (through feeder switching), fault on adjacent feeders, insulator leakage current (ILC) and harmonic load. The proposed method has employed multi-resolution morphological gradient (MMG) for extraction of the time-based features from three half cycles of the post-disturbance current waveform. Then, according to these features, three multi-layer perceptron neural networks are trained. Finally, the outputs of these classifiers are combined using the average method. Applying the data for HIF, ILC and harmonic load from field tests and for other similar phenomena from simulations has shown high security and dependability of the proposed method. Also, a comparison between the features from the proposed MMG-based procedure and the features from discrete Fourier transform, discrete S-transform, discrete TT-transform and discrete wavelet transform is made in the feature space.