Abstract

In this paper, a single fault-diagnosis framework of electrical systems operating under non-stationary conditions is proposed. The algorithm is a hybrid of spectral analysis with Short-Time Fourier Transform (STFT) and adaptive time-frequency analysis with Hilbert-Huang Transform (HHT) to obtain complimentary fault-relevant features. They are combined features which are fed to a trained neural-network classifier in a coordinated learning protocol. In contrast to traditional methods that use discrete signal-processing or learning methods, the suggested framework incorporates multi-resolution time-frequency representation into a hierarchical diagnostic process. The method is more sensitive to non-persistent and non-stationary faults and is also more robust to classification. The effectiveness of the proposed unified protocol to provide accurate and reliable fault identification is confirmed through the validation of its effectiveness through extensive simulation.

Key words: Active distribution grid; power distributed network; fault location; machine learning model; renewable energy.