Machine Learning-Based Image Pattern Recognition Using Histogram of Oriented Gradient for Islanding Detection

A vital issue faced by the distribution network is the occurrence of unintentional islanding. The failure to identify unintentional islanding results in significant implications for both the power system and human lives. In this paper, a novel machine learning islanding detection method (IDM) based on image classification utilizing the histogram of oriented gradient (HOG) feature is proposed. In particular, the set of parameters are utilized, namely total harmonic distortion (THD) of both three phase currents and voltages, and rate of change of negative sequence voltage, are first transformed into time-frequency representations (i.e., spectrograms via the short time Fourier transform, and scalograms through continuous wavelet transform). Then, the HOG features are extracted from these images and used to train the machine learning (ML) algorithms to distinguish between occurrences of islanding and non-islanding events. Performance metrics including F1 score, recall, accuracy, precision and misclassification error are employed in the assessment process. Numerical results show that our image-based detector achieves faster detection times and higher detection accuracy versus state-of-art methods, thus confirming the validity of such approach for identifying islanding events.