Machine Learning Based Approach by Image Recognition for Detection of Islanding

Excessive penetration of DG sources poses a significant difficulty known as islanding. The act of islanding has the potential to inflict harm upon both customers and their equipment. As per the IEEE 1547 DG interconnection criteria, it is necessary to identify islanding within a two-second timeframe and then deactivate the DG. The present study introduces a novel methodology for identifying islanding through the utilisation of visual pattern recognition and Machine Learning classification techniques. The Histogram of Oriented Gradient (HOG) is employed as a feature descriptor in order to perform pattern recognition on images, with the objective of identifying both non-islanding and islanding occurrences. Therefore, the spectrogram images are acquired from the time-series signal that exists at the point of common coupling. The provided photos are employed for the purpose of extracting features using the HOG algorithm. These extracted features are subsequently used as input for the k-Nearest Neighbors (KNN) classifier, both for training and testing purposes. The spectrogram image is obtained by using the rate of change of the negative sequence voltage component. The efficacy of the KNN classifier is evaluated using 5,10,20, and 25 -fold cross-validations. The classification findings indicate that the utilisation of the KNN classifier and HOG feature for visual pattern recognition yielded exceptional outcomes, with an accuracy rate of 96.75% and a detection time of 211 ms.