Intelligent islanding detection framework for smart grids using wavelet scalograms and HOG feature fusion

Unintended islanding presents substantial operational and safety risks in modern electrical distribution networks, particularly as distributed generation (DG) sources increasingly match or nearly match local load requirements. Conventional islanding detection schemes (IDS) often fail under balanced load-generation conditions, resulting in significant undetected events, commonly referred to as the non-detection zone (NDZ). This research addresses these critical limitations by introducing a novel, highly reliable, and robust machine learning-based islanding detection scheme. The proposed approach innovatively utilizes Histogram of Oriented Gradient (HOG) features derived from scalogram images, which are generated through Continuous Wavelet Transform (CWT) of the total harmonic distortion (THD) signals from three-phase voltages and currents. The HOG descriptors effectively capture the intricate patterns and subtle signal changes associated with islanding conditions. A Random Forest classifier, selected for its robustness against noise and minimal parameter tuning, is trained and validated extensively using these descriptors. Comprehensive performance assessments under various noise conditions and challenging scenarios demonstrate that this methodology significantly surpasses existing state-of-the-art methods in terms of accuracy, precision, recall, F1-score, and reduced misclassification errors. Furthermore, real-time testing using the OPAL-RT platform confirms the practical applicability, reliability, and robustness of the proposed system. This work significantly advances the capabilities of islanding detection, providing a highly effective solution to enhance grid stability and ensure safety in contemporary power systems.