Machine Learning-Based Islanding Detection Technique for Hybrid Active Distribution Networks

Passive Islanding Detection Techniques (PIDT) are renowned for their simplicity, absence of power quality degradation, and cost-effectiveness in implementation. However, these techniques are not without challenges, notably in the form of threshold setting limitations and a considerable non-detection zone (NDZ). To address these limitations, this research paper introduces a novel machine learning-based islanding detection technique proficient of detecting islanding events even with zero NDZ and without encountering threshold limitations. The study involves an in-depth analysis of the sensitivity of 16 indices commonly utilized in passive techniques for detecting islanding and non-islanding events. Data is meticulously collected by simulating diverse islanding and non-islanding cases on a Hybrid Active Distribution Network (HADN). For each scenario, measurements are taken for the 16 parameters to identify the most suitable minimum passive parameters that yield the highest accuracy. Three prominent classifiers—Multilayer Perceptron (MLP), Decision Tree (DT), and Random Forest (RF)—are selected for the classification task. The simulation results reveal that the dV/dP emerges as a highly effective parameter in distinguishing islanding events without the constraints of threshold limitations and zero power mismatch. Remarkably, the dV/dP parameter achieves 100% accuracy, precision, recall and F-Score with the MLP classifier, followed by RF and DT classifiers. This research offers valuable comprehensions into the potential of machine learning techniques for enhancing islanding detection in hybrid active distribution networks, offering a promising avenue for overcoming the limitations associated with traditional passive techniques.