Abstract
This paper develops an islanding classification mechanism to overcome the problems of non-detection zones in conventional islanding detection mechanisms. This process is achieved by adapting the support vector-based data description technique with Gaussian radial basis function kernels for islanding and non-islanding events in single phase grid-connected photovoltaic (PV) systems. To overcome the non-detection zone, excess and deficit power imbalance conditions are considered for different loading conditions. These imbalances are characterized by the voltage dip scenario and were subjected to feature extraction for training with the machine learning technique. This is experimentally realized by training the machine learning classifier with different events on a 5 kW grid-connected system. Using the concept of detection and false alarm rates, the performance of the trained classifier is tested for multiple faults and power imbalance conditions. The results showed the effective operation of the classifier with a detection rate of 99.2% and a false alarm rate of 0.2%.