Investigation of a fast islanding detection methodology using transient signals

A novel approach for fast detection of power islands in a distribution network using the transient signals generated during the islanding event is investigated. Performance of several pattern recognition techniques in classifying the transient generating events as islanding or non-islanding was examined. Discrete wavelet transform of the transient current signals are utilized to extract feature vectors for the classifiers. Samples of the feature vectors corresponding to various islanding and non-islanding events are applied to train (i) a decision tree classifier, (ii) a probabilistic neural network classifier, and (iii) a support vector machine classifier for recognizing the transient patterns originating from the islanding events. The trained classifiers were then tested with unseen test current waveforms. The test results demonstrated that the investigated technique can potentially provide a new way for identification of islanding in distribution systems.