Autocorrelation Aided Islanding Detection Using bi-directional Long-short Type Memory Network

In the present work, autocorrelation aided deep learning framework for islanding detection in grid connected distributed generation system is proposed. For this purpose, islanding along with other transient events were simulated on a grid connected power system network with DG penetration. For each case, negative sequence voltage signals obtained at the point of common connection were used to determine the sequence components of the autocorrelation function. From the autocorrelation sequences representing each type of transient event, 36 features were extracted. The obtained feature vectors were fed as inputs to a bi-directional longshort type memory network classifier for classification of islanding and other events. It has been examined that the suggested methodology has resulted in 99.01% accuracy in discriminating islanding from non-islanding events. Besides, for the multiclass classification, a mean accuracy of 98.50% is obtained. Comparative studies with machine learning classifiers indicated that the result of the suggested methodology is better. The proposed model can be used for accurate prediction and classification of islanding and other transient events in power system network.