An Intelligent Deep Convolutional Neural Networks-Based Islanding Detection for Multi-DG Systems

Abstract

Unintentional islanding is a problem in electrical distribution networks; it happens when the central utility is unintentionally separated from the rest of the distributed power system. The islanding detection problem becomes severe in non-detection zones. We propose an intelligent islanding detection technique with zero non-detection zone for a hybrid distributed generation system. It is based on the computation of frequency spectrum variations over time using short-term Fourier transform and convolutional neural networks. For various islanding and non-islanding occurrences, the three-phase voltage at the point of common coupling is monitored, and time-series data is collected. Then computations for a set of multiple frequencies on scaled time-series data are carried out, and complex numbers are split into magnitude and phase values. To detect islanding and non-islanding occurrences, a modified convolutional neural network with forward propagation was utilized. For the IEC 61850-7-420 test system, several islanding and non-islanding scenarios are created and deployed to train the convolutional neural network for the proposed approach. The efficacy of the proposed islanding detection learning model is assessed using 5-fold cross-validation. The findings reveal that under normal and noisy conditions, the proposed methodology has zero non-detection zone with original dataset, excellent accuracy, selectivity, and sensitivity.