Unsupervised learning in islanding studies: Applicability study for predictive detection in high solar PV penetration distribution feeders

Abstract

Unintentional islanding is a pressing issue associated with integration of distributed solar photovoltaic generation with a distribution network. The probability of its occurrence is usually dominated by the photovoltaic penetration however, as a direct consequence of this, load-inverter dynamic interactions alongside grid-side disturbances can also lead to anomalous instances that can become responsible for accidental island creation and one such anomaly has been described in this work. Given such dynamic behaviours occurring on photovoltaic inverter-integrated distribution feeders, threshold based classical islanding detection can not suffice and hence machine learning based techniques have began to be researched and adopted. However the orientation can be directed towards predictive approaches leveraging knowledge extraction from huge event data available in smart grids. Furthermore, unsupervised learning can be explored for real-time applications to enable self-learning and acting systems. This paper presents preliminary results of application of a self-organizing map neural network for preemptive detection of unintentional islanding by classifying the discovered islanding precursor from other power system events. Classification of a three phase short-circuit fault at the point of common coupling was found to be invariant to input feature reduction however the same gives contrasting results for the other two test cases investigated.