ABSTRACT

The use of distributed generations as an integral part of conventional distribution networks is an aid to deal with sharply increasing demand of electricity. rt stabilizes power quality and provides an additional power supply to distribution network. The employment of DGs in power system networks is advantageous to power utilities, DG owners, and customers in terms of power quality, reliance and economics. Along with several advantages there are also some technical issues to be solved to fully adopt DG technology in power systems. The "Islanding" condition is one of the key issues in this regard. Islanding takes place due to power system disruption viz. faults, line and generator outages or any other disorder which can result in division of the system into some islanded networks. Up to now, several remote and local islanding detection methods have been proposed. Local islanding detection techniques are further divided into passive, active and hybrid detection methods. Remote detection techniques depend upon communication b/w the utility and the DG site, whereas, local techniques depend upon the measurement of the system parameters at the DG site. However, all existing techniques suffer from different limitations which cause
factual error in islanding detection. To address this issue, this research proposes a computational intelligence based techniques for islanding detection to overcome
limitation of these techniques.

The proposed islanding detection technique utilizes adaptive nuero-fuzzy inference system. In this technique, rate of change frequency, rate of change voltage, rate of
change of active power and rate of change of reactive power are used as input parameters for ANFIS. In order to test its effectiveness, a test system consisting of an existing Malaysian distribution network is simulated in PSCAD and diverse islanding & non islanding conditions are created to test and train ANFIS. The ANFIS training and testing results confirm that this technique is capable to accurately identify islanding and non-islanding cases. Furthermore, the need for threshold setting is also eliminated in the proposed technique. This high class accuracy makes it fit for implementing it in real systems.