Fault Location in Distribution Grid using Machine Learning

Foo, Kheng Hong (2019) Fault Location in Distribution Grid using Machine Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Foo Kheng Hong.pdf Restricted to Registered users only Download (1MB)

Abstract

Distribution system is the section after the transmission system that is designed to deliver electrical energy to consumers in an area. Automated power system protection schemes in a distribution grid can be complemented using machine learning (ML) based fault location system. Such systems can perform very well if there are no major changes to the configuration of the distribution grid. Due to the advancement of technology, distributed generation (DG) can be employed to generate power near the end-users with the goal of reducing transmission losses. However, this may cause the power to flow in both directions in the distribution grid which may negatively affect any existing ML-based fault location schemes. This effect needs to be investigated to develop suitable countermeasures. In this work, a distribution grid is simulated with different fault types at various location to train an artificial neural network (ANN) for fault location. DIgSILENT PowerFactory was used to simulate a modified IEEE 16-bus test system. Instead of using all three feeders, feeder 3 was excluded since it is essentially a mirror of feeder 1. The main feeders were energised at 23-kV, and each bus was connected to a balanced 1 MVA, 0.85 lagging power factor load. One 4.5-MVA, three-phase PV was used as DG unit in this investigation and was installed at one bus at a time. A total of four types of faults were considered which are single-line-to-ground, line-to-line, double-line-to-ground, and three-phase-to-ground. All four faults were simulated at all nine sections at every 5 % interval of the section length. The fault resistance used was varied from 0 to 25 Ω at an interval of 0.1 Ω. The ANNs were trained in MATLAB using fault data consists of root mean square (RMS) values of voltage sag magnitude and voltage angle of all buses at a 70:30 ratio for training and testing. Results show that the addition of PV has negligible impact on the fault type classification, minor reduction in fault section detection accuracy and major increase in fault distance prediction error. Upon investigating, it was identified that section that still transfer power in a unidirectional fashion after the additional of PV will not significantly affect the ANN prediction results. Most of the error occurs at sections that transfer power in both directions after the addition of PV. Therefore, to overcome this issue, simulation of fault data when the PV is connected should be collected to train a separate, new ANN. Whenever the PV is connected, the active ANN trained without PV should be replaced with the new ANN trained when PV was connected.