Fault Detection Using Machine Learning in the Southwest Part of Bangladesh Power System

Fault identification is crucial for the fast restoration of power system to pre-fault state. The coastal south-west region of Bangladesh is frequently affected by several forms of natural disaster which has made the transmission system of this area particularly vulnerable to faults. In this paper, we present a model of this power-system region with actual generation, distribution, and line data along with distance information obtained from google maps. Using our realistic model, we perform fault identification study on this system with three different machine learning classifier algorithms. The performance of the algorithms were tested for the case of fault occurrence at the four major power plant locations of this area. When trained with the dataset of individual target area, all the models show more than 99% accuracy in fault identification, with the highest accuracy 99.999%. The effect of distance and branching of the power system was also observed in cross-evaluation results, where the trained data set was obtained from one surrounding power plant of the target area. It is observed that the distance between test and train areas plays a less significant role than the number of branches between those areas.