A comparison framework for fault classification methods in power system distribution network

The supply of electricity is an increasingly essential service in modern society and, like this one, has passed through several changes during its evolution. The increased complexity of the electrical energy systems from the generation to the distribution has brought with it several adapables, practical and legal impositions, restrictions, as well as more attention to certain areas of research. Since the continuity of the electricity supply one of its main characteristics, many researches turn their attention to problems involving the breaking this continuity, being the faults one of the well-known reasons. When a fault is detected, the protection systems must act quickly so that the system stability is preserved. Given the importance of the fault event, several studies are focused on its detection, classification and location. This work compared the faults classification methods most used in modern Electricity Distribution Systems like Neural Networks, Trees, Logistic Regression and others. The accuracy and robustness of the algorithms were performed using the data collected in simulations of different types of fault such as line to ground (LG), double Line (LL), double line to ground (LLG) and three phase fault involving ground (LLLG). An analysis of possible interference in the acuracies of algorithms was also carried out in the face of variations in the intensity of the fault impedance, presence of distributed generation, noisy data and reduced dimensionality. All fault scenarios have been modeled using the OpenDSS software and IEEE-13 bus scenario was used. After the simulations, it was noticed that the Neural Networks showed a better ability to generalize the problem even in scenarios with noisy data. Dimensionality reduction was more effective only in the noiseless scenarios, decreasing the accuracy of the methods when the data were noisy.