Non-linear high impedance fault distance estimation in power distribution systems: A continually online-trained neural network approach

Abstract

This paper presents a new methodology for high impedance fault (HIF) location in overhead power distribution systems. A polynomial function was used to model the voltage at fault point as function of the fault current. Additionally, a continuously trained neural network after the occurrence of HIF was used. This neural network is used to estimate unknown parameters present in the equations that model the feeder during an HIF. The proposed algorithm uses measurements of voltage and current taken only at the substation together with the feeder parameters. A typical 13.8kV distribution system is used to test and validate the proposed scheme. The performance of the method was evaluated according to the fault current amplitude and the system load level. In addition, a comparative analysis with a state-of-the-art method was also performed. The HIF distance estimation errors remained below 1% in 86% of the tested cases. The maximum error obtained was 2.3%. Hence, such good performance along with the simplicity of the method and low cost of implementation, make this methodology suitable for real distribution feeder.