A Method for Fault Section Identification of Distribution Networks Based on Validation of Fault Indicators Using Artificial Neural Network

Abstract

A fault section in Korean distribution networks is generally determined as a section between a switch with a fault indicator (FI) and a switch without an FI. However, the existing method cannot be applied to distribution networks with distributed generations (DGs) due to false FIs that are generated by fault currents flowing from the load side of a fault location. To identify the false FIs and make the existing method applicable, this paper proposes a method to determine the fault section by utilizing an artificial neural network (ANN) model for validating FIs, which is difficult to determine using mathematical equations. The proposed ANN model is built by training the relationship between the measured A, B, C, and N phase fault currents acquired by numerous simulations on a sample distribution system, and guarantees 100% FI validations for the test data. The proposed method can accurately distinguish genuine and false Fis by utilizing the ability of the ANN model, thereby enabling the conventional FI-based method to be applied to DG-connected distribution networks without any changes to the equipment and communication infrastructure. To verify the performance of the proposed method, various case studies considering real fault conditions are conducted under a Korean distribution network using MATLAB.

Application of back propagation neural network in complex diagnostics and forecasting loss of life of cellulose paper insulation in oil-immersed transformers

“…Twenty-six specific transformers utilized in distribution substations in South Africa, Mpumalanga, were exposed to the proposed diagnostics. The diagnostics were divided into five phases, which are described below 25 , 65 – 68 : Assessment of the transformer’s current condition: involves visual inspection and all DGA and chemical tests conducted on transformer oil. Gathering historical data: involves gathering information from prior faults, repairs, and maintenance.…”

“…The overall efficiency of the BPNN network is expressed by the value of R and the best BPNN network is identified based on its closest relationship value to 1 68 , 95 . The Levenberg–Marquardt (LM) training approach 64 , 65 was used since it is recommended as the preferable supervised algorithm in the MATLAB environment due to its fast training rate 116 . Figure 7 a,b show the coding of the two algorithms.…”