Fault Location Method for Active Distribution Network Based on SVM and Feature Search Algorithm

As a large number of distributed generators (DGs) are connected to the power grid, the traditional distribution network begins to transform to an active distribution network, correspondingly the traditional fault location method is no longer applicable. Therefore, a new fault location method based on support vector machine (SVM) combined with Sequential Forward Selection (SFS) and Sequential Backward Selection (SBF) is proposed. The method comprises four steps: the first step is data extraction and optimal M-dimensional feature selection, the second is SVM model training, the third is to determine the optimal feature quantity and SVM classifier model, the fourth is to identify suspected fault areas. Firstly, when each section of the active distribution network has faults, the phase voltages of all measuring points are extracted respectively, and the phase voltages are decomposed into fundamental frequency and 2
^nd
-7
^st
harmonic components by Fourier analysis. Secondly, SFS and SBS algorithms are combined with distance-based discrimination criterion to select the best M-dimensional feature quantity. Then, the training samples composed of the optimal M-dimensional feature quantity are input into the SVM classifier, and the optimal feature quantity and SVM classifier model are determined by comprehensively considering the two factors of economy and accuracy. Finally, the test samples are input into the SVM classifier model, and then the fault line segments shown in the results are extended to each direction of their connection respectively to obtain the boundary of the fault area. Then the suspected fault area is obtained and the location of the suspected fault area of the active distribution network based on data correlation is realized. In the end of the paper, based on PSCAD/EMTDC software, an IEEE33-node active distribution network model is established, and the simulation results verify the effectiveness of the method.