Improved Orthogonal Genetic Algorithm Based On Support Vector Machine And Application In Recognition Of Gas Insulated Substation Partial Discharge

Genetic algorithm(GA) has been used in many fields as an optimal algorithm. However, there still exist some problems in basic GA such as the search is too blind, easy to fall into local optimum and slow convergence speed and so on. This paper takes advantage of orthogonal experimental design to design the initial population and crossover of GA, and use chaotic map to produce stochastic disturbance among a part of the best chromosomes so as to improve the ability of local search, obtains an improved orthogonal GA called OCGA. At last, we compared and analyzed the OCGA, basic orthogonal GA(denote by OGA/Q) and classical GA(denote by Simple GA) through the optimal experiments of the typical test functions, Calculated the optimal value, mean value, variance and the average times of convergence of the function value(precision <=1.0e-5) of 10 times experiments that on the three algorithms in each test function. Gained the experiments results as below:(1)In the whole index of every test function, OCGA and OGA/Q are better than Simple GA and improved by several or even more than 1.0e+20 times; it is especially obvious that at multimodal functions, variance and the average times of convergence.(2) OCGA could converge on 4 test functions except Schwefel function, however, OGA/Q can only converge on 2 test functions, it indicates that OCGA is better than OGA/Q on the expansibility. There is great progress on the performance of the other 2 multimodal test functions especially at the aspects of variance and the average times of convergence. The average frequency of convergence of OCGA is OGA/Q’s ten times, the mean value and variance are reduced 5 and 10 orders of magnitude than OGA/Q respectively(decimalism).Support vector machine(SVM) is a widespread applied learning machine, but the choice of the related parameters values has no mature theory to support. On the other hand, feature selection plays an important role in the classification performance of SVM. Therefore, this paper proposes a method based on improved GA for feature selection and the parameters optimization of SVM(denoted by OCGA-SVM), it make use of the global optimization ability of GA to optimize the parameters of SVM and to select suitable features at the same time. After that, we has carried on the classification experiments on two UCI data sets, and the results shows that the classification accuracy of the new algorithm compared with basic SVM was improved by about 7.25 and 33.6 percent point respectively. The number of features has been decreased by more than 35%; these indicate that the new algorithm is obviously better than the old algorithm.Gas insulated substation(GIS for short) is one of the large important equipment of power system. Because of the merit of small occupation area and high operation reliability, GIS has been largely applied in electric system. It plays a vital role in safety running of power system. Hence, it will be of great importance in detection and to assess the internal defects of GIS before the breakdown in GIS. This paper mainly studied the four types of typical defects of GIS partial discharge, and extracts the statistical parameter features of every typical defects of partial discharge, lays the foundation for late pattern recognition of GIS partial discharge. At last, in this paper, the improved OCGA-SVM is applied to the GIS partial discharge pattern recognition, and the experiments of OCGA-SVM and the SVM of no parameters optimizing have been conducted, and after that the results have been analyzed and compared. As a conclusion, the classification accuracy of the new algorithm compared with basic SVM was improved by 5 percent point, and the number of features has been reduced by 30%; we verified the effectiveness of the algorithm of this paper.