Improved Genetic Simulated Annealing Algorithm Based On BP Neural Networks And Application In Recognition Of GIS Partial Discharge

Genetic algorithm could be well applied to function optimization problems, the choice ofcrossover rate and mutation rate will produce a significant impact on the efficiency andvalidity of it. Generally accepted range of crossover rate Pc is between0.4and0.99, andmutation rate Pm is between0.0001and0.1, but no one has made a systematic, in-depth studyon these ranges of science and rationality. In this paper, by genetic algorithm to solve theglobal optimal solution required minimum average number of iterations as a goal, we mainlymake a systematic study on the choice of the crossover rate and mutation rate for a class ofperiodic function y=A sin(ω x+). The conclusions as follows:(1) After a large number of experiments and analysis, we have obtained the optimalrange of mutation rate Pm is0.005to0.026from using genetic algorithm to solve the globaloptimal solution required minimum average number of iterations;(2) By lots of tests and variance analysis, we have proved, no matter what the mutationrate value we take within the range of0.005to0.026, the crossover rate makes no significantdifference to obtain the global optimal solution required minimum average number ofiterations by using genetic algorithm;(3) As derived from a large number of experimental data and statistical analysis, thecompare of the mutation rate Pm value within the range of0.005to0.026and outside thisinterval value, the global optimal solution required the average number of iterations of theformer reduce more than70%than the latter. The compare of the mutation rate Pm valuewithin the range of0.005to0.026and in previous recommended value within the range of0.0001to0.1, it reduces more than60%.Since the1990s, the pattern recognition method was applied to the identification of thetype of partial discharge (PD). In recent years, artificial neural networks as a commonly usedpattern recognition methods become the research focus of the PD pattern recognition, themethod greatly improves the reliability and practicality of the identification.Back-propagation (BP) neural network is currently used most widely, however, due to thereare several disadvantages of BP including sensitive to the initial weights and threshold, possibility of being trapped at locally minimum value. This article proposes genetic simulatedannealing (GSA) algorithm to optimize back propagation neural network (BPNN) via adetailed analysis of principles and characteristics of the genetic algorithm, BP neural networkand simulated annealing algorithm. And it was applied to the Gas Insulated Substation (GIS)partial discharge pattern recognition, the result of identify four type of typical GIS defectsshow that, compared with BP neural network, this model can achieve better classificationaccuracy.