A Modified PSO-SVM And Application In Recognition Of GIS Partial Discharge

In recent years, Gas Insulated Substation (GIS)has a low failure rate, small footprint,insulation, and other characteristics, has been widely used in the power system. The fault typeidentification of GIS partial discharge has become a hot research, pattern recognition methodof BP algorithm is widely used, but BP algorithm is the existence of the initial weights andthresholds sensitive, easy to fall into local optimum problems. Through the analysis of thesupport vector machine (SVM) and particle swarm (PSO) principles and characteristics, thispaper presents an optimized SVM based on improved particle swarm pattern recognitionmethod and apply it to a gas insulated substation partial discharge UHF pattern recognitionfault types.Particle swarm algorithm is a group of intelligent search algorithm, with few parameters,simple principle, fast convergence, easy to implement and so on, PSO is widely used incombinatorial optimization, parameter optimization, engineering applications and otheraspects, suitable for parameter optimization of SVM. Inertia weight of PSO has a great impacton the performance of the algorithm. Currently, inertia weight of PSO algorithmimprovements mainly in two aspects, one is based on the improved fitness value, and theother is based on the number of iterations to improve. The main results of this study are asfollows:(1)Based on the inertia weight of PSO algorithm, we propose a particle convergence andco-regulation based on the number of iterations of nonlinear adaptive particle swarm weightPSO (MPSO).(2)By experimental analysis optimization precision on four typical test functions,obtained MPSO parameter ranges, and with improved particle swarm optimization (PSO1)based on particle evolution degree of polymerization, based on the dynamic fitness valueadjust the weights PSO (PSO2) optimizing the performance of comparative analysis to verifythe effectiveness of the improved particle swarm algorithm. (3)By compared performance of MPSO-SVM, GA-SVM, PSO-SVM algorithm on UCIdatasets,the results of MPSO-SVM is best.(4)Through analysis of the statistical characteristics of the four typical types of GISpartial discharge, principal component analysis of the statistical characteristics ofdimensionality reduction method, the algorithms of PSO1-SVM, PSO2-SVM, MPSO-SVM,BP on GIS partial discharge pattern recognition, experiments the results show thatMPSO-SVM can achieve better recognition rate.