CombinedTraining ofRecurrent Neural Networks withParticle SwarmOptimization andBackpropagation Algorithms forImpedance Identification
Ganesh Kumar Venayagamoorthy, A. Corzine · 2007
A recurrent neural network(RNN)trained witha combination of particle swarm optimization (PSO)and backpropagation (BP)algorithms isproposed inthis paper. The network isusedasadynamic system modeling tool toidentify the frequency-dependent impedances ofpowerelectronic systems suchasrectifiers, inverters, andDC-DC converters. As a categoryof supervised learning methods,the various backpropagation training algorithms developed forrecurrent neural networks usegradient descent information toguidetheir search foroptimal weights solutions thatminimize theoutput errors. Whiletheyprovetobeveryrobustandeffective in training manytypes ofnetwork structures, theysuffer fromsome serious drawbacks suchasslowconvergence andbeing trapped atlocalminima.Inthispaper,a modified particle swarm optimization technique isusedin combination withthe backpropagation algorithm totraverse inamuchlarger search spacefortheoptimal solution. Thecombined methodpreserves theadvantages ofbothtechniques andavoids their drawbacks. Themethodisimplemented totrain a RNN thatsuccessfully identifies theimpedance characteristics ofathree-phase inverter system. Theperformance oftheproposed methodiscompared to thoseofbothBP andPSO whenusedseparately tosolve the problem, demonstrating itssuperiority. I.INTRODUCTION