<正>Neuralnetworkhastheabilitiesofself-studying,self-adapting,faulttoleranceandgeneralization.Buttherearesomedefaultsinitsbasicalgorithm,suchaslowconvergencespeed,localextremes,anduncertainnumberofimpliedlayerandimpliednotes.Thispaperpresentsasolutionforovercomingtheseshortagesfromtwoaspects.Oneistoadoptprinciplecomponentanalysistoselectstudysamplesandmakesomeofthemcontainsamplecharacteristicsasmanyaspossible,theotheristotrainthenetworkusingLevenberg-Marquardtbackwardpropagationalgorithm.Thisnewmethodwasprovedtobevalidandpracticableinsiteselectionofpracticalgarbagepowergenerationplants.
