基于混合优化的 RBF神经网络集成的降水预报模型

摘要

针对传统的单个 RBF神经网络集成中个体的隐节点个数和初始参数难以客观确定的不足，为了提高泛化能力，提出一种以高斯核函数的混合优化的 RBF神经网络的方法，首先引入正交最小二乘法动态客观的获取数据中心的个数、数据中心及权值；然后通过计算隐层中心点间最小距离作为扩展常数；最后使用剃度法调节权值、中心及扩展常数使网络参数和结构达到最优。该方法结合了正交最小二乘法和剃度算法的优点，通过从结构和算法两方面的调整提升了单个的传统的 RBF网络的性能。并将上述优化混合的 RBF神经网络与主成分分析方法相结合建立模型。本文以广西5月逐日降水事先初选的众多预报因子进行主成分分析算法提取有效的几个综合因子，然后使用混合算法优化的径向基网络建立降水预测模型。结果表明，该模型具有较好的收敛效果和泛化能力，在预报性能上明显优于同期的 T213降水预报，具有一定的普遍适用性。%　　Thispaper,aimingatthedeficienciesofbeingdifficulttoobjectivelydeterminethetraditionalsingleRBFneuralnet -workintegrationoftheindividualinthenumberofhiddennodesandinitialparameters,putsforwardakindoptimizationofRBF neuralnetworkmethodwhichisonGausskernelfunctiontoimprovethegeneralizationability.Itfirstlyintroducestheorthogonal leastsquaremethoddynamicobjectivedataacquisitioncenternumber,datacenterandweight ; thencomputestheminimum dis -tancebetweenhiddenlayercenter,whichisconsideredasanextensionofconstant ; finallyusesgradientmethodtoadjustweights, centersandextendedconstantinordertomakethenetworkparametersandstructureoptimal.Themethod,combiningtheadvan -tagesoftheorthogonalleastsquaresmethodandgradientalgorithm,promotestheperformanceofthesingletraditionalRBFnet -workfromtheadjustmentofthetwoaspectsofstructureandalgorithmandestablishesmodelbycombiningthetwomethodsof theoptimizationofmixedRBFneuralnetworkandprincipalcomponentanalysis.Itprimarilychoosesnumerousforecastfactors fromthedailyprecipitationofGuangxiinMayandprincipalcomponentanalysisalgorithmforextractingeffectiveseveralcompre -hensivefactors,thenbuildsmodelforrainfallforecastbyusingahybridoptimizationalgorithmofRBFneuralnetwork.Theresult showsthat,themodelhasagoodconvergenceandgeneralizationabilityandacertainuniversalapplicabilityinthepredictionper -formancewhichisobviouslysuperiortotheearlierT213precipitationforecast.