复共线性关系对逐步回归预报方程的影响研究金龙黄小燕史旭明广西气象减灾研究所,南宁,530022摘要针对气象预报中常用的逐步回归预报建模方法,由于没有直接考虑筛选出的预报因子之间可能存在复共线性关系会影响气象预报方程的预报性能问题,提出了在初选的大量气象预报因子(自变量)中,采用条件数计算分析方法,选择复共线性关系小的预报因子组合建立预报模型的方法。以重要气象灾害的预报难点——台风预报为例,用大样本分别建立了12个台风移动经度、纬度的条件数预报方程和逐步回归预报方程。对比分析结果表明,由于条件数计算分析有效控制了预报因子间的复共线性关系,因此,在相同的预报因子(自变量)和预报对象(因变量)条件下,分月建立的条件数台风移动路径预报方程,虽然历史建模样本的拟合精度略低于逐步回归预报方程,但是对独立样本的预报精度明显提高,其中7、8和9月条件数预报方程的预报误差平均为153.9km,而相应的逐步回归预报误差平均为229.2km,两者相差75.3km。进一步研究发现,在F值分别取1.0、2.0和3.0的情况下,建立的台风移动路径的逐步回归预报方程,其预报误差也明显大于条件数预报方程。另外,由于预报因子组合的复共线性的影响,逐步回归方程还出现了在个别点预报误差极大的不合理情况。关键词复共线性,气象预报,逐步回归资助课题:国家自然科学基金项目(40675023)和国家科技部社会公益性研究专项(2004CB418306)。作者简介:金龙,主要从事非线性人工智能气象预报技术研究。Email:激nlong01@1632007-07-31收稿,2007-12-29改回.中图法分类号P457.8Astudyonimpactofmulticollinearityonstepwiseregressionpredictionequation激NLongHUANGXiaoyanSHIXumingGuangxiResearchInstituteofMeteorologicalDisastersMitigation,Nanning530022,ChinaAbstractTheaccuracyoftraditionalstepwiseregressionmeteorologicalpredictionequation(SRMPE)islimitedbytheexistenceofmulticollinearityamongpredictorsoftheequation,thispaperintroducesconditionalnumberintothepredictionmodelingtominimizeitinthetraditionalSRMPE.InthepredictionmodelingofnovelSRMPE,theconditionalnumberisusedtodeterminethepredictorsetwhichhasthelowestmulticollinearityamongthepossiblesetsfromanumberofpreliminaryscreeningoutpredictors(independentvariables),andisthenusedtoconstructthenovelSRMPE.Thenovelpredictionmodelingbasedonconditionnumberisexampledwithtyphoontrackprediction,whichisawellknownnodusinmeteorologicaldisasterprediction.12typhoonstracklatitude/longitudestepwiseregressionpredictionequationshavebeenbuiltemployingboththetraditionalandnovelpredictionmodelingmethods,respectively,butusingalargenumberofidenticalsamples.Andthecomparisonandanalysisresultsindicatethatundertheconditionofsamepredictors(independentvariable)andpredictands(dependentvariables),despitethefittingaccuracyoftyphoontracksofthenovelpredictionmodeltothehistoricalmodelingsamplesisslightlylowerthanthatofthetraditionalmodel,thepredictionaccuracytotheindependentsamplesisobviouslyimproved,withanaveragedpredictionerrorofthenovelmodelfo---本文来源于网络,仅供参考,勿照抄,如有侵权请联系删除---rJuly,August,andSeptemberbeing153.9km,75.3kmsmallerthanthatofthetraditionmodel(areductionof33%),duetotheeffectivelyminimizingofmulticollinearitybythecomputationandanalysisofconditionnumberinmodeling.ItisfurthershownthatwhenF=1.0,2.0and3.0,thepredictionerrorsofthetraditionalstepwiseregressionpredictionequationsarealsoobviouslylargerthanthoseofthenovelmodel.Furthermore,theextremelylarge/unreasonableerrorsoccurredattheindividualpointsoftyphoontracksintheindependentsamplepredictionexperimentsofthetraditionalpredictionmodelduetotheimpactofthemulticollinearityinitspredictorset.KeywordsMulticollinearity,Meteorologicalprediction,Stepwiseregression---本文来源于网络,仅供参考,勿照抄,如有侵权请联系删除---