1、PDF外文：http:/ 5120 字 EnergyandBuildings39(2007)5258 Usingintelligentdataanalysistodetectabnormalenergyconsumptioninbuildings JohnE.Seem* JohnsonControls,Inc.,507EastMichiganStreet,Milwaukee,WI53202,USA Received31October2005；receivedinrevisedform11March2006；accep
2、ted18March2006 Abstract Thispaperdescribesanovelmethodfordetectingabnormalenergyconsumptioninbuildingsbasedondailyreadingsofenergyconsumptionandpeakenergyconsumption.Themethodusesoutlierdetectiontodetermineiftheenergyconsumptionforaparticulardayissignificantlydifferentthanprevious
3、energyconsumption.Forbuildingswithabnormalenergyconsumption,theamountofvariationfromnormalisdeterminedusingrobustestimatesofthemeanandstandarddeviation.Thisnewdataanalysismethodwillreduceoperatingcostsbydetectingproblemsthatpreviouslywouldhavegoneunnoticed.Also,operatorsshouldsavetimebynothavingtoma
4、nuallydetectfaultsordiagnosefalsealarms.Thenewdataanalysismethodhassuccessfullydetectedhigh-energyconsumptioninmanybuildings.Thispaperpresentsfieldtestresultsforbuildingsthathadthefollowingproblems:(1)chillerfailureandapoorcontrolstrategy,(2)poordesignofventilatingandair-conditioningequipment,and(3)
5、improperoperationofequipmentfollowingachangeintheelectricalpanel. #2006ElsevierB.V.Allrightsreserved. Keywords:Energyconsumption；Faultdetection；Outlieranalysis；Performancemonitoring；Robuststatistics 1. Introduction Energymanagementandcontrolsystemscancollectandstoremassivequ
6、antitiesofenergyconsumptiondata.Facilityoperatorscanbeoverwhelmedwiththequantityofdata.Formanyoperators,itisnotpossibletodetectequipment,design,oroperationproblemsbecauseofdataoverload.Modernbuildingmanagementsystemshavetwosystemstohelptheoperatorswiththisdataoverload:alarmandwarningsystemsanddatavi
7、sualizationprograms.Today,operatorsmustselectthethresholds foralarmsandwarnings.Thisis adifficulttask.Ifthethresholdsaretootight,thenanumberoffalsealarmsareissued,andifthethresholdsaretooloose,thenequipmentorsystemfailurescangoundetected.Thedatavisualizationprogramscanhelpbuildingoperatorsdetectandd
8、iagnoseproblems,butalargeamountoftimecanbespentdetectingproblems.Also,theexpertiseofbuildingoperatorsvariesgreatly.Neworinexperiencedoperatorsmayhavedifficultydetectingfaultsandtheperformanceofanoperatorcanvarywiththetimeofdayordayoftheweek. *Tel.:+14145244677；fax:+14145245810. E-mailad
9、dress:. 0378-7788/$seefrontmatter#2006ElsevierB.V. Allrightsreserved.doi:10.1016/j.enbuild.2006.03.033 Theresearchcommunityhasdevelopedanumberofmethodsfordetectingfaultsinbuildingsandheating,ventilating,andair-conditioningsystems.TwomajorresearcheffortshavebeensponsoredbytheInternationalEnergy
10、Agency:Annex251,2andAnnex343.Therearetwobasicapproachestofaultdetectionanddiagnosticsinbuildings:acomponentlevel(bottom-up)approachandawhole-building(top-down)approach.Thecomponentlevelapproachlooksforfaultsinindividualsystemssuchasvariable-air-volumeboxes,air-handlingunits,chillers,orboilers.Thewho
11、le-buildingapproachlooksforunusualbehaviorinhigh-levelmeasure-mentssuchasthewhole-buildingcooling,heating,orelectricalconsumption. Claridgeetal.4describeanenergyconsumptionreportmethodthathelpsbuildingoperatorsandfacilitymanagersidentifyifthebuildingsystemsareworkingproperly.Thereportcontainsscatter
12、plotsofdailychilledwaterenergyconsumptionversusaveragedailytemperatureanddailyhotwaterconsumptionversusaveragedailytemperaturefora3-monthperiod.Forthelastmonth,thescatterplotusesletters(M,T,W,H,F,S,U)toidentifythedaysoftheweek.Thelettershelpsbuildingoperatorsidentifyoutliersinenergyconsumptionforapa
13、rticularday.Thereportalsocontainstwo-andthree-dimensionaltime seriesplotsofchilledwaterconsumptionand J.E.Seem/EnergyandBuildings39(2007)5258 53 Nomenclature a2B a isan elementofset B a2=B aisnotanelementofsetB i indexusedinforloopinFig.1n numberofelementsinsetXnout numberofoutlie
14、rsinsetX p righttailareaprobabilityfort-distribution Ri extremestudentizeddeviateforithextreme s standarddeviationforelementsinsetX srobust robustestimate ofstandarddeviationfor ele-mentsinsetX tn,pcriticalvalue(tn,p)fortheStudents t-distributionwithndegreesoffreedomandarighttailareaprob
15、abilityofp xe,i valueofithextreme xj valueofjthobservationinsetX x average of elements in set X xrobust robustestimateofaverageofelementsinsetXX setofobservations thatcontainoutliersandnon- outliers Xnon-out setofobservationsthatcontainnooutliers Xout setofobservationsthatcontainoutliers zm mo
16、difiedz-score(standardscore) setofobservationsorelements j suchthat Greekletters a probabilityofdeclaringanormalvalueanoutlier li criticalvalueforRosnersgeneralizedESDmany-outlierprocedure throughthetediousprocessofmanuallyinspectinggraphstodetectabnormalenergyconsumption.Instead,theoperatoror
17、maintenanceoperatorcaninvestigateonlybuildingswithabnormalenergyconsumption.Themethodaccountsforweeklyvariationinenergyconsumptionbygroupingdaysoftheweekwithsimilarpowerconsumption.Arobustoutlierdetectionmethodisusedtodetermineiftheenergyconsumptionissignificantlydifferentthanpreviousenergyconsumpti
18、on.Fortimeperiodswithabnormalenergyconsumption,theamountofdeviationfromnormalisdeterminedusingrobuststatistical methods. 2. Overviewofdataanalysismethod Fig.1showsthemajorstepsrequiredtoidentifyabnormalenergyconsumptioninbuildings.Thefeatureextractionblockdeterminesfeaturessuchastheaver
19、agedailyconsumptionorpeakdemandforadayfromenergydatasuchasthewhole-buildingelectricalconsumption.Thefeaturesarethensortedintogroupsbasedondaysoftheweekwithsimilarenergyconsumptionprofiles.(Inthispaper,thetermdaytypereferstodaysoftheweekwithsimilarconsumptionprofiles.)Afterthedataisgroupedbasedondayt
20、ype,outlieridentificationisusedtodeterminethefeaturesthataresignificantlydifferentfromthefeaturesforthesamedaytype.Ifanyoutliersareidentified,thenamodifiedz-score9isusedtodeterminetheamountanddirectionofvariationfromanormalobservation.(z-Scoresarealsocalledstandardscores10.)Next,detailsonarobustoutl
21、ieridentificationmethodandarobustmethodfordeterminingtheamountofvariationfromnormalarepresented. whole-buildingelectricconsumption.Byinspectingtheseplots,buildingoperatorscanidentifydaysofabnormalenergyconsumption.HaberlandAbbas5,6reviewseveralnewgraphicaldisplaysforviewingbuildingenergydata.
22、 DodierandKreider7presentamethodfordetectingwhole-buildingenergyproblemsforthefollowingenergyuses:whole-buildingtotalelectricenergy,whole-buildingtotalthermalenergy,HVAC-other-than-chillerelectricenergy,andchillerenergyusage.TheyusedanEnergyConsumptionIndex(ECI)todetermineiftheenergyconsumptionwashi
23、gherthannormal,normal,orlowerthannormal.TheECIistheratioofactualenergyconsumptiontoexpectedenergyconsumptionasdeterminedfrom aneuralnetwork. Iftheratioislargerthananupperlimit(e.g.,1.125)thenthestateofthesystemishigherthannormal.Iftheratioislowerthanalowerlimit(e.g.,0.875),then the stateofthe system
24、 is lowerthan normal. If the ratioisbetweenthelowerlimitandupperlimit,thenthestateofthesystemisnormal.GraphsoftheECIwillhelpbuildingoperatorsidentifymajorchangesinenergyconsumption.FigurespresentedbyDodierandKreider7showaweeklycycleofECI. Thispaperpresentsanintelligentdataanalysismethod8forautomatic
25、allydetectingabnormalenergyconsumptioninbuildings.Withthismethod,operatorswillnothavetogo Fig.1.Blockdiagramfordetectingabnormalenergyconsumption. 54 J.E.Seem/EnergyandBuildings39(2007)5258 Pn n i 1；p Pn 2 s 3. Outlieridentification:GESDmany-outlierprocedure Anoutlie
26、risanobservationthatappearstobeinconsistentwiththemajorityofobservations inadataset. For example,in Block1:Setnout= 0.Thisstepisusedtoinitializethenumberofoutlierstozero. Block2:Computeaverage(x)ofelementsinsetX.The averageisdeterminedfrom thedataset1,2, 1,0,3,2,101, 2,theobservation101 appearstobea
27、noutlier.Datasetsmaycontainmorethanoneoutlier.Forexampleinthedataset1,2, 1,0,3,2,101, 2, x j1xj n (1) 96,2,0, 209,theobservations101,96,and209appeartobeoutliers. BarnetandLewis11providedetailsonseveralcommonoutlieridentificationmethods.Aftercomparingseveralpopularoutlieridentificat
28、ionmethods,IglewiczandHoaglin9highlyrecommendthegeneralizedextremestudentizeddeviate(ESD) wherexj isamemberofsetXandnequalsthenumberofelementsinsetX. Block3:Computestandarddeviation(s)ofelementsinsetX. Thestandarddeviationisdeterminedfrom s j1xj x many-outlier procedure that was proposed by Rosner 1
29、2 s becauseitworkswellunderavarietyofconditions. n 1 (2) ThegeneralizedESDmany-outlierprocedurecanidentitytheelementsinasetthatareoutliers.Fig.2isaflowchartfordeterminingoneormoreoutliersfromasetofnobservationsX 2x1,x2,x3,.,xn.Theuserneedstospecifytheprobability,a,ofincorrectlydeclaringoneormoreoutl
30、ierswhennooutliersexistandanupperbound,nu,onthenumberofpotentialoutliers.Careyetal.13saidtheupperbound(nu) Block4:s=0.Thisblockchecksifthestandarddeviationof theelementsinsetXiszero.Ifthestandarddeviationequalszero,thentheelementsinsetXallhavethesame valueandtherearenooutliersintheremainingelementsi
31、nsetX.(Duringfield-testingofthismethod,severaldatasetshadastandarddeviationofzero.)TopreventadividebyzeroinBlock6,executiongoestoBlock10whenthestandard deviationdeterminedinBlock3equalszero. couldbedeterminedbyfindingthelargestintegerthatsatisfies thefollowinginequality:nu0.5(n 1).Followingaredetail
32、s Block5:Findithextreme(xe,i )insetX.Theextremeelement, onthenumberedblocksinFig.2. xe,i,istheelementinsetXthatisfurthestfromx.Ofallthe elementsinsetX,theextremeelementxe,imaximizesthe functionjxj xjwherexi isanelementofsetX. Block6:ComputeithextremestudentizeddeviateRi.The extremestudentizeddeviate
33、isdeterminedfrom xe；i xj Rij (3) whereRiisanormalizedmeasureofhowfartheithextremeisfromtheaveragevalue(x)determinedinBlock2. Block7:Computeithcriticalvalueli.Rosner12developedthefollowingequationfordeterminingthecriticalvalue: n itn i 1；p liq (4) n i1n i 1t2 wher
34、etn i 1,pistheStudentst-distributionwith(n i1)degreesandthetailareaprobabilitypisdeterminedfrom a p2n i1 (5) Fig.2.FlowchartforimplementingRosnersgeneralizedmany-outlierpro-cedure. Abramowitzand Stegun 14reviewequations forestimatingtheStu
35、dentst-distribution. Block8:Ri>li.Thisblockdeterminesiftheithextremestudentizeddeviate,Ri,determinedinBlock6isgreaterthantheithcritical value,li,determinedinBlock7. Block9:Setnout=i.Thisblocksetsthenumberofoutliers, nout,equaltoi. Block10:Removeextremeelementxe,ifromsetX.Theextremeelementxe,iisremovedfromsetXandafterremovingtheextremeelementxe,i,thenumberofelementsinSetXisni.Ifiequalsnu,thenexecutiongoestoBlock11；otherwise,returntotheforlooponi.
