TheSmartThermostat:UsingOccupancySensorstoSaveEnergyinHomesJiakangLu - PDF document

TheSmartThermostat:UsingOccupancySensorstoSaveEnergyinHomesJiakangLu†,TamimSookoor†,VijaySrinivasan†,GeGao†,BrianHolben†,JohnStankovic†,EricField‡,KaminWhitehouse††DepartmentofComputerScience,UniversityofVirginia‡SchoolofArchitecture,UniversityofVirginiaAbstractHeating,ventilationandcooling(HVAC)isthelargestsourceofresidentialenergyconsumption.Inthispaper,wedemonstratehowtousecheapandsimplesensingtechnol-ogytoautomaticallysenseoccupancyandsleeppatternsina 9AM 10AM 11AM 12PM 1PM 0 5 10 15 20 Power (kW) 55 60 65 70 75 Temperature (F) Temperature Setpoint Power Home Away Shallow setback wasteComfort missVacant house waste (a)ProgrammableThermostatOperation 9AM 10AM 11AM 12PM 1PM 0 5 10 15 20 Power (kW) 55 60 65 70 75 Temperature (F) Slow reaction wasteShallow setback wasteReaction waste (b)ReactiveThermostatOperationFigure1.Bothprogrammableandreactivethermostatscausesubstantialenergywasteanddiscomfort.frequentequipmentcycling.Furthermore,alongertime-outperiodisnotanadequatesolutionbecauseitwouldwasteenergybyconditioningunoccupiedspaces;thesmartther-mostatrequiresoccupancymonitoringthatisbothquickandreliable.Toaddressthisproblem,weuseanovelalgorithmthatanalyzespatternsinthesensordatatoquicklyrecognizeleaveandsleepevents,allowingthesystemtorespondwithinminuteswithoutincreasingfalsedetectionrates.ThesecondmainchallengeofthisaproachistodecidewhentoturntheHVACsystembackon.Preheatingthehousecouldwasteenergyifthesystemisactivatedtooearly.Ontheotherhand,heatingonlyinresponsetooccupantar-rivalcouldalsowasteenergybecause,atthatpoint,thehousemustbeheatedveryquickly;manymulti-stageHVACsys-temshaveahighlyefcientheatpumpthatcanbeusedforslowlypreheating,butalowerefciencyfurnaceorelectricheatingcoilsmustbeusedtoheatthehousequickly.Sincethesmartthermostatcannotpredictexactlywhenoccupantswillarrive,itisdifculttodecidewhichapproachwillbemoreefcientonanygivenday.Instead,thesystemusesahybridapproachthatminimizesthelong-termexpecteden-ergyusagebasedontheoccupancypatternsofthehouse:itslowlypreheatsthehousewithhighefciencyequipmentatatimeand,iftheoccupantsreturnbeforethattime,itquicklyrespondsbyheatingthehomewiththeloweref-ciencyequipment.Thetimeischosenbasedontheequip-mentefcienciesandthehistoricaldistributionofoccupantarrivals,balancingtheexpectedcostsofpreheatingtooearlyandpreheatingtoolate.Toevaluateourapproach,wedeploysensorsin8homesusingX10motionanddoorsensorsthatcostabout$5each[12]andcanbeeasilyinstalledwithdouble-sidedtape.Wealsocollectempiricalmeasurementsofthetemperatureresponseandenergyconsumptionofahomewithatypi-calheatingsystem.WeconstructasimulationmodelofthishomeusingtheEnergyPlushomeenergysimulationframe-work[13]andvalidatethattheenergypredictionsmatchourempiricalenergymeasurementsofthehome.Then,usingthismodel,wecalculatetheenergycostofheatingeachofthese8householdsusingthesmartthermostatalgorithmanddemonstratea28%energysavingusing12-20sensorsperhome,foratotalcostoflessthan$100.Ouranalysisshowsthatsimilarresultswouldbeachievedwithasfewas3-5sen-sorsperhome,acostofabout$25.Forcomparison,weusethesamehomemodelandweathertracestoalsoevaluateareactivealgorithmthatturnsthesystemoninresponsetomotionsensorordoorsensorvalues,andturnsthesystemoffinresponsetoaperiodofinactivity.Thisapproachiscommonlyusedbyoccupancy-basedlightingsystemsandhasrecentlybeenadoptedbyoff-the-shelfthermostatsthatclaimtosaveenergybyrespondingtooccupancy[14,15].However,ourstudiesshowthatwithoutoursensoranalysisandcontrolalgorithms,thisapproachonlyachievesa6.8%energysavingonaverageinthese8homes.Infact,itac-tuallyincreasesenergyusageinfourofthehomesduetoitsinabilitytorespondquicklytooccupants,asexplainedinSection2.2BackgroundandRelatedWorkProgrammablethermostatshavebeenapillarofenergyconservationprogramssinceshortlyaftertheirinventionin1906,over100yearsago.ThebasicideaistocontroltheHVACequipmentbasedonasetbackschedule:thehouseisconditionedtoasetpointtemperaturewhentheoccupantsaretypicallyactiveandoatstoamoreenergy-efcientset-backtemperaturewhentheoccupantsaretypicallyawayorasleep.Thenotionthatenergycouldbesavedinthisman-nerhasbeenpartoftheU.S.collectiveconsciousnesssincePresidentCarterfamouslydonnedacardiganandturnedthetemperatureoftheWhiteHousedownto55°Fatnightduetotheenergycrisisofthe1970s.However,thisapproachwastesenergyinseveralways,asillustratedbyFigure1(a).First,theoccupantsleavethehomeshortlyafter9AM,butthesystemwastesenergybecauseitisscheduledtocontinueheatingthehomeuntil10AM(leftside).Second,theset-backtemperatureiswellabovethesafetylimitforthehouse,causingenergyconsumptionevenwhilethehouseisvacant(center).Thistypeofshallowsetbackistypicallyusedtoreducetheriskofcomfortloss,incasethebuildingbecomesoccupiedatthattime.Third,theoccupantsbecomeuncom-fortablewhentheyreturnshortlyafter1PMbecausethesys- 9AM 10AM 11AM 12PM 1PM 0 5 10 15 20 Power (kW) 55 60 65 70 75 Temperature (F) Temperature Setpoint Power Home Away Fast reaction Pre-heating Deep setback Figure2.ThegoalofthesmartthermostatistoautomaticallyturnofftheHVACequipmentassoonastheoccupantsleave,useadeepsetbacktemperaturewhiletheyaregone,andpreheatimmediatelybeforetheoccupantsreturn.temisnotscheduledtoheatthehouseuntilmuchlater.Onthesurface,thislastproblemappearstobeonlyacomfortissue,butisinfactanimportantcauseofenergywaste:thestaticsetbackschedulesusedbyprogrammablethermostatscannotcapturethehighlydynamicoccupancypatternsofmosthomesandwillinevitablycausesomelossofcomfort.Thisriskofcomfortlosscausespeopletoreducetheiruseofsetbackschedules,orstopusingthemaltogether.Over50%ofhouseholdsthathaveprogrammablethermostatsarereportedtonotusesetbackperiodsatnightorduringtheday[5].Incontrast,householdswiththesimplerdial-typethermostatscaneasilyadjusttemperaturesettingsbeforego-ingtosleeporleavingthehouseand,asaresult,actuallysavemoreenergyonaveragethanuserswithprogrammablethermostats[5,6].Inthepreliminarywork,twooftheau-thorsdesignedandevaluatedtheself-programmingthermo-stattoxthisproblembyautomaticallychoosingtheoptimalsetbackschedulebasedonhistoricaloccupancydata[16].However,thatsystemstillproducesastaticscheduleand,sinceoccupancypatternschangeeveryday,anystaticsched-ulemustsacriceeitherenergyorcomfort.Inthispaper,weusereal-timesensordatatodynamicallycontroltheHVACsystemastheoccupancystatusofthehousechanges.Analternativetotheprogrammablethermostatisthere-activethermostat,whichusesmotionsensors,doorsensors,orcardkeyaccesssystemstoturntheHVACequipmentonandoffbasedonoccupancy[17,15,18,19].However,ourpreliminarystudiesfoundthatreactivethermostatssavelessenergythanprogrammablethermostatsinresidentialbuild-ings,andin4outof8householdsactuallyincreaseen-ergyusagebyupto10%[16].Weidentiedthreesourcesofenergywaste,whichareillustratedbyFigure1(b),col-lectedfromahomeusingtheBAYwebbrandreactivether-mostat[14]withvemotionanddoorsensors.First,theoccupantsleavethehouseatabout9:30AM,butthesystemwaitsuntil10:30AMtostopheating(leftside).Thislongdelayisusedtoreducetheriskofturningofftheheatwhilethebuildingisstilloccupied(acommonproblemwithlight-ingsystemsthatusemotionsensors).Second,evenwhenthesystemisfairlycondentthatthebuildingisunoccupied,itstillwastesenergybymaintainingatemperaturethatiswellabovethebuildingsafetylevel(middle),inordertoreducethebuildingresponsetimeoncetheoccupantsreturn.Third,whentheoccupantsdoarriveshortlyafter1PM,thesystemmustwasteenergybyusinganinefcientstageofheating:itrstrespondswithanenergy-efcientheatpump,butafterdetectingthatthetemperatureisrisingtooslowlyitswitchestoaveryinefcientauxiliaryheatertoraisethetemperaturemorequickly.Thissamephenomenonhasbeenobservedinpreviousstudiesofprogrammablethermostats[20,21].Insummary,theenergysavingpotentialofreactivether-mostatsislimitedbytheirinabilitytorespondquicklytobuildingoccupants.Thesmartthermostatpresentedinthispaperaddressesthislimitationbydevelopingnewalgorithmstoquicklyturnoffthesystemwhennotneeded,andtoturnonthesystematatimethatminimizeslong-termexpectedenergyconsumptionbasedonoccupancypatterns.3TheSmartThermostatThesmartthermostatusesoccupancysensorstosaveen-ergybyautomaticallyturningofftheHVACwhenoccupantsaresleepingoraway.Thesystemusescheap,simplemo-tionanddoorsensorsinstalledthroughoutthehome(Sec-tion3.1).Basedonthesesensors,thesystememploysthreeenergysavingtechniques,asillustratedinFigure2.First,thefastreactionalgorithmusesaprobabilisticmodeltopro-cessthesensordataandquicklyestimatewhetheroccupantsareactive,sleeping,oraway(Section3.2).Thisalgorithmcantypicallyrespondwithinminutesoftheoccupantsleav-ingthehouse,withoutintroducingfalsevacancydetections.Second,thesystemcombineshistoricaloccupancypatternswithon-linesensordatatodecidewhethertopreheatthehomeortoheataftertheoccupantsarrive(Section3.3).Fi-nally,thesystemsavesadditionalenergybyallowingthetemperaturetodriftfurtherfromthesetpointtemperaturewhenitiscondentthatthehomeisunoccupied.Wecallthisadeepsetback(Section3.4).Thesethreetechniquesallowthesystemtoautomaticallysaveenergywithoutsacricingoccupantcomfort.3.1InstrumentingtheHomeInordertorespondtotheresidents,thesmartthermostatrequirestwotypesofsensorstoidentifywhenoccupantsareinthehomeandwhentheyaresleeping:passiveinfrared(PIR)motionsensorsinroomsandmagneticreedswitchesonentryways.PIRsensorsandmagneticreedswitchesarecheapandeasytoinstall:wedeployoff-the-shelfwirelessX10sensors[12]asshowninFigure3thatcanbepurchasedforapproximately$5eachandeasilyinstalledbyattachingthemtothewallordoorusingdouble-sidedtape.Incon-trasttoothersmarthomeapplicationssuchasmedicalmon-itoringandsecurity,thedomainofenergyconservationcantolerateasmalllossinaccuracyinfavorofcostandeaseofuse.Therefore,thesmartthermostatdoesnotrequirecam-erasorwearabletagsthatmaybeconsideredintrusivetotheuser[22,23]ormoresophisticatedsensingsystemsusedforne-grainedtrackingandactivityrecognition[24,25,26].Forourexperiments,wedeployedmoresensorsthanwe (a)MotionSensor (b)DoorSensorFigure3.Thesmartthermostatusesmotionsensors(left)andcontactswitchesondoors(right).thoughtnecessaryinordertoanalyzethesensitivityofourapproachtothenumberofsensors.Weinstalledatleastonemotionsensorineveryroomandamagneticreedswitchonexteriordoorwaytothehome.Thistypeofdeploymentwouldrequirelessthan20minutesand,at$5persensor,wouldcostlessthan$100formosthomesandabout$50foranaveragehomewith9roomsandonemainentrance.Thisissimilartothecostofpurchasingandinstallingatypicalprogrammablethermostat,andapproximately35%ofhomesintheU.S.alreadycontainasimilarsetofsensorsaspartofahomesecuritysystem[27].Furthermore,ouranalysisinSection6.3.1showsthatonly3-5strategicallyplacedsen-sorsareactuallyneededineachhometoachieveenergysav-ings.Analternativetomotionsensorswouldbetosensethehome'selectricalandmechanicalsystemstodetectoc-cupancy[28,29,30],althoughtheeffectofthesesystemsoncostandtheabilitytoquicklyandreliablydetecthomeoccupancyhasnotbeendemonstrated.3.2TurningtheHVACSystemOffOnekeychallengeofthesmartthermostatistodecidewhentheoccupantshaveleftthehomesothatitcanturnofftheHVACsystem.Beingtooaggressivecancausetheequipmenttoshutofftooearly,causingoccupantdiscomfort,wastingenergyduetorapidequipmentcycling,andshorten-ingthelifeoftheequipment.Ontheotherhand,beingtooconservativecanwasteenergybyconditioningunoccupiedspaces.Inordertoachieveanebalance,thesmartther-mostatusesaHiddenMarkovModel(HMM)toestimatetheprobabilityofthehomebeingineachofthreestates:(i)Awaywhenthehomeisunoccupied,(ii)Activewhenthehomeisoccupiedandatleastoneresidentisawake,and(iii)Sleepwhenalltheresidentsinthehomearesleeping.Oncethesystemdetectsastatetransitionwithhighprobability,itrespondsbyswitchingthetemperaturesetpointappropri-ately.TheHMMisdepictedinFigure4(b).Thehiddenvariable(yt)isadistributionoverthehomestate:Away,ActiveandSleepandtheHMMtransitionstoanewstateeveryvemin-utes.Theobservedvariablesxtareavectorofthreefeaturesofthesensordata:(i)thetimeofdayat4-hourgranularity,(ii)thetotalnumberofsensorringsinthetimeintervaldT,and(iii)binaryfeaturestoindicatepresenceoffrontdoor,bedroom,bathroom,kitchen,andlivingroomsensorringsinthetimeintervaldT.TherstfeaturehelpstheHMMusehistoricaloccupancyateachtimeofdaytohelpestimatecur-rentoccupancy.Thesecondfeaturesimplyindicateswhethertheoccupantsarehighlyactive.Thethirdfeaturehelpsde- thresholdlast ! lastthresholdlast " thresholdlast (a)ReactiveStateMachine 1 ty1 txtx1!tx ty 1!ty = Sensor Features = Occupant Activities (b)HMMFigure4.ReactivethermostatsuseastatemachinetoswitchbetweenstatesbasedonTlast,thetimeelapsedsincethelastsensorring.ThesmartthermostatusesaHMMtoswitchbetweenstatesbasedonaprobabilisticmodelofstatetransitionsandsensordata.tectwhethertheoccupantshaveopenedorclosedadoorre-cently,andalsohelpslteroutmotionsensorswithhighfalsepositives,e.g.thosethatarenearawindow.TotraintheHMMusingadatatracefromahomewithknownoccupancystates,thesefeaturesarerstcalculatedeveryveminutes.TheMarkovstatetransitionprobabilitiesP(ytjyt1)andtheemissionprobabilitiesfortheobservedsensorfeaturesP(xtjyt)arerepresentedusingadiscretecon-ditionalprobabilitytableandarebothcalculatedusingfre-quencycounting.However,frequencycountingresultsinverylowprobabilitiesforseveralvaluesoffeature(ii)be-causethetotalnumberofringspertimeunithasalargerdomainthanfeatures(i)and(iii).Therefore,webuildagen-erativeGaussianmodelforP(xtjyt)tosmooththeprobabilitydistributionforfeature(ii).Additionally,weexplicitlydis-ablesleepstatesinthemorningaftercontiguoushoursofsleepstatesaredetectedduringthenight.Thiseffectivelyencodesthatapersonhasrecentlywokenupandisunlikelytosleepagain,whichimprovestheaccuracyofourHMMbycorrectlyclassifyingidleperiodsafterapersonhaswokenupasawaytimes.3.3TurningtheHVACSystemOnSincetheoccupantarrivaltimesarenotknown,akeychallengeofthesmartthermostatistodecidewhetherandwhentopreheatthehouse.Preheatingtooearlycanwasteenergybymaintainingthesetpointfortoolong,whilepre-heatingtoolatecanwasteenergybyincreasingthechanceofneedingtoreacttooccupantarrivalswithaninefcientheatingstage,asdescribedinSection2.Inordertomanagethisdelicatetrade-off,thesmartthermostatchoosestheop-timalpreheattimethatminimizesthelong-termexpectedenergyusage.Itslowlypreheatsthehousewithhighef-ciencyequipmentatatimeand,iftheoccupantsreturnbeforethattime,ituseshighercapacitybutlowerefciencyequipmentinordertoquicklyheatthehome.Twostepsarerequiredtoderivethevalueof:(i)characterizethecapacityandefciencyforeachstageofthehome'sHVACsystem,and(ii)analyzehistoricaloccupancypatternsofthehome.Weempiricallymeasuretheefciencyofathree-stageHVACsystemthatincludesa2-stageheatpumpandathirdstageelectricheaterinthehouseshowninFigure9.Foreachstage,wepreformfourexperimentsbyturningoffthesystem,whichallowsthehousetocooldowntobelow65°F,andthenheatingthehousetoatargettemperature.Bycor- Stage 1 Stage 2 Stage 3 0 0.36 0.71 1.07 1.42 1.78 Average energy to heat 1 degree Fahrenheit (kW h) HVAC Stage 0 5.28 10.56 15.84 21.12 26.40Average time to heat 1 degree Fahrenheit (min) Figure5.EnergyefciencyandlagtimevaryamongthemultiplestagesofHVAC.Thesmartthermostatusesthemostenergyefcientstageforpreheatinginordertore-ducethereactionenergywaste.relatingthethermostatoperationlogsandpowermeasure-ments,wecalculatetheaverageenergyusedandthetimetakenbyeachstagetoraisethetemperatureby1°Fonaver-age,asdepictedinFigure5.TheresultsshowthatStage2isthemostenergyefcientstage,buthasalongresponsetime;at15minutesperdegree,thisstagewouldrequire2hourstorecoverfromatypical8degreesetbacktemperature,whichwouldbetoolongforanoccupiedbuilding.Ontheotherhand,Stage3hasthefastestresponsetimebutaveryhighenergycost.NotethatforthisequipmentStage1usesthesamecompressorandfanasStage2,butoperatesatalowerpowerlevelandspeed.Itislessefcientbutmoreeffectiveatmaintainingaconstanttemperature.Weusethesemeasurementstochoosetheoptimalpreheattimegivenasetaofobservedarrivaltimesatthehome.Forallpossibletargetpreheattimest:min(a)tmax(a),wecalculatetheexpectedenergycostbyaveragingthewasteforeacharrivaltimea2a.ThewasteforarrivaltimeaisdenedtobetheenergyrequiredtoheatwithStage3ifat.Other-wise,ifat,thewasteisdenedtobetheenergyrequiredtopreheatwithStage2andthenmaintainthetemperatureusingStage1fortimeat,untiltheoccupantsarrive.Oncetheexpectedenergycostsarecalculated,wesettobethetimetwiththelowestexpectedenergycost.Toillustratethisoptimizationprocess,wecalculatedtheoptimalpreheattimegiventheempiricalarrivaltimesfoundinthepubliclyavailableTulumdataset(excludingweek-ends),whichwascreatedbymonitoringtheoccupantsofahomeforapproximatelyonemonth[31].Figure6showstheexpectedenergycostforallpreheattimesbetween4:45PMand6:45PM.Thisgureillustratesthat,ifthesystempre-heatstooearly(leftside),itwastesenergyduetomaintainingahighsetpointtemperaturetoolong.Ifthesystempreheatstoolate(rightside),itwastesenergybecauseitmustreactwiththeinefcientbutfast-reactingStage3heatingsystem,ifoccupantsarrivebeforepreheatingiscomplete.Thesys-temcanachievetheminimumenergyusagebychoosinga 16:52 17:18 17:43 18:09 18:35 0 0.5 1 1.5 2 2.5 3 Target preheat timeExpected Energy Usage (kWh) Maintain(Stage1) Preheat(Stage2) React(Stage3) Figure6.Thesmartthermostatselectsthetargetpreheattimethatoptimizesthelong-termexpectedenergyusage. Setback period (min)Temperature difference between setback and setpoint (°F) 0 30 60 90 120 150 180 210 240 270 300 330 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 0 10 20 30 40 50 60 Figure7.Adeepersetbackdegreehasalargerimpactonenergysavingsthanalongersetbackperiod.targetpreheatingtimeof6:04PM.Preheatingtothistimerequires9%lessenergyonaveragethanapurelyreactivethermostat.Itisworthnotingthatoptimalpreheattimeistypicallynotthesameastheexpectedarrivaltimeoftheoc-cupant;itchangesbasedontheefciencylevelsoftheHVACequipment.3.4UsingDeepSetbacksThetypicalsetbacktemperatureis8degreesfromtheset-point,whichiswellabovethesafetylimitforahouseandcausesenergyconsumptionevenwhenthehouseisvacant.Shallowsetbacktemperaturesaretypicallyusedtoreduceriskofcomfortloss,incasethebuildingbecomesoccupiedatthattime.Becausethesmartthermostatrespondstooccu-pancyevents,itcanincreaseenergysavingsbyusingdeepersetbacksduringperiodswhenthebuildingisunoccupiedandoccupantsarehighlyunlikelytoreturn.Specically,givenahistoricalsetofarrivaltimesa,thesmartthermostatusesadeepsetbackassoonasthebuildingisdetectedtobeunoccu-pied,andswitchestoatypicalshallowsetbackattheearliest Figure9.Forrealisticenergycalculationsandpredic-tions,wecreatedandempiricallyvalidatedanenergymodelofthehouseandHVACsystemshownhere.previouslyobservedarrivaltimemin(a),whichreducesthetimerequiredtorecoveracomfortabletemperatureoncetheoccupantsreturn.Figure7illustratesthatdeepersetbacktemperatureshavealargerimpactonenergysavingsthanlongersetbackperi-ods:avedegreeincreaseinsetbacktemperatureforanhourhasthesameeffectasanadditionalvehoursofsetbacktimethatusesthenormalsetbacktemperature,eveninamoderateclimatelikeWashington,D.C..Sincethesmartthermostateitherpreheatsthehomeorquicklyrespondstooccupantar-rivals,itcanexploitthelargeenergysavingsmadepossiblebydeepsetbackswithoutsacricingoccupantcomfort.Toillustratetheconceptofadeepsetback,Figure8showsthedistributionsofleaveandreturntimesforthepubliclyavailableKasteren[25]andTulum[31]homemonitoringdatasets,excludingweekends.TheindividualintheTu-lumstudyisconsistentlyawayfromhomeforalongerpe-riodoftime,andwillthereforebenetmorefromdeepset-backs.Thelengthofadeepsetbackdependsonthemini-mumarrivaltimeofahousehold,andsotheindividualintheKasterendatasetdoesnotobtainalargerbenetfromdeepsetbackseventhoughhe/shesometimesreturnsverylateintheevening.4ExperimentalSetupInthissection,wedescribethedatacollectionprocessandthesimulationframeworkusedtoevaluatethesmartthermo-stat.4.1CollectingOccupancyDataOccupancypatternsplayasignicantroleintheperfor-manceofthesmartthermostat.Toinvestigatetheimpactofoccupancypatternsontheperformanceofsmartthermostat,weuseoccupancypatternscollectedthroughthreemeans:(i)theempiricaldatatracesfrom8instrumentedhomes,(ii)theoccupantsurveysof41homes,and(iii)twopublicsmarthomedatasets.Weusetheempiricalsensordatatoevaluateallthreephasesofsmartthermostat,butcannotevaluatefastreac-tionusingthedatacollectedthroughtheothertwosources #Residents #Rooms #Motion #Doors #Door #Weeks Sensors Sensors 1 7 7 5 3 2 1 3 3 3 2 1 1 4 4 3 1 1 1 5 4 3 1 1 2 5 5 3 1 2 3 5 5 4 2 1 3 4 4 3 1 1 2 5 5 4 2 1 Table1.Detailsofthe8homesusedindeploymentsbecausetheylackthesensordatarequiredforfastreaction.Therefore,theyareusedtoevaluateonlythedeepsetbackandpreheating.4.1.1SensorDeploymentsWedeployX10motionsensorsanddoorsensorsin8homestocollectoccupancyandsleepinformation.Thesehomesincludebothsingle-personandmulti-personresi-dences,andthepeoplelivinginthehomeincludestudents,professionalsandhomemakers.Forexample,onehomein-cludesagraduatestudentcouplealongwithanelderlyres-ident,twootherhomesincludeyoungworkingprofession-als,andanotherhomeincludesthreegraduatestudents.Thedurationofthesensordeploymentsvariesfromonetotwoweeks.Ingeneral,wedeployonemotionsensorineachroomandonedoorsensoroneachentrywaytothehome,andsomeinnerdoors.However,wedonotinstrumentroomsorentrywaysthatareveryinfrequentlyused.Table1summa-rizestheinformationaboutthehomes.Wecollectgroundtruthusingamanualpost-processingofthedataanddailyinterviewswiththeresidentstoclar-ifyambiguousorquestionabledata.Thegroundtruthvaluesusedforthisstudyarebestestimatesbylabelinguseractiv-itiesmanually,butarenotexpectedtobeperfectlyaccurate.Groundtruthinhomemonitoringexperimentsisverydif-culttocollect,andpreviousstudieshaveusedawidevari-etyofapproachesrangingfromselfreportstovideocamerarecordingstohavingaproctorphysicallyonsitetomonitorhomeactivities[32,25,33].Noneoftheseschemesforcre-atinggroundtruthareexpectedtobeperfect.4.1.2SurveysandDataCollectionToaugmentthehomedeployments,wecollectdatafromanother41householdsforfourweeksusingsurveys:eachindividualisinstructedtowritedowntheirsleep,wake,leave,andarrivetimeseveryday,andthedataarecollectedthroughperiodictelephonecalls.Theperiodofthetelephonecallsrangesfromonceperdaytoonceperweek.Thesur-veyedindividualsrangefromstudentstoprofessionalstore-tirees.Thehouseholdscompriseavarietyofsingle-personandmulti-personresidencesfromvariouspartsoftheeast-erncoastinU.S..Thetimescollectedthroughthesesurveysareexpectedtobeprecisewithin15minutes,sincemanyresidentsreporttimesin15-minuteintervals.Overall,theoccupantscanbecategorizedintovedifferentlifestyles:re-tirees,students,professionals,youngprofessionalsandfam-ilies. 6AM 9AM 12PM 3PM 6PM 9PM 12AM 0 1 2 3 4 5 Time of dayNumber of occurrences Leave event Arrive event Deep Setback (a)ShortDeepSetback(Kasteren) 6AM 9AM 12PM 3PM 6PM 9PM 12AM 0 5 10 15 20 Time of dayNumber of occurrences Leave event Arrive event Deep Setback (b)LongDeepSetback(Tulum)Figure8.Thedurationofdeepsetbackvariesamongdifferentpeople,anddependsonlyontheearliestobservedarrivaltime;latearrivalsdonothaveadditionalbenetfromdeepsetbacks.Inadditiontosurveys,weanalyzetheleave,return,wake,andsleeptimesfromtwopublicly-availabledatasetsthatcontainhomeoccupancyinformationfortwoindividualsoverthecourseofapproximatelyonemontheach.Thesedatasetsarecollectedbymanuallylabelingactivitiessuchassleeping,eating,andbathing,andleavinghome.WecallthesetheKasteren[25]andTulum[31]datasets,respec-tively.Forthepurposesofthisstudy,weonlyusetheleavehome,arrivehome,andsleepeventlabels.4.2SimulationFrameworkThepracticalperformanceofthesmartthermostatcanbeaffectedbymanyfactorssuchasoutdoortemperature,airleakageandhouseinsulation.Therefore,itisimpor-tanttoevaluatethesmartthermostatwithvariousclimatesandbuildingconditions.However,large-scaleexperimentsareextremelydifcultduetoresourceconstraints.Toad-dressthisproblem,wehavemodeledthehomeinFigure9andvalidatedthemodelbycomparingempiricalenergymea-surementsandenergypredictionsgeneratedusingtheEner-gyPlussimulator.Thisvalidatedmodelallowsustoevaluatethesmartthermostatundervariousconditions,suchasdif-ferentclimates,thatcannotbeeasilydoneempirically.4.2.1EnergyPlusSimulatorInourexperiments,weusewhole-housethermalsimu-lationmodelingprovidedbytheU.S.DepartmentofEn-ergy'sEnergyPlussimulatorasaframeworktoevaluatedif-ferentthermostatalgorithmsunderdifferenthousingcondi-tionsandclimates.EnergyPlusisdevelopedanddistributedbytheU.S.DepartmentofEnergy'sEnergyEfciencyandRenewableEnergydivision,derivedfromandextendingtheearlierDOE-2andBLASTsimulators.IthaswonawardsforR&D,TechnologyTransferandTechnicalExcellence,andiswidelyregardedasthepremierbaselineenergyperformancesimulationtoolintheindustry.Inthesimulation,amodelisdescribedwhichinte-gratesthephysicaldescriptionofabuilding(includingwalls,oors,roofs,windowsanddoors,eachwithassociatedcon-structionpropertiessuchasR-Valueofmaterialsused,sizeofwalls,locationandtypeofwindows)withthedescrip-tionsofmechanicalequipment(heatingandcooling),me-chanicalventilationschedules,occupancyschedules,otherhouseholdequipment,andsoon.Thesimulationapplies 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 HVAC Simulation Results (kWh)HVAC Empirical Results (kWh) Daily Measurement Linear Regression Figure10.Theobservedenergyusagecloselymatchesthevaluespredictedbyourmodel.thismodeltoatime-seriesthermalcalculationusingwell-knownthermaltransferequationsandaggregateclimateandweatherdatafromlocalairportsandweatherstations.Thecalculationsoutputinteriorandexteriorairtemperatures,en-ergyconsumption,heatingandcoolingloadsandindicesforhumancomfort,amongnumerousotherresults.Thesimula-tionisperformedforextremeheatingandcoolingperiodstoestablishmechanicalequipmentsizingandperformancere-sponse,andcanbecarriedoutforafullyearorpartoftheyeartoobtaincomprehensiveorspecicresults.4.2.2SimulationModelValidationTocreaterealisticenergycalculationsandpredictions,weinstrumentedandmodeledatwo-story,1700squarefootres-identialbuildingequippedwithathree-stageHVACsystem,asillustratedinFigure9.Thebuildingcontainsover100sen-sorstomonitorbuildingoperationandresponse,including80temperatureand40humiditysensors,15motionsensors,7doorsensors,electricpowermetering,andaWeb-enabledthermostatthatprovidesbothcontrolandoperationallogs.Wecreateadetailedmodelofthesystemthatincludesthebuildinglocation,constructionpropertiesandmultistageop- WallInsulation AirInltration R-value(ft2Fh/BTU) AirChangesperHour(ACH) Poor 3.6 1.5 Moderate 13.2 0.8 Well 25.7 0.25 Table2.Buildingconditionsusinginouranalysis ClimateZones Locations Zone1 Minneapolis/St.Paul,MN Zone2 Pittsburgh,PA Zone3 Washington,D.C./Stirling,VA Zone4 SanFrancisco,CA Zone5 Houston,TX Table3.WeatherconditionsusedinouranalysiserationsoftheHVACsystem.Inordertovalidatethedelityofourmodel,werunthesamecontrolandoperationallogsoftherealsysteminthesimulationandcomparetheresultswiththeempiricalmea-surements.Inordertoperformthesimulationunderthesameweatherconditions,wecollecttheactualweatherrecordsoftheweekwhenthedatacollectiontookplacefromthelocalairportweatherstationthatprovideshourlydataresolution.Also,toincreasethecredibility,wepickaweekinthewinterwithuctuatingoutdoortemperaturethatcausestheHVACsystemtoreactindifferentways.Weperformaregressionanalysisonthesimulationandempiricalresults,asshowninFigure10,andndthattheaveragedailyerrorofHVACenergyusageis1.80kWh,smallerthantheaccuracyofthepowermeter.Therefore,theseresultsindicatethatthesim-ulationaccuratelyrepresentstheempiricalenergyconsump-tionasobservedinourrealtestbed.4.2.3SimulationCongurationsUsingthevalidatedmodel,weevaluatedoursystemusingallsetsofoccupancymeasurementsundermultipledifferentbuildingconditionsandclimatezones.Table2liststhreetypesofbuildinginsulations,eachofwhichisdecidedbythecombinationofwallconstructionandairleakage.Table3liststhelocationsthatrepresentstheveclimatezonesintheU.S.,rangingfromcoldMinneapolis,MinnesotatohotHouston,Texas.Inthesesimulations,wefocusontheenergyusageofHVAC.Anyadditionalinternalloadsfromarticiallights,appliances,andradiantheatfromoccupantsareintentionallyexcluded,astheywouldrenderresultsoftimeandtemper-aturesetbackstudiesambiguous.Naturalandarticialairventilationarealsointentionallyexcludedtokeepresultsfo-cusedonchangesinthermostat.Allthesimulationoutputaretabulatedataone-minutetimestep.5EvaluationInthissection,werstdescribethebaselinealgorithmsandtheevaluationmetricsthatareusedforevaluationandcomparison.Then,wepresenttheperformanceofthesmartthermostat.5.1BaselineandOptimalAlgorithmsWecomparethesmartthermostatagainstthereactivethermostatdescribedinSection2thatinfersthreeoccupantstatesfromsensordata,asshowninFigure4(a).Thereac-tivealgorithmswitchestotheActivestatewheneveritsensesmotionringsfromahome.Thealgorithmthenwaitsforasilentperiod(Tlast)atleastthresholdminuteslongbeforeswitchingtotheIdlestate.TheIdlestateisclassiedasAwayduringtheday,andasSleepduringthenightfrom10PMto10AM(xedtimeinterval).Thereactivethermostatsonthemarketuseproprietaryalgorithmsthatarenotpubliclyavail-able,sowecreatedabest-effortreplicaofthesystembasedonmarketingliteratureandempiricalobservationsofarealsysteminaction[14,15].Asastandard,weusetheEner-gyStarsetpointtemperatureswhenevertheoccupantwakesorarrives,andtheEnergyStarsetbacktemperaturewhenevertheoccupantleavesorsleeps[34].Programmablethermostatsdonotreacttooccupancyatall,andsotheyalwaysachievethesameenergysavings,butthecomfortsacricechangesperhome.Thismakestheen-ergysavingsdifculttointerpret,becauseitisunclearhowmuchoftheenergysavingisduetoeliminatedwasteandhowmuchisduetosacricedcomfort.Forthisreason,wedonotincludetheprogrammablethermostatinthecompari-son.Inourcomparisonwiththereactivethermostat,weuseathresholdofveminutesbecauseitproducesasimilarcomfortsacricetothesmartthermostat,whichmakestheenergysavingsmorecomparable.AsmentionedinSec-tion3.2,theHMMinourfastreactiontechniquealsousesave-minutetimeintervalwithwhichtodecidestatetransi-tions.Inactualcommercialproducts,suchastheBAYwebreactivethermostat[14],alargerthresholdsuchas60min-utesisusuallyusedbydefault[35].Ourcomparisonwithve-minutethresholdisconservative,becauseusinghigherthresholdvalueswouldonlydecreaseenergysavings.Wecompareoursystemwithanoptimalalgorithmthatprovidesthetheoreticalupperboundonenergysavings.Weassumethattheoptimalschemeknowsthestatesofthehomeatalltimes,andthatthereisnolagtimeinthetemperatureadjustmentatthestateswitch.Thisimpliesthatthemisstimeoftheoptimalschemewillalwaysbezero.Theop-timalalgorithmappliesdeepsetbackwheneverthehomeisunoccupied,andusesthesametemperaturesettingsasEner-gyStarwheneverthehomeisoccupied.Thus,noalgorithmcouldachievehigherenergysavingsthantheoptimalalgo-rithmwithoutsacricingcomfort.Thesmartthermostatrunsoverthedatatracesofeitherthehomedeploymentsorthesurveysandotherdatasets.Inordertomaintainthevalidationwithlimitednumberofdata,weperformleave-one-outcrossvalidationoverthenumberofdaysofthedeploymentwhentrainingtheHMMofon-lineinferencealgorithminthesmartsystem.Forexample,givenndaysofdeployment,wetesttheHMMoneachdayusingtheremainingn1daysoflabeledgroundtruthdataastrainingdata.5.2EvaluationMetricsWeevaluatethetrade-offbetweenenergyefciencyandusercomfortintheexperimentalresultswithtwoquantita-tivemetrics:energysavingandmisstime.Energysavingisdenedasthepercentageofsavingbytheschemeoverthecostofcontinuouslymaintainingthesetpointtempera- HomeA HomeB HomeC HomeD HomeE HomeF HomeG HomeH -10 0 10 20 30 40 50 60 Energy Savings (%) Reactive Smart Optimal (a)HomeEnergySavings HomeA HomeB HomeC HomeD HomeE HomeF HomeG HomeH 0 20 40 60 80 100 120 Average Daily Miss Time (min) Reactive Smart (b)HomeMissTimeBenchmarkFigure11.Basedondatacollectedin8homes,thesmartthermostatsavesmoreenergythanreactivesystemsandsacriceslesscomfort.ture.Misstimeisdenedasthetotaltimewhenthehomeisoccupiedbutthetemperaturehasnotreachedthesetpointtemperature.Inordertoaddresssmalltemperatureuctua-tions,ourmetrictolerates1°C(1.8°F)temperaturedifferencebetweentheactualtemperatureandthesetpoint.Thisvalueiswithintheboundsofsensornoise.5.3HomeDeploymentsEvaluationWeevaluatethesmartthermostatagainstthebaselineandoptimalalgorithmsinthe8homedeploymentsusingourval-idatedhousemodel.Inthesimulation,werun14daysinJanuaryandJulyusingtheclimateinCharlottesville,VAtoevaluatebothcoolingandheating.Wesetthedeepsetbacksto10°C(50°F)forheatingand40°C(104°F)forcooling,whicharesafetemperaturesthatdonotcausedamagetoahouseinreallife.Tofurtherimprovethecredibility,weran-domlymapeachdayofoccupancydatatracestoeachdayofweatherdatatraces.Thesesimulationsareusedtocalculatetheaverageofheatinginthemiddleofwinterandcoolinginthemiddleofsummer.Figure11(a)showstheresultsofenergysavingsofthe8homesusingsensordeployments.Thesmartthermo-statoutperformsthereactivethermostatinallthe8homes.HomesA-Dhaveregularoccupancypatternssothatoursys-temachievesmoreenergysavings,withanaverageof16.3kWh(38.4%),whilethereactivethermostatsaves8.7kWh(20.6%).Incontrast,homesE-Haretypicallyoccupiedformostoftheday.Theaverageenergysavingofthesmartthermostatdecreasesto7.4kWh(17.4%).However,there-activethermostatwastesenergyduetothefrequentreactions,whicharecostlybecausetheymustusethehighercapacitybutlowerefciencystageofHVACoperation.Theaverageenergywasteis-2.9kWh(-6.9%)andthemaximumwasteiscloseto4.2kWh(10.0%).Onaverage,thereactivethermo-statsaves2.9kWh(6.8%)whilethesmartthermostatsaves11.8kWh(27.9%),whichapproachestheoptimalsavingat15.2kWh(35.9%).Thus,thesmartthermostatcanreduceenergyconsumptioninawiderangeofhomesthathavedif-ferentoccupancypatterns.Figure11(b)showsthemisstimeofthesamethreeschemesinthe8homes.Comparedtothereactivethermo-stat,thesmartthermostatisbetterwiththreehomes(B,CandE),thesamewithonehome(G),andslightlyworsewithfourhomes(A,D,FandH).Themisstimesofreactivethermo-stat,however,aremuchmorevariablethanthoseofthesmartthermostat.Onaverage,thesmartthermostathas48minutesofmisstime,whilethereactivethermostathas60minutes.Thus,thesmartthermostatactuallyreducesmisstimeby12minutesonaverage.Weconcludethatthetwoapproachesareroughlycomparableintermsofmisstime,sincethisisasmallaveragedailyimprovementthatwouldnotlikelybenoticedinmosthomes.Ontheotherhand,extremecasessuchasHomesBandCprobablywouldbenoticeable:thesmartthermostatdecreasetheirdailymisstimesby55and80minutes,respectively.6AnalysisInthissection,weanalyzehowmucheachcomponentofthesmartthermostatalgorithmcontributestoitsenergysav-ings.Wealsodiscusstheimpactofthenumberofsensors,climatezones,andbuildingtypesontheperformanceofthesmartthermostat.Finally,weuseacombinationofcensusdata,weatherdataandhousingdatatoweighteachoftheseparameterstogenerateaweightedsumofexpectedenergysavingsifthesmartthermostatwereappliedacrosstheen-tireU.S..6.1InferenceAccuracyWeevaluatetheaccuracywithwhichourHMMapproachtracksoccupancyofthehomeandcompareittothecom-mercialreactivealgorithmdescribedinSection5.1.Weex-pecttheHMMapproachtooutperformthenaivereactiveal-gorithm,sincetheHMMincorporatesrichsemanticinfor-mationfromthedeployment,incontrasttothecommercialreactivealgorithmthatonlyusesthenumberandtimingofsensorrings. HMM React5 React30 React60 React90 React120 0 10 20 30 40 50 60 70 80 90 100 Percentage of Events Detected (%) Correctly Labeled Active as Inactive Inactive as Active Figure12.OurHMMhashigheraccuracyofon-lineoc-cupancyinferencethanbaselineapproaches.Figure12showstheaccuracyofourHMMapproachininferringthestatesinthehomedeployments,incompari-sontothenaivereactivealgorithm.Asparameterizedbythreshold,weevaluatethereactivealgorithmwithdifferentvaluesofthreshold,rangingfrom5to120minutes.LowerthresholdvaluesresultinaggressiveimplementationswithfasterreactiontoAwayevents,butalsoincreasestheproba-bilityofidleActiveeventsbeingclassiedasAway;higherthresholdvaluesresultinmoreconservativeimplementa-tions.Tobefairtothereactivealgorithm,wegroupSleepandAwayeventstogetherastheInactivestate,sincethere-activeschemeusesasimplexedtimewindowtodifferen-tiateSleepfromAwayevents.Weobservethatthereactivealgorithm'saccuracywaslowerwhenSleepandAwayarenotgroupedtogether.Weshowthreeevaluationmetrics,namely,percentageofeventscorrectlylabeled,percentageofeventswhereActiveisclassiedasInactive,andviceversa.TheresultsinFigure12showthattheHMMapproach(88%correctlylabeled)outperformsthebestreactivealgo-rithm(React5withonly78%correctlylabeled).WeseethattheHMMclassiesfeweractiveeventsasinactiveandfewerinactiveeventsasactive,leadingtolowermisstimesandhigherenergygainsforthesmartthermostat.Thus,ourHMMapproachisabletoachievehigheraccuracythanreac-tiveschemesbecauseitisabletoautomaticallyincorporatesemanticinformationaboutwhichsensorsarebeingredinthehome(livingroom,bathroom,kitchenorbedroom),andotherusefulcontextnotcurrentlyusedincommercialim-plementations.Wealsoobservethat,asthresholdincreases,thereactivealgorithmisabletoreduceoreliminatethenum-berofactiveeventsthatarelabeledasinactive,byessen-tiallywaitingforalongersilenceperiodtoensurehighcon-denceintheInactivestateprediction.However,increasingthethresholdalsoleadstomoreInactiveeventsbeingclas-siedasActive,sincethealgorithmhastowaitforalongerthresholdperiodintheActivestate,beforechangingthestatetoInactive. 0 7.44 14.88 22.32 29.76 37.20 Energy Savings (kWh) ReactiveFast ReactionFast Reaction 0.00 23.56 47.11 70.67 94.22 117.78Miss Time (min) Figure13.Allcomponentsofthesmartthermostatcon-tributetothereducedenergyusageandmisstime(dataisfromHomeB).6.2EffectofEachComponentThesmartthermostatconsistsofthreemaincomponents:fastreaction,deepsetback,andpreheating.Toinvestigatetheeffectofeachcomponent,weruntheexperimentswhileaddingthesecomponentsinanaccumulativefashion,start-ingfromfastreaction,thenaddingdeepsetbackandnallyaddingpreheating.WeusedatafromhomeBforthisanal-ysis,andtheenergysavingandmisstimeareshowninFig-ure13.Theresultsindicatetheeffectofeachcomponentontotalenergysavingandmisstime.First,weseethatfastreactionoutperformsthereactivethermostatinenergysaving,from22.0%to23.2%,andmisstime,from107minutesto56min-utes.Thisisbecauseouron-lineinferencealgorithmismuchmoreeffectivethanthesimplethresholdusedinthereac-tivealgorithminbothresponsivenessandaccuracy.Oncethedeepsetbackisadded,thesmartthermostatsaves8.6%moreenergywhilethemisstimeincreasesslightlyby2minuteduetothelargertemperatureoffsetinthereactions.Finally,bypreheatingwhenpossible,thesmartthermostatcanachieveenergysavingof34.0%andimprovemisstimeof51min-utes.6.3SensitivityAnalysisWeperformasensitivityanalysistoidentifyhowsensordeployments,occupanttypesandclimatesaffecttheperfor-manceofthesmartthermostat.6.3.1SensitivitytoNumberofSensorsTheevaluationofouron-lineHMMinferencealgorithminFigure12usesdatafromallthesensorsinstalledinour8homes.ThisincludesmoresensorsthanshowninTable1,andincludessensorsondailyuseobjectssuchasthefridge,microwave,stove,sink,andshower,deployedforactivityrecognitionpurposes[36].Inthissection,weperformasim-pleanalysisofhowmanysensorsareactuallyrequiredforourproposedsmartthermostat.Inparticular,weconsidertwosetsofsensors:(i)thefullsetofsensors(12-20sensors)includingmotionsensors,door HomeA HomeB HomeC HomeD HomeE HomeF HomeG HomeH 0 10 20 30 40 50 60 70 80 90 100 Percentage of Events Detected (%) Select sensors (3-5) All sensors (12-20) Figure14.TheHMMeventdetectionaccuracyisrobustevenwithonlyasmallnumberofsensors.sensorsandreedswitchesoneverydayobjects,and(ii)theselectsetofsensors(3-5sensors),includingonlythemotionsensorinthelivingroom,bathroom,bedroom,andkitchen,andthefrontdoorswitchsensor,forallour8homes.Theselectsetischosenbasedonintuitionaboutwhichareasinthetesthomeswouldbemostindicativeofthethreeactiv-itystates,andthesameanalysiscouldbedoneon-sitebyatrainedtechnicianatthetimeofinstallation.Weonlychosetheselectsetonceforeachhouse,anddidnotchooseandevaluatemultiplesetsorusepost-factooptimization.Figure14showsthepercentageofresidentstatescor-rectlyidentiedbytheHMMforourtwosensingchoices.Weobservethatthedifferenceininferenceaccuracyforthesetwoschemesisalmostnegligibleforallhomes.Ourselectedsensorsetsufcientlycapturesresidentactivityinour8homesforthepurposeofaccuratelyinferringoccu-pancyandsleepinformation.Figure15illustrateshowthesmartthermostatperformsforthe8householdswithtwodif-ferentsensorsets.Theresultsindicatethatthesmartthermo-stat,forbothsensorsets,providessimilarenergysavingandmisstime.Theselectedsensorsetsavesenergyby28.9%onavarage,whiletheseningchoicewithallsensorsachievestheaverageenergysavingof23.6%.Theaveragemisstimeoftheselectedsensorsis54minutes,whilethesetofallsen-sorshas48minutesofmisstimeonaverage.Thus,usingourdeploymentsin8homes,weshowthepotentialofusingasmallsetof3-5sensorsatlowcosts(lessthan$25)toac-curatelyinferresidentstateforenergymonitoringpurposes.6.3.2SensitivitytoOccupancyPatternsWedividetheoccupantsofsurveysintohomeswithpe-riodicandaperiodicschedulesbyanalyzingtheoccupancypatternsineachcase.Foreachcategory,weusethevalidatedhousemodelandrunthesimulationforheatingandcoolingfor14daysinJanuaryandJuly,respectively,usingthesameweatherleofourcity.Also,foreachsimulationday,werandomlypickonedayoutofthedeploymentdaysanduseitsoccupancydataforthesimulation.Whenthesimulationnishes,wesumtheresultsofallsimulationdaysandthengettheaveragevalueforenergyusageandmisstime.Figure16illustrateshowthesmartthermostatperformsforoccupantswithtwomaindifferentoccupancypatterns.Theresultsindicatethat,forbothoccupanttypes,thesmartthermostatprovidesmuchhigherenergysavingandlowermisstimethanthereactivethermostat.Foraperiodicpeople,smartachieves26.4%energysavingand44minutesformisstime,whilereactiveprovides20.0%energysavingand60minutesformisstime.Forperiodicpeople,smartachieves32.4%energysavingand38minutesformisstime,whilere-activeprovides23.0%energysavingand65minutesformisstime.Weobservethatthesmartthermostatbenets”peri-odic”morethan”aperiodic”.Thisisbecausetheoccupancydynamicsof”periodic”islowerthan”aperiodic”,makingiteasiertopreheatforthesepeople.Ingeneral,thesmartthermostatisbetterthanthereactiveschemeacrossdifferentcategoriesofoccupancypatterns,inbothenergysavingandmisstime.6.3.3SensitivitytoClimateZonesWeevaluatethesmartthermostatalgorithmineachofthevetypicalclimatezonesacrosstheU.S..Foreachclimatezone,weusethevalidatedhousemodelandrunthesimula-tionforheatingandcoolingfor14daysinJanuaryandJuly,respectively.Also,werandomlymapthedaysofdeploy-mentstothesimulationdays.Alltheresultsareaveragedbythenumberofsimulationdays.Figure17showstheeffectofclimateontheperformanceofthesmartthermostat,whenusedwithprofessionaloccu-pants.Theseresultsindicatethatthesmartthermostatpro-videshigherenergysavingandlowermisstimethanthere-activethermostat.AnotherobservationisthatastheclimatebecomeswarmerfromMNtoTX,thesmartthermostatap-proachestheoptimalschemeintermsofpercentenergysav-ing.Thisisduetotworeasons.First,thedeepsetbacksusedbythesmartthermostatarebenecialduringtheday,butarenotusedatnightwhentheoccupantsaresleeping.Thishelpswarmclimatesmore,wherepeakloadsaretypicallymid-day.Incontrast,coldclimateshavepeakloadsatnight.Thesecondreasonwhythesmartthermostathelpsmoreisthatthetotalenergyusedishigherincoldclimates,buttheenergysavedbythesmartthermostatremainsroughlycon-stant:inawarmclimate,loweringthesetpointtemperatureby5-8degreesmaybeenoughtogetthehome'sheattoturncompletelyoff,saving100%oftheenergy.Inacoldcli-mate,ontheotherhand,loweringthesetpointbythesame5-8degreeswillonlyreducetheenergybillbyafraction.6.4ProjectedNationwideSavingsBasedonthepercentageofenergysavedbythesmartthermostatineachclimatezone(Figure17(a))andtheU.S.EnergyInformationAdministration'sdata(Table4)ontheamountofenergyusedforheating[37]andcooling[38]byresidencesineachzone,weestimatetheamountofenergythatcouldbesavedifthesmartthermostatweredeployedinallhomeswithHVACsystemsacrosstheUnitedStates.Weestimatetheenergysaved,Ezinzonezas:Ez=(Hz+Cz)Pz(1)whereHzistheenergyusedforheating,Czistheenergyusedforcooling,andPzisthepercentagesavedbyusingthesmart HomeA HomeB HomeC HomeD HomeE HomeF HomeG HomeH 0 10 20 30 40 50 60 Energy Savings (%) Smart(All sensors) Smart(Selected sensors) Optimal (a)Energy HomeA HomeB HomeC HomeD HomeE HomeF HomeG HomeH 0 10 20 30 40 50 60 70 80 90 100 Average Daily Miss Time (min) Smart (All sensors) Smart (Selected sensors) (b)MissTimeFigure15.SensitivitytoNumberofSensors Periodic Aperiodic 0 5 10 15 20 25 30 35 40 Energy Savings (%) Reactive Smart Optimal (a)Energy Periodic Aperiodic 0 10 20 30 40 50 60 70 Average Daily Miss Time (min) Reactive Smart (b)MissTimeFigure16.SensitivitytoOccupancyPatterns MN PA VA CA TX 0 5 10 15 20 25 30 35 40 45 50 Energy Savings (%) Reactive Smart Optimal (a)Energy MN PA VA CA TX 0 50 100 150 Average Daily Miss Time (min) Reactive Smart (b)MissTimeFigure17.SensitivitytoClimateZones ClimateZone Heating Cooling energysaving (billionkWh) (billionkWh) (%) 1 9 6 25.1919 2 24 25 25.8860 3 34 33 32.4408 4 23 31 40.2601 5 25 88 47.7498 Total 115 183 Table4.Energyusageforheatingandcoolingineachoftheveclimatezones.thermostatinthatclimatezone.Theprojectedenergysavednationally,NE,is:NE=5åz=1Ez(2)andthepercentageofenergysavednationally,PE,is:PE=NE å5z=1(Hz+Cz)(3)Thesecalculationsgivetheprojectednationwidesavingstobe113.9billionkWhor38.22%oftheelectricityusedforheatingandcooling.7LimitationsandFutureWorkInthecurrentwork,weassumethattherearenopetsorplantsinthebuilding.However,theexistenceofpetsandplantswouldmakeadifferenceinthesystemdesign.Forex-ample,weneedtotaketherequirementsofpetsandplantsintoconsiderationwhendecidingthesetbacktemperatures,andwouldneedtoaccountforpetswhenanalyzingoccu-pancysensordata.Inthefuturework,wewillprovideaninterfacetoallowuserstosetcustomizedsetbacktempera-tures.Futureimprovementswillalsobeneededtodifferen-tiatebetweenpetsandpeopleforoccupancysensing.Anotherlimitationofthecurrentworkisthatweonlyevaluateasingletypeofequipment;differentequipmenttypeshasdifferentefciencies,andwillofferdifferentrisksandbenetsforpreheatingandprecooling.Infuturework,weplantoinvestigatemoreequipmenttypestodoasensitiv-ityanalysis,whichwillgiveusamoreprecisevisionofthepotentialimpactofthesmartthermostatifdeployedatlargescale.Weplantoextendthesmartthermostatandfurtherim-provetheenergyefciencywiththeuseofzoning.Zoningsystemshavelongbeenusedtostabilizethetemperaturesindifferentpartsofahome,suchastherstandsecondoor,butthesesystemsareallthermostaticallycontrolled.Inpre-liminaryanalysis,wendthatonlyhalfoftheroomsareusedforupto60%ofthetimethatahomeisoccupied,andtheseroomsaresomewhatpredictablebasedonongoingac-tivitiesandthetimeofday.Thisindicatesthepotentialforsubstantiallymoresavingsbycombiningthesmartthermo-statwithzoningsystems.8ConclusionsInthispaper,wepresenttheconceptofasmartthermo-statthatsensesoccupancystatisticsinahomeinordertosaveenergythroughimprovedcontroloftheHVACsystem.Thissystemusesacombinationoflong-termoccupancyandsleeppatternswithreal-timesensordatatocontroltheHVACsystem.Weevaluatethesmartthermostatbyanalyzing51datasets,8ofwhichweregeneratedbydeployingarealsen-sornetworkinhomestocollectthewake,leave,arrive,andsleeptimesoftheoccupants.Ourresultsindicatethatthesmartthermostatcanprovidelargerenergysavingsandmorecomfortthanexistingbaselinesolutions.Thisapproachhasaverylowinitialcostoflessthan$25perhome,andcansave28%ofresidentialHVACenergyconsumptiononav-erage,withoutsacricingcomfort.Thissolutionservesanimportantneedforlow-costenergyconsumption.Thisprojecthasthepotentialforalargeimpactbecauseofitslowcost.Theimpactofmanyotherwiseeffectiveenergy-savingtechnologiesislimitedbyhighinitialcost,becausetheycantakeyearsorevendecadestoproduceapositivere-turnoninvestment.Studieshaveshownthatenergy-savingtechnologiesshouldproduceareturnoninvestmentwithintwoyearsinordertoachievewidespreadadoption[39].IntheU.S.,theaverageexpenditureperhouseholdforspaceheatingandelectricair-conditioningis$677[8]annuallyor$56.42permonth.Therefore,oursystemshouldcostabout$230tobenanciallyviable,includingthecostsofhard-ware,installation,conguration,andmaintenance.Ouranal-ysisshowsthatthesensing-basedsolutionspresentedherecanbeeffectivewithacostoflessthan$25perhomeinoff-the-shelfhardware,anditwillalsobeeasytoretrottoexistinghomesandbuildings.Recently,theAmericanRe-coveryandReinvestmentActof2009allocated$5billiontowardhelpinglow-incomefamiliesimprovetheweatheriza-tionoftheirhome.However,thismoneyisonlyexpectedtoachieveasmallpercentageofthenationaltargetforenergyreduction,andachievingtheactualtargetswillrequiremanybillionsmore.Thecostproleofthesmartthermostatwouldgivemorebangforthebucktofederalstimulusmoney:weexpectacostoflessthan$10billioninhardwaretoequipall130millionhomesintheU.S.withoursystem,savinganestimated113.9billionkWhnationwideperyear.Duetoitslowcost,thisresearchwillhelppropelthenationtowardsitsgoalofa25%improvementintheenergyefciencyofexist-ingbuildingsacrossthecountryby2030,asdenedbytheArchitecture2030Challenge[40]andreiteratedbyPresidentObama[41].9AcknowledgementsThismaterialisbaseduponworksupportedbytheNationalScienceFoundationunderGrantNo.0845761,1038271andIIS-0931972.10References[1]EnergyInformationAdministration.2005ResidentialEnergyConsumptionSurvey.http://www.eia.doe.gov/emeu/recs/contents.html.[2]EnergyPolicyBranchEnergySectorEnergyForecastingDivision.Canada'sEnergyOutlook,1996-2020.NaturalResourcesCanada,1997.[3]KRathouseandBYoung.DomesticHeating:UseofControls.Tech-nicalReportRPDH15,BuildingResearchEstablishment,UK,2004.[4]IowaEnergyCenter.Lowerenergybillswithaset-backthermostat.http://www.energy.iastate.edu/news/pr/energysavingideas/setbacktherm.htm. [5]EnvironmentalProtectionAgency.SummaryofResearchFindingsFromtheProgrammableThermostatMarket.Washington,DC:OfceofHeadquarters,2004.[6]HSachs.ProgrammableThermostats.ACEEE,2004.[7]EnergyStar.ProgrammableThermostatsSuspensionMemo.http://www.energystar.gov/ia/partners/prod_development/revisions/downloads/thermostats/Spec_Suspension_Memo_May2009.pdf.[8]U.S.EnergyInformationAdministration.Consumption&expendituredatatables.http://www.eia.doe.gov/emeu/recs/recs2001_ce/ce1-1e_climate2001.html,2001.[9]EnergyInformationAdministration.Naturalgas1998:Issuesandtrends.http://www.eia.doe.gov/oil_gas/natural_gas/analysis_publications/natural_gas_1998_issues_and_trends/it98.html,April1999.[10]U.S.EnergyInformationAdministration.Annualenergyreview(aer):Primaryenergyproductionbysource.http://www.eia.doe.gov/aer/pdf/pages/sec1_7.pdf,August2010.[11]G.Hardin.TheTragedyoftheCommons.Science(NewYork,NY),162(859):1243,1968.[12]X10HomeSecurityHomeAutomationElectronics.http://www.x10.com.[13]D.B.Crawley,L.K.Lawrie,C.O.Pedersen,F.C.Winkelmann,M.J.Witte,R.K.Strand,R.J.Liesen,W.F.Buhl,YJHuang,RHHenninger,etal.EnergyPlus:anupdate.ProceedingsofSimBuild,pages4–6,2004.[14]BAYwebThermostat.http://www.bayweb.com/.[15]ProThermostats.VerdantV8-BB-7SLineVoltageHeatOnlyThermostat-Programmable.http://www.prothermostats.com/product.php?p=verdant_v8-bb-7s&product=173694.[16]G.GaoandK.Whitehouse.TheSelf-ProgrammingThermostat:Op-timizingSetbackSchedulesbasedonHomeOccupancyPatterns.InProceedingsoftheFirstACMWorkshoponEmbeddedSensingSys-temsforEnergy-EfciencyinBuildings,pages67–72.ACM,2009.[17]TelkonetSmartEnergy.http://www.telkonet.com.[18]Viconics.VT7000V-PIRPassiveInfraredMotionDetector.http://www.viconics.com/vt7000.php.[19]Peco.SmartEnergyManagementforClassroomsandPortables.http://www.pecomanufacturing.com/products/datasheets/School_salessheet_REV_00_web.pdf.[20]M.P.Bouchelle,D.S.Parker,M.T.Anello,andKMRichardson.Fac-torsInuencingSpaceHeatandHeatPumpEfciencyfromaLarge-ScaleResidentialMonitoringStudy.InProceedingsofthe2000Sum-merStudyonEnergyEfciencyinBuildings,volume1001,page15,Washington,DC,2000.AmericanCouncilforanEnergy-EfcientEconomy(ACEEE).[21]CharlesE.Bullock.EnergySavingsthroughThermostatSetbackwithResidentialHeatPumps.ASHRAETransactions,83(AL-78-1):352–361,1978.[22]J.R.Smith,K.P.Fishkin,B.Jiang,A.Mamishev,M.Philipose,A.D.Rea,S.Roy,andK.Sundara-Rajan.RFID-basedtechniquesforhuman-activitydetection.CommunicationsoftheACM,48(9):44,2005.[23]H.Nait-CharifandS.J.McKenna.Activitysummarisationandfallde-tectioninasupportivehomeenvironment.InProceedingsofthe17thInternationalConferenceonPatternRecognition,volume4,2004.[24]E.M.Tapia,S.S.Intille,andK.Larson.ActivityRecognitionintheHomeUsingSimpleandUbiquitousSensors.PervasiveComputing,pages158–175,2004.[25]TimvanKasteren,AthanasiosNoulas,GwennEnglebienne,andBenKr¨ose.AccurateActivityRecognitioninaHomeSetting.InProceed-ingsofthe10thinternationalconferenceonUbiquitouscomputing,pages1–9,NewYork,NY,USA,2008.ACM.[26]R.J.OrrandG.D.Abowd.Thesmartoor:Amechanismfornatu-raluseridenticationandtracking.InCHI'00extendedabstractsonHumanfactorsincomputingsystems,page276.ACM,2000.[27]ParksAssociatesResearchandAnalysisforDigitalLiving.HomeSecuritySystemForecasts:2005andBeyond,November2005.[28]S.Patel,M.Reynolds,andG.Abowd.DetectingHumanMovementbyDifferentialAirPressureSensinginHVACSystemDuctwork:AnExplorationinInfrastructureMediatedSensing.PervasiveComput-ing,pages1–18,2008.[29]ShwetakN.Patel,ThomasRobertson,JulieA.Kientz,MatthewS.Reynolds,andGregoryD.Abowd.AttheFlickofaSwitch:Detect-ingandClassifyingUniqueElectricalEventsontheResidentialPowerLine.InProceedingsofthe9thinternationalconferenceonUbiqui-touscomputing,pages271–288,Berlin,Heidelberg,2007.Springer-Verlag.[30]J.E.Froehlich,E.Larson,T.Campbell,C.Haggerty,J.Fogarty,andS.N.Patel.HydroSense:Infrastructure-MediatedSingle-PointSens-ingofWhole-HomeWaterActivity.InProceedingsofthe11thIn-ternationalConferenceonUbiquitousComputing,pages235–244.ACM,2009.[31]DJCookandM.Schmitter-Edgecombe.AssessingtheQualityofAc-tivitiesinaSmartEnvironment.Methodsofinformationinmedicine,48(5):480,2009.[32]StephenS.Intille,JohnRondoni,CharlesKukla,IsabelAncona,andLingBao.AContext-AwareExperienceSamplingTool.InCHI'03:CHI'03extendedabstractsonHumanfactorsincomputingsystems,pages972–973,NewYork,NY,USA,2003.ACM.[33]S.S.Intille,K.Larson,J.Beaudin,E.M.Tapia,P.Kaushik,J.Nawyn,andTJMcLeish.ThePlaceLab:Alive-inlaboratoryforpervasivecomputingresearch(video).InProceedingsofPERVASIVE2005VideoProgram,May2005.[34]EnergyStar.ProgrammableThermostatsSetpointTimes&Temper-atures.http://www.energystar.gov/index.cfm?fuseaction=find_a_product.showProductGroup&pgw_code=TH.[35]BAYwebThermostatOwner'sManual.http://www.bayweb.com/portal/docs/WebThermostatOwnersManual.pdf.[36]V.Srinivasan,J.Stankovic,andK.Whitehouse.ProtectingyourDailyIn-HomeActivityInformationfromaWirelessSnoopingAttack.InProceedingsofthe10thinternationalconferenceonUbiquitouscom-puting,pages202–211.ACMNewYork,NY,USA,2008.[37]U.S.EnergyInformationAdministration.TableCE2-1c.Space-HeatingEnergyConsumptioninU.S.HouseholdsbyClimateZone.http://www.eia.doe.gov/emeu/recs/recs2001_ce/ce2-1c_climate2001.html,2001.[38]U.S.EnergyInformationAdministration.TableCE3-1c.ElectricAir-ConditioningEnergyConsumptioninU.S.HouseholdsbyCli-mateZone.http://www.eia.doe.gov/emeu/recs/recs2001_ce/ce3-1c_climate2001.html,2001.[39]AlfredA.Marcus,PaulSommers,andBonnieBerk.Barrierstotheadoptionofanenergyefcienttechnology.EnergyPolicy,10(2):157–158,1982.[40]E.MazriaandK.Kershner.The2030Blueprint:SolvingClimateChangeSavesBillions.CaseStudy,2030,Inc./Architecture,2030.[41]H.A.WaxmanandE.J.Markey.TheAmericanCleanEn-ergyandSecurityActof2009(DiscussionDraftSummary).http://energycommerce.house.gov/Press_111/20090331/acesa_summary.pdf.