Citation for this paper:
Westermann, P., & Evins, R. (2021). Using Bayesian deep learning approaches for
uncertainty-aware building energy surrogate models. Energy and AI, 3, 1-13.
https://doi.org/10.1016/j.egyai.2020.100039.
UVicSPACE: Research & Learning Repository
_____________________________________________________________
Faculty of Engineering
Faculty Publications
_____________________________________________________________
Using Bayesian deep learning approaches for uncertainty-aware building energy
surrogate models
Paul Westermann & Ralph Evins
March 2021
© 2021 Paul Westermann & Ralph Evins et al. This is an open access article distributed under the terms of the Creative Commons Attribution License. https://creativecommons.org/licenses/by-nc-nd/4.0/
This article was originally published at:
https://doi.org/10.1016/j.egyai.2020.100039
ContentslistsavailableatScienceDirect
Energy
and
AI
journalhomepage:www.elsevier.com/locate/egyai
Using
Bayesian
deep
learning
approaches
for
uncertainty-aware
building
energy
surrogate
models
Paul
Westermann
∗,
Ralph
Evins
Energy and Cities Group Department of Civil Engineering, University of Victoria, Canada
h
i
g
h
l
i
g
h
t
s
g
r
a
p
h
i
c
a
l
a
b
s
t
r
a
c
t
•Developinguncertainty-aware engineer-ingsurrogatemodels.
•Comparing deep Bayesianneural net-worksandGaussianprocessmodels.
•Uncertaintyestimatescanidentifyand mitigateerrorsinsurrogatemodels.
•Aconcepttohybridizeengineering mod-elsanddata-drivenmodels.
a
r
t
i
c
l
e
i
n
f
o
Article history: Received 6 October 2020
Received in revised form 17 November 2020 Accepted 7 December 2020
Keywords:
Surrogate modelling Metamodel
Building performance simulation Uncertainty
Bayesian deep learning Gaussian Process Bayesian neural network
a
b
s
t
r
a
c
t
Fastmachinelearning-basedsurrogatemodelsaretrainedtoemulateslow,high-fidelityengineeringsimulation modelstoaccelerateengineeringdesigntasks.Thisintroducesuncertaintyasthesurrogateisonlyan approxi-mationoftheoriginalmodel.
Bayesian methods canquantifythat uncertainty,anddeeplearning modelsexist thatfollow the Bayesian paradigm.Thesemodels,namelyBayesianneuralnetworksandGaussianprocessmodels,enableustogive predic-tionstogetherwithanestimateofthemodel’suncertainty.Asaresultwecanderiveuncertainty-awaresurrogate modelsthatcanautomaticallyidentifyunseendesignsamplesthatmaycauselargeemulationerrors.Forthese samplesthehigh-fidelitymodelcanbequeriedinstead.ThispaperoutlineshowtheBayesianparadigmallows ustohybridizefastbutapproximateandslowbutaccuratemodels.
Inthispaper,wetraintwotypesofBayesianmodels,dropoutneuralnetworksandstochasticvariationalGaussian Processmodels,toemulateacomplexhighdimensionalbuildingenergyperformancesimulationproblem.The surrogatemodelprocesses35buildingdesignparameters(inputs)toestimate12annualbuildingenergy perfor-mancemetrics(outputs).Webenchmarkbothapproaches,provetheiraccuracytobecompetitive,andshowthat errorscanbereducedbyupto30%whenthe10%ofsampleswiththehighestuncertaintyaretransferredtothe high-fidelitymodel.
1. Introduction
Awealthofconceptsexisttoexplorethedesignofnewandexisting buildingstoimprovethebuildingsector’slargeclimatefootprint[1]. Scalingthemischallenging,asusuallyeachbuildingisdesigned individ-ually,respondingtotheculturalcontext,climaticconditions, surround-ingbuildingsanddesignpreferences.Thisimpedesthedistributionof
Abbreviations:BDL,Bayesiandeeplearning;BNN,Bayesianneuralnetwork;SVGP,stochastic-variationalGaussianProcess;DoE,design-of-experiment;ReLU, rectifiedlinearunit.
∗Correspondingauthor.
E-mailaddresses:pwestermann@uvic.ca(P.Westermann),revins@uvic.ca(R.Evins).
centrally-deriveddesignparadigmstothelevelofindividualbuilding projects.
Architectsandengineers playa vitalroleinbridgingthegap be-tweenhigh-levelideasandindividualbuildingprojects.Oftentheyuse buildingperformancesimulation(BPS)toolstoassesstheenergyand environmentalperformanceofvariousdesignoptionsandbalancethem againstdesignpreferences.Thecomputationalexpenseandassociated
https://doi.org/10.1016/j.egyai.2020.100039
2666-5468/© 2020TheAuthor(s).PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Fig.1.Distributionoferrorsofasurrogatemodel.Theplotshowstheerror ofasurrogatemodelwhichemulatesthesimulationoftheheatingdemandofan officebuilding(seecasestudyinSection4).Whiletheaverageabsoluteerror𝐴𝐸 andabsolutepercentageerror𝐴𝑃𝐸arelow(indicatedbytheredlines),large errorscanoccur.Thisstudyaimstoidentifythelargeerrorsusingestimatesof surrogatemodeluncertainty.
waitingtime,however,prohibitsexhaustive designspaceexploration andoptimization.This hasledresearcherstotrainmachinelearning modelsonsimulationinputandoutputdatatoemulatebuilding simu-lationmodels[2].
Thecomputationalspeedoftheseso-called‘surrogatemodels’has beenthebasisforarangeofinnovationsinthefieldofbuilding simu-lation,forexample,interactiveearly-stagedesigntools(e.g.ELSA[3], BuildingPathfinder[4],Net-ZeroNavigator[5]),fasteroptimization al-gorithms[6],anddetaileddesignsensitivityanduncertaintyanalysis [7][8].Arecentsurveyofbuildingdesignersconfirmedthatthosewho receivedrealtimefeedbackfrom asurrogatemodelarrivedat higher performingbuildingdesigns[9].
Thegrowinguseofsurrogatemodelsturnsattentiontothe robust-nessoftheiraccuracy.Theaccuracyofasurrogatemodelismeasuredby theerrorofthesurrogatemodeltoestimatethephysics-based simula-tionresults,whichisconsideredthegroundtruth.1Studieshaveshown
satisfactoryaverageaccuracyontestdata[11]whichcanbeinfluenced bythetypeandthecomplexityofinputs[12]andtheselectionof out-puts[5].
Nonetheless,averageerrorscomputedontestdatacanbedeceiving (seeFig.1).Testdatausuallyconsistsofdesignsamplesdistributed uni-formlyinthedesignspaceandmaynotreflecttheportionofthespace thebuildingdesignerisinterestedin.Largeerrorsonspecificbuilding designsmayoccur(i.e.heteroscedasticityoftheerrors),affecting impor-tantdesignchoicesandpotentiallyloweringtheenergyperformanceof thefinalbuildingdesign.
Bayesianmethods offer a frameworkto quantify theuncertainty stemmingfromtheinadequacyofanapproximatemodel(epistemic un-certainty)andrecentdevelopmentsin Bayesiandeep learning(BDL)
1 Pleasenote,thatthesurrogatemodelaccuracydoesnotreflecthowwell
theunderlyingsimulationmodelmatchesareal-worldbuilding.Thereaderis referredto[10]andmanyotherstudies,thataddressthegapbetweensimulation modelandtherealbuilding.
managed tointegrateBayesianconceptsinto largemachine learning models[13,14].AsaresultBDL-basedsurrogatemodelscan express forwhichinputstheirestimatesareuncertain.Inourcase,aBayesian surrogatemodelproducesabuildingperformanceestimateasa prob-abilitydistribution,wheretheentropyorvarianceofthatdistribution allowustoquantifytheuncertainty.Thearchitectorbuildingdesigneris thereforeprovidedwithalevelofconfidenceintheperformanceresults andthuscandefineuncertaintythresholdsabovewhichthehigh-fidelity model,heretheBPStool,isqueriedtoguaranteehighconfidenceresults (seeFig.2).
Inthisstudy,weexploretwodifferentBayesianmodels,Bayesian neuralnetworks[15]andstochasticvariationalGaussianprocess mod-els [16], toquantify epistemic uncertaintyin surrogatemodels(see Section2).Bothmodelswerechosenastheyscalewelltolarge sur-rogate modelling problemswith many inputsandoutputswhich re-quirestotrainthemodelsonlargedatasets.Webenchmarkthe over-allaccuracyagainstnon-Bayesiansurrogatemodels,validatethe qual-ityoftheuncertaintyestimate,andquantifyhowahybridizationoffast butapproximateandslowbutaccuratemodelsreducestheerrorofa surrogatemodelwhilecomputationalcostsincreaseonlyslightly(see Section5ff.).
2. Background
2.1. Motivationforsurrogatemodelling
Thecoremotivationtoemulateaphysics-basedhigh-fidelitymodel iscomputationalefficiency;simulationoutputscanbeestimatedmany ordersofmagnitudefaster,effectivelyinreal-time.Thisallowsa holis-ticdesignspaceanalysiswhichwouldbeinfeasiblewithaslow simu-lationmodel.Variousapplicationsofsurrogatemodellingarefoundin thebuildingdomainaswellasotherdomains[18,19]:
• Generaldesignspaceexploration:Therelationshipbetweendesign parametersandperformanceisinteractivelyexploredtoimprovethe user’sunderstandingofthedesignproblem [9,20]. Thiscan hap-penonthesinglebuildinglevelorontheurbanlevel[21].Oftena parallel-coordinatesplotisusedtovisualizethemulti-dimensional problemspace[5].
• Designoptimization:Thesurrogate modelis trainedandqueried toaccelerateiterativeoptimizationalgorithms[22–24].Adaptively trainingthesurrogatemodelonnewsimulationsamplescollected ateachoptimizationiterationcanfurtherincreaseoptimization per-formance[6].
• Sensitivityanalysis:Thesurrogatemodelisusedtoruntheextensive sampling(thousandsofsimulationruns)requiredforglobal sensitiv-ityanalysismethods[7].
• Designuncertaintyanalysis:Severaltypesofuncertaintiesexist dur-ingthebuildingdesignprocess-causedbyundetermineddesign pa-rameters,uncertaincontextualparameters(e.g.surrounding build-ings,carbonfactors,etc.),andvaguedesignconstraints[25].This uncertaintyisoftenquantifiedusingMonteCarlosamplingmethods, wheresamplesfrom uncertainparameter distributionsaredrawn andsimulatedtoquantifyhowthatparameteruncertainty propa-gatestobuildingperformanceuncertainty.Withasurrogatemodel, theseuncertaintiescanrapidlybecalculatedandupdated through-outthedesignprocess[8].
• Simulationmodelcalibration:Anaccuratecalibrationofa simula-tionmodelisrequiredtoassessretrofitdesignchoicesforan exist-ingbuilding.Thecalibration,i.e.theprocessofdetermining uncer-tainbuildingparameters,oftenrelieseitheroniterative optimizia-tionalgorithms[26],oronBayesiancalibrationoftheseuncertain parameters[27]. Inboth cases simulationsareiteratively run to closelymatchsimulationoutputswithmeasuredsensordataby ad-justingtheunknownparameters.Onecanusesurrogatemodelsto reducethecomputationallimitationsoftheseapproaches.Notethat
Fig.2. Uncertaintyestimatestolinkhigh-fidelitymodelandasurrogatemodel.Thesurrogatemodelprovidesbothaperformanceestimatê𝑦𝑠𝑢𝑟𝑟𝑜𝑔𝑎𝑡𝑒andan
uncertaintyestimatê𝜎𝑠𝑢𝑟𝑟𝑜𝑔𝑎𝑡𝑒.Iftheuncertaintyislarge,ahigh-fidelitymodel(e.g.abuildingenergysimulation)isqueriedtoproduceaccurateestimates𝑦𝑠𝑖𝑚ofan
engineeringdesign(e.g.abuilding).Pleasecompareto[17]whointroducedasimilarconcept.
simulationmodelcalibrationcanbedonebothforaspecific build-ing[28]ormultiplebuildings[29].Thelattercommonlyrequiresan archetypemodelwhoseparametersarerepeatedlycalibratedusing measurementsoftheconsideredbuildings[30].
2.2. Surrogatemodelderivation
Insurrogatemodelling,wefitamachinelearningmodeltoa simu-lationdataset𝐷={𝑥𝑛,𝑦𝑛}𝑁𝑛
=1=(𝑋,𝑌)consistingof𝑁 samples,where
theinputs𝑥𝑛 correspondtothesimulationparametersand𝑦𝑛 to
real-valuedoutputsofthesimulationrunrecordedforsample𝑛[19].2Inthe
caseofbuildingenergysurrogatemodels,thesimulationparametersare thebuildingdesignparameters(e.g.insulationvalueofthewalls)and theoutputsarethesimulatedbuildingperformancemetricslikethe ag-gregatedannualenergyconsumptionorgreenhousegasemissions[2]. Studiesalsoexistwithtimeseriesoutputs,likehourlyenergydemand [21].
Forderivingthesurrogatemodelthemodellerfirstneedstocarefully specifythedesignproblem,whichincludeschoosingthefreedesign pa-rametersandtheperformanceobjectivesaswellasallotherimportant contextualparameters(surroundingbuildings,etc.).Thensimulations areruntocreatethesimulationdataset𝐷.Theideaistogainmaximum informationaboutthedesignspace(thecollectionofallpossible pa-rametercombinations)persimulationrun.Tailoredsamplingschemes exist,calleddesign-of-experimentmethods[31],e.g. Latin-Hypercube-samplingthatuniformlydistributessamplesinthemultidimensional in-putspace.Thenumberofsamplesmustbespecified(e.g.10-1000 sam-plesperparameterdimension[2])andisadjustedifmodelaccuracyon testsamplesistoolow.
2.3. Accuracyinsurrogatemodelling
Theaccuracyof a surrogatemodelis quantified byhow wellits buildingperformanceestimatesmatchtrue,physics-based simulation outputs.Weassumethesimulationmodelasourground-truthmodel, anddisregardthemismatchbetweenthesimulationmodelandthe real-worldbuildingwhencalculatingthesurrogate’s accuracythroughout thepaper.
Metricslikethecoefficientofdetermination(𝑅2),themeanabsolute
percentageerror (𝑀𝐴𝑃𝐸), or theroot-mean-squared-error (𝑅𝑀𝑆𝐸) can be used to quantify accuracy[32]. Basedon [5,11], accuracies of𝑅2>0.99arefeasiblewhenestimatingannuallyaggregated
perfor-mancemetrics,e.g.heatingdemand,buttheycanbesignificantlylower whenmorecomplexperformancemetricsareestimated.
As mentioned above, surrogatemodel accuracy is commonly re-portedasonemetric,implyinghomoscedasticerrors.Thismaynot al-wayshold,i.e. theerrors maydependon thechoice of inputs (het-eroscedasticity).ByusingBayesiandeeplearning[13],weaimtotrain surrogatesthatareawareofwhereinthedesignspace,i.e.forwhich
2 Alsocategoricaloutputscanbeconsideredbutpracticalexamplesare
lack-inginbuildingsimulationliterature.
buildingdesigns𝑥∈𝑋,themodelisuncertainandmayproducelarge errors.
2.4. Uncertaintyinsurrogatemodels
Amathematicalfunction𝑓 ofthesimulationisnotexplicitly avail-able. We use thesurrogatemodel tofindan estimate 𝑓̂to approxi-mate thatfunction. Themost importantcause ofuncertaintyin sur-rogate modellingishowplausiblethedetermined𝑓̂is(model uncer-taintyorepistemicuncertainty)[13].Forthemostpart,thisuncertainty is causedbythetrainingset𝐷=(𝑋,𝑌)which containsonlya finite setofpointswithinthespaceofpossiblesimulationparameter combi-nations𝑋 (thedesignspace) andassociatedbuildingperformance𝑌. Theoretically,epistemicuncertaintycanbereducedtozerogivenmore andmoredata[13].
Weconsidertheproblemofsurrogatemodellingasfreeofaleatoric uncertainty,whichrepresentsnoiseorotherunknowns impactingthe observations.3Therefore,weonlydealwithepistemicuncertainty.We
proposethatquantifyingthisuncertaintycanbeapowerfulaidin sur-rogatemodellingasitacknowledgesthatwehavetotrainourmodel withalimitednumberofsimulationsamplesthatrepresentafractionof thedesignspace,whichmakesthesurrogatemodeluncertain.Bayesian modellingnowallowsustoreasonunderthatuncertainty,whilestill benefitingfromtheadvantagesofsurrogatemodelling,i.e.the compu-tationalefficiencyforlargescaledesignspaceexploration.
2.4.1. Othersourcesofuncertaintyinbuildingperformancesimulation
Thescopeofthisstudyisspecificallysetonestimatingthe uncer-taintycausedby trainingasurrogatemodeltoemulateasimulation model(seeFig.3).Itdoesnotconsiderorcomputeanyothersourcesof uncertaintyprevailinginbuildingperformancemodelling,whichmay includeuncertaintyindesignparameterandmodelspecification, uncer-taintyinthepropertiesofthefinalconstructionanduncertainty stem-mingfromassumptionsofinternal(e.g.occupantbehaviour)and exter-nal(e.g.climate)conditions[25].Whereuncertaintyinsurrogate mod-ellingispurelycausedbythemodellingprocess(epistemic),uncertainty inspecifyingasimulationmodelisaleatoric.Formoreinsightsonthe uncertaintiestacklingthemismatchbetweenthesimulationmodeland theconstructedbuilding,thereaderisreferredto[34]instead.
3. Bayesianmodellingforsurrogatemodels
Bayesian probability theory offersus grounded tools toquantify modeluncertainty[35].
To understand thecoreidea of Bayesian modelling, we consider a parametricmodel 𝑦=𝑓(𝑥,Θ), where 𝑥 is the input, 𝑓 isa space of possiblemodels(seeFig.4)andΘisthesetof modelparameters
3Inthecaseofsensordata,thiscancorrespondtosensornoise.Here,we
considersimulationrunstobedeterministic,i.e.theimpactofnumericalnoise tobesmall.Inthecaseofnumericalbuildingsimulation,hereEnergyPlus[33], thiscorrespondstothenumericalnoiseofsolvingthethermodynamic-based differentialequations.
Fig.3. Uncertaintyinsurrogatemodelling,anduncertaintyinbuildingperformancesimulation.
Fig.4. HeatingdemandestimatedwithaBayesianneural network,andtheassociatedepistemicuncertainty.In par-ticular,theuncertaintyofthesurrogatemodelislargewhenthe buildinghasawallthicknesswiderthan1𝑚,whichiswiderthan allsamplescontainedinthetrainingdata(out-of-sample).
(forexample,theweightsina neuralnetwork). Insteadof findinga singleΘ,inBayesianmodellingwesearchforacollectionofΘ,which likelyhasproducedtheoutput𝑌 given𝑋.Inourcasewesearchfora collectionofsurrogatemodelswithdifferentweights.
TheBayesiantheorem,asshowninEq.(1),isappliedtofinda collec-tionwhichlikelyhasproduce𝑌given𝑋.Basedonourpriorknowledge onthedistributionofthemodelweights𝑝(Θ)andcombinedwiththe likelihoodfunction 𝑝(𝑌|𝑋,Θ)=∏𝑁𝑛=1𝑝(𝑦𝑛|𝑥𝑛,Θ),which quantifiesthe probabilitythataspecificmodelparametersetgeneratedthe observa-tions(𝑋,𝑌),theposteriorofthemodelparameterscanbecomputed.
𝑝(Θ|𝑌,𝑋)= 𝑝(𝑌|𝑋,Θ)𝑝(Θ)
𝑝(𝑌|𝑋) (1)
where𝑝(𝑌|𝑋)iscalledthemarginallikelihood.Itrepresentsthe proba-bilityoftheobserveddatagiventhemodel𝑓 withallpossiblemodel pa-rameters.Itisascalarthatnormalizestheposterior.Giventheposterior, wecannowinferaboutfuturedatainformofapredictivedistribution:
𝑝(𝑦∗|𝑥∗,𝑋,𝑌)=∫ 𝑝(𝑦∗|𝑥∗,Θ)𝑝(Θ|𝑋,𝑌)𝑑Θ (2)
Themeanandvarianceorentropycanbederived,wherethelattertwo provideinformationontheuncertaintyintheestimatedvalues.Inthe buildingsurrogatemodellingsetting,wepredictanexpectedbuilding performance,e.g.annualheatingdemand,andanassociateduncertainty givenbuildingdesign parameters,e.g.thethicknessof thewall(see Fig.4).
3.1. Variationalinference
Thetrueposterioroftheweights𝑝(Θ|𝑌,𝑋)however,iscommonly intractable.This isparticularlythecaseinthebigdataregimewhen morecomplexmodelsarerequired[16].Inthesmalldataregime (be-lowafewthousandsamples)posteriorinferencewithastandard Gaus-sianProcessBayesianmodelisfeasibleandwassuccessfullyappliedfor buildingsurrogatemodels[28,36].However,withincreasing complex-ity,forexamplemoreinputsandoutputs(e.g.[12]),standardGPshave majorshortcomings:
• Themodelcomplexityislimitedasitonlyconsistsofonelayer,i.e. theoutputsoftheGParenotusedasinputstoanotherGP.This prohibitsmodelinghierarchicalstructuresandabstractinformation [14].
• Computationalcostincreasewiththecubically((𝑛3))withthe
num-berofsamples𝑛.Thisprohibitsincreasingthesizeofthesurrogate modeltrainingsettoimprovethemodelaccuracy(forexample,to trainacomplex,tailoredkernelwithmanyhyperparameters[35]). Instead,recentadvancesinvariationalinference(VI)allowusto ap-proximatethetrueposteriorofΘinbigdataproblems[37].Wepick anapproximatevariationaldistributionoverthe(latent)model param-eters𝑞𝜈(Θ)withitsownvariationalparameters𝜈.Nowwesearchfor𝜈
thatminimizesthedivergencetothetrueposteriorwhichisquantified bytheso-calledKullback-Leibler(KL)divergence.Therebythe marginal-ization,i.e.theintegrationrequiredtocalculatethetrueposterior,is turnedintoanoptimizationproblemwhichisofteneasiertosolve.The
approximativedistributionof𝑞 canbeusedtoformpredictionsabout unseensamples.
3.1.1. Variationalinferencefortrainingscalablesurrogatemodels
Scalablevariational inferencemethodshavebeendeveloped both todoapproximative inferencewithBayesianneuralnetworks(BNN) [13]andwithGaussianprocessmodels[38].Wepickedoneapproach ofeachtype(BNN,GP)thatcanbeused”off-the-shelve”,thatis scal-ableto10’000andmoretrainingsamples, andthathasshownhigh performanceinpreviouspublications[16,17].Theyareintroducedin thefollowingsections.
The interested reader is referred to [39] for an introduction to Bayesiandeeplearningapproaches.Pearceetal.[40]providesa com-parisonof variousBNNtypes;different Gaussianprocessmodeltypes whichrelyonvariationalinferenceareexplainedin[38].
3.2. DeepBayesianneuralnetworks
Theconceptof aBayesianneuralnetwork(BNN) isan extension ofstandard networkarchitectures (e.g.feed-forward neuralnetwork, convolutionalneuralnetwork,orrecurrentneuralnetwork)tofollow theBayesianmodellingparadigm[41].InaBNNwesampletheneural networkweightsfromapriordistributionratherthanhavingasingle fixedvalueasinnormalneuralnetworks,forexample,froma Gaus-sianΘ ∼𝑁(0,𝐼)[39]. Insteadof optimising thenetworkweights di-rectly, we average over all possible weights, called marginalisation. GiventhestochasticoutputoftheBNN𝑓Θ(𝑥),wereceiveamodel
like-lihood𝑝(𝑦|𝑓Θ(𝑥)).Basedonthedataset𝐷,Bayesianinferenceisusedto
computetheposteriorovertheweights𝑝(Θ|𝑋,𝑌).Thisposterior cap-turesthesetofallplausiblemodelparameters.Thisdistributionallows predictionsonunseendata.
Asmentionedabove theexactposterior isintractable, and differ-entapproximationsexist[15,40].Intheseapproximateinference tech-niques,theposterior𝑝(Θ|𝑋,𝑌)isfittedwithasimpledistribution𝑞(Θ). HereweconsidertheDropoutvariationalinferenceapproachasithas showngreat performancewhenbenchmarkedagainstother methods [15,17].
3.2.1. Dropoutvariationalinference
Dropoutvariational inferenceisa variationalinferenceapproach, i.e.itallowstofinda𝑞∗
𝜈(Θ)thatminimisestheKullback-Leibler
diver-gencetothetruemodelposterior,thatneitherrequirestochangethe architectureofcommonnetworkarchitecturesnortochangethe opti-misationalgorithmfortrainingthenetwork[39].Theinferenceofthe posteriorisdonebytrainingamodelwhichusesstochasticdropouton everyneuronlayer[42](seeFig.5).Thisstochasticdropoutisalsoused toremoveneuronswhenperformingpredictions.Byrepeatingthe pre-dictions(stochasticforwardpasses),wecreateadistributionofoutputs, whichwasshowntominimizetheKLdivergence[39].
This KL divergence objective is formally given in thefollowing, whereweapproximate𝑝(Θ|𝑋,𝑌)with𝑞(Θ)[13,39]:
Ł(Θ,𝑝)=−1 𝑁 𝑁 ∑ 𝑖=1 log𝑝(𝑦𝑖|𝑓̂Θ𝑖(𝑥 𝑖))+12−𝑁𝑝||𝜃||2 (3)
with𝑁 datapoints,dropoutprobability𝑝,weightsampleŝΘ𝑖∼𝑞∗ 𝜈(Θ),
and𝜃 thesetofthesampledistribution’sparameterstobeoptimised (weightmatricesinthedropoutcase).Notethatforeachdatapointin thetrainingsetdropoutisapplied,whichprovidesuswith𝑁 samples
ofΘ𝑖.
Whenperformingdropoutvariationalinferencethe𝑇stochastic for-wardpassesprovideuswiththeepistemicuncertaintygivenbythe vari-ance𝑉𝑎𝑟(𝑦): 𝑉𝑎𝑟(𝑦)≈ 1 𝑇 𝑇 ∑ 𝑡=1 𝑓̂Θ𝑡(𝑥)𝑇𝑓̂Θ𝑡(𝑥 𝑡)-𝐸(𝑦)𝑇𝐸(𝑦) (4)
withpredictionsinthisepistemicmodeldonebyapproximatingthe pre-dictivemean:𝐸(𝑦)≈𝑇1∑𝑇𝑡=1𝑓̂Θ𝑡(𝑥).Notethatinthisformulationwe
as-sumednonoiseinherentinthedataandtherefore,𝑉𝑎𝑟(𝑦)iszerowhen wehavenoparameteruncertainty.
3.3. Gaussianprocessesinthebigdataregime
Gaussian Processes models are attractive for non-parametric Bayesian modelling [35]. They use a Gaussian Process prior for a stochastic, latent function 𝑓 to describetherelationship between 𝑋
and𝑌 (seeFig.5).Thefunctionvalues 𝑓(𝑥)areassumedtobe sam-pledfromthatGaussianwithzeromeanandcovariancematrix𝐾,i.e.
𝑓∼(0,𝐾).Thechoiceofcovariancefunctionimpactsvariousaspects of theGP modelandalso determineswhich modelparameters Θto betuned.ThesemodelparametersareoptimizedwhentrainingtheGP model.
However,giventheabove-mentionedlimitationsofstandard Gaus-sianProcessmodels(seeSection3.1),sparseGPapproximationshave beendevelopedtohandlelargedatasetsbyloweringthecomputational complexityto(𝑛𝑚2)[38,43].4Theyrelyontheuseofinducing
vari-ables(orpseudo-inputs),i.e.areducedsetoflatentvariableswithsize
𝑚<<𝑛torepresenttheactualdataset𝐷 with𝑛samples.The𝑚inducing pointsareGPrealisations𝑢=𝑓(𝑧)attheinducinglocations𝑍 whichare inthesamespaceastheobservedinputs𝑋 (butnotnecessarilypartof
𝑋).WhentrainingtheSVGP,thelocationsoftheinducingpoints𝑍 and
thecovarianceparametersΘareoptimallychosentominimizetheKL divergence.Importantisthatthelocations𝑍 areparameterstoshape thevariationalapproximatedistribution𝑞(𝑓),ratherthanbeingpartof themodelparametersΘ,i.e.thecovariancefunctionwithparametersΘ arecalculatedfortheinducinglocations𝑍.
IncomparisontosparseGPs[43],stochasticvariationalGPs[16] al-lowmini-batchtrainingwhichfurtherreducescomputational complex-ityto(𝑛𝑏𝑎𝑡𝑐ℎ𝑚2).Since[16]andothers,multi-layereddeepGaussian
Processmodelshavebeendeveloped,too,butarenotconsideredinthis studyasourcasestudydatasetisstilloflimitedsizeandcomplexity [14,44].However,ourSVGPmodelmayberegardedasaone-layered deepGP[45].
4. Casestudy:surrogatemodelsforthedesignofnet-zeroenergy buildings
4.1. Objective
Weuseacasestudyonapopulartopicinthebuildingdomain,the designof buildingswithnet-zeroenergydemand,totrainandassess thetwoBayesianmodeltypesintroduced above.Itshallserveas an exampleshowcasingtheuseofbothmodeltypesforbuildingsurrogate modelling,butshouldnotbeconsideredasanexhaustivecomparisonof thetwo.Forthatpurposethereaderisreferredtootherstudiesinstead, e.g.[17,44].
4.2. Casestudybuilding
Weemulatethesimulationoutcomesofonearchetypebuilding con-tainedin theNet-Zeronavigatorproject[5]. AspartoftheNet-Zero navigatorproject,buildingsimulationsurrogatemodelsarehostedona web-platformwhichallowsuserstoreceivebuildingenergy consump-tionofarchetypebuildingsgivenalargesetofbuildingdesign param-etersinrealtime.Sofartheplatformreliedoncommondeterministic neuralnetworksurrogates,whosebuildingperformanceestimation ac-curacywasvalidatedonseparatebuildingdesignsnotcontainedinthe
4This blog post provides a summary on the history on sparse
Gaus-sian Process models: https://www.prowler.io/blog/sparse-gps-approximate-the-posterior-not-the-model.
Fig. 5. Considered variational-inference approachestoturnexistingsurrogate mod-ellingarchitecturesintoscalableBayesian models[15,16].
trainingdata.Allthesimulationrunsfortrainingandtestingwere col-lectedusingthewell-knownbuildingperformanceassessmentprogram EnergyPlus[46].Currently,deterministicsurrogatemodelsareused.
Inthiscasestudy,webuildasurrogatemodelofamediumoffice archetypebuilding,where35designparametersarefreetochooseand thebuildingenergyperformanceisquantifiedby12 separate perfor-mancemetrics(seeFig.6).Theofficearchitectureisbasedonworkfrom theUSDOECanmet-Energywhichderivedcommercialprototype build-ingmodels.Thedevelopmentoftheparameterset,thechoiceof per-formancemetrics,andsoftwaretogeneratethe(parametric)simulation dataset,however,wasdevelopedindividuallyforthatproject,where theparameterrangesaredirectlybasedonrequirementsinthe Cana-dianbuildingsector[47].Themechanicalsystemsareparametrizedto captureawidevarietyofconfigurationsallowingdirectmanipulation oftheair-sidesystem(incl.heatrecoveryventilation,variouspump ef-ficiencies)andplantequipmentperformanceofvarioussystems(heat pump,electricresistance heater,biogasfurnace,natural gasfurnace, airconditioningsystem).ThisallowsustoexplorealargeHVACsystem designspaceonahigh-level(incl.multi-systemsetups).Alldetailson thebuildingmaybefoundin[5].
4.2.1. Datasetandtransformations
Wesamplethedesignspaceusing10’000simulationruns,wherethe individualparametercombinationsinthedatasetarepickedusingthe space-fillingLatin-Hypercube-sampling(LHS) [31]. Similarly,werun additional3000simulationsanduseitasaseparatetestset.Thenumber
ofsimulationsrunsrequiredtofitanaccuratesurrogatemodelwas pre-viouslystudiedin[5],whereitwasfoundthat10’000runsaresuitable fortheconsideredbuilding.Eachbuildingsimulationruntook approxi-mately2minand10susing1CPUand4GBRAM,butvarieddepending ontheparameterchoices.
Priortotraining,westandardizedtheuniformlydistributedinputs withdifferentrangestobenormallydistributedwithzeromean. Fur-thermore,wetransformedthe12outputvariablestoalsobeclosetoa normaldistribution.Therefore,adaptiveBox-Coxtransformationswas applied[48].Itadaptivelyfindstransformationparameterstotransform variouskindsofdistributions(hereof12differentoutputs)tonormal distributions.This,inparticular,increasedtheaccuracyofthe multi-outputneuralnetworkcomparedtoothertransformations.
4.3. Modelarchitectures
InthissectionweprovidedetailsonthedropoutBayesianneural net-workandthestochasticvariationalGaussianProcessmodelwetrained toemulatethesimulationmodelofthecasestudybuilding.
4.3.1. BNNmodelarchitectureandimplementation
WeimplementedadropoutneuralnetworkusingtheKeras Tensor-flow API[49,50] basedon thework fromGalandGahramani[15]. Our network is a feed-forward neural network with 2 hidden lay-ers of 512 neurons which are activated with a leaky rectified lin-ear(ReLU) function. Trainingwasdonewithin 1200epochs usinga
Fig.6. Overviewofthecasestudybuilding.Thebuildingdesignparameterscorrespondtothesurrogatemodelinputsandtheannualperformancemetricstothe surrogatemodeloutputs.
batch sizeof 128 samples. A dropout rateof 5% was set.All men-tionedparameters (𝑛𝑙𝑎𝑦𝑒𝑟𝑠∈ [1,2,3], 𝑛𝑛𝑒𝑢𝑟𝑜𝑛𝑠=[256,512,1024], dropout
rate∈ [5%,10%,20%])wereanalysedina5-foldcross-validation.The modelwiththehighestaccuracyonthetestsetwaspicked. Further-more,weanalysedtheimpactofthedropoutrateontheuncertainty quality(seeSection4.4),butnosignificantchangeintheperformance wasobserved,whichagreeswiththeobservationfrom[15],thatthe un-certaintyestimatesofmodels,thatusedifferentdropoutrates,converge withthetrainingprogress.
4.3.2. GPmodelarchitectureandimplementation
We built a stochastic variational Gaussian Process model based on [16]using theGPy implementation[51]. Thefinal modelhas a Matern32covariancefunctionwithafixednoiseterm(≈ 0.001%ofthe meanabsolute valueoftherespectiveoutput)anditusesaGaussian likelihoodfunction.Weappliedoneseparatelengthscaleperoutputfor thecovariancefunction.OursparseGaussianprocessmodelused400 in-ducingpoints,whichweinitializedrandomlydrawingfromauniform distribution.Trainingwasperformedonmini-batchesof100samples usingtheAdadeltaoptimizer.
Thecovariancefunctionwaspickedafterrunninga5-foldcross val-idation(bothsquared-exponential,andMatern32kernelswere consid-ered).Althoughtheobserveddatasetisdeterministic,weconsidereda fixednoiselevelin themodel(≈ 0.001%of themeanabsolutevalue oftheoutputs)asitproducedmuch moreaccuratemodels.This im-pliesthatvarianceoftheonelayeredGaussianprocessmodelin[16]is toosmallandadeepGaussianprocessmaybeabetterchoiceforour problem.
4.4. Evaluationcriteria
Weevaluatethemodelswithregardtomultipleobjectives:(i)the model accuracy, (ii) uncertainty accuracy, (iii) the effectiveness of uncertainty-estimate-basedissue-raising.
4.4.1. 𝑅2score,𝑀𝐴𝑃𝐸 and𝐴𝑃𝐸
90scoretoquantifypredictionaccuracy
Ourerrormetricscovertwooftenusedmetricsinthefield,i.e.the R2[11]andtheMeanAbsolutePercentageError(MAPE)[52].
R2(𝑌,̂𝑌)=1 -∑𝑛 𝑖=1(𝑦𝑖-̂𝑦𝑖) 2 ∑𝑛 𝑖=1(𝑦𝑖-̄𝑌)2 (5) MAPE(𝑌,̂𝑌)=1 𝑛 𝑛 ∑ 𝑖=1 |𝑦𝑖-̂𝑦𝑖| 𝑦𝑖 (6)
where ̂𝑌 correspondstothematrixofpredictedvalues,𝑌 isthematrix of simulatedbuildingperformancevalues.Whentheerrorterm,𝑌-̂𝑌 approacheszero,R2approachesone,andMAPEgoestozero.
Thegiventwoerrormetricsprovideinsightintotheoverall perfor-manceofthemodels.However,theymaydisguiselargeerrorswhich occurfor fewsamples.Therefore,weaddedthe𝐴𝑃𝐸90 error.It
rep-resentsthe90thpercentileoftheabsoluteerrorssortedbyascending magnitude,andtherefore,allowstoestimatemaximummodelerrors whileaccountingforpossibleoccurrencesofoutliers.
4.4.2. Accuracyoftheuncertaintyestimate
Inawell-calibratedBayesianmodeltheuncertaintyestimates cap-turethetruedatadistribution,forexample,a95%posteriorconfidence
intervalalsocontainsthetruesimulationoutcomein95%ofthetimes [53]. Quantifyingthelevelofcalibrationis awell-knownconceptin classification[54]buthasalsobeenusedforregressionproblems re-cently[53,55].
Formally,we say that theuncertainty estimatesof thesurrogate modelarewell-calibratedif
∑𝑁 𝑛=1{𝑦𝑡≤𝐹
-1 𝑡 (𝑝)}
𝑁 → 𝑝forallp∈ [0,1] (7)
where 𝐹𝑡 is the cumulated density function targeting 𝑦𝑡 and 𝐹𝑡-1=
𝑖𝑛𝑓{𝑦∶𝑝≤𝐹𝑡(𝑦𝑡)}isthequantilefunction.Hereweconsidereach pre-dictionasastandard,symmetricGaussiandistribution(𝜇(𝑋),𝜎(𝑋)).5
Theconfidenceintervalscanbecomputedusingtheinversecumulated densityfunction.Toassessthecalibrationquality,wecountthe frac-tionofobservationsinthetestdatafallinginthepredictionconfidence intervalsderivedfromthequantilefunction(seeFig.8,left).
We show the level of calibration of the Bayesian models in Fig.8(left),whereperfectlycalibrateduncertaintyestimateswouldbe alignedwiththediagonal.Toquantitativelycomparedifferent calibra-tioncurves,onecanalsocomputetheabsolutedifferencebetweenthe confidencecurveandthediagonal,calledthecalibrationerrororthe areaunderthecurve(AUC)[55].Theproblemofassessingthe calibra-tionqualitybasedonthecalibrationplotisthatitcansuggestperfect qualitywithhomoscedasticuncertaintyestimates,i.e.constant uncer-taintyestimatesforanyinput.Therefore,wealsoquantifythesharpness
oftheuncertaintyestimatesbycalculatingtheoverallvarianceinthe uncertainty[53](seeSection5).
4.4.3. Discard-rankingtoquantifytheeffectivenessofuncertainty estimatesforsurrogatemodelapplication
While havingaccurate uncertaintyestimates is the one thing,in buildingsurrogatemodellingwearemostly concernedwithwarning modelusers,whenthemodelisuncertainandrecommendtoratherrun asimulationinstead(seeFig.2).Therefore,wederivearankingofthe samplesinthetestsetbasedonthemagnitudeoftheiruncertainty.This providestwoconclusions.First,ifitstronglyoverlapswiththeactual surrogatemodelerrortheuncertaintyestimatesareaneffective het-eroscedasticwarningmechanism.Second,we canuse therankingto calculatehowmuchtheaverageerrorcanbereducedwhenreferring acertainpercentageofmostuncertainsamples(here10%or20%)to thehigh-fidelitysimulationprogramthanprocessingitwithasurrogate model.
Bothaspectsareaddressedwhenplottingthemeanerrorcomputed ondiscretepercentiles ofthetestdata,wherethetestdataissorted bythemagnitude oftheuncertainty. Wecan comparethat curveto themeanerrorcomputed usingtestdatasortedbythemagnitudeof thecomputederror(oracleranking).Alargedistancebetweenthetwo curvescantellusthatthesurrogate’suncertaintyestimatesarenot help-fultopredictwhenitisinaccurate.Furthermore,bylookingattheslope ofthecurve,wecanseebyhowmuchthemeanerrorcanbereducedif wediscardallsampleswithuncertaintiesaboveacertainthreshold.
5. Results
Inthissection,weshowtheresultsofthecasestudywherewe de-riveduncertainty-awaresurrogatemodelstoreplacedbuildingenergy simulationmodels.
Inthecasestudy,wetrainedtwodifferentBayesianmachine learn-ing models to provide epistemic uncertainty estimates, i.e. a deep Bayesiandropout neural network(here abbreviated byBNN) and a stochasticvariational Gaussian Process model(SVGP) approach. We scrutinizetheperformanceofbothapproachesbycomparingtheir pre-dictiveaccuracy,bycomparingthequalityoftheuncertaintyestimates,
5 Thisisnotnecessarilytrueandpossiblyarecalibrationstepisrequired[53].
andbyquantifyinghoweffectivelytheuncertaintyestimatesallowus toidentifypossiblesurrogatepredictionerrors.
5.1. Modelaccuracyanduncertaintyquality 5.1.1. Accuracy
Webenchmarktheaccuracyofthetwomodeltypes,dropoutneural networksandSVGPmodels.Theperformancewasquantifiedusingthree performancemetricsasintroducedabove(seeSection4.4).Eachmodel wastrainedfivetimestogeneraterobustresults.Theresultsareshown inFig.7andTable1intheAppendix;detailsonthemodellayoutand trainingprocesscanbefoundinSections4.3.1and4.3.2).
Bothconsideredmodelsreachanaccuracyof𝑅2>0.97onallthe
outputs,whenpredictingbuildingperformanceofbuildingscontained in thetestdata. TheBNN is more accuratewith𝑅2⩾ 0.99(also see
Table1).Meanpercentageerrorsof𝑀𝐴𝑃𝐸<13.2%fortheGPmodel and𝑀𝐴𝑃𝐸<9.82%forBNNwerefound.Thelargesterrorsoccurwhen estimatingtheenergydemandprovidedbydifferentheatingsources(i.e. thedifferentfueltypes),andtheair-sidesystemenergydemand.Small surrogatemodelerrorsarefoundfortheotherbuildingperformance tar-getslikethephotovoltaic(PV)generation,orenergydemandforinterior lightsandequipment.
Toproverobustnessof surrogatemodelestimates,wespecifically lookatthelargesterrorsitproduces.Therefore,wecomplementour analysisofthemeanabsolutepercentageerrorwithananalysisofthe distributionoftheabsolutepercentageerrorsobservedforeachsample inthetestdata.Weextractthe90-thpercentileofthedistributionasa proxyofthelargesterrorfoundwhileignoringoutliers.Weabbreviate thismetricwith𝐴𝑃𝐸90.𝐴𝑃𝐸90errorsarefoundreachingupto22.3%
(30.5%)fortheBNN model(GPmodel),highlightingthedemandfor increasingtherobustness.
5.1.2. Uncertaintycalibration
Whenuncertaintyestimatesareperfectlycalibrated,thederived con-fidenceinterval,e.g.the90%confidenceinterval,containsthetrue out-comeintherightnumberofcases,i.e.90%ofthetimesforthegiven example.ThisisillustratedinFig.8,wherewecountedforhowmany timesthetruesimulationoutcomewascontainedintheestimated confi-denceinterval.WithaperfectlycalibratedBayesianmodeltheestimated confidenceandfractionofthetestsampleswithinthatintervalshould perfectlyalign(dashedline).Theregionbelowthedashedlineindicates anoverlyconfidentmodel(i.e.confidencebandsaretoonarrow),the regionabovethedashedlinemeansthatthemodelistoocarefulhaving toolargeconfidencebands.
WefindthattheBNNmodeliswell-calibrated,whiletheGPmodelis overlyconfident(Fig.8,left).Thelowqualityofuncertaintyestimates oftheGPmodelcanalsobeseenontheright,wherewedisplaythe distributionofalluncertaintyestimatescollectedforpredictionsofthe testdatasamples.TheaveragemagnitudeofuncertaintyintheGPmodel indicatesitstoohighconfidence,andthesmallwidthofthedistribution indicatesthattheuncertaintyestimatestendtobehomoscedastic,i.e. asimilar uncertaintyispredictedindependentlyofthemodelinputs. Thiswidthofthedistributionisalsocalledthesharpnessofuncertainty estimates(seeSection4.4).IncaseoftheBNN,thesharpnessisbetter anduncertaintyestimatesdepictasignificantlevelofvariance.
WecanconcludethattheuncertaintyestimatesoftheBNNare well-calibratedandprovideheteroscedasticuncertaintyestimates.
5.1.3. Usinguncertaintyestimatestoincreaserobustness
Inthissectionwestudyhoweffectivetheepistemicuncertainty es-timatesaretopredictinaccuraciesofthesurrogatemodel.
Theconceptisasfollows.Wesorttheuncertaintyestimatesonthe testdatabyscale,whereweassumethatsurrogatemodelestimatesare moreinaccuratewhenitisuncertain.Thesampleswithhighuncertainty willbeevaluatedbythehighfidelitysimulationprograminsteadofthe surrogatemodel(seeFig.2).Asa consequence,thesurrogatemodel
Fig.7.Summaryofresultsontheuseofdeep,uncertainty-awaresurrogatemodels.Theplotshowstheaccuracy,quantifiedusingthreedifferenterrormetrics, ofbothBayesianlearningapproachesforalltwelveoutputsconsideredinthecasestudy.Thefiguresalsoincludeperformancemetricswhenweusetheuncertainty estimatestoidentifyerror-pronesamplesinthetestdata(texturedbars,fordetailsseeSection5.1.3).
Fig.8. VisualizationofthequalityofuncertaintyestimatesoftheBNNandtheSVGP.Thequalityisquantifiedbyhowwell-calibratedandsharptheuncertainty estimatesare.Inbothregards,theBNNoutperformstheSVGPinthisstudy.
user,hereabuildingdesigner,isprovidedwithestimatesproducedby thesurrogatemodelonlywhenithashighconfidence,andwithactual simulationresultswhenthesurrogatemodelhaslowconfidence.The numberofsamplesprocessedbythecomputationallyexpensive simula-tionmodelshouldbetraded-off againstanincreaseinruntime.Here,we
handlethistrade-off bydefininganuncertaintythresholdabovewhich thesimulationprogramisqueried.
Wedefinethisthresholdasthe90th-or80th-percentileofall uncer-taintiesobservedonourtestdataset.Therationalebehindthatchoice is thatonly10%(or20%) ofallsamplesaretransferredtotheslow
Table1
ResultsoftheaccuracyoftheBayesianmodels.
𝑅 2 𝑀𝐴𝑃 𝐸 𝐴𝑃 𝐸90
BNN SVGP BNN SVGP BNN SVGP
Pumps [MWh/y] 0.990 ± 0 . 001 0 . 983 ± 0 . 001 7.180 ± 0 . 180 8 . 530 ± 0 . 260 14.830 ± 0 . 510 17 . 950 ± 0 . 610 Heating supply, Other [MWh/y] 0.990 ± 0 . 003 0 . 977 ± 0 . 001 9.820 ± 0 . 350 12 . 490 ± 0 . 430 22.300 ± 0 . 750 29 . 300 ± 1 . 480 Fans [MWh/y] 0.991 ± 0 . 004 0 . 988 ± 0 . 001 8 . 630 ± 0 . 380 8.530 ± 0 . 250 18.120 ± 0 . 770 18 . 280 ± 0 . 540 Heating supply, Elec. [MWh/y] 0.992 ± 0 . 001 0 . 986 ± 0 . 000 7.150 ± 0 . 290 8 . 670 ± 0 . 360 15.130 ± 0 . 290 18 . 260 ± 0 . 900 Heating supply, Gas [MWh/y] 0.992 ± 0 . 002 0 . 973 ± 0 . 001 9.400 ± 0 . 380 13 . 230 ± 0 . 220 21.440 ± 0 . 620 30 . 480 ± 0 . 520 Cooling supply, Elec. [MWh/y] 0 . 992 ± 0 . 002 0.998 ± 0 . 000 3 . 550 ± 0 . 200 2.820 ± 0 . 100 7 . 490 ± 0 . 560 5.820 ± 0 . 200 Heating demand [MWh/y] 0 . 995 ± 0 . 001 0.996 ± 0 . 000 3 . 960 ± 0 . 330 3.710 ± 0 . 080 8 . 040 ± 0 . 710 7.800 ± 0 . 250 Cooling demand [MWh/y] 0.997 ± 0 . 000 0.997 ± 0 . 000 2 . 440 ± 0 . 050 2.270 ± 0 . 060 4 . 980 ± 0 . 090 4.700 ± 0 . 110 Interior lights [MWh/y] 0 . 998 ± 0 . 000 0.999 ± 0 . 000 2 . 410 ± 0 . 100 1.590 ± 0 . 080 5 . 050 ± 0 . 160 3.150 ± 0 . 270 Interior equipment [MWh/y] 0.998 ± 0 . 000 0.998 ± 0 . 000 2 . 790 ± 0 . 100 1.410 ± 0 . 120 5 . 650 ± 0 . 200 2.600 ± 0 . 250 Water heating, Gas [MWh/y] 0 . 999 ± 0 . 000 1.000 ± 0 . 000 1 . 220 ± 0 . 130 0.250 ± 0 . 070 2 . 590 ± 0 . 260 0.430 ± 0 . 090 PV Generation [MWh/y] 0.999 ± 0 . 000 0.999 ± 0 . 001 3 . 030 ± 0 . 090 1.290 ± 0 . 090 6 . 040 ± 0 . 100 2.200 ± 0 . 150
Fig.9. Recordedsurrogatemodelerrorreductionaftertransferringuncertainsamplestothehigh-fidelitysimulationmodel.Thedatashowstheerrorif either100%,90%or80%ofthebuildingdesignsamplesareprocessedbythesurrogatemodelandtherestprocessedbythehigh-fidelitymodel.Inthatway,the averageerrorofsamplesprocessedbysurrogatemodelscanbedecreased(herequantifiedbythe90-percentileabsolutepercentageerror).
simulationprogram.Findingasuitablethresholdismoredifficultand shouldalsobebasedonthepreferencesofthebuildingdesigner.
InFig.9,thedecreaseintheerror ofthesurrogatemodel predic-tionsisillustratedforthethreetargetvariablescoveringtheheatsupply ofdifferentfuelsources.Thesetargetsproducedthelargesterrors(see Section5.1.1)andthus,wefocusonincreasingthesurrogaterobustness particularlyforthem.Discardingthe10%sampleswiththehighest un-certaintyonthetestdata,wecandecreasethe𝐴𝑃𝐸90errorinestimating
theannualheatingsupplywithagasfurnacefrom21.44%to16.66%.6
Thisisequivalenttoareductionof≈ 22%.
The𝑀𝐴𝑃𝐸 errorontheothersurrogatemodeloutputswasreduced by4%to18%,andthe𝐴𝑃𝐸90by5%to25%(seeFig.9).Inparticular,
thesignificantreductionofthe𝐴𝑃𝐸90errorproofstheincreaseinthe
robustnessofthesurrogatemodelpredictions.
6. Discussion
Surrogatemodelshaveshowntohelparchitectsandbuilding de-signerstorapidlyassesstheenergyperformanceoftheirdesigns[9]. However,bybeingonlyapproximative,concernsabouttherobustness
6 Tocalculatetheseerrors,weexcludethe10%or20%mostimportant
sam-plesfromEqs.5and6.Forexample,the16.66%errorwascomputedonthe 90%remainingsamplesinthetestset.
ofthesurrogatemodelaccuracyarise.ABayesianapproachfor surro-gatemodelling,allowstonotonlyprovideaperformanceestimatebut alsoinformabouttheconfidenceoftheapproximatingsurrogatemodel andpotentially,toidentifypartsofthedesignspacewherethesurrogate modelmayprovideinaccurateresults.
ThisfirstanalysisoftheuseofBayesiansurrogatemodelsrevealed essential properties on therobustnessof surrogatemodels,and how Bayesianmodellingcanbeanaidforeffectivereasoningontheenergy performanceofbuildingsundertheepistemicuncertaintyofsurrogates. Thegoalwastoaugmentsurrogatessuchthatwecanmaintainthe ben-efitsofsurrogatemodelswhileminimizingtheriskassociatedwiththe uncertaintyofsurrogatemodels.
6.1. Lackingrobustnessofsurrogatemodels
Surrogatemodelaccuracyisoftenreportedwitherrormetricslike the𝑅2 or𝑀𝐴𝑃𝐸 scores.Theyareimportantbutcanbedeceiving.A
highcoefficientofexplainedvariance(𝑅2)oralowmeanabsolute
per-centage error 𝑀𝐴𝑃𝐸,may disguisethat thesurrogatemayproduce quitelarge errorsin certainfractions ofthedesignspace. For exam-ple,wefoundthatthe90-percentileabsolutepercentageerrorcanbe ashighas22.3%althoughan𝑅2=0.99suggestsveryhighperformance
(seeTable1).Thismotivates,thatindeedmeasurestoidentifysurrogate inaccuraciescouldlessentheriskassociatedwithsurrogatemodelling.
Fig.10. ConvergenceofBNNestimateswithanincreasingnumberofMonteCarlodropoutsamples.TheplotshowsBNNheatingdemandestimatesand uncertaintyestimateswithincreasingnumberofMCsamples(seecasestudyinSection4).Bothapproximatelyconvergeafterconducting30randomdropoutruns, whichtakesaround0.8s(withoutparallelization).
6.2. Bayesianlearningtoexpresssurrogateconfidence
Resultsonthequalityofuncertaintyestimatesofthedropoutneural networkvalidatedthatitcanbeusedtoeffectivelyexpressconfidence onitspredictions,e.g.onecanformulatethattheheatingdemandfor abuildingwithawallof1𝑚thicknessisbetween220𝑀𝑊ℎ∕𝑦𝑒𝑎𝑟and 230𝑀𝑊ℎ∕𝑦𝑒𝑎𝑟witha90%confidence(seeFig.4).
Ontheotherhand,whilebeingalmostasaccurateastheneural net-workmodel,wefoundthatthestochasticvariationalGaussianProcess modelproducesmiscalibrateduncertaintyestimates.Pleasenote,that thisfindingcannotbegeneralizedasmethodsexisttocalibrate uncali-bratedestimates[53],andinotherstudiesdeepGaussianprocess mod-elswerefoundtoproducealargervarianceintheuncertaintyestimates [44].Nonetheless,theresultsontheSVGPmodelshighlightthat assess-ingthequalityofBayesianuncertaintyestimatesisimportant.
6.3. PracticalissuesofBayesiansurrogatemodels
WeleveragedtheuncertaintyestimatesoftheBNNtoraisewarnings whenthesurrogatemodelishighlyuncertain.Bydefiningathreshold, herethe90-percentileor80-percentileoftheuncertaintyestimateson thetestdata,wecouldreducethe𝐴𝑃𝐸90errorbyupto40%.Thisisa
significantfirststeptowardsthehybridizationoffast,low-fidelity,and slow,high-fidelitymodels.
Still,practicalissueshavetobesolved.Forexample,thequestion arisesonhowtoimplementtheroutingbetweenthesurrogatemodel andhigh-fidelitymodelruns.Simulationscouldbecarriedoutinthe backgroundwhiletheuserwouldbeworkingwiththeuncertain surro-gatemodelestimatesasastart.Inourcasetheresultswouldbeupdated after2minutesand10seconds,whichcorrespondstotheapproximate runtimeofonesimulation.
AnotherissueisthatthecomputationalcostofevaluatingaBayesian modelincreasescomparedtoadeterministicsurrogatemodel.When us-ingdropoutBNNs,weperformMonteCarlo(MC)dropout,i.e.we re-peatedlyevaluatetheBNNwhereasineachrunthesetof”dropped” neu-ronschangesandtherewith,theoutputsofthenetworkchange.Mean
𝜇 andstandarddeviation𝜎 oftheestimatesconvergewithincreasing numbersofMCevaluations,whichisshowninFig.10.Weperformed between10and2000MCevaluationsandreportedthemeanandthe standarddeviationoftheresultingestimates.Weconsiderbothmean
andstandarddeviationtohaveconverged,whentheyremainwithina bandof±1%ofthemeanweobservedafter2000MCdropoutruns.
Intheplotwevisualizedtheconvergenceoftheheatingdemand estimatesforasinglebuildingdesign.Theplotimpliesthatittakes ap-proximately0.8s,whichcorrespondsto30MCdropoutruns,forboth themeananduncertaintyestimatestoconverge.Without paralleliza-tion,thiswouldmeanthatMCdropoutsamplingofaBNNis30times slowerthantheevaluationofacommonfeed-forwardneuralnetwork, anditwouldpreventinteractivebuildingdesignprocesses.However, theindependentMCdropoutrunscaneasilybeparallelizedto multi-plecores.Pleasenotethattheconvergenceratedependsonthespecific buildingdesignparameters(surrogatemodelinputs)ortheconsidered buildingperformanceoutput(surrogatemodeloutputs).Afirst heuris-ticcheckforvariousinputsandoutputsindicatedthatestimatesalways convergedwithin100orlessMCdropoutruns.
Theseandotherquestionshavetobestudiedinmoredetailbefore integratingBayesiansurrogatemodelsintosoftwareproductsfor build-ingdesigners.
6.4. AccuracyoftheBayesianmodelcomparedtoadeterministicsurrogate model
Wecan comparetheresultsof thisstudytoa non-Bayesian feed-forwardneuralnetworktrainedonthesamedataset(seeTable2inthe Appendix).Detailsonthenon-bayesiannetworkusedcanbefoundin [5].IthasaverysimilarlayouttothedropoutBNN(2hiddenlayers with512neurons,leakyrectifiedlinearunitactivationfunction)and wastrainedusingthesamecostfunctionandoptimizer(1200training epochswithAdamoptimizer).
The𝑅2,𝑀𝐴𝑃𝐸 and𝐴𝑃𝐸
90scoresofthedeterministicmodel
com-putedonthetestdataarebetterformostoutputswhennouncertainty basedsamplefilteringisapplied(seeTable2).However,whenusing uncertaintythresholdstheBayesianmodelproduceslower𝑀𝐴𝑃𝐸 and 𝐴𝑃𝐸90errorsproposingthattheBNNisausefulmeanstoincreasethe
robustnessofsurrogatemodels.7
7Here,weusedauniformlydistributedsetofbuildingdesignsamplesasour
testdata.However,thismaynotberepresentativeofactualdesignprocesses.In future,acomparisonofbothneuralnetworktypes(Bayesiansurrogatemodel,
Table2
ComparisonofBayesiandropoutneuralnetwork(BNN)andnon-bayesiandeterministicneural network(ANN).Theperformanceofthedropoutneuralnetwork(BNN)isprovidedwithandwithout theapplicationofuncertainty-basedthresholding(90%/80%).
(i) 𝑅 2 -score
ANN BNN BNN 90% BNN 80%
Pumps [MWh/y] 0.992 ± 0 . 000 0 . 990 ± 0 . 001 0 . 989 ± 0 . 001 0 . 989 ± 0 . 001 Heating supply, Other [MWh/y] 0.995 ± 0 . 001 0 . 990 ± 0 . 003 0 . 989 ± 0 . 004 0 . 988 ± 0 . 004 Fans [MWh/y] 0.994 ± 0 . 002 0 . 991 ± 0 . 004 0 . 990 ± 0 . 004 0 . 989 ± 0 . 004 Heating supply, Elec. [MWh/y] 0.994 ± 0 . 000 0 . 992 ± 0 . 001 0 . 992 ± 0 . 001 0 . 992 ± 0 . 001 Heating supply, Gas [MWh/y] 0.995 ± 0 . 001 0 . 992 ± 0 . 002 0 . 992 ± 0 . 002 0 . 991 ± 0 . 002 Cooling supply, Elec. [MWh/y] 0.994 ± 0 . 001 0 . 992 ± 0 . 002 0 . 993 ± 0 . 001 0 . 992 ± 0 . 002 Heating demand [MWh/y] 0.996 ± 0 . 000 0 . 995 ± 0 . 001 0 . 995 ± 0 . 001 0 . 993 ± 0 . 002 Cooling demand [MWh/y] 0.997 ± 0 . 000 0 . 997 ± 0 . 000 0 . 996 ± 0 . 000 0 . 995 ± 0 . 000 Interior lights [MWh/y] 0.999 ± 0 . 000 0 . 998 ± 0 . 000 0 . 997 ± 0 . 000 0 . 997 ± 0 . 000 Interior equipment [MWh/y] 0.999 ± 0 . 000 0 . 998 ± 0 . 000 0 . 998 ± 0 . 000 0 . 997 ± 0 . 000 Water heating, Gas [MWh/y] 1.000 ± 0 . 000 0 . 999 ± 0 . 000 0 . 998 ± 0 . 000 0 . 998 ± 0 . 001 PV Generation [MWh/y] 1.000 ± 0 . 000 0 . 999 ± 0 . 000 0 . 998 ± 0 . 000 0 . 998 ± 0 . 000
(ii) 𝑀𝐴𝑃 𝐸
ANN BNN BNN 90% BNN 80%
Pumps [MWh/y] 6 . 480 ± 0 . 170 7 . 180 ± 0 . 180 6 . 200 ± 0 . 130 5.850 ± 0 . 130 Heating supply, Other [MWh/y] 8 . 550 ± 0 . 630 9 . 820 ± 0 . 350 8 . 380 ± 0 . 310 7.480 ± 0 . 410 Fans [MWh/y] 7 . 610 ± 1 . 000 8 . 630 ± 0 . 380 7 . 300 ± 0 . 470 6.690 ± 0 . 540 Heating supply, Elec. [MWh/y] 6 . 530 ± 0 . 370 7 . 150 ± 0 . 290 6 . 070 ± 0 . 270 5.670 ± 0 . 320 Heating supply, Gas [MWh/y] 8 . 040 ± 0 . 220 9 . 400 ± 0 . 380 7 . 880 ± 0 . 370 7.190 ± 0 . 400 Cooling supply, Elec. [MWh/y] 3 . 280 ± 0 . 260 3 . 550 ± 0 . 200 3 . 320 ± 0 . 200 3.150 ± 0 . 170 Heating demand [MWh/y] 3 . 710 ± 0 . 290 3 . 960 ± 0 . 330 3 . 550 ± 0 . 370 3.410 ± 0 . 370 Cooling demand [MWh/y] 2.240 ± 0 . 160 2 . 440 ± 0 . 050 2 . 310 ± 0 . 050 2 . 250 ± 0 . 060 Interior lights [MWh/y] 1.830 ± 0 . 170 2 . 410 ± 0 . 100 2 . 290 ± 0 . 090 2 . 180 ± 0 . 070 Interior equipment [MWh/y] 2 . 810 ± 0 . 390 2 . 790 ± 0 . 100 2 . 290 ± 0 . 080 2.130 ± 0 . 090 Water heating, Gas [MWh/y] 0.660 ± 0 . 060 1 . 220 ± 0 . 130 1 . 110 ± 0 . 130 1 . 050 ± 0 . 120 PV Generation [MWh/y] 1.650 ± 0 . 120 3 . 030 ± 0 . 090 1 . 900 ± 0 . 150 1 . 660 ± 0 . 180
(iii) 𝐴𝑃 𝐸 90
ANN BNN BNN 90% BNN 80%
Pumps [MWh/y] 12 . 450 ± 0 . 530 14 . 830 ± 0 . 510 12 . 280 ± 0 . 310 11.480 ± 0 . 230 Heating supply, Other [MWh/y] 20 . 400 ± 1 . 480 22 . 300 ± 0 . 750 17 . 160 ± 0 . 580 15.240 ± 0 . 610 Fans [MWh/y] 15 . 810 ± 1 . 540 18 . 120 ± 0 . 770 14 . 950 ± 0 . 910 13.800 ± 1 . 050 Heating supply, Elec. [MWh/y] 13 . 790 ± 0 . 810 15 . 130 ± 0 . 290 12 . 470 ± 0 . 490 11.670 ± 0 . 640 Heating supply, Gas [MWh/y] 18 . 320 ± 0 . 640 21 . 440 ± 0 . 620 16 . 660 ± 0 . 610 14.970 ± 0 . 690 Cooling supply, Elec. [MWh/y] 6 . 780 ± 0 . 560 7 . 490 ± 0 . 560 6 . 920 ± 0 . 460 6.540 ± 0 . 320 Heating demand [MWh/y] 7 . 670 ± 0 . 550 8 . 040 ± 0 . 710 7 . 260 ± 0 . 740 6.940 ± 0 . 770 Cooling demand [MWh/y] 4 . 620 ± 0 . 300 4 . 980 ± 0 . 090 4 . 710 ± 0 . 090 4.610 ± 0 . 090 Interior lights [MWh/y] 3.840 ± 0 . 330 5 . 050 ± 0 . 160 4 . 790 ± 0 . 170 4 . 560 ± 0 . 170 Interior equipment [MWh/y] 5 . 320 ± 0 . 960 5 . 650 ± 0 . 200 4 . 780 ± 0 . 200 4.450 ± 0 . 240 Water heating, Gas [MWh/y] 1.340 ± 0 . 100 2 . 590 ± 0 . 260 2 . 350 ± 0 . 270 2 . 210 ± 0 . 250 PV Generation [MWh/y] 2.460 ± 0 . 320 6 . 040 ± 0 . 100 4 . 120 ± 0 . 300 3 . 530 ± 0 . 350
7. Conclusionandoutlook
Inthisstudyweproposedtoaugmentandhybridizephysics-based simulation software with Bayesian (deep) learning surrogate mod-els.Byquantifying thesurrogate model(epistemic) uncertainty, the Bayesianparadigmacknowledgesthat surrogatemodelsare approxi-mations of original simulationmodels, andit offersa tool to effec-tivelyreasonunderthatincurreduncertaintywhileexploitingthemuch fasterruntimeofsurrogatemodelstoproduceengineeringperformance estimates.
Ina casestudy weshowcased the applicationof Bayesian surro-gatemodelsforthedesignofnet-zeroenergybuildings.Wefoundthat dropoutneuralnetworkmodelsprovidedwell-calibrateduncertainty es-timates,whichcanbeusedtoidentifybuildingdesignchoicesforwhich thesurrogatemodelproduceslargeerrors.Thelatterenablesusto re-ferthosedesignstothehigh-fidelityenergysimulationtooltoassure accurateestimatesforthearchitectorbuildingdesigner.Thatreferral processsignificantlyloweredtheerrorsincomparisontoacommon de-terministicsurrogatemodel.
non-bayesiansurrogatemodel)thattakesarchitecturaldesignpreferencesinto accountwhenchoosingthetestdatashouldbeconsidered.
Althoughallfindingsareboundtothecasestudyofabuilding sim-ulationsurrogate,resultsmotivatetoapplyBayesianlearningtoother fieldswheresurrogatemodelsarecommonlyused[19].
Infuture,weforeseethatBayesianmodelswillallowustohybridize
data-drivensurrogatemodelsandhigh-fidelitysimulationmodels[18]. This particularlyrequiresstudiesonhowhybridmodelscanworkin practiceinasurrogatemodel-baseddesignprocess.
Apartfromthat,futureresearchcouldmakeuseofBayesian surro-gate modelsforgeneralizingsurrogatemodelstocovermorebuilding designproblems[12,56].TheBayesianparadigmcouldhelp identify-ingwhenthesurrogatemodelisusedfordesignproblemsitwasnot trainedfor.Finally,Bayesianlearningformsafoundationforadaptively samplingsimulationruns,forwhichthesurrogatemodelisparticularly uncertain.Thisprogress,calledactivelearning,willbeexploredinan upcomingstudy[57].
CodeandDataavailability
Theentiresourcecodeofthiswork,theEnergyPlusdescriptionfile (.idf)ofthebuildingtemplate,andinstructionsonhowtodownloadthe datausedinthisstudyareavailableinaGitLabrepository.8
DeclarationofCompetingInterest
Theauthorswishtoconfirmthattherearenoknownconflictsof in-terestassociatedwiththispublicationandtherehasbeennosignificant financialsupportforthisworkthatcouldhaveinfluenceditsoutcome.
Acknowledgement
ThisresearchwassupportedbygrantfundingfromCANARIEviathe BESOSproject(CANARIERS-327).
References
[1] John Dulac CD, Abergel T. Tracking buildings. Tech. Rep.. Internation Energy Agency; 2019 . URL: https://www.iea.org/reports/tracking-buildings
[2] Westermann P, Evins R. Surrogate modelling for sustainable building design – a review. Energy Build 2019;198:170–86. doi: 10.1016/j.enbuild.2019.05.057 .
[3] Jusselme T . Data-driven method for low-carbon building design at early stages. EPF Lausanne; 2020. Ph.D. thesis .
[4] Open Technologies. The building pathfinder. URL http://www.buildingpathfinder. com/ .
[5] Paul Westermann, David Rulff, Kevin Cant, Gaelle Faure, Ralph Evins. Net-zero navigator: a platform for interactive net-zero building design using surrogate mod- ellingURL http://www.enerarxiv.org/page/thesis.html?id = 1975 .
[6] Waibel C , Wortmann T , Evins R , Carmeliet J . Building energy optimization: an ex- tensive benchmark of global search algorithms. Energy Build 2019;187:218–40 .
[7] Rivalin L , Stabat P , Marchio D , Caciolo M , Hopquin F . A comparison of methods for uncertainty and sensitivity analysis applied to the energy performance of new commercial buildings. Energy Build 2018;166:489–504 .
[8] Hester J, Gregory J, Kirchain R. Sequential early-design guidance for residential single-family buildings using a probabilistic metamodel of energy consumption. En- ergy and Buildings 2017;134:202–11. doi: 10.1016/j.enbuild.2016.10.047 . URL < Go to ISI > ://WOS:000390624800018
[9] Brown NC . Design performance and designer preference in an interactive, data– driven conceptual building design scenario. Des Stud 2020 .
[10] De Wilde P . The gap between predicted and measured energy performance of build- ings: a framework for investigation. Autom Constr 2014;41:40–9 .
[11] Ostergard T, Jensen RL, Maagaard SE. A comparison of six metamodeling techniques applied to building performance simulations. Applied Energy 2018;211:89–103. doi: 10.1016/j.apenergy.2017.10.102 . URL < Go to ISI > ://WOS:000425075600008 [12] Westermann P , Evins R . Using a deep temporal convolutional network as a
building energy surrogate model that spans multiple climate zones. Appl Energy 2020;264:114715 .
[13] Kendall A , Gal Y . What uncertainties do we need in Bayesian deep learning for computer vision?. In: Advances in neural information processing systems; 2017. p. 5574–84 .
[14] Damianou A , Lawrence N . Deep Gaussian processes. In: Artificial intelligence and statistics; 2013. p. 207–15 .
[15] Gal Y , Ghahramani Z . Dropout as a Bayesian approximation: representing model un- certainty in deep learning. In: International conference on machine learning; 2016. p. 1050–9 .
[16] Hensman J , Fusi N , Lawrence ND . Gaussian processes for big data. In: Uncertainty in artificial intelligence. Citeseer; 2013. p. 282 .
[17] Filos A, Farquhar S, Gomez AN, Rudner TG, Kenton Z, Smith L, et al. A systematic comparison of Bayesian deep learning robustness in diabetic retinopathy tasks. arXiv preprint arXiv:1912.10481 2019.
[18] Reichstein M , Camps-Valls G , Stevens B , Jung M , Denzler J , Carvalhais N , et al. Deep learning and process understanding for data-driven earth system science. Nature 2019;566(7743):195–204 .
[19] Wang GG , Shan S . Review of metamodeling techniques in support of engineering design optimization. J MechDes 2007;129(4):370–80 .
[20] Ritter F , Geyer P , Borrmann A . Simulation-based decision-making in early design stages. In: 32nd CIB W78 conference, Eindhoven, The Netherlands; 2015. p. 27–9 .
[21] Vazquez-Canteli J , Demir AD , Brown J , Nagy Z . Deep neural networks as surro- gate models for urban energy simulations. Journal of Physics: Conference Series 2019;1343:012002 . IOP Publishing
[22] Prada A , Gasparella A , Baggio P . On the performance of meta-models in building design optimization. Appl Energy 2018;225:814–26 .
[23] Eisenhower B, O’Neill Z, Narayanan S, Fonoberov VA, Mezic I. A method- ology for meta-model based optimization in building energy models. Energy and Buildings 2012;47:292–301. doi: 10.1016/j.enbuild.2011.12.001 . URL < Go to ISI > ://WOS:000301989800034
[24] Bre F , Roman N , Fachinotti VD . An efficient metamodel-based method to carry out multi-objective building performance optimizations. Energy Build 2020;206:109576 .
[25] Hopfe CJ , Hensen JL . Uncertainty analysis in building performance simulation for design support. Energy Build 2011;43(10):2798–805 .
[26] Coakley D , Raftery P , Keane M . A review of methods to match building energy sim- ulation models to measured data. RenewSustainEnergy Rev 2014;37:123–41 .
[27] Manfren M, Aste N, Moshksar R. Calibration and uncertainty analysis for computer models - a meta-model based approach for integrated building energy simulation. Applied Energy 2013;103:627–41. doi: 10.1016/j.apenergy.2012.10.031 . URL < Go to ISI > ://WOS:000314669500059
[28] Heo Y , Choudhary R , Augenbroe G . Calibration of building energy models for retrofit analysis under uncertainty. Energy Build 2012;47:550–60 .
[29] Sokol J , Davila CC , Reinhart CF . Validation of a Bayesian-based method for defining residential archetypes in urban building energy models. Energy Build 2017;134:11–24 .
[30] Kristensen MH , Hedegaard RE , Petersen S . Hierarchical calibration of archetypes for urban building energy modeling. Energy Build 2018;175:219–34 .
[31] Garud SS , Karimi IA , Kraft M . Design of computer experiments: a review. Comput Chem Eng 2017;106:71–95 .
[32] Roman ND , Bre F , Fachinotti VD , Lamberts R . Application and characterization of metamodels based on artificial neural networks for building performance simulation: a systematic review. Energy Build 2020:109972 .
[33] Crawley DB , Lawrie LK , Winkelmann FC , Buhl WF , Huang YJ , Pedersen CO , et al. En- ergyplus: creating a new-generation building energy simulation program. Energy Build 2001;33(4):319–31 .
[34] Tian W , Heo Y , De Wilde P , Li Z , Yan D , Park CS , et al. A review of uncertainty anal- ysis in building energy assessment. Renew Sustain Energy Rev 2018;93:285–301 .
[35] Rasmussen CE . Gaussian processes in machine learning. In: Advanced lectures on machine learning. Springer; 2004. p. 63–71 .
[36] Østergård T, Jensen RL, Maagaard SE. Building simulations supporting de- cision making in early design–a review. Renewable and Sustainable Energy Reviews 2016;61:187–201 . URL https://www.sciencedirect.com/science/article/ pii/S136403211600280X
[37] Blei DM , Kucukelbir A , McAuliffe JD . Variational inference: a review for statisticians. J AmStat Assoc 2017;112(518):859–77 .
[38] Bauer M , van der Wilk M , Rasmussen CE . Understanding probabilistic sparse Gaus- sian process approximations. In: Advances in neural information processing systems; 2016. p. 1533–41 .
[39] Gal Y . Uncertainty in deep learning. University of Cambridge 2016;1(3) .
[40] Pearce T, Zaki M, Brintrup A, Anastassacos N, Neely A. Uncertainty in neural net- works: Bayesian ensembling arXiv preprint arXiv:1810.05546
[41] Neal RM . Bayesian learning for neural networks, 118. Springer Science & Business Media; 1995 .
[42] Srivastava N , Hinton G , Krizhevsky A , Sutskever I , Salakhutdinov R . Dropout: a simple way to prevent neural networks from overfitting. JMachLearnRes 2014;15(1):1929–58 .
[43] Titsias M . Variational learning of inducing variables in sparse gaussian processes. In: Artificial intelligence and statistics; 2009. p. 567–74 .
[44] Salimbeni H , Deisenroth M . Doubly stochastic variational inference for deep Gaus- sian processes. In: Advances in neural information processing systems; 2017. p. 4588–99 .
[45] Svendsen DH , Morales-Álvarez P , Ruescas AB , Molina R , Camps-Valls G . Deep Gaus- sian processes for biogeophysical parameter retrieval and model inversion. ISPRS J Photogramm Remote Sens 2020;166:68–81 .
[46] Crawley DB , Pedersen CO , Lawrie LK , Winkelmann FC . Energyplus: energy simula- tion program. ASHRAE J 2000;42(4):49 .
[47] National Energy Code of Canada for Buildings 2017. National Research Council Canada (NRCan); 2017. URL https://nrc.canada.ca/en/certifications-evaluations- standards/codes-canada/codes-canada-publications/national-energy-code-canada- buildings-2017 .
[48] Box GE , Cox DR . An analysis of transformations. J R Stat Soc 1964;26(2):211–43 .
[49] Chollet F., et al. Keras. 2015.
[50] Abadi M , Barham P , Chen J , Chen Z , Davis A , Dean J , et al. Tensorflow: a system for large-scale machine learning.. In: OSDI, 16; 2016. p. 265–83 .
[51] GPy. GPy: A gaussian process framework in python. URL http://github.com/ SheffieldML/GPy , since; 2012.
[52] Edwards RE, New J, Parker LE, Cui B, Dong J. Constructing large scale surrogate models from big data and artificial intelligence. Applied Energy 2017;202:685–99. doi: 10.1016/j.apenergy.2017.05.155 . URL < Go to ISI > ://WOS:000407188500055 [53] Kuleshov V , Fenner N , Ermon S . Accurate uncertainties for deep learning us- ing calibrated regression. In: International conference on machine learning; 2018. p. 2796–804 .
[54] Platt J , et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. AdvLarge Margin Classifiers 1999;10(3):61–74 .
[55] Scalia G , Grambow CA , Pernici B , Li Y-P , Green WH . Evaluating scalable uncertainty estimation methods for deep learning based molecular property prediction. J Chem Inf Model 2020 .
[56] Geyer P , Singaravel S . Component-based building performance prediction using sys- tems engineering and machine learning. Appl Energy 2017;228:1439–53 .
[57] Westermann P, Evins R. Adaptive sampling for building simulation surro- gate model derivation using the Lola-Voronoi algorithm. In: International Building Performance Association (IBPSA), editor. Proceedings of the inter- national building performance simulation association, 16; 2019. p. 1559–63. doi: 10.26868/25222708.2019.211232 .