Seite - 392 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 242 (ANN)were applied to thehigh-resolutionpredictionof building energy consumption, and their experimental results demonstrated that both models have comparable predictive power. In [2], a hybridmodel combining differentmachine learning algorithmswas presented for optimizing energyconsumptionof residentialbuildingsunder theconsiderationofbothcontinuousanddiscrete parametersofenergy. In [3], theextremelearningmachine(ELM)wasusedtoestimate thebuilding energyconsumption,andsimulationresults indicatedthat theELMperformedbetter thanthegenetic programming(GP)andtheANN.In[4], theclusterwiseregressionmethod,alsoknownas the latent class regression,which integratesclusteringandregression,wasutilizedto theaccurateandstable predictionofbuildingenergyconsumptiondata. In [5], the feasibilityandapplicabilityof support vectormachine (SVM)forbuildingenergyconsumptionpredictionwereexamined ina tropical region. Moreover, in[6–9],avariationofSVM,thesupportvectorregressor(SVR)wasproposedforforecasting thebuildingenergyconsumptionandtheelectric load. Furthermore, in [10], anovelmachine learning modelwasconstructedforestimatingthecommercialbuildingenergyconsumption. Thehistoricalbuildingenergyconsumptiondatahavehighlevelsofuncertaintiesandrandomness dueto the influenceof thehumandistribution, the thermalenvironment, theweatherconditionsand theworkinghours inbuildings. Thus, therestill exists theneedto improve thepredictionprecisionfor thisapplication. Torealize thisobjective,wecantake twostrategies intoaccount. Thefirst strategy is toadopt themorepowerfulmodelingmethods to learn the informationhiddenin thehistoricaldata, while theotherone is to incorporate theknowledgeorpatterns fromourexperienceordata into the predictionmodels. Ontheonehand, thedeeplearningtechniqueprovidesusoneverypowerful tool forconstructing thepredictionmodel. Inthedeeplearningmodels,morerepresentativefeaturescanbeextractedfrom thelowest layer to thehighest layer [11,12].Until today, thismiraculoustechniquehasbeenwidelyused invariousfields. In[13],anovelpredictor, thestackedautoencoderLevenberg–Marquardtmodelwas constructedfor thepredictionof traffic flow. In[14],anextremedeeplearningapproachthat integrates thestackedautoencoder (SAE)withtheELMwasproposedforbuildingenergyconsumptionprediction. In [15], thedeeplearningwasemployedasanensemble techniqueforcancerdetection. In[16], thedeep convolutionalneuralnetwork (CNN)wasutilized for facephoto-sketch recognition. In [17], adeep learningapproach, theGaussian–Bernoulli restrictedBoltzmannmachine (RBM)wasapplied to3D shapeclassificationthroughusingspectralgraphwaveletsandthebag-of-featuresparadigm. In[18], the deepbelief network (DBN)was applied to solve the natural languageunderstandingproblem. Furthermore, in[19], theDBNwasutilizedtofusethevirtuesofmultipleacoustic featuresfor improving therobustnessofvoiceactivitydetection.Asonepopulardeeplearningmethod, theDBNhasshownits superiority inmachine learningandartificial intelligence. Thisstudywilladoptandmodify theDBNto makeitbesuitable for thepredictionofbuildingenergyconsumption. Ontheotherhand,knowledgeorpatternsfromourexperiencecanprovideadditional information for the design of the prediction models. In [20–22], different kinds of prior knowledge were incorporated into the SVM models. In [23], the knowledge of symmetry was encoded into the type-2 fuzzy logicmodel to enhance its performance. In [24,25], the knowledge ofmonotonicity was incorporated into the fuzzy inference systems to assure themodels’monotonic input–output mappings. In [26–29],howtoencodetheknowledge intoneuralnetworkswasdiscussed.Asshown inthesestudies, throughincorporatingtheknowledgeorpattern, theconstructedmachine learning modelswillyieldbetterperformanceandhavesignificantly improvedgeneralizationability. From the above discussion, both the deep learningmethod and the domain knowledge are helpful for thepredictionmodels’performance improvement. Followingthis idea, this studytries to presentahybridmodel thatcombines theDBNmodelwith theperiodicityknowledgeof thebuilding energyconsumption to further improve thepredictionaccuracy. Thefinalpredictionresultsof the proposedhybridmodelareobtainedbycombiningtheoutputs fromthemodifiedDBNmodeland theenergy-consumingpatternmodel.Here, theenergy-consumingpatternrepresents theperiodicity propertyofbuildingenergyconsumptionandcanbeextractedfromtheobservedhistoricalenergy 392