Seite - 391 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

energies Article DeepBeliefNetworkBasedHybridModelfor BuildingEnergyConsumptionPrediction ChengdongLi1,∗,ZixiangDing1, JianqiangYi2 ID ,YishengLv2 ID andGuiqingZhang1 1 Schoolof InformationandElectricalEngineering,ShandongJianzhuUniversity, Jinan250101,China; zixiang.ding@foxmail.com(Z.D.);qqzhang@sdjzu.edu.cn(G.Z.) 2 InstituteofAutomation,ChineseAcademyofSciences,Beijing100190,China; jianqiang.yi@ia.ac.cn (J.Y.); yisheng.lv@ia.ac.cn (Y.L.) * Correspondence: lichengdong@sdjzu.edu.cn;Tel.: +86-0531-8636-1056 Received: 15December2017;Accepted: 16 January2018;Published: 19 January2018 Abstract: To enhance the prediction performance for building energy consumption, this paper presentsamodifieddeepbeliefnetwork(DBN)basedhybridmodel. Theproposedhybridmodel combines theoutputs fromtheDBNmodelwith theenergy-consumingpattern toyield thefinal predictionresults. Theenergy-consumingpattern in this studyrepresents theperiodicitypropertyof buildingenergyconsumptionandcanbeextractedfromtheobservedhistoricalenergyconsumption data. The residual data generatedby removing the energy-consumingpattern from theoriginal data are utilized to train themodifiedDBNmodel. The training of themodifiedDBN includes two steps, the first one ofwhich adopts the contrastive divergence (CD) algorithm to optimize thehiddenparameters inapre-trainway,while thesecondonedetermines theoutputweighting vectorbythe least squaresmethod. Theproposedhybridmodel isappliedto twokindsofbuilding energyconsumptiondatasets thathavedifferentenergy-consumingpatterns (daily-periodicityand weekly-periodicity). Inordertoexaminetheadvantagesoftheproposedmodel, fourpopularartificial intelligencemethods—thebackwardpropagationneuralnetwork (BPNN), thegeneralizedradial basis functionneuralnetwork (GRBFNN), theextreme learningmachine (ELM), and thesupport vector regressor (SVR)arechosenas thecomparativeapproaches. Experimental resultsdemonstrate that theproposedDBNbasedhybridmodelhasthebestperformancecomparedwiththecomparative techniques. Another thing tobementioned is that all thepredictors constructedbyutilizing the energy-consumingpatternsperformbetter thanthosedesignedonlybytheoriginaldata. Thisverifies theusefulnessof the incorporationof theenergy-consumingpatterns. Theproposedapproachcan alsobeextendedandappliedtosomeothersimilarpredictionproblemsthathaveperiodicitypatterns, e.g., the trafficflowforecastingandtheelectricityconsumptionprediction. Keywords:buildingenergyconsumptionprediction;deepbeliefnetwork; contrastivedivergence algorithm; least squares learning;energy-consumingpattern 1. Introduction With the growth of population and the development of economy,more andmore energy is consumedintheresidentialandofficebuildings. Buildingenergyconservationplaysan important role in the sustainabledevelopment of economy. However, someubiquitous issues, e.g., thepoor buildingmanagementandtheunreasonable taskscheduling,are impedingtheefficiencyof theenergy conservationpolicies. To improve the buildingmanagement and the task scheduling of building equipment,onewayis toprovideaccuratepredictionof thebuildingenergyconsumption. Nowadays, numerous data-driven artificial intelligence approaches have been proposed for buildingenergyconsumptionprediction. In [1], therandomforestandtheartificialneuralnetwork Energies 2018,11, 242;doi:10.3390/en11010242 www.mdpi.com/journal/energies391