Field Tests of an Adaptive Model-predictive Heating Control System

ABSTRACTModel-predictivecontrol(MPC)hasshowninthepastgreatpotentialforoptimisingtheoperation of heating control systems in buildings, but the major drawback has alwaysbeentheautomaticidentificationofthesystemitself. InthisworkwereportonfieldtestsofaheatingcontrolsystemderivedfrompreviousresearchworkatEPFL,implementingMPC with an adaptive model, i.e. a model that identifies automatically its parameters.Thesefieldtestsinvolved10sites,mostofthemsingle-familyhouses. Byalternatingona regular basis (typically every two weeks) between the original control system and themodel-predictive one, we have derived estimates for the possible energy savings; theseestimates range from 10% to 40%, with a marked improvement in the stability of theindoortemperature.INTRODUCTIONSpaceheatingisoneofthelargestconsumersofenergyinbuildings,butevenprofessionalheatinginstallersfinditremarkablydifficulttoproperlyconfigureacentralheatinginstal-lation. Furthermore,thereislittleeconomicincentiveforthemtodoso: fewcustomerswillbeabletoprovethatabuildingcoulduselessenergyifitwerebetterparameterised.Thisisespeciallytrueforsmallerinstallationssuchassingle-familydwellings. Withlittleinformationattheirdisposal,mostend-usersaresatisfiedprovidedthattheindoorcom-fort is maintained. Consequently, the energy demand of much of the existing buildingstockissignificantlyhigherthanneeded,althoughthereislittleresearchonthesubject.Asolutionistheso-calledModel-Predictive Control (MPC)classofalgorithms,whereamathematicalrepresentationofthebuilding,togetherwithamodelofthefutureclimateconditions, let the system compute the flow temperature that minimises the consumedenergy while preserving thermal comfort. MPC has attracted much interest because,provided the model is accurate and provided the prediction of future perturbations iscorrect, it is not possible to significantly outperform such a system. Furthermore, bychoosing a suitable formulation of the objective function, it is possible to incorporatedesirableattributessuchastime-varyingtariffs;futurechangesinsetpoint;night-setback;constraintsoncontrolvariables;andconstraintsontherateofchangeofcontrolvariables.Thereisnosignificantadditionalcomputationalcostforincludingsuchconstraints.ThepresentworktracesitsrootstotheNeurobat swissresearchproject[1,2,3,4,5],an early proposal for a so-called adaptive model-predictive control of heating systems.ThealgorithmsenabledanefficientMPCforHVACwithoutrequiringtheusertoprovidean identified model; the model itself, being provided with sensor data, was capable ofidentifying its own parameters while running. However, computing costs at that timemadeitscommercialimplementationimpractical.