Page - 203 - in Emerging Technologies for Electric and Hybrid Vehicles

Energies 2016,9, 594 2.1. Control-BasedDSM Thefirst classofapproaches is control-basedDSM.Theseapproachesuseanestimationof the currentstateof thesystem,andmakeonlinecontroldecisionsbasedonthis state.As theseapproaches donot take futuredecisions into account, theymaydeployflexibility at an early stage,while this flexibilitymightbeofmoreuseata later time. Asa result, it canoccur thata largepeakcannotbe preventedbecause thesystemhasalreadyusedmostof itsflexibility. Example1. Atypical example for this is the control of abattery incombinationwithphotovoltaic (PV)panels inahouse.Acontrol-basedDSMapproachwouldcharge thebatterystartingat the timemoreenergy isproduced by thePVthan is consumed inside thehouse. Onasunnysummerday, this generally leads to a full battery beforenoon.Asaconsequence, thehighPVpeakcannotbe reducedby thebatteryand the localdistributiongrid mayget capacityproblems. InGermany, this isnowadaysalreadyaseriousproblem(see, e.g., [7]). Anexampleofcontrol-basedDSMis thePowerMatcher (see,e.g., [8,9]). Recently, someof these control-basedapproacheshavebeenadaptedto incorporatesomeformofpredictions tomitigate to someextent thementioneddisadvantages (see,e.g., [10]). 2.2. Planning-BasedDSM Thesecondclassofapproaches,planning-basedDSM,makesamoredetailedpredictionof the futurepowerproduction/consumption (e.g., onedayahead in time) anduses this information to planthecontroldecisions forsmartappliances (aplanning) toattainagivengoal (e.g.,peak-shaving). Thestrengthof thisapproach, comparedtocontrol-basedDSM, is thatflexibilitycanbepreservedfor whenit is requiredthemost.Adisadvantagemaybethat tomaketheneededplannings, specialized algorithmsarerequiredatdevice level (e.g., [4])and/or forgroupsofappliancesorhouses (e.g., [2]). In the approaches described in [2,3], each house makes a prediction of its average power consumptionforeach15minintervalwithin theupcomingday(i.e., 96 intervals). Thesepredictions aresent toaneighborhoodcontroller, andsummeduptoobtain thepredictedneighborhoodprofile. Thispredictedneighborhoodprofile indicateswhenpeaksoccur,andgivesalsohintsonhowthehouse profilescouldbeadoptedtocounter thesepeaks. Basedonthis, theneighborhoodcontroller requests someorallhouses to followanewdesired(ordifference)profile [2],or it sends incentives tohouses toadapt theirprofile [3]. Eachhouseonits turnusesboththe informationsentbytheneighborhood controlleranditspreviouspredictionsandplanningtomakeanewplanningforall its controllable appliances. Thisprocedure is repeated iterativelyuntil theneighborhoodcontroller is satisfiedwith theresultingplannedneighborhoodprofile. Thedrawbackof these typeofapproaches is that theyaresensitive topoorpredictions.Whenthe predictions are not accurate, the derived planning may be of a low quality or even not valid. Althoughsomeinformation is relativelyeasytopredictat theneighborhoodlevel [11,12], it isoften hardtopredict thesameinformationat thehousehold level [11]. Furthermore, theerrormadedueto poorpredictionsat thehousehold levelmayaccumulateat theneighborhoodlevel. Thiseffect canbe explainedthebestbyasimpleartificialexample: Example2. Householdpowerpredictionproblem.Consider agroupof100houses,where eachhouse containsa televisionandwhere eachof the television isusedwitha90%probabilityat8p.m.Whenahouse controllerhas todecide if it incorporates theTVwithin itsplanningat8p.m., it shoulddo this since theprobability is close to 1.Within theplanningat theneighborhood level, this implies thatall 100 televisionsare likely tobe switched onat8p.m.However,whenonewouldestimate, onaneighborhood level, howmanytelevisionsare turnedon at8p.m., thebest estimator is the expectedvalue,which is90TVs. This shows that if predictionsarebasedon probabilities, predictionerrorsmayaccumulate at theneighborhood level. 203