Seite - 73 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 The electrical load profile of the office building is usually light onweekends compared to weekdays because energy is consumed according to the business schedule. In contrast to those of residential loadprofilepatterns, energy the increase anddecrease times of office building load profilesarerelatedtocommutetimeandhavesimilardailycharacteristics. Figure4showsatypical profile for building electricity loadover oneweek, fromwhich a clearweekly seasonalitypattern canbeobserved. Theweekly loadpattern isquite similarover fourweeks,withaweeklyaverage correlationof0.93. Therefore,manystudieshaveproposedloadforecastingmethodsusingweekly statisticalmethodsordividingthe timeseriesdata intoholidays,weekends,andweekdays [4,5]. However, theprocessofdividing the timeseriesdata inadatabase intoweekdays,weekends, andholidays is inefficientbecause the calendar informationmaynotbeprovided inadvance, and eachconsumergroupmayhavedifferentdaysoff.Moreover, thesimplemethodofdividingthedata intoweekdaysandholidayscannotcapture theperiodicityof the loadprofilesuchas thecommute timeandperiodicpower systemon/off states. InFigure 4, the fourthweek loadpatterndeviates somewhat fromthepreviouspattern,withsignificantpeak loadshift in theafternoon,particularlyon WednesdayandFriday(averagecorrelations forWednesdayandFridayare0.82&0.84, respectively). Asthepatternsdeviatedgreatlyonweekends(theweekendaveragecorrelation is0.71), it isdifficult to predictaccuratelyenergyconsumptionusingdailystatisticaldataalone. Therefore, featureextraction fromthe loadprofile is required tocaptureperiodic components causedbycommuting time,meal times, thermalcontrol change,elevatorsystemoperation,etc. Figure4.Thetypical loadprofileof thebusinessbuilding. 5.2. ComparisonofDecompositionPerformance Figures5and6showthe loadprofilesofFigure4decomposedbyEMDandVMD,respectively, whereeachIMFofeach loadprofilecovers fourweeks. Toanalyzevarious frequencycomponentsand preserve thesignalenergy, inEMD, thestandarddeviationas thestopcriterion isdeterminedas0.1%; hence, theweekly loadprofilesaredecomposedinto10 IMFs. AsEMDdecomposes the signalusingextremaenvelopes (Figure5), the results are similar to thoseobtainedwitha lowpassfilter.However,VMDissimilar toahighpassfilter, as itdecomposes the loadprofile fromlowfrequencycomponents. VMDIMFs(VMFs)arebandlimited;hence, they aresimilar toharmoniccomponents. Therefore,VMDefficiently identifiesperiodiccharacteristics in non-linearandnon-stationarysignalscomparedtoEMDIMFs(EMFs). 73