Page - 11 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2226 3. ForecastingResults 3.1.Dataset ofExperimentalExamples Totest theperformanceof theproposedLS-SVR-CQFOAmodel, thispaperemploystheMELdata fromanislanddataacquisitionsystemin2014(IDAS2014) [45]andthedataofGEFCom2014-E[46] to carryoutanumerical forecast. Taking thewhole timeof 24has the sampling interval, the load datacontains168-hour loadvalues in total, i.e., from01:0014 July2014 to24:0020 July2014 in IDAS 2014 (namely IDAS2014), andanother two loaddatasetswith the same168-hour loadvalues, i.e., from01:001 January2014 to24:007 January2014(namelyGEFCom2014(Jan.)) andfrom01:001 July 2014 to24:007 July2014 (namelyGEFCom2014(July)) inGEFCom2014-E, respectively. Theprecisenessandintegrityofhistoricaldatadirectly impact the forecastingaccuracy. Thedata of thehistorical loadare collectedandobtainedbyelectrical equipment. To someextent, thedata transmissionandmeasurementwill leadtosome“baddata”inthedataofhistorical load,whichmainly includesmissingandabnormaldata. If thesedataareusedformodeling, theestablishmentof load forecastingmodelandthe forecastingwillbringadverseeffects. Thus, thepreprocessingofhistorical data is essential to load forecasting. In this paper, before the numerical test, thedata of theMEL arepreprocessed, including: completing themissingdata; identifyingabnormaldata; eliminating and replacing unreasonable data; andnormalizingdata. When the input of anLS-SVRmodel is multidimensionalwitha largedatasize (e.g., severalordersofmagnitude), itmayleadtoproblems whenusingtherawdatato implementmodel trainingdirectly. Therefore, it isessential that thesample dataarenormalizedforprocessing, tokeepall thesampledatavalues inacertain interval (this topic limits [0,1]), ensuringthatallof thedatahavethesameorderofmagnitude. Thenormalizationof loaddata isconvertedaccordingtoEquation(31),where i=1,2, . . . ,N (N is thenumberofsamples);xi andyi represent thevaluesofbeforeandafter thenormalizationofsample data, respectively; andmin(xi) andmax(xi) represent theminimal andmaximalvaluesof sample data, respectively. yi= xi−min(xi ) max(xi)−min(xi) (31) After the end of the forecasting, it is necessary to use the inverse normalization equation to calculate theactual loadvalue,asshowninEquation(32): xi=(max(xi)−min(xi))yi+min(xi). (32) Thenormalizeddataof thevalues in IDAS2014,GEFCom2014 (Jan.) andGEFCom2014 (July)are collectedandshowninTables1–3, respectively. Duringthemodelingprocesses, the loaddataaredividedinto threeparts: the trainingsetwith the former120h, thevalidationsetwith themiddle24h,andthe testingsetwith the latter24h. Then, therolling-basedmodelingprocedure,proposedbyHong[18,47], isappliedtoassistCQFOAto look forappropriateparameters, (γ,σ),ofanLS-SVRmodelduringthetrainingstage.Repeat thismodeling procedureuntilall forecasting loadsarereceived. Thetrainingerrorandthevalidationerrorcanbe calculatedsimultaneously. Theadjustedparameters, (γ,σ),wouldbeselectedas themost suitable parametersonlywithboth thesmallestvalidationandtestingerrors. The testingdataset isneverused during the trainingandvalidationstages; itwill onlybeused tocalculate the forecastingaccuracy. Eventually, the24h’s loaddataare forecastedbytheproposedLS-SVR-CQFOAmodel. 11