Page - 125 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 electrical loaddata. However, theLSTMnetworks can reflectyesterday’s information for thenext day’s forecast. Since thedaily load forecasting of a smart grid aims to be scheduleduntil after a week,LSTMnetworksarenotsuitable forapplicationto thedaily loadforecastingbecause there isa gapofsixdays. Furthermore, if theprediction isnotvalid, theLSTMmodelmethodcangiveabad result. For instance,nationalholidays,quickclimatechange,andunexpected institution-relatedevents canproduceunexpectedpowerconsumption. Therefore, theLSTMmodelalone isnot enough for short-termloadforecastingdueto its simplestructureandweakness involatility. Eventually,asimilar lifepatterncanbeobserveddependingonthedayoftheweek,whichinreturngivesasimilarelectrical loadpattern. This studyuses theLSTMnetworksmethod to showthe repeatingpatternofpower consumptionsdependingonthedayof theweek. The inputvariableperiodof the trainingdataset iscomposedof theelectrical loadfrom2012to2015andthedependentvariableof the trainingset is composedof theelectrical loadfrom2013to2015.Weperformed10-foldcross-validationonarolling basis foroptimalhyper-parameterdetection. 3.3.DiscoveringSimilarTimeseriesPatterns Sofar,diversemachine learningalgorithmshavebeenproposedtopredictelectrical load[1,3,6,14]. However, they showed different prediction performances depending on the various factors. For instance, for timeseriesdata,onealgorithmgives thebestpredictionperformanceononesegment, while for other segments, another algorithm can give the best performance. Hence, oneway to improve theaccuracy in thiscase is tousemore thanonepredictivealgorithm.Weconsiderelectrical loaddataas timeseriesdataandutilizeadecisiontree toclassify theelectrical loaddatabypattern similarity. Decision trees [26,40] can handle both categorical and numerical data, and are highly persuasivebecause theycanbeanalyzedthrougheachbranchof the tree,whichrepresents theprocess of classificationorprediction. Inaddition, theyexhibit ahighexplanatorypowerbecause theycan confirmwhich independentvariableshaveahigher impactwhenpredicting thevalueofadependent or targetvariable.Ontheotherhand,continuousvariablesused in thepredictionofvaluesof the time seriesareregardedasdiscontinuousvalues,andhence, thepredictionerrorsare likely tooccurnear theboundaryofseparation.Hence,using thedecision tree,wedividecontinuousdependentvariables intoseveral classeswithasimilarelectrical loadpattern. Todothis,weuse the trainingdataset from theprevious threeyears.Weusethedailyelectrical loadas theattributeofclass labelordependent variableandthecharacteristicsof the timeseriesas independentvariables, representingyear,month, day,dayof theweek,holiday,andacademicyear.Detailsontheclassificationof timeseriesdatawill beshownintheexperimental section. 3.4. BuildingaHybridForecastingModel Toconstructourhybridpredictionmodel,wecombinebotharandomforestmodelandmultilayer perceptronmodel. Randomforest isarepresentativeensemblemodel,whileMLPisarepresentative deep learningmodel;both thesemodelshaveshownanexcellentperformance in forecastingelectrical load[5,12,15–17]. Random forest [41,42] is an ensemblemethod for classification, regression, and other tasks. It constructsmanydecision trees that canbeused to classify anew instance by themajority vote. Eachdecisiontreenodeusesasubsetofattributesrandomlyselectedfromtheoriginal setofattributes. Randomforest runsefficientlyonlargeamountsofdataandprovidesahighaccuracy[43]. Inaddition, comparedtoothermachine learningalgorithmssuchasANNandSVR, it requires lessfine-tuningof itshyper-parameters [16]. Thebasicparametersof randomforest includethe totalnumberof trees to begenerated(nTree)andthedecision tree-relatedparameters (mTry), suchasminimumsplit andsplit criteria [17]. In this study,wefindtheoptimalmTryandnTree forour forecastingmodelbyusingthe trainingsetandthenverify theirperformanceusingthevalidationandtest set. Theauthors in [42] suggestedthatarandomforest shouldhave64 to128 treesandweuse128 trees forourhybridSTLF model. Inaddition, themTryvaluesusedfor this studyprovidedbyscikit-learnareas follows. 125