Page - 126 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 • Auto:maxfeatures=nfeatures. • Sqrt:maxfeatures=sqrt (n features). • Log2:maxfeatures= log2(nfeatures). A typicalANNarchitecture, knownasamultilayerperceptron, is a typeofmachine learning algorithmthat isanetworkof individualnodes, calledperceptrons,organized inaseriesof layers [5]. Each layer inMLP is categorized into three types: an input layer,which receives featuresused for prediction; ahidden layer,wherehidden featuresare extracted; andanoutput layer,whichyields thedeterminedresults.Amongthem, thehiddenlayerhasmanyfactorsaffectingperformance, such as thenumberof layers, thenumberofnodes involved,andtheactivationfunctionof thenode[44]. Therefore, thenetworkperformancedependsonhowthehiddenlayer is configured. Inparticular, the numberofhidden layersdetermines thedepthorshallownessof thenetwork. Inaddition, if there aremore than twohidden layers, it is called adeepneural network (DNN) [45]. To establish our MLP,weuse twohidden layerssincewedonot requiremanyinputvariables inourpredictionmodel. In addition,weuse the same epochs andbatch size as theLSTMmodelwedescribedpreviously. Furthermore,asanactivationfunction,weuseanexponential linearunit (ELU)without therectified linearunit (ReLU),whichhasgainedincreasingpopularityrecently.However, itsmaindisadvantage is that theperceptroncandie in the learningprocess. ELU[46] isanapproximate function introduced toovercomethisdisadvantage,andcanbedefinedby: f(x)= { x ifx≥0 α(ex−1) ifx<0 . (2) Thenext importantconsideration is tochoose thenumberofhiddennodes.Manystudieshave beenconductedtodetermine theoptimalnumberofhiddennodes foragiventask [15,47,48], andwe decidedtousetwodifferenthiddennodecounts: thenumberof inputvariablesand2/3of thenumber of inputvariables. Sinceweusenine inputvariables, thenumbersofhiddennodeswewilluseare 9and6. Sinceourmodelhas twohidden layers,wecanconsider three configurations,depending onthehiddennodesof thefirstandsecond layers: (9, 9), (9, 6), and(6, 6). As in therandomforest, weevaluate these configurationsusing the trainingdata for eachbuildingcluster and identify the configuration thatgives thebestpredictionaccuracy.After that,wecompare thebestMLPmodelwith therandomforestmodel foreachcluster type. 3.5. TimeseriesCross-Validation Toconstructa forecastingmodel, thedataset isusuallydivided intoa trainingsetand test set. Then, the trainingset isused inbuildingaforecastingmodelandthe test set isused inevaluating the resultingmodel.However, intraditional timeseriesforecastingtechniques, thepredictionperformance ispooreras the intervalbetweenthe trainingandforecastingtimes increases. Toalleviate thisproblem, weapplythetimeseriescross-validation(TSCV)basedontherollingforecastingorigin[49].Avariation of this approach focuses ona singlepredictionhorizon for each test set. In this approach,weuse various trainingsets, eachcontainingoneextraobservationthanthepreviousone.Wecalculate the predictionaccuracybyfirstmeasuring theaccuracy foreachtest setandthenaveragingtheresults of all test sets. Thispaperproposes aone-week (sumfrom145h to168h) look-aheadviewof the operation forsmartgrids. For this, a seven-step-aheadforecastingmodel isbuilt to forecast thepower consumption at a single time (h+ 7+ i− 1) using the test setwith observations at several times (1,2, . . . ,h+ i−1). Ifhobservationsarerequiredtoproduceareliable forecast, then, for the totalT observations, theprocessworksas follows. For i=1 toT−h−6: (1) Select theobservationat timeh+7+i−1 for the test set; (2) Consider theobservationsat several times1,2, · · · , h+ i−1 toestimate the forecastingmodel; 126