Page - 133 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 5.3. ComparisonofForecastingTechniques Toverifythevalidnessandapplicabilityofourhybriddaily loadforecastingmodel, thepredictive performanceofourmodel shouldbecomparedwithothermachine learning techniques, including ANNandSVR,whichareverypopularpredictivetechniques[6]. Inthiscomparison,weconsidereight models, includingourmodel, as showninTable9. In the table,GBM(GradientBoostingMachine) isa typeofensemble learning technique that implements thesequentialboostingalgorithm.Agridsearch canbeused tofindoptimalhyper-parametervalues for theSVR/GBM[25]. SNN(ShallowNeural Network)has three layersof input,hidden,andoutput,anditwas foundthat theoptimalnumberof thehiddennodes isnine forall clusters. Tables9–11comparethepredictionperformanceintermsofMAPE,RMSE,andMAE,respectively. Fromthetables, thepredictedresultswith thebestaccuracyaremarkedinboldandweobserve that ourhybridmodelexhibitsasuperbperformance inall categories. Figure7showsmoredetailof the MAPEdistributionforeachclusterusingaboxplot.Wecandeduce thatourhybridmodelhas fewer outliersandasmallermaximumerror. Inaddition, theerror rate increases in thecaseof longholidays inKorea. For instance,duringthe10-dayholiday inOctober2017, theerror rate increasedsignificantly. Anothercauseofhigherrorrates isduetooutliersormissingvaluesbecauseofdiversereasons, such asmalfunctionandsurge. Figure8compares thedaily loadforecastsofourhybridmodelandactual dailyusageonaquarterlybasis.Overall,ourhybridmodelshowedagoodperformanceinpredictions, regardlessofdiverseexternal factorssuchas longholidays. Table9.MAPEdistributionforeachforecastingmodel. ForecastingModel Cluster# ClusterA ClusterB ClusterC MR 7.852 8.971 4.445 DT 6.536 8.683 6.004 GBM 4.831 6.896 3.920 SVR 4.071 5.761 3.135 SNN 4.054 5.948 3.181 MLP 3.961 4.872 3.139 RF 4.185 5.641 3.216 RF+MLP 3.798 4.674 2.946 Table10.RMSEcomparisonforeachforecastingmodel. ForecastingModel Cluster# ClusterA ClusterB ClusterC MR 5725.064 6847.179 1757.463 DT 6118.835 7475.188 2351.676 GBM 4162.359 5759.276 1495.751 SVR 3401.812 5702.405 1220.052 SNN 3456.156 4903.587 1236.606 MLP 3381.697 4064.559 1170.824 RF 4111.245 4675.762 1450.436 RF+MLP 3353.639 3894.495 1143.297 Nevertheless,wecansee that thereare several timeperiodswhen forecastingerrors arehigh. For instance, from2013to2016,ClusterBshowedasteadyincrease in itspowerconsumptiondueto buildingremodelingandconstruction. Eventhoughtheremodelingandconstructionarefinishedat thebeginningof2017, the inputvariable forestimatingtheyear-aheadconsumption isstill reflecting such an increase. Thiswas eventually adjustedproperly for the third and fourth quarters by the timeseriescross-validation.Ontheotherhand,duringtheremodeling, theoldheating,ventilation, andairconditioning(HVAC)systemwasreplacedbyamuchmoreefficientoneandthenewsystem started itsoperation inDecember2017. Eventhoughourhybridmodelpredictedmuchhigherpower consumptionfor thecoldweather in the thirdweek, theactualpowerconsumptionwasquite lowdue 133