Page - 121 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 forconstructingourhybridelectrical loadforecastingmodel indetail. InSection4,wedescribeseveral metrics forperformanceacomparisonof loadforecastingmodels. InSection5,wedescribehowto evaluate theperformanceofourmodelviaseveralexperimentsandshowsomeof theresults. Lastly, inSection6,webrieflydiscuss theconclusion. 2.RelatedWork So far, many researchers have attempted to construct STLF using variousmachine learning algorithms. Vrablecová et al. [7] developed the suitability of an online support vector regression (SVR)methodtoshort-termpower loadforecastingandpresentedacomparisonof10state-of-the-art forecastingmethods in termsofaccuracy for thepublic IrishCommissionforEnergyRegulation (CER) dataset. TsoandYau[26]conductedweeklypowerconsumptionpredictionforhouseholds inHong Kongbasedonanartificial neural network (ANN),multiple regression (MR), andadecision tree (DT).Theybuilt the inputvariablesof theirpredictionmodelbysurveyingtheapproximatepower consumption fordiverse electronic products, such as air conditioning, lighting, anddishwashing. Jainetal. [27]proposedabuildingelectrical loadforecastingmodelbasedonSVR.Electrical loaddata werecollectedfrommulti-familyresidentialbuildings locatedat theColumbiaUniversitycampus in NewYorkCity.Grolingeretal. [28]proposedtwoelectrical loadforecastingmodelsbasedonANN andSVRtoconsiderbothevents andexternal factors andperformedelectrical load forecastingby day,hour, and15-min intervals fora largeentertainmentbuilding. Amberetal. [29]proposed two forecastingmodels,geneticprogramming(GP)andMR, to forecast thedailypowerconsumptionof anadministrationbuilding inLondon.Rodriguesetal. [30]performedforecastingmethodsofdaily andhourlyelectrical loadbyusingANN.Theyusedadatabasewithconsumptionrecords, loggedin 93realhouseholds inLisbon,Portugal. Efendietal. [31]proposedanewapproachfordeterminingthe linguisticout-sample forecastingbyusingthe indexnumbersof the linguisticsapproach. Theyused thedaily loaddata fromtheNationalElectricityBoardofMalaysiaasanempirical study. Recently, ahybridprediction schemeusingmultiplemachine learningalgorithmshas shown a better performance than the conventional prediction scheme using a single machine learning algorithm [14]. The hybridmodel aims to provide the best possible prediction performance by automaticallymanagingthestrengthsandweaknessesofeachbasemodel. Xiaoetal. [18]proposed twocombinationmodels, thenonegative constraint theory (NNCT) and the artificial intelligence algorithm,andshowedthat theycanalwaysachieveadesirable forecastingperformancecomparedto the existing traditional combinationmodels. Jurado et al. [24] proposed a hybridmethodology that combines feature selection based on entropies with soft computing and machine learning approaches (i.e., fuzzy inductivereasoning, randomforest, andneuralnetworks) for threebuildings in Barcelona.Abdoosetal. [20]proposedanewhybrid intelligentmethodforshort-termloadforecasting. Theydecomposedtheelectrical loadsignal into twolevelsusingwavelet transformandthencreated the training inputmatrices using thedecomposed signals and temperaturedata. After that, they selected thedominant features using theGram–Schmidtmethod to reduce thedimensions of the inputmatrix. They used SVMas the classifier core for learning patterns of the trainingmatrix. Dongetal. [21]proposedanovelhybriddata-driven“PEK”model forpredicting thedaily total load of thecityofShuyang,China. Theyconstructedthemodelbyusingvarious functionapproximates, includingpartialmutual information(PMI)-based inputvariableselection,ensembleartificialneural network (ENN)-basedoutput estimation, andK-nearest neighbor (KNN) regression-basedoutput error estimation. Lee andHong [22] proposed a hybridmodel for forecasting the electrical load severalmonthsaheadbasedonadynamic (i.e., air temperaturedependencyofpower load)anda fuzzy time series approach. They tested their hybridmodelusingactual loaddataobtained from theSeoulmetropolitanarea, andcompared itspredictionperformancewith thoseof theother two dynamicmodels. Previousstudiesonhybridforecastingmodelscompriseparameterselectionandoptimization technique-basedcombinedapproaches. Thisapproachhas thedisadvantages that it isdependenton 121