Seite - 46 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2019,12, 164 costofhighexecution time (slowconvergence rate)due tohighcomplexity.Whereas, theMarkov chain-basedmodels [14–16] have lowexecution time, but at the cost of reduced forecast accuracy. Furthermore, the stochasticdistribution-basedmodels [17–20]need improvement in termsofboth accuracy and execution time. The fuzzyANN-basedmodels [21–26] achievemoderate accuracy, butat thecostofhighexecution time. Finally,hybridANN-basedmodels improve theaccuracyof ANN-basedmodels toanextent,butat thecostofhighexecutiontime.AmongthehybridANN-based models, reference [27] selects featuresviaMI techniqueandANN-basedprediction to forecast the day-ahead load (DAL)of SGs. To improve theaccuracyof [27], theauthors in [28] addaheuristic optimization-basedtechniquewith [27]. Similarly,anotherhybridstrategy ispresented in [29] subject toDALFofSGs.However, reference[27,29]achieverelativelyhighforecastaccuracywhile takinghigh timetoexecute thealgorithm. Furthermore, theforecasterrorof theexistingworks[28,29]significantly increasesduetometeorologicalvariables (suchasdewpoint temperature,drybulb temperature, etc.), andexogenousvariables (suchascultural andsocial events, human impact, etc.). Thus,weaimat improvingthe forecastaccuracyofDALFmodelswithout increasingtheirexecutiontime,andinthe presenceofmeteorologicalandexogenousvariables. In our proposedwork, a hybrid ANN-basedDALFmodel for SGs is presentedwhich is a multi-model forecastingANNwithasupervisedarchitectureandMARAfor training. Theproposed model followsamodular structure (it has three functionalmodules): apre-processor, a forecaster, andanoptimizer. Given thecorrelated lagged loaddataalongwith influentialmeteorological and exogenousvariablesas inputs, thefirstmoduleremoves twotypesof features fromit: (i) redundant; and (ii) irrelevant. Given theselected features, thesecondmoduleemploysANNtopredict future valuesof load. TheANisactivatedbysigmoidfunctionandtheANNis trainedbyMARA.Wefurther minimize the forecast/predictionerrorbyusinganoptimizationmodule inwhichaaheuristics-based optimization technique is implemented. The proposed DALF strategy for SGs is validated via simulationswhichshowthatourproposedstrategyforecaststhefutureloadofSGswithapproximately 98.76%accuracy. Tosumup, thispaperhas the followingcontributions/advantages: • Theproposedmodel takes intoaccountexternalDALFinfluencingfactorssuchasmeteorological andexogenousvariables. • Duetobetteraccuracyandlessexecutiontime,wehaveusedMARAfor trainingwhichnoneof theexistingforecastmodelshasusedfor training. • To improve the forecast accuracyandminimize the executionof the forecastmodel,wehave performedlocal trainingwhichnoneof theexistingforecastmodelshasused. • Wehaveusedourmodifiedversionof theEDEintheerrorminimizationmodule. Theexisting Bi-level strategy[28]hasusedEDEalgorithmintheerrorminimizationmodule. • We have tested our proposed model on the datasets of two USA grids: DAYTOWN and EKPC.Forevaluationandvalidationpurposes,wehavecomparedourproposedmodelwith twoexisting forecastmodels (bi-level forecastandMI+ANNforecast)andprovidedextensive simulationresults. Pleasenote that thiswork iscontinuationofourpreviouswork in[30,31],where inboth[30,31] wehavenotconsideredexogenousandmeteorologicalvariables. Therestof thepaper isorganized as follows. Section2discusses recent/relevantDALFworks, Section3brieflydescribes thenewly proposedANNandmodifiedevolutionaryalgorithm-basedDALFmodel forSGs, simulationresults arediscussed inSection4,andSection5states theconcludingpointsdrawnfromthisworkalongwith futurework. 2.RelatedWork For thesakeofbetterunderstanding, theexisting techniquesarediscussed in twoclasses (linear andnon-linear)accordingto the typeofmodelused[9]. Themodel tobeusedis totally thechoiceof researcherduetospecificdesignconsiderations. 46