Page - 289 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 time.Amongthesecategories, short-termloadforecasting(STLF) isanessential tool for theplanning andoperation [1,2]ofenergysystemsandithas thusbeenamajorareaof researchduring thepast fewdecades. According to existing research, concern mostly focuses on the point forecasting of STLF. Additionally, therelativealgorithmscanbemainlyclassifiedinto threemajorcategories: traditional statistical techniques, computational intelligentmethods,andmultimodulehybridmodels [3]. In theearlystagesof research, traditional statistical techniqueswereextensivelyemployedfor point forecastingofSTLF, suchas linear regressionmethods [4,5], exponential smoothing[6],Kalman filters [7], andother linear time-seriesmodels. Ingeneral,mostof the traditional statisticalapproaches havebeeninvolvedinlinearanalysisandhavemainlyconsideredlinearfactorsintimeseries.However, theshort-termloadseriesareamixtureofmultiplecomponentswhich include linearandnon-linear factors. Therefore, the traditional statisticalapproachesencounterdifficultieswhendealingwith the STLF,andthe forecastingaccuracy isoftenunsatisfactory.With thedevelopmentofmachine learning andartificial intelligence,an increasednumberofnon-linearcomputational intelligentmethodshave beenapplied toSTLF, suchasneuralnetworkmodels (NN) [8,9], expert systems [10] andsupport vectormachines (SVM)[11,12]. Theseapproacheshavebeenprovedtohaveadvantages indealing with thenon-linearproblemsofSTLFcomparedto traditional statisticalmethods, therebyeliciting improvedperformances inmostcases.Most importantly,akeypoint that influences theperformance ofcomputational intelligentmethods is thesettingofrelatedparameters inalgorithms.At this time, efficienthybridmodelsappeared. Inhybridmodels,differentmoduleswere introducedto improve theperformanceandaccuracyofSTLF[13–19].Amongexistingreviews in the literature, twopopular andefficientmodules include thedatapreprocessingandoptimizationmodules. In thecaseof the datapreprocessingmodules,amultiwavelet transformwasusedincombinationwithathree-layer feed-forwardneuralnetworktoextract thetrainingdataandpredict theloadin[13]. Fanetal. [14]used empiricalmodedecomposition (EMD) todecomposeelectric loaddata,generatinghigh-frequency seriesandresiduals for the forecastingofsupportvectorregression(SVR)andautoregression(AR), respectively. Theresultsshowedthatthehybridmethodscanperformwellbyelicitinggoodforecasting accuracyandinterpretability. In thecaseof theoptimizationmodules,AlRashidietal. [15]employed theparticleswarmoptimizer (PSO) tofine-tunethemodelparameters,andtheforecastingproblem waspresented inastatespace form.Wangetal. [16]proposedahybrid forecastingmodelcombining differential evolution (DE)andsupportvector regression (SVR) for load forecasting,where theDE algorithmwasusedtochoose theappropriateparameters forSVR. However, asmentioned above, the current research on STLFmainly concentrates on point forecasting inwhich theaccuracy isusuallymeasuredbytheerrorsbetweenthepredictedandthe targetvalues.Withpowersystemgrowthandthe increase in itscomplexity,point forecastingmight notbeable toprovideadequate informationsupport forpowersystemdecisionmaking.Anincreasing numberof factors, suchas loadmanagement, energyconversion, spotpricing, independentpower producersandnon-conventionalenergy,makepoint forecastingundependable inpractice. Inaddition to the fact thatmostof thesepoint forecastingmodelsdonotelicit the requiredprecision, theyare alsonotadequatelyrobust. Theyfail toyieldaccurate forecastswhenquickexogenouschangesoccur. Othershortcomingsarerelatedtonoise immunity,portability,andmaintenance [20]. Ingeneral,point forecastingcannotproperlyhandleuncertaintiesassociatedwith loaddatasets inmost cases. To avoid such imperfection, interval prediction (IP) of STLF is an efficientway to dealwith the forecast uncertainty in electrical power systems. Prediction intervals (PIs) not only providearange inwhich targetsarehighly likely tobecovered,but theyalsoprovidean indicationof theiraccuracy,knownas thecoverageprobability. Furthermore, thePIscantake intoaccountmore uncertain informationand the result of (PIs) commonly formadoubleoutput (upperboundsand lowerbounds)whichcanreflectmoreuncertain informationandprovideamoreadequatebasis for powersystemplanning. 289