Seite - 418 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 213 managed.Accordingto theMckinseyGlobal Institute [14], theAIcouldbeapplied in theelectricity industry forpowerdemandandsupplyprediction,becauseapowergrid loadforecastaffectsmany stakeholders. Based on the short-term forecast (1–2 days ahead), power generation systems can determinewhichpowersources toaccess in thenext24h,andtransmissiongridscantimelyassign appropriate resources to clients basedon current transmission requirements. Moreover, using an appropriatedemandandsupply forecast, electricity retailers can calculate energypricesbasedon estimateddemandmoreefficiently. The powerful data collection and analysis technologies are becomingmore available on the market, sopowercompaniesarebeginningtoexplorea feasibilityofobtainingmoreaccurateresults usingAI inshort-termloadforecasts. For instance, in theUnitedKingdom(UK), theNationalGrid iscurrentlyworkingwith theDeepMind[15,16],aGoogle-ownedAIteam,which isusedtopredict thepowersupplyanddemandpeaks in theUKbasedonthe informationfromsmartmetersandby incorporatingweather-relatedvariables. Thiscooperationtends tomaximize theuseof intermittent renewable energyandreduce theUKnational energyusageby10%. Therefore, it is expected that electricitydemandandsupplycouldbepredictedandmanagedinreal timethroughdeeplearning technologiesandmachines,optimizing loaddispatch,andreducingoperationcosts. The loadforecastingcanbecategorizedbythe lengthof forecast interval.Althoughthere isno official categorization in thepower industry, thereare four loadforecasting types [17]: veryshort term loadforecasting(VSTLF), short termloadforecasting(STLF),mediumtermloadforecasting(MTLF), andlongtermloadforecasting(LTLF).TheVSTLFtypicallypredicts loadforaperiod less than24h, STLFpredicts load foraperiodgreater than24huptooneweek,MTLFforecasts loadforaperiod fromoneweekuptooneyear,andLTLFforecasts loadperformanceforaperiodlonger thanoneyear. Theloadforecastingtypeischosenbasedonapplicationrequirements.Namely,VSTLFandSTLFare appliedtoeverydaypowersystemoperationandspotpricecalculation,sotheaccuracyrequirement ismuchhigher thanfora longtermprediction.TheMTLFandLTLFareusedforpredictionofpower usageovera longperiodof time,andtheyareoftenreferencedinlong-termcontractswhendetermining systemcapacity, costsofoperationandsystemmaintenance,andfuturegridexpansionplans. Thus, if thesmartgridsareintegratedwithahighpercentageofintermittentrenewableenergy, loadforecasting willbemore intense thanthatof traditionalpowergenerationsourcesduetothegridstability. Inaddition, the loadforecastingcanbeclassifiedbycalculationmethodintostatisticalmethods andcomputational intelligence (CI)methods.With recentdevelopments in computational science andsmartmetering, the traditional load forecastingmethodshavebeengradually replacedbyAI technology. Thesmartmeters forresidentialbuildingshavebecomeavailableonthemarketaround 2010,andsince then,variousstudiesonSTLFfor residential communitieshavebeenpublished[18,19]. Whencomparedwiththetraditionalstatisticalforecastingmethods,theabilitytoanalyzelargeamounts ofdata inaveryshort timeframeusingAI technologyhasdisplayedobviousadvantages [10]. Someof frequentlyusedloadforecastmethods include linear regression[5,6,20], autoregressive methods[7,21], andartificialneuralnetworks [9,22,23]. Furthermore, clusteringmethodswerealso proposed[24]. In [20,25] similar timesequenceswerematchedwhile in [24] the focuswasoncustomer classification. A novel approach based on the support vectormachinewas proposed in [26,27]. Theother forecastingmethods, suchasexponential smoothingandKalmanfilters,werealsoapplied infewstudies [28].Acareful literaturereviewof the latestSTLFmethodcanbefoundin[8]. In [13], itwas shownthat accuracyofSTLF is influencedbymany factors, suchas temperature, humidity, windspeed,etc. Inmanystudies, theartificialneuralnetwork(ANN)forecastingmethods[9–11,29] havebeenproventobemoreaccurate thantraditional statisticalmethods,andaccuracyofdifferent ANNmethodshasbeen reviewedbymany researchers [1,30]. In [31], amulti-model partitioning algorithm(MMPA)forshort-termelectricity loadforecastingwasproposed.Accordingto theobtained experimental results, theMMPAmethod is better thanautoregressive integratedmovingaverage (ARIMA)method. In [17], authorsused theANN-basedmethod reinforcedbywavelet denoising algorithm.Thewaveletmethodwasusedtofactorizeelectricity loaddata intosignalswithdifferent 418