Seite - 338 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1282 consideringthetemperatureastheinputvariablemaybenotenough[23–25],andothermeteorological factors such as humidity, visibility and air pressure etc. also should be taken into consideration. Besides, it is necessary toanalyzeandpretreat the influence factorson thepremiseof considering the influence factors synthetically soas toachieve thegoal of improving thegeneralizationability andtheprecisionof thepredictionmodel. Therefore, thispaperappliedfactoranalysis (FA)andthe similar-dayapproach(SDA)for inputdatapre-processing,where the former isutilizedtoextract the latent factors thatessentiallyaffect the loadandtheSDAisadoptedtoexcavate thesimilardays that havecommonfactorswith the forecastday. To sumup, the load forecasting process of the ELMoptimized by the bat algorithm can be elaborated infoursteps. Firstly,basedon22original influence factors, factoranalysis isadoptedto extract the latent factorswhichessentially affect load. To further explore the relationshipbetween historical loadandcurrent load,apartialautocorrelationfunction(PCAF) isappliedtodemonstrate thesignificanceofpreviousdata. Then, inaccordancewith the latent factorsandthe loadsofeachday, antcolonyclustering isusedtodivide the loadtodifferentclusters. Therestof thepaper isorganizedas follows: Section2givesabriefdescriptionabout thematerial andmethods, includingbatalgorithm(BA),extremelearningmachine (ELM),antcolonyclustering algorithm (ACC) aswell as the frameworkof thewholemodel. Data analysis andprocessingare considered in Sections 3 and4whichpresent an empirical analysis of thepower load forecasting. Finally, conclusionsaredrawninSection5. 2.Methodology 2.1. BatAlgorithm Basedontheecholocationofmicro-bats,Yang[26]proposedanewmeta-heuristicmethodand called it thebatalgorithm,onethatcombines theadvantagesboththegeneticalgorithmandparticle swarmoptimizationwith the superiority of parallelism, quick convergence, distribution and less parameter adjustment. In theddimensionsof search spaceduring theglobal search, thebat i has thepositionof xti, andvelocity v t i at the timeof t,whoseposition andvelocitywill beupdatedas Equations (1)and(2), respectively: xt+1i =x t i+v t+1 i ; (1) vt+1i =v t i+ ( xti− xˆ ) ·Fi (2) wherexˆ is thecurrentglobaloptimalsolution;andFi is thesonicwavefrequencywhichcanbeseen inEquation(3): Fi=Fmin+(Fmax−Fmin)β (3) whereβ is a randomnumberwithin [0, 1]; Fmax andFmin are themaxandminsonicwave frequency of thebat I. In theprocessof flying, each initialbat is assignedonerandomfrequency in linewith [Fmin, Fmax]. In local search,onceasolution isselected in thecurrentglobaloptimalsolution,eachbatwould produceanewalternativesolution in themodeofrandomwalkaccordingtoEquation(4): xn(i)=x0+μAt (4) wherex0 isa solution that is chosen incurrentoptimaldisaggregationrandomly;At is theaverage volumeof thecurrentbatpopulation;andμ isaDdimensionalvectorwithin in [−1,1]. Thebalanceof bats is controlledby the impulsevolumeA(i) and impulse emission rateR(i). Oncethebat lockstheprey, thevolumeA(i)willbereducedandtheemissionrateR(i)willbe increased at thesametime. TheupdateofA(i)andR(i)areexpressedasEquations (5)and(6), respectively: At+1(i)=γAt(i) (5) 338