Page - 19 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2226 mapping function todeterminemoreappropriateparametersof anLS-SVRmodel to improve the forecastingaccuracy. Comparing theLS-SVR-QFOAmodelwith theLS-SVR-FOAmodel, the forecasting accuracy of theLS-SVR-QFOAmodel is superior to thatof theLS-SVR-FOAmodel. Thisdemonstrates that theQCMempowers the fruitflytohavequantumbehaviors, i.e., theQFOAfindmoreappropriate parametersofanLS-SVRmodel,which improves the forecastingaccuracyof theLS-SVR-FOAmodel inwhich theFOAishybridizedwithanLS-SVRmodel. Forexample, inTable4, theusageof theQCM inFOAchanges the forecastingperformances (RMSE=15.93,MAPE=2.48%,MAE=15.63)of the LS-SVR-FOAmodel to themuchbetterperformance (RMSE=14.87,MAPE=2.32%,MAE=14.61) of theLS-SVR-QFOAmodel. Similar results aredemonstrated in theGEFCom2014 (Jan.) and the GEFCom2014(July) fromTables5and6, respectively. ForforecastingperformancecomparisonbetweentheLS-SVR-CQFOAandLS-SVR-QFOAmodels, thevaluesofRMSE,MAPE,andMAEfor theLS-SVR-CQFOAmodelaresmaller thanthoseof the LS-SVR-QFOAmodel. This reveals that the introductionofcatchaoticmappingfunction intoQFOA playsapositiverole insearchingappropriateparameterswhenthepopulationofQFOAalgorithmis trappedinto localoptima. Then, theCQFOAfindsmoreappropriateparameters.Asaresult, asshown inTable4,employingCQFOAtoselect theparameters foranLS-SVRmodelmarkedly improves the performance (RMSE=14.87,MAPE=2.32%,MAE=14.61)of theLS-SVR-QFOAmodel to themuch betterone (RMSE=14.10,MAPE=2.21%,MAE=13.88)of theLS-SVR-CQFOAmodel. Similar results are illustrated in theGEFCom2014(Jan.) andtheGEFCom2014 (July) fromTables5and6, respectively. Comparing the time-consumingproblemduring theparameter searchingprocesses in all the IDAS2014, theGEFCom2014 (Jan.), andtheGEFCom2014 (July)datasets, theproposedCQFOAis less thanthatof theCQGAandCQBAalgorithms,butmore thanthatof theCQPSOandCQTSalgorithms. However, consideringthe timerequirementsof theactualapplication, the increase in timecompared withCQPSO(more than7s)andCQTS(more than23s) isacceptable. Finally, some limitations should be noticed. This paper only employs an existing dataset to establishtheproposedmodel; thus, fordifferentseasons,months,weeks,anddates, theelectricity load patternsshouldbechangedseasonbyseason,monthbymonth,andweekbyweek. Forreal-world applications, thispapershouldbeagoodbeginningtoguideplannersanddecision-makerstoestablish electricity load forecastingmodels overlapping the seasons,months, andweeks to achievemore comprehensive results. Thus, our planned future research direction is to explore the feasibility of hybridizingmore powerful novel optimization frameworks (e.g., chaotic mapping functions, quantum computingmechanism, and hourly, daily, weekly, monthly adjustedmechanism) and novelmeta-heuristicalgorithmswithanLS-SVRmodel toovercomethedrawbacksofevolutionary algorithmstoachieveexcellent forecastingaccuracy. 5.Conclusions Thispaperproposesanovelhybrid forecastingmodelbyhybridizinganLS-SVRmodelwith theQCM, the cat chaoticmapping function, and the FOA. The forecasting results show that the proposedmodelachievesbetterperformance thanthealternative forecastingmodels,byhybridizing chaoticmapping function,QCM,andotherevolutionaryalgorithmswithanLS-SVR-basedmodel. Employingthecatchaoticmappingfunctiontoenrichthediversityofsearchingscopeandenhancethe ergodicityof thepopulationcouldsuccessfullyavoid trapping into localoptima,and,alsoproves that applyingQCMtoovercomethe limitationsof the fruitfly’ssearchingbehaviorsempowers the fruit flytoundertakequantumsearchingbehaviors, therebyachievingmoresatisfactoryresults forMEL forecasting. Theglobalchaoticperturbationstrategybasedonthecatmappingfunctionisemployedto jumpoutof localminimawhile thepopulationofQFOAsuffers fromprematureconvergence,andalso helps to improvethe forecastingperformance. AuthorContributions:M.-W.L.andW.-C.H.conceivedanddesignedtheexperiments;G.J. andZ.Y.performed theexperiments;M.-W.L.andW.-C.H.analyzedthedataandwrote thepaper. 19