Page - 143 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 2.2.1.HistoricalLoadAvailability Themost important inputofa load-forecastingmodel is itspastbehavior. Apersistentmodel that only takes into accountpreviousvaluesmayprovide, in some cases, a valid baseline to start developingamorecomplexone.However, inmanysituations,andespecially in industryapplications, theavailabilityofsuchhistoricaldata isnotasdeepasdesiredandit is restricteddueto thequantity or thequalityof thestoreddata. Insomecases, thedataacquisitionsystemhasnotbeenrunning long enough,orachange in itsconfigurationmaycauseolddata tobeobsolete. The question of howold thedata thatweuse in our forecasting systemshould be is a valid one. The inclusionofdata fromtoofarbackmaycause themodel to learnobsoletebehavior thathas changedover theyearsandthat isnotcurrentlyaccurate: the incrementofairconditioningsystems mayincrease thesensitivityof loadto temperature increasewhile theuseofmoreefficient lighting maydecrease the load inafter-sunsethours. On theotherhand, there are certainphenomena like extremetemperaturesorspecialdays thatdonothappenfor longperiodsof timeand, therefore, if the database isnotdeepenough, itmaynothaveenoughexamples toshapethis typeofbehavior. Ourresearchproposesusingdata fromthe last3,5and7years to trainbothmodels. Thegoalof theseexperiments is todeterminewhichoneof themrequiresadeeperdatabase,orwhichonecan benefit themost fromsuchdataavailability. Thedatawillbebrokendownintoseparate typesofdays inorder todeterminewhichcategory isaffectedbythiscondition. 2.2.2. TemperatureLocationsAvailability Temperature is themost importantexogenous factor for loadforecastingof regulardaysasboth extremesof the temperaturerange increaseelectricityconsumption. Loadforecastingofsmallareas in whichtemperature ishomogenousmayrequireonlyoneseriesof temperaturedata to learn thearea’s behaviors thatarerelatedto temperature.However, if theregion is largerandtheweatherpresents highervariability, it isnecessary todeterminewhich locationsprovidearelevant temperatureseries thatcouldmodel the localarea’sbehaviorrelatedtoweather.Needless tosay,notall localareaswillbe equallyaffectedbytemperatureandtherelevanceofeachareawithin theoverall loadfor theregion willvarydependingonthe lowerorhigherelectricitycapacityofeacharea. Theelectricitycapacity normallyrelates to thearea’sgrossproduct. Inourcase,Spain isa largecountrywithawideweatherdiversity. Inaddition, thepopulation distribution also causes ahighvariability of power consumptionamongareas. According to this, themodel used at REE includes data fromfive locations that represent the fiveweather regions: North-Atlantic (Bilbao),Mediterranean(Barcelona),Upper-Center (Zaragoza),Lower-Center (Madrid) andSouth (Sevilla). These cities, shown inFigure 2 are themostpowerdemandingareas in each weatherregion. %$5&(/21$ 0$'5,' =$5$*2=$ %,/%$2 6(9,//$ Figure2.Locationof thefivetemperatureseriesanddistributionof theweatherregions inSpain. 143