Page - 147 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 discardedaseitherhotor cold. Alldays thatdonotbelong tooneof thecategories (special, hotor cold)areconsideredasregulardays. 3.DataAnalysis It is important todescribe thecharacteristicsof thedataseries relevant to the forecastingprocess inorder tounderstand the forecastingproblemandwhetherornot its conclusionsmayapply toa differentcase: 3.1. Load The load data series covers from 2010 to 2017 and it includes hourly values of electricity consumption in theSpanish inlandsystem. The long-termtrendof the series shown inFigure5 is related toeconomicgrowth,efficiency improvementsandbehavioral shifts like theuseofACsystems. Onashorter termscale, the factorsdrivingthe loadinSpainare temperatureandsocialevents andholidays,whichareexplainedin the followingsubsections. 2004 2006 2008 2010 2012 2014 2016 2018 0 0.2 0.4 0.6 0.8 1 Evolution of national electricity demand and Gross National Product Normalized GNP Normalized load Figure5. Evolutionof 52weeksmovingaverage loadandGrossNationalProduct. Both series are normalized[0,1]. 3.2. Temperature The temperature data available includes series from59 stations scattered across the country. Realdataofdailymaximumandminimumdata iscollectedalongwithdaily forecastsofupto ten daysahead. Therefore, it ispossible tosimulatereal timeconditions if forecastsareusedinsteadof realdata. As it was explained before, the national forecast only uses information from five stations selectedfromthe59available. This selection ismadethroughanempiricalevaluation. Inaddition, the temperature fromupto fourpreviousdays isalsoused inorder tocapture thedynamicsof the temperature-load relation. Thenon-linearity of the relation ismodeledusing theCDDandHDD approachalreadydiscussed. Figure6showsthescatterplotofnational loadat18honweekdayagainst temperatureat the threemost relevant locations. TheHDDandCDDlinearization isalsoplottedfor each locationalongwith theMeanAveragePercentageError (MAPE)betweentheactual loadandthe linearizedone. 147