Page - 150 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 4.2. TemperatureLocations The results for testing the availability of temperature data series fromdifferent locations are included inTable3. Inaddition,Figure9shows theevolutionof theoverallRMSEofbothmodels fromhavingonly location to including all five. Locations are included sequentially frommost to least relevant. Table3.Forecastingerror (RMSE)withavailable temperature locationfrom1to5. TypeofDay MAD MAD,BAR MAD,BAR,VIZ MAD,BAR,VIZ,SEV MAD,BAR,VIZ,SEV,ZAR AR NN AR NN AR NN AR NN AR NN Overall 1.63% 1.61% 1.53% 1.59% 1.48% 1.54% 1.46% 1.54% 1.45% 1.55% Regular 1.59% 1.53% 1.48% 1.50% 1.43% 1.45% 1.41% 1.44% 1.40% 1.44% Special 1.84% 2.22% 1.86% 2.21% 1.81% 2.15% 1.80% 2.17% 1.81% 2.31% Hot 1.83% 2.02% 1.63% 1.91% 1.52% 1.84% 1.55% 1.94% 1.55% 1.93% Cold 2.00% 1.61% 1.81% 1.49% 1.83% 1.47% 1.76% 1.48% 1.72% 1.48% Testconditions: 7YearsTraining(7YT),10N,10RN,12MF,7/14LAG. 1.35% 1.40% 1.45% 1.50% 1.55% 1.60% 1.65% 1 2 3 4 5 Number of Avaliable Locations Accuracy vs temperature availability Figure9.Overall forecastingerror (RMSE)withavailable temperature locationfrom1to5. TheNNoutperformstheARmodelwhenonlyonelocationisavailable. Bothmodelsbenefit from havingmoredataseriesincluded,buttheARmodelobtainsamoreaccurateforecastwithfivelocations. In fact, theNNmodelobtainsa largererrorwithfive locations thanitdoeswith four. Thiscould imply that the linear restrictionontheARmodelallows it tocorrectly include this information in themodel. Theexcessiveavailabilityof information,however, seemsto increase theriskofNNmodeloverfitting the trainingdataand, therefore, losingforecastingcapabilities. 4.3. TemperatureTreatment Thepreprocessingofthetemperaturedataisakeyaspectoftheforecastingsystem.Thethresholds needtobeproperly tunedso that the linearizationof therelation iscorrect.However, these thresholds mayshiftover timeasconsumers’behaviorregardingtemperaturechanges. Therefore, robustness to thisconfiguration isalso important. Theresultswereobtainedusingone locationeachtimeandvaryingHDDandCDDthresholds from13to25 ◦C.Table4showstheoverall results forshifting theHDDthresholdforBarcelonaalong with thehotandcoldcategoriesas thespecialdaysarenot relevant to this test. Theeffect of adjusting the threshold ismore clearly shown inFigure10, inwhich forecasting accuracyofbothmodelsusingtemperature fromZaragozaandBarcelona isplotted. Thegraphshows 150