Seite - 149 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 2080 4.Results Theresultsexpressed in this sectioncorrespondto the forecastingperiodof2017. Eachsubsection presents theaccuracyofboth techniques (ARandNN)whenthecorrespondentparameterorexternal condition changes. In addition, these results have been analyzedunder the categories described inSection2.3. 4.1.HistoricalLoadAvailability The results shown inTable 2 represent the effect of increasing the number of previous years considered in the trainingof themodel from3to7 forbothmodels. Theresults showagenerallymore accurateperformancebytheARmodelespeciallywith feweryearsofdata (1.50%vs. 2.17%). TheNN model,however,benefitsmore fromtheavailabilityofmoredataandthisdifference is reducedto0.1% whensevenyearsareused. TheARmodel showsvery little improvement from3 to7yearswhile theNNmodelappears tobeable tobenefit fromevenlonger trainingdataas itsperformanceonall categoriescontinues to improvefrom5to7years (seeFigure8).Unfortunately, theavailabledatabase isnotyetdeepenoughto test this. Table2.Forecastingerror (RMSE)with trainingperiods from3to7years. TypeofDay 3-Years 5-Years 7-Years AR NN AR NN AR NN Overall 1.50% 2.17% 1.52% 1.72% 1.45% 1.55% Regular 1.44% 1.96% 1.47% 1.57% 1.40% 1.44% Special 1.91% 3.62% 1.81% 2.71% 1.81% 2.31% Hot 1.63% 2.65% 1.53% 2.08% 1.55% 1.93% Cold 1.61% 2.79% 1.73% 1.81% 1.72% 1.48% Testconditions: 10neurons (10N),10redundantnetworks (10RN),5 temperature locations (5TL),12monthtraining freq(12MF),7 lags forARand14forNN(7/14LAG). 0.00% 0.50% 1.00% 1.50% 2.00% 2.50% 3-yrs 5-yrs 7-yrs Accuracy vs training period length AR NN Figure8.Overall forecastingerror (RMSE)with trainingperiods from3to7years. Regardingthecategorizedresults, regulardaysobtainalmost thesameresultwhile inhotand special days theARoutperforms theNNmodel. However, colddays are clearly forecastedmore accuratelybytheNNmodel. Thiscouldimplythat the linearrestrictionpresent intheARmodel limits its capacity tomodel thebehaviorof the loadwith thedata treatmentused. 149