Seite - 312 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 Table6.Sensitivityanalysies resultsofdifferenthyper-parameters. Metrics TheNumberofSalpPopulations inMOSSA 10 20 30 40 50 60 70 80 90 100 CP 95.21 95.32 97.01 96.66 98.69 98.33 98.50 97.85 97.46 98.10 PINAW 17.34 17.63 13.52 13.89 13.10 13.02 13.64 14.05 13.82 13.72 PINRW 18.28 18.05 14.18 14.35 13.84 13.75 14.36 14.92 14.67 14.53 Time(s) 425 452 472 524 548 593 668 734 869 10.45 Metrics TheInitialThresholdofParameters [−0.5,0.5] [−1,1] [−2,2] [−3,3] [−5,5] CP 97.65 98.84 99.00 98.26 96.89 PINAW 14.36 12.82 12.80 13.12 13.68 PINRW 15.32 13.46 13.42 13.94 14.36 Time(s) 433 450 461 453 484 5.3. ConsistencyAnalysis In this section, inorder toverify theconsistencyofourproposedmodel,newdatasets involving latestdatesare introduced. Inaddition, severalbasic comparedmodels including longshort-term memory (LSTM)networks, functionfittingneural networks (FITNET), and least squares support vectormachine (LSSVM)whichhavebeenprovedtoprovidegoodresults forSTLFareemployedto verify theadvantagesof theproposedmodel. WechoseNSWandVICrandomlyasexamples. Thenewdatasetsarecollectedfrom1January 20180:30amto30May20180:00amand the totalnumberof samples is 7152. The samples in the secondquarter inNSWandthefourthquarter inVICarechosenascompareddatasets.Accordingto theresultsshowninTable7, theproposedmodelalsohasagoodperformanceonthenewdatasets. TheCP is almost 90%,whichmeans the predicted interval can cover 90% target loadvalue. The consistencyof theproposedcanbeguaranteed,andthechangeof thedatesofdatasetwillnot risk alteringthefinalconclusion. Consideringdifferentbasicmodels forSTLF,wechose threewidelyusedartificial intelligence models (LSTM,FITNET,andLSSVM)ascomparators toverify thesuperiority.AsshowninTable7, theproposedmodelsprovidea largerCPandsmallerPINAWcomparedwith theother threemodels. Inparticular,LSTMrevealsdesirednarrowerPINAWandPINRW,but theCPsarenot satisfactory. Moreover, theproposedmodeloutperformedthanotherbasicmodelsinAWD.Therefore, theproposed approachhaveadistinctadvantage in theperformanceofshort-termpower loadinterval forecasting. It isable toprovideasatisfactoryCPandrestrict the intervalwidthat thesametime,which is themost importantaspectof superiorityof theproposedmodel. Table7.Consistencyanalysis resultsof somebasicmodelsandnewdatasets. Models NSW-2018-NEW VIC-2018-NEW CP PINAW PINRW AWD Time(s) CP PINAW PINRW AWD Time(s) Proposed 89.58 15.51 16.58 0.023 593.29 89.08 11.50 12.66 0.065 564.55 LSSVM 78.67 15.95 17.64 0.677 495.32 86.67 12.16 13.01 0.026 486.85 FITNET 72.08 16.25 17.24 0.043 405.52 74.33 11.66 12.95 0.087 300.72 LSTM 44.00 5.47 5.92 0.382 1199.04 59.83 5.28 5.79 0.250 947.78 Models NSW-2017-2Q VIC-2017-4Q CP PINAW PINRW AWD Time(s) CP PINAW PINRW AWD Time(s) Proposed 100.00 16.27 16.66 0.000 543.20 82.08 6.90 7.78 0.001 526.39 LSSVM 94.42 16.67 17.12 0.038 409.24 71.58 7.60 10.59 0.097 435.50 FITNET 94.33 15.83 17.29 0.012 402.12 74.33 7.88 8.94 0.076 504.31 LSTM 70.67 6.01 6.37 0.101 753.60 65.33 3.10 3.55 0.248 732.21 312