Seite - 376 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1138 Figure 2. Inner structureofGRU,where all arrows represent theweightsbetweengates andunits andtheunitsof f andφare theactivationfunctions. Theparametersareexplainedindetailafter the Equations (4)–(10). whereu is thenumberofupdategatevector; r is thenumberof resetgatevector;h is thenumber ofhiddenvectorsat t timestep;h′ is thenumberofhiddenvectorsat t−1 timestep; f andφare the activationfunctions; f is thesigmoidfunctionandφ is the tanhfunctiongenerally;and s˜th′means the newmemoryofhiddenunitsat t timestep. AccordingtoFigure2, thenewmemory s˜th′ isgeneratedbythe inputx t i at thecurrent timestep andthehiddenunit state st−1h at the last timestep,whichmeans thenewmemorycancombine the new informationand thehistorical information. The reset gatedetermines the importanceof st−1h to s˜th′. If thehistorical information s t−1 h isnot related tonewmemory, the resetgatecancompletely eliminate the information in thepast. Theupdategatedetermines thedegreeof transfer from st−1h to sth. If s t u≈1, st−1h isalmostcompletelypassedto sth. If stu≈0, s˜th′ ispassedto sth. Thestructureshown inFigure2 results ina longmemory inGRUneuralnetworks. Thememorymechanismsolves the vanishinggradientproblemoforiginalRNNs.Moreover, comparedtoLSTMnetworks,GRUneural networksmerge the inputgateandforgetgate,andfuse thecellunitsandhiddenunits inLSTMblock. Itmaintains theperformancewithsimplerarchitecture, lessparametersandlessconvergence time[33]. Correspondingly,GRUneuralnetworksare trainedbyback-propagationthroughtimeasRNNs[35]. 2.2.DataDescription Thereal-world loaddataof individualcustomers inWanjiangarea is recordedfromDongguan PowerSupplyBureauofChinaSouthernPowerGridinGuangdongProvince,Chinaduring2012–2014. ThetopologystructureofWanjiangarea isshowninFigure3. Thereare36 feedersconnectingto the loadsides in theWanjiangarea, i.e., Feeders1—36. Theactivepower isextractedfor loadforecasting from these feeders. The samplingperiod is 15min as themeter recorddata. The load curve of a customer,No. 53990001, fromFeeder2duringamonth isshowninFigure4,where thedifferent load characteristicsof thecustomeroneachdaycanbeconcluded. 376