Seite - 420 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies 2018,11, 213 Figure2.Theone-dimensional (1D)convolutionandpooling layer. The other popular solution of the forecastingproblem is Long Short TermMemorynetwork (LSTM) [33]. The LSTM is a recurrent neural network,which has beenused to solvemany time sequenceproblems. ThestructureofLSTMisshowninFigure3,anditsoperation is illustratedbythe followingequations: ft=σ(Wf · [ht−1,xt]+bf) (1) it=σ(Wi · [ht−1,xt]+bi) (2) C˜t= tanh(WC · [ht−1,xt]+bC) (3) Ct= ft×Ct−1+ it× C˜t (4) ot=σ(Wo · [ht−1,xt]+bo) (5) ht= ot× tanh(Ct) (6) wherext is thenetwork input,andht is theoutputofhiddenlayer,σdenotes thesigmoidal function, Ct is thecell state, and C˜tdenotes thecandidatevalueof thestate. Besides, thereare threegates in LSTM: it is the inputgate, ot is theoutputgate, and ft is the forgetgate. TheLSTMisdesigned for solvingthe long-termdependencyproblem. Ingeneral, theLSTMprovidesgoodforecastingresults. tC Figure3.TheLongShortTermMemorynetwork(LSTM)structure. 3.TheProposedDeepNeuralNetwork Thestructureof theproposeddeepneuralnetworkDeepEnergy is showninFigure4.Unlike the general forecastingmethodbasedontheLSTM,theDeepEnergyuses theCNNstructure. The input layerdenotes the informationonpast load,andtheoutputvaluesrepresent the futureenergy load. Therearetwomainprocesses inDeepEnergy, featureextraction,andforecasting. Thefeatureextraction 420