Seite - 291 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1561 empiricalmodedecomposition(EMD)[45]isextensivelyused,whichisanadaptivemethodintroduced toanalyzenon-linearandnon-stationarysignals. Inorder toalleviatesomereconstructionproblems, suchas“modemixing”ofEMD,someotherversions [46–48]areproposed. Particularly, theproblem ofdifferentnumberofmodes fordifferent realizationsofsignalandnoiseneedtobeconsidered. Summinguptheabove, in this study,ahybrid intervalpredictionsystemisproposedtosolve the STLF problem based on themodified Lower andUpper bound estimate (LUBE) technique, by incorporatingtheuseofadatapreprocessingmodule,anoptimizationmodule,andaprediction module. Inorder toverify theperformanceof theproposedmodel,wechooseas theexperimentalcase thepower loadsof fourstates inAustralia. Theelicitedresultsarecomparedwith those frombasic benchmarkmodels. Insummary, theprimarycontributionsof this studyaredescribedbelow: (1) AmodifiedLUBEtechnique isproposedbasedona recurrentneural network,which isable to consider previous information of former observations in STLF.The contest layer of ENNcan store theoutputsofa formerhidden layer, andthenconnect the input layer in thecurrentperiod. Comparison of the traditional interval predictivemodelwith the basic neural network, this mechanismcanimprovetheperformanceof timeseries forecastingmethods, suchasSTLF. (2) Amore convincingoptimization techniquebasedonmulti-objectiveoptimization isproposed forLUBE. InLUBE,besidesCP,PIWshouldalsobeconsidered in theconstructionof thecost function. In thisstudy, thenovelmulti-objectiveoptimizationmethodMOSSAisemployedin theoptimizationmodule tobalance theconflictbetweenhigherCPandlowerPIW,andto train theparameters inENN.With thismethod, thestructureofneuralnetworkscanprovideabetter performance in intervalprediction. (3) A novel and efficient data preprocessing method is introduced to extract the valuable information fromrawdata. Inorder to improve thesignalnoise ratio (SNR)of the inputdata, anefficientmethodisusedtodecomposetherawdata intoseveralempiricalmodal functions (IMFs).Accordingto theentropytheory, the IMFswith littlevaluable informationare ignored. Theperformanceof theproposedmodel trainedwithprocesseddata improvessignificantly. (4) The proposed hybrid system for STLF canprovide powerful theoretical andpractical support for decisionmaking andmanagement in power grids. This hybrid system is simulated and tested depending on the abundant samples involving different regions and different times, which indicate its practicability and applicability in the practical operations of power grids comparedtosomebasicmodels. The rest of this study is organized as follows: The relevant methodology, including data preprocessing, Elman neural network, LUBE, andmulti-objective algorithms, are introduced in Section2. Section3discussesourproposedmodel indetail. The specific simulation, comparisons andanalysesof themodelperformancesareshowninSection4. Inorder to furtherunderstandthe featuresof theproposedmodel, severalpointsarediscussed inSection5.Accordingto theresultsof ourresearch, conclusionsareoutlined inSection6. 2.Methodology In this section, the theory of the hybrid interval prediction model is elaborated, and the methodologyof the components inhybridmodels, including complete ensemble empiricalmode decomposition with adaptive noise (CEEMDAN), Elman neural networks, LUBE, and MOSSA, areexplainedindetail. 2.1.DataPreprocessing TheEMDtechnique [45]usuallydecomposesasignal into severalnumbersof IMFs. Foreach IMF, the series have to fulfill two conditions: (i) the number of extrema (maxima andminima) and thenumber of zero-crossingsmust be equal or differ atmost byone; and (ii) the localmean, definedas themeanof theupper and lower envelopes,must be zero. In order to alleviatemode 291