Seite - 67 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3433 AMIonly, i.e.,withoutexternaldata. Inaddition, the three-stepregularizationprocessremoves the problemofdataprocessing indeeprunningandimprovesLSTM.Theproposedmethodsimulates loadforecastingwithina fewminutes (USTLF) toseveraldays (STLF)usingreal-worldbuildingdata andshowstheadvantages thatLSTMhasover the traditionalmodels. Therestof thepaper isorganizedas follows. Section2 introduces theproposedfeatureextraction methodandprovidesbackgroundinformation. Section3presentsdeep learning. Section4 introduces theexperiments, andSection5presents theanalysis resultsusingtheproposedmulti-decomposition. InSection6, thepredictionresultswithdifferentmodelsarecompared,andSection7summarizesand concludes thepaper. 2.TheProposedMulti-DecompositionforFeatureExtraction 2.1. EnhancedAMI forSmall-ScaleLoadandRealTime Loadforecastingaimstodetermine the futurepowerplanbasedonaseriesofgivenhistorical datasets. Forefficientpowerplanning,aminimumweekly loadmustbepredictedaccordingto the timescaleof the task,e.g.,demandsidemanagement,economicdispatch,andenergyscheduling[2]. As theprediction timescaleandloadscalebecomesmaller, thenon-linearityproblemmustbesolved throughamoresophisticatedpredictionmethod. State-of-the-artAMIwith5-minsamplingprovides moresamplesperhourthanconventional15-minAMI.Asaresult,powerconsumptionmeasurements that are close to real-timemeasurements are achieved. However, as the amountofdata increases, conventionalmachine learningcausesproblemssuchasoverfitting, thevanishinggradientproblem, the long-termdependencyproblem,andincreasedcalculationtimes. 2.2. EmpiricalModeDecomposition Decompositionmethodsarewidelyusedtoanalyzesimilarsignalsandextract features. TheEMD decompositionmethoduses extreme signal values, and theVMDmethoddecomposes the signal byreflectingfrequencycharacteristics tocompensate for theweaknessesEMD.Bothmethodswere employed to analyze time series data in [22]. TheEMDmethodpreprocessesdata by recursively detecting localminimaandmaximainasignalandestimating lowerandupperenvelopesbyspline interpolationof theextremevalues, thenremovinglowerandupperenvelopeaverages. Todecompose a signal into a sum of intrinsic mode functions (IMFs), the following two conditions must be satisfied[18–22]: • In the entire dataset, the number of zero crossingsmust either be equal to ordiffer from the numberofextremabynomore thanone; • The lower andupper envelopemeans, definedby interpolating the local signalminimaand maxima, respectively,mustequalzero. 2.3.VariationalModeDecomposition The goal of VMD is to decompose a signal into a discrete number of sub-signals (modes) thathave specific sparsitypropertieswhile reproducing the signal. VMDreplaces themost recent definitionof IMFs; forexample,anEMDmodeisdefinedasasignalwhosenumberof localextrema andzero-crossingsdiffer atmost byone or asAM-FMsignals by the correspondingnarrowband property [23]. Variationalmodedecompositionprovidesananalyticalexpressionthat relatesAM-FMparameter descriptors to the estimated signal bandwidth, i.e., eachmode k is required tobemostly compact aroundacenterpulsation,wk, that isdeterminedalongwith thedecomposition. This IMFdefinition complements theweaknessofEMDof lackingamathematicaldefinition. VMDalso reducesEMD end-point effectsbecause itdecomposes the signal into kdiscrete IMFs,whereaseach IMF isband limited in thespectraldomain[23–25]. 67