Page - 230 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 1893 Ref. [5]mentions thatSwedenandItalyhaveachievedfulldeploymentand[6] that Italiandistribution systemoperatorsareplanningthesecondwaveof roll-outs. Thisresults innewopportunitiessuchas localoptimisationof thegrid,demandsidemanagement andsmart control of storagedevices. Exploiting the smart grid efficiently requires advanceddata analyticsandoptimisationtechniques to improveforecasting,unit commitment,andloadplanning atdifferentgeographical scales.Massivedatasetsareandwillbeproducedasexplainedin[7]: data fromenergyconsumptionmeasuredbysmartmetersatahighfrequency(everyhalfminute insteadof every6months);data fromthegridmanagement (e.g.,PhasorMeasurementUnits);data fromenergy markets (pricesandbidding, transmissionanddistributionsystemoperatorsdata, suchasbalancing andcapacity);data fromproductionunitsandequipments for theirmaintenanceandcontrol (sensors, periodicmeasures...).A lotofeffortsaremadebyutilities todevelopdatalakesandITstructures to gatherandmakethesedataavailable for theirbusinessunits inreal time.Designingnewalgorithms toanalyseandprocess thesedataat scale isakeyactivityandareal competitiveadvantage. We will focus on individual consumption data analysis which plays a major role for energymanagement and electricity load forecasting, designingmarketing offers and commercial strategies,proposingnewservicesasenergydiagnosticsandrecommendations,detectandprevent non-technical losses. 1.2. IndividualElectricalConsumptionData:AState-of-the-Art Individualconsumptiondataanalysis is,accordingtothedevelopmentofsmartmeters,apopular andgrowingfieldofresearch.Composinganexhaustivesurveyofrecentrealizations is thenadifficult challengenotaddressedhere.Asdetailed in [3], individualconsumptiondataanalyticscoversvarious fields of statistics andmachine learning: time series, clustering, outlier detection, deep learning, matrixcompletion,online learningamongothers. Givenadatasetof individualconsumptions,afirstnaturalstepisexploratory: clustering,whichis themostpopularunsupervised learningapproach. Thepurposeofclustering is topartitionadataset intohomogeneoussubsetscalledclusters (see [8]).Homogeneity ismeasuredaccording tovarious criteriasuchas intraandinterclassvariances,ordistance/dissimilaritymeasures. Theelementsofa givenclusterare thenmoresimilar to thoseof thesamecluster thantheelementsof theotherclusters. Timeseriesclustering isanactivesubfieldwhereeach individual isnotcharacterisedbyasetof scalar variablesbutaredescribedbytimeseries, signalsor functions, consideredasawhole,openingtheway forsignalprocessingtechniquesor functionaldataanalysismethods (see [9,10] forgeneral surveys). Clustering methods for electricity load data have been widely applied for profiling or demand responsemanagement. Refs. [11,12] give an overview of the clustering techniques for customergrouping, findingpatterns into electricity loaddataordetectingoutliers andapply it to 400 non-residentialmediumvoltage customers. Clustering canbe seen as longitudinalwhen the objective is tocluster temporalpatterns (e.g.,daily loadcurves) fromasingle individualor transversal whenthegoal is tobuildclustersof customersaccording to their loadconsumptionprofileand/or side information. Themainapplicationof clustering is loadprofilingwhich is essential for energy management,gridmanagementanddemandresponse (see [13]). Forexample, in [14]datamining techniques are applied to extract load profiles from individual load data of a set of low voltage Portuguesecustomers,andthensupervisedclassificationmethodsareusedtoallocatecustomerstothe differentclasses. In[15], loadprofilesareobtainedbyiterativeself-organizingdataanalysisonmetered dataanddemonstratedonasetof660hourlymeteredcustomers inFinland. Ref. [16]proposesan unsupervisedclusteringapproachbasedonk-meansonfeaturesobtainedbyaverageseasonalcurves usingminutemetereddata from103homes inAustin,TX.Correspondencebetweenclusters, their associatedprofiles andsurveydataarealso studied. Authorsof [17] suggest ak-means clustering to derive daily profiles from 220,000 homes and a total of 66millions daily curves in California. Otherapproachesbasedonmixturemodelsarepresented in [18] for customerscategorizationand loadprofilingonadatasetof2613smartmeteredhouseholdfromLondon. 230