Event-Based Machine Learning in NILM: A Hardware-in-the-Loop Approach

Abstract

The transitiontowardintelligentDemand-SideManagement(DSM)reliesheav- ily onaccurate,appliance-levelenergydisaggregation.Whilecontinuous,cloud-based Non-IntrusiveLoadMonitoring(NILM)leveragingDeepLearningoffershightheoret- ical accuracy,itintroducesseverelatency,bandwidthconstraints,anddataprivacy vulnerabilities. Thisthesisproposesaparadigmshifttowardanevent-based,edge- computing methodology,detailingthedesignanddeploymentof“Nesnas”—afully autonomous Edge-AIenergy-monitoringinstrumentcapableofidentifyinghousehold appliances usinglocalizedmachinelearning. Nesnas integratescustomdata-acquisitionhardwaretocapturehigh-precision grid telemetry.Tobypassprocessingbottlenecks,amulti-threadedsoftwarearchi- tecture isolateshardwarepollingfromhigh-levelinference.Thesystemextractskey electrical featuresandbuffersthemintoadeterministictemporalslidingwindow.Gov- erned byalocalizedFiniteStateMachine(FSM),thisarchitectureactsasanintelligent gatekeeper,triggeringacomputationallylightweightRandomForestclassifiertoiden- tify thespecificapplianceonlyupondetectingphysicalstatetransitions. Empirical validationusingmulti-stateinductiveloads(e.g.,hairdryers)against purely resistiveloads(e.g.,kettles)demonstratesthatthelocalizedmachinelearn- ing modelaccuratelyidentifiesappliancesandresolvesclassificationerrorsinherent to continuousenergymonitoring.TheNesnasarchitectureprovesthatdeterministic, zero-latency NILMclassificationcanbeachievedentirelyattheedge,offeringahighly secure, privacy-preservingsolutionformodernSmartGridenvironments.