Research on Smart Grids has recently focused on the energy monitoring issue, in which one of the hottest topic is represented by Non-Intrusive Load Monitoring (NILM): it refers to ecomposing the consumption aggregated data acquired at a single point of measurement into the consumption profiles of appliances. This work reports an up-to-date state of the art of most promising NILM methods, with an overview of the public available dataset used. Within all the proposed methods, the Hidden Markov Model (HMM) based and the Deep Neural Network (DNN) based ones have been detected as the most performing. In the HMM based approaches, the Additive Factorial Approximate MAP (AFAMAP) algorithm is nowadays regarded as a reference model. The AFAMAP algorithm has been extended, by means of a differential forward model. In a second step, an alternative formulation of the same algorithm is presented, in order to deal with bivariate HMM, whose emitted symbols are the joint active-reactive power signals. The experiments are conducted on the AMPds dataset, in noised and denoised conditions. Additionally, a user-aided footprint extraction procedure is presented in real scenario. In the DNN based approaches, the Denoising Autoencoder (dAE) represents one of the most performing approaches. In this work, this method is extended and improved by conducting a detailed study on the topology of the network. The experiments have been conducted on the AMPds, UK-DALE, and REDD datasets in seen and unseen scenarios both in presence and in absence of noise. Furthermore, the same method is explored when the input size is increased, including the reactive power component near the active power consumption. Finally, similar computational intelligence techniques are applied in other field, i.e. the smart water and gas grid, and audio application.
La ricerca sulle Smart Grids si è concentrata sulla questione del monitoraggio energetico, in cui il Non-Intrusive Load Monitoring (NILM) rappresenta uno degli argomenti di maggiore interesse: si riferisce alla scomposizione dei dati aggregati di consumo acquisiti in un singolo punto di misurazione nei profili degli elettrodomestici. Questo lavoro riporta uno stato dell'arte aggiornato dei metodi più performanti, con una panoramica dei dataset pubblici disponibili. Tra tutti i metodi proposti, quelli basati su Hidden Markov model (HMM) e su Deep Neural Network (DNN) risultano tra i più performanti. Tra gli approcci basati su HMM, l'algoritmo Additive Factorial Approximate MAP (AFAMAP) è considerato come un modello di riferimento. L'algoritmo AFAMAP è stato esteso, per mezzo di un modello differenziale in avanti. In una seconda fase, viene presentata una formulazione alternativa dello stesso algoritmo, al fine di trattare con HMM bidimensionali, i cui simboli emessi sono i segnali di potenza reattivi attivi congiunti. Gli esperimenti sono condotti sul dataset AMPds, in condizioni di assenza e presenza di rumore. Inoltre, una procedura agevolata di estrazione dell'impronta è presentato in uno scenario reale. Tra gli approcci basati su DNN, il Denoising Autoencoder (dAE) rappresenta uno degli approcci più performanti. In questo lavoro, questo metodo è esteso e migliorato conducendo uno studio dettagliato sulla topologia della rete. Gli esperimenti sono stati condotti su AMPds, UK-DALE e REDD in scenari seen ed unseen in presenza e in assenza di rumore. Inoltre, lo stesso metodo è esplorato quando la dimensione dell'ingresso viene aumentata, includendo la componente di potenza reattiva di consumo di energia. Infine, tecniche simili di intelligenza computazionale sono applicate in altri campi, ossia nella Smart Grid per la distribuzione idrica e gas e in applicazioni audio.
Machine Learning approaches for Non-Intrusive Load Monitoring / Bonfigli, Roberto. - (2018 Mar 26).
Machine Learning approaches for Non-Intrusive Load Monitoring
Bonfigli, Roberto
2018-03-26
Abstract
Research on Smart Grids has recently focused on the energy monitoring issue, in which one of the hottest topic is represented by Non-Intrusive Load Monitoring (NILM): it refers to ecomposing the consumption aggregated data acquired at a single point of measurement into the consumption profiles of appliances. This work reports an up-to-date state of the art of most promising NILM methods, with an overview of the public available dataset used. Within all the proposed methods, the Hidden Markov Model (HMM) based and the Deep Neural Network (DNN) based ones have been detected as the most performing. In the HMM based approaches, the Additive Factorial Approximate MAP (AFAMAP) algorithm is nowadays regarded as a reference model. The AFAMAP algorithm has been extended, by means of a differential forward model. In a second step, an alternative formulation of the same algorithm is presented, in order to deal with bivariate HMM, whose emitted symbols are the joint active-reactive power signals. The experiments are conducted on the AMPds dataset, in noised and denoised conditions. Additionally, a user-aided footprint extraction procedure is presented in real scenario. In the DNN based approaches, the Denoising Autoencoder (dAE) represents one of the most performing approaches. In this work, this method is extended and improved by conducting a detailed study on the topology of the network. The experiments have been conducted on the AMPds, UK-DALE, and REDD datasets in seen and unseen scenarios both in presence and in absence of noise. Furthermore, the same method is explored when the input size is increased, including the reactive power component near the active power consumption. Finally, similar computational intelligence techniques are applied in other field, i.e. the smart water and gas grid, and audio application.File | Dimensione | Formato | |
---|---|---|---|
Tesi_Bonfigli.pdf
accesso aperto
Descrizione: Tesi_Bonfigli.pdf
Tipologia:
Tesi di dottorato
Licenza d'uso:
Creative commons
Dimensione
99.14 MB
Formato
Adobe PDF
|
99.14 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.