The present thesis is focused on short-term prediction of air-conditioning (AC) loads of residential building by only employing the data obtained from a conventional smart meter. Accordingly, the AC load is determined at each time stamp through energy disaggregation and the obtained values are next employed for short-term prediction of AC loads. In the disaggregation procedure, also called Non-Intrusive Load Monitoring, the cumulative power consumption of the building, obtained from the smart meter, is disaggregated into appliance-by-appliance consumption. The power consumption corresponding to the AC equipment, determined in the previous step, along with the available weather data are utilized in the prediction step as input features. Accordingly, machine-learning algorithms are employed in the second step to conduct hour-ahead and day-ahead AC load predictions utilizing those features. Combinatorial Optimization was used for conducting the disaggregation step while the prediction step is carried out utilizing several machine-learning algorithms including Artificial Neural Networks (ANNs), Support Vector Machines (SVMs) and Random Forests Decision Trees. The main advantage of the present methodology with respect to conventional prediction algorithms is separating the AC consumption from the other devices. The harsh alterations in the consumption of AC equipment can be predicted employing the short-term weather forecast while the consumption of other residential appliances is largely dependent upon the occupant’s behaviour, which is more difficult to predict without intrusive methods. Hence, this methodology facilitates predicting notable changes in the AC consumption, with a higher accuracy, only employing the aggregate consumption measured by a smart meter. However, it should be noted that a very short measurement period, with additional meters for the main devices, is needed; the requirement that can eventually be evaded once abundant measured data including different combinations of various residential devices is available. The mentioned prediction enables the utility companies and grid managers to estimate the loads of residential buildings, which contribute a significant portion of energy demand and better manage the energy dispatch.
La presente tesi è focalizzata sulla previsione a breve termine dei carichi della climatizzazione degli edifici residenziali solo utilizzando i dati ottenuti da un contatore intelligente convenzionale. Il carico del condizionatore viene determinato, ad ogni istante, attraverso la disaggregazione energetica e i valori ottenuti vengono successivamente utilizzati per la previsione a breve termine di suddetti carichi. Nella procedura di disaggregazione, anche chiamata monitoraggio non intrusivo del carico, il consumo cumulativo di energia dell'edificio, ottenuto dal contatore intelligente, viene disaggregato in consumi separati per dispositivo. Il consumo di energia corrispondente al condizionatore, determinato nella fase precedente, insieme ai dati meteo disponibili, viene utilizzato nella fase di previsione come dati di input. Successivamente, gli algoritmi di Machine Learning vengono impiegati per svolgere le previsioni di carichi della climatizzazione dell'ora e del giorno successivi. L'ottimizzazione combinatoria (combinatorial optimization) è stata utilizzata per svolgere la fase di disaggregazione mentre il passaggio di previsione viene eseguito utilizzando diversi algoritmi di Machine Learning, tra cui “Artificial Neural Networks (ANNs)”, “Support Vector Machines” (SVM) e “Random Forests Decision Trees”. Il principale vantaggio della metodologia utilizzata in questa tesi rispetto agli algoritmi di previsione convenzionali, è la separazione del consumo del condizionatore dagli altri dispositivi. Le notevoli variazioni di consumo del condizionatore possono essere prevedute utilizzando la previsione meteorologica a breve termine mentre il consumo degli altri elettrodomestici dipende in gran parte dal comportamento dell'inquilino che è più difficile da prevedere. Questa metodologia consente quindi di prevedere notevoli cambiamenti nel consumo del condizionatore, con una precisione più elevata, utilizzando solo il consumo aggregato misurato da un contatore intelligente. Occorre tuttavia notare che è necessario un periodo di misurazione molto breve, con contatori per i dispositivi principali. La misurazione è un requisito che può eventualmente essere evitato una volta che sono disponibili dati abbondanti misurati che includono diverse combinazioni di vari dispositivi residenziali. La suddetta previsione consente alle aziende di servizi di rete e ai gestori di rete di valutare i carichi di edifici residenziali che contribuiscono ad una quota significativa della domanda energetica e gestiscono meglio la rete.
Machine learning based short-term prediction of air-conditioning loads through smart meter analytics
MANIVANNAN, MANOJ
2016/2017
Abstract
The present thesis is focused on short-term prediction of air-conditioning (AC) loads of residential building by only employing the data obtained from a conventional smart meter. Accordingly, the AC load is determined at each time stamp through energy disaggregation and the obtained values are next employed for short-term prediction of AC loads. In the disaggregation procedure, also called Non-Intrusive Load Monitoring, the cumulative power consumption of the building, obtained from the smart meter, is disaggregated into appliance-by-appliance consumption. The power consumption corresponding to the AC equipment, determined in the previous step, along with the available weather data are utilized in the prediction step as input features. Accordingly, machine-learning algorithms are employed in the second step to conduct hour-ahead and day-ahead AC load predictions utilizing those features. Combinatorial Optimization was used for conducting the disaggregation step while the prediction step is carried out utilizing several machine-learning algorithms including Artificial Neural Networks (ANNs), Support Vector Machines (SVMs) and Random Forests Decision Trees. The main advantage of the present methodology with respect to conventional prediction algorithms is separating the AC consumption from the other devices. The harsh alterations in the consumption of AC equipment can be predicted employing the short-term weather forecast while the consumption of other residential appliances is largely dependent upon the occupant’s behaviour, which is more difficult to predict without intrusive methods. Hence, this methodology facilitates predicting notable changes in the AC consumption, with a higher accuracy, only employing the aggregate consumption measured by a smart meter. However, it should be noted that a very short measurement period, with additional meters for the main devices, is needed; the requirement that can eventually be evaded once abundant measured data including different combinations of various residential devices is available. The mentioned prediction enables the utility companies and grid managers to estimate the loads of residential buildings, which contribute a significant portion of energy demand and better manage the energy dispatch.File | Dimensione | Formato | |
---|---|---|---|
2017_10_Manivannan.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Master Thesis Report
Dimensione
2.83 MB
Formato
Adobe PDF
|
2.83 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/136434