Gaussian Mixture Models (GMMs) are probabilistic models representing the combination of multiple Gaussian distributions. Despite their popularity, their modelling capabilities over the control of heating systems are relatively unexplored, since conventional approaches are usually white-box models, but they result complex to handle. This thesis aims to address this gap by employing a black-box modelling approach using Non-linear Auto-Regressive Exogenous (NARX) models. The Silhouette algorithm is implemented to choose the optimal the number of mixture components for GMM modelling. Gaussian Processes (GPs) are then considered to assess the performance of GMMs. The study begins with an examination of a water-heating benchmark system, where hot water is produced by regulating the gas flow heating a plate and controlling the water flow into a tank. Both GMMs and GPs are trained to model the system, demonstrating accurate regression performance as measured by the Mean-Squared Error (MSE) metric. Subsequently, attention shifts to the control domain, where a Model Predictive Control (MPC) approach is implemented to regulate the gas flow to maintain the desired water temperature despite disturbances from the water flow. The cost function designed aims to track the reference water temperature by controlling the gas flow and reject changes of the water flow. GP models are also integrated for online prediction, revealing similar control performance but longer computation times compared to GMMs. Lastly, the thesis explores the application of GMMs to a District Heating System (DHS), an energy plant responsible for heating various loads through water pipelines (District Heating Network or DHN). Following similar steps to the benchmark case, the modelling phase reveals slightly less accurate results. However, a new MPC is designed to optimize the electrical costs of the DHS, resulting in successful cost reduction compared to fixed boiler temperature values, with rapid optimization achieved.
I Gaussian Mixture Models (GMMs) sono modelli probabilistici formati dalla combinazione di più distribuzioni gaussiane. Nonostante la loro popolarità, le capacità di modellizzazione dei sistemi di riscaldamento sono relativamente inesplorate. Questa tesi mira a colmare questa lacuna con un approccio black-box, utilizzando modelli non-lineari auto-regressivi esogeni (NARX), dato che gli approcci convenzionali sono solitamente modelli white-box, ma risultato difficili da gestire. All'inizio il Silhouette algorithm è applicato ai dati per scegliere il numero ottimo di componenti del mixture per la modellizzazione dei GMM. Anche i Processi Gaussiani (GPs) sono considerati per valutare le prestazioni dei GMMs. Lo studio inizia con il test di un sistema di benchmark per il riscaldamento dell'acqua, dove l'acqua calda è prodotta regolando il flusso di gas che riscalda una piastra mentre scorre un flusso d'acqua nel serbatoio. Sia i GMMs che i GPs sono addestrati per modellare il sistema, dimostrando ottime prestazioni, grazie all' errore quadratico medio (MSE). L'attenzione, poi, si sposta sul campo del controllo, dove viene implementato un modello di controllo predittivo (MPC) per regolare il flusso di gas, mantenendo la temperatura desiderata dell'acqua nonostante le variazioni del flusso d'acqua. La funzione di costo implementata mira a tracciare la temperatura dell'acqua di riferimento controllando il flusso di gas e a respingere i cambiamenti del flusso d'acqua. Anche i modelli GP sono integrati per la predizione online, rivelando una prestazione di controllo simile ma tempi di calcolo più lunghi rispetto ai GMMs. Infine, la tesi esplora l'applicazione dei GMM a un District Heating System (DHS), un impianto energetico responsabile del riscaldamento di vari carichi tramite condutture d'acqua (District Heating Network o DHN). Seguendo passi simili al caso di benchmark, la fase di modellizzazione rivela risultati leggermente meno accurati. Tuttavia, un nuovo MPC viene implementato per ottimizzare i costi elettrici del DHS, con conseguente riduzione dei costi rispetto ai valori fissi della temperatura della caldaia, e con una rapida ottimizzazione.
Modeling and predictive control of heating systems via Gaussian Mixture Models
LOBRIGLIO, LUCA
2022/2023
Abstract
Gaussian Mixture Models (GMMs) are probabilistic models representing the combination of multiple Gaussian distributions. Despite their popularity, their modelling capabilities over the control of heating systems are relatively unexplored, since conventional approaches are usually white-box models, but they result complex to handle. This thesis aims to address this gap by employing a black-box modelling approach using Non-linear Auto-Regressive Exogenous (NARX) models. The Silhouette algorithm is implemented to choose the optimal the number of mixture components for GMM modelling. Gaussian Processes (GPs) are then considered to assess the performance of GMMs. The study begins with an examination of a water-heating benchmark system, where hot water is produced by regulating the gas flow heating a plate and controlling the water flow into a tank. Both GMMs and GPs are trained to model the system, demonstrating accurate regression performance as measured by the Mean-Squared Error (MSE) metric. Subsequently, attention shifts to the control domain, where a Model Predictive Control (MPC) approach is implemented to regulate the gas flow to maintain the desired water temperature despite disturbances from the water flow. The cost function designed aims to track the reference water temperature by controlling the gas flow and reject changes of the water flow. GP models are also integrated for online prediction, revealing similar control performance but longer computation times compared to GMMs. Lastly, the thesis explores the application of GMMs to a District Heating System (DHS), an energy plant responsible for heating various loads through water pipelines (District Heating Network or DHN). Following similar steps to the benchmark case, the modelling phase reveals slightly less accurate results. However, a new MPC is designed to optimize the electrical costs of the DHS, resulting in successful cost reduction compared to fixed boiler temperature values, with rapid optimization achieved.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Lobriglio.pdf
accessibile in internet per tutti
Descrizione: Executive Summary + Testo Tesi
Dimensione
6.02 MB
Formato
Adobe PDF
|
6.02 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/10589/218453