Optimizing energy consumption in buildings while maintaining thermal comfort is a strategic objective in the design of intelligent HVAC systems. This work is based on the analysis of a real dataset collected from a building located in Denmark, which includes thermal data such as indoor and outdoor temperatures, as well as information on the operation of heat pumps and thermal storage systems. A thorough data quality analysis was conducted initially, applying validation rules to identify anomalies, measurement errors, and inconsistent behaviour of sensors and actuators. This step was essential to ensure the reliability of the predictive models. Subsequently, a simplified grey-box model of the building was developed, based on resistive-capacitive thermo-dynamic equations. The model includes thermal components such as internal thermal capacity (Cz), wall thermal capacity (Cwall), heating power (Ph), and heat loss (Udisp). Parameter identification was performed using Least Squares Estimation (LSE), and the model was validated by comparing simulations with real data. All models, including the grey-box and black-box approaches, were implemented and tested within the MATLAB environment. For comparison purposes, two black-box models were also implemented, based on Gaussian Processes (GP) and Neural Networks (NN), respectively. These models estimate the building’s thermal behavior directly from data, without relying on explicit physical structures. Although the three approaches showed similar overall trends, significant differences in accuracy and adaptability distinguished their comparative performance. For the black-box models, a validation system based on clustering was adopted, involving a structured split between test and validation sets to evaluate the predictive capacity under different and independent operating conditions. This approach helped better identify the generalization limits of data-driven models. The combined use of physical and data-driven approaches highlights the potential of predictive models in building energy management, suggesting future developments that aim to integrate advanced control strategies, continuous adaptation, and dynamic forecasting of thermal behaviour.
L’ottimizzazione dei consumi energetici negli edifici, mantenendo al contempo condizioni di comfort termico, rappresenta un obiettivo strategico nella progettazione di sistemi HVAC intelligenti. Il presente lavoro si basa sull’analisi di un dataset reale, acquisito da un edificio situato in Danimarca, contenente dati termici relativi a temperature interne ed esterne, nonché informazioni sul funzionamento delle pompe di calore e dei sistemi di accumulo termico. Nella fase preliminare è stata condotta un’accurata analisi della qualità dei dati, applicando regole di validazione per l’identificazione di valori anomali, errori di misura e comportamenti incoerenti dei sensori e degli attuatori. Questo processo si è rivelato fondamentale per garantire l’affidabilità dei successivi modelli predittivi. Successivamente, è stato sviluppato un modello grey-box semplificato dell’edificio, basato su equazioni termo-dinamiche di tipo resistivo-capacitivo. Il modello include componenti termici come la capacità termica interna (Cz), la capacità delle pareti (Cwall), la potenza di riscaldamento (Ph) e la dispersione verso l’esterno (Udisp). L’identificazione dei parametri è stata effettuata tramite Least Squares Estimation (LSE), e il modello è stato validato confrontando le simulazioni con i dati reali. Tutti i modelli, inclusi quelli grey-box e black-box, sono stati implementati e testati all’interno dell’ambiente MATLAB. A fini comparativi, sono stati implementati anche due modelli black-box, basati rispettivamente su Gaussian Processes (GP) e Neural Networks (NN). Questi modelli hanno permesso di stimare il comportamento termico dell’edificio direttamente dai dati, senza utilizzare una struttura fisica esplicita. L’analisi dei risultati ha evidenziato prestazioni comparabili tra i tre approcci, con differenze significative in termini di accuratezza e capacità di adattamento. Per i modelli black-box è stato adottato un sistema di validazione basato su clustering con una suddivisione strutturata tra set di test e set di validazione, al fine di valutare la capacità predittiva dei modelli su condizioni operative diverse e indipendenti. Questo approccio ha permesso di identificare con maggiore precisione i limiti di generalizzazione dei modelli basati su dati. L’impiego combinato di approcci fisici e data-driven ha messo in luce il potenziale dei modelli predittivi nella gestione energetica degli edifici, suggerendo possibili sviluppi futuri orientati all’integrazione con strategie di controllo avanzato, adattamento continuo e previsione dinamica del comportamento termico.
Data-driven thermal modeling for buildings: from grey-box to black-box methods
SERRA, GIUSEPPE;Zito, Alessandro
2024/2025
Abstract
Optimizing energy consumption in buildings while maintaining thermal comfort is a strategic objective in the design of intelligent HVAC systems. This work is based on the analysis of a real dataset collected from a building located in Denmark, which includes thermal data such as indoor and outdoor temperatures, as well as information on the operation of heat pumps and thermal storage systems. A thorough data quality analysis was conducted initially, applying validation rules to identify anomalies, measurement errors, and inconsistent behaviour of sensors and actuators. This step was essential to ensure the reliability of the predictive models. Subsequently, a simplified grey-box model of the building was developed, based on resistive-capacitive thermo-dynamic equations. The model includes thermal components such as internal thermal capacity (Cz), wall thermal capacity (Cwall), heating power (Ph), and heat loss (Udisp). Parameter identification was performed using Least Squares Estimation (LSE), and the model was validated by comparing simulations with real data. All models, including the grey-box and black-box approaches, were implemented and tested within the MATLAB environment. For comparison purposes, two black-box models were also implemented, based on Gaussian Processes (GP) and Neural Networks (NN), respectively. These models estimate the building’s thermal behavior directly from data, without relying on explicit physical structures. Although the three approaches showed similar overall trends, significant differences in accuracy and adaptability distinguished their comparative performance. For the black-box models, a validation system based on clustering was adopted, involving a structured split between test and validation sets to evaluate the predictive capacity under different and independent operating conditions. This approach helped better identify the generalization limits of data-driven models. The combined use of physical and data-driven approaches highlights the potential of predictive models in building energy management, suggesting future developments that aim to integrate advanced control strategies, continuous adaptation, and dynamic forecasting of thermal behaviour.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_07_Serra_Zito.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Tesi di laurea magistrale Serra Zito
Dimensione
2.86 MB
Formato
Adobe PDF
|
2.86 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/10589/240352