A district heating network is a system that combines different energy sources to generate heat and distribute it via a network to different users, such as private households or commercial offices. District heating networks are considered a key technology for the energy transition. They involve complex nonlinear dynamical equations not prone to control design. Moreover, they rely on a large number of physical parameters, often unknown. With the advent of recurrent neural networks (RNNs), reliable models, that exploit operational data to infer the dynamics of the system, have been developed. Although advantageous in terms of model learning capability, neural networks are highly dependent on the quality and quantity of the data used for training. This makes them unable to represent dynamical behavior they have not previously seen in the training set. The aim of this thesis is to develop online model adaptation algorithms to improve the accuracy of the network by adapting it using data collected in real time. This objective is achieved through the implementation of two online adaptation techniques: Gaussian Processes and Bayesian Neural Networks, tested for the prediction of district heating network dynamics. For both techniques, online improvement of the model has positive effects not only in terms of identification, but also of control performances. In fact, the developed models have been integrated into a model predictive controller to optimize network operation in order to minimize energy costs. Online adaptation of the model, on the other hand, improves control performance and reduces generation costs. The algorithms have been experimentally tested on the Test facility of RSE for over four hours, obtaining enhanced performances. In conclusion, the overall aim of this thesis is to make a data-based model of the district heating network more accurate through online learning in order to improve network operation and minimize generation costs.
Una rete di teleriscaldamento è un sistema energetico che combina diverse fonti di energia per generare calore e distribuirlo tramite una rete a diversi utenti, siano essi abitazioni private o uffici commerciali. Le reti di teleriscaldamento sono considerate una tecnologia chiave per la transizione energetica. Esse comportano complesse equazioni dinamiche non lineari che non si prestano alla progettazione di sistemi di controllo. Inoltre, si basano su un gran numero di parametri fisici, spesso sconosciuti. Con l’avvento delle reti neurali ricorrenti (RNN), sono stati sviluppati modelli affidabili che sfruttano i dati operativi per dedurre le dinamiche del sistema. Sebbene vantaggiose in termini di capacità di apprendimento del modello, le reti neurali dipendono fortemente dalla qualità e dalla quantità dei dati utilizzati per l’addestramento. Ciò le rende incapaci di rappresentare comportamenti dinamici che non hanno precedentemente osservato nel set di addestra- mento. Lo scopo di questa tesi è sviluppare algoritmi di adattamento del modello online per migliorare l’accuratezza della rete adattandola ai dati raccolti in tempo reale. Questo obiettivo è raggiunto attraverso l’implementazione di due tecniche di adattamento online: processi gaussiani e reti bayesiane, testate per la previsione delle dinamiche delle reti di teleriscaldamento. Per entrambe le tecniche, il miglioramento online del modello ha effetti positivi non solo in termini di identificazione, ma anche di prestazioni di controllo. In- fatti, i modelli sviluppati sono stati integrati in un controllore predittivo per ottimizzare il funzionamento della rete al fine di ridurre al minimo i costi energetici. L’adattamento online del modello, d’altra parte, migliora le prestazioni di controllo e riduce i costi di generazione. Gli algoritmi sono stati testati sperimentalmente sulla Test Facility per oltre quattro ore, ottenendo dei risultati consistenti con la simulazione. In conclusione, l’obiettivo generale di questa tesi è quello di rendere più accurato un modello data-based della rete di teleriscaldamento attraverso l’apprendimento online, al fine di migliorare il funzionamento della rete e ridurre al minimo i costi di generazione
Online learning and predictive control of district heating networks: design and experimental validation
CAPPELLO, MICHELA
2024/2025
Abstract
A district heating network is a system that combines different energy sources to generate heat and distribute it via a network to different users, such as private households or commercial offices. District heating networks are considered a key technology for the energy transition. They involve complex nonlinear dynamical equations not prone to control design. Moreover, they rely on a large number of physical parameters, often unknown. With the advent of recurrent neural networks (RNNs), reliable models, that exploit operational data to infer the dynamics of the system, have been developed. Although advantageous in terms of model learning capability, neural networks are highly dependent on the quality and quantity of the data used for training. This makes them unable to represent dynamical behavior they have not previously seen in the training set. The aim of this thesis is to develop online model adaptation algorithms to improve the accuracy of the network by adapting it using data collected in real time. This objective is achieved through the implementation of two online adaptation techniques: Gaussian Processes and Bayesian Neural Networks, tested for the prediction of district heating network dynamics. For both techniques, online improvement of the model has positive effects not only in terms of identification, but also of control performances. In fact, the developed models have been integrated into a model predictive controller to optimize network operation in order to minimize energy costs. Online adaptation of the model, on the other hand, improves control performance and reduces generation costs. The algorithms have been experimentally tested on the Test facility of RSE for over four hours, obtaining enhanced performances. In conclusion, the overall aim of this thesis is to make a data-based model of the district heating network more accurate through online learning in order to improve network operation and minimize generation costs.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Cappello_Executive_Summary.pdf
accessibile in internet per tutti
Dimensione
935.97 kB
Formato
Adobe PDF
|
935.97 kB | Adobe PDF | Visualizza/Apri |
|
2025_12_Cappello_Tesi.pdf
accessibile in internet per tutti
Dimensione
6.51 MB
Formato
Adobe PDF
|
6.51 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/10589/246879