The objective of this thesis is to design hybrid Model Predictive Control (MPC) algorithms to optimize the behaviour of the cooling station of a large business center. Two different modelling approaches are exploited to identify the cooling station: (i) a physical-based method, where the single components of the cooling station are separately analysed with the aim of building an overall simulator in the MATLAB/Simulink environment. (ii) a black-box method, where the identification is performed based on input/output data using linear and nonlinear models. Specifically, two particular types of Recurrent Neural Networks are investigated, that are Echo State Newtorks (ESN) and Long Short-Term Memory (LSTM) networks. They represent two different ways of overcoming the so called "vanishing/exploding gradient problem", which is a big issue of the training procedure of a recursive neural network. Firstly, the architecture of these networks is presented and a mathematical model is derived to describe their dynamical behaviour. Moreover, the training algorithms of both the networks are illustrated. Then, the black-box identification is performed using linear models and neural networks, and a comparison between their performances is made to decide which model to use for control purposes. The last part of the thesis relates to the implementation of model predictive control schemes. In particular, the limitations of nonlinear MPC are emphasised and an alternative formulation of the optimization algorithm based on successive linearizations is presented. A particular attention is given also to another issue of the control problem, i.e. the fact that the control variables assume only integer values. Two main control problems are considered, one to control the physical-based model and one to control the LSTM network model. The performances of the designed control systems are tested via simulation in the MATLAB/Simulink environment and numerical results are provided.
L'obiettivo di questa tesi è la progettazione di algoritmi di controllo predittivo basati su modello (MPC) per ottimizzare il funzionamento dell’impianto di raffreddamento dell’acqua di un centro commerciale formato da cinque edifici. Vengono seguiti due diversi approcci di modellazione per identificare il sistema di raffreddamento: (i) Un metodo basato sulla fisica del sistema, dove i singoli componenti dell’impianto vengono analizzati separatamente con lo scopo di realizzare un simulature complessivo nell’ambiente di lavoro MATLAB/Simulink. (ii) Un modello a scatola nera (black-box), dove il modello è identificato basandosi su dati ingresso/uscita utilizzando modelli lineari e non lineari. In particolare, vengono analizzati due tipi di reti neurali ricorsive (RNN), cioè reti “Echo State” (ESN) e “Long Short-Term Memory” (LSTM). Queste due reti rappresentano due modi diversi di far fronte al cosiddetto problema della “scomparsa /esplosione” del gradiente, che è uno dei principali problemi che influenzano la procedura di addestramento di una rete neurale ricorsiva. Prima di tutto, viene presentata l’architettura delle due reti e viene derivato un modello matematico per descriverne il comportamento dinamico. Successivamente, vengono analizzati gli algoritmi di addestramento di entrambi i modelli. Infine, modelli lineari e reti neurali vengono utilizzati per identificare il sistema tramite metodi a scatola nera, e le loro prestazioni vengono confrontate per stabilire quale modello sia più adatto ad essere utilizzato per fini di controllo. L’ultima parte di questa tesi riguarda l’implementazione di schemi di controllo predittivo MPC basati su modello. In particolare, vengono evidenziati i limiti di un MPC non lineare a viene proposta una formulazione alternativa basata sulla linearizzazione del sistema lungo le traiettorie degli stati. Particolare attenzione viene data al fatto che le variabili di controllo all’interno del problema di ottimizzazione assumano valori interi. Vengono considerati due principali problemi di controllo, uno per controllare le equazioni che rappresentano la fisica del sistema, e uno per controllare il modello basato su reti neurali LSTM. I sistemi di controllo implementati vengono testati nell’ambiente di lavoro MATLAB/Simulink e le loro prestazioni vengono valutate mostrando i risultati numerici delle simulazioni.
Modeling and learning-based hybrid predictive control with recurrent neural networks of the cooling station of a large business center
SACCANI, DANILO;BONETTI, TOMMASO
2018/2019
Abstract
The objective of this thesis is to design hybrid Model Predictive Control (MPC) algorithms to optimize the behaviour of the cooling station of a large business center. Two different modelling approaches are exploited to identify the cooling station: (i) a physical-based method, where the single components of the cooling station are separately analysed with the aim of building an overall simulator in the MATLAB/Simulink environment. (ii) a black-box method, where the identification is performed based on input/output data using linear and nonlinear models. Specifically, two particular types of Recurrent Neural Networks are investigated, that are Echo State Newtorks (ESN) and Long Short-Term Memory (LSTM) networks. They represent two different ways of overcoming the so called "vanishing/exploding gradient problem", which is a big issue of the training procedure of a recursive neural network. Firstly, the architecture of these networks is presented and a mathematical model is derived to describe their dynamical behaviour. Moreover, the training algorithms of both the networks are illustrated. Then, the black-box identification is performed using linear models and neural networks, and a comparison between their performances is made to decide which model to use for control purposes. The last part of the thesis relates to the implementation of model predictive control schemes. In particular, the limitations of nonlinear MPC are emphasised and an alternative formulation of the optimization algorithm based on successive linearizations is presented. A particular attention is given also to another issue of the control problem, i.e. the fact that the control variables assume only integer values. Two main control problems are considered, one to control the physical-based model and one to control the LSTM network model. The performances of the designed control systems are tested via simulation in the MATLAB/Simulink environment and numerical results are provided.File | Dimensione | Formato | |
---|---|---|---|
Tesi.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Thesis text
Dimensione
6.21 MB
Formato
Adobe PDF
|
6.21 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/10589/148005