Many industrial processes, as well as systems found in nature, have a dynamics characterized by a strong non-linearity. In recent years, to deal with this complexity, attention has turned to data-driven methods for analysis and control, which allow to obtain models not based on physical laws, but derived from data directly collected during normal operations or through specific experiments. Among these data-driven algorithms, Neural Networks have emerged as particularly effective tools. This thesis focuses on a specific type of Neural Network: Gated Recurrent Units, applied to a challenging benchmark, namely the pH neutralization process. In the context of non-linear systems, it is common to have to handle very complex Neural Network models characterized by many layers and neurons. To cope with this complexity, the idea has been to reduce the size of models by training simpler networks and combining them to achieve performance comparable to that of more complex models. In Machine Learning this approach is the basis of ensemble methods, which propose several algorithms inspired by this idea. However, these methods have not found wide applications in the field of control. Therefore, this thesis wants to explore these existing algorithms and extend the concept of developing smaller models describing the dynamics in different operating regions. The goal is then to combine these simple models and to compare their performance with the one of a unique model trained for all the operating conditions of interest. However, combining models raises the question of how to carry out this combination. To this end, this thesis proposes and compares various techniques, assessing to what extent simpler combined models can compete with the larger model. Once the best models to describe the system have been identified, a crucial step is the implementation of a control strategy. To this end, a non-linear Model Predictive Control (MPC) strategy based on solving an optimization problem is presented. MPC is implemented both on a complex model and on several simpler models combined by means of the Mahalanobis distance, enabling a comparison of performance in terms of efficiency and computation time.
Molti processi industriali, così come i sistemi presenti in natura, hanno una dinamica caratterizzata da una forte non linearità. Negli ultimi anni, per affrontare questa complessità, l'attenzione si è rivolta ai metodi data-driven per l'analisi e il controllo, che permettono di ottenere modelli non basati su leggi fisiche, ma derivati da dati raccolti direttamente durante le normali operazioni o attraverso esperimenti specifici. Tra questi algoritmi data-driven, le reti neurali sono emerse come strumenti particolarmente efficaci. Questa tesi si concentra su un tipo specifico di rete neurale: Gated Recurrent Units, applicate a un benchmark impegnativo, ovvero il processo di neutralizzazione del pH. Nel contesto dei sistemi non lineari, è comune dover gestire modelli di reti neurali molto complessi, caratterizzati da molti strati e neuroni. Per affrontare questa complessità, si è pensato di ridurre le dimensioni dei modelli addestrando reti più semplici e combinandole per ottenere prestazioni paragonabili a quelle di modelli più complessi. Nel Machine Learning questo approccio è alla base dei metodi ensemble, che propongono diversi algoritmi ispirati a questa idea. Tuttavia, questi metodi non hanno trovato ampie applicazioni nel campo del controllo. Pertanto, questa tesi vuole esplorare questi algoritmi esistenti ed estendere il concetto di sviluppare modelli più piccoli che descrivano la dinamica in diverse regioni operative. L'obiettivo è poi quello di combinare questi modelli semplici e di confrontare le loro prestazioni con quelle di un unico modello addestrato per tutte le regioni operative di interesse. Combinare i modelli solleva, però, la questione di come effettuare tale combinazione. Pertanto, questa tesi propone e confronta diverse tecniche, valutando la competitività dei modelli combinati più semplici rispetto al modello più grande. Una volta identificati i modelli migliori per descrivere il sistema, un passo cruciale è l'implementazione di una strategia di controllo. A tal fine, viene presentata una strategia di Controllo Predittivo (MPC) non lineare basata sulla risoluzione di un problema di ottimizzazione. L'MPC è implementato sia su un modello complesso che su diversi modelli più semplici combinati con la distanza di Mahalanobis, consentendo un confronto delle prestazioni in termini di efficienza e tempo di computazione.
Ensemble approaches for estimation and control of systems with multiple operating points
VIGANO', LISA
2023/2024
Abstract
Many industrial processes, as well as systems found in nature, have a dynamics characterized by a strong non-linearity. In recent years, to deal with this complexity, attention has turned to data-driven methods for analysis and control, which allow to obtain models not based on physical laws, but derived from data directly collected during normal operations or through specific experiments. Among these data-driven algorithms, Neural Networks have emerged as particularly effective tools. This thesis focuses on a specific type of Neural Network: Gated Recurrent Units, applied to a challenging benchmark, namely the pH neutralization process. In the context of non-linear systems, it is common to have to handle very complex Neural Network models characterized by many layers and neurons. To cope with this complexity, the idea has been to reduce the size of models by training simpler networks and combining them to achieve performance comparable to that of more complex models. In Machine Learning this approach is the basis of ensemble methods, which propose several algorithms inspired by this idea. However, these methods have not found wide applications in the field of control. Therefore, this thesis wants to explore these existing algorithms and extend the concept of developing smaller models describing the dynamics in different operating regions. The goal is then to combine these simple models and to compare their performance with the one of a unique model trained for all the operating conditions of interest. However, combining models raises the question of how to carry out this combination. To this end, this thesis proposes and compares various techniques, assessing to what extent simpler combined models can compete with the larger model. Once the best models to describe the system have been identified, a crucial step is the implementation of a control strategy. To this end, a non-linear Model Predictive Control (MPC) strategy based on solving an optimization problem is presented. MPC is implemented both on a complex model and on several simpler models combined by means of the Mahalanobis distance, enabling a comparison of performance in terms of efficiency and computation time.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Viganò_Tesi.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Tesi
Dimensione
10.62 MB
Formato
Adobe PDF
|
10.62 MB | Adobe PDF | Visualizza/Apri |
2024_12_Viganò_ExecutiveSummary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive Summary
Dimensione
1.4 MB
Formato
Adobe PDF
|
1.4 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/10589/230454