Cement production plays a significant role in global CO2 emissions. Advanced control algorithms could reduce its environmental impact by improving the efficiency of the process. Nonlinear Model Predictive Control (NMPC) is a technique particularly fit for this role, since it minimizes a cost function while satisfying a set of constraints. A key element required by NMPC is an accurate mathematical model of the controlled system. However, its derivation could be challenging, especially for complex systems such as cement production plants. As an alternative, learning-based approaches are being investigated. They leverage historical data to design the entire controller or part of its components. In this work, Gaussian Processes regression is used to obtain a black-box model of key system variables of a clinker production plant. A CV-informed GP-NOE model was trained on historical data and compared to a MIMO transfer function model. The results show slight improvements in long multi-step-ahead predictions. The developed model has the potential to be implemented within a learning-based NMPC framework to control the clinker production plant.
La produzione di cemento svolge un ruolo significativo nelle emissioni globali di CO2. Algoritmi di controllo avanzati potrebbero ridurne l'impatto ambientale tramite un miglioramento dell'efficienza del processo. Il controllo predittivo non lineare (NMPC) è una tecnica particolarmente adatta a questo ruolo, poiché minimizza una funzione di costo soddisfacendo una serie di vincoli. Un elemento chiave richiesto dal NMPC è un modello matematico accurato del sistema controllato. Tuttavia, la sua derivazione può risultare onerosa, soprattutto per sistemi complessi come gli impianti di produzione del cemento. Una possibile alternativa sono gli approcci learning-based, attualmente in fase di studio. Essi sfruttano i dati storici per lo sviluppo dell'intero controllore o di alcuni suoi componenti. In questo elaborato, tecniche di regressione gaussiana vengono utilizzate per ottenere un modello black-box delle variabili chiave di un impianto di produzione di clinker. Un modello CV-informed GP-NOE è stato addestrato su dati storici e confrontato con una funzione di trasferimento MIMO. I risultati mostrano leggeri miglioramenti nelle predizioni multipasso a lungo termine. Il modello sviluppato ha il potenziale per essere implementato all'interno di un approccio learning-based NMPC per il controllo dell'impianto di produzione di clinker.
Implementation of a black-box Learning-based Nonlinear Model Predictive Control for a clinker production plant
SCOLA, DAVIDE
2021/2022
Abstract
Cement production plays a significant role in global CO2 emissions. Advanced control algorithms could reduce its environmental impact by improving the efficiency of the process. Nonlinear Model Predictive Control (NMPC) is a technique particularly fit for this role, since it minimizes a cost function while satisfying a set of constraints. A key element required by NMPC is an accurate mathematical model of the controlled system. However, its derivation could be challenging, especially for complex systems such as cement production plants. As an alternative, learning-based approaches are being investigated. They leverage historical data to design the entire controller or part of its components. In this work, Gaussian Processes regression is used to obtain a black-box model of key system variables of a clinker production plant. A CV-informed GP-NOE model was trained on historical data and compared to a MIMO transfer function model. The results show slight improvements in long multi-step-ahead predictions. The developed model has the potential to be implemented within a learning-based NMPC framework to control the clinker production plant.File | Dimensione | Formato | |
---|---|---|---|
Scola_Davide.pdf
accesso aperto
Dimensione
3.71 MB
Formato
Adobe PDF
|
3.71 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/40471