Morabito, Bruno (2024)
Risk-aware and Robust Approaches for Machine Learning-supported Model Predictive Control for Iterative Processes.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00026499
Ph.D. Thesis, Primary publication, Publisher's Version
Text
main_darmstadt.pdf Copyright Information: CC BY-SA 4.0 International - Creative Commons, Attribution ShareAlike. Download (38MB) |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Risk-aware and Robust Approaches for Machine Learning-supported Model Predictive Control for Iterative Processes | ||||
Language: | English | ||||
Referees: | Findeisen, Prof. Dr. Rolf ; Lucia, Prof. Dr. Sergio | ||||
Date: | 23 January 2024 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xv, 162 Seiten | ||||
Date of oral examination: | 20 December 2023 | ||||
DOI: | 10.26083/tuprints-00026499 | ||||
Abstract: | The recent advances in machine learning have catalyzed a renewed interest in machine-learning-supported model predictive control. Machine learning promises to facilitate modeling and improve the process' performance. Nevertheless, it brings some challenges: For instance, as the connection with physics law is (partially) lost, machine learning models can provide wildly inaccurate results. It is therefore necessary to provide control methods that take the model uncertainty of these models into account. Uncertainties are even more important for iterative processes - processes that do not operate at a steady state - due to the large changes in the process conditions during operation. In this work, two methods for data-driven uncertainty modelling are proposed. The first method uses Gaussian processes to learn the model uncertainty and neural networks to learn the nominal model. It provides an simple way to summarize the uncertainty of the model into a single parameter, which can be used by a model predictive controller to make risk-aware decisions. This method, while being simple, does not guarantee constraint satisfaction. The second method is based on tube-based model predictive control and can guarantee constraint satisfaction. It is based on the concept of the "safe set": a set where a tube-based MPC has a feasible solution. We show that, under some assumptions, the safe set enlarges at every iteration of the process, potentially allowing increased performance. Finally, a novel Python library for machine-learning-based model predictive control, called HILO-MPC, is presented. This library interfaces with TensorFlow and PyTorch and provides easily-accesible tools for defining control and estimation problem using machine learning model. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-264993 | ||||
Classification DDC: | 600 Technology, medicine, applied sciences > 620 Engineering and machine engineering | ||||
Divisions: | 18 Department of Electrical Engineering and Information Technology > Institut für Automatisierungstechnik und Mechatronik > Control and Cyber-Physical Systems (CCPS) | ||||
Date Deposited: | 23 Jan 2024 13:05 | ||||
Last Modified: | 24 Jan 2024 08:29 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/26499 | ||||
PPN: | 514935073 | ||||
Export: |
View Item |