Doctoral Thesis Defense - Learning-Based MPC with Adaptive and Economic Features - Thomás Pinto

Описание: Doctoral Thesis defense of the Graduate Program in Electrical Engineering at UFMG
Abstract:
This thesis focuses on learning-based model predictive controllers (LB-MPC), a class of control strategies that combines the capabilities of two prominent scientific fields: Model Predictive Control (MPC) and Machine Learning (ML). While LB-MPC has gained increasing attention, significant challenges remain. This work addresses two important gaps in these frameworks. The first concerns the challenge of ensuring enhanced long-term performance when applied to processes that are subject to frequent changes in operating conditions and system dynamics. The second involves providing formal and rigorous stability and feasibility guarantees, theoretical foundations often overlooked in practical implementations.

To address these challenges, two LB-MPC frameworks are proposed. Both incorporating adaptive learning mechanisms and formal guarantees, and are designed for reference tracking, a common requirement in real-world applications. Recursive feasibility is ensured through a tracking formulation that incorporates artificial references into the optimal control problem. The first framework introduces a dual-learning approach, in which two key components of an MPC scheme are performed through learning techniques: the prediction model and the selection of tuning parameters. It employs a nonlinear autoregressive exogenous (NARX) model for prediction, while a supervised machine learning algorithm adjusts selected controller parameters online. The second framework presents an adaptive tuning mechanism designed to enhance the economic performance of LB-MPC when tracking changing economic targets. This is achieved by switching the values of parameters in the predictive scheme using a supervised learning technique that accounts for the current system conditions. A Lyapunov analysis of the switched system establishes asymptotic stability under a switching terminal cost and the proposed switching policy.

The proposed frameworks are evaluated through simulations and case studies in process control applications. Results demonstrate that the methods improve tracking performance compared to conventional MPC schemes. By bridging the gap between learning mechanisms and control guarantees, this thesis contributes to advancing LB-MPC toward broader industrial applicability, especially in scenarios where adaptability, performance, and safety are critical.