Adaptive modelling, optimization and control of large-scale systems using machine learning

This thesis explores the intersection of machine learning (ML), optimization, and control within process systems engineering (PSE), focusing on the real-time optimization (RTO) and control layers to advance industrial autonomy. The primary goal is to enhance adaptability, robustness, and efficiency through ML-driven frameworks. To address key challenges—plant-model mismatch, computational complexity, and robustness to disturbances—this work develops novel methods across three pillars: data-driven modelling, adaptive systems, and large-scale process control. The research begins with adaptive RTO using Gaussian Processes (GPs) to address model mismatch. This is extended to Adversarially Robust Real-Time Optimization and Control (ARRTOC), which shifts the robustness burden from control to RTO, improving operability under uncertainty. The combined use of GP adaptation and ARRTOC addresses both mismatch and robustness. In dynamic systems, we conduct a rigorous evaluation of data-driven models using defined metrics to guide model selection and assessment. This study highlights trade-offs across the modelling lifecycle, offering practitioners practical insights. The thesis further investigates deep learning-based Koopman operator methods to linearize nonlinear dynamics for efficient model-based control. For high-dimensional systems, we introduce ARRO-MPC, an Adversarially Robust Reduced-Order Model Predictive Control framework that uses adversarial training to improve ROM robustness, ensuring efficient yet reliable control. Overall, the findings show ML can significantly enhance adaptability and robustness in PSE. However, successful deployment hinges on integrating ML with domain expertise. This work bridges ML, control, and optimization to provide a pathway toward more autonomous and resilient industrial systems.