Physics-guided time-series machine learning for dynamic modeling, control, and resilience in process systems

Abstract

Real-time predictive process control and optimization rely fundamentally on accurate dynamic process models. Across chemical, agricultural, and food processing systems, several strategies have been extensively studied for constructing such models. First-principles, white-box models are among the principal approaches, derived from mass, energy, and momentum balances that yield systems of partial differential equations (PDEs) or ordinary differential equations (ODEs). Using these models for real-time purposes, however, faces certain limitations, including unknown parameters and high computational costs. Moreover, as process systems become increasingly integrated, the governing equations grow limited in their ability to describe the full system dynamics, introducing unmodeled physics and discrepancies that adversely affect real-time performance. Time-series machine learning (ML) modeling has been introduced to address these challenges; however, purely black-box models, in which a learning-based model is developed from process data alone, suffer in turn from limited generalizability and extrapolation capability. Bridging the gap between these two modeling approaches, this dissertation explores different methods of integrating process domain knowledge with advanced time-series ML techniques, constructing knowledge-guided modeling frameworks that can be embedded within real-time control and optimization schemes and further employed for process safety and resilient operation. Depending on the form in which process knowledge is available, ranging from high-fidelity simulations and reduced-order physical models to expert assessments, imaging, and real-time sensor measurements, dedicated frameworks are developed, including surrogate models for model predictive control, hybrid and physics-informed architectures, multimodal modeling, and ML-based cyberattack detection and recovery. The proposed frameworks are evaluated through several case studies in chemical, agricultual and food process systems.