A Framework for Implementation of Dynamic Discrepancy Reduced-Order Modeling in Advanced Process Control

This paper introduces a novel framework for implementing dynamic discrepancy reduced-order modeling in advanced process control. This framework balances computational complexity and model accuracy by constructing a gray-box model that combines first-principles components with black-box functions to represent the critical process dynamics. Unlike conventional methods that correct plant-model mismatches based on output discrepancies, the proposed approach focuses on discrepancies in the rates of change within the reduced-order model. This method compensates for the reduced model's loss of dynamic information, resulting in a more accurate model for process control. Three criteria are proposed for constructing dynamic discrepancy functions in gray-box models for model predictive control (MPC). Because differences in rates of change are not directly observable through the outputs, moving horizon estimation is used to guide data generation and collection for dynamic discrepancy functions. This technique allows flexible integration of dynamic discrepancies into the reduced-order model. Bayesian inference is employed to calibrate hyperparameters and apply the Occam's razor principle to simplify the discrepancy functions. The approach is validated through a simulation involving a Fischer-Tropsch synthesis slurry bubble column reactor, demonstrating that the dynamic discrepancy approach can improve model and computational performance to be used in MPC applications.