Integration of economic mpc and modifier adaptation in slow dynamic processes with structural model uncertainty

Résumé

Real-Time Optimization, known by its acronym RTO, uses a steady-state nonlinear model of the process to optimize a plant's economic objective subject to process constraints. This is the technology currently used in commercial RTO applications. However, no model is a perfect representation of reality, and structural and parametric model uncertainties make the optimum calculated by RTO do not match those of the actual process. One way to address this problem is to modify the optimization problem so that the Necessary Conditions of Optimality (NCO) of the problem match those of the actual plant. This strategy is known as Modifier Adaptation (MA) methodology. The MA methodology requires the gradient values of the real plant and the model to calculate the modifiers. There are several ways to accurately estimate model gradients, but estimation of the real process gradients are more difficult. In addition, the need to use stationary data is a limitation of RTO with MA, especially for slow dynamic systems. This thesis focuses on ways to mitigate the weaknesses of RTO and MA unification that we consider most critical for its application in industry. To this end, it is proposed to couple the RTO and control layers with the concepts of the Modifier Adaptation (MA) methodology by estimating process gradients or directly the MA modifiers using transient data.