Tracking necessary condition of optimality by a data-driven solution combining steady-state and transient data

Abstract

One of the difficulties in practical implementations of the classic Real-Time Optimization (RTO) strategy is the integration between optimization and control layers, mainly due to the differences between the models used in each layer, which may result in unreachable setpoints coming from optimization to the control layer. In this context, Economic Model Predictive Control (EMPC) is a strategy where optimization and control problems are solved simultaneously. However, this strategy is based on the assumption that a nonlinear dynamic model is available, which may not be valid. Also, when considering a first-principles nonlinear model, the computational cost and convergence may be relevant issues. The present work presents an RTO framework based on an EMPC structure considering a Hammerstein model for the plant. This modeling approach can be applied even in the absence of first-principles models. The proposed EMPC considers the minimization of the economic objective function gradient calculated through a steady-state model based on a Gaussian Process. This strategy was applied to the Willians–Otto Reactor benchmark and presented superior results than the classic RTO and Hybrid RTO (H-RTO) approaches in closed-loop and a lower average iteration time than these other approaches.