Learning-Based MPC-Architecture Satisfying Constraints During Open Loop Identification

A popular approach to ensure economic operation of processes is to employ Model Predictive Control (MPC) combined with an economic set-point optimizer. To track the economic optimal set-points, the MPC utilizes a dynamical model of the system to coordinate the system inputs and outputs. However, the closed loop performance depends on the model accuracy. So to ensure high model accuracy, a typical approach is to learn the model parameters from experimental process data. This learning-based approach requires that the process is excited during the experiments. During economic operation, the process is very often close to its constraints, so the excitation might drive the process across the constraints; so constraint handling is of high importance. Since the loop is closed with an MPC, it is not possible to use typical closed loop identification methods. These methods often require linear controllers where the control law is explicitly known. To this end, we propose an MPC-architecture that allows for open loop identification while handling constraints. We further provide proofs for that the MPC-architecture does not interfere with the open loop identification, when the process is not at its constraints. We illustrate the proposed MPC-architecture with simulations on a simple example.