Automatic Tuning Methods for MPC Environments

Abstract

Model predictive control is a powerful method for controlling multivariable systems, in particular when constraints have to be taken into account. However, MPC also has negative attributes, like a high computational effort but also a non intuitive tuning. While the first issue can be addressed by use of numerically efficient optimizers, the non intuitive tuning still remains. To this end, an approach for efficient tuning of a MPC environment consisting of controller and state observer is proposed. The idea is to provide an automatic tuning strategy, such that even an unexperienced user can design a satisfactory controller within reasonable time. The proposed tuning of the state observer is done by a combination of multi model and adaptive estimation methods and for the weight tuning of the MPC objective function an additional optimization loop is applied which also accounts for the numerical condition. Finally an example is presented, where the proposed strategies were used to tune the MPC for controlling a pipeline compressor natural gas engine in a nonlinear simulation environment, yielding promising results.