Tube-based distributionally robust model predictive control for nonlinear process systems via linearization

Model predictive control (MPC) is an effective approach to multivariable dynamic systems with constraints. Most real models are however affected by plant-model mismatch and process uncertainties, which can lead closed-loop performance deterioration constraint violations. Methods such as stochastic MPC (SMPC) have been proposed alleviate these problems; however, the resulting state trajectory might still significantly violate prescribed constraints if system deviates from assumed disturbance distributions made during controller design. In this work we propose a novel data-driven distributionally robust scheme for nonlinear systems. Unlike SMPC, requires exact knowledge of distribution, our decides action respect worst distribution ambiguity set. This set defined Wasserstein ball centered at empirical distribution. Due potential model errors that cause off-sets, also extended leveraging offset-free method. The favorable results demonstrated empirically verified mass spring CSTR case study.