Online monitoring of MPC disturbance models using closed-loop data

In model predictive control applications, knowledge about the disturbances affecting the system can be incomplete or non-existent. Typically, the state noise and sensor noise covariances used in the state estimation problem are not known. Additionally, an integrated white noise disturbance is often used to remove offset. Imperfect or non-existent knowledge of this additional covariance is another source of uncertainty in the controller design. In this paper, methods are developed to employ closed-loop MPC data to identify the unknown covariances and fully realize the potential of the state estimator. Two simulation examples are provided, demonstrating the benefits of using such techniques.