Multistage Model Predictive Control with Online Scenario Tree Update using Recursive Bayesian Weighting

This work deals with a nonlinear multistage model predictive control (MPC) formulation, where the future propagation of the uncertainty in the prediction horizon is represented via a discrete scenario tree. The scenario tree is often generated using finite realizations of the uncertainty sampled from an uncertainty set or a probability distribution function. Once the scenarios are chosen, the scenario tree is often kept fixed for all the iterations. In this paper, we propose to update the different discrete realizations of the uncertainty in the scenario tree using a recursive Bayesian weighting approach. We show that by gradually shrinking the uncertainty set, we can further reduce the conservativeness of the closed-loop solution. The effectiveness of the proposed method is demonstrated using an oil and gas production optimization case study.