Two-phase model predictive control and its application to the Tennessee Eastman process

Abstract

This paper considers a two-phase model predictive control (MPC) which utilize a parsimonious parametrization of the future control moves to decrease the number of the degrees of freedom of the optimization and, thereby, to reduce the computations. Namely, the future control actions of the dynamic optimization are split into two stages. In the first phase, the control moves are considered as individual degrees of freedom. In the second phase, which is defined as the period between the end of the first phase and the prediction horizon, the control actions are determined using a weighted sum of some open-loop controls selected at the MPC design stage. With this parametrization, the bounds on the manipulated variables need to be treated as linear constraints. Alternatively, this paper estimates the maximum and the minimum of the future control trajectory that allows one to limit the number of the constraints representing the bounds on the MPC inputs. Thus, an additional reduction of the computations is achieved. To test the two-phase MPC, an MPC-based control strategy for the Tennessee Eastman challenge problem is developed, and a comparison of the two-phase MPC with MPC using Laguerre functions and MPC with move blocking is presented.