Feedback Through Steady-State Target Optimization for Nonlinear Model Predictive Control

Abstract

Model predictive control (MPC) algorithms use an explicit process model to predict future plant behavior and select an optimal control sequence based on a user-defined objective function. The optimal sequence is implemented until new data are available, at which time the data are incorporated as feedback and the calculation is repeated. One of the defining features of MPC is the repeated optimization of an objective that incorporates the most recent feedback from the process. Since most MPC algorithms used in industry are based on input-output models, state feedback is rarely used in practice. Feedback is usually incorporated by including an output error term in the objective function that includes the effects of disturbances and model mismatch. When internal states are included in the process model, other feedback options become available. This paper presents a promising new feedback approach in which the output error term is used to precalculate steady-state targets for state and control variables by means of an initial nonlinear program. The target values are then used to formulate the dynamic MPC objective. Results show that the new approach can provide stability and robustness properties equivalent to those of conventional MPC feedback formulations, with as much as an order of magnitude decrease in overall computation time. This makes nonlinear, state-based MPC much more attractive for on-line implementation.