(184m) Dual Control Framework with Multistep Ahead Prediction Model

In this work, a dual control framework, which is able to simultaneously track setpoint and online re-identify process model, is presented. Moreover, a new optimization algorithm is integrated into this framework to minimize the multistep prediction error for the auto regressive with exogenous inputs (ARX) model.

Though the model predictive control (MPC) has shown great capabilities to achieve the optimal solution for setpoint tracking, its real performance may degrade rapidly due to model mismatch and parameters drift. To overcome this drawback, the dual control scheme, in which control and model identification are solved in parallel, has been received great attentions. In such scheme, the persistent excitation (PE) condition becomes significantly important, because the identification from closed-loop data without such condition will lead to biased model. However, how to strike a balance between PE condition and setpoint tracking is still outstanding because these two objectives are naturally conflicted. An analytical solution is hardly developed [1] and numerical optimization should be used. However, the resulting problem is usually non-convex [2]. Therefore it can be very complex and difficult to implement for real time regulation.

Our first contribution is the development of a tractable two-step formulation to decide the control input such that the information from input/output data sequence is maximized while the stability of the process is still guaranteed. The resulting optimization is relatively easy to solve. Our second contribution is the development of a simple optimization scheme for ARX model identification by explicitly minimizing the multistep prediction error. Hence, the resulting process model is very suitable for process prediction for a long time horizon.

Reference

[1] J. Sternby, A simple dual control problem with an analytical solution, IEEE Transactions on Automatic Control 21 (6) (1976) 840-844.

[2] H. Genceli, M. Nikolaou, New approach to constrained predictive control with simultaneous model identification, AIChE Journal 42 (1996) 2857–2868.