(659e) Error-Triggered on-Line Model Identification for Model-Based Feedback Control | AIChE

(659e) Error-Triggered on-Line Model Identification for Model-Based Feedback Control


Alanqar, A. - Presenter, University of California, Los Angeles
Durand, H., University of California, Los Angeles
Christofides, P., University of California, Los Angeles
Empirical models obtained from model identification techniques are used in the chemical process industry when a first-principles model is unknown or too complex to use. Once a process model is obtained, it can be used in model-based control strategies like model predictive control (MPC) [1]-[2]. Many methods for the identification of linear empirical models based on input/output data exist such as the canonical variate algorithm (CVA) [3] and the multivariable output error state-space algorithm (MOESP) [4]. Linear empirical models can be used to model nonlinear process dynamics within a limited region. If plant variations occur or the process state moves out of the region for which the linear empirical model was valid for the nonlinear process, the linear model may need to be updated.

In this work, we propose a prediction error metric that can be used to trigger on-line model identification to update a linear empirical model for use in model-based control. A moving horizon error detector is developed based on the total relative error between the measured states and the states predicted using the current linear empirical model throughout the time period corresponding to the length of the moving horizon. Once a threshold value of the relative error metric is exceeded, model re-identification is triggered using the most recent input/output data, and the model-based controller is updated to include the new model. This prevents constant updating of the model and the controller, but allows for updates when the model no longer captures the nonlinear dynamics of the process. The method is demonstrated using Lyapunov-based economic model predictive control [5] and shown to improve the profit and to reduce plant-model mismatch compared to using one linear empirical model for the entire period of operation both for the case of plant variations (modeled through catalyst deactivation) and an expansion of the region of process operation.

[1] Alanqar A, Ellis M, Christofides PD. Economic model predictive control of nonlinear process systems using empirical models. AIChE Journal. 2015;61:816-830.

[2] Qin SJ, Badgwell TA. A survey of industrial model predictive control technology. Control Engineering Practice. 2003;11:733-764.

[3] Larimore WE. Canonical variate analysis in identification, filtering, and adaptive control. In: Proceedings of the 29th IEEE Conference on Decision and Control. Honolulu, HI, 1990;596-604.

[4] Verhaegen M, Dewilde P. Subspace model identification Part 1. The output-error state-space model identification class of algorithms. International Journal of Control. 1992;56:1187-1210.

[5] Heidarinejad M, Liu J, Christofides PD. Economic model predictive control of nonlinear process systems using Lyapunov techniques. AIChE Journal. 2012;58:855-870.