(438c) A Novel MPC Strategy by Adapting Disturbance Models

Model predictive control (MPC) is one of the advanced process control techniques that are widely used in industry. It calculates optimal values for manipulated variables based on future behaviors predicted by a model it maintains. When there are unmodeled disturbances, predictions of controlled variables made by the model may not be accurate. The bias between the current measured and estimated outputs is usually added to future predicted values to reduce prediction errors [1]. It is studied that this strategy could minimize prediction errors only when disturbance is a sequence of integrating white noise. An alternative approach would rather consider these disturbances as a result of model mismatch, and then applies robust MPC [2]. A limitation of the method is that unmeasured disturbances often bear dynamics; the uncertainty range could be very large. Recently, some adaptive MPC schemes have been proposed. Fukushima et al. [3] combines robust MPC with online parameter identification. Dougherty and Cooper [4] construct a controller composed of multiple linear MPCs adapting to changing operating points of nonlinear systems. While most adaptive MPC methods focus on models, rare attention has been paid to disturbances.

In this work, we propose a new MPC scheme that adapts the disturbance model. Usually process models are more complicated than disturbance dynamics; accurate identification and prediction of the process can hardly be achieved under closed-loop. Therefore, in our approach, only the disturbance model is adapted without the need for external perturbations. The disturbance model is updated by re-estimating a Kalman gain to give better predictions of future disturbances. MPC then takes advantage of these predictions to calculate optimal manipulated variables. Implementation of the proposed MPC strategy in commercial MPC packages such as DMCplus is discussed. In this case the objective is the ease of implementation rather than trying to be the most general form of formulation.

Performance of the proposed MPC strategy is studied by simulations based on the Wood-Berry distillation model. Results show that as compared with traditional MPC great improvements in controlled variables could be made. It is also found that process model mismatch can be compensated to some extent by adapting the disturbance model. Performance deteriorations caused by nonlinearities could be reduced.

References

[1] C. R Cutler, D. L. Ramaker. Dynamic matrix control?a computer control algorithm. Proceedings of the JACC 1980. San Francisco, CA.

[2] A. Bemporad, M. Morari. Robust model predictive control: A survey. Lecture notes in control and information sciences, pages 207?226. Springer-Verlag, 1999.

[3] H. Fukushima, T.-H. Kim, T. Sugie. Adaptive model predictive control for a class of constrained linear systems based on the comparison model. Automatica. 2007; 43:301-308.

[4] D. Dougherty, D. Cooper. A practical multiple model adaptive strategy for single-loop MPC. Control Engineering Practice. 2003; 11:141-159.