Model predictive control of a polyolefin reactor
Abstract
Two models for the copolymerization of propylene and ethylene in a fluidized bed reactor (FBR) were developed. A simple model for control; and a more advanced model which served as both a replacement for logged industrial plant data, and as a structurally different simulator to test the robustness of the controller. To ensure the best possible consistency between the two models, a set of parameters in the control model (CM) was estimated by fitting it to the plant replacement model (PRM). Additionally, an unscented Kalman filter (UKF) was set up to further mitigate the discrepancies between the models.
A nonlinear model predictive control (NMPC) strategy was applied to transition between two different polypropylene (PP) grades. Hard constraints were imposed on the manipulated variables (MVs), while soft constraints were applied on the controlled variables (CVs). The effect of utilizing inert feed as an MV was studied with the CM as the simulator, i.e., no mismatch between the model of the controller and the process simulator. By employing the inert feed, the control of pressure improved; however, more catalyst was required to keep the production at the desired level.
In order to demonstrate the effectiveness and robustness of NMPC, despite structural differences between the model and the plant, grade transitions were simulated with the PRM as the process simulator. The controller performed reasonably well, however, further tuning of both the controller and the UKF would result in more effective grade transitions. Nevertheless, the controller proved to be robust and coped well, notwithstanding the model mismatch.