A large-scale set of differential and algebraic equations (DAEs) is used to model and control an industrial gas phase polymerization reactor. Moving horizon estimation, preceding the nonlinear control optimization, provides an estimate of the current states and unmeasured disturbances. MHE is compared to Implicit Dynamic Feedback (IDFTM)*. With MHE, there is improved estimation of unmodeled disturbances in the UNIPOLTM** polyethylene plant. The polymerization model is converted to algebraic equations by orthogonal collocation and solved with the MHE objective function in a simultaneous optimization. NOVATM, an active-set sparse NLP solver, is used to converge the problem that has 46,870 variables. This large, sparse optimization problem is initiated every 5 minutes to update the model as new plant measurements become available and prior to the control optimization. The same plant model is used for nonlinear model predictive control (MPC) with 10 manipulated variables (MVs) and 26 controlled variables (CVs). In this case, MHE improves the control by better estimating reactor compositions of hydrogen, the condensing agent, and other measured state variables.
Ramhal et al. (Sat,) studied this question.