| This thesis describes the research that was performed to investigate the modelling and control of an industrial scale distillation column located at The Dow Chemical Company's Michigan Division. The overall objective of this thesis was to determine if advanced control would be beneficial to the column under study. This objective led to two major themes in the research: (1) develop a dynamic, multivariable model of the column, and (2) propose an alternate model based control scheme and test it through simulation.;Least squares parameter estimation of the identified model is based on real operating data. It was found that the concentration on the 57th tray exhibited inverse response to changing reflux flow. This was totally unexpected and has not been mentioned in the chemical engineering literature. Discussion of the Linde column model shows that other aspects of the model were very consistent with the experience of the operating personnel.;The alternate control strategy selected is a feedforward version of Internal Model Control, IMC. Two peculiarities of the column required extensions of existing theory on IMC, they are: (1) multirate sampling of the two product concentrations and (2) the inverse response already mentioned. The multirate sampling problem is addressed by a unique implementation of feedforward control. A reduced order controller design technique is developed to handle the non-minimum phase behavior.;The multivariable controller out performs the conventional controllers in load change simulations. Improved disturbance rejection was achieved in both product streams.;The model that is developed is a 2 input, 2 output matrix of discrete time transfer functions. A novel identification procedure is developed which greatly reduces the number of parameters necessary to describe the behavior of multivariable systems with multiple time delays and is based on a new "delayed polynomial matrix" representation of discrete systems. A computer algorithm for the delayed polynomial matrix method is described in a way that makes it suitable for interactive use. A simulated example demonstrates the ability of this new identification method to perform with up to 20% additive noise on the data. |
