Doctoral Dissertations

Date of Award

5-1991

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Chemical Engineering

Major Professor

C. F. Moore

Committee Members

Elden DePorter, Duane Bruns, Pete Counce, E. Vogel

Abstract

Plants in the chemical process industry have become more difficult to control as they have increased in complexity and come to produce a wider variety of products. Within the last two decades, the chemical process industry has attempted to deal with some of the more difficult control problems through model based control schemes. This dissertation develops a nonlinear model predictive control scheme and addresses some of the practical issues related to using feedback in such schemes. This work examines several areas of continued research with model predictive control (MFC) schemes; feedback analysis, process signal filtering, and unmeasured state variable estimation. The work presented in this dissertation deals with these areas using a technique called feedback cultivation. Feedback cultivation extracts temporal (time-related) information from a historical database to reduce the amount of data to be examined and to enhance the dynamic characteristics present in the historical data. Other tools used in conjunction with feedback cultivation include neural networks and adaptive filtering. Neural networks are one form of pattern recognition that has seen increased use recently. Adaptive signal filters remove excessive noise from the process signal measurements. The results presented in this work demonstrate that the feedback cultivation algorithm successfully incorporates feedback information into the nonlinear optimal control algorithm. Using the neural network. process-to-model mismatch and unmeasured disturbances can be diagnosed and treated accordingly. In the case of process-to-model mismatch, a model parameter re-identification algorithm updates the inaccurate model parameter. The correction for unmeasured disturbances involves determining new values for model output variable bias terms using the model parameter re-identication algorithm. The model parameter re-identification algorithm involves the solution of a modified form of the nonlinear optimal control problem. The results presented in this work involve applications of the nonlinear optimal control algorithm with feedback cultivation on a simple first-order process with a first-order model and on a continuous stirred tank reactor with a first principle nonlinear model.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS