Adaptive control and optimization with unknown disturbances

This thesis deals with the application of adaptive control and optimization techniques to chemical processes. One of the main theoretical results of this thesis was to decrease the dependence of the stability of adaptive controllers on additional knowledge about process dynamics and disturbances using a new approach to select useful data and restart a converged estimator.; Two industrial case studies are studied in this thesis. The first is a glass furnace and the second is a silicon reactor. Plant data is used to study offiine model structure determination and adaptive control and optimization strategies. In addition, an industrial implementation of a simple adaptive predictive controller was applied to glass furnace temperature control. The large changes in the parameter estimates illustrate the need for an algorithm to stop the estimation when excitation is lacking.; Self-Tuning with Selective Memory is developed to prevent the problems of parameter drift and bursting effects with least squares adaptive control in the presence of unknown disturbances. The main idea is to measure the information available in each data point. Only data that leads to an increase in the information content are used. Stability proofs are also presented, and it is shown that the algorithm eventually stops estimating.; A method was needed to restart the estimation for the cases where the real parameters are not constant. The key here was to realize that the criteria for restart must be based on comparison with past performance using the current model. We call this Internal Model Audit (IMA). This contrasts published methods on multi-model adaptive control that are based on External Review (ER) of several models running in parallel. We show that ER leads to continuous switching which defeats the purpose of stopping the estimation.; An example using models identified from the glass furnace data is presented to show that the Adaptive Control with Selective Memory algorithm can track changes in plant parameters. The Selective Memory algorithm was also applied to the adaptive optimization of the silicon furnace. A steady state optimization technique was used along with the adaptive estimation of a nonlinear dynamic model.