Process control systems enable plants to meet objectives such as maintaining product quality and minimizing energy consumption, which in turn maximize the plant's profitability and rate-of-return or investment. Because they enable a rapid response to change, process control systems play a vital part in a company's ability to remain profitable in an uncertain economic climate. The first step in the design of an advanced control system is to build a model that represents the dynamics of the plant. Most plants are too complex or the underlying processes too poorly understood to be adequately modeled using first principles. The most reasonable way to obtain reliable dynamic models is from data generated through well designed experiments. In the petrochemical and refining industries, black-box models obtained from experiments are by far the most common means of obtaining dynamic models. The task of obtaining dynamic models from data is referred to as system identification. Control-relevant system identification is motivated by the desire to increase the utility and acceptance of advanced identification concepts in the process industries. It takes into consideration the closed-loop control objectives and the skill level of the user. Control-relevant identification therefore offers the opportunity for the migration of advanced identification concepts to nonexpert users and for the development of computer-aided design tools for identification that can be used by practicing engineers with a B.S. level of education. The main objective of this project is to investigate the subject of control-relevant identification. The basis for the control-relevant approach is the relationship between the design variables of the identification problem and the performance objective of the control problem. To obtain this, bias and variance expressions in the frequency domain and representations of the control problem in terms of linear fractional transformations are used. This analysis leads to a systematic procedure for prefilter design that substantially improves the performance of prediction-error algorithms without demanding substantial increases in skill from the user. In addition, use of the Structured Singular Value leads to a model validation procedure for identified models that provides a clear picture of model limitations to achievable control performance. By providing the theoretical basis for improved computer-aided design tools, these results should make the application of advanced identification concepts a more commonplace practice by engineers in the process industries.
