Control-oriented modeling of nonlinear process systems

This dissertation addresses the control-oriented modeling problem of process systems. It is well known in linear system identification that integrating system identification with control design has a synergistic effect on the overall closed-loop performance. Yet control-relevance in nonlinear system identification remains an open issue. A main focus of this dissertation is on the development of a systematic procedure for control-oriented identification of nonlinear process systems. Relevant issues in linear system identification and first principles modeling are also addressed.; As the first step, a detailed first-principles dynamic model of a pilot distillation column is developed. The model is built via tray-by-tray mass and energy balances. Required physical properties are estimated using an equation of state model. Fidelity of the column model is verified by experiment results on the real column. The model, along with the real column, provides an "industry-oriented" testbed for subsequent studies.; Tools required for nonlinear model control-relevance analysis are studied. These include inverse Volterra series models, internal model control design of nonlinear systems, and Volterra series representation of closed-loop systems. Using these tools, a control-relevant model reduction algorithm is presented.; Motivated by the need for simple nonlinear control strategies meaningful to process industries, a control-relevant nonlinear system identification method for restricted complexity models is further proposed. The method consists of two stages. In the first stage, using the orthogonal least squares method a Nonlinear AutoRegressive with eXogenous input (NARX) model is estimated from the noisy process measurements. An equivalent Volterra series model is then generated from the NARX model through a realization algorithm. In the second stage, the desired restricted complexity model is estimated through the control-relevant model reduction. A corresponding model validation procedure is implemented to ensure the closed-loop robust stability. Practical guidelines for the identification experiment design are also provided. It is shown that control design based on restricted complexity models often leads to a familiar proportional-integral-derivative (PID) controller augmented by some zero-memory nonlinear blocks. Applications to several industry-relevant systems indicate that using the proposed method leads to the similar synergistic effect.; The bias/variance trade-off and getting informative process measurements are two central issues in linear system identification practice. Geared for chemical/petrochemical process system applications, input/output variable selection poses another challenge. To address these crucial issues, correlation analysis of multivariable linear systems is investigated. Under mild assumptions, it is shown that the resulting nonparametric estimator forms a linear equation set with a symmetric parameter matrix. A consistent estimate of the system impulse responses can be readily obtained from the estimator, and all the input information necessary to characterize a linear system can be readily extracted from the parameter matrix. Applications of the estimator for bias/variance trade-off, online identification experiment monitoring and input/output variable selection are formulated.