Description |
-
Abstract: Mathematical models play an essential role for the purposes of process optimization and control. There are two major information sources for the development of these models: the knowledge of the process inner workings and the input-output data set. The model estimated using the detailed knowledge are called the knowledge-driven model. However, the inner workings of many industrial proces... read moreses are not always fully understood to enable the development of accurate knowledge-driven models. In such a situation, the data-driven model, relying on the input-output data, is an attractive alternative. Among varieties of data-driven modeling approaches, the Design of Dynamic Experiments (DoDE), a generalization of the traditional Design of Experiments (DoE) approach, has been demonstrated as an effective modeling methodology for optimizing nonlinear processes. When time-resolved data are obtainable during the experiments, developing a Dynamic Response Surface Methodology (DRSM) model is more favorable. As the estimated DRSM model with time-varying parameters captures the process dynamics, it has the potential to be applied for not only the process optimization but also the process control purposes. The main goal of this research work is to further advance and improve the two data-driven methodologies, the DoDE and the DRSM, to model, optimize and control nonlinear processes. We first proposed ways to incorporate prior process knowledge to improve the design of the input domain, in which the time-varying input of the DoDE experiments are selected. Improved process performance has been achieved in the refined input domain. In addition, as process optimization is usually under budgetary and time constraint, we developed an evolutionary DoDE approach to optimize the processes in a timely manner. The size of the initial set of experiments has been dramatically reduced while the achieved optimal process performance is similar to the one obtained using the original DoDE approach. To extend the applicability of the original DRSM approach (DRSM-1) to deal with processes with various and infinite time duration, we proposed a new DRSM approach (DRSM-2). The novelty of the DRSM-2 rests on a nonlinear transformation of time, the independent variable. Comparing to the DRSM-1, the new method has the following advantages. It is capable of 1) Modeling both continuous as well as batch processes, handling semi-infinite as easily as finite time domains 2) Using data that are not equidistant in time 3) Using data segments that are of varied durations due to possible strong nonlinearities in dynamics We also developed a single model approach, using the DRSM model, for both process optimization and control purposes. The proposed method reduces the experimental effort comparing to the current practices which use separate models for process optimization and control purposes, respectively. When the number of measurements is small, the proposed approach provides better control performance compared to the performance achieved using a model estimated with Pseudo Random Binary Signal (PRBS) data.
Thesis (Ph.D.)--Tufts University, 2018.
Submitted to the Dept. of Chemical and Biological Engineering.
Advisor: Christos Georgakis.
Committee: Eric Miller, David Schmidt, and Emmanuel Tzanakakis.
Keyword: Chemical engineering.read less
|
This object is in collection