Date of Award
5-1999
Document Type
Thesis
Abstract
Linear Identification and control strategies suffer from the inadequacy of capturing the inherently nonlinear phenomena of chemical processes, and are valid only in a narrow region around a chosen operating point.
Hence, the goal of this dissertation is be to design nonlinear identification and control strategies using neural networks. Neural networks are identification paradigm that mimics the brain and can capture the underlying nonlinearities of a given data.
The Methodology undertaken is to decompose a complex nonlinear problem into less complex nonlinear problems that can easily be handled by multiple neural network models. Two strategies are developed in this dissertation: operating regime decomposition for single-input single-output (SISO) processes and dynamic neural network partial least squares for multiple-input multiple-output (MIMO) processes.
Operating regime decomposition addresses the problem of identification of a nonlinear process operating in multiple regimes. The global space is divided into multiple local regimes, a nonlinear model is developed for each regime, and a quadratic programming based algorithm is
used to ensure smooth transition between the regimes on-line. The use of nonlinear models as opposed to linear models reduces the number of regimes needed. The combination of the nonlinear models with the switching algorithm improves transient performance. The identification strategies also show positive potential on the control of a reactor around its unstable steady state using nonlinear model predictive controllers.
A dynamic neural network partial least squares is developed to address the problem of identification and control of nonlinear MIMO processes: using a transformation, the MIMO problem is decomposed into SISO problems in a latent space. Nonlinear SI SO identification and control theories using neural network, that are easier to apply than their MIMO counterpart, are then applied to the individual SISO problems. The transformation decouples the MIMO process while capturing the necessary interactions of the process in the latent space.
The last contribution of this dissertation is the usage of the dynamic neural network partial least square in the open-loop identification of a MIMO process from closed-loop generated data. Satisfactory open loop models are obtained for a 2x2 polymerization reactor and a 3x3 numerical process.
Recommended Citation
Adebiyi, Olufemi, "Decomposition Strategies for Process Identification and Control Using Neural Networks" (1999). LSU Historical Dissertations and Theses. 8362.
https://repository.lsu.edu/gradschool_disstheses/8362
DOI
10.31390/gradschool_disstheses.8362