Nonlinear Multivariable Predictive Controls with Laguerre-based Wiener Neural Network Models and Partial Least Squares Model Reduction

Antti Pelkola

Research output: ThesisDoctoral ThesisMonograph

Abstract

This thesis considers dynamic modeling and control of nonlinear multivariable systems. Practical nonlinear control methods for industrial processes have been generated based on an orthogonal Laguerre-based Wiener-NN model structure. a priori information of the process can be included in the model structure by selecting a suitable orthogonal Laguerre basis, which gives a low dimensional and robust model structure with only a few parameters to be identified. Moreover, it gives in practice a solid basis for using NN's to describe the static nonlinear mapping in the Wiener model structure. This modeling approach is especially suitable for empirical modeling in the process industry, where a priori information can be included in the model structure to describe the dominant dynamics of the process as the basis for the actual identification. Based on this model structure a nonlinear control method has been generated, where the existing field-proven Linear Multivariable Predictive Control methods (LMPC's) can be utilized to produce corresponding Nonlinear Multivariable Predictive Control methods (NMPC's). The integrated nonlinear control method has been realized with two separate field-proven LMPC-methods; Model Algorithm Control (MAC) and Dynamic Matrix Control (DMC) methods, which have been modified into a Laguerre-based form to make the realizations straightforward. The integrated nonlinear control structure yields a good approach to solve nonlinear control problems in real-life processes using the existing familiar LMPC methods by adding, even gradually, identified nonlinearities into a "running" control structure. To achieve a more radical model reduction, the data-based orthogonalized Partial Least Squares (PLS) model structure has also been integrated with the same Wiener-NN model structure. Based on this model structure, the corresponding PLS-based nonlinear control methods have also been achieved via the integration with the existing LMPC methods (MAC and DMC). Using the PLS structure the control implementation can be realized using only separate SISO controllers according to the defined PLS dimensions. In the event of facing a lack of computing power caused by, for example, a required short control cycle time, this approach allows us to accomplish good practical control solutions. Comprehensive control performance tests have been executed utilizing nonlinear simulation models of a 2 x 2 ethanol-water distillation unit, an industrial scale neutralization process and a 5 x 7 heavy oil fractionator.
Translated title of the contributionEpälineaarisia ennustavia monimuuttujasäätöjä käyttäen Laguerre-pohjaisia Wiener neuroverkkomalleja ja osittaisen pienimmän neliösumman malliyksinkertaistusta
Original languageEnglish
QualificationDoctor's degree
Awarding Institution
  • Aalto University
Supervisors/Advisors
  • Visala, Arto, Supervising Professor
Publisher
Print ISBNs978-952-60-6755-1
Electronic ISBNs978-952-60-6756-8
Publication statusPublished - 2016
MoE publication typeG4 Doctoral dissertation (monograph)

Keywords

  • orthogonality
  • Laguerre approximation
  • Wiener model
  • multivariable predictive control
  • nonlinear control
  • model reduction
  • partial Least Squares
  • neural Networks
  • binary distillation
  • neutralization process
  • heavy oil fractionator

Fingerprint Dive into the research topics of 'Nonlinear Multivariable Predictive Controls with Laguerre-based Wiener Neural Network Models and Partial Least Squares Model Reduction'. Together they form a unique fingerprint.

Cite this