Subspace identification methods

Abstract

This thesisdealswith the subspace identificationmethods andmainly focuses on the utilization of prior information into their concept. Thework ismotivated by the fact, that in the realworld identification, there is often a strong prior information about the identified system, which can significantly improve the quality of the identified model and its compliance with the physical reality. The idea comes fromthe possibility to interpret subspace identification as a criterionmini- mization problemresulting to a state spacemodel,which is an optimalmulti-step predictor on the experimental data. The problems of predictor non-causality and over-parametrization are addressed and solved. Further, the optimal predictor is reformulated in the Bayesian frame- work, allowing to combine the available prior information with the information contained in the experimental data. Several types of prior information are considered and transformed to the probability den- sity function of prior parameters estimate by a covariance matrix shaping. A recursive version of algorithmis also presented and themain proposed identification algorithmis demonstrated on the practical application with the experimental data fromthe oil fired steamboiler with the rated effective power of 100MW.