Improved Identification Of Error-In-Variable Models For Industrial Process

This paper is concerned with the identification of linear parametric models of multivariate processes using Total Least Squares (TLS). Since the recorded reference data of the industrial process such as in coal fired power plants are usually corrupted by measurement uncertainty, this identification represents an Error-in-Variable (EIV) problem. TLS is known to address the EIV problem but fails to estimate the covariance matrix of the variable set that is used to predict the process output variables. In this work, a novel TLS technique is introduced, which (i) provides a consistent estimation of this matrix and (ii) is computationally more efficient than conventional TLS algorithms, and thus could be used for online data measurement and processing. These benefits were illustrated using a simulated SISO example and by application to in two application studies that relate to the simulation of a single-input single-output first order lag and recorded data from a power plant boiler process.