Challenges in the Specification and Integration of Measurement Uncertainty in the Development of Data-Driven Models for the Chemical Processing Industry

• Published in: Industrial & engineering chemistry research 2015-09, Vol.54 (37), p.9159-9177
• Main Authors: Reis, Marco S, Rendall, Ricardo, Chin, Swee-Teng, Chiang, Leo
• Format: Article
• Language: English
• Subjects: Chemical processing industry; Computer simulation; Flexibility; Handles; Least squares method; Mathematical models; Regression; Uncertainty
• ISSN: 0888-5885; 1520-5045
• DOI: 10.1021/ie504577d

Obtaining and handling measurement uncertainty information is still a challenge in the chemical processing industries (CPI). From our experience, among the variables most affected by uncertainty, one typically finds the process outputs, comprising concentrations (main product and subproducts, reactants, etc.), measurements of quality properties (mechanical, chemical, etc.), or other relevant information about the end use of the product. With the increasing flexibility of processing units, these quantities can easily span different orders of magnitude and present rather different uncertainties associated with their measurements. This means that heteroscedasticity in the process outputs is a rather prevalent feature in CPI, which must be properly managed and integrated in all the tasks that make use of process data. In this article we address this critical problem, by considering two challenges: (i) specifying measurement uncertainty in heteroscedastic contexts; (ii) integrating uncertainty information on the outputs in model development. We present a critical review of the relevant literature regarding both challenges, and we address them through industrial case studies as well as an extensive Monte Carlo simulation study. For the first challenge, we provide solutions to handle the problem of specifying measurement uncertainties near physical limits (namely at very low concentrations, where intervals can easily include negative concentrations) and how to develop uncertainty models for heteroscedastic measurement devices spanning several orders of magnitude of the measurand. Regarding the second challenge, we conclude that prediction performance can be significantly improved by replacing tacitly homoscedastic modeling methodologies (ordinary least squares, partial least squares, and principal components regression) by alternative approaches that are able to explicitly handle the uncertainty in the outputs. Several important aspects of model development under heteroscedastic noisy conditions in the outputs are also highlighted and discussed.