Dynamic Surrogate Modeling for Multistep-ahead Prediction of Multivariate Nonlinear Chemical Processes

• Published in: Industrial & engineering chemistry research 2020-09, Vol.59 (35), p.15634-15655
• Main Authors: Shokry, Ahmed, Baraldi, Piero, Zio, Enrico, Espuña, Antonio
• Format: Article
• Language: English
• Subjects: Process Systems Engineering
• ISSN: 0888-5885; 1520-5045
• DOI: 10.1021/acs.iecr.0c00729

This work proposes a methodology for multivariate dynamic modeling and multistep-ahead prediction of nonlinear systems using surrogate models for the application to nonlinear chemical processes. The methodology provides a systematic and robust procedure for the development of data-driven dynamic models capable of predicting the process outputs over long time horizons. It is based on using surrogate models to construct several nonlinear autoregressive exogenous models (NARX) with each one approximating the future behavior of one process output as a function of the current and previous process inputs and outputs. The developed dynamic models are employed in a recursive schema to predict the process future outputs over several time steps (multistep-ahead prediction). The methodology is able to manage two different scenarios: (1) one in which a set of input–output signals collected from the process is only available for training and (2) another in which a mathematical model of the process is available and can be used to generate specific datasets for training. With respect to the latter, the proposed methodology includes a specific procedure for the selection of training data in dynamic modeling based on design of computer experiment (DOCE) techniques. The proposed methodology is applied to case studies from the process industry presented in the literature. The results show very high prediction accuracies over long time horizons. Also, owing to the flexibility, robustness, and computational efficiency of surrogate modeling, the methodology allows dealing with a wide range of situations, which would be difficult to address using first-principles models.