A new termination criterion for sampling for surrogate model generation using partial least squares regression. (2nd February 2019)

• Record Type: Journal Article
• Authors: Straus, Julian; Skogestad, Sigurd
• Is Part Of: Computers & chemical engineering. Volume 121(2019)
• Page Start: 75
• Page End: 85
• Publication Date: 2019-02-02
• Subjects: Partial least squares regression -- Incremental sampling -- Surrogate model -- Optimization -- Integrated processes -- Machine learning -- Design of computer experiments -- Grey box model; Chemical engineering -- Data processing -- Periodicals; 660.0285
• DOI: 10.1016/j.compchemeng.2018.10.008
• Languages: English
• ISSNs: 0098-1354

Abstract

Highlights: Partial least square regression as novel termination criterion for surrogate model generation. Significant components in partial least square regression converge with increasing number of sampling points. This convergence corresponds to a convergence of the error of the fitted surrogate model and can be used as termination criterion. The application to three case studies shows similar behaviour with respect to the parameters of the termination criterion. Combination of the resulting surrogate models for optimization shows similar results as the detailed model. Abstract: This paper proposes a new incremental sampling method for the generation of surrogate models based on the application of partial least squares regression (PLSR) as a termination criterion. Compared to existing incremental and adaptive methods, the proposed method allows the sampling algorithm to stop without needing to fit a surrogate model at each iteration step. The proposed procedure was applied to a motivating pipe model and two case studies; the reaction and the separation section of an ammonia synthesis loop. In all cases, the new sampling method allows a small number of sampling points, corresponding to a regular grid with less than two points in each independent variable. The two surrogate models of the ammonia loop are combined for overall optimization. The optimum for the combined surrogate models is close to the optimum obtained with the original model.