A novel run-to-run optimization algorithm for batch processes using localized partial least squares regression models

Driving a process to optimal conditions under various uncertainties is a key issue for meeting objectives of productivity and quality of batch or fed-batch product. To overcome a limitation of two-step approaches unable to cope with nonparametric or large uncertainty, several gradient based iterative optimization methods have been proposed. Among these, latent variable model based approaches have a strong point that it can estimate plant gradient with less number of excitation than common model free ones when the variables are correlated. In this paper, a novel run-to-run optimization approach using partial least squares regression is proposed. Being different from the last latent variable model based ones, the proposed method excites and utilizes only interpolating data set for gradient estimation by introducing trust region. And consideration of input constraints are also added in the iterative updates for optimization. Finally, a case study in which a fed-batch bio-reactor's feeding rates are optimized validates the proposed method's utility.