Identification of pseudo-State Space Models for Batch Processes using Multivariate Statistical Methods

Abstract

A new methodology to identify models in a pseudo-state space form for batch/fed-batch processes is proposed. The methodology employs historical data from previous batch runs, where a few intermittent measurements of product quality were made, and multivariate statistical methods in order to identify data-based models. Multivariate statistical methods, such as principal components analysis (PCA) and partial least squares (PLS), are being increasingly employed for batch processes model identification due to the advantages they offer over more difficult and time-consuming first-principle modelling techniques. In the proposed model identification approach, predictors are obtained employing PCA and PLS algorithms. Then, after a new vector of pseudo-states is defined, a pseudo-state space model is identified by performing an algebraic manipulation of the PCA and PLS statistical models. The ability of the pseudo-state space models to accurately predict future process variable trajectories is demonstrated by means of a simulation benchmark for penicillin production.