Data-Driven Modeling of Batch Processes

Abstract

A one dimensional grid of interdependent linear models obtained from operation data is proposed for modeling repeated finite horizon, non linear and nonstationary process operations. Such finite horizon process operations include startups, grade transitions. shut-downs, and of course batch, semi-batch and periodic processes. The model grid is identified from data using a novel interpretation of generalized ridge regression that penalizes weighted discrepancies betweell one linear model and the models in its neighborhood. It is furthermore outlined how different representations of such a model grid may be used off-line as well as Oil-line. for prediction. monitoring, control, and optimization. Among these representations is a linear time-varying state space model which may be used for design in established linear monitoring and control methodologies.