Data-Driven Modeling of Nonlinear and Time-Varying Processes

Abstract

In this contribution, it is proposed using a novel interpretation of generalized ridge regression to identify a 1-dimensional grid of interdependent linear models from operation data. Such 1-dimensional model grids can be used to model repeated finite horizon. nonlinear and non-stationary process operations. These finite horizon process operations include chemical process operations such as start-ups, grade transitions, shut-dovnls, and of course batch, semi-batch and periodic processes. Explicitly, it is proposed to identify sets of interdependent linear lnodels using modified generalized ridge regression/Tikhonov regularization that penalizes weighted discrepancies between one linear model and the models in its neighborhood. Penalizing weighted discrepancies between neighboring linear models induce both rnodel interdependency and designed model properties.