Selecting model features that lead to linear models of bi-product distillation towers

Abstract

This work introduces a methodology to develop linear data-driven steady-state models of bi-product distillation towers for use in operations monitoring and control. Linear monitoring models derived from measurements that are typically used for composition control (reboiler duty/bottoms flow, reflux flow/distillate flow, and one temperature measurement per tower section) have low accuracy due to the nonlinear behavior of distillation towers. We show that the addition of a judiciously selected second temperature measurement in each section leads to a highly accurate linear model. These temperatures are selected via an iterative procedure based on analyzing the residuals’ coefficient plots of PLS models. Remarkably, the models accurately predict product compositions, even without knowing the feed composition. Prediction of product composition for the multicomponent mixture feeds requires that the product flows to be included in the model. Our results demonstrate that the nonlinearity of the data-driven models can be reduced or eliminated by the proper selection of the model features.