Linear hybrid models of distillation towers

Abstract

In this work, we present a steady-state hybrid model composed of a simplified first principles model (SFPM) and an error correction model. The SFPM represents the internal structure of equipment and linear correction model, which is different from recent research using simpler forms of SFPM and nonlinear data-driven models. Retaining distillation tower structure in SFPM enables accurate calculation of sensitivities, composition correction model to be linear, and errors due to changes in tower efficiency can be corrected by an additional linear correction. This work uses multiplicative correction terms and shows that additive corrections often cause hybrid models to calculate negative concentrations. Due to the structure of the model, the parameters are estimated using PLS since PLS predictions were more accurate than neural networks. The correction models are linear, simple to update iteratively, and ensure accurate predictions for up to 10% changes in tower efficiency without updating the SFPM parameters.