8th International IFAC Symposium on Dynamics and Control of Process Systems INFERRING DISTILLATION PRODUCT COMPOSITION: A HYBRID SOFT SENSOR APPROACH I.Y. Smets ∗ S. Boon ∗ T. Boelen ∗∗∗ J. Espinosa ∗∗∗,∗∗∗∗ J.F. Van Impe ∗,∗∗ ∗ BioTeC, Dept. of Chemical Engineering, K.U.Leuven, W. de Croylaan 46, B-3001 Leuven, Belgium ∗∗ Corresponding author: jan.vanimpe@cit.kuleuven.be ∗∗∗ IPCOS, Technologielaan 11-0101, B-3001 Leuven, Belgium ∗∗∗∗ Faculty of Mines, Universidad Nacional de Colombia, Cra.80 No.65-223, Medell´ ın, Colombia Abstract: Adequate process monitoring and optimal control of distillation columns relies heavily on accurate and preferably on-line estimates of the product composition. Hence, inferring the product composition from easily accessible and abundantly available process measurements has become a key element for successful operation. This paper compares a hybrid soft sensor approach, based on the General Distillation Shortcut method introduced by Friedman in 1995, with a pure black box approach. On the basis of two industrial multicomponent distillation case studies, it can be concluded that the hybrid GDS approach outperforms the black box one if (i ) a temperature measurement is available that is sensitive for the to be predicted concentration and (ii ) if that concentration is present in a substantial amount with respect to the other components. The black box soft sensors do not suffer from that last drawback but, once again, their lack of extrapolative power is clearly illustrated. Copyright c 2007 IFAC. Keywords: distillation columns, soft sensing, monitoring, hybrid modeling, black box modeling, industrial case study 1. INTRODUCTION For adequate process monitoring and control, an accurate estimation of the product compo- sitions during distillation is a prerequisite. Al- though product composition can be measured on- line, most analyzers, like gas chromatographs and NIR (Near-Infrared) analyzers are expensive and difficult to maintain. Furthermore, they entail sig- nificant measurement delays precluding in time control actions. Hence, inferring the product composition from easily accessible and abundantly available pro- cess measurements has become a key element for successful operation. The development of such inferential or soft sensor controllers is far from new but remains highly relevant as witnessed by recent publications in this domain. The type of model on which the soft sensors rely, varies from first principles, mechanistic models to black box models. As correctly formulated by Kano et al. (2000) a first principles model is preferred as far as it is available and provides sufficient accuracy with reasonable computational load. Predominantly, strategies based on Extended Kalman Filters are proposed (e.g., (Baratti et al., 1995; Baratti et al., 1997; Lee and Morari, 1992; Oisiovici and Cruz, 2001)), but, more recently, also a combi- nation of the wave propagation equation with static mass and energy balances has been reported (Roffel et al., 2003). If, however, no fundamental model appropriate for real-time use exists, an empirical model must be derived from process data. With the huge amount of data that is nowadays stored in computers, this black box modeling represents a feasible challenge. At present, techniques based on multivariate statistics such as Partial Least Squares (PLS) Preprints Vol.3, June 6-8, 2007, Cancún, Mexico 169