(472b) Hybrid Models for Optimization of Hydrocarbon Separation Equipment

Objective of our work is to develop highly accurate and at the same time simple models that can be used for monitoring of plant operation, for optimization of operating conditions, or for production scheduling and planning. Models developed in this work (so called ?hybrid models?) consist of two parts:

(i) material and energy balances that determine external and internal material and energy flows in the equipment and

(ii) linear PLS models that predict product properties based on the properties of the feed and the internal equipment flows.

These models that can be used for both operations improvement and for production planning.

Prior work on approximate modeling of distillation towers can be classified into models that aim to improve actual operations and models that aim to improve planning results.

Mejdell and Skogestad (1991) used all tray temperature measurements to estimate product compositions, in order to emphasize that all variable measurements should be used and to exclude the influence of measurement selection. Linear multivariate Principal Component Regression (PCR) and Partial Least Squared (PLS) estimators performed well, even for multicomponent mixtures, pressure variations, and nonlinearity caused by changes in operating conditions. To deal with nonlinearity they found that logarithmic transformation of temperatures and compositions significantly improve the estimator performance.

Hybrid models of a binary distillation tower presented by Safavi, Nooraii, and Romagnoli (1999) combined a column separation factor with mass and energy balances, thereby enabling optimization of the tower operation. The separation factor was represented via a wave-net model.

Lee, Lee and Park (2000) focused on predicting qualities of the products from a multi-product distillation tower. They approximated TBP of component of feed and products by probability functions and found that the choice of the probability function is rather arbitrary and others various distribution functions or combined functions can also be used. They correlated parameters of the probability function with operating conditions by using an inferential modeling technique such as Partial Least Squared (PLS) regression analysis.

Li,Hui and Li (2005) optimized cutpoints of a crude distillation unit (CDU) by determining the size of the swing cut expressed by Weight Transfer Ratio (WTR) which is calculated from empirical procedure described by Watkins (1979).

Current work addresses separation of multicomponent mixtures (feedstocks and products) that can be described by distillation curves, such as true boiling point (TBP) curve. PLS model describing separation uses as independent variables selected point on the feed TBP curve and internal reflux ratio at selected trays. The model predicts points on the products TBP curves. Material and energy balances are used to calculate internal flows in the equipment, thereby linking PLS separation models with the overall product yields. Hybrid models have been developed for typical separation equipment (absorber, stripper, two product distillation, and for an atmospheric pipestill). The models predict product properties and internal equipment flows with approx. 0.5% error with respect to rigorous models of the same equipment. Methodology for building such models has been generalized and can be readily applied to specific separation equipment.

Since these models are mostly linear, they are suitable for use in production planning, scheduling, on-line monitoring, and for on-line optimization. The models offer a significant accuracy improvement over the models currently used in production planning. Moreover, they enable building of planning (or scheduling) models that include energy balances (in addition to material balances). Using the same model will shorten the implementation time, simplify maintenance, and ensure consistency of answers between different business applications.