Industrial & Engineering Chemistry Research
30 (12), pp 2555 - 2564
Results are given for the implementation of a static partial least-squares (PLS) regression estimator for product compositions on a high-purity pilot-plant distillation column. Temperatures on all 11 trays are used as inputs to the estimator. Several estimators were tested off line to compare their performance, and one estimator was used on line for dual composition control. It was found that the estimators perform very well when appropriate logarithmic transforms and scalings are used. Since the estimator is static, the implementation is straightforward. An estimator based only on experimental data gave excellent performance over a wide range of operating points. Estimators based on simulations did not perform quite as well, and the bias had to be adjusted when a change was made from one operating point to another. Nevertheless, since it may be difficult to obtain good experimental data in an industrial setting, this estimator is probably most useful in practice. In this paper we also discuss how to combine information from simulations (basic modeling) and experiments.