SOFT SENSOR MODELS FOR A FRACTIONATION REFORMATE PLANT USING SMALL AND BOOTSTRAPPED DATA SETS

Abstract In refinery plants key process variables, like contents of process stream and various fuel properties, need to be continuously monitored using adequate on-line measuring devices. Such measuring devices are often unavailable or malfunction and, hence, laboratory assays which are irregular and time consuming and therefore not suitable for process control are inevitable alternatives. This research shows a comparison of different soft sensor models developed from a small industrial data set with soft sensor models developed from data generated by a bootstrap resampling method. Soft sensors were developed applying multiple linear regression, multivariable adaptive regression splines (MARSpline) and neural networks. The purpose of the developed soft sensors is the assessing of benzene content in light reformate of a fractionation reformate plant. The best results were obtained by the neural network-based model developed on bootstrapped data.