Development of inferential distillation models using multivariate statistical methods

Chemical processes often have many variables that are being monitored every minute or every second. This can result in "data overload" and useful information that is buried within the collection of data is lost. Techniques that provide a quick method to extract information from large sets of data can prove to be very beneficial. In many cases, however, the data collected from processes are redundant, or highly correlated. In the paper, inferential models for estimating product compositions are built by using partial least squares (PLS) regression, based on simulated time series data. The PLS method removes the correlation problem by projecting the original variable space to an orthogonal latent space. A debutanizer column is used as a case study and the results of the PLS method are compared to another two multivariate statistical methods, which are multiple linear regression and principal components regression