APPLICATIONS OF MULTIVARIATE MODELING IN THE HYDROCARBON INDUSTRY

In large process plants the process control computer systems are the depository of large amounts of
operational data, rich in knowledge content. The information obtained from such data can be used for a
variety of purposes such as inferential measurements and model predictive control. However careful
variable selection and data preprocessing is required for developing adequate models from this data. The
objective of this paper is to examine in detail the methods to be adopted for developing successful
empirical models from plant data. Three case studies have been presented from the hydrocarbon
industry. The first case study deals with the development of a heat exchanger model by neural networks
to be used in model predictive control. The second case study deals with the development of a soft
sensor for predicting propane concentration in a depropaniser column. The third case study deals with
development of a heat exchanger fouling model to be used as part of a preventive maintenance tool. In
all the cases statistical model adequacy tests showed that careful selection of variables, data
preprocessing and post modeling analysis helped in developing models which were adequate for the
intended purposes.