With process computers routinely collecting measurements on large numbers of process variables, statistical methods for the analysis, monitoring and diagnosis of process operating performance have received increasing attention in recent years. Producing quality products is an extremely important process control objective. However, achieving this objective can be very difficult in a continuous process, especially when quality measurements are not available on-line or have long time delays. Recent approaches to multivariate statistical process control, which utilise not only product quality data, but also the available process data are based on multivariate projection methods such as principle component analysis (PCA) and partial least squares (PLS).
Control of the flash point of light cycle oil (LCO) at the Caltex Lytton Refinery can be achieved through a process model, as online sensors are unavailable. Manipulation of the flash point can be achieved in a steam stripper unit, which removes the light ends of the LCO and improves the flash point. The existing multiple linear regression (MLR) flash point model has had limited success in flash point prediction, because the model suffers from the restrictions of an MLR model such as correlated variables. The model has been built upon several months of daily average process data and is used for both a model predictive controller and for assistance of operator control.
Large multivariate processes are difficult to monitor by traditional methods, and this project presents methods for the improvement of the existing MLR model through the application of a first principles model, a steady state PLS model and, a dynamic parametric model.
The success of each model type depends on the limitations of the control system and flash point sampling requirements. A design of experiments may be necessary to map the entire range of operation. Rather than risk producing large amounts of off-specification products, the steady state reference set has been developed by monitoring the process over a large period of time (6 months), so that most types of operating conditions are encountered.
A steady state PLS model has been constructed around the normal operating point of the process unit, and initial results have shown clear performance improvements over the existing MLR model. This should allow faster, more responsive controller performance and result in economic savings and product quality improvements. A HYSYS process simulation has also been created and methods for the future implementation of the model are discussed. In addition a number dynamic sampling programs have been investigated, although could not be carried out due to the excessive demand on resources at Caltex.