A Comparative Study of Adaptive Soft Sensors for Quality Prediction in an Industrial Refining Hydrocracking Process

Abstract

Soft sensors have played indispensable roles in modern refining industry, which can provide significant information for process modeling, control, monitoring and optimization. However, the prediction performance often gradually deteriorates due to process time-varying problem caused by reasons like catalyst deactivation. Therefore, it is very important to update the inferential models regularly in order to keep good prediction performance. In this paper, a comparative study of adaptive soft sensors is carried out for quality prediction in a real hydrocracking process. Recursive partial least squares (RPLS), moving window RPLS (MWRPLS), locally weighted partial least squares (LWPLS) and moving window LWPLS (MWLWPLS) models are built to predict the 10% boiling point of the aviation kerosene product. The results show that RPLS and MWRPLS can provide better prediction performance.