Development of soft-sensors for online quality prediction of sequential-reactor-multi-grade industrial processes

Abstract

Reliable online quality prediction of sequential-reactor-multi-grade (SRMG) chemical processes often encounters different challenges, including process nonlinearity, input variable selection/extraction, sequential relationship in reactors, and multiple grades in a production line. A novel just-in-time sequential nonlinear modeling method is proposed. It integrates input variable selection/extraction and quality prediction into a unified framework. First, the input variables in the previous reactors are substituted by “virtual” quality variables via least squares support vector regression (LSSVR) transform models. Then, the sequential relationship in a sequential-reactor process can be captured by a global sequential LSSVR model using an efficient training strategy. Furthermore, for a new test sample, an improved model is constructed by integrating just-in-time learning and the proposed sequential LSSVR model. Consequently, shifting into operating modes for multiple grades can perform better than a single global model. Finally, the proposed just-in-time sequential LSSVR (JS-LSSVR) model shows sequential, global–local, and quality-relevant characteristics for an SRMG process. The JS-LSSVR modeling method is applied to online prediction of melt index in an industrial polymerization production process in Taiwan. The prediction results show its superiority in terms of high prediction accuracy and reliability in comparison with other approaches.