Experiences implementing the extended Kalman filter on an industrial batch reactor

Abstract

Model-based estimation techniques applied at the industrial scale present many difficulties that simulation studies typically do not address adequately. This work candidly assesses the industrial-scale feasibility of an on-line estimator for state estimation by applying it to an industrial pilot 1 m 3 semibatch reactor. A checklist is proposed to aid critical assessment of this and previous published industrial estimation applications. In this application, prediction of the batch end-point could be achieved using only standard instruments coupled with a simple fitted model, while independently the open-loop kinetic model gave satisfactory concentration estimates throughout the batch without need for measurement correction. However, combining heat-generaon measurement information with the dynamic reactor model, in the form of an extended Kalman filter, did not significantly improve the already reasonable state estimates. This was due to wide disparity between the qualities of the process and measurement models, long periods of poor system observability, and erratic quality of the state measurements necessary for model development, tuning, and verification. The experiences gained cast serious doubts on the usefulness of on-line estimation in industrial situations, compared to simple open-loop prediction or mere measurement model inversion, due to frequent predominance of either the dynamic or the measurement model. Further, employment of heat information for providing the measurement correction is deemed overly ambitious in view of excessive uncertainty incurred in closing a heat balance on industrial equipment. The paper concludes with several other less critical implications for industrial-scale estimation, which may serve as general recommendations and cautions for future industrial applications.