TABLE OF CONTENTS Start-up idiosyncrasies, catalyst activity uncertainties, upsets, and other unplanned events that occur during the course of any petroleum or chemical plant run are the bane of plant engineers and managers. These unintended yet persistently occurring run anomalies make it difficult to predict the remaining course of the run as changing conditions considerably reduce the confidence one can place in pre-existing models. However, these difficulties can be alleviated by dynamically modeling reactor behavior "on the fly" during each run using an adaptive learning system. Such a system continually updates reactor models comprised of neural nets initially trained on historical plant data. The extrapolative capability of these models is enhanced by embedding constraints within them based on engineering judgment and operating experience. Constraining neural nets in this manner is essential for most industrial applications. This case study deals with predicting catalyst deactivation in a hydroprocessing reactor.
[1]Bhagat, P.M. & Ellis, E.S. (2000) In-Situ Learning while Refining: Reaction Engineering in the 21st Century, North American Symposium on Chemical Reaction Engineering (NASCRE 1), Houston, January 2000.
A hydroprocessing reactor is one in which hydrogen is added to the feed in the presence of a catalyst and then heated in order to obtain a desirable product. One of the main goals is to decrease the level of sulfur in the product. The catalyst in the reactor deactivates with time during this process. This deactivation is influenced by many factors, making its mechanistic determination extremely difficult. Catalyst deactivation...
