TABLE OF CONTENTS This case deals with a refinery unit that included a number of reactors in parallel. The feed lines to these reactors did not have individual flow controls; hence there was maldistribution in the flow rates through them. This flow maldistribution was manifested in the form of a temperature maldistribution, resulting in additional reagent consumption in the reactors and poorer product quality control. The question to be answered was whether or not the process credits realized by correcting the flow maldistribution would justify the cost of installing the needed individual flow controllers.
The chemical reactions involved and the parallel reactors' hydrodynamics were too complicated for a cost effective and timely mechanistic modeling and simulation solution to be developed. Conventional statistical methods were unable to identify the effects of maldistribution on the reagent consumption from historical plant data.
Screening the historical plant data revealed significant noise in them. The filling of the reagent holding tanks was performed manually at the operators' discretion, resulting in considerable uncertainty as to the actual amount of reagent consumed at any given time. The neural net model developed to correlate the reagent consumption with the temperature maldistribution and other key process operating conditions was able to surmount the difficulties posed by the noise in the data. Fitting noise in the data was prevented by embedding constraints in the net while it was learning to extract patterns from the data in the process of developing its internal model. These constraints were based on rules...
