TABLE OF CONTENTS The proof of the pudding is in the eating. However intriguing a technology may be, what sells in industry are solutions, not technology. This chapter is a prelude to the case studies that follow by way of outlining general modes of application, many of which are already in use.
Adaptive learning models that capture the operating behavior of process plants and units are very rapidly constructed by extracting underlying patterns from historical plant data. The phenomena governing process unit behavior are very complex, involving complicated unit-specific hydrodynamics, chemical reactions and catalyst-specific kinetics, thermodynamic and physical properties associated with the wide variety of feeds usually encountered in plant operations, etc. Conventional models and simulations are not only expensive and time-consuming to develop, but they are also not easy to modify quickly and cost-effectively when dealing with configuration and operational changes. By contrast, neural net-based adaptive learning models are very rapidly developed if adequate plant operating data are available, and can be used to predict performance robustly when engineering judgment and experience are used in constraining their behavior appropriately.
Neural nets are able to predict conversion and selectivity when trained on historical data comprising of operating conditions, e.g. flow rates, treat rates, temperatures, pressures, and easily obtained properties characterizing the feed, such as density, cut points, aniline point, sulfur and nitrogen content, etc. (see Figure 11.1).
Of particular interest and use is...
