(601d) A New Big Data Benchmark Problem: Fluid Catalytic Cracker Under Model Predictive Control

The chemical industry has been for decades a major generator of large-scale data sets. Yet, the means for analyzing these data have started to come of age fairly recently. Their application in transforming the “four Vs” of data (volume, velocity, veracity, variety) to a fifth and most important “V” – Value – is still lagging. In effect, many process operators feel that they are “drowning in data,” being overwhelmed by the sheer amount of data and the challenge of extracting information from them [1].

This work stems from the authors’ assumption that advances in big-data applications in the chemical industry would benefit considerably from the availability of a set of benchmark problems. We base this assumption on the tremendous impact that the definition of a set of “standardized” control problems, such as the Tennessee Eastman challenge process [2] and the Pensim simulator [3] have had on the development of plant-wide control techniques. Further, we note that the simulation models of these problems have been “borrowed” by the data analysis community and turned into machines for generating data sets used to develop and test new algorithms.

While these benchmark problems typically provide sufficient dynamic complexity and the opportunity to simulate multiple types of process faults, we note that they feature rather simple control approaches (in its standard form, the Tennessee Eastman plant is an open-loop system). As such, a key advance of the present work is to propose a benchmark model together with an advanced control system of the type used in industry, i.e., MPC. We believe that this development will provide a much richer environment for developing new data-driven approaches, such as plant-model mismatch detection, controller performance monitoring, etc.

Our focus is on modeling a Fluid Catalytic Cracking process (FCC). The starting point is the Model IV industrial cracker described by McFarlane et al.[4] and the computational implementation of this model by Ali [5]. The original system consists of a regenerator, preheater, reactor and air blowers. To make the model more realistic, the following extensions are made in this work. First, we add a fractionator (modeled as a steady state flash unit) to separate the cracking product. Further, we model the reactor (riser) as a dynamic unit and incorporate a complete model of the cracking reaction system, that captures the temperature dependence of the reaction rates and of the product distribution. Further, we model the dynamics of the most important actuators and implement a regulatory control system based on those found in practice. Subsequently, we carry out a system identification effort to build a state-space model of the entire system, and formulate a state-space model predictive controller. Several scenarios are tested in closed-loop, including setpoint tracking, disturbance rejection and the ability to deal with plant model mismatch.

The model and the controller are implemented in MATLAB (www.mathworks.com) and are available upon request and free of charge to researchers interested in using them for research in the big data realm.

References

[1] Venkatasubramanian, Venkat. "DROWNING IN DATA: Informatics and modeling challenges in a data‐rich networked world." AIChE J. 55, no. 1 (2009): 2-8.

[2] Downs, James J., and Ernest F. Vogel. "A plant-wide industrial process control problem." Computers & chemical engineering17, no. 3 (1993): 245-255.

[3] Birol, Gülnur, Cenk Ündey, and Ali Cinar. "A modular simulation package for fed-batch fermentation: penicillin production." Computers & Chemical Engineering 26, no. 11 (2002): 1553-1565.

[4] McFarlane, Randy C., Ralph C. Reineman, James F. Bartee, and Christos Georgakis. "Dynamic simulator for a model IV fluid catalytic cracking unit." Computers & chemical engineering 17, no. 3 (1993): 275-300.

[5] E. Ali, http://faculty.ksu.edu.sa/Emad.Ali/Pages/SimulinkModule.aspx3. Last accessed November 21, 2017