(560b) Machine Learning-Based Model Predictive Control of Distributed Chemical Processes

Many important commodity and specialty chemicals are produced from highly exothermic processes in packed-bed tubular reactors, such as phthalic anhydride synthesis, ammonia synthesis and light olefin synthesis, at chemical manufacturing industry for several decades. In today's saturated markets where continuous improvement of process efficiency to reduce the production costs is necessary to keep chemical manufacturing plants from being obsoleted by global competition, packed-bed modeling and control have become an active research area for both academia and chemical manufacturing industry. Open literature reports suggest that the intrinsic characteristics of packed-bed tubular reactors with exothermic processes, i.e., nonlinearity, nonstationary (due to reactor fouling and catalyst deactivation), nonlinear distributed dynamics and limited on-line measurements with large delays, pose substantial difficulties to process control [1, 2, 3], and model-based control algorithms developed from more high-fidelity models yield better performance in practice [4]. Hence, the need for high-fidelity computationally efficient models to simulate the dynamic behavior of packed-bed tubular reactors controlled variables (e.g., the hot-spot temperature and product composition) as a function of the packed-bed tubular reactors manipulated inputs (e.g., the jacket temperature profile) becomes apparent. Traditional mechanistic models and computational fluid dynamics (CFD) models, although often yield adequate descriptions of the process dynamics, would not be suitable for designing model-based controllers because solving these models repeatedly in real-time as part of a model predictive control (MPC) optimization problem that seeks to reduce the hot-spot temperature and maximize the product composition may require significant computational time. Data-based modeling is an appealing alternative as data-driven models are computationally inexpensive and can have reasonable accuracy.

Motivated by the above considerations, the present work introduces a statistical-based machine learning algorithm that generates a computationally efficient Bayesian recurrent neural network (RNN) to simulate the dynamics of the hot-spot temperature and product composition at the exit of the packed-bed reactor as a function of the jacket temperature profile and the feed conditions. Next, a CFD model for the commercial-scale packed-bed tubular reactor used in phthalic anhydride synthesis is developed [1, 2, 3, 5] and is used to create a sufficiently large database, from which the Bayesian RNN is derived. Then, the Bayesian RNN is integrated within an MPC framework to create an MPC that optimizes the jacket temperature profile to maximize the product composition and to prevent the hot-spot temperature from exceeding a critical temperature value. Subsequently, the present work outlines a novel procedure that creates the communication pathway between a CFD solver and a nonlinear programming (NLP) solver so that CFD model for the packed-bed tubular reactor can be used to simulate the process dynamics of the physical reactor in the closed-loop (under machine-learning MPC) system. Finally, the closed-loop control system is subjected to various feed disturbances to demonstrate the effectiveness of the data-based control scheme proposed in this work.

[1] Hua, X., Jutan, A., 2000. Nonlinear inferential cascade control of exothermic fixed-bed reactors. AIChE journal 46, 980-996.

[2] Chen, C.Y., Sun, C.C., 1991. Adaptive inferential control of packed-bed reactors. Chemical engineering science 46 , 1041-1054.

[3] Wu, W., Huang, M.Y., 2003. Nonlinear inferential control for an exothermic packed-bed reactor: Geometric approaches. Chemical engineering science 58, 2023-2034.

[4] Logist, F., Smets, I.Y. and Van Impe, J.F., 2008. Derivation of generic optimal reference temperature profiles for steady-state exothermic jacketed tubular reactors. Journal of Process Control 18, 92-104.

[5] De Klerk, A., 2003. Voidage variation in packed beds at small column to particle diameter ratio. AIChE journal 49, 2022-2029.