(434c) Smart Manufacturing and Optimal Operation of an Industrial Air Separation Unit Via Surrogate Modeling and Multiparametric Programming | AIChE

(434c) Smart Manufacturing and Optimal Operation of an Industrial Air Separation Unit Via Surrogate Modeling and Multiparametric Programming


Avraamidou, S., Texas A&M University
Ganesh, H. S., McKetta Department of Chemical Engineering, The University of Texas at Austin
Wang, Y., Linde PLC
Leyland, S., Process Systems Enterprise
Pistikopoulos, E., Texas A&M Energy Institute, Texas A&M University
Cao, Y., McMaster University
Air separation via cryogenic distillation is the basis for most industrially produced air products such as pure oxygen, nitrogen, and argon. These products are extensively used in the chemical industries for inert blanket gases in process vessels and feedstocks. On the industrial scale, Air Separation Units (ASU) typically involve cryogenic distillation processes and highly interconnected recycle loops to reduce utility costs. This process is highly complex and energy-intensive, with ASUs in the US accounting for more than 2\% of all US manufacturing sectors' electrical energy consumption [1].

This work focuses on improving the operational efficiency of an ASU plant though integrated scheduling and control. Firstly, a high-fidelity model of the ASU plant is developed, based on [2], from thermophysical relations to recreate the real-life ASU’s behaviour to act as the digital twin of the process. This digital twin is used for optimization of scheduling and optimal control applications as a part of the Greater CESMII Project. Due to the complexity of the digital twin (on the order of 10,000 variables) the online solution of a dynamic optimization problem is prohibitive. To this end, we create surrogate models to represent the original digital twin. The surrogate modeling approaches allow us to bypass the model complexity of the high-fidelity model in the optimization stage while still faithfully optimizing the system. This is done with a state-space and ReLU neural net surrogate models, which are embedded into a model predictive control (MPC) formulation to develop explicit MPCs for the entire process using multiparametric programming [6]. The developed controllers reduce the online burden of solving the optimal control problem by solving the entire map of solutions offline [4]. We then close the loop by deploying the developed controllers on the detailed high-fidelity model for tuning with prospects of employing them on the real industrial plant. The performance of the surrogate model based MPC controllers is demonstrated in this work. This approach is then coupled a plant wide scheduling optimizer that sets production targets on the day time scale, where the MPC tracks the production targets that the optimal scheduler while also abiding by purity and process constraints.

[1] United States Energy Information Administration, 2014, “Consumption of Energy for All purposes”, EIA, Washington, DC, accessed April 28, 2020,https://www.eia.gov/consumption/manufacturing/data/2014/#r2

[2] Pistikopoulos, E. N.; Diangelakis, N. A.; Oberdieck, R.; Papathanasiou, M.; Nascu I.; Muxin S.; PARAC – An integrated framework and software platform for the optimization and advanced model-based control of process systems Chemical Engineering Science 2015, 136, 115-138

[3] Pistikopoulos E. N. Perspectives in multiparametric programming and explicit model predictive control. AIChE Journal 2009, 55 (8), 1918-1925.

[4] Oberdieck, R.; Diangelakis, N. A.; Pistikopoulos, E. N. Explicit Model Predictive Control: A connected-graph approach. Automatica 2017, 76, 103-112.

[5] Cao, Y; Swartz, L. E. Christopher, Flores-Cerrillo, Jesus; Ma, Jingran Dynamic modeling and collocation-based model reduction of cryogenic air separation units. AIChE Journal 2016, 62 (5), 1602-1615

[6] Katz, J; Pappas, I; Avraamidou, S; Pistikopoulos, E. N; Integrating deep learning models and multiparametric programming. Computers & Chemical Engineering 2020, 136