(752d) Modeling of Thermal Cracking of Natural Gas Liquids: Optimization with Support Vector Machine-Based Constraint Handling Scheme for Stiff ODEs

Rapid increase in shale gas production in the Appalachian Basin (Marcellus and Utica Shales) for the past 5 years has led to a significant growth in natural gas liquids (NGLs) production as well as a projected increase of these petrochemical feedstocks in the upcoming years [1, 2]. In this perspective, many existing ethylene crackers have expanded capacity and new crackers are becoming online to benefit from this unique opportunity [3]. Hence, the mathematical modeling and optimization of this process emerges as a necessity so as to determine the optimal operating conditions and the reactor configurations for the steam cracker, in such a way that the profit from thermal cracking of NGLs is maximized.

In this work, we are demonstrating (i) the comprehensive first-principles modeling of an adiabatic propane steam cracking process as a 1D plug flow reactor with coking effects [4] and (ii) the data-driven optimization of this reactor model using the ARGONAUT [5-6] framework. The mathematical model for thermal cracking includes the reaction rates expressed with Arrhenius equation, as well as the ODEs for continuity, energy, and momentum balances. The kinetics of thermal cracking of propane is adapted from Sundaram and Froment (1979) and the cracking model is validated using industrial data [7]. This simulation model is later used in the data-driven optimization phase to collect input-output data. However, the stiffness in the reactor models poses an additional complexity in retrieving the optimal operating conditions for the cracking process, especially when the data-driven optimization strategies strongly rely on collecting numerically stable samples from the reactor simulator. To this end, a novel Support Vector Machine (SVM)-based filtering approach is applied to eliminate candidate sampling points that are susceptible to the stiffness in the model and that will result with an abrupt termination of the cracking simulation due to high pressure drop. The SVM classifier identifies and removes the numerically unstable candidate points a priori to simulation execution by formulating an optimal separating hyperplane between stable/unstable points. This enables the data-driven algorithm to consistently converge to a profitable and feasible reactor configuration for the propane cracking process with an improved objective function value.

References

[1] Energy Information Administration. Today in Energy: Appalachia, Permian, Haynesville drive U.S. natural gas production growth, August 2018. URL https://www.eia.gov/todayinenergy/detail.php?id=36934. (accessed 04.10.2019).

[2] Energy Information Administration. Short-term energy outlook, April 2019. URL https://www.eia.gov/outlooks/steo/. (accessed 04.10.2019).

[3] Energy Information Administration. U.S. ethane consumption, exports to increase as new petrochemical plants come online, February 2018. URL https://www.eia.gov/todayinenergy/detail.php?id=35012. (accessed 02.26.2018).

[4] O. Onel. (2017). Advances in Modeling, Synthesis, And Global Optimization of Hybrid Energy Systems Toward the Production of Liquid Fuels and Olefins. Princeton University.

[5] F. Boukouvala and C. A. Floudas. (2017). ARGONAUT: AlgoRithms for Global Optimization of coNstrAined grey-box compUTational problems. Optimization Letters, 11(5), 895-913.

[6] B. Beykal, F. Boukouvala, C. A. Floudas, N. Sorek, H. Zalavadia, E. Gildin. (2018). Global optimization of grey-box computational systems using surrogate functions and application to highly constrained oil-field operations. Computers & Chemical Engineering, 114, 99-110.

[7] K. M. Sundaram and G. F. Froment. (1979). Kinetics of coke deposition in the thermal cracking of propane. Chemical Engineering Science, 34, 635-644.