(711h) Machine Learning-Based Model Predictive Control of an Electrified Steam Methane Reformer

Hydrogen gas is an important resource in the chemical industry offering various applications including facilitating decarbonization and electrification of various industrial applications [1]. Steam Methane Reforming (SMR), one of most widely used hydrogen production methods in industry can be improved by electrifying the heating for the highly endothermic reaction. At UCLA, an electrically heated SMR experimental setup was built to investigate the dynamic of an electrified SMR production process. In our previous work, a first-principal model was built utilizing a systematic framework based on the fundamental physical and chemical principles [2]. Model predictive control (MPC) was applied on the first-principal model to optimize the process performance. Despite this progress, it is important for reactor scale-up purposes to develop model predictive control schemes, relying fully on time-series data.

To capture the dynamic of nonlinear complex reaction process while improving the accuracy, machine learning methods like neural networks (NN) take advantage of determining unknown correlations between the reaction states through directly learning from experimental datasets generated under different reaction conditions (i.e., temperature, current) [3]. This study focuses on constructing a recurrent neural network (RNN) process model to capture the nonlinear complex dynamic behavior of the electrified SMR process, which enables the use of an MPC for real-time control over the current applied to the experimental reactor, thereby optimizing the hydrogen production. Specifically, the long short-term memory (LSTM) neural network structure is utilized within the RNN model that enables the capture of time dependencies within the sequential data, thereby enhancing the learning process. Additional techniques like dropout and L2 regularization are employed to further enhance the learning process. Eventually, the RNN process model is trained using actual experimental data along with simulated data. This RNN model is used for the implementation of an MPC which will optimize the hydrogen production by adjusting the current applied to the SMR reactor to reach an economically optimal operating condition. The controller design is evaluated through simulations and through experimental implementation on the UCLA electrified SMR process.

References:

[1] Ramachandran, R., & Menon, R. K. (1998). An overview of industrial uses of hydrogen. International journal of hydrogen energy, 23, 593-598.

[2] Çıtmacı, B., Cui, X., Abdullah, F., Richard, D., Peters, D., Wang, Y., ... & Christofides, P. D. (2024). Model predictive control of an electrically-heated steam methane reformer. Digital Chemical Engineering, 10, 100138.

[3] Nazemzadeh, N., Malanca, A. A., Nielsen, R. F., Gernaey, K. V., Andersson, M. P., & Mansouri, S. S. (2021). Integration of first-principle models and machine learning in a modeling framework: An application to flocculation. Chemical Engineering Science, 245, 116864.