This work focuses on developing a learning-based model
predictive
control (MPC) approach for regulating an ammonia synthesis reactor’s
hotspot temperature and outlet concentration dynamics. We create a
detailed multiscale-multiphysics dynamic model of the reactor, integrating
thermodynamics, kinetics, and transport phenomena to characterize
its spatiotemporal behavior at different scales. Despite its accuracy,
this model is computationally expensive, rendering it unsuitable for
real-time optimization-based decision-making methods like MPC. To
address this limitation, we investigate the advantages of deep recurrent
neural networks in alleviating the existing computational bottleneck
through model reduction. Specifically, we employ a long short-term
memory (LSTM) network to approximate the system dynamics and reduce
the computational load of the multiscale model. A large-scale time-series
simulation data set is generated for LSTM-based model training using
a multivariate parallel time-series approach. The LSTM-based model
is then utilized for MPC design, while the high-fidelity multiscale
model represents the reactor’s spatiotemporal dynamics. Using
the LSTM-based model comes at the cost of reduced prediction accuracy.
However, frequent feedback from the high-fidelity model at each sampling
time enables resetting the LSTM-based model and mitigates the prediction
errors.