Learning-based estimation and predictive control of an ammonia synthesis reactor

This study introduces a dynamic data-driven model predictive control (MPC) framework for an ammonia synthesis packed-bed reactor (PBR). In response to the computational demands associated with high-fidelity modeling of the PBR, we propose a surrogate model based on long short-term memory (LSTM). This LSTM-based model adeptly captures the nonlinear dynamics of the process by leveraging information from offline computational fluid dynamics (CFD) simulations. While this surrogate model effectively mitigates the computational challenges associated with MPC, constraints emerge due to limitations in real-time measurements of maximum temperature and outlet ammonia concentration. To address these constraints, we present a feedforward neural network (FNN)-based state observer that utilizes temperature sensor values spatially distributed throughout the reactor’s length to estimate maximum temperature and ammonia outlet concentration. The FNN-based state observer effectively estimates these key controlled outputs, facilitating MPC implementation and achieving optimal closed-loop performance in scenarios where real-time measurements are challenging to obtain.