Integrated learning-based estimation and nonlinear predictive control of an ammonia synthesis reactor

Abstract

This paper presents an advanced machine learning-based framework designed for predictive modeling, state estimation, and feedback control of ammonia synthesis reactor dynamics. A high-fidelity two-dimensional multiphysics model is employed to generate a comprehensive dataset that captures variable dynamics under various operational conditions. Surrogate long short-term memory neural networks are trained to enable real-time predictions and model-based control. Additionally, a feedforward neural network is developed to estimate the outlet ammonia concentration and hotspot temperature using spatially distributed temperature readings, thereby addressing the challenges associated with real-time concentration and maximum temperature measurements. The machine learning-based predictive modeling and state estimation methods are integrated into a model predictive control architecture to regulate ammonia synthesis. Simulation results demonstrate that the machine learning surrogates accurately represent the nonlinear process dynamics with minimal discrepancy while reducing optimization costs compared to the high-fidelity model, ensuring adaptability and effective guidance of the reactor to desired set points.