Dynamic analysis using NARX-based surrogate model for natural gas and hydrogen liquefaction process

초록

Data-driven surrogate models are increasingly recognized as effective tools for overcoming the limitations of conventional input-intensive dynamic modeling and for reducing computational costs associated with real-time decision-making and design improvements. In complex chemical systems, it is essential to select models that offer robustness and flexibility for accurate prediction and dynamic analysis. In this study, nonlinear autoregressive models with exogenous input (NARX)-based Gaussian process (GP) and neural network (NN) models were developed and evaluated using transient data from high-fidelity simulations. Cryogenic natural gas (NG) and hydrogen (H₂) liquefaction processes served as case studies. Comparative analysis showed that GP-NARX models required longer training times and achieved lower accuracy (16.48 %) than NN-NARX models (89.96 %). NN-NARX models provided acceptable accuracy and were effective for operability studies such as plant turndown and capacity expansions in NG liquefaction, with accurate predictions within trained data ranges. In H₂ liquefaction, closed-loop NN-NARX achieved moderate success for correlated profiles in multi-step-ahead predictions, but performed poorly for highly nonlinear or weakly correlated variables. NN-NARX was effective for one-step-ahead predictions in open-loop mode, making it useful for real-time applications, but faced challenges in closed-loop scenarios. The proposed NARX-based surrogate modeling framework offers a structured, scalable solution for dynamic process analysis. While demonstrated on cryogenic systems, the methodology is process-agnostic and can be applied to other nonlinear unit operations, providing a practical alternative to traditional high-fidelity dynamic modeling.

키워드