LSTM-Based Prediction of System Dynamics in a Commercial AEM Water Electrolyzer Under Target Hydrogen Production Rates

Commercial anion exchange membrane water electrolyzers (AEMWE) typically do not provide detailed information about internal system parameters or stack characteristics, making physics-based modeling difficult and limiting their integration into customized digital simulation environments. Moreover, electrochemical and multiphysics models often require extensive computational resources, and parameter optimization can be prohibitively expensive. This study investigates the use of a long short-term memory (LSTM) network to predict the dynamic response of a commercial AEMWE under varying hydrogen production rates. The goal is to establish a data-driven model that replicates the operational logic of the real system, enabling the prediction of stack current, stack voltage, and hydrogen flow rate directly from the target production rate with significantly lower computational cost while preserving the transient behavior of the device. Model validation shows that the proposed LSTM achieves RMSE values of 0.48 for stack current, 0.92 for stack voltage, and 5.22 for hydrogen flow rate. The model successfully captures the overall dynamic trends and demonstrates that hydrogen flow rate and stack voltage can be inferred from the target production rate and stack current. Training is computationally efficient, allowing rapid model development. Although undershoot occurs during rapid decreases in production rate due to the smoothing characteristics of LSTM, the overall prediction error remains within acceptable bounds. The results highlight the potential of data-driven modeling for fast and practical AEMWE system representation, supporting future development of digital twins and accelerating applications in green hydrogen technologies.