Use of Machine Learning for Development of Data-Driven Predictive Control of PEM Electrolyzers

Résumé fourni par la source

This study examines performance loss in PEM electrolyzers under different operating regimes and proposes a predictive control framework. Systematic experiments assessed degradation through standardized profiles: stationary baselines, gradually dynamic conditions, and accelerated stress protocols that simulate operation with renewable energy sources. In addition to real-time data monitoring, analysis of polarization curves, electrochemical impedance spectroscopy, and cyclic voltammetry were used for performance diagnostic purposes. The results show that the performance loss is influenced by prolonged exposure to potentials greater than 1.95 V, which causes catalyst dissolution, and by extended dwell times at fixed potentials, even at nominal levels of 1.75 V, leading to an increase in membrane resistance. Machine learning used a part of the operating parameters measurements from the generated empirical data set to train a model, which were then used to predict the hydrogen production on the remaining part of the data set. The predictive data-based framework integrates operating parameters dynamics to forecast the efficiency losses under varying voltage profiles. The model achieved more than 95% accuracy in the prediction of hydrogen production, enabling optimization and adaptive control of PEMWE. Future work will focus on validation of the developed framework under real-world renewable profiles and scalability to multi-stack systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.