Digital Twin and Machine Learning-Based Diagnostics for PEM Electrolyzer

The degradation of the health state of Proton Exchange Membrane (PEM) water electrolyzer, caused by power supply variability, operating temperature changes, and other chemical factors, represents a major challenge for green hydrogen production efficiency. This paper presents an advanced hybrid system combining a digital twin and machine learning, enabling real-time anomaly detection of a PEM electrolyzer. This intelligent approach allows for the real-time prediction of operating parameters, namely current, voltage, and hydrogen flow rate, via Azure Machine Learning, and their visualization within the system's digital twin via Azure Digital Twins. Furthermore, the comparison between simulated data from the digital twin and those predicted by machine learning enables the anticipation of PEM electrolyzer anomalies. The selected prediction models rely on the Extreme Random Trees algorithm for current and voltage estimation, and on the Elastic Net algorithm for hydrogen flow rate prediction. The obtained results confirm the robustness of the proposed approach, with coefficients of determination of 0.99820, 0.99693, and 0.99665 for current, voltage, and hydrogen flow rate respectively, associated with Normalized Root Mean Square Errors (NRMSE) of 0.00870, 0.011278, and 0.11087. This high accuracy provides the digital twin with the capability to anticipate failures and extend the PEM electrolyzer's lifespan, with a view to optimizing the global efficiency of green hydrogen production.