EARLY DETECTION OF HYDROGEN CROSSOVER IN PEM ELECTROLYZER USING EXPLAINABLE MACHINE LEARNING

This study addresses the safety and reliability challenges of proton exchange membrane water electrolysis integrated with intermittent renewable energy. Using a public dataset of a 1.25 MW commercial electrolyser, a framework was developed to identify transitions from steady state to dynamic crossover regimes. When prioritizing upstream operational precursors, we compared the Random Forest and XGBoost architectures, with the latter achieving a 96.36% higher recall and 3.6x faster inference. In addition to predictive performance, explainability methods were employed to validate the physical logic of the model. The results identify water resistivity, inlet temperature, and current drive variability as the main drivers of crossover risk. This alignment between AI inference and electrochemical theory demonstrates the potential of explainable AI for real-time monitoring of the integrity of green hydrogen systems.