Green hydrogen production via PEM electrolysis is a cornerstone of the energy transition; however, its integration with highly dynamic renewable sources imposes severe challenges to the system's structural integrity. Extreme load variability makes it difficult to distinguish between normal transient responses and early indicators of catastrophic failures, such as gas crossover or external leakages. This work proposes a monitoring methodology based on "soft sensors" to enhance operational reliability. Utilizing the large-scale dataset from the National Laboratory of the Rockies (NLR), supervised machine learning models were applied to predict hydrogen mass flow in real-time, conditioned on process variables and oscillating input power. Residual analysis between the model's prediction and the physical Coriolis meter measurement allows for the identification of performance deviations and sensor failures. The results demonstrate that this data-driven monitoring layer enables a more dynamic Probabilistic Safety Assessment (PSA), reducing false alarms caused by intermittency and providing quantitative metrics for risk management and predictive maintenance of green hydrogen (GH2) systems.