Conditional invertible neural network for online-capable monitoring of polymer electrolyte membrane fuel cells

The operational management of proton exchange membrane fuel cells (PEMFCs) in alternative propulsion technology can enhance efficiency and reduce hydrogen consumption and degradation. This requires advanced measurement and control technology due to dynamic load changes, necessitating the recording of various state variables such as pressure, temperature, and humidity. A new online method of monitoring the behaviour of fuel cell systems was developed using a 0D–1D grey box model in MATLAB/Simulink, which was validated against real measurements. To enable online-capable prediction, a conditional invertible neural network (cINN) was trained on simulation data, effectively correlating simulation parameters with sensor outputs. Hyperparameter optimisation indicated that a configuration with 10 coupling blocks and 512 layers minimized errors. The neural network demonstrated the ability to detect anomalies with a resolution limit of 2.83σ and can predict voltage on low-power hardware, potentially complementing traditional monitoring methods and enabling virtual sensors for hard-to-measure variables.