Degradation prediction of a fuel cell-based CHP-system under dynamic load using feed forward neural networks

A combined heat and power (CHP) system with natural gas as primary energy and a hydrogen fuel cell as power source is modeled by implementing a feedforward neural network (FNN). It is shown that only the load profile and the resulting voltage are already sufficient to provide an accurate prediction of the voltage and its degradation under normal operation, excluding anomalies. The short-term reversible degradation is accurately modeled with high fidelity, whereas the irreversible long-term degradation remains more challenging to predict. The influence on the prediction is analyzed for different input features. Additionally, the size of the training dataset is varied and as a physical parameter the gas composition is incorporated into the model, allowing it to more accurately predict anomalies in the data. The presented approach is further tested with data from a system operated in a residential area. • A FNN voltage prediction model for a PEM fuel cell CHP-system is developed. • An approach to model dynamic operation using FNN is shown. • A good prediction for normal operation can be obtained without physical data. • Physical data can be used to find the causes for deviations from normal behavior.