State of health prediction for proton exchange membrane fuel cells combining semi-empirical model and machine learning

Durability is one of the main reasons limiting the large-scale application of fuel cells. Accurate prediction of state of health can help improve fuel cells safety and lifetime. This paper performs accurate health state prediction based on fuel cell semi-empirical model and machine learning algorithm. Firstly, based on the semi-empirical model, multiple stochastic conditions are imported and the voltage outputs are obtained. Voltage is selected as a health indicator characterizing the state of health for the fuel cells and is input to the neural network for training. The state of health prediction model is initially obtained. The model is then further validated and adjusted using real output voltages based on endurance test data. In the training process using model data and endurance test data, not only the output voltages under different operating conditions are used, but also the effects of different training weights on the prediction results are explored. This enables the selection of the most suitable training time and training accuracy. In the end, the lowest root mean squared error using simulated data is 2.48 mV and using test data is 8.88 mV.