A hybrid method for estimation of PEMFC state of health

To estimate the state of health (SOH) of proton exchange membrane fuel cell (PEMFC), a hybrid method fusing the model-based particle filter method and data-based long short-term memory (LSTM) neural network method is proposed. The hybrid method is applied to estimate the PEMFC aging rate and power-current curve based on power degradation. The estimated results of LSTM neural network are considered as the observation in the prediction stage of particle filter method, and the state parameters are updated by the posterior estimation of semi-empirical power degradation model and particle filter method. Using a sliding window, the hybrid method can be updated iteratively with the accumulation of operating data. The results show that the average output power root mean square error (RMSE) of model-based method and hybrid method are 0.393 W and 0.203 W respectively in ten prediction windows. Moreover, the average output power RMSE is reduced by 36.0% in ten prediction windows. The average RMSE of power-current curve between estimated results and experimental data of hybrid method is 45% at 100 h, 32% at515h and 14% at 666h less than that of model-based method, respectively.