Kinetic losses prediction of fuel cell by deep learning fusing characteristic frequency impedance

Online monitoring of kinetic processes in fuel cells is significant for extending their service life. However, traditional impedance spectrum measurements require a lengthy sweep time. This paper proposes a hybrid deep learning method for kinetic loss prediction that integrates characteristic-frequency impedance. First, based on impedance spectrum training data, 10 Hz and 316.23 Hz are identified as the characteristic frequencies through correlation analysis. The impedance magnitude and phase angle at these frequencies are fused with time-domain signals to construct the model input. Subsequently, the scSETCNTrans model, which integrates the temporal convolutional network (TCN), spatial-channel attention mechanism (scSE), and Transformer backbone is developed. Experimental results demonstrate that the model achieves optimal overall performance with an input sequence length of 60 s, yielding mean absolute percentage errors as low as 0.852%, 3.400% and 5.614%, for the ohmic loss, charge transfer loss and mass transfer loss, respectively.