Multi-scenario long-term degradation prediction of PEMFC based on generative inference informer model

The long-term degradation prediction of proton exchange membrane fuel cell (PEMFC) has always been a difficult problem in fuel cells since it has experienced complex operating conditions during use. Although serial recurrent neural networks represented by long short-term memory (LSTM) can solve the long-term prediction problem in static operation or simple working condition, the forecasting accuracy under compound dynamic conditions still needs to be improved. To compensate for the shortcomings of the existing serial prediction methods, Informer, a generative inference prediction model, is introduced into the PEMFC life prediction for the first time. The model realizes parallel multi-step output by employing an encoder-decoder structure based on the multi-head ProbSparse self-attention mechanism. The cycle durability test and the urban bus road driving data validate the model's performance. The results show that the predicted MAPE of this model on two datasets is 0.2732 % and 0.5538 %, respectively, which has decreased by 77.37 % and 55.07 % based on LSTM's MAPE. In addition, the model has a strong information extraction ability and can predict future data that is 16 times longer with only a small amount of historical data, which has a high practical application value in the long-term degradation prediction of PEMFC.