To enhance the accuracy and robustness of estimating the remaining useful life (RUL) of proton exchange membrane fuel cells (PEMFC)

an error-weighted stacked ensemble model is proposed in this paper. The data quality is improved by preprocessing with Savitzky-Golay (SG) filter and feature selection using eXtreme Gradient Boosting (XGBoost). The model integrates the deep belief network (DBN)

the gated recurrent unit (GRU)

and the temporal convolutional network (TCN). Prediction weights are adjusted based on the errors of each constituent model and combined by using random forest (RF) to achieve optimal results. Comparative experimental analysis shows that this integrated approach significantly improves the prediction accuracy of the degradation trend and RUL of PEMFC.