Sensitivity analysis of a CNN–LSTM prognostic framework for PEMFCs: Effects of sliding window size and training data allocation

For proton exchange membrane fuel cell (PEMFC) prognostics, deploying deep learning models in real applications depends not only on the network architecture but also on carefully chosen hyperparameters and training strategies. Hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) models can attain high predictive accuracy, yet their sensitivity to practical implementation choices has not been systematically quantified. This work addresses that gap by performing a methodological assessment of a representative CNN–LSTM framework rather than proposing a new architecture. Using the static-load IEEE PHM 2014 FC1 dataset as a controlled benchmark, we examine how two key factors—sliding window length and training data partitioning—jointly affect short-term accuracy and long-term forecast stability. Our results show that the hybrid CNN–LSTM reduces the Root Mean Square Error (RMSE) of short-term voltage prediction by 49% compared with a standalone LSTM baseline. For recursive long-horizon lifetime forecasting, a moderately sized sliding window of 40 hours is identified as optimal, keeping the remaining useful life prediction error within ±5 hours. In addition, the model exhibits strong robustness with respect to training set size: once more than 500 hours of data are used for training, the variation in RMSE remains below 0.001. By confining the analysis to a static-load test case, we isolate the influence of these implementation parameters and provide practical, data-driven guidance for configuring CNN–LSTM-based prognostic models in PEMFC health management. The proposed validation methodology can be extended in future work to dynamic-load scenarios.