Data-driven prognostic algorithm based on Bi-LSTM transformer model for proton exchange membrane fuel cell

The complex interaction of internal physical phenomena makes predicting the degradation of Proton Exchange Membrane Fuel Cells (PEMFCs) under dynamic operating conditions a critical but persistent challenge. To address this, the present work develops a hybrid model based on Bidirectional Long Short-Term Memory (Bi-LSTM) layers with a multi-head self-attention mechanism to capture degradation dependencies across both short and long time-frames. When benchmarked against state-of-the-art architectures, including attention-based LSTM variants and the Temporal Fusion Transformer (TFT), the model achieves a reduction of up to 52.7% in Root Mean Square Error (RMSE). The model’s accuracy is confirmed across single-step, multi-step, and long-term forecasting scenarios, where it consistently maintains high predictive accuracy (R$^2$ ≥ 0.998) for prediction horizons of up to 175 steps ahead. Furthermore, this high performance is achieved with notable data efficiency, requiring a training dataset of 300 h. These combined strengths in accuracy, long-term reliability, and data efficiency make the proposed Bi-LSTM Transformer an effective tool for predictive health management under dynamic load cycles.