Enhanced multi-feature and multi-step degradation prediction of proton exchange membrane fuel cells using a TCN-LSTM model with particle swarm optimization

Accurate long-horizon performance prediction of proton exchange membrane fuel cells (PEMFC) is crucial for enabling timely maintenance and ensuring reliable operation in real-world deployments. However, most existing data-driven models are restricted to single-step prediction, which fails to capture long-range degradation dynamics. This study develops a physically interpretable, non-recursive multi-step prediction framework that integrates Temporal Convolutional Networks (TCN) and Long Short-Term Memory (LSTM) networks, with hyperparameters adaptively optimized via Particle Swarm Optimization (PSO). Unlike conventional recursive or sliding-window approaches, the proposed non-recursive strategy enables direct multi-horizon prediction, effectively eliminating cumulative error propagation. A key novelty lies in the introduction of air outlet flow as a rarely explored yet physically meaningful degradation sensitive feature, which is systematically validated to enhance prediction accuracy and interpretability. The PSO-based adaptive optimization further improves model stability across different degradation stages. The PSO-based adaptive optimization further improves model stability. Extensive experiments on a 7-h voltage prediction task demonstrate that the proposed model improves root mean square error (RMSE) by 29.33%, mean absolute error (MAE) by 49.31%, and R 2 by 133.33% compared to baseline LSTM. The framework is further validated under 20-step look-ahead prediction and across three heterogeneous PEMFC degradation datasets, confirming its generalizability, robustness, and potential for real-time deployment. These results highlight the advantage of combining physics-informed feature engineering with advanced sequence modeling for accurate and stable long-term PEMFC health prognosis.