Scholarlink

<p>Accurate and reliable degradation prediction is essential for ensuring the long-term performance and safety of proton exchange membrane fuel cells (PEMFCs) operating under dynamic conditions. Existing data-driven approaches often struggle to capture long-range degradation dependencies, neglect non-stationary aging behavior, or provide only deterministic predictions without uncertainty information. To address these limitations, this paper proposes an uncertainty-aware degradation prediction framework that integrates a Patch-Based Transformer (PatchTST) with a degradation-aware weighted quantile random forest. PatchTST is employed to learn robust temporal degradation representations from multivariate operational data, while the proposed degradation-aware weighting strategy emphasizes late-life aging behavior and quantile regression enables probabilistic prediction with calibrated uncertainty intervals. Experimental evaluation on a publicly available PEMFC dataset demonstrates that the proposed method achieves superior performance, with an RMSE of <b>0.0018</b>, an MAE of <b>0.0022</b>, and a MAPE of <b>0.028</b>, outperforming representative deep learning and tree-based baselines. In addition, the proposed framework attains a prediction interval coverage probability of <b>0.91</b> with a narrow mean prediction interval width of <b>0.063</b>, confirming its reliability for uncertainty-aware prognostics. The results indicate that the proposed approach provides an effective and robust solution for PEM fuel cell degradation prediction and health monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Uncertainty-aware degradation prediction of PEM fuel cells using PatchTST and weighted quantile random forests

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月21日

查看全文
  • Zainab Imad Al-Tamimi,
  • Abdullahi Ibrahim

摘要

Accurate and reliable degradation prediction is essential for ensuring the long-term performance and safety of proton exchange membrane fuel cells (PEMFCs) operating under dynamic conditions. Existing data-driven approaches often struggle to capture long-range degradation dependencies, neglect non-stationary aging behavior, or provide only deterministic predictions without uncertainty information. To address these limitations, this paper proposes an uncertainty-aware degradation prediction framework that integrates a Patch-Based Transformer (PatchTST) with a degradation-aware weighted quantile random forest. PatchTST is employed to learn robust temporal degradation representations from multivariate operational data, while the proposed degradation-aware weighting strategy emphasizes late-life aging behavior and quantile regression enables probabilistic prediction with calibrated uncertainty intervals. Experimental evaluation on a publicly available PEMFC dataset demonstrates that the proposed method achieves superior performance, with an RMSE of 0.0018, an MAE of 0.0022, and a MAPE of 0.028, outperforming representative deep learning and tree-based baselines. In addition, the proposed framework attains a prediction interval coverage probability of 0.91 with a narrow mean prediction interval width of 0.063, confirming its reliability for uncertainty-aware prognostics. The results indicate that the proposed approach provides an effective and robust solution for PEM fuel cell degradation prediction and health monitoring.