Scholarlink

<p>Accurate remaining useful life (RUL) prediction of proton exchange membrane fuel cells (PEMFCs) is important for health monitoring and maintenance planning. To improve degradation prediction under different operating conditions, this study proposes a CNN-BiLSTM-Attention hybrid framework with Bayesian hyperparameter optimization. Correlation analysis and random forest importance evaluation are used for feature selection, while CNN, BiLSTM, and attention modules extract local fluctuation, temporal dependency, and key degradation information, respectively. The proposed model is validated on the IEEE PHM 2014 fuel cell aging dataset under stationary and dynamic operating conditions. Compared with a grid search-optimized LSTM baseline, the proposed framework achieves lower RMSE, MAE, MAPE, and RUL percentage error in most tested cases. For RUL estimation, the percentage error is reduced from 12.93% to 7.82% at 700&#xa0;h under the stationary condition and from 12.74% to 7.40% at 600&#xa0;h under the dynamic condition. These results suggest that the proposed CNN-BiLSTM-Attention framework can provide more accurate degradation trend tracking and RUL estimation for PEMFCs within the tested stationary and ripple-current scenarios, while broader operating profiles should be further examined in future work.</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

Research on fuel cell remaining useful life prediction based on bayesian-optimized CNN-BiLSTM-attention model

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Yuan Gao,
  • Jiangqin Li

摘要

Accurate remaining useful life (RUL) prediction of proton exchange membrane fuel cells (PEMFCs) is important for health monitoring and maintenance planning. To improve degradation prediction under different operating conditions, this study proposes a CNN-BiLSTM-Attention hybrid framework with Bayesian hyperparameter optimization. Correlation analysis and random forest importance evaluation are used for feature selection, while CNN, BiLSTM, and attention modules extract local fluctuation, temporal dependency, and key degradation information, respectively. The proposed model is validated on the IEEE PHM 2014 fuel cell aging dataset under stationary and dynamic operating conditions. Compared with a grid search-optimized LSTM baseline, the proposed framework achieves lower RMSE, MAE, MAPE, and RUL percentage error in most tested cases. For RUL estimation, the percentage error is reduced from 12.93% to 7.82% at 700 h under the stationary condition and from 12.74% to 7.40% at 600 h under the dynamic condition. These results suggest that the proposed CNN-BiLSTM-Attention framework can provide more accurate degradation trend tracking and RUL estimation for PEMFCs within the tested stationary and ripple-current scenarios, while broader operating profiles should be further examined in future work.