融合注意力机制的CNN-GRU燃料电池老化趋势预测

Aging trend prediction of fuel cells based on CNN-GRU with attention mechanism

精确预测燃料电池老化趋势不仅能为电池健康管理和剩余寿命估计提供可靠依据,而且在提高电池安全性方面具有重要意义。提出一种融合注意力机制的CNN-GRU燃料电池老化趋势预测模型。首先根据电池老化特性和皮尔逊相关系数,构建由堆电压、电压最大偏差率和电流组成的混合老化指标作为输入。然后利用注意力机制对CNN卷积特征进行进一步权重评估,凸显重要特征,弱化次要特征。同时,探究SE-Block、ECA-Block、CBAM-Block三种注意力机制模块对预测精度的影响。实验结果表明,所提出的混合老化指标作为输入能够获得更贴近实际老化趋势的预测效果。对比GRU基线模型可知,融合注意力机制之后,平均绝对值误差(MAE)和均方根误差(RMSE)分别最少降低30.01%、29.39%,其中CBAM-Block模块性能最佳,MAE和RMSE分别减少72.72%、63.14%。

Accurate prediction of aging trend of fuel cells not only provides a reliable foundation for Prognostics Health Management and estimation of remaining useful life,but also plays an important role in enhancing safety.This paper proposes a CNN-GRU prediction model with attention mechanism.According to aging mechanism and Pearson correlation coefficient,an aging index comprising voltage,maximum voltage deviation rate and current is built as input.Then,the weight of CNN convolutional features is evaluated based on the attention mechanism to highlight important features and weaken minor features.Meanwhile,the impact of three attention mechanism modules(the squeeze and excitation module,efficient channel attention module and convolutional block attention module)on prediction performance is investigated.Our experimental results show the aging index proposed effectively enhances prediction accuracy.Compared with the baseline GRU model,the incorporation of the attention mechanism leads to a marked reduction in MAE and RMSE by at least 30.01%and 29.39%.Notably,the CBAM-Block module performs the best with MAE and RMSE down by 72.72%and 63.14%.