基于多特征融合的PEMFC剩余使用寿命预测
Remaining Useful Life Prediction of PEMFC Based on Multi-feature Fusion
-
摘要: 质子交换膜燃料电池(Proton Exchange Membrane Fuel Cell, PEMFC)作为直接将化学能转换为电能的新能源，凭借独特的工作原理在新能源汽车和分布式发电等领域有着极其广泛的应用。然而PEMFC在现实工况中受多种因素影响使得其使用寿命有限，因此对剩余使用寿命(Remaining Useful Life, RUL)进行预测是非常必要的。本研究基于深度学习模型对PEMFC的RUL预测，首先，对具有多特征的数据进行皮尔逊相关系数分析，筛选出相关性强的特征信号；然后运用自适应噪声完备集合经验模态分解(Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, CEEMDAN)对数据进行去噪并基于统计特征(均值和方差)来筛选本征模态函数(Intrinsic Mode Functions, IMFs)；紧接着，使用改进的北方苍鹰优化(Improved Northern Goshawk Optimization, INGO)算法优化通道注意力机制(Channel Attention Mechanism, CAM)对数据进行特征融合；最后，利用改进的成吉思汗鲨鱼优化(Improved Genghis Khan Shark Optimizer, IGKSO)算法优化双向长短时记忆网络(Bidirectional Long Short-Term Memory, BiLSTM)对融合数据进行时序预测，以决定系数(R2)为最终预测的评价指标。实验结果表明，该模型预测精度达到99.86%。与传统的长短时记忆网络(Long Short-Term Memory, LSTM)相比，该预测模型在迭代速度、迭代稳定性以及预测精度上都有较显著的提升。Abstract: Proton exchange membrane fuel cells (PEMFCs), which directly convert chemical energy to electricity, are widely applied in new energy vehicles and distributed power generation. But real - world factors limit their lifespan, making remaining useful life (RUL) prediction essential. This research presents a deep - learning - based PEMFC RUL prediction approach. First, Pearson correlation analysis picks relevant features from multi - feature data. Then, CEEMDAN denoises the data, and mean/variance statistical features screen IMFs. Next, the INGO algorithm optimizes the CAM for feature fusion. Finally, the IGKSO algorithm optimizes the BiLSTM network for time - series prediction. With R2 as the metric, the model reaches 99.86% accuracy. Compared to the traditional LSTM network, it shows marked improvements in iteration speed, stability, and prediction accuracy.
下载:
