Remaining useful life prediction method of PEMFC based on joint feature extraction of LSTM and improved Transformer

结合LSTM和改进Transformer联合特征提取的PEMFC剩余使用寿命预测方法

¹ Automation and Electrical Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
² College of Mathematics and Physics, Lanzhou Jiaotong University, Lanzhou 730070, China
³ Bailie Mechanical Engineering, Lanzhou City University, Lanzhou 730070, China

Received: 27 March 2025

Abstract

Proton exchange membrane fuel cells (PEMFC) are a crucial component of modern sustainable clean energy generation technology. Accurate prediction of performance degradation is key to enhancing the performance of PEMFC systems and is also an important step in promoting this clean energy technology for broader applications. Traditional methods for predicting performance degradation typically achieve their aims through mechanistic models and forecasting algorithms, refining model parameters and algorithm structures to improve accuracy. However, these methods often fall short in fully considering the detailed characteristics implied by aging data over long-time scales and the phenomenon of voltage recovery. Therefore, this paper proposes a predictive model that combines long short-term memory (LSTM) networks with an enhanced Transformer for joint feature extraction to achieve precise predictions of PEMFC output voltage. Initially, based on traditional Transformer architecture, an optimized design is performed to build an improved Transformer model suitable for PEMFC remaining useful life (RUL) prediction. Secondly, the improved Transformer model is embedded into the conventional LSTM framework, constructing a combined LSTM and improved Transformer joint feature extraction model. Under steady-state, dynamic and pseudo-dynamic datasets, finally, the LSTM, convolutional neural networks (CNN), CNN-LSTM, the improved Transformer, and the joint feature extraction model were evaluated for output voltage prediction and quantitatively compared. The results indicate that the improved Transformer and the joint prediction model show significant improvements over other comparative models in evaluation metrics such as RMSE, MAE, and R². This confirms that the proposed prediction model can enhance the RUL prediction accuracy of PEMFC to some extent.

摘要

质子交换膜燃料电池(proton exchange membrane fuel cell, PEMFC)性能退化的精确预测是提升PEMFC系统性能的关键。传统性能退化预测方法一般通过机理模型和预测算法实现, 并通过优化模型参数和算法结构以提高预测精度, 但是对老化数据长时间尺度所隐含的细节特征考虑不够充分。因此提出一种长短期记忆网络(long short-term memory, LSTM)结合改进Transformer联合特征提取模型以实现PEMFC输出电压的精确预测。基于传统Transformer结构进行优化设计, 构建适用于PEMFC RUL预测的改进型Transformer模型。在传统LSTM的基础上嵌入改进Transformer模型, 构造LSTM结合改进Transformer联合特征提取模型。在稳态、动态和伪动态工况下分别对LSTM、卷积神经网络(convolutional neural networks, CNN)、CNN-LSTM、改进型Transformer和联合特征提取模型进行输出电压预测及量化对比。结果表明, 改进Transformer和联合预测模型在均方根误差E_RMS、平均绝对误差E_MA和R²等评估指标上相较于其他对比模型均有显著提升, 验证了所提预测模型能够在一定程度上提高PEMFC的RUL预测精度。