引用本文: | 高锋阳,刘庆寅,赵丽丽,等.改进灰狼优化算法优化CNN-LSTM的PEMFC性能衰退预测[J].电力系统保护与控制,2025,53(13):175-187.[点击复制] |
GAO Fengyang,LIU Qingyin,ZHAO Lili,et al.PEMFC performance degradation prediction based on optimized CNN-LSTM using improved grey wolf optimization algorithm[J].Power System Protection and Control,2025,53(13):175-187[点击复制] |
|
摘要: |
为进一步提高车用质子交换膜燃料电池(proton exchange membrane fuel cell, PEMFC)电堆性能衰退预测与剩余使用寿命预测精度,提出一种改进灰狼优化算法优化卷积神经网络-长短期记忆(convolutional neural network-long short-term memory, CNN-LSTM)的车用PEMFC性能衰退预测方法。首先,通过稳定小波变换对数据集去噪重构,使用改进灰狼算法对实测PEMFC电堆衰退数据进行分析,获得CNN-LSTM最优超参数。其次,利用最优超参数训练CNN-LSTM网络模型进行PEMFC性能衰退预测,并计算PEMFC电堆剩余使用寿命。最后,在电堆静态和动态工况下,将所提方法与传统长短期记忆循环网络、门控循环单元循环网络和未经优化的CNN-LSTM等模型预测进行比较。结果表明:在静态工况中,当训练集占比为60%时,所提方法相比传统CNN-LSTM预测结果均方根误差缩小59.02%,当训练集占比为70%时,PEMFC剩余使用寿命预测与实际相差1.16 h;在动态工况中,当训练集占比为40%时,平均绝对误差缩小18.78%。 |
关键词: 质子交换膜燃料电池 改进灰狼优化算法 卷积神经网络-长短期记忆 衰退预测 剩余使用寿命 |
DOI:10.19783/j.cnki.pspc.240871 |
投稿时间:2024-03-31修订日期:2024-08-25 |
基金项目:中车“十四五”科技重大专项计划项目资助(2021CXZ021) |
|
PEMFC performance degradation prediction based on optimized CNN-LSTM using improved grey wolf optimization algorithm |
GAO Fengyang1,LIU Qingyin1,ZHAO Lili2,QI Fengxu1,LIU Jia1 |
(1. School of Automation and Electrical Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China;
2. CRRC Tangshan Co., Ltd., Tangshan 063035, China) |
Abstract: |
To further improve the prediction accuracy of stack degradation and remaining useful life of proton exchange membrane fuel cell (PEMFC) for vehicles, a novel vehicle PEMFC degradation prediction method is proposed based on convolutional neural network-long short-term memory (CNN-LSTM) optimized by improved grey wolf algorithm (IGWO). First, the data set is denoised and reconstructed using stationary wavelet transform, and IGWO is then used to analyze the measured PEMFC stack degradation data to obtain the optimal hyperparameters for the CNN-LSTM model. Next, the CNN-LSTM network model is trained with the optimal hyperparameters to predict PEMFC performance degradation and calculate the remaining useful life of the PEMFC stack. Finally, under both static and dynamic operating conditions, the proposed method is compared with traditional LSTM, gated recurrent unit network, and unoptimized CNN-LSTM models. Results show that under static conditions with a 60 % training set ratio, the proposed method reduces the root mean square error by 59.02 % compared with the traditional CNN-LSTM. With a 70% training set ratio, the predicted remaining service life differs from the actual value by only 1.16 hours. Under dynamic conditions with a 40% training set ratio, the average absolute error is reduced by 18.78 %. |
Key words: proton exchange membrane fuel cell improved grey wolf optimization algorithm convolutional neural network-long short-term memory degradation prediction remaining useful life |