电工技术学报  2024, Vol. 39 Issue (24): 7933-7949    DOI: 10.19595/j.cnki.1000-6753.tces.231644
电能存储与应用 |
带降噪自编码器和门控递归混合神经网络的电池健康状态估算
陈媛1, 段文献2, 何怡刚3, 黄小贺1
1.安徽大学人工智能学院 合肥 230601;
2.汽车仿真与控制国家重点实验室(吉林大学) 长春 130022;
3.武汉大学电气工程与自动化学院 武汉 430000
State of Health Estimation of Lithium Ion Battery Based on Denoising Autoencoder-Gated Recurrent Unit
Chen Yuan1, Duan Wenxian2, He Yigang3, Huang Xiaohe1
1. College of Artificial Intelligence Anhui University Hefei 230601 China;
2. State Key Laboratory of Automotive Simulation and Control Jilin University Changchun 130022 China;
3. School of Electrical Engineering and Automation Wuhan University Wuhan 430000 China
全文: PDF (3935 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 准确的电池健康状态估计可以保证锂离子电池可靠安全运行,减少系统不必要的维护成本。采用机器学习算法虽然能够获得精确的电池健康状态(SOH),但是其估计精度严重依赖于算法中的参数,普适性差,且会受到传感器噪声的影响。该文提出一种结合降噪自编码器(DAE)和门控递归单元的递归神经网络(GRU-RNN)的混合模型进行电池的SOH估计,以提高算法估计精度及抗干扰能力。首先,利用电压-容量模型来重构电池恒流充电和放电阶段的电压曲线,以减小传感器噪声对SOH估计的影响;其次,从电压曲线和增量容量(IC)曲线中提取相关特征作为SOH估计模型的输入;再次,利用DAE对带有噪声的输入特征进行无监督的训练,可以增强模型的鲁棒性;最后,在输入特征含有噪声的情况下,利用提出的DAE-GRU-RNN算法与其他SOH估计算法进行对比验证。结果表明,该文提出的算法精度更高,相对误差比GRU-RNN和深度神经网络(DNN)模型小6.39%~23.23%。利用部分电压曲线获得的特征数据进行电池SOH预测时,该算法依然具有较高的电池SOH估计精度。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
陈媛
段文献
何怡刚
黄小贺
关键词 电池健康状态估计降噪自编码器门控递归单元的递归神经网络无监督训练    
Abstract:The state-of-health (SOH) estimation of lithium-ion batteries is a key technology in battery management systems, which can ensure the safe operation of the battery system. In practical applications, sensors for voltage and current collection are susceptible to external environmental interference, and the collected data often contains a large amount of noise. The estimation accuracy of the battery SOH model largely depends on the parameters in the algorithm, and some methods are usually not universal. In addition, the charging and discharging conditions of the battery are not complete, so the extracted feature information can easily deviate from the degradation trajectory. Such issues can lead to inaccurate estimation of SOH. This paper proposes a new framework based on the denoising autoencoder (DAE) and gated recurrent unit recurrent neural network (GRU-RNN) with the gated recurrent unit.
Firstly, the voltage-capacity (VC) model is used to reconstruct the voltage curve during the constant current charging stage, and the least squares method is used for model parameter identification. The incremental capacity (IC) curve is obtained from the reconstructed voltage curve. Subsequently, relevant features are extracted from the reconstructed voltage and IC curves as inputs of the SOH estimation model. The reconstructed voltage curve can easily obtain the IC curve of the battery and identify its feature information. Even when the noise ratio is 0.03, the reconstructed voltage curve is consistent with the real voltage curve, which shows that this method has strong noise anti-jamming ability.
Then, the reconstructed voltage curve is divided into different voltage segments, and relevant features are extracted from different voltage segments and IC curves. The Pearson correlation coefficient method is used, and features with high correlation are selected as inputs for the estimation model. Subsequently, a denoising autoencoder is used to train input features with noise in an unsupervised manner and learn more robust and valuable feature information from the damaged data.
Finally, two experiments are performed to verify the effectiveness of the proposed method. When the features contain noise, the proposed algorithm battery has high SOH prediction accuracy, and the relative error is 6.39% to 23.23% lower than that of the GRU-RNN and deep neural network (DNN) models. When the battery SOH is predicted using all the features, the battery’s MAE and RMSE errors of the two models are less than 2%, and the IA is more than 99.4%. When the features are obtained from part of the voltage and IC curve, the prediction error of the battery SOH increases. Two experiments show that the proposed method can learn more robust feature information from noisy data to improve the model’s prediction accuracy and has strong generalization ability. A model can be trained on multiple test sets without adjusting the model’s parameters.
Key wordsState-of-health estimation    denoising autoencoder    gated recurrent unit recurrent neural network    unsupervised training   
收稿日期: 2023-10-02     
PACS: TM912  
基金资助:国家自然科学基金重点项目(51637004)和国家重点研发计划“重大科学仪器设备开发”项目(2016YF0102200)资助
通讯作者: 段文献 男,1988年生,博士,研究方向为电池管理系统状态估计与故障诊断。E-mail:dwx342977542@126.com   
作者简介: 陈 媛 女,1990年生,博士,硕士生导师,研究方向为电池管理系统状态估计与故障诊断。E-mail:cumtjiangsucy@126.com
引用本文:   
陈媛, 段文献, 何怡刚, 黄小贺. 带降噪自编码器和门控递归混合神经网络的电池健康状态估算[J]. 电工技术学报, 2024, 39(24): 7933-7949. Chen Yuan, Duan Wenxian, He Yigang, Huang Xiaohe. State of Health Estimation of Lithium Ion Battery Based on Denoising Autoencoder-Gated Recurrent Unit. Transactions of China Electrotechnical Society, 2024, 39(24): 7933-7949.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.231644          https://dgjsxb.ces-transaction.com/CN/Y2024/V39/I24/7933