摘要
锂电池健康状态(state of health,SOH)的精确预测评估对电池设备安全稳定运行极为重要,通过对SOH的快速准确预测,可以提高电池设备的安全性并降低出现故障的风险。针对难以精确预测锂离子电池SOH的问题,本文采用电池容量作为SOH的指标,提出一种利用平均欧几里得距离(average euclidean distance,AED)和互补集合经验模态分解(complementary ensemble empirical mode decomposition,CEEMD)方法建立基于Transformer网络结构的锂离子电池健康状态估计算法。首先,我们利用AED评估电池数据库中的电池与待预测电池初期循环容量之间的相似度,并选出具有相似容量退化趋势的电池作为训练集以提高模型的训练速度,然后利用CEEMD方法将电池容量曲线分解为容量再生部分以及退化趋势部分,将各个分量分别使用Transformer网络来建立锂电池退化模型,进而得到锂离子电池的SOH预测结果。本文使用分别来自斯坦福大学与马里兰大学的两个具有不同充放电策略与不同测试环境下的锂离子电池数据集来验证了所提出的电池预测算法的准确性。本文所提模型的均方根误差均能控制在4%以内,具有较好的精确性,并通过与基于LSTM、RNN、GRU的常用锂离子电池健康状态估计算法结果的比较,验证了所提出估计方法的优越性。
The accurate prediction and assessment of the state of health(SOH) of lithium-ion batteries are extremely important for the safe and stable operation of the battery equipment.Quickly and accurately predicting the SOH can enhance the safety of battery devices and reduce the failure risk.This study proposes an algorithm for estimating the health status of lithium-ion batteries based on the transformer network structure to address the challenge of accurately predicting their SOH.This algorithm utilizes the battery capacity as the SOH indicator,incorporating the average Euclidean distance(AED) and complementary ensemble empirical mode decomposition(CEEMD) methods.First,the AED is used to assess the similarity between the initial cycle capacities of the batteries in the battery database and the battery to be predicted.The batteries in the battery database with similar capacity degradation trends are selected as the training set for improving the model's training speed.The CEEMD method is then employed to decompose the battery capacity curve into the capacity regeneration and degradation trend parts.The degradation models for the lithium-ion batteries are separately established using the transformer network for each component.As a result,the predictions for the SOH of lithium-ion batteries are obtained.This study validates the accuracy of the proposed battery prediction algorithm using two lithium-ion battery datasets from Stanford University and the University of Maryland.These datasets comprise batteries tested under different chargedischarge strategies and testing environments.The root mean square error of the proposed model can be controlled within 4%,demonstrating its high accuracy level.The superiority of the proposed estimation method is validated by comparing it with the commonly used lithiumion battery health estimation algorithms based on the long short-term memory,recurrent neural network,and gated recurrent unit.
作者
陈锐
丁凯
祖连兴
许青松
王宗标
罗大思
苏敬江
胡圣
毛冀龙
CHEN Rui;DING Kai;ZU Lianxing;XU Qingsong;WANG Zongbiao;LUO Dasi;SU Jingjiang;HU Sheng;MAO Jiong(CYG SUNRI CO.,LTD.,Zhuhai 519000,Guangdong,China;State Key Laboratory of Advanced Electromagnetic Engineering and Technology(Huazhong University of Science and Technology),Wuhan 430074,Hubei,China)
出处
《储能科学与技术》
CAS
CSCD
北大核心
2023年第10期3242-3253,共12页
Energy Storage Science and Technology