锂离子电池的健康状态(state of health,SOH)是电池管理系统的重要功能,对于电池的可靠运行和使用寿命具有重要意义.为了进一步提高数据驱动方法对锂离子电池SOH估计的精度,提出一种卷积Fastformer模型的SOH估计方法.首先,提取锂离子电池多个充电阶段的每次循环电压曲线、电流曲线,每个阶段各个曲线转换为统计健康特征来表征锂离子电池老化特性,并使用Pearson相关系数对所选统计特征进行了相关性分析,筛选出与容量相关性高的健康特征,消除特征冗余性.随后,融合卷积神经网络和具有线性复杂度的Fastformer神经网络的特点,使用卷积神经网络强大的特征提取能力挖掘健康特征的局部信息,利用Fastformer的多头附加注意力机制可以更高效地在复杂的长序列中总结全文信息.然后,为减少模型训练时间,利用正交实验法对模型超参数进行优化.最后,采用公开数据集将所提方法与CNN、GRU、RNN模型进行对比,验证卷积Fastformer模型的准确性,结果表明,平均绝对误差、均方根误差最大仅为0.25％,0.29％,相对误差在0.8％以内,具有较高的估计精度和稳定性.
SOH estimation of lithium-ion batteries using a convolutional Fastformer
The state of health(SOH)of lithium-ion battery batteries is a crucial parameter for battery management systems,playing a substantial role in ensuring reliable operation and extending battery lifespan.This study aims to introduce an estimation method based on convolutional Fastformer model to improve the accuracy of data-driven SOH estimation for lithium-ion batteries.Initially,voltage and current curves from various charging stages of the lithium-ion battery are extracted for each cycle.These curves are then transformed into statistical health features,providing insights into the battery's aging characteristics.The Pearson correlation coefficient is used to analyze the relationship between selected statistical features and capacity,facilitating the identification of highly correlated health features and the elimination of redundant ones.Based on the strengths of convolutional neural networks and the linear complexity of Fastformer neural network,our approach combines the feature extraction capability of the former with local information mining of health features.The Fastformer's multihead attention mechanism efficiently summarizes contextual information within lengthy sequences.Model training time is reduced by optimizing hyperparameters using an orthogonal experimental method.Finally,a publicly available dataset is used for comparative evaluations,pitting our approach against other models such as CNN,GRU,and RNN.The results validate the accuracy of the convolutional Fastformer model,with maximum mean absolute error and root mean square error at only 0.25％and 0.29％,respectively,and a relative error within 0.08％.These findings demonstrate the high accuracy and stability achieved by the proposed method for SOH estimation.
万方数据
维普
