贝叶斯正则化优化BP神经网络估算SOH  

Estimation of SOH based on Bayesian regularization optimized BP neural network

作　　者：朱聪聪 郭晟 常海涛[2] 路密 ZHU Congcong;GUO Sheng;CHANG Haitao;LU Mi(School of Materials Science and Technology,Xiamen University of Technology,Xiamen 361024,Fujian,China;Key Lab for High Efficiency Battery of Fujian Province,Fujian Nanping Nanfu Battery Co.,Ltd.,Nanping 353000,Fujian,China)

机构地区：[1]厦门理工学院材料科学与工程学院,福建厦门361024 [2]福建南平南孚电池有限公司,福建省高效能电池重点实验室,福建南平353000

出　　处：《电池》2025年第1期25-31,共7页Battery Bimonthly

基　　金：国家自然科学基金(21975212);建省高校产学研联合创新项目(2022H6010)。

摘　　要：为提高锂离子电池健康状态(SOH)估算的精度,采用基于贝叶斯正则化算法优化的反向传播(BP)神经网络模型。该模型的核心是,引入先验分布约束BP网络权重参数,以减少过拟合风险;并引入后验分布评估参数的不确定性,提升模型对数据噪声的适应性。以充电全过程提取健康特征验证模型精度;以放电片段数据提取健康特征模拟实际工况。训练后的模型在充电全过程提取特征时的均方根误差(RMSE)和平均绝对误差(MAE)均小于1.65%,采用放电片段提取特征时的RMSE和MAE均小于3.85%,相较于未优化的BP神经网络,两种方式的估算误差分别降低18%和41%以上。In order to improve the accuracy of state of health(SOH)estimation of Li-ion battery,a Bayesian regularized algorithm optimized back propagation(BP)neural network model is used.The core of the model is to introduce a priori distribution to constrain the weight parameters of BP network to reduce the risk of over-fitting.The uncertainty of the posterior distribution evaluation parameters is introduced to improve the adaptability of the model to data noise.The accuracy of the model is verified by extracting health features from the entire charging process.The health features are extracted from discharge fragment data to simulate actual working conditions.The root mean square error(RMSE)and mean absolute error(MAE)of the trained model are less than 1.65%when extracting features during the entire charging process,the RMSE and MAE are less than 3.85%when extracting features with discharge segments.The errors of two methods are reduced by more than 18%and 41%,respectively,compared with that of the unoptimized BP neural network.

关 键 词：锂离子电池 健康状态(SOH) 贝叶斯正则化算法 反向传播(BP)神经网络 健康特征 先验分布 后验分布 

分 类 号：TM912.9[电气工程—电力电子与电力传动]