锂离子电池荷电状态(state of charge,SOC)估计技术是电池管理系统(battery management system,BMS)里的关键性技术之一,其精度要求随着锂离子电池应用领域的不断拓宽而越来越高.由此,该文提出一种基于门控循环单元(gated recurrent unit,GRU)编解码器(encoder decoder,ED)的估计方法;在编解码器框架下,首先利用双向GRU网络对可测量变量序列双向捕获依赖关系并将相关信息编码成语境向量,然后使用单向GRU网络完成对语境向量的解码.相比之前提出的循环神经网络,此类端到端模型可以从输入序列中更完整地学习到序列信息以建立出更精确的非线性SOC估计模型.实验数据验证表明,相较于3种双向循环神经网络,该文提出的门控循环单元编解码器(gated recurrent unit encoder decoder,GRU-ED)模型在固定环境温度下取得了最佳的SOC估计效果;同时还在变温环境下实现了小误差的SOC估计,得到的平均绝对误差(mean absolute error,MAE)与最大误差(maximum error,MAX)分别为 0.92％与 4.96％.
Lithium-ion Battery State of Charge Estimation Based on Gated Recurrent Unit Encoder-decoder
The state of charge(SOC)estimation technology matters in battery management systems(BMS),and its accuracy requirements have been increasing as the range of uses for lithium-ion batteries expands.In order to achieve more accurate SOC estimation,this paper proposes an SOC estimation technique based on the gated recurrent unit(GRU)encoder-decoder(ED).With the ED framework,the dependencies of the input sequence are bi-directionally captured by the encoder using a bi-directional GUR network,and the encoder condenses the related information of the input sequence into a context vector,which is subsequently unlocked by the decoder using a unidirectional GRU network.Compared to the previously proposed recurrent neural networks,such end-to-end models can better learn the sequence information from the input sequences to build a more accurate nonlinear SOC estimation model.The simulation experiments demonstrate that the proposed GRU-ED model achieves the best SOC estimation under a fixed temperature compared to 3 kinds of bidirectional recurrent neural networks.Moreover,it accurately estimates the SOC with a low mean absolute error(MAE)and maximum error(MAX)of 0.92％and 4.96％under the changing ambient temperatures.
lithium-ion batterystate of charge estimationgated recurrent unitencoder-decoder
国家科技期刊平台
NSTL
万方数据
维普
