文献类型：期刊文献

中文题名：基于物理信息引导的多尺度卷积注意力模型在锂电池荷电状态估计中的应用

英文题名：Physics-informed multi-scale convolutional attention model for state of charge estimation of lithium-ion batteries

作者：彭鑫[1];单劲航[1];练成[2];谭帅[1];钟伟民[1]

机构：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237;[2]华东理工大学化学工程与低碳技术全国重点实验室,上海200237

年份：2025

卷号：55

期号：11

起止页码：1926

中文期刊名：中国科学:技术科学

外文期刊名：Scientia Sinica(Technologica)

收录：;EI(收录号：20254719577059);北大核心:【北大核心2023】；

基金：国家杰出青年科学基金(编号:61925305)项目资助。

语种：中文

中文关键词：锂离子电池;荷电状态;数据驱动;物理信息神经网络;注意力机制

外文关键词：lithium-ion battery;state of charge;data-driven;physics-informed neural network;attention mechanism

摘要：锂离子电池荷电状态的精确估计是电池管理系统的核心任务,但现有数据驱动方法因难以解析多尺度动态特征、忽略放电数据中的局部上下文依赖及缺乏物理规律约束,导致其在复杂工况下的鲁棒性与泛化能力不足.为此,本文提出一种物理信息引导的多尺度卷积注意力模型(physics-informed multi-scale convolutional attention network,PI-MSCAN).首先,设计时域多尺度分解融合架构,通过自适应分解与重构放电数据的瞬态波动与长期趋势,有效捕捉跨时域多尺度特征.在此基础上,构建全局通道注意力机制,量化多源特征的贡献,增强关键通道的表达.其次,针对传统自注意力机制无法有效识别放电曲线中的局部变化趋势这一不足,提出卷积自注意力机制,通过卷积操作提取时间序列的局部特征并将其融入注意力计算,识别放电曲线的局部变化趋势.进一步,通过将电荷守恒方程嵌入损失函数,联合优化数据驱动误差与物理残差约束,确保模型输出遵循电池动态演化规律.实验结果表明,该模型能够有效提升荷电状态估计的精度与可靠性.
Accurate estimation of the state of charge(SOC)for lithium-ion batteries is a core task in battery management systems.However,existing data-driven methods face limitations in robustness and generalization under complex operating conditions due to challenges in resolving multi-scale dynamic features,neglecting local contextual dependencies,and lacking physical constraints.To address these issues,this paper proposes a physics-informed multi-scale convolutional attention network(PI-MSCAN).First,a temporal multi-scale decomposition and fusion architecture is designed to adaptively decompose and reconstruct transient fluctuations and long-term trends in discharge data,effectively capturing characteristics across temporal scales.Building on this,a global channel attention mechanism is developed to quantify contributions from multi-source features and enhance the representation of critical channels.Second,to address the insufficiency of traditional self-attention mechanisms in modeling local contextual information,a convolutional self-attention mechanism is proposed,which incorporates the local perception capability of convolutional kernels into correlation computation to identify local variation trends in discharge curves.Furthermore,by embedding the charge conservation equation into the loss function,the model jointly optimizes data-driven errors and physical residual constraints,ensuring outputs adhere to battery dynamic evolution principles.Experimental results demonstrate that the proposed model significantly improves SOC estimation accuracy and reliability,exhibiting promising application prospects.