邱健蓄电池基于脉冲电化学阻抗谱变化增量驱动的铅酸蓄电池健康状态评估方法

鉴于在缺乏历史运行数据时评估退役铅酸电池健康状态的难度，本文提出一种基于脉冲前后电化学阻抗谱中频区动态变化的快速SOH评估方法。研究发现，EIS中频区半圆弧的演变趋势与电池劣化程度高度相关。在短时脉冲激励下，健康电池的中频半圆弧显著收缩，而劣化电池则呈现扩张现象，对应于增量Δ的方向性变化电荷转移阻抗增量Δ_ct脉冲后欧姆内阻短时恒流脉冲放电瞬态压降Δ_ct短时恒流脉冲充电瞬态压升Δ以及开路电压恢复₀基于此，本研究设计了一种由脉冲电化学阻抗谱变化驱动的电池健康状态评估模型。实验结果表明，该方法可在无需历史充放电数据的情况下实现快速健康状态（SOH）评估，有效区分健康电池与落后电池，平均SOH误差为6.7%。该方法显著提升了退役铅酸电池的筛选效率与可靠性，为无运行数据储能设备的健康评估提供了一种新颖且具有工程可行性的解决方案。V_neg of short-time constant current pulse discharge, transient voltage recovery ΔV_pos of short-time constant current pulse charging, and open-circuit voltage recovery U_OCV, etc. Based on this, a battery health status assessment model driven by pulsed electrochemical impedance spectrum changes was designed. The experimental results demonstrate that this method enables rapid SOH assessment without requiring historical charge and discharge data, effectively distinguishing between healthy and lagging cells, with an average SOH error of 6.7%. It significantly improves the screening efficiency and reliability of retired lead-acid batteries, providing a novel and engineering-feasible solution for the health assessment of energy storage devices without operational data.

引言

铅酸蓄电池因其成本低廉与技术成熟，在变电站直流电源系统中占据重要地位。其主要在交流电源中断时为继电保护、自动装置、应急照明等关键装备提供直流电源[1][2]。用作后备电源的铅酸蓄电池通常工作于浮充模式，这意味着电池与充电器并联连接，充电器持续提供电流以补偿电池的自放电。蓄电池始终保持满电状态，仅在交流断电期间经历短暂的充放电过程。然而在这种特定运行机制下，铅酸蓄电池仍存在一系列加速单体电池健康状况劣化的问题，导致电池组性能下降甚至提前失效[3]。铅酸蓄电池设计的理论寿命通常为10至15年。在实际应用中，制造差异、环境温度变化以及电池内部电解液浓度与分布不均等因素会导致单体电池间的电化学反应速率不一致。因此，部分单体电池会在达到设计寿命前提前失效[4]。铅酸电池主要呈现三种失效模式：正极板栅腐蚀、负极板硫酸盐化以及电解液分布不均[5]。部分单体电池因过充电导致正极板加速腐蚀，而另一些则因欠充电引发负极板硫酸盐化，从而加速单体电池的性能衰退。除电池自身物理特性及不当充放电操作外，单体电池间的不一致性还会导致充放电过程中电池容量失衡。这会损害过充或过放的电池单体，最终影响整个电池组。
在退役铅酸电池组中，准确识别健康状态恶化的单体电池至关重要。现有的铅酸电池健康评估方法主要依赖表征电池健康状态的特征参数，由此发展出两类判别方法：基于模型的判别方法和数据驱动的判别方法。基于模型的判别方法通过构建电池模型（如等效电路模型、数学模型或电化学模型）来识别模型参数与健康状态（SOH）之间的关联性，从而间接估算电池的剩余健康状态。Guasch等[6]建立了包含电压、电流等参数的经验模型来评估电池健康状况，但由于参数数量庞大且电池个体差异显著，这类模型精度较差。针对参数过多导致的过拟合问题，Eduardo等[7]采用元启发式方法实现了电池荷电状态(SOC)与电压的精确估算。刘等人[8]针对变电站铅酸蓄电池建立了两种不同的工况专用模型：负载运行时的放电电压-健康状态(SOH)关联模型与浮充状态下的内阻-SOH关联模型。当这两个评估模型协同工作时，可将蓄电池健康状态估算误差控制在3%以内。Piłatowicz等人[9]采用电化学阻抗谱与Butler-Volmer方程电学模型评估电池健康状态，将模型参数与老化机制建立关联。K等[10]开发了保留电池主要电化学特性的半经验模型，该模型具有较高复杂度。Reza等[11]通过堆叠式双向长短期记忆网络结合极限梯度提升算法预测容量衰减轨迹，在降低模型复杂度的同时提升了性能。A等[12]提出名为CNN-LSTM-PSO的混合模型，可实现电池剩余寿命的多步超前预测。Y等[13]设计了两阶段联邦迁移学习框架，用于构建能捕捉单体电池特性的个性化模型。
建模方法能够检测、估计和预测电池的内部状态。然而，实际测量数据含有显著噪声，这使得浅层模型难以提取统计特征。数据驱动的分类方法无需对电池内部复杂的电化学过程进行建模，而是直接从大量实验数据中学习健康状态特征，识别电池参数，并提取有效数据特征。这种方法可实现简单、快速且实时的电池退化识别[14]。Haider等学者[15]提出K-shape聚类方法能有效检测数据中心铅酸蓄电池的早期异常，且无需构建电池电化学模型。Murariu等[16]发现等效电路中的恒相位元件参数在电池寿命后半程呈现线性衰减，可作为电池健康状态评估与寿命预测的核心指标。Zhang等[17]提出基于深度学习的在线电池验证与预后方法，通过结合可变输入维度的LSTM网络，利用额外标注样本促进网络训练。Wang等[18]提出新型分段预后模型以精确模拟电池性能衰退。Yang等[19]采用核主成分分析法将输入空间的复杂非线性数据映射至高维特征空间，从而构建能够实现早期故障精准检测与隔离的数据驱动模型。Xue等[20]运用K-means聚类算法、Z-score法与3σ筛选方法实现异常电池单体检测与定位。Shang等[21]将铅酸蓄电池的特征CDF现象与数据驱动方法相结合，包括线性回归、回归树、支持向量机、高斯过程和神经网络。通过利用215个特征，他们实现了高达0.96的电池健康状态预测精度。Talha等人[22]提出了一种基于神经网络的简化在线估计方法，仅需实时电压和电流数据即可实现对电池SOC和SOH的快速、低误差估计，无需获取电池内部参数。₁ parameter of the constant-phase element in the equivalent circuit exhibits linear decay during the latter half of the battery's lifespan, serving as a core metric for battery SOH assessment and lifespan prediction. Zhang et al. [17] proposed a deep learning-based online battery validation and prognosis method, combining LSTM with variable input dimensions to facilitate network training using additional labeled samples. Wang et al. [18] proposed a novel segmented prognostic model for accurately simulating battery degradation. Yang et al. [19] employed nuclear principal component analysis to map complex nonlinear data in the input space onto a high-dimensional feature space, thereby constructing a data-driven model capable of achieving accurate early fault detection and isolation. Xue et al. [20] employed the K-means clustering algorithm, Z-score method, and 3σ screening approach to detect and localize abnormal cells. Shang et al. [21] combined the characteristic CDF phenomenon of lead-acid batteries with data-driven methods, including linear regression, regression trees, support vector machines, Gaussian processes, and neural networks. By utilizing 215 features, they achieved a battery health state prediction accuracy as high as 0.96. Talha et al. [22] proposed a simplified online estimation method based on neural networks that achieves rapid, low-error estimation of battery SOC and SOH using only real-time voltage and current data, without requiring internal battery parameters.
相较于基于模型的分类方法，数据驱动方法显著降低了模型开发所需的专业能力要求。然而，传统数据驱动分类需要大量标注数据。在实际工程应用中，采集足量标注样本具有挑战性，且机器学习训练过程常与支配电池健康衰退的底层物理逻辑相脱节。然而，数据驱动方法可作为补充工具，在实验、计算理论、建模、电化学测试等传统表征方法间形成协同增强效应[23]。Sular等[24]指出，将机器学习与物理模型相融合，可从包含显著噪声的数据中实现电池剩余寿命估计与二次寿命状态评估。Finegan等[25]通过实验优化降低数据需求，整合多维运行数据建立高质量数据库，并从数据驱动预测中提取可解释性洞见，以指导更安全电池及运行条件的设计。这要求同步发展基于物理的模型体系。Mohsin等[26]提出基于电化学阻抗谱的健康状态评估方法，采用以模型驱动为主、数据驱动为辅的研究范式。他们构建了一个基于铅酸电池老化机制加权的健康状态（SOH）预测公式（腐蚀占比20%，硫化及其他因素占比80%）。Naha等学者[27]将随机森林分类器与物理驱动的等效电路模型相结合，用于预测机械滥用电池的内部短路。尽管现有研究尝试融合这两种方法，但仍存在特征选择有限、模型复杂度与实时性能难以平衡，以及宽荷电状态（SOC）范围内适应性不足等挑战。
针对传统铅酸蓄电池健康状态评估模型存在依赖大量标注数据、老化特征提取难度大、评估方法缺乏物理意义等问题，本文提出一种基于脉冲电化学阻抗谱变化驱动的铅酸蓄电池健康状态快速评估模型。通过探究电化学阻抗谱中频区半圆变化与电荷转移阻抗增加的机理关联，构建了具有物理解释性的健康状态敏感特征体系。建立融合特征选择与模型Fusion的统一建模框架，实现了铅酸蓄电池不同健康状态区间的自适应预测。本文主要贡献如下：

• 1)
  It reveals the physical correlation laws between the evolution of the mid-frequency region in the EIS before and after the pulse and the battery's health status from the perspective of the aging mechanism of floating charging. Through conducting short-term pulse experiments and dual electrochemical impedance spectroscopy tests on retired lead-acid batteries, this paper discovered and verified that the change in direction and amplitude of the semi-circle in the mid-frequency region of EIS before and after the pulse has significant discriminatory capabilities for the deterioration state of the battery. This change behavior is mainly dominated by the electrode interface dynamic process, providing a sensitive characteristic with clear physical meaning for the SOH assessment of lead-acid batteries under float charging aging conditions.
• 2)
  A unified nested cross-validation framework is constructed to ensure methodological fairness and eliminate information leakage. Within this framework, mechanism-guided pulse–EIS features are integrated with nonlinear regression models (SVR and XGBoost). Compared with linear baselines, nonlinear learners significantly improve prediction accuracy. Furthermore, a fusion strategy with optimized weight ω^⁎ is introduced and validated through out-of-fold optimization, bootstrap confidence intervals, paired statistical testing, and permutation tests.
• 3)
  For small samples of retired lead-acid batteries lacking historical operation data, a rapid method for classifying and screening the health status of these batteries has been developed. This method only requires one short-time pulse experiment and two EIS measurements to determine the battery's SOH. It offers the advantages of a short testing time, low cost, and eliminates the need for full-life monitoring and unpacking inspection. It can be directly applied to the large-scale screening and secondary utilization scenarios of retired batteries, providing a feasible technical path for engineering circulation.