基于超参数优化和混合神经网络的锂硫电池健康状态估计

doi:10.19799/j.cnki.2095-4239.2025.0608

摘要/Abstract

摘要：

锂硫（Li-S）电池以其高比容量、高能量密度、环保友好等优点得到了广泛的应用，其中健康状态（State of health，SOH）是设计安全可靠的电池管理系统（Battery management systems，BMS）的关键参数。然而，由于Li-S电池内部电化学性质的复杂性和外部工作环境的不确定性，实现准确的SOH估算是一个很大的挑战。在本研究中，我们提出了一种新的基于超参数优化和CNN-Bi-LSTM混合神经网络的SOH估计框架，该框架无需复杂的人工预处理步骤即可从原始传感器数据中提取空间特征和通道特征，然后使用双向长短期记忆神经网络与自注意力机制进行数据拟合。本文在实验室自主搭建的Li-S电池测试平台上对四块不同容量、不同结构的电池进行了实验，当采用不同比例的训练数据集，SOH估计结果的最大RMSE不超过0.57%，MAE不超过0.39%，证明所提出方法的精度较高，鲁棒性较强。

关键词: 锂硫电池, 健康状态估计, 混合神经网络, 超参数优化

Abstract:

Lithium-sulfur (Li-S) batteries have been widely used due to their high specific capacity, high energy density, and environmental friendliness, among which state of health (SOH) is the key parameter for designing safe and reliable battery management systems (BMS). However, due to the complexity of the internal electrochemical properties of Li-S cells and the uncertainty of the external operating environment, achieving accurate SOH estimation is a big challenge. In this study, we proposea novel SOH estimation framework based on hyperparameter optimization and CNN-Bi-LSTM hybrid neural networks, which can extract spatial features and channel features from raw sensor data without complex manual preprocessing steps, and then use bidirectional long short-term memory neural network with self-attention mechanism for data fitting. In this paper, experiments are carried out on four batteries with different capacities and different structures on the Li-S battery test platform built by the laboratory, and when different proportions of training datasets are used, the maximum RMSE of the SOH estimation results is not more than 0.57%, and the MAE is not more than 0.39%, which proves that the proposed method has high accuracy and robustness.

Key words: Lithium-sulfur battery, state of health estimation, hybrid neural networks, hyperparameter optimization

中图分类号:

TK 02

时洪雷, 刘喜军, 高竟译, 苏志伟, 宫羽, 王凯. 基于超参数优化和混合神经网络的锂硫电池健康状态估计[J]. 储能科学与技术, doi: 10.19799/j.cnki.2095-4239.2025.0608.

HongLei SHI, Xijun LIU, JingYi GAO, ZhiWei SU, Yu GONG, Kai WANG. State of Health Estimation of Lithium-Sulfur Batteries Based on Hyperparameter Optimization and Hybrid Neural Networks[J]. Energy Storage Science and Technology, doi: 10.19799/j.cnki.2095-4239.2025.0608.

图/表 15

表1

图1

表2

图2

图3

图4

表3

图5

图6

表4

表5

表6

表7

图7

图8

参考文献 20

[1]	王萍, 弓清瑞, 张吉昂, 等. 一种基于数据驱动与经验模型组合的锂电池在线健康状态预测方法[J]. 电工技术学报, 2021, 36(24):5201-5212. DOI:10.19595/j.cnki.1000-6753.tces.210385.
[2]	闫正义, 赵康, 王凯. 基于强化学习的新型电力系统优化策略应用综述[J]. 发电技术, 2025, 46(03):508-520.
[3]	逯云杰. 电力汽车储能系统控制技术研究[J]. 储能科学与技术, 2024, 13(02):608-610. DOI:10.19799/j.cnki.2095-4239.2024.0013.
[4]	Bi J, Lee J-C, Liu H. Performance Comparison of Long Short-Term Memory and a Temporal Convolutional Network for State of Health Estimation of a Lithium-Ion Battery using Its Charging Characteristics[J]. Energies. 2022, 15(7): 2448.
[5]	尹杰, 刘博, 孙国兵, 等. 基于迁移学习和DAE-LSTM的锂离子电池剩余寿命预测[J/OL]. 电工技术学报, 1-14[2025-08-17]. https://doi.org/10.19595/j.cnki.1000-6753.tces.221890.
[6]	赵康, 闫正义, 王凯. 光纤传感技术在电池荷电状态/健康状态监测中的研究进展[J/OL]. 发电技术, 1-12[2025-08-17].https://link.cnki.net/urlid/33.1405.TK.20250311.0955.002.
[7]	Zhang YX, Cao YD, Fan LL, et al. Structure-defined viologen-polyoxometalate modified separator dominating endurable Li-S batteries by a synergistic adsorption-electrocatalysis mechanism[J]. CHEMICAL ENGINEERING JOURNAL. 2024, 482.
[8]	Jiao C, Zhao CR, Zhang L, et al. Electrochemical properties of high-loading sulfur-carbon materials prepared by in situ generation method[J]. RARE METALS. 2023, 42(11): 3877-3885.
[9]	张效伟, 衣振晓, 王凯. 基于改进自适应蜜獾优化算法优化时间卷积网络的车载锂离子电池健康状态估计[J/OL]. 发电技术, 1-13[2025-08-17]. https://link.cnki.net/urlid/33.1405.TK.20240719.1509.002.
[10]	史宏思, 孙新伟, 王凯. 基于电化学阻抗谱的锂离子电池健康状态估计[J/OL]. 发电技术, 1-15[2025-08-17]. https://link.cnki.net/urlid/33.1405.tk.20240716.1502.008.
[11]	Soni R, Hu J, Robinson JB, et al. Predicting cell failure and performance decline in lithium-sulfur batteries using distribution of relaxation times analysis [J]. CELL REPORTS PHYSICAL SCIENCE. 2024, 5(2).
[12]	Hua X, Zhang T, Offer GJ, et al. Towards online tracking of the shuttle effect in lithium sulfur batteries using differential thermal voltammetry [J]. JOURNAL OF ENERGY STORAGE. 2019, 21: 765-772.
[13]	李鹏举, 陈晓宇, 谢佳, 等. 锂离子电池功率状态预测方法研究进展[J/OL]. 储能科学与技术, 1-8[2025-08-17].https://doi.org/10.19799/j.cnki.2095-4239.2025.0549.
[14]	李英顺, 阚宏达, 郭占男, 等. 基于数据预处理和VMD-LSTM-GPR的锂离子电池剩余寿命预测[J]. 电工技术学报, 2024, 39(10):3244-3258. DOI:10.19595/j.cnki.1000-6753.tces.230210.
[15]	姜寒, 王凯. 面向变电站无人化巡检的多源融合即时定位与建图方法综述[J]. 广东电力, 2025, 38(03): 55-68.
[16]	Cui ZH, Kang L, Li LW, et al. A combined state-of-charge estimation method for lithium-ion battery using an improved BGRU network and UKF[J]. Energy. 2022, 259: 124933.
[17]	Gao J, Yang D, Wang S, et al. State of health estimation of lithium-ion batteries based on Mixers-bidirectional temporal convolutional neural network[J]. Journal of Energy Storage. 2023, 73: 109248.
[18]	尚玉朝, 刘春豪, 王凯. 基于融合优化算法的超级电容器健康状态预测模型[J/OL]. 发电技术, 1-11[2025-08-17].https://link.cnki.net/urlid/33.1405.TK.20250117.1141.002.
[19]	陈媛, 章思源, 蔡宇晶, 等. 融合多项式特征扩展与CNN-Transformer模型的锂电池SOH估计[J]. 储能科学与技术, 2024, 13(09):2995-3005. DOI:10.19799/j.cnki.2095-4239.2024.0465.
[20]	张朝龙, 陈阳, 刘梦玲, 等. 一种基于ICA-T特征和CNN-LA-BiLSTM的锂离子电池健康状态估计方法[J]. 储能科学与技术, 2025, 14(03):1258-1269. DOI:10.19799/j.cnki.2095-4239.2024.1124.

电池	充电截至电压	放电截至电压	充电速率	放电速率	温度	循环次数	寿命终止条件
B1 (1.1 A.h)	2.6V	1.7V	1C	1C	25℃	225	SOH≤80%
B2 (1.1 A.h，CNF)	2.6V	1.7V	1C	1C	25℃	241	SOH≤80%
B3 (1.3 A.h)	2.6V	1.7V	1C	1C	25℃	208	SOH≤80%
B4 (1.3 A.h，CNF)	2.6V	1.7V	1C	1C	25℃	232	SOH≤80%

算法	平均收敛代数	最佳适应度	适应度方差
改进FPA	62	0.0215	0.0012
PSO	100	0.0287	0.0038
GA	107	0.0304	0.0045

结构	参数
DSC	Window-width=2 Window-number =4 Strides=1 Pdding=1 Window-width=1 Window-number =4 Strides=1 Polling=2
Conv2D
Concatenate layer	Input neurons-number = 240, Dropout = 0.2
Bi-LSTM	Neurons-number=240, Dropout=0.1, hidden_size=128
Attention	num=10,nhead=8

电池	MAE (%)	RMSE (%)
B1	0.42	0.67
B2	0.19	0.29
B3	0.57	0.72
B4	0.33	0.45

电池	模型	MAE (%)	RMSE (%)
B1	CNN-Improved-Bi-LSTM	0.19	0.27
	CNN-LSTM	0.40	0.66
	CNN-GRU	0.39	0.61
	CNN	0.55	0.99
B2	CNN-Improved-Bi-LSTM	0.13	0.21
	CNN-LSTM	0.37	0.58
	CNN-GRU	0.39	0.69
	CNN	0.57	0.83
B3	CNN-Improved-Bi-LSTM	0.23	0.33
	CNN-LSTM	0.27	0.45
	CNN-GRU	0.32	0.54
	CNN	0.59	0.79
B4	CNN-Improved-Bi-LSTM	0.21	0.37
	CNN-LSTM	0.42	0.69
	CNN-GRU	0.39	0.54
	CNN	0.52	0.81