Scholarlink

<p>Lithium-ion batteries are extensively employed as critical energy storage devices in mobile electronics, electric vehicles, and renewable energy systems. However, the reliability, performance, and safety of these batteries are highly contingent on precise state of charge (SOC) estimation. To address this challenge, this paper proposes a novel SOC estimation approach based on a hybrid neural network combined with variational mode decomposition (VMD). Initially, the battery data is decomposed using VMD, effectively separating global trends from local fluctuations. This decomposition mitigates the effects of capacity regeneration and noise interference, thereby facilitating the extraction of multi-dimensional features and capturing data patterns across different time scales. Subsequently, a hybrid neural network architecture is constructed to estimate the SOC, integrating a depthwise separable convolutional neural network (DSCNN), a bidirectional long short-term memory network (BILSTM), and an attention mechanism (AT). Specifically, DSCNN reduces computational complexity while efficiently extracting key spatial features from the decomposed data. BILSTM, with its internal memory units, captures sequential dependencies in both forward and backward directions, effectively modeling long-term temporal relationships. The AT adaptively re-weights input features, allowing the framework to focus on the most relevant information for SOC estimation. Finally, the superiority and generalization ability of the proposed method are comprehensively validated on three lithium-ion battery datasets, including A123 18,650, INR18650 and Panasonic 18650PF datasets. Multiple evaluation experiments are carried out, covering conventional working condition tests, cross-dataset domain adaptation, and cross-temperature robustness verification. Experimental results demonstrate that the proposed method achieves remarkable estimation accuracy and stable generalization performance under various scenarios. At 25&#xa0;°C, the MAE and RMSE are limited within 0.0067 and 0.0085, respectively, and the R<sup>2</sup> value remains higher than 99.9%. These results fully verify that the proposed method can significantly improve the SOC estimation accuracy, stability and robustness of lithium-ion batteries, and possesses good application potential in complex temperature and cross-dataset practical scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid neural network combined with variational mode decomposition for estimating state of charge of lithium-ion batteries

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Di Zheng,
  • Longji Qin,
  • Wenjun Deng,
  • Xifeng Guo,
  • Yi Ning,
  • Tianke Ma

摘要

Lithium-ion batteries are extensively employed as critical energy storage devices in mobile electronics, electric vehicles, and renewable energy systems. However, the reliability, performance, and safety of these batteries are highly contingent on precise state of charge (SOC) estimation. To address this challenge, this paper proposes a novel SOC estimation approach based on a hybrid neural network combined with variational mode decomposition (VMD). Initially, the battery data is decomposed using VMD, effectively separating global trends from local fluctuations. This decomposition mitigates the effects of capacity regeneration and noise interference, thereby facilitating the extraction of multi-dimensional features and capturing data patterns across different time scales. Subsequently, a hybrid neural network architecture is constructed to estimate the SOC, integrating a depthwise separable convolutional neural network (DSCNN), a bidirectional long short-term memory network (BILSTM), and an attention mechanism (AT). Specifically, DSCNN reduces computational complexity while efficiently extracting key spatial features from the decomposed data. BILSTM, with its internal memory units, captures sequential dependencies in both forward and backward directions, effectively modeling long-term temporal relationships. The AT adaptively re-weights input features, allowing the framework to focus on the most relevant information for SOC estimation. Finally, the superiority and generalization ability of the proposed method are comprehensively validated on three lithium-ion battery datasets, including A123 18,650, INR18650 and Panasonic 18650PF datasets. Multiple evaluation experiments are carried out, covering conventional working condition tests, cross-dataset domain adaptation, and cross-temperature robustness verification. Experimental results demonstrate that the proposed method achieves remarkable estimation accuracy and stable generalization performance under various scenarios. At 25 °C, the MAE and RMSE are limited within 0.0067 and 0.0085, respectively, and the R2 value remains higher than 99.9%. These results fully verify that the proposed method can significantly improve the SOC estimation accuracy, stability and robustness of lithium-ion batteries, and possesses good application potential in complex temperature and cross-dataset practical scenarios.