Scholarlink

<p>The accurate estimation of the state of charge (SOC) of lithium-ion batteries is a core technological challenge in battery management systems. However, data-driven SOC estimation methods have insufficient feature representation capabilities, and the attention mechanism in traditional Transformer models has limitations in capturing local dynamic features, making it difficult to differentiate the importance of features and to overly concentrate weight distribution in peak areas. To address these issues, this paper proposes an enhanced feature extraction method for lithium-ion battery SOC estimation. This method uses Singular Spectrum Analysis (SSA) to perform trend decomposition on input features and constructs a six-dimensional time-series feature matrix by combining the original input features to enhance the input representation ability. Additionally, an improved Transformer model, CCHformer (Causal Convolution-Channel Attention-Hilly Attention Transformer), is introduced, which incorporates a DCC Block (Dilated Causal Convolution with Channel Attention) in its architecture—this block extends the receptive field via dilated causal convolution and integrates a channel attention mechanism to enhance the model’s capability of capturing multi-scale local features and key information—while replacing the traditional self-attention mechanism with a Hilly Attention mechanism that optimizes weight distribution through a power-law transformation. Comparative experiments on a publicly available battery dataset show that the root mean square error (RMSE) of SOC estimation using this method remains stable below 0.85<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation> at under three temperature conditions and two testing scenarios, with the coefficient of determination (R<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(^{2}\)</EquationSource> </InlineEquation>) consistently above 99.88<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation>.</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 lithium-ion battery SOC estimation method integrating singular spectrum analysis and an improved transformer architecture

  • 期刊论文

  • 发表日期：

    2025年11月28日

查看全文
  • Houwen Shen,
  • Zhanying Li,
  • Hao Xu,
  • Wenhao Fu,
  • Mingyu Wang

摘要

The accurate estimation of the state of charge (SOC) of lithium-ion batteries is a core technological challenge in battery management systems. However, data-driven SOC estimation methods have insufficient feature representation capabilities, and the attention mechanism in traditional Transformer models has limitations in capturing local dynamic features, making it difficult to differentiate the importance of features and to overly concentrate weight distribution in peak areas. To address these issues, this paper proposes an enhanced feature extraction method for lithium-ion battery SOC estimation. This method uses Singular Spectrum Analysis (SSA) to perform trend decomposition on input features and constructs a six-dimensional time-series feature matrix by combining the original input features to enhance the input representation ability. Additionally, an improved Transformer model, CCHformer (Causal Convolution-Channel Attention-Hilly Attention Transformer), is introduced, which incorporates a DCC Block (Dilated Causal Convolution with Channel Attention) in its architecture—this block extends the receptive field via dilated causal convolution and integrates a channel attention mechanism to enhance the model’s capability of capturing multi-scale local features and key information—while replacing the traditional self-attention mechanism with a Hilly Attention mechanism that optimizes weight distribution through a power-law transformation. Comparative experiments on a publicly available battery dataset show that the root mean square error (RMSE) of SOC estimation using this method remains stable below 0.85 \(\%\) at under three temperature conditions and two testing scenarios, with the coefficient of determination (R \(^{2}\) ) consistently above 99.88 \(\%\) .