Scholarlink

<p>Accurate prediction of battery capacity, a key indicator of State of Health (SOH), is of significant engineering value for ensuring system safety, optimizing maintenance strategies, and extending equipment lifespan. However, the battery degradation process exhibits complex characteristics such as strong non-linearity, stochasticity, and capacity regeneration, posing severe challenges to the accuracy of traditional models in predicting the Remaining Useful Life (RUL) of lithium-ion batteries. To address these challenges, this paper proposes an RUL prediction method for lithium-ion batteries based on Variational Mode Decomposition (VMD) and a cascaded BiLSTM-Transformer network, which leverages frequency-time domain collaborative modeling and a dynamic attention-focusing mechanism. Initially, VMD is utilized to adaptively decompose the original battery capacity sequence in the frequency domain, effectively isolating the dominant degradation trend while suppressing non-stationary perturbations and the capacity regeneration phenomenon. Subsequently, a Bidirectional Long Short-Term Memory (BiLSTM) network models the long-term bidirectional dependencies of the decomposed sequences to construct rich temporal representations. Finally, the Transformer’s Multi-Head Self-Attention mechanism is introduced, using the final hidden state as a query vector to achieve precise contextual focusing and weighted modeling of critical degradation stages, thereby enhancing the model’s capability to perceive end-of-life decay features. Test results on the CALCE lithium-ion battery dataset show that the proposed VMD-BiLSTM-Transformer model outperforms mainstream baseline models such as VMD-LSTM and LSTM-Transformer across multiple evaluation metrics, including RMSE, MAE, and R<sup>2</sup>. We conducted out-of-domain validation on an external dataset, strictly reusing the evaluation protocols and hyperparameter settings from the primary domain; the results show that the model still outperforms mainstream models on RMSE, MAE, and R<sup>2</sup> metrics. It demonstrates excellent prediction accuracy, stability, and strong robustness against noise and degradation non-linearity.</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

Remaining life prediction of lithium-ion batteries based on vmd decomposition and cascaded bilstm-transformer network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月7日

查看全文
  • Bin Zou,
  • Ruisong Li,
  • Longyu Ling

摘要

Accurate prediction of battery capacity, a key indicator of State of Health (SOH), is of significant engineering value for ensuring system safety, optimizing maintenance strategies, and extending equipment lifespan. However, the battery degradation process exhibits complex characteristics such as strong non-linearity, stochasticity, and capacity regeneration, posing severe challenges to the accuracy of traditional models in predicting the Remaining Useful Life (RUL) of lithium-ion batteries. To address these challenges, this paper proposes an RUL prediction method for lithium-ion batteries based on Variational Mode Decomposition (VMD) and a cascaded BiLSTM-Transformer network, which leverages frequency-time domain collaborative modeling and a dynamic attention-focusing mechanism. Initially, VMD is utilized to adaptively decompose the original battery capacity sequence in the frequency domain, effectively isolating the dominant degradation trend while suppressing non-stationary perturbations and the capacity regeneration phenomenon. Subsequently, a Bidirectional Long Short-Term Memory (BiLSTM) network models the long-term bidirectional dependencies of the decomposed sequences to construct rich temporal representations. Finally, the Transformer’s Multi-Head Self-Attention mechanism is introduced, using the final hidden state as a query vector to achieve precise contextual focusing and weighted modeling of critical degradation stages, thereby enhancing the model’s capability to perceive end-of-life decay features. Test results on the CALCE lithium-ion battery dataset show that the proposed VMD-BiLSTM-Transformer model outperforms mainstream baseline models such as VMD-LSTM and LSTM-Transformer across multiple evaluation metrics, including RMSE, MAE, and R2. We conducted out-of-domain validation on an external dataset, strictly reusing the evaluation protocols and hyperparameter settings from the primary domain; the results show that the model still outperforms mainstream models on RMSE, MAE, and R2 metrics. It demonstrates excellent prediction accuracy, stability, and strong robustness against noise and degradation non-linearity.