Scholarlink

<p>Accurate prediction of the state of health (SOH) of lithium-ion batteries is essential for safe operation and predictive maintenance in battery management systems. However, existing data-driven methods still suffer from cumulative error in long-horizon autoregressive prediction, attention bias toward early positions in long sequences, and insufficient dynamic modeling of heterogeneous physical channels across degradation stages. To address these issues, this paper proposes GST-Net (Gated Sequential Transformer Network), a dual-branch architecture that integrates gated attention with causal large-kernel convolution. The local branch uses a causal large-kernel depthwise separable convolution block to capture cycle-level degradation patterns, while the global branch adopts a bidimensional gated attention block to model long-range temporal dependencies and stage-aware channel reweighting. The two branches are adaptively combined through a gated fusion mechanism. In addition, an instance-aware post-hoc revision head and an autoregressive consistency loss are introduced to reduce cumulative error during long-horizon rollout. Experiments on three public lithium-ion battery degradation datasets, including CALCE, NASA, and TJU, show that GST-Net reduces the average MAE on CS2_35, CS2_37, B0005, and CY25_1 by 36.0%, 25.6%, 30.9%, and 35.7%, respectively. GST-Net achieves the lowest or near-lowest AE among the compared baselines, with mean AE values remaining below one cycle on B0005 and within a few cycles on CS2_37 under the adopted EOL threshold. Ablation studies further confirm the effectiveness of the proposed components.</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

GST-Net: a dual-branch fusion network with gated attention and causal large-kernel convolution for lithium-ion battery state-of-health prediction

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Yuchang Gu,
  • Taiyu Lu,
  • Biao Liu

摘要

Accurate prediction of the state of health (SOH) of lithium-ion batteries is essential for safe operation and predictive maintenance in battery management systems. However, existing data-driven methods still suffer from cumulative error in long-horizon autoregressive prediction, attention bias toward early positions in long sequences, and insufficient dynamic modeling of heterogeneous physical channels across degradation stages. To address these issues, this paper proposes GST-Net (Gated Sequential Transformer Network), a dual-branch architecture that integrates gated attention with causal large-kernel convolution. The local branch uses a causal large-kernel depthwise separable convolution block to capture cycle-level degradation patterns, while the global branch adopts a bidimensional gated attention block to model long-range temporal dependencies and stage-aware channel reweighting. The two branches are adaptively combined through a gated fusion mechanism. In addition, an instance-aware post-hoc revision head and an autoregressive consistency loss are introduced to reduce cumulative error during long-horizon rollout. Experiments on three public lithium-ion battery degradation datasets, including CALCE, NASA, and TJU, show that GST-Net reduces the average MAE on CS2_35, CS2_37, B0005, and CY25_1 by 36.0%, 25.6%, 30.9%, and 35.7%, respectively. GST-Net achieves the lowest or near-lowest AE among the compared baselines, with mean AE values remaining below one cycle on B0005 and within a few cycles on CS2_37 under the adopted EOL threshold. Ablation studies further confirm the effectiveness of the proposed components.