DpMamba: a dual-path state space model for reliable remaining useful life prediction of lithium-ion batteries

Abstract

Lithium-ion batteries are widely used in transportation and energy storage systems, where accurate remaining useful life (RUL) prediction is essential for safety management and maintenance planning. However, battery degradation is often affected by capacity regeneration, measurement noise, and non-monotonic local fluctuations, which makes it difficult for single-scale sequence models to simultaneously capture short-term capacity variations, long-term degradation trends, and end-of-life (EOL) risk. To address these issues, this paper proposes DpMamba, a dual-path selective state-space framework for reliable battery RUL prediction. The proposed model constructs a local path at the original temporal resolution and a trend path at a downsampled scale, and adaptively integrates the two representations through a gated residual fusion mechanism. In addition to capacity regression, DpMamba introduces an explicit hazard head to estimate EOL-related risk at the target cycle, enabling the model to identify risks within the threshold neighborhood before the EOL. A hybrid loss function combining weighted SMAPE loss and risk-alignment loss has been designed to improve capacity prediction and EOL localization capabilities. Experiments on four public battery degradation datasets, including TJU, NASA, OXFORD, and CALCE, demonstrate the effectiveness of the proposed method under different prediction starting points. Compared with the strongest competing baselines, DpMamba reduces MAE by 20.0% to 35.3% on TJU, 58.8% to 62.6% on NASA, 10.3% to 14.0% on OXFORD, and 26.4% to 28.0% on CALCE. In addition, ablation studies and computational cost analysis show that DpMamba can improve the reliability of degradation modeling while maintaining practical inference efficiency.