<p>Accurate prediction of lithium-ion battery remaining useful life (RUL) is of great significance for managing and evaluating the health status of lithium batteries in various fields. However, the degradation noise and capacity regeneration phenomenon during the lithium battery capacity degradation process pose significant challenges to precise RUL prediction. To this end, this study proposes a Multi-Scale Window Attention with Attention Sinks Transformer (MWASFormer) network. By introducing Rotational Position Encoding (RoPE) and a Multi-Scale Window Attention with Attention Sinks (MWAS) mechanism, and based on the PatchTST architecture, this study builds a forecasting framework with the ability to perceive local features and model global degradation. In addition, MWASFormer also integrates Huber Loss and Reversible Instance Normalization (RevIN) techniques to construct a full-link robustness enhancement mechanism from feature preprocessing to loss optimization, thus mitigating the impact of noise interference and distribution shift on prediction stability. Extensive experiments were conducted on three public lithium battery degradation datasets covering different chemical compositions and operating conditions. Simulation results show that MWASFormer exhibits superior prediction performance compared to other advanced deep learning models, with significantly lower fluctuations in prediction errors, and maintains stable high accuracy in early prediction scenarios with sparse historical data and noisy conditions, demonstrating strong generalization ability and robustness. In addition, the ablation studies also demonstrate the necessity of the key components and their synergistic effects. Thus, the proposed network offers a novel solution for RUL prediction of various lithium - ion battery types under complex working conditions.</p>