Hierarchical Attention-Enhanced Transformer Framework for Prognostic Degradation Modeling and Capacity Estimation of Lithium-Ion Batteries

Precise prognostics of the Remaining Useful Life (RUL) of Lithium-ion (Li-ion) batteries are imperative for augmenting their reliability, safety, and operational efficacy, particularly in domains such as electric vehicles (EVs), renewable energy storage, and portable electronic devices. Conventional approaches frequently produce suboptimal battery health assessments because they fail to capture complex long-term correlations among essential metrics, such as voltage, current, and temperature. To address these challenges, this paper integrates a Denoising Auto-Encoder (DAE) and a Mixture of Expert (MoE) paradigms to present an improved Transformer-based neural architecture for battery capacity estimation and RUL prediction. The Transformer network effectively predicts short-term and long-term degradation patterns, while the DAE reduces sensor noise to enable improved feature extraction. Furthermore, by dynamically utilizing several specialized sub-models for improved representation learning, the MoE framework improves predictive accuracy. Data augmentation strategies are used to address the issues of data scarcity, guaranteeing a training sample that is more diverse and robust. Compared to standard models, the suggested approach can reduce Relative Error (RE) by up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 0 \%}-\mathbf{2 0 \%}$</tex>, demonstrating its resilience in battery health prognostics.