BATTER-MoE: A Sparse Mixture-of-Experts Model for Accurate and Efficient Remaining Useful Life Prediction of Lithium-Ion Batteries

Accurately estimating the remaining useful life (RUL) of batteries from degradation trajectories is critical for system safety but remains challenging due to the complex and heterogeneous nature of degradation processes. Existing dense predictors achieve strong accuracy but process heterogeneous degradation trajectories through a shared full-capacity pathway, which may limit deployment flexibility in resource-constrained settings due to their computational cost. To address this limitation, we propose BATTER-MoE, a sparse Mixture-of-Experts (MoE) framework that couples battery-oriented temporal representation learning with sparse conditional computation. The framework combines time-aware multi-scale tokenization and cross-scale feature fusion with temporal reweighting, a sparse MoE backbone with RoPE-driven self-attention and a shared expert mechanism, and time-aware pooling for prediction. Experimental results on three public datasets (NASA, GOTION, and TJU), covering different chemistries and capacities, demonstrate that BATTER-MoE achieves competitive performance compared with representative baselines. Across multiple settings, it achieves the best or tied-best MAE/RMSE in most evaluations while maintaining a competitive balance between predictive accuracy and computational cost. This study suggests that sparse conditional computation is a promising direction for battery prognostics when both predictive accuracy and computational efficiency are important.