A hybrid and efficient architecture for enhanced battery capacity degradation prediction with zero-shot generalization

The study presents a novel hybrid architecture integrating Mixture of Experts (MOE) and Efficient Multi-Scale Attention (EMSA) modules to address the challenge of battery capacity prediction amid diverse and complex characteristics. Zero-shot prediction refers to the model's ability to predict battery states or performance for datasets or conditions not seen during training, without additional fine-tuning, leveraging knowledge from other battery datasets. This demonstrates strong cross-domain generalization, crucial for battery management with diverse battery types and operating conditions. The model adapts to varying data distributions via MOE's gating mechanism, while EMSA enhances multi-scale feature extraction. Residual connections and Switchable Normalization (SN) enable hierarchical feature learning and robust performance across heterogeneous datasets. Experimental evaluations on four datasets (HUST, MIT, TJU, XJTU) demonstrate that the proposed model outperforms the Physics-Informed Neural Network (PINN) baseline. Specifically, the 2-layer EMSA variant reduces the overall Mean Absolute Error (MAE) by 8 % compared to PINN, zero-shot variant achieves a 3 % MAE reduction on unseen data. Ablation studies confirm that EMSA modules significantly improve prediction accuracy yielding up to 2.2 % higher R2 on the XJTU dataset, and Pareto optimization enhances multi-loss balance, boosting the Base model's average accuracy by 1.5 %. These results highlight the model's high accuracy and robust zero-shot generalization, providing a promising solution for advanced battery management systems.