Physical-Context Guided Mixture of Experts with LSTM for Accurate and Robust State-of-Health Estimation of Lithium-Ion Batteries in Electric Vehicles

Accurate state-of-health estimation is critical for the safe and efficient operation of electric vehicle batteries. To address the limited cross-segment generalization capability and vulnerability to data noise in state-of-health estimation for electric vehicle batteries, this paper proposes a physicalcontext guided mixture-of-experts with long short-term memory (MoE-LSTM). The method integrates multi-dimensional charging features with a dynamic expert routing mechanism to improve the accuracy and robustness of capacity estimation. Based on large-scale real-world vehicle charging data, a three-tier feature system is constructed, encompassing shallow electrochemical response, medium-level statistical distributions, and deep usage patterns. The model employs multiple long short-term memory networks as expert units and designs a physically driven context router to dynamically allocate expert weights. A gated residual bias correction network is further embedded to suppress systematic deviations. In three-fold vehicle-level cross-validation, MoELSTM consistently outperforms all baseline models across all evaluation metrics, achieving a mean absolute percentage error of 1.13%, a mean absolute error of 1.22 Ah, a root mean square error of 1.50 Ah, and a coefficient of determination of 0.38 for real vehicle batteries with nominal capacity of 120 Ah. Ablation studies confirm that the synergistic combination of the multiexpert architecture, physical context router, and bias correction module yields the best overall performance. Robustness tests under additive noise further demonstrate that MoE-LSTM maintains the most stable performance across varying noise levels. The framework requires no additional hardware sensors and provides efficient and reliable technical support for cloud-based intelligent battery health diagnosis and predictive maintenance.