Battery Capacity Prediction Method Based on Jamba Model

Accurate prediction of lithium-ion battery health state is a core challenge to ensure the safe operation and maintenance of new energy vehicles, and efficient modeling methods are urgently needed for the risk of capacity degradation and thermal runaway caused by battery aging. Existing prediction methods, such as the Transformer model, are not capable of modeling long sequences, and the complexity of the secondary computation limits the real-time performance. In this paper, we propose a battery performance prediction method based on the Jamba hybrid architecture, which combines the linear complexity of state-space models (SSMs) for long-sequence modeling with the local feature focusing property of the attention mechanism. By alternately stacking the Transformer layer (detecting mutation events) and the Mamba layer (modeling battery capacity degradation), and introducing mixture of experts (MoE) to balance the synergistic modeling of long-term slow-change and short-term mutation. Experiments show that compared with Transformer, Mamba and Mamba-2 models, this paper's method performs well in terms of root-mean-square error, F1-Score, and inference delay in battery capacity prediction, which verifies the advantages of this paper's method in terms of accuracy and efficiency. In addition, the ablation experiments confirm that the SSM module contributes more to the long-term prediction, while the attention mechanism focuses on improving the mutation detection performance. The method is informative for the realization of low latency and high robustness for on-board battery management systems.