A Capacity Prediction Method for Lithium-Ion Batteries in the Production Stage Based on Multiscale Interactive Attention and Hybrid Driving Methods

Abstract

In the production process of lithium-ion batteries, it is necessary to assemble cells with similar capacities into battery packs. However, traditional capacity measurement methods require a significant amount of time, energy, and cost. Therefore, fast and accurate prediction of the capacity for each lithium-ion battery cell in the production stage is of crucial importance. To address these issues, this article proposes a hybrid-driving method based on multiscale interactive attention (MIA) for lithium-ion battery capacity prediction. This method extracts capacity-related features from three perspectives (temporal, frequency, and thermal imaging). The features are then further extracted and optimized through the backbone network embedded with MIA. Finally, a bidirectional gated recurrent unit network is employed to establish global dependency relationships and obtain the capacity prediction for each cell. The effectiveness of MIA is validated through ablation experiments and testing with real-world production data, while comparative analysis with state-of-the-art models demonstrated the superior performance of the proposed capacity prediction framework in this study.