Abstract:Accurate estimation of lithium-ion battery state of health is crucial for the reliable operation of new energy vehicles. To address the dynamic changes in the importance of multidimensional features during battery degradation and the limitations of traditional models in capturing both overall trends and local fluctuations, this paper proposes a state-of-health estimation method based on an integrated framework combining a dual dynamic attention mechanism network with LightGBM. A convolutional neural network combined with a bidirectional long short-term memory network is employed to extract degradation features from battery charge-discharge processes. Feature attention and temporal attention mechanisms are introduced to adaptively weight degradation information. Furthermore, a stacking heterogeneous ensemble model integrating CNN-BiLSTM-DAM and LightGBM is constructed to achieve collaborative complementarity among different models. The results show an average absolute error of 0.0040 on the NASA dataset and an average absolute error of 0.0137 on the CALCE dataset. The proposed method offers high prediction accuracy and can provide an effective reference for lithium-ion battery health state assessment.
