Hybrid Neural Networks for Lithium-Ion Batteries State of Health Estimation and Remaining Useful Life Prediction

Abstract Lithium-ion batteries deteriorate due to long-term charging and discharging, leading to performance degradation and safety hazards, so precisely evaluating the state of health (SOH) and remaining useful life (RUL) of lithium-ion batteries is essential for battery safety assurance. To address the issue of poor accuracy of current SOH estimation and RUL prediction methods, this paper proposes a hybrid VMD-BiLSTM-Transformer model. First, a variational modal decomposition is applied to reduce noise and segment the raw data, eliminating capacity regeneration. Second, a combination of bidirectional long short-term memory network (BiLSTM) and Transformer is employed to accurately capture features and fuse global information by utilizing its bidirectional information processing and self-attention mechanisms. Finally, the proposed model is verified on CALCE and Oxford datasets to demonstrate superior prediction accuracy compared to the benchmark models. Specifically, for SOH prediction, the root mean square error is 1.27% for CALCE and 0.23% for Oxford, while the mean absolute error is 0.99% for CALCE and 0.45% for Oxford. Similarly, for RUL prediction, the mean squared error is 0.02% for CALCE and 0.0006% for Oxford. This fully demonstrates that the proposed model exhibits excellent generalization capability and robustness.