State of health estimation for lithium-ion batteries: A residual-optimized and attention-enhanced bidirectional gated recurrent unit autoencoder approach

The state of health (SOH) of lithium-ion batteries must be estimated with high accuracy to guarantee safe use. Nonetheless, the wide diversity of battery types and operating conditions poses substantial challenges. To address this issue, this paper proposes a residual bidirectional gated recurrent unit-attention autoencoder (ResBiGRU-AttnAE) model. The approach leverages the CC–CV charging protocol as a basis. First, degradation descriptors such as the duration of the constant-current phase, the length of the constant-voltage phase, and the peak identified in the incremental capacity curve are extracted from charge data. Then, a BiGRU encoder–decoder network is used to represent extended temporal relationships throughout the aging process. Residual connections are introduced to enhance deep feature propagation and alleviate the vanishing gradient problem, while an attention mechanism adaptively assigns weights to different time steps to focus on critical stages of degradation. Experiments on the public TJU and XJTU datasets demonstrate that the proposed model performs excellently under diverse operating conditions: the MAPE values on the TJU/XJTU datasets are 0.30 %/0.34 %, representing an average reduction of 73.04 %/75.64 % (TJU) and 76.62 %/72.38 % (XJTU) compared with traditional LSTM/GRU models. Beyond these baselines, additional head-to-head comparisons with BiGRU-AttnAE, BiLSTM-AttnAE, and ResBiLSTM-AttnAE show that ResBiGRU-AttnAE consistently attains the lowest MAE/RMSE/MAPE and exhibits tighter error distributions across both datasets. The obtained results highlight the model's effectiveness in accurate prediction and broad applicability, reinforcing its value in battery health management.