State of health estimation for lithium-ion batteries using attention gated recurrent unit network with partial charging curve reconstruction

Accurate battery state of health (SOH) estimation is crucial for the reliable operation of electric vehicles (EV). The lack of electric vehicle charging voltage data due to user charging behavior undoubtedly poses challenges for estimating SOH. This paper proposes a battery voltage reconstruction method to address the issue of missing battery data in real-world scenarios. Reconstructed data is considered a prerequisite for extracting aging features. Health indicators are derived from complete battery voltage data to characterize the aging state. In addition, feature selection is performed based on the relevance of features to SOH. A self-attention mechanism is introduced into the gated recurrent unit (GRU) network to capture local details of SOH degradation. Experimental validation is conducted on four different battery datasets. The proposed model shows an average decrease of 24.0% and 21.8% in MAE and RMSE, respectively, compared to Temporal Convolutional Network (TCN). In comparison to long short-term memory network (LSTM), the average decrease in MAE and RMSE is 49.4% and 48.5%, respectively. Compared to GRU, the average decrease in MAE and RMSE is 33.5% and 29.0%, respectively. Experiments demonstrate that the method offers high estimation accuracy, strong robustness, and generalizability across different batteries.