Lithium-ion batteries state of charge accurate estimation based on fusion deep learning models considering the noises and mechanical properties

Abstract

This paper aims to improve the generalization ability of neural networks (NN) for lithium batteries state of charge (SOC) estimation across different datasets and evaluates the impact of pressure as a new feature. First, the noise of current and voltage is considered to expand the input to the NN and assess its effect on SOC estimation performance. The results indicate that models trained with noise-extended inputs perform better when the sampling accuracy decreases. Then, the pressure characteristics during charging/discharging processes of lithium batteries are studied and introduced as a new feature to explore its impact on the SOC estimation performance. The comparative analysis demonstrates that pressure is an effective feature for SOC estimation, as it varies monotonically with SOC. Finally, several models (including BP, Transformer, XGBoost, etc.) are compared for lithium batteries SOC estimation. After validation with multiple datasets, the results indicate that recurrent neural networks (RNNs) outperform others. Based on this, a hybrid CNN-LSTM model is proposed for high-precision SOC prediction across different lithium batteries datasets. The trained network has been validated with datasets from three different NCM batteries and the SOC estimation results with the maximum root mean square error (RMSE) and mean absolute error (MAE) less than 0.39 % and 0.33 %, respectively. In conclusion, considering noise of current and voltage, along with press properties, have positive influence for networks on SOC estimation, which provides valuable insights for future big data platform development.