State of Charge Estimation of Lithium-Ion Battery Based on GA-LSTM and Improved IAKF
Jianfeng Wang,Zhiwen Zuo,Yili Wei,Yongkai Jia,Bowei Chen,Yuhan Li,Na Yang
DOI: https://doi.org/10.1016/j.apenergy.2024.123508
IF: 11.2
2024-01-01
Applied Energy
Abstract:Intelligent adaptive extended Kalman filter (IAEKF) is based on the maximum likelihood (ML) method and is widely used to estimate the state of charge (SOC) of EV lithium batteries. The method changes the length of the error innovation sequence (EIS - composed of several error innovation points) in real-time by monitoring whether the error innovation point (EIP), which is the difference between the predicted and observed values in the Kalman filter, at k-1 time changes the distribution of EIS. Subsequently, the classical noise adaptive formula can be used to update the covariance R and Q values at k time. However, in cases where the EIP at k-1 time does not cause changes in the EIS distribution but the EIP at k time results in alterations in the EIS distribution, employing the EIP at k-1 time to assess changes in the EIS distribution and adjusting the EIS length will yield significant errors. Moreover, IAEKF is greatly affected by the accuracy of model parameters and has high requirements for parameter identification and precision of various sensors. To address these limitations, this paper proposed an estimation method based on genetic algorithm and long short term memory (GA-LSTM) and improved Intelligent adaptive Kalman filter (IAKF). The output state of charge (SOC) value of the LSTM neural network is used to replace the observed terminal voltage value in IAEKF, and the initial parameter of the LSTM network is obtained by GA. The improved IAKF employs a symmetric assignment method to determine whether the EIP at k time causes changes in the EIS distribution; furthermore, the noise adaptive formula was modified based on the EIS distribution variation to greatly improve the estimation accuracy and robustness of SOC. Finally, under DST, FUDS and US06 conditions, the maximum errors (ME) were 1.87%, 1.87% and 1.84%, and the root mean squared error (RMSE) was 0.77%, 0.78% and 0.95%, respectively. The results revealed that the proposed method had good accuracy and robustness.