A novel extended Kalman filter-guided long short-term memory algorithm for power lithium-ion battery state of charge estimation at multiple temperatures

The power lithium-ion battery's state of charge (SOC) is critical for electric vehicles. However, current neural network-based SOC estimation algorithms fail to consider the effects of feature enhancement in reducing the network training time and improving the output stability. Aiming at the above problems, this paper proposes a SOC estimation algorithm based on the Extended Kalman Filter-Guided Long Short-Term Memory Network (EKF-GLSTM). First, the algorithm utilizes the EKF algorithm to address the problem of few input features and strong nonlinearity caused by measurement limitations, and achieves the purpose of extracting enhanced features from the original battery measurement data. Second, in order to guide the LSTM neural network to better capture the timing dependence and dynamic characteristics of battery behavior, the algorithm inputs the enhanced features extracted by the EKF into the LSTM neural network along with the original measurement data, thereby enhancing the regression prediction capability of the neural network. In addition, the fusion of EKF and LSTM reduces the effects of measurement noise and model uncertainty. The EKF-GLSTM algorithm is validated under multiple temperatures and operating conditions. It can improve the maximum error by more than 17.24 % and 58.09 % compared to the EKF algorithm and LSTM algorithm, respectively.