A novel state of charge estimation method for LiFePO4 battery based on combined modeling of physical model and machine learning model

Abstract

In recent years, the machine learning (ML)-based method has become popular for the battery state of charge (SOC) estimation. However, its application in the LiFePO₄ (LFP) battery dynamic discharge conditions suffers from dramatic estimated SOC fluctuations, which is attributed to the limited nonlinear mapping ability. Therefore, this paper proposes a novel SOC estimation method for LFP battery based on combined modeling of physical model and ML model. Our work includes proposing a physical model based on the reverse first-order RC (RFORC) equivalent circuit to fit and shield the explainable dynamics, and using a ML model based on the long short-term memory recurrent neural network (LSTM-RNN) to fit and establish the nonlinear mapping between the residual unexplainable dynamics of battery and the SOC. By leveraging the strengths of the physical model and the ML model, this method reduces the battery dynamics that the ML model needs to fit for indirectly improving its ability to fit the unexplainable dynamics, and consequently reduces the estimated SOC fluctuations and improves the accuracy. The experimental results show that the proposed combined model RFORC-LSTM can reduce the RMSE of LFP battery SOC estimation by at least 51%, and the maximum error (MAXE) by at least 62% compared to the standalone LSTM-RNN, proving that the RFORC-LSTM is highly effective. The method improves the interpretability of the SOC estimation model based on direct mapping, especially the estimation accuracy at low temperatures, and provides a foundation for further improving the performance of SOC estimation based on ML method.