State-of-Charge Estimation of Lithium-ion Batteries Using LSTM Deep Learning Method

The effects of ambient temperature and the flat form characteristics of the open circuit voltage state-of-charge (SOC) curve for lithium iron phosphate batteries are the major issues that influence the accuracy of the SOC estimation, which is critical for estimating the driving range of electric vehicles, and the optimal charge control of batteries to prevent the sudden loss of power in battery-powered systems. We proposed a SOC estimation method by using a long short-term memory (LSTM)-recurrent neural network (RNN) to reduce the SOC estimation errors, and to develop a model for the sophisticated battery behaviors under varying ambient temperatures, including time-variable current, voltage, and temperature conditions. The proposed method was evaluated using data from the LiFePO4 battery obtained by the dynamic stress test. The experimental results show that the proposed method can accurately learn the influence of ambient temperatures on the battery and also estimate the battery's SOC under varying temperatures with root mean square errors less than 1.5% and mean average errors less than 1%. Moreover, the proposed method also provides a sufficient SOC estimation under other temperature conditions. The main contribution of this study is the comprehensive explanation and implementation process of the data-based DL approach for the SOC estimation of the LIBs in the following aspects, (1) An LSTM-RNN was trained to model the complex battery dynamics under varying ambient temperatures. (2) The proposed method is model-free and data-driven approach, which means there is no need to construct OCV-SOC lookup tables under varying temperatures in order to pick an appropriate equivalent circuit model. The proposed method can be extended for the SOC estimation of other types of lithium batteries.