State-of-charge estimation of lithium-ion batteries using LSTM and UKF

Abstract

For lithium iron phosphate battery, the ambient temperature and the flat open circuit voltage - state-of-charge (SOC) curve are two of the major issues that influence the accuracy of SOC estimation, which is critical for driving range estimation of electric vehicles and optimal charge control of batteries. To address these problems, this paper proposes a long short-term memory (LSTM) – recurrent neural network to model the sophisticated battery behaviors under varying temperatures and estimate battery SOC from voltage, current, and temperature variables. An unscented Kalman filter (UKF) is incorporated to filter out the noises and further reduce the estimation errors. The proposed method is evaluated using data collected from the dynamic stress test, federal urban driving schedule, and US06 test. Experimental results show that the proposed method can well learn the influence of ambient temperature and estimate battery SOC under varying temperatures from 0°C to 50°C, with root mean square errors less than 1.1% and mean average errors less than 1%. Moreover, the proposed method also provides a satisfying SOC estimation under other temperatures which have no data trained before.