An LSTM-SA model for SOC estimation of lithium-ion batteries under various temperatures and aging levels

Accurate state-of-charge (SOC) estimation under various ambient temperatures and aging levels remains a challenge for lithium-ion batteries. In this work, a model combining a long short-term memory network and self-attention mechanism (an LSTM-SA model) is derived to enhance traditional LSTM model and improve its process capability. The performance of proposed model is demonstrated using data of two types of lithium-ion batteries collected under various loading conditions, temperatures, and aging levels. Compared with traditional LSTM model, the proposed LSTM-SA model provides more accurate estimation at both normal and high ambient temperatures, and presents much better performance in estimation at an untrained low ambient temperature. In case of inaccurate initial SOCs, the LSTM-SA model shows faster convergence to the true SOC, with mean average errors within 2% in average and average root mean square errors within 3%. The proposed model is further trained and tested under a variety of temperature combinations and aging levels, and performs well in SOC estimation under all of tested conditions, compared to traditional feed-forward machine learning methods.