An enhanced lithium-ion battery state-of-charge estimation method using long short-term memory with an adaptive state update filter incorporating battery parameters

The advancements in deep learning strategies offer a promising solution for accurate state-of-charge (SOC) estimation, which serves as a foundation for ensuring the reliable and safe operation of lithium-ion batteries. However, the erratic dynamics and ignorance of the battery state compromise the performance of these methods. Therefore, in this paper, a novel method that incorporates battery domain knowledge sequences is proposed for SOC estimation. First, the current and voltage sequences are decoupled into an adaptive multi-timescale identification strategy (AMIS) with frequency feature decomposition to identify the dynamic battery parameters of a lithium iron phosphate battery. Second, the current and voltage are augmented with the identified dynamic parameters by the AMIS and used as inputs into a multi-layered long short-term memory (LSTM) network, as ALSTM. Finally, to mitigate the negative effect of temperature uncertainties, an adaptive squared-gain unscented Kalman filter (ASGUKF) is proposed to eliminate noise and optimize the final SOC by ignoring the high time dependence of the battery system. The results show that the proposed ALSTM-ASGUKF method is effective and has an optimal mean absolute error and root mean square error of 0.0806% and 0.0986%, respectively, even at low temperatures using two batteries, with only a slight increase in computational cost. Furthermore, its validations and applications at various temperatures demonstrate its effectiveness and the potential of battery domain knowledge to improve the SOC performance of lithium-ion batteries.