A hybrid-aided approach with adaptive state update for estimating the state-of-charge of LiFePO4 batteries considering temperature uncertainties

Abstract

Developing an accurate state of charge (SOC) estimation method is crucial for proper monitoring and management of electric vehicles (EVs). As deep learning methods advance, it is critical to design a network structure for SOC estimation that is accurate, flexible, and adaptable to different driving conditions. However, these methods are hindered by high computational costs and trial-and-error training approaches, which compromise their performance. Therefore, this paper proposes a variational epoch selector for a multi-layered long short-term memory (LSTM) network with an adaptive weighted extended Kalman filter (AWEKF) for SOC estimation of lithium-ion batteries. First, a random weight (RW) algorithm is proposed to variably select the required number of training epochs suitable to train the LSTM network for SOC estimation using three domain knowledge, which improves stability, accuracy, robustness against uncertainties, etc., with a significant reduction in model computational training costs and errors. Second, the AWEKF with feedback correction ability is proposed to optimize the estimations of the network to map the nonlinear characteristics and minimize the output SOC fluctuations and errors. The estimations critically investigate the various key factors for the proposed RWLSTM strategy at different temperatures under real-world simulated driving cycles using a lithium iron phosphate battery. Finally, the results show that the mean absolute error and root mean square error of the proposed RWLSTM and RWLSTM-AWEKF strategies are <0.6% and 0.2%, respectively, under various driving conditions, showing their efficiency in estimating the SOC by utilizing battery domain knowledge for BMS applications.