A combined robust approach based on auto-regressive long short-term memory network and moving horizon estimation for state-of-charge estimation of lithium-ion batteries

State-of-charge (SOC) of battery is one of the important evaluation indexes for the battery management system (BMS) application and driving range estimation of electric vehicles. However, acquiring accurate battery SOC information is subject to the indeterminacy of initial SOC value, as well as uncertainties and inconsistencies of the battery pack. To deal with those problems, a combined method based on the auto-regressive long short-term memory network (ARLSTM) and moving horizon estimation (MHE) is put forward. Through training on two typical standard tests offline, the ARLSTM is used to learn the sophisticated dynamics behaviors of the battery. The moving horizon optimal algorithm based on the equivalent circuit model (ECM), whose parameters could be obtained from the improved test, is designed to estimate SOC online combined with ARLSTM. The proposed approach, compiled in MATLAB/Simulink, is evaluated under several current loading tests in AMESim. The results demonstrate the ARLSTM can well simulate the battery behavior and estimate battery SOC offline with relatively high precision (0.5%). Compared with the traditional methods based on MHE or LSTM, the proposed method for estimation precision is almost less than 0.2% and the convergence time reached about 500 seconds. Novelty Statement A combined method based on ARLSTM and MHE is investigated in battery SOC estimation in terms of uncertainty or large deviation of initial SOC value. That combined approach can make full use of the data from historical working conditions and historical data of current working conditions. The ARLSTM network is introduced to simulate the dynamic behavior of battery offline using the data collected from the battery dynamic charging/discharging standard test. The trained ARLSTM network achieves satisfactory SOC predicted results. Not only the method has high estimation accuracy of SOC, but also has good generalization ability, that is, adaptability to various working conditions. The proposed method also has an excellent performance in convergence time.

Automated summary

The article introduces a combined method based on ARLSTM and MHE for battery SOC estimation, which effectively addresses the uncertainty of battery SOC initial value. After offline training, ARLSTM can accurately simulate battery behavior and estimate SOC values with relatively high precision; the method shows excellent performance in estimation accuracy and convergence time.