A highly effective and robust structure-based LSTM with feature-vector tuning framework for high-accuracy SOC estimation in EV

Abstract

This paper describes a new dual long short-term memory (LSTM) model for accurate estimation of the state of charge (SOC) of lithium-ion batteries in electric vehicles. The proposed network has highly effective and robust structure combining a mainstream (m-) LSTM and gradient (g-) LSTM in parallel, which can capture both datatemporal dependency and variability in battery's time-series. The g-LSTM possessing a gradient function consists of very few unit-cells corresponding to about 3 % of m-LSTM cells, and helps prevent the decrease of SOC accuracy caused by sudden changes of current and voltage during charging and discharging. Experimental results show that due to the gradient-tuning effect of feature vectors, the proposed model offers an innovative approach to predicting the SOC patterns with extraordinary precision, resulting in remarkably improved accuracy, on average 12.02 % higher than that of the vanilla LSTM. Further, the proposed dual LSTM demonstrates a fast convergence speed in the training process, and achieves highly accurate SOC estimation, even on unexpected data. Consequently, the computationally efficient and effective g-LSTM collaboration provides a highly robust and strong LSTM network structure to accurately estimate battery SOC, which helps maintain stable performance.