Improved Long Short-Term Memory: Statistical Regression Model for High Precision SOC Estimation of Lithium-Ion Batteries Adaptive to Complex Current Variation Conditions

Abstract

Lithium battery health management is of great significance to promote its wide application. Its accurate battery modeling and state prediction can ensure the safe start-up and stable operation of battery management system. A new method for estimating the charge state of lithium-ion batteries based on phase space reconstruction was proposed by combining long and short term memory network and statistical regression. Compared with the traditional method, the improved LSTM improves the accuracy of prediction by adding data feature dimension through phase space reconstruction, and the segmentation prediction reduces the complexity of data and improves the learning speed. By combining neural network with Kalman filter, it is more consistent with the continuity of lithium battery SOC and further improves the prediction accuracy. Finally, in order to verify the accuracy of the algorithm, an estimation test is carried out using ternary lithium battery. The results show that in BBDST conditions, the prediction ability of the proposed method is significantly improved compared with other algorithms. After 400 cycles of charge and discharge, the prediction error is less than 2.21%, which further indicates that this method has good estimation ability.