Accurate state of charge (SOC) estimation is crucial for the lithium-ion batteries management system in electric vehicles. This paper proposed a novel method to improve the estimation performance of SOC from the three aspects. First, to overcome the dependence of model on the internal parameters of battery, this paper uses the NARX neural network to build the experimental model. Compared to the feedforward neural network, the NARX model with timing characteristics shows a better performance in state of charge estimation. Second, analyzing the characteristics of the experimental data and combining with the moving window method, the real-time segmentation and output of the data are carried out, the overall model is simplified and the modeling accuracy and speed are improved. Third, the superiority of the newly trained model is verified under several driving cycles by comparing with other existed model.