Charge state of battery is one of the important parameters of battery. Accurate estimation of SOC can improve battery safety, prolong battery life and improve battery utilization efficiency. Straightforward valuation of battery SOC is very essential for retired electric vehicle batteries. The traditional SOC estimation method based on extended Kalman filter (EKF) is excessively dependent on the accurate battery model, and requires that the system noise must be subject to Gaussian white noise, and the SOC prediction error is large. To address the mentioned above existing problems, a method based on Recurrent Neural Networks (RNN) to correct the SOC error of Extended Kalman Filter is proposed to realize the accurate estimation of SOC. By comparing the experimental results, it can be obtained that its algorithm can accurate to predict SOC, and the error is within 1.3 %. and has good convergence and robustness, which has high practical application value.