Li-ion battery has been broadly used in industrial and commercial areas in recent years for its no leaks, no pollution and no noises. The currently used state of charge (SOC) estimation methods based on extended Kalman filter (EKF) doesn't have a very good accuracy due to the modeling error, which could influence the performance of the battery management system (BMS) and the control of the host machine. Considering about this, Randles model was adopted which has a good precision and exponential function and more orders of the polynomial were introduced to increase modeling precision. In the experiment, the standard deviation of the estimation error based on the improved model declined 64.43% comparing to the original model. The experiment result shows through improving the battery model the SOC estimation accuracy basing on EKF is improved greatly, which is significant for the performance of BMS and the host machine.