For the problem of low accuracy of lithium battery state of charge (SOC) estimation by a single algorithm, this paper proposes an algorithm combining extreme learning machine (ELM) and cubature Kalman filter (CKF). By dynamically tracking the Thevenin equivalent circuit model parameters by recursive least squares (RLS), the initial estimate of SOC is obtained by the CKF algorithm. The battery operating voltage and current are used as input, corresponding to the estimation error of the CKF algorithm as output, using the ELM algorithm to build an error prediction model based on data-driven. Then the output of the ELM error prediction model is used to compensate for SOC initial estimation value to achieve the purpose of reducing the estimation error. The simulation results show that the proposed ELM-CKF algorithm effectively improves the SOC estimation accuracy compared with the CKF algorithm.