Aimed at the problem that the state-of-charge (SOC) of batteries estimated by a combination algorithm is not ideal when there exits deviation from the initial value, a BP-multi-timescale unscented Kalman filter (UKF) optimization algorithm is proposed in this paper, which applies the multi-timescale theory to the combination algorithm. First, the multi-timescale UKF is used to preliminarily estimate the SOC value. Then, a BP model is constructed to train the data generated by the filtering algorithm, and the self-learning and adaptive capability of the BP algorithm is used to compensate and further reduce the error. The multi-timescale extended Kalman filter (EKF) modeling can not only reduce the SOC estimation error caused by changes in parameters, but also provide reliable data for error compensation, thus improving the model’s generalization capability and reducing the data dependence by the data-driven part. Simulation results show that the proposed algorithm depends on less data, and it outperforms other algorithms in terms of estimation accuracy, generalization capability and robustness. The root mean square error and mean absolute error are less than 1%, and the maximum absolute error is about 1%.