Abstract: | Advanced battery management technology relies on the near-future prediction of state of charge (SOC). However, the accumulation of errors and the diminished time-dependency lead to a decline in prediction accuracy. In this paper, the machine learning algorithms combined with multi-step prediction strategies are proposed to improve the accuracy of SOC over multiple steps ahead. The effects of different multi-step prediction strategies are studied based on actual lithium battery data. The results show that the actual lithium battery SOC prediction exhibits a significant linear characteristic during the charging phase,and conversely, a nonlinear characteristic in the discharging process. Furthermore, with the prediction step size of 15, the MAPEs of the LR, KNN, and RF models are below 6%, and the R2 values are greater than 0.90. It is found that the LR combined with MIMO shows the greatest potential for practical applications. |