Considering the real-time capacity of lithium batteries at low temperatures is hard to be estimated, the instantaneous current and voltage largely influence the change in transient battery capacity at low temperatures . The deep feedforward BP network model with the Dense fully connected layer as the main body is studied, and the influence of different added layers on the predicted and actual values of the model is analyzed; the BP network with three hidden layers [11-9-12] is used for higher accuracy; Nadam optimization algorithm and the optimization model of Log-cosh loss function that accelerate convergence with the momentum method and the adaptive learning rate method based on SGD expansion are adopted. Overfitting are reduced by regularization for better network generalization. The model is trained and tested based on the data at 0 °C under the HPPC working condition. Consequently, the prediction error of soc at different voltages and currents is about 0.04.