Since the battery state of charge (SOC) cannot be measured directly, and the traditional SOC estimation methods have low accuracy. To improve the SOC estimation accuracy of lithium-ion batteries, this paper compares the effects of different deep-learning network models applied to SOC estimation and proposes a SOC estimation method for lithium-ion batteries based on the DRSN-CW-LSTM network. The method is based on long-short-term memory (LSTM) and deep residual shrinkage networks with channel-wise thresholds (DRSN-CW), using the data information of lithium-ion battery voltage, current, temperature, and capacity extracted in the deep residual shrinkage networks, and the time series data trends are further fitted by LSTM to achieve the prediction of SOC of lithium-ion battery during its service life. In the residual shrinkage module of the DRSN-CW network, an adaptive noise data processing function can be implemented to eliminate the negative impact of lithium-ion battery data stream quality on SOC prediction. In this paper, the proposed network is trained using the lithium-ion battery public dataset and the prediction effects of three neural network models on the two datasets are compared. The experimental results show that the MAE and RMSE of the deep learning model proposed in this paper are controlled within 5% of the average value on both public datasets, and have better noise immunity and prediction performance with high estimation accuracy compared with the other three deep learning models.