Based on the data set of hourly soil moisture automatic observation at 10 cm depths from 2016 to 2019 in Changsha Hydrometric Station, the neural network of Long/Short Term Memory (LSTM) combined with random sampling learning was used to carry out multi-time prediction of soil moisture. The results showed that RMSE of prediction in 6, 12, 24, 48 h was 0.22%, 0.28%, 0.38%, 0.54%, and coefficient of determination (R²) was 0.99, 0.99, 0.98, 0.96, respectively. The prediction accuracy was better than Autoregressive Integrated Moving Average (ARIMA) model except the 6 h, the deviation was stable and no abnormal value appeared, the prediction accuracy was far better than relevant studies. The results prove that the feasibility of accurately predicting in soil moisture dynamics based on LSTM model, provide computer technology and means for accurate irrigation and drought warning, and data support for the formulation of water resource management policies by government and research institutions.