The unsaturated zone's water transfer is an important component of the hydrological cycle. In arid and semi-arid regions, limited water resources constrain vegetation productivity and drought resistance in ecosystems. Therefore, accurately predicting soil moisture (SM) in the unsaturated zone is crucial for the rational use of water resources. Although data-driven deep learning methods have been widely applied in various fields, their application in hydrology has been limited due to the lack of physical mechanisms. To address this, this study utilized measured data from different soil profiles and meteorological observations, applying the Hydrus-1D model for soil hydraulic parameter inversion to generate a physically meaningful SM dataset, which was then used to train a Long Short-Term Memory (LSTM) model. The model was then validated using automatic station data and employed for future predictions. The results show that the LSTM model trained on the physically meaningful SM dataset significantly improves the accuracy of SM prediction, especially in short-term forecasts. The model performed particularly well in predictions at depths of 60-80 cm and 80-100 cm, with R2 values of 0.9989 and 0.9971, respectively, MAE values of 0.0003 and 0.0005, and RMSE values of 0.0004 and 0.0008. When using observed station data for predictions, the model's accuracy was similarly impressive, with an R2 value of 0.8547. The framework proposed in this study, which combines physical mechanisms with data-driven methods for unsaturated zone SM prediction, provides new theoretical foundations and practical references for water resource management and ecological protection in arid and semi-arid regions.