Accurate prediction of soil moisture is crucial for optimizing crop planting quality and irrigation schemes of jujube trees (Ziziphus jujuba Mill.). This study established a high-precision soil moisture prediction model to improve the irrigation management of jujube trees in southern Xinjiang. Based on hourly datasets of soil moisture content, meteorological data, and irrigation volume for jujube trees during the entire growing seasons of 2021 and 2022 at soil depths of 20, 40, 60, and 80 cm, a long short-term memory (LSTM) neural network model was used to perform multi-step predictions of soil moisture for each soil layer. To expand the model’s prediction range and improve prediction accuracy, a multihead LSTM (M-LSTM) model consisting of four individual LSTM models was introduced. k-fold cross-validation combined with the sparrow search algorithm (SSA) was used for hyperparameter tuning of each individual model to ensure the model’s generalization ability and accuracy. Finally, the final prediction result was obtained by performing a weighted average of the outputs from each individual model. The results show that the M-LSTM model improved the coefficient of determination (R2) of the soil moisture at 1, 12, 24, and 48 h to 0.951, 0.932, 0.870, and 0.815, respectively, according to the dataset of soil moisture content averages from four soil layers. The M-LSTM model effectively enhanced the medium- and long-term prediction accuracy of soil moisture for jujube trees, with particularly significant improvements in predictions at 24 and 48 h. These findings can provide a strong support for the precise irrigation management of jujube trees, thus improving water use efficiency and avoiding unnecessary water waste.