The spatiotemporal prediction of deep landslide deformation is crucial for landslide early warning and prevention. Traditional methods face challenges in capturing the nonlinear spatiotemporal evolution characteristics of landslide deformation. To address this issue, this study proposes deep learning-based models, including Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Gated Recurrent Unit (GRU), for spatiotemporal prediction of deep landslide deformation. Using monitoring data from inclinometer holes in a reservoir landslide as the research subject, an in-depth analysis of spatiotemporal data was conducted. The dataset spans from 2015 to 2023, comprising 147 sets of spatiotemporal displacement data with a total of 8,742 data points. The dataset was divided into training and testing sets, and the models were trained using the Adam optimizer. The performance of the models was comprehensively evaluated using five metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination (R2). The results demonstrate that the LSTM, BiLSTM, and GRU models exhibit high prediction accuracy, with most prediction errors concentrated within the range of 0 to 1 mm. These models effectively capture the spatiotemporal nonlinear characteristics of landslide displacement, particularly excelling in identifying abrupt deformations caused by weak structural planes. This highlights the adaptability and robustness of deep learning models in handling complex spatiotemporal data. This study provides reliable technical support for short-term landslide deformation prediction and early warning, while also suggesting that future improvements in long-term prediction accuracy could be achieved through “data-model” assimilation techniques. The research outcomes not only hold significant practical value for landslide early warning and engineering management but also provide scientific foundations and technical support for safety monitoring in hydraulic engineering and the construction of digital twin models, demonstrating broad application potential.