<p>With the frequent occurrence of landslide accidents, the prediction of landslide displacement using monitoring data has become a major focus of attention. Previous research on landslide displacement prediction based on deep learning (DL) mainly focuses on the temporal evolution of landslides, ignoring the spatial correlation between features and displacements. And the characteristics of black-box models make their prediction results lack interpretability and generality. To solve these problems, a new landslide displacement prediction model GAT-TAM-LSTM is established in this study, which is based on the Long Short-Term Memory (LSTM) model by introducing the Graph Attention Network (GAT) and the Temporal Attention Mechanism (TAM). The model first passes the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to decompose the landslide displacement into trend and period terms. Using TAM and GAT, LSTM is able to capture landslide evolution patterns in the spatio-temporal dimension and explain the spatio-temporal dependence of the predictions. In this paper, three landslides with different geographical locations, data volumes and monitoring frequencies are studied and the results show that the model consistently outperforms other prediction models and that the model is able to pay attention to the information of different temporal nodes through the weights of the TAM and the self-learning property of graph neural networks, and to describe the spatial dependence between the slope and its stability influencing factors as a graph structure. In summary, the GAT-TAM-LSTM model can be widely used in a variety of scenarios and provide accurate interpretation and analysis of the prediction results.</p>