Abstract: Reservoir landslides occur frequently in the Three Gorges Reservoir area. Predicting the deformation of the landslides is an important measure to reduce the risk. This paper constructs a Chinese reservoir landslide knowledge graph combined with multivariate Taylor series Kalman filter and proposes a knowledge graph optimized Kalman filter model KG-MTKF. Taking the
Xinpu landslide in Fengjie County in the Three Gorges Reservoir area as an example, the effectiveness of the model was verified by using on-site monitoring data. Results show that compared with the monitoring data, the multivariate Taylor series Kalman filter model (MT-KF) and the knowledge graph optimized Kalman filter model (KG-MTKF) show good consistency in the stationary stage when used to predict reservoir landslides. In the initial stage and step-like stage of the landslide deformation, the prediction accuracy of KG-MTKF is higher. The error of the two models in the initial stage is relatively large, which is caused by the small initial deformation value of the landslide and the more significant influence of system noise. The errors of the two models are
relatively small in the step-like stage, the stationary stage, and the entire monitoring cycle, and KG-MTKF has higher accuracy than MT-KF. For nonlinear dynamic systems like Xinpu landslides, the KG-MTKF can maintain high accuracy and strong robustness at different positions and deformation stages.

Key words: reservoir landslide, deformation prediction, knowledge graph, Kalman filter, prediction accurac