Abstract: Reservoir impoundment-induced valley contraction deformation poses a significant threat to the overall safety of dams, making the prediction of valley width deformation crucial for dam safety management. This study focuses on the valley deformation characteristics of the Xiluodu Hydropower Station and proposes a physics-guided intelligent prediction model to conduct valley deformation forecasting. Based on the analysis of valley deformation patterns, a statistical regression physics-based model (STPM) is developed to quantify the displacement components contributed by individual triggering factors to valley deformation. Building upon this foundation, a machine learning prediction framework (STPM-GRA-LSTM-RF) integrating grey relational analysis (GRA), long short-term memory networks (LSTM), and random forest algorithms (RF) is constructed with displacement components as inputs, incorporating uncertainty analysis in valley deformation prediction. Key findings include: (1) The valley deformation at Xiluodu is primarily caused by viscoelastic deformation, viscoplastic deformation, and effective stress-induced deformation, contributing 67.71%, 29.75%, and 2.51%, respectively, to the total displacement; (2) Compared with the STPM, LSTM, SVM, and XGBoost models, the proposed STPM-GRA-LSTM-RF model delivers higher predictive accuracy, moreover, by incorporating the physico-mechanical mechanisms of valley deformation, it significantly enhances the reliability of the predictions. This research provides valuable insights for deformation prediction and safety control of high reservoir slopes in analogous projects.