Abstract: The excavation phase of large-scale port basin projects is the most hazardous stage in the entire construction process.To enhance the accuracy of deformation prediction during this critical phase and ensure construction safety,this study proposes a Temporal Fusion Transformer (TFT) model integrating multi-source information such as construction progress and groundwater level.The proposed model introduces historical deformation data and critical engineering features into a hybrid deep learning framework,accurately predicting future deformation trends in a mega port basin excavation project.Experimental results indicate that the model exhibits excellent predictive performance,achieving a mean absolute error (MAE) of 0.3755mm,mean squared error (MSE) of 0.2597mm²,and mean absolute percentage error (MAPE) of 0.7971%.Ablation experiments further validate the model’s applicability and effectiveness,it provides a reliable technical solution for deformation monitoring of similar large-scale hydraulic structures.