Objective Deformation prediction in foundation pit engineering is of great significance for ensuring the construction safety. Traditional prediction methods have limitations in handling the complex spatiotemporal evolution characteristics of foundation pit deformation. Therefore, it is necessary to conduct in-depth analysis of spatiotemporally coupled prediction methods for foundation pit deformation.

Method Taking the foundation pit project of Longdong Avenue Station on Shanghai Metro Line 21 as a case study, a deformation prediction method for foundation pits based on ConvLSTM（convolutional long short-term memory) network is proposed. The general situation of the foundation pit project of the case station is analyzed. Taking 3 measuring points in the foundation pit Area C as the research objects, the collection and preprocessing methods of monitoring data are expounded. A structural diagram of the foundation pit deformation prediction model based on ConvLSTM is constructed, and the Adam optimi-zer is used to train the model. The measured and predicted va-lues of the horizontal displacement of the underground dia-phragm wall at these 3 measuring points during different excavation stages are compared. Three indicators, namely RMSE (root mean square error), MAE (mean absolute error) and Goodness of Fit (R²), are adopted to evaluate the prediction accuracy of the results, so as to verify the effectiveness and practicality of the proposed method.

Result & Conclusion The proposed method can simultaneously capture spatial features and temporal sequence properties of foundation pit deformation, achieving significantly higher accuracy than traditional methods. At the three monitoring points, the RMSE values are 1.32 mm, 1.20 mm, and 1.26 mm, respectively. The established model accurately identifies key deformation features such as maximum displacement locations and inflection points, and can effectively support real-time monitoring and early warning in foundation pit engineering.