Deformation Prediction of Deep Foundation Pit Support Piles Based on a CNN-LSTM-Transformer Model with Spatiotemporal Feature Fusion

Monitoring data from deep foundation pits exhibit significant nonlinear, nonstationary, and spatiotemporal coupling characteristics. Traditional methods struggle to simultaneously characterize their spatial correlations and temporal evolution patterns. To address these issues, on the basis of measured data from a typical deep foundation pit project in Beijing, this paper proposes a spatiotemporal feature-fused CNN-LSTM-Transformer prediction model for support pile deformation. By constructing a spatiotemporal matrix of monitoring data, the model achieves the synergistic fusion of spatial feature extraction, temporal dependency modeling, and global correlation perception. The results of the comparative analysis indicate that the proposed model demonstrates stable predictive performance across different pile locations and operating conditions. Its root mean square error (RMSE) and mean absolute error (MAE) are reduced by approximately 20% compared with those of the CNN and CNN-LSTM models. Particularly in areas with severe deep deformation and during high-fluctuation stages, the model effectively mitigates prediction lag and error accumulation, demonstrating a superior response capability to local abrupt changes. The findings suggest that this method can provide a reliable data-driven approach for the dynamic prediction of support structure deformation during deep foundation pit construction.