AI-Driven Time-Series Prediction of Retaining Wall Deformation: A Case Study in Korea

This study evaluates the field applicability of a multi-resolution Convolutional Long Short-Term Memory (ConvLSTM) framework for predicting time-series retaining wall deformations during staged excavation using field measurements from various excavation sites in South Korea. The proposed framework integrates three ConvLSTM models trained with different temporal input resolutions to capture deformation characteristics at multiple time scales. Their multi-step predictions are subsequently refined using a stacking ensemble strategy with a neural network–based meta-learner, which mitigates error accumulation commonly observed in recursive long-horizon forecasting and enhances overall prediction stability and accuracy.

To generate a comprehensive training database, numerical analyses were conducted on a wide range of excavation cross-sections with varying final excavation depths, wall tip restraint conditions, and initial groundwater levels, reflecting diverse geotechnical and structural configurations. The geotechnical and structural properties were defined probabilistically to account for inherent uncertainties in ground conditions and structural stiffness. In total, 4,000 numerical analysis cases were generated and further augmented into 16,000 training datasets through Gaussian noise injection to improve model generalization ability.

For field validation, 34 time-series displacement measurements collected from 11 excavation sites in South Korea were employed to assess the predictive performance of the proposed framework under real construction conditions. When lateral displacement data obtained from earlier excavation stages were provided as inputs, the model predicted retaining wall deformation induced by an additional excavation depth of 5.0 m, achieving an average coefficient of determination (R²) of 0.85 and a mean absolute error (MAE) of 5 mm. Furthermore, the framework demonstrated an average inference time of 0.92 s, confirming its suitability for near–real-time prediction and potential integration with field monitoring systems. These results indicate that the proposed multi-resolution ensemble framework is practically applicable to real-world excavation projects and offers a robust tool for predictive decision-making in excavation safety management.

 

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2023R1A2C1007635).