First World Congress of International Consortium on Geo-disaster Reduction (1st ICGdR World Congress)

Excavation of coastal soft-soil harbor basins often triggers retaining-wall displacement and settlement under groundwater fluctuation and stage-by-stage construction loading. This study develops an artificial intelligence-enabled deformation prediction framework based on the Temporal Fusion Transformer (TFT) for a super-large harbor basin excavation in southeastern China. Historical deformation, groundwater level and construction-stage information were organized as observed, dynamic and known future inputs after outlier screening, cubic interpolation and Min-Max normalization. The model was tested under 3, 7, 14 and 20 d forecasting horizons and under rolling forecasting. For the 14 d horizon, the TFT achieved MAE, MSE and MAPE values of 0.7397 mm, 0.8060 mm² and 1.5725%, respectively, outperforming SimpleRNN, LSTM and a standard Transformer. In rolling forecasting, the errors decreased to 0.4844 mm, 0.3910 mm² and 1.0554%. Removing groundwater or construction-stage inputs increased MAE to 0.8763 mm and 0.9188 mm. These results show that multi-source TFT forecasting gives a 14 d deformation estimate for warning-threshold checks in marine geotechnical works.

Geo-disaster reduction; Artificial intelligence; Deformation forecasting; Coastal soft soil; Temporal Fusion Transformer; Early warning