open access

https://hdl.handle.net/2115/100463

Effect of Fold Number K in Repeated K‐Fold Cross‐Validation on Deep‐Learning‐Based Consolidation Settlement Prediction

英語

cross-validation

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

Predicting consolidation settlement in large-scale land reclamation remains difficult because oedometer data are spatially sparse and natural marine clays exhibit strongly non-linear structured behavior. This study examines the influence of the fold number K in repeated K-fold cross-validation on uncertainty quantification and consolidation settlement prediction using a deep neural network framework. The framework was trained using 49 boreholes and 634 oedometer-test samples from Kobe Airport, Japan. The repeated K-fold partitioning was performed at the oedometer-test-sample level, and model uncertainty, σmodel, and predictive uncertainty for observed oedometer data, σpred, were evaluated separately. Independent blind tests at two boreholes excluded from training showed that more than 95% of the laboratory pressure-void ratio data points were included in the 95% prediction interval. The uncertainty of the predicted e–log p′ relationships was then propagated to a coupled soil–water FEM analysis at monitoring point KC-1 through Monte Carlo simulation. Within the investigated range of K = 3–10, no clear dependence of the predicted void ratio uncertainty or final settlement uncertainty on K was observed. The simulated settlement reproduced the early-to-mid-term field monitoring data for the modeled Holocene clay layer, while later discrepancies were attributed to settlement of the underlying Pleistocene strata. These results suggest that moderate K values, such as K = 5, can be a practical compromise between computational cost and uncertainty evaluation for the present dataset and modeling conditions.

会議発表論文

Version of Record