Forecasting Offshore Caisson Tilt via Deep Learning: A Numerical Simulation-Based Approach Accounting for Geotechnical Uncertainty

This study presents a comparative analysis of time-series forecasting models to predict caisson tilt using early-stage monitoring data. To establish a training dataset that accounts for inherent geotechnical uncertainty, 1,000 2D numerical simulations were performed using PLAXIS2D, based on an actual design case in South Korea. To incorporate spatial variability, the subsurface was discretized into 61 independent zones: Deep Cement Mixing (33 zones), foundation rubble (6 zones), backfill rubble (10 zones), and underlying heaving soil (12 zones). Geotechnical parameters including elastic modulus (E), undrained shear strength (S_u), and interface strength reduction factor (R_inter), were varied by up to 50% of their design values. Latin Hypercube Sampling (LHS) was used to assign geotechnical properties to each zone. Each case simulated a 28-stage construction sequence, with caisson tilt extracted at each stage to generate time-series data.

Four forecasting models such as ARIMA, LSTM, Temporal Convolutional Network (TCN), and an encoder-only Transformer, were evaluated. The dataset was split into 680 simulations for training, 170 for validation, and 150 for testing. Forecasting performance was assessed across varying initial observation lengths (cut = 3, 5, 10, 15, and 20 stages) to predict all remaining future stages. Results indicate that while the statistical baseline (ARIMA) showed consistently high errors regardless of observation length, with RMSE values of approximately 0.09 at cut = 3 and 0.08 at cut = 10. In contrast, deep learning models exhibited clear error reductions as more initial observations became available. Among the tested models, the TCN achieved the highest accuracy, with RMSE values of approximately 0.006 at cut = 10 and 0.004 at cut = 15. The encoder-only Transformer model also maintained stable performance for cut ≥ 10, with RMSE values below 0.01.

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