Deep learning-based forecasting and model comparison of caisson tilt using numerical simulation–derived time-series data

Abstract Accurate forecasting of caisson tilt during construction is important for ensuring structural stability and effective construction management. However, reliable prediction remains challenging when only limited early-stage observations are available. In this study, the caisson tilt forecasting problem was formulated as an early-stage limited observation forecasting task using time-series data generated from numerical simulations. A staged construction model was developed using PLAXIS2D, and construction-related uncertainties in deep cement mixing (DCM)-improved ground, riprap, and heaving soil were incorporated through Latin Hypercube Sampling, resulting in a total of 1,000 analysis cases. Stage-wise tilt time series over 28 construction stages were extracted. A masking strategy, in which only the initial observations were provided as input was applied to compare the forecasting performance of a statistical model (ARIMA) and deep learning models (LSTM, TCN, and Transformer). Model performance was evaluated over the forecast horizon using MAE, RMSE, and $${R}^{2}$$ . The deep learning models consistently outperformed ARIMA under all conditions, and the TCN model showed stable forecasting performance even under highly limited early-stage observations. The LSTM model demonstrated improved accuracy as more observations became available.