Short-term prediction of relative vertical motions at connection points of adjacent moored vessels based on motion-response data and temporal convolutional networks

Accurate short-term prediction of relative motion between adjacent moored vessels is important for safe close-proximity offshore operations. Based on motion-response data from two-vessel mooring tank tests, this study develops a forecasting framework for the vertical relative motion at vessel connection points under bow-to-stern and side-by-side configurations. Multi-step samples are constructed using sliding historical windows, and a temporal convolutional network (TCN) with causal and dilated convolutions is compared with six established forecasting methods, including LSTM, GRU, Transformer, XGBoost, persistence, and constant-velocity extrapolation. The effects of input history length, forecast horizon, and operating condition are examined, and cross-case performance is evaluated using leave-one-case-out testing. With an input history of approximately 27 s and a forecast horizon of approximately 6 s, the TCN achieved the best overall accuracy in all four tested cases, with a case-level macro MAE of approximately 923 mm and R 2 of 0.877. In leave-one-case-out evaluation, it maintained positive R 2 values of 0.73–0.88 for all held-out cases. Physically relevant inputs are selected based on connection geometry and rigid-body motion relationships. The framework shows potential for rolling forecasts and operational warning in close-proximity marine operations.