Uncertainty-Aware Short-Horizon Warning of Large-Inclination Exceedance in Small Fishing Vessels: A Simulation-Based Multi-Model Benchmark

A large heel can develop within seconds on a small fishing vessel, so a short-horizon forecast of inclination is useful for safety only if it reports both the predicted angle and its uncertainty and is turned into an explicit warning decision. We present an uncertainty-aware early-warning benchmark and decision layer for the total inclination angle, framing the task as warning of large-inclination exceedance—that the heel will cross a fixed operational threshold (15∘/20∘/25∘)—rather than predicting a vessel-specific capsize or dynamic-stability limit. Models are trained and evaluated on 90 time series generated from 6-DOF simulations spanning five tonnages, three sea states, and six independent wave-phase realizations under a leakage-safe, multi-seed protocol. The strongest in-distribution forecaster (an LSTM) reaches R2=0.677 (RMSE 3.51∘) pooled over the 1–5 s horizon, and converting its predictive distribution into a probabilistic exceedance alarm lowers the event-level false-alert burden at matched recall; the receiver-operating advantage is threshold-specific—clear at 15∘, marginal at 20∘, and reversed at the rarest 25∘—so the alarm is not uniformly better than a point alarm. Leave-one-realization-out folds confirm that forecasting is robust to wave phase, whereas the smallest (10 t) vessel remains the dominant generalization failure mode. Because nominal interval coverage drops in the large-inclination tail, a conformal recalibration layer is required at deployment. The study is a simulation benchmark: operational value remains conditional on validation against measured seakeeping data and on mapping the thresholds to vessel-specific stability criteria.