Probabilistic prediction of operable sea states with a random forest ensemble

• Machine learning model to integrate forecast uncertainty into response predictions. • Enable decision-making based on detailed wave descriptions during offshore operations. • Exemplified for suction anchor deployment with 5 years of forecast/hindcast data. • Achieves high accuracy for the 90th percentile compared to deterministic benchmark. • Integration of load predictions with limit-state reliability methods discussed. A Probabilistic Sea State Classification method (PSSC) is proposed to determine the operability of sea states from the forecasted wind and swell sea parameters. Unlike similar models, the proposed method produces an ensemble of possible outcomes from a random forest model, enabling a probabilistic assessment of operability that accounts for the model’s predictive error, the forecast uncertainty, and the random variability in the wave loads. The model was tested through a case study, considering crane wire loads during the subsea deployment of a suction anchor, comparing it against benchmark decisions based on the true sea state and conventional decisions based on the forecasted total wave parameters. Unlike conventional methods, the PSSC Model performed consistently across seasons with varying wave characteristics. Applied to an Ultimate Limit State (ULS) criterion, the long-term distribution of load exceedance over the five-year test set closely matched the benchmark distribution at the target 10 –1 level. However, incorporating forecast uncertainty and modeling errors in the probabilistic assessment altered the distribution of the extreme loads, such that the benchmark was exceeded at the 10 –4 probability level. More work is needed for practical application, establishing a revised load factor, or considering detailed reliability analyses using Monte Carlo simulations. The latter requires extensive training since the model cannot extrapolate beyond known data points.