From Benchmarking to Explainability: Machine Learning Algorithms for Rogue Wave Prediction

Rogue waves pose a major hazard to offshore structures and maritime operations, yet reliable prediction remains challenging due to their rarity and the nonlinear processes involved. We investigate data-driven prediction of the abnormality index, a metric of wave extremeness, using a metocean dataset collected at the K14 radar station in the southern North Sea. Five machine learning models, Elastic-Net regression (ENR), support vector regressor (SVR), random forests (RF), XGBoost (XGB), and a feed-forward neural network (FFNN), are benchmarked. Explainable AI is then applied to examine the physical consistency of learnt feature dependencies. XGB and RF achieve the highest predictive performance (r s = 0.98); however, RF yields a lower recall (79%) than XGB (85%), while both models maintain low inference cost. In contrast, the SVR exhibits strong detection capability (r s = 0.97, recall of 89%) but with higher false alarms, while the FFNN achieves only moderate performance (r s = 0.91, recall of 77%). Both SVR and FFNN involve substantially higher computational cost. ENR performs the worst among all models. Feature importance analysis consistently identifies the dimensionless water depth as the dominant predictor across models, with additional contributions from atmospheric pressure and gustiness. Moreover, spectral parameters exhibit mixed influence, suggesting that rogue waves may arise from multiple overlapping physical regimes rather than a single universal mechanism. Our findings demonstrate that interpretable and computationally efficient machine learning models provide a robust framework for advancing physical understanding of rogue waves and support both future research and operational early-warning applications.