Evaluation of Reduction and Validation Strategies in the Prediction of Extreme Ocean Events

Introduction - Climate change is intensifying extreme ocean phenomena, such as increased wave height, posing significant risks to marine infrastructure. This study aims to improve the prediction of maximum wave height (Hmax) and its ratio to significant wave height (Hmax/Hs), using real buoy data from Bilbao-Vizcaya, Cabo de Peñas, Estaca de Bares, and Villano-Sisargas.

Methods -Three predictive models were applied: Linear Regression (LM), Support Vector Regression (SVR), and Random Forest (RF). The study was divided into two phases. In the first, data reduction techniques were analyzed, including instance reduction (through ordered removal of rows) and variable reduction (by eliminating features with low correlation to the target variable). In the second phase, validation techniques were evaluated, specifically Walk-Forward and Rolling Window, testing different window sizes and training set compositions (with observed or predicted values).

Results - Results showed that reducing both the number of instances and variables is feasible without significantly impacting performance metrics (MSE, MAE, RMSE, R²). LM and SVR yielded the best results. While both validation strategies performed similarly, Rolling Window proved faster and more effective with larger windows. However, incorporating predicted values into the training set notably degraded model performance.

Conclusion - This work demonstrates the feasibility of deploying predictive models in real-world settings, enabling early warnings of potentially destructive wave events using real-time buoy data, thereby improving planning and safety in ocean engineering.