Machine learning prediction of wave characteristics: Comparison between semi-empirical approaches and DT model

Abstract

Prediction of wave characteristics plays a crucial role in design and performance assessment of various coastal projects. The computational complexity and time-consuming procedures have limited the applications of numerical models for wave predictions. This study develops a new model for wave prediction using the capabilities of the M5p Decision Tree (DT) algorithm. The wind speed was employed as the model input parameter, and satellite measured altimeter data was used to train and test the developed models. The proposed DT-based model is compared with the existing semi-empirical wave prediction methods recommended by Coastal Engineering Manual (CEM). For a comprehensive assessment of the model's performance, four scenarios with different input data and modelling approaches are investigated. It was shown that the locally calibrated CEM formula can provide the best performance amongst the modified semi-empirical formulations. Comparison results show that M5p DT models' prediction are far more accurate and closely match the satellite measured altimeter data than the semi-empirical models. Furthermore, the ‘Short-term’ M5p tree model is shown to have the best predictive results, and the obtained results highlight that the proposed M5p model can provide a robust alternative for wave prediction across large spatial and temporal scales.