Wave field predictions using a multi-layer perceptron and decision tree model based on physical principles: A case study at the Pearl River Estuary

Abstract

In this study, a novel data-driven model is proposed for efficient spatial–temporal forecasting of the significant wave height (SWH), mean wave period (MWP), and mean wave direction (MWD) in the Pearl River Estuary, located in Southern China, using machine learning. The model utilizes machine learning techniques and employs a multi-layer perceptron (MLP) and decision tree (DT) based architecture following a matrix-to-matrix mapping strategy to establish the underlying spatial–temporal relationship between the wind and wave fields for four-day and ocean-scale forecasts. The training and testing of the model were carried out using long-term hindcast wave and wind datasets released by ERA5. A rectangular sea region covering an area of 5° × 5°in the Pearl River Estuary was regarded as the modeling domain. The data-driven model was validated against a 10 years of historical hindcast data, and the global errors were analyzed. The forecasting performance of the established model for tropical cyclone wave events was investigated in terms of both the spatial pattern and temporal variance of the SWH, MWP, and MWD. Extreme waves were generated using the proposed model and compared with the original hindcast results. The data-driven model showed promising potential for accurately capturing ambiguous patterns and features that are variant in both spatial and temporal dimensions, and exhibited obvious superiority over numerical wave modeling in terms of computational efficiency.