Wave prediction using Graph Neural Network at Darwin Harbour, Australia

Abstract

Darwin Harbour (DH), Australia, is a flood-dominated estuary in which the waves substantially influence ocean dynamics, especially in the outer harbour. Hence, the prediction of waves is crucial for coastal activities and management in DH. This paper uses a graph neural network (GNN) model to forecast wave characteristics in the nearshore zone of DH. The model was used for next-frame prediction of the wave parameters such as significant wave height, peak period, wavelength, wave velocity, and wave direction, with the model configured to project future values by assessing parameters at a specific domain and in specific time frames. The GNN framework is intended to identify graph dependence via message passing between the nodes. The input to the model is the wave outputs from a 62-day simulation from the SWAN model. The study was carried out using 7194 nodes, with each node linked to 5 neighbour nodes, to forecast dependencies accurately. The data were split 80/20 % for training and testing purposes. Furthermore, the significance of the number of hours as an input on anticipated outcomes was investigated. The performance of the model was also evaluated for short-term (1 h, 3 h and 6 h) and longer-term (12 h and 24 h) predictions; 17 scenarios were investigated, with one scenario for single-step forecasting, and the others for multi-step forecasting. The GNN model forecasts the wave parameters well, with root-mean-squared errors ranging from 0.06 to 0.13, and correlation coefficients R² ranging from 0.96 to 0.84, considering all parameters together. Moreover, it demonstrated that increasing the number of hours for input and time steps for forecasting reduced the model’s performance. Finally, the model was compared with the other three popular ML models. The result indicates that the proposed GNN model is useful for spatiotemporal wave prediction and can be integrated with numerical ocean modelling to examine coastal phenomena.