Multi-step ahead significant wave height forecasting using global and local view graph representation based on GRU model

Multi-step ahead significant wave height forecasting using global and local view graph representation based on GRU model

Significant wave heights (SWH) are important to be predicted accurately for clean wave energy production, and beach erosion risks. The existing models lack the ability to analyse the dynamic behaviour of oceanic drivers, as a result, they cannot predict SWH at different forecasting horizons. In this paper, an innovative modelling scheme (termed as GLG-DL) based on graph deep learning which integrated global and local graph features has been designed to predict SWH. The global and local graph leaners enable GLG-DL model to capture the global information, and the local trends from oceanic drivers. The extracted graph representations are then used into the gated recurrent unit (GRU) based encoder and decoder to predict multistep ahead SWH for Palm Beach, Gladstone, and Albatross Bay stations, Australia. The GLG-DL model was compared with Auto-regression model (ARM), Auto-regressive multilayer perceptron (AR-MLP), Recurrent neural network (RNN), RNN based attention mechanism (RNN-AM), RNN based Long Short-term Memory (RNN-LSM), Spatial-temporal attention mechanism (STAM), and improved recurrent neural networks (SNN) models. The results demonstrated that the GLG-DL attained higher performance to forecast multistep ahead SWH for all stations. The GLG-DL model is beneficial in the application and optimization of clean energy resource generations worldwide.