Significant wave height prediction at multiple sites using sequence decomposition and dynamic spatiotemporal graph neural networks
Abstract
With the continuous development of artificial intelligence technology, accurate prediction of significant wave height (SWH) has become increasingly important in the application of marine energy and renewable energy systems, especially in marine energy development, energy network optimization and response to extreme weather events. This study proposes a sequence decomposition dynamic spatiotemporal graph neural network (DSTGNN) prediction model to improve the prediction performance of significant wave height. The model decomposes the data into trend and seasonal components, uses frequency domain multi-layer perceptron to capture the temporal dynamic characteristics, and combines adaptive dynamic graph neural network to effectively model the complex spatial correlation between multiple buoy sites. Experimental results show that the DSTGNN model significantly outperforms traditional methods on data sets in the southeastern United States and the western Atlantic Ocean, the Gulf of Mexico and the Caribbean Sea, the northeastern United States, the coastal Pacific Ocean of the United States, and the Hawaiian Islands. In the prediction task of the next 6 h, its MAPE values reach 0.1228, 0.2368, 0.3094, 0.2576 and 0.0798 respectively. This model can not only accurately capture the temporal and spatial dependence characteristics of wave data, but also provide higher accuracy in the prediction of extreme wave events, providing strong support for the development and prediction of marine energy and having significant energy application value.