Adaptive graph neural network based South China Sea seawater temperature prediction and multivariate uncertainty correlation analysis

摘要

Changes in ocean temperature over time have important implications for marine ecosystems and global climate change. With the development of artificial intelligence, research on ocean temperature prediction has made some progress. However, existing methods are mostly limited to temperature prediction of isolated points on ocean surface, with less vertical studies. And existing graph neural network-based methods typically use predefined graphs, which cannot adap-tively capture unknown associations between data. In this paper, we propose a new adaptive spatiotemporal dynamic graph convolution network to predict three-dimensional sea water temperature. Combined with adaptive graph learning and k nearest neighbor clustering methods, the network can automatically mine unknown dependencies between sequences based on raw data without any prior knowledge. Temporal and spatial dependencies in time series are further captured using temporal convolution and graph convolution. The graph learning, graph convolution, and temporal convolution modules are jointly learned in an end-to-end framework for the three-dimensional seawater temperature prediction task. In this paper, the prediction experiment is carried out using the high-resolution three-dimensional temperature and salt datasets from the Copernicus Global Ocean Physics Reanalysis. The results showed that our method achieved the best predictive performance on all prediction scales compared to current mainstream methods, with MAE increasing by an average of 21.7 and RMSE of 23.3.