Attention Adaptive Temporal Graph Convolutional Network for Long-Term Seasonal Sea Surface Temperature Forecasting

Abstract

Sea surface temperature (SST) plays a crucial role in the global meteorological system, particularly as long-term seasonal variations are significant for analyzing SST anomalies and supporting long-term climate decision-making. Current forecasting methods are primarily focused on short-term or fine-grained predictions and often fail to effectively capture long-term seasonal trends. To address this, we introduce a novel attention adaptive temporal graph convolutional network (AA-TGCN) specifically designed for long-term seasonal SST forecasting. Unlike traditional adaptive graph convolutional networks, the AA-TGCN incorporates an attention mechanism to capture internode correlations and utilizes a pruning strategy from SGAT to eliminate noisy connections, thereby improving the model's inductive learning capabilities. Moreover, the network employs a TCN-like architecture to expand its receptive field, enhancing its ability to grasp long-term trends, and employs differential embedding to further refine the prediction accuracy of seasonal fluctuations. Practical applications in the Bohai Sea and parts of the South China Sea demonstrate that AA-TGCN outperforms existing technologies on multiple scales, particularly achieving significant improvements in the South China Sea regions.

Keywords

• Attention mechanism pruning
• dynamic graph
• long-term seasonal forecasting
• sea surface temperature (SST)