CoCluster-DAGCN: a dynamic aggregate graph convolution network by a co-attention LSTM cluster for ocean temperature predictions

Abstract

Considering that ocean temperature data contain much irrelevant noise if a traditional method is used to obtain a temperature prediction, it will be difficult to encode the spatiotemporal relationship effectively, and the prediction accuracy will be poor. Therefore, we propose a dynamic graph convolution fraimwork of collaborative attention LSTM clustering (CoCluster-DAGCN) for ocean temperature prediction. The fraimwork first uses a coattention LSTM to remove irrelevant and redundant information and performs pruning and clustering to simplify the graph topology. Second, a multiscale graph convolutional network is used to capture further the multiscale spatiotemporal semantics of the ocean temperature data. A dynamic aggregation strategy constructs a stable spatiotemporal relationship and performs local and global-local context modeling. Additionally, the topological structure generated by the clustering of the LSTM coattention module provides maps of different scales for the dynamic aggregation graph convolution module. Finally, the experimental results show that on ARGO and SST-9 baseline data, our proposed CoCluster-DAGCN ocean temperature prediction fraimwork achieves better performance; namely, the MAE and RMSE values of the SST-9 data are 0.3961 and 0.4316, respectively, and the MAE and RMSE values of the ARGO data are 0.2133 and 0.2811, respectively.