Temporal Temperature Profile Prediction Using Graph Convolutional Networks and Inverted Echosounder Measurements

Abstract

Ocean temperature prediction is a prominent research topic in current ocean science. The empirical modal method, which is based on the inverse echosounder, is one of the most significant methods for analyzing the physical environment of the deep-sea sound layer. This method effectively inverts the temperature profile of the research area. In this article, we propose the time- and self-attention mechanism graph convolutional neural network (ASeTGCN) that uses the inverted data. Unlike traditional time-series forecasting methods, ASeTGCN utilizes graph convolutional networks to capture the inherent spatial correlation of the research area. It also employs self-attention mechanisms to address the nonuniformity of temperature profiles at varying depths. Lastly, it uses time-attention mechanisms to analyze the correlation of temperature profile sequences sampled at daily, weekly, and monthly frequencies. We conducted multiple comparative experiments and related ablation experiments on the proposed model, and our results indicate that the model can effectively extend the time series of temperature profiles in the research area with a root-mean-square error of only 0.19.