Time domain convolutional network ocean surface temperature prediction model based on multi-feature fusion attention mechanism

Ocean surface temperature is a key factor in the interaction between the ocean and the atmosphere, and rising ocean temperature has the potential to change the global atmospheric circulation system, which in turn has profound effects on the global climate. Accurate prediction of sea surface temperature is of great significance to Marine ecosystems and global climate. Based on the time domain convolutional network, this paper first proposes an ATCN (Attention-based temporal convolutional networks) model of the time domain convolutional network based on the attention mechanism to study the one-dimensional time series. Then, a multi-feature fusion model AM-TCN (With attention-based multi-feature fusion temporal convolutional network) model is proposed to study multi-dimensional time series for the prediction of sea surface temperature in the Yellow Sea and Bohai Sea. In this paper, a total of 15997 sea surface temperature data of the Yellow Sea and the Bohai Sea from 30°N to 42°N and 117° E to 128 ° E for 43 years are obtained by using the reanalyzable values from the website of the European Centre for Medium Range Numerical Weather Prediction (ECMWF) from January 1979 to December 2021. Use the last 60 months (1,860 days) to predict the temperature for the next 12 months (372 days). In the study of one-dimensional time series, this paper compares the improved ATCN model with several mainstream deep learning models, and verifies that the deep learning model has the ability to predict a few anomalies. In the study of multi-dimensional time series, a multi-feature fusion model AM-TCN model is proposed, and the characteristics of El Nino-Southern Oscillation Index (Nino3.4 Index) and ANOM value are added for multi-variable collaborative prediction. The improved AM-TCN model not only occupies less memory, but also has the advantages of short running time. It also improves prediction accuracy. Four model evaluation indexes were adopted in the experiment, namely Mean Squared Error, Root Mean Squared Error and Mean Absolute Error. Coefficient of Determination-R-squared (R
²
) was used to analyze the degree of fit between the predicted sea surface temperature and the real sea surface temperature. The experimental results show that the accuracy of the improved AM-TCN model can reach 98.54%. The prediction performance of the improved model increased by 1.03%, and the overall performance was stable and the accuracy was high.