Research On Sea Surface Temperature Prediction Method Based On Deep Learning

Sea surface temperature is an important engine for the entire sea conveyor belt,and sea temperature changes have a greater impact on weather forecasting,underwater communications,and sea pastures.Therefore,the effective prediction of sea surface temperature has become an internationally important scientific research topic.This topic uses deep learning's powerful data mining capabilities and feature learning capabilities to conduct the following research on sea temperature prediction methods:First of all,in view of the fact that global sea surface temperature data has correlation characteristics both in time and space,this paper transforms the prediction problem of sea surface temperature into the prediction problem of time and space sequence.The convolutional long short term memory network(ConvLSTM)and the convolutional gated recurrent unit network(ConvGRU)are used to mine the time and space features of the 3D data of the space-time sequence,which avoids the prediction results caused by the traditional time regression model ignoring the influence of space discontinuity in space.In the experimental test,the actual global sea surface temperature remote sensing observation data from 2007 to 2016 was used as the training data set,and the data of one year in 2017 was used as the test set to predict the temperature in the next ten days.Through the analysis of the spatial error and time series error of the prediction and observation results,it can be concluded that the prediction results maintain the spatial profile and position information of the sea temperature observation data in the overall space,and the various error indicators also meet the actual prediction needs.Secondly,using the sparse convolution and feature extraction capabilities of the convolutional neural network,the paper built a three-dimensional convolutional autoencoder network architecture(CAE)for temperature prediction.Compared with convolutional recurrent networks,the network has the characteristics of fewer network parameters,shorter training time,and easier optimization.Experimental test results show that the shorter the prediction time of the network,the smaller the prediction error.The maximum root mean square error of its one-day to ten-day forecast for the whole year is 0.26 ° C,and the forecast curve fits well with the real curve even at some places where the details are rich.Comparing the three networks of CAE,ConvLSTM and ConvGRU,it can be seen that the training time of CAE is the shortest,followed by ConvGRU and ConvLSTM.Judging from the prediction error results,the spatial prediction error and the temporal prediction error of CAE are the smallest.ConvLSTM and ConvGRU predict the error from 1 day to 4 days is similar,but as the number of prediction days increases,ConvGRU performance is more stable than ConvLSTM.Compared with the traditional forecasting method,the prediction of the Pacific region near the equator for one week shows that the RMS error of this paper is 0.39?,and the traditional method is close to 0.5?,which shows that the network constructed in this paper has better prediction performance.Finally,the paper discusses the feasibility of multi-element collaborative prediction of sea surface temperature by using three sea elements,sea surface height,sea surface wind field data and sea surface temperature.The prediction results show that the overall collaborative prediction improves the prediction accuracy,but the improvement is lower,and the effect is not obvious.