Research On Sea Surface Temperature Prediction Based On Non-stationary Time Series

Sea Surface Temperature(SST),as an important indicator to balance the surface energy and measure the heat of the sea,has a significant impact on the changes in the marine climate and ecosystem,and is associated with many natural disasters,such as El Nino,Nanila,and Red Tide.etc.Therefore,improving the accuracy of SST prediction is of great significance to global climate,marine disaster prevention and mitigation,marine environment and fisheries.SST observation data is a typical time series data.Many scholars use time series regression methods to perform statistics and analysis on historical SST data to obtain the development trend of SST and make predictions.It is difficult to use traditional methods to predict sea surface temperature(SSTP),and the accuracy is low because the SST sequence shows obvious non-stationarity under extreme climatic conditions.SST prediction based on deep learning methods has become a hot topic in current research with the wide application of deep learning in the fields of natural language processing,stock prediction,image processing,and target detection in recent years,and it shows better advantages such as high prediction accuracy and better robustness.Due to the limitations of its own conditions such as insufficient reliability of the observation equipment and poor anti-interference,the SST sequence shows significant non-stationarity,and there are certain errors,especially the SST sequence under special climatic conditions(El Nino,Nanila,etc.),resulting in low prediction accuracy of SST.In addition,SST data has strong regional correlation,effectively using the spatial characteristics of SST data,and fusing them with time characteristics for SST prediction,which can greatly improve the prediction accuracy of SST.Therefore,in view of the non-stationarity and strong regional correlation characteristics of SST data,this topic is to improve the prediction accuracy of SST according to different prediction models.The main research contents of this paper as follows:(1)For single-point SST time series,due to the obvious non-stationarity of SST series,it affects the prediction accuracy of single-point SST time series.Therefore,reducing the non-stationarity of data is the key to improving SST prediction accuracy.This paper comprehensively considers the advantages of Empirical Mode Decomposition(EMD)and Gated Recurrent Unit(GRU)neural network,and proposes a sea surface temperature prediction model based on EMD-GRU,which greatly improves the SST prediction effect.First,use EMD to decompose the SST sequence to reduce the non-stationarity of SST,and then input the processed sub-sequences into RNN,LSTM,and GRU networks for experiments,and use the original sequence to directly input the three kinds of neural networks.The prediction accuracy is compared.The results show that the EMD-GRU model proposed in this paper has the best effect.Its MSE and MAE are 0.236 and 0.371,respectively,while the MSE and MAE that directly use GRU are 0.874 and 0.691 respectively.(2)SST time series not only have time characteristics,but also have spatial and regional correlation.The article fully considers the impact of spatial correlation on prediction accuracy,combines EMD and ConvLSTM neural network,and proposes an EMD-ConvLSTM model to analyze regional SST sequences.Make predictions.First,use the EMD method to process the SST matrix sequence,decompose it into several imf matrices and a res matrix,and perform a separate prediction for each sub-matrix value sequence.In the prediction part,in order to fully obtain the spatial regional correlation of the SST time series,the local spatial features of the SST are first extracted through the convolution operation;then the prediction model is constructed using ConvLSTM,and the predicted value of the sub-matrix sequence is output.At the same time,in order to dig out the influence of time distance on the prediction results and highlight the key information in the time dimension,the attention mechanism in the time dimension is adopted in this article to predict the sub-sequences.Finally,the prediction results of each sub-sequence are superimposed to obtain the experimental results of the original SST sequence.The experimental results show that when using 10 days to predict 5 days,the RMSE and PACC of the experimental results are 0.32 and99.03%,respectively,which is a certain improvement compared with 0.47 and 98.79%when using CNN-ConvLSTM directly.Through the above research content,this subject has made some progress in improving the accuracy of SST prediction.There are two main contributions: First,for the non-stationary characteristics of the SST sequence,the EMD algorithm is introduced to reduce the non-stationarity of the SST sequence,and combined with the deep learning method for prediction,and good prediction results have been obtained;secondly,comprehensive Considering the regional characteristics and non-stationary characteristics of SST sequence,an EMD-ConvLSTM prediction method is proposed to improve the prediction accuracy of regional SST.