Research On Spatiotemporal Prediction Of SST Based On ConvLSTM Model

With the development of Chinese marine economy and the improvement of marine strategic position,the requirements for sea surface temperature(SST)prediction are gradually increasing.The realization of fast and accurate prediction of SST can provide a solid and reliable guarantee for the development of various marine industries and marine defense security in China.The traditional prediction of SST mostly uses numerical forecast method,which has complex physical equations and limited prediction time.The statistical prediction of SST based on mathematical statistics has low prediction accuracy.At present,the deep learning models are widely used in the prediction of SST,but most of them only consider the time characteristics of SST,and can only realize the prediction of single-point SST time series,lacking the research on the spatial characteristics.In fact,SST not only has time characteristics such as periodicity,seasonality and trend,but also has obvious spatial characteristics.In order to fully consider the spatio-temporal characteristics of SST,this paper constructs a spatio-temporal prediction model of SST based on convolution long-short term memory(ConvLSTM)neural network.The convolution operation is used to extract the spatial characteristics of SST.The long-short term memory(LSTM)neural network is used to learn the time characteristics of SST.The high-precision spatio-temporal prediction of SST in the South China Sea is realized.Firstly,through the research of the temporal variation characteristics,spatial distribution characteristics,time series autocorrelation and global spatial correlation of SST in the South China Sea,it is found that there are obvious temporal and spatial characteristics of SST,and SST is not only affected by its own historical temperature,but also affected by the temperature of the surrounding location.The feasibility and necessity of spatio-temporal prediction of SST are proved.Secondly,by constructing the LSTM model,the time series prediction of single-point SST is realized.The backpropagation(BP)neural network model and the recurrent neural network(RNN)model are constructed as the comparison models.By comparing the prediction effects of the three models,the LSTM model is better than the other two comparison models.The LSTM model is more suitable for dealing with long-term and complex time series of SST.However,the prediction results of LSTM model also have two problems.The prediction effect at different locations is quite different and the prediction error is large at some extreme points of SST.Finally,by constructing the ConvLSTM model,the spatio-temporal prediction of SST in some areas of the South China Sea is realized.The prediction results show that the spatio-temporal prediction effect of SST based on the ConvLSTM model is better than the time series prediction effect of the LSTM model,and the problems existing in the LSTM model are well solved,which proves that it is necessary to consider the spatial characteristics of SST for the prediction of SST.The model by connecting convolution neural network(CNN)and LSTM in series,which called CNN+LSTM model,is constructed as comparison model.A spatio-temporal empirical orthogonal function(EOF)model is also constructed as comparison model.The prediction results show that the prediction effect of ConvLSTM model is better than comparison models,which proves that the spatio-temporal prediction model of SST based on ConvLSTM constructed in this paper has excellent prediction effect.