Research On Sea Surface Temperature Prediction Method Based On Attention-LSTM Neural Networ

The ENSO phenomenon is an important sea-air coupling phenomenon that occurs in the equatorial Pacific Ocean and plays an important role in global climate change.When the sea surface temperature(SST)in the eastern equatorial Pacific experiences abnormal warming/cooling for more than 5 consecutive months,an El Ni(?)o/La Ni(?)a phenomenon occurs.Therefore,studying and predicting the SST dynamics in this region has important scientific significance.This thesis establishes an LSTM neural network prediction model with additional attention mechanism in the input layer,abbreviated as Attention-LSTM.Based on the Attention-LSTM model,SST prediction is performed using ocean data monitored by ocean observation buoys.The main research content of this thesis is as follows:Firstly,based on the Attention-LSTM model,single variable input data is applied for SST prediction research.On the basis of exploring the impact of training set length on training results,the Attention-LSTM model is used to predict the El Ni(?)o and La Ni(?)a SST data obtained from equatorial Pacific buoy stations at multiple time periods and stations for one year.Research has found that in the prediction of SST at experimental sites,the mean square error of the LSTM model is around 0.5 ℃,while the mean square error of the Attention-LSTM model does not exceed 0.31 ℃,proving that the prediction accuracy of the Attention-LSTM model is higher than that of the traditional LSTM model;The Attention-LSTM model also has a certain improvement effect on the Spring Predictability Barrier(SPB)phenomenon predicted by SST at different stations in the eastern Pacific Ocean during the year of ENSO phenomenon.Secondly,Sparrow Search Algorithm(SSA)is used to optimize the key hyperparameter in the Attention-LSTM model,and multivariable input data is used for SST prediction.Using Pearson correlation coefficient,correlation analysis was conducted on the marine meteorological and hydrological factors that may affect SST,identifying four characteristic factors with moderate to strong correlation with SST(10 meter wind speed,sea surface salinity,2 meter temperature,and sea surface density),and forming a5-dimensional multivariate time series with SST sequence.Input a 5-dimensional multivariate time series into the Attention-LSTM model for SST prediction.The prediction results were compared with the results of the single variable input data model in Chapter 3.The experimental results showed that the prediction accuracy of the multi variable input data model was higher than that of the single variable input data model;At the same time,it has been verified that the SSA can improve the prediction accuracy of multivariate input data models;The repeated experiments at multiple stations in the equatorial Pacific also confirmed the above results,and demonstrated the universal applicability of the model in predicting the SST of El Ni (?) o and La Ni (?) a years obtained at multiple stations in the eastern Pacific region;There is also further improvement in the SPB phenomenon predicted by SST in El Ni (?) o and La Ni (?) a years.Finally,based on the ocean data obtained from the multi parameter ocean monitoring buoy independently developed by the research group,the multivariate input data model proposed in Chapter 4 is used for daily SST prediction application.Comparing the prediction results of this model with the LSTM and SSA-Attention-LSTM univariate input data models,the experimental results show that the model has the highest prediction accuracy,indicating that the model also has good application effect in predicting SST daily data.This thesis not only cites the Attention mechanism to further improve the LSTM model to improve the prediction accuracy of SST for El Ni (?) o and La Ni (?) a years,but also adds data features.The SST data is combined with four closely related features to form a five dimensional time series,which is input into the neural network model constructed in this thesis.The SSA algorithm is used to optimize model parameters,thereby improving the prediction accuracy of SST.