Spatio-temporal Information Mining And Prediction Of The Influence Of SST On Precipitation In China Based On Machine Learning

The ocean occupies about 71% of the earth's surface area and is one of the main factors determining the development of the earth's climate.As one of the most important factors of marine hydrological environment parameters,sea surface temperature has always been the focus of oceanographers' research.The main factors affecting precipitation in China are monsoons and typhoons driven by the sea temperature of the Indian Ocean and the Pacific,and coupled shocks of sea temperature and air at different time scales.Most previous studies have used temperature anomalies in a specific ocean area or coupled sea temperature and air-sea shocks in a certain ocean area to study the impact of sea temperature on precipitation in China.However,whether it is an intraseasonal shock,an interannual or an interdecadal shock,it is impossible to fully explain the precipitation activity in a certain period of time.This article is different from the traditional research methods,taking the sea surface temperature of 30°E-70°W & 50°S-50°N as the research scope,which is also the main source and active sea area of monsoons,typhoons and Kuroshio that affect precipitation in China.Use machine learning methods to explore the temporal and spatial effects of sea temperature on precipitation in China,and finally use Ensemble Empirical Mode Decomposition-Focused Time Delay Neural Network and XGBoost(Extreme Gradient Boost)models to try Using sea temperature to predict the precipitation in China after clustering in the medium and long term,a prediction model suitable for each region is obtained.This paper uses the k-means clustering algorithm to cluster the 36-year weekly precipitation data in China into 7 regions.Among them,Region 1,Region 2,Region 4,and Region 7 are in the monsoon region,and the precipitation is affected by monsoon,typhoon,El Ni?o and southern waves.The impacts of ENSO and intra-seasonal oscillations(MJO)are greater,while the majority of regions 3,5,and 6 belong to the temperate continental climate and plateau mountain climate,and precipitation is less affected by ocean activities.After the sea surface temperature is processed by Principle Component Analysis(PCA)dimensionality reduction,the predictive factors of the input model are determined through correlation analysis and feature engineering.After exponentially smoothing the weekly precipitation in each area,it is input into the Focused Time Delay Neural Network(FTDNN),and the relationship between it and the sea surface temperature after PCA dimensionality reduction is solved,and the results are: Area 1,Area 2,Area 4 and The optimal delay time for areas subject to greater ocean activity such as area 7 is greater than the optimal delay time for areas far away from the ocean and less affected by the ocean.Furthermore,the optimal delay time for predicting precipitation in each area is obtained,and the sea surface temperature is established.The main component of and the time delay relationship of precipitation in each region.In the prediction of each area for multiple consecutive weeks,it is concluded that the pure FTDNN model has a better prediction effect in areas that are less affected by ocean activities,but the prediction accuracy is worse in areas where precipitation fluctuations are relatively large,and it is further combined with EEMD In the future,the mean square error of each region and the average absolute percentage error of most regions have been significantly reduced,and the performance of the model has been greatly improved,which shows that the ensemble empirical mode decomposition(EEMD)can improve the prediction accuracy of the FTDNN network The above has played an obvious role,giving full play to the advantages of EEMD decomposition.Compared with the previous two models,the prediction accuracy of the XGBoost model selected by feature engineering has been further improved.In particular,the average absolute percentage error(MAPE)has been significantly reduced,showing that the gradient boosting decision tree is the core The superiority of the XGBoost model.This paper establishes the relationship between sea surface temperature and precipitation in China,which provides a basis for the prediction of precipitation in areas with a lack of stations and a small number of stations,and is helpful to the study of the climate in areas with long-term absence of precipitation records.