Geographically And Temporally Weighted Least-Square Support Vector Regression Machine And Its Application

Observation data in many disciplines such as geography,environment,meteo-rology,ecology,economics,finance,population,and epidemics are always collected at a specific time and(geographically)spatial location.Spatiotemporal data set of spatial and temporal attributes.Due to its broad application prospects,along with the advancement of computer computing capabilities in recent years and the im-provement of spatio-temporal data production capabilities in various fields,spatio-temporal data statistical analysis methods are becoming a frontier hot-spot in the study of spatial statistics in the emerging branch of statistics.In the 21st century,the spatial data collection and storage methods based on computer and Internet technology are unprecedentedly powerful,and spatio-temporal data presents com-plex features such as unstructured,non-normal and multi-scale,which makes the widely used classic spatial regression models difficult adapting to the complexity of spatial data sets and unable to effectively model,the spatial regression analysis method established under the classic statistical framework is facing new challenges.The combination of statistical machine learning techniques and spatial regression models significantly improves the effectiveness,reliability,and interpretability of traditional spatial regression methods,and provides valuable new ideas for the re-search of regression analysis of complex spatial data.At present,there has been some research on the study of spatial data through a machine learning method-support vector machine(SVM).However,the research on modeling and analysis of spatio-temporal data using support vector machine algorithm is few and inadequate.At first,based on the previous work on spa-tial data,this paper combines the geographically and temporally weighted regres-sion(GTWR)method with the least-squares support vector regression(LS-SVR)to establish the geographically and temporally weighted least-squares support vec-tor Regression(GTW-LS-SVR)model,which studies the local geographically and temporally weighted least squares support vector estimation method of regression coefficient function and the cross-validation selection method of the optimal hyper-parameters of the model.Then,through the random simulation experiment of the geographically and temporally weighted least-squares support vector regression model and the traditional geographically and temporally weighted regression model,the mean square error and the sum of squared residuals of the model’s predicted values under the two estimation methods are compared and studied.The fitting ac-curacy of the geographically and temporally weighted least-squares support vector regression model is higher than that of the traditional geographically and temporally weighted regression model.Finally,the geographically and temporally weighted least-squares support vec-tor regression model was used to analyze the 1991-1998 precipitation spatial and temporal data sets of 54 meteorological observatories in Xinjiang,and the spatial and temporal distribution characteristics of precipitation and the spatial heterogene-ity of the influence of altitude on precipitation were studied.The geographically and temporally weighted least-squares support vector regression model was used to pre-dict the precipitation of 54 measuring stations in Xinjiang from 1991 to 1998 for grade analysis.The empirical study found that:1)The distribution of total an-nual precipitation in Xinjiang has significant spatio-temporal heterogeneity.2)The change trend of precipitation in different spatiotemporal locations is different.3)The analysis of the precipitation level in Xinjiang shows that with the geographical location and time Changes,there are great differences between the total annual pre-cipitation in Xinjiang.The results of empirical research provide scientific decision-making basis for the effective use of Xinjiang’s water resources to prevent floods and droughts.