Research On Spatio-temporal Series Hybrid Predict Model And Empirical Based On Least Squares Support Vector Machine

High-precision spatio-temporal series prediction is an important guarantee for revealing the spatial-temporal distribution of regional variables and its temporal and spatial evolution.Effective prediction technology is the key to improve the prediction accuracy of research elements.Spatio-temporal series prediction is the important parts of spatio-temporal data mining.However,with the increasing volume of spatio-temporal datasets,the traditional prediction model has inadequate consideration of complex time and space,the nonlinearity of the underground hydrological system and parameter estimation of the model,which greatly affect the prediction accuracy.Hybrid prediction modeling can overcome the deficiencies of temporal-spatial series characteristics for single-prediction model.Hybrid prediction modeling is to build a model with only one final predictor by combining different models with different theories.The hybrid prediction model can take into account the characteristics of the spatio-temporal variables from multiple angles and aspects,and improve the prediction accuracy through the coordinated operation of the various theories.Integrating neural network,machine learning,and time series analysis to establish a hybrid prediction model is an important research method for improving the prediction accuracy at present.Least squares support vector machines have a strong ability to fit nonlinear complexity problems and have been widely used in complex engineering fields.Due to the large fluctuations in the data collected by the regional variables in different spatial points,only a least squares support vector is used as a training model,which cannot exert its own nonlinear characteristics.Due to the large number of series in spatio-temporal datasets,traditional time series methods can only predict one or several sequences,and it is difficult to meet the requirements of spatio-temporal series sequence modeling.Based on the idea of clustering and First Law of Geography,the spatio-temporal datasets are firstly clustered.Each type of spatio-temporal series is fitted with training by least squares support vector,and an intelligent optimization algorithm is used to establish the best time-space sequence hybrid prediction model for each class of dateset.This not only can fully fit the nonlinear characteristics of the regional spatio-temporal variables,but also can well consider the impact between each monitoring point and other relevant sites in the space,so as to improve the prediction accuracy of the model.This is the focus of the study.For a long time,unreasonable over-exploitation of groundwater resources in the oasis has led to a continuous increase in the degree of groundwater mineralization,accelerated the deterioration of groundwater quality,and triggered a series of ecological environmental problems in the Minqin County arid region.Accurate prediction of groundwater elements and effective management measures have important practical value for the rational sustainable development and utilization of groundwater resources in the study area.Regional variables are spatio-temporal variables that vary with space and time.The spatio-temporal data analysis is an important approach to rational analysis and prediction of the degree of groundwater mineralization.The main works of his paper are follows:Due to the complexity and uncontrollability of experiment factors of groundwater system,the measured data will inevitably be affected by noise,and it will result in missing data values and abnormal,which will have a great impact on the results of modeling validation experiments.In order to obtain high-quality research datasets,this paper is to interpolate spatial-temporal datasets by spatial-temporal sequence missing value interpolation methods,to perform noise reduction processing by wavelet analysis,to optimize parameters for selection by cross-validation techology.Considering the spatio-temporal correlation and nonlinearity of spatio-temporal sequences,based on the classification modeling of clustering ideas,the least square support vector machine is applied to fit each type of nonlinear features.Different training modes are set for each type of series,and each type of spatio-temporal sequence hybrid prediction model is established.Grey wolf algorithm has a very good search mechanism.In this paper,the optimal selection process of two-parameters for least squares support vector machine is designed to use the grey wolf algorithm.Setting different search ranges for each type of parameters based on multiple experimental results,each type of optimal space-time sequence mixed prediction model SOM-GWO-LSSVM is established.This paper takes the space-time sequences of groundwater mineralization obtained from 74 monitoring stations as experimental data in Minqin Oasis.Experimental results show that the spatial-temporal hybrid prediction model built in this paper improves the prediction accuracy compared to a single prediction method.It validates the effectiveness of the new model.