Research On PM_2.5 Concentration Prediction Method Based On CART-LSTM

Due to the interaction of various factors,the complex variation pattern of PM_2.5 brings difficulties to the prediction of PM_2.5,and the changeable atmospheric reaction system also makes the accurate prediction of PM_2.5 concentration face severe challenges.For the prediction of PM_2.5 concentration per hour,it is very beneficial to accurately predict PM_2.5by considering the correlation between pollutant factors and analyzing the variation law of PM_2.5 concentration pattern.Based on this,a prediction model based on classification regression tree（CART）and long and short term memory neural network（LSTM）was proposed in this paper.The data set was divided into multiple subsets in a hierarchical way,and the prediction model was established for each leaf.The prediction model fully considers the multi-mode characteristics of PM_2.5 concentration variation,and solves the global-local duality problem.First of all,the air quality data were analyzed to explore the influencing factors and trends of PM_2.5.Taking Beijing area as an example,PM_2.5 concentration is affected by local pollutant concentration and meteorological conditions.Using data analysis technology,PM_2.5 concentration change trend visualization and analysis of PM_2.5 has a variety of change patterns.Secondly,features were selected based on grey correlation analysis and random forest model.For all the data samples,the grey correlation analysis algorithm was used to calculate the degree of independent influence of a single feature on PM_2.5.The random forest model was used to evaluate the coupling effect of multiple features on PM_2.5 concentration.The correlation between pollutant concentration and meteorological characteristics and PM_2.5concentration was determined by comprehensively considering the two assessment results,and the input variables of the prediction model were finally determined.Thirdly,based on various variation patterns of PM_2.5 concentration,the method of classification regression tree and long short-term memory neural network was constructed to predict PM_2.5 concentration.Classified regression tree is used to divide the whole data into several subsets with similar properties according to the form of hierarchical tree to capture a variety of variation patterns of PM_2.5 concentration,that is,each leaf may be an existence pattern.The neural network was trained with its own training samples on each node,and then the prediction model of each leaf was selected from one global and several local models on the path from root node to leaf node according to the validation error of leaf nodes.Finally,the rationality of the proposed model is verified by constructing the prediction problem for the next 24 hours.The method proposed in this paper is verified by experiments.The experimental data were real air pollutant concentration and meteorological data in Beijing area on January 1,2016 and October 31,2017.