Prediction Of PM2.5 Concentration In Changchun Based On Machine Learning

With the continuous progress and development of society,PM2.5 pollution has become increasingly serious,attracting extensive attention from all walks of life.Taking Changchun city as an example,this study studies the change trend of PM2.5 concentration and the prediction of PM2.5 concentration,aiming to improve the accuracy of prediction,provide a solid theoretical basis for improving PM2.5 pollution,and provide reference for strengthening PM2.5 pollution control,relevant policy formulation and personal protection.There are three steps to study PM2.5 concentration in Changchun.The first step is to analyze the basic situation of PM2.5 concentration at the present stage.The second step is to build a single prediction model,analyze the time series data,and build the LSTM prediction model and GRU prediction model.IDW method and KNN algorithm were used to weight and filter spatial data,and CNN model and Ada Boost model were constructed to capture spatial information.Multi-source features(meteorological factors,other air quality factors and timestamp factors)affecting PM2.5 concentration were collected to establish The Random Forest model and GBDT model,and the importance of features was ranked.The third step is to further improve the prediction accuracy of the single prediction model.By integrating the three perspectives,the combined prediction model LSTM-CNN-GBDT and the Blending model of the Blending theory are first constructed,and then the Blending theory prediction model is improved,that is,a three-layer Blending theory prediction model is constructed.After a series of experiments and studies,the following findings are mainly obtained:Firstly,there is a certain relationship between PM2.5 concentration and heating cycle in Changchun,and generally it is a law of high concentration in winter and low concentration in summer.Secondly,LSTM prediction model based on PM2.5concentration time series data is more powerful and flexible than GRU prediction model.In the spatial dimension,PM2.5 concentration has a large correlation with surrounding cities,but a small correlation with distant cities.In the prediction of PM2.5 concentration,other air quality factors have higher characteristic importance than meteorological factors and time stamp factors.Among meteorological factors,temperature and humidity have higher significance,and among time stamp factors,month has the highest characteristic importance.Thirdly,the prediction effect of the improved Blending theory model,which integrates time,space and multi-source characteristics,is better than that of all single prediction models,combined prediction model LSTM-CNN-GBDT and Blending theory model,indicating that the three-layer Blending theory prediction model can effectively reduce the prediction error and improve the accuracy of prediction.