Fine-Grained Prediction And Inference Of PM2.5 Concentration Based On Multi-source Data And Deep Learning

In recent years,China’s society and economy have developed rapidly,but the problem of air pollution is still serious.Many cities and even urban agglomerations are often and for a long time attacked by haze,especially in Beijing,Tianjin and Hebei and their surrounding areas.PM2.5 is the main component of haze and the main cause of aggravating haze weather pollution.The key of haze control is to control PM2.5concentration.It is of great significance for people’s life decision-making and the implementation of relevant government policies and measures to effectively master the spatio-temporal evolution of PM2.5 concentration and predict accurately.The goal of this paper is to achieve fine-grained prediction and inference of PM2.5concentration throughout the region,and the target is divided into site prediction task and grid inference task.In the prediction of PM2.5 concentration,most of the existing studies just use a single model,without extracting features from different levels,and seldom take into account the spatio-temporal evolution of PM2.5 concentration.In grid PM2.5 concentration inference,most of the previous studies use interpolation models,but due to various factors,the concentration values of different geographical locations show a non-linear relationship,moreover,the interpolation method is restricted by the spatial distribution of points.In order to solve the above problems,the research contents of this paper are divided into three aspects:(1)Spatio-temporal evolution analysis of PM2.5 concentration in Beijing,Tianjin and Hebei.Using hourly PM2.5 concentration monitoring data from 101 air quality monitoring stations in Beijing,Tianjin and Hebei over the years,this paper analyzed the evolution of PM2.5 concentration in Beijing,Tianjin and Hebei from time and space dimensions by using classical mathematical statistical method and the method based on spatial interpolation and gravity center shift for time and space dimension respectively.The analysis shows that PM2.5 concentration in Beijing,Tianjin and Hebei has a certain periodicity in the annual,seasonal,monthly,weekly and intra-day periods,and from2016 to 2018,the overall pollution level has decreased year by year,indicating that the three governments’s haze control policies and measures have achieved remarkable results.(2)Hourly PM2.5 concentration forecast for air quality monitoring stations in the next 24 hours.Starting from two dimensions of time and space,this paper establishes time predictor and space predictor based on LSTM.The former models the monitoring station’s own factors and concentration,while the latter is responsible for describing the internal relationship between the concentration of the station and the concentration of its surrounding stations under the influence of various factors,finally,the predictive output of the time predictor and the spatial predictor is fused dynamically according to the meteorological conditions of the station by a stacking integrated model based on regression tree,and the final prediction result of the station is obtained.Analysis shows that the prediction results after dynamic fusion of Stacking integrated model have better prediction accuracy in the whole target time series,and have higher practical value.(3)Hourly PM2.5 concentration inference for each 1 km~2 grid in the next 24 hours.In this paper,the Beijing-Tianjin-Hebei experimental area is divided into non-overlapping 1 km~2 grids.The meteorological features,road network features,POI features and spatial relationship features of each grids are obtained from multi-source data.The inference model based on BP full-connection network is obtained by training the features of the grids belonging to monitoring stations.Precision analysis shows that the inference results of the inference model in this paper fit well with the actual results,and have better practical reference value compared with the results of the interpolation model.