Research On PM2.5 Spatiotemporal Network Prediction Model Integrating Multi-source Environmental Factors

Particulate matter ≤2.5μm(PM2.5)is a very important indicator in the assessment of environmental pollution and exposure.At present,PM2.5 ground monitoring stations basically realize the full coverage of urban space,and satellite remote sensing data has become an important means of PM2.5 monitoring in a wide range.At the same time,with the rapid development of deep learning and intelligent monitoring technology,the application of massive historical environmental data has become an important research content.As one of the applications of deep learning technology in the field of environmental governance,the prediction of environmental factors plays an important role in environmental monitoring,disaster warning and analysis,traffic control and other fields.Timely acquisition of time series data of environmental factors within the region is made use of deep learning technology for prediction,providing basis for resource allocation in traffic control and environmental governance decision-making.At present,most PM2.5 prediction methods are based on convolutional neural network,and the accuracy of prediction results has been significantly improved.However,there are still many problems in practice: traditional PM2.5 prediction methods fail to take into account time and space dependence,single data source,waste of data value and other problems.Starting from the above problems,this paper takes the deep learning model to predict PM2.5 as the main research object,and the research content is as follows:(1)This paper designs an analysis and research on the variation characteristics of PM2.5 in Beijing from the perspective of temporal variation characteristics(annual and seasonal),spatial distribution characteristics and correlation analysis.To understand the temporal and spatial distribution characteristics and influencing factors of PM2.5 in Beijing,Pearson correlation analysis was used to explore the correlation between PM2.5 and meteorological factors(humidity,temperature,atmospheric pressure,wind speed,rainfall(snow)amount)and other pollution factors(PM10,SO2,NO2,O3,CO).It provides the basis for subsequent collaborative forecasting.The results show that: in recent years,the annual concentration of PM2.5 shows a decreasing trend,and there is a significant seasonal,weekly and daily variation trend.Among them,the temporal variation law of the seasonal unit is obvious.Specifically,the concentration is low in summer,followed by autumn,and relatively high in spring and winter.Temperature and relative humidity were positively correlated with PM2.5 concentration.Atmospheric pressure,wind speed and rainfall(snow)are negatively correlated with PM2.5.It is positively correlated with PM10,SO2,NO2,CO,and negatively correlated with O3.Thanks to environmental control measures in recent years,PM2.5 concentration is negatively correlated with green coverage,population density,transportation passenger volume,GDP and motor vehicle ownership in the city(2)A PM2.5CNN-LSTM hybrid neural network prediction model considering climate factors is proposed.First,CNN network is used to abstract the characteristics of climate factors and seasonal factors,as an additional input in the prediction process,and it is analyzed collaboratively with LSTM network.The pollution data and weather data collected from monitoring stations in Beijing from 2010 to 2014(sampling interval is 1 hour)are used for experiments,and CNN-LSTM model is compared with other prediction models.The results show that: Compared with the LSTM model RMSE decreased by 10.71%,compared with the multi-source data fusion of the LSTM model RMSE decreased by 5.52%,the study shows that the multi-source data fusion of CNN-LSTM model proposed in this paper has better prediction ability.(3)This paper proposes a PM2.5 distance-weighted spatio-temporal graph convolution region PM2.5 prediction model,abstracts the monitoring stations within the region into a distance-weighted graph structure,introduces the spatio-temporal graph convolution model,uses time convolution to capture the time features coherentially,and uses the spatial convolution to capture the spatial features.Finally,the fully connected layer is used to synthesize the spatio-temporal features and generate PM2.5 prediction results.The PM2.5 concentration data of Beijing were used in the experiment and compared with K nearest neighbor model(KNN)and Gaussian process regression model(GPR).MAE increased by 16.59%,12.35% and 11.05% compared with GPR and KNN,respectively,in the 24 h prediction scale.RMSE increased 12.84 percent and 4.52 percent,respectively.Experimental results show that the model used in this paper can achieve better prediction accuracy.(4)Based on the prediction model,this paper introduces the technology into the monitoring and analysis of environmental factor changes in practical applications;An environmental factor data set was developed and the method was verified.The visual analysis of PM2.5 changes within the study area was carried out,and a system for real-time PM2.5 monitoring,display and prediction was designed.