Air Quality Prediction Based On Intensive Micro-enviromental Monitoring Stations

Grid,fine and diversified air quality prediction is of great significance for improving the government’s precise control and innovating management of environmental pollution.However,due to the sparsity of the deployment of air quality monitoring stations,the current air quality feature analysis and prediction are mainly regional large-scale spatial studies,which primarily focus on single pollutant such as PM_2.5,and there are few feature analysis and prediction for small-scale,intensive and multi-pollutants.In recent years,in order to promote the refined management of air pollutants,Lanzhou Municipal Government has deployed a large number of air monitoring stations,forming intensive station scenes.These stations provide data support for the accurate prediction of air quality in small-scale areas.In this research,typical street areas in Lanzhou are selected as the research object,and the data collected from microenvironmental quality monitoring stations are used to carry out the research on multi-site and multi-pollutants prediction model under intensive site scenarios.Specific researches are as follows:1.Study on the temporal and spatial relationships of urban pollutants and their influencing factors.In order to carry out the multi-task predictive analysis under the background of intensive stations,it is necessary to input the information of strongly correlated stations.Therefore,the correlation between stations and pollutants is firstly analyzed.Nonlinear Granger causality test was used to analyze the nonlinear relationship of PM_2.5concentration series between stations,and it was found that there was a spatiotemporal correlation between stations.By using the generalized additive model,we find that the cross-action of SO₂and NO₂has a significant effect on the PM_2.5concentration.According to the experimental results,the site data can be used as the input feature of the multi-task model,so as to better mine the dynamic spatiotemporal relationship between the sites and improve the prediction performance of the model.2.Multi-site PM_2.5concentration prediction based on multi-task learning.The PM_2.5concentration prediction model MTL-Res Net-LSTM（MRL）was constructed by designing a shared encoder module and a specific time-series feature decoder module.Different feature selectors were selected to verify the validity of the results of the nonlinear Granger test,and the effects of different task numbers on the model performance were compared.The results showed that the MAE and RMSE of MRL model were lower than that of the single-task model.Meanwhile,the average MAE and RMSE at 9 sites were increased by 31.8%and 26.2%than LSTM,respectively.Compared with the two random input results,the non-linear Granger causality test provided reliable guidance for MRL to select characteristic information from the sites,and the multi-task model performed best when six tasks were predicted at the same time.3.Research on the joint prediction model of multi-site and multi-pollutants.The MTL-Station-Pollutant（MSP）model was constructed by designing shared encoders,We predict the three highly correlated pollutants of PM_2.5,NO₂,and SO₂at the same time on six sites.The results show that the MSP model has better overall prediction performance than the multi-task model without the attention mechanism,which verify the effect of the attention mechanism in the predictive model MSP、MRL.Compared with single-task models,MSP and MRL models have lower MAE and RMSE in prediction tasks than single-task models.which proves that the multi-task model can achieve good information sharing and is suitable for air quality prediction in dense site scenarios.