| Air pollution has a major impact on human beings and the ecological environment on which they depend.Among them,PM2.5is the main component of haze,endangering human health,and increasing the incidence and mortality of cardiovascular,cerebrovascular and respiratory diseases.The timely and accurate prediction of PM2.5concentration will help scientifically prevent and effectively reduce the losses caused by haze incident.With the emergence of a large number of observational data in recent years,PM2.5concentration modeling and prediction based on deep learning has gradually become a research hotspot.Based on multi-source spatiotemporal data such as PM2.5,meteorology,land use types,and considering the sparsity and non-European distribution characteristics of environmental monitoring sites,as well as the nonlinear spatiotemporal characteristics of PM2.5and its impact factors,this study builds a deep learning model（GNN-GRUA）to simulate and predict PM2.5concentration.On the basis of using graph neural network（GNN）to capture spatial interaction and gated recurrent unit network（GRU）to capture time correlation,integrating attention mechanism to capture global information improves the modeling ability of time series information and further improves model prediction accuracy.In order to evaluate the effectiveness of the proposed model,this study collected four years of observational data on PM2.5concentration,meteorology,land use,etc.in291 cities,and performed data preprocessing,selection of characteristic factors,and data set division to transform the geographic environment observation data of the monitoring site into the input data characteristics required by the deep learning method.In the 3-hour prediction task in the study area of 291 city nodes,the probability of detection of GNN-GRUA on each data set reached more than 87%,the root mean square error remained between 6.0-12.7 on each data set,and false alarms The rate is below 8.6%,and all indicators are better than the GNN model that originally did not use the attention mechanism,which reflects the superiority of the attention mechanism for time-dependent information capture.In the 72-hour prediction task,GNN-GRUA showed better prediction results on each data set,and the probability of detection on data set 3 reached more than 79.4%,and the probability of detection of the GNN-GRUA model using land use type data reaching more than 80%,it is about 2%higher than the GNN model,and a considerable performance improvement has been achieved on the basis of the improved model,indicating that land use type data is beneficial for PM2.5concentration prediction. |
