Fine-grained PM[formula omitted] prediction in Lanzhou based on the spatiotemporal graph convolutional network

Abstract

Urban fine-scale PM2.5 concentrations can be predicted that considers both spatial and temporal correlation, owing to the widespread installation of air quality micro-monitoring stations in cities. The traditional method only uses the data of a single monitoring station while ignoring the spatial diffusion characteristics of the influence between stations. Recently, urban PM2.5 predictions have often been insufficiently refined due to coarse-grained air quality data and poor consideration of spatial dependence in modelling. Here, the influence relationship of PM2.5 flow between irregular monitoring stations is abstracted into an unstructured correlation graph, combined with the PMPre-GCN model to forecast the PM2.5 concentration of the city’s intensive monitoring stations. The validity of this method was verified by using air quality monitoring data from 456 dense micro-stations in Lanzhou City as an example. The results illustrate that: (1) The construction of the correlation graph significantly improved the model’s prediction accuracy. Compared with the fully connected and distance-based graphs, the prediction results’ root means square error (RMSE) was reduced by 10.09% and 8.42%, respectively. (2) The density of monitoring sites makes the precision of fine-grained prediction higher. Using data from 456 stations resulted in a 36.31% reduction in RMSE compared to the selection of data from 228 stations to make inferences about air quality in the city. (3) PMPre-GCN showed the best prediction performance when predicting PM2.5 concentration from T+1 to T+8 h at all stations. Especially at T+8 h, RMSE and the mean absolute error are 11.83 μg/m3 and 8.21 μg/m3. Overall, the method is capable of fine-grained predictions of urban PM2.5 concentrations, thereby supporting resident health management and fine-grained government decision-making on the urban air environment.