Fine particulate matter concentration prediction based on hybrid convolutional network with aggregated local and global spatiotemporal information: A case study in Beijing and Chongqing

Air pollution is a highly concerned environmental issue that have serious impacts on human health and the ecological environment. Accurate air quality prediction can help people effectively deal with the threats posed by air pollution. Most previous work mainly focused on temporal modeling of air quality data at monitoring stations. In recent years, a number of works have used graph convolution to model the spatial dependencies between neighboring sites to extract spatial features and combined temporal model to extract temporal features for predicting PM2.5. However, these models only considered the local spatial relationships of adjacent sites and ignored sites with longer geographic distance. Therefore, this study proposes a spatio-temporal hybrid model based on convolution and attention (named Attentive Graph Convolution and 1D Convolution Network, AGCC) to predict multi-site and multi-step PM2.5 concentration. This method not only models the local spatial relationships of adjacent sites, but also combines spatial attention and graph convolutional network (GCN) to model the global spatial relationships of sites. The local and global spatial features are obtained respectively and are fused to obtain spatial features with richer semantic information. Meanwhile, the temporal dependency relationship is modeled through a temporal module composed of temporal attention and one-dimensional convolutional neural network (Conv 1D). AGCC was compared to six baseline models at different prediction horizon based on the air quality datasets of Beijing and Chongqing. The experimental results demonstrate that the model proposed by our achieves the best performance and verifies the feasibility of the model.