Effective PM2.5 concentration forecasting based on multiple spatial–temporal GNN for areas without monitoring stations

With rapid industrial developments, air pollution has become a hot issue globally. Accurate prediction of PM2.5 (a category of particulate pollutant with a diameter of less than 2.5μm) has been a critical topic, as it can provide valuable information for government decision-making and policy control in environmental management affairs. In this paper, we propose a deep learning model based on graph neural networks (GNNs) to predict the next 48hr PM2.5 concentration in Taiwan. In this model, monitoring stations are regarded as nodes and edges are the distances between monitoring stations. Hence, the distribution of the stations can be perceived as a graph. GNNs are promising in processing non-grid structure data that can be represented as a graph. By incorporating the GNN and gated recurrent units (GRUs), this model can effectively capture the long-term spatial–temporal features in air quality time-series data. In addition, we also investigated the problem of predicting PM2.5 concentrations in the areas without monitoring stations or at sites far away from the stations. This problem has not captured researchers' attention whose methods are based on GNN. The problem is, however, quite challenging as these areas do not have historical air quality data, leading to low prediction quality. Finally, we performed experiments to verify the effectiveness of the proposed model based on actual data sources obtained in Taiwan. The results show that the proposed model exhibits satisfactory prediction performance compared to existing models.