Interpretable spatio-temporal graph neural networks for air quality prediction under real-world multi-station monitoring data

Air quality forecasting has a crucial place in environmental management, public health safety, and sustainability of urban areas. The dynamic nature of air pollution leads to both spatial relationships between monitoring Stations as well as nonlinear relationships between meteorological and air pollutant Factors at specific times, which complicates accurate prediction. In this research, we present an interpretable Spatio-Temporal Graph Neural Network (CIST-GNN), which utilizes Graph Attention Networks (GATs) to capture spatial dependencies from the air pollution monitoring stations, Gated Recurrent Units (GRUs) to learn temporal patterns in the features of the Time Series Data, and a feature attention mechanism that improves the interpretation of feature contributions. The effectiveness of the proposed framework is evaluated using the actual dataset (i.e., the UCI multi-site air quality measurements from Beijing), consisting of 420,768 observations collected from 12 separate monitoring sites for the years 2013 through 2017. Experimental evaluation demonstrates that the CIST-GNN achieves an R 2 score of 93.23%, RMSE of 21.66 μg/m 3 , and MAE of 12.08 μg/m 3 on the test set. The model uses pollutant concentrations and meteorological variables, including PM2.5, PM10, NO2, CO, SO2, O3, temperature, humidity and wind speed information. In addition, the feature attention mechanism highlights which environmental variables provide interpretable feature importance to the model’s predictions. The proposed framework has additional advantages over GRU (which uses only temporal information), such as dealing with spatial dependence, providing feature attribution, and being robust to different types of spatial monitoring networks. These results demonstrate that the proposed CIST-GNN provides an effective and interpretable framework for air quality forecasting.