GraPhy: Graph-Based Physics-Guided Urban Air Quality Modeling for Monitoring-Constrained Regions

Shangjie Du, Zhizhang Hu, Shijia Pan

Abstract: This work presents GraPhy, a graph-based physics-guided learning scheme for accurate and high-resolution air quality modeling for urban regions with extremely constrained monitoring data. Finegrained air quality information is essential for mitigating public exposure to pollutants. However, monitoring data is often rare for socioeconomically disadvantaged regions, which directly limits the accuracy and resolution of air quality modeling. We tackle this problem by using physics-guided graph neural networks with neural network layers and graph edge features customized for low-resolution monitoring data. We conduct experiments with data from California's socioeconomically disadvantaged San Joaquin Valley. GraPhy achieves the lowest mean absolute error (MAE) with a 15.1%-33.6% absolute error reduction compared to baselines.