Author(s)

SHEEN MCLEAN CABANEROS, CARROW MORRIS-WILTSHIRE, PHILIP JAMES

Abstract

Urban air pollution forecasting remains challenging due to non-stationary spatial interactions driven by time-varying atmospheric transport and heterogeneous local emission processes. Practical constraints on sensor deployment also result in incomplete spatial coverage and uneven data availability across monitoring networks. These limitations highlight the need for resilient and uncertainty-aware forecasting tools, particularly at sparsely monitored locations. Recent spatio-temporal deep learning models have shown strong predictive performance yet often combine heterogeneous spatial dependencies and offer limited insight into predictive uncertainty. This study examines short- and long-term forecasting of hourly NO₂ concentration levels using data from 19 monitoring stations across Newcastle upon Tyne, UK, collected between 2023 and 2024. Lagged pollutant concentration levels and meteorological variables are used as model inputs. A structured local–global spatio-temporal graph neural network model that separately learns stable local interactions from dynamic long-range dependencies combined with a probabilistic framework is employed. Experiments further assess robustness under reduced monitoring coverage and spatial extrapolation to unseen sites. Overall, the proposed model outperforms benchmark models across forecasting horizons, achieving about 84% index of agreement for 1-hour-ahead prediction under standard training, while retaining stronger performance at longer forecasting horizons. Under sparse monitoring, performance declines only modestly with the smallest root mean square error (RMSE) increase (9.75%) among benchmark models. Under the site-holdout setting, the model generally achieves the lowest RMSE at unseen stations, although errors remain higher at spatially isolated locations. Uncertainty quantification results show that the 95% prediction intervals remain well calibrated, with empirical coverage slightly above 0.96 under sparse monitoring scenarios, while interval widths increase in more challenging settings. Overall, the findings show that the proposed forecasting framework can improve the reliability of urban NO₂ prediction under realistic spatial sparsity and monitoring limitations.

Keywords

air pollution, deep learning, uncertainty quantification, air pollution forecasting