G2RL: grid-aware gated recurrent learning for spatiotemporal urban air quality forecasting

Accurate urban air-quality forecasting is critical for addressing the health and environmental impacts of urbanization. Reliable urban air-quality prediction is challenging due to the complex nonlinear dynamics of pollutant concentrations and significant spatial variability influenced by local emissions and changing weather conditions. This paper presents grid-aware gated recurrent learning (G 2 RL), a framework that uses grid-based spatial representation and gated recurrent units (GRUs) for effective spatiotemporal forecasting of air pollutant interactions. Existing forecasting methods have notable limitations. Conventional recurrent models excel in temporal dependencies but struggle with localized spatial interactions. CNN-based approaches require structured spatial data, while graph-based methods rely on complex topologies that are hard to establish for irregular networks. Additionally, complex spatiotemporal architectures often demand substantial computational resources, limiting their scalability and real-time application in urban monitoring systems. G 2 RL overcomes limitations in current frameworks by effectively capturing localized spatial heterogeneity and temporal pollutant dynamics without manual graphs. The study aims to: (i) develop a learnable grid-based model for localized pollutant interactions; (ii) jointly learn spatial and temporal dependencies among pollutant concentrations and meteorological variables; and (iii) assess forecasting accuracy and robustness under realistic monitoring conditions, including missing data. G 2 RL outperforms traditional CNN-GRU and graph-based methods by using a learnable grid embedding to capture localized pollutant interactions. It combines pollutant concentrations and meteorological variables to forecast accurately with sparse sensor data. Tested on 34 New Delhi stations, G 2 RL achieved an RMSE of 14.2, MAE of 9.8, and an R 2 -score of 0.92, outperforming various models. Statistical tests confirm the significance of these improvements, indicating that the observed performance gains are not attributable to random variation. The framework maintains high performance amid sensor dropout, noise, and spatial issues, enabling real-time predictions and making it ideal for smart-city air-quality management.