AI-Driven Air Pollution Forecasting and Early Warning Systems
This project forecasts ozone hours to days ahead for densely populated cities, so air quality authorities can warn people before a pollution episode arrives.
Olmo Zavala-Romero Pedro Segura Pablo Camacho Jorge Zavala-HidalgoRosario Romero-Centeno
Goals
Improve operational air quality forecasting in densely populated urban areas so government authorities can issue timely alerts and reduce exposure of vulnerable populations during pollution episodes. This research line focuses on machine learning frameworks that integrate historical monitoring-network data, regional meteorological forecasts, and real-time atmospheric observations to predict ozone and related pollutants hours to days in advance.
Methods
We develop neural-network forecasting systems trained on long-term governmental air quality and atmospheric monitoring records, combined with meteorological reanalysis and forecast fields from regional Weather Research and Forecasting (WRF) implementations. Input features include cyclical time encodings (day of week, hour, day of year), data augmentation strategies, and multi-station architectures for simultaneous predictions across monitoring networks. Trained models ingest real-time station data and regional weather forecasts to produce 1–24 hour ozone forecasts, evaluated against held-out observations using mean absolute error and index of agreement metrics. Systems are deployed operationally through web interfaces for public and regulatory use.
Key findings
The 2025 Mexico City operational ozone forecasting system (Atmospheric Environment; doi:10.1016/j.atmosenv.2024.121017) achieved a mean absolute error of 9.81 ppb and an index of agreement of 0.91 across all stations; for the top 10 stations, MAE fell below 8.7 ppb and agreement exceeded 0.93. Integrating WRF meteorological forecasts and cyclical time features improved prediction skill over baselines, and performance is comparable to ML forecasting systems in other major metropolitan areas while representing an improvement over prior Mexico City operational systems. The pilot system is live at https://aire.atmosfera.unam.mx/.