Forecasting urban air quality is important for protecting public health, but current model forecasts are often limited by an inaccurate prescription of pollutant emissions from human activities. We developed a new approach that improves air quality forecasts by adjusting emission prescription based on observed concentrations in urban agglomerations for key pollutants such as nitrogen oxides, sulfur dioxide, carbon monoxide, particulate matter, and volatile organic compounds. Applying this new approach to the São Paulo metropolitan area, Brazil, we compared forecasted and observed pollutant concentrations (from 6 February to 17 April 2023). Using adjusted emission significantly improved air quality forecasts for São Paulo, especially for ozone levels after adjusting estimates of volatile organic compound emissions. However, the forecast of particulate matter concentrations remained challenging due to their links with gaseous pollutants. Our study demonstrates the potential of using observed concentrations in urban agglomerations to improve air quality forecasts. Extending this approach to other urban agglomerations can help refine emission estimates and improve regional air quality forecasts, enabling better decision making for health protection.

An intercomparison of four air quality models is performed in the tropical megacity of Sao Paulo with the perspective of developing an air quality forecasting system based on a regional model ensemble. During three contrasting periods marked by different types of pollution events, we analyze the concentrations of the main regulated pollutants (Ozone, CO, SO2, NOx, PM2.5 and PM10) compared to observations of a dense air quality monitoring network. The modeled concentrations of CO, PM and NOx are in good agreement with the observations for the temporal variability and the range of variation. However, the transport of pollutants due to biomass burning pollution events can strongly affect the air quality in the metropolitan area of Sao Paulo with increases of CO, PM2.5 and PM10, and is associated with an important inter-model variability. Our results show that each model has periods and pollutants for which it has the best agreement. The observed day-to-day variability of ozone concentration is well reproduced by the models, as well as the average diurnal cycle in terms of timing. Overall the performance for ozone of the median of the regional model ensemble is the best in terms of time and magnitude because it takes advantage of the capabilities of each model. Therefore, an ensemble prediction of regional models is promising for an operational air quality forecasting system for the megacity of Sao Paulo.