Physics-Informed Machine Learning for Urban Nitrogen Dioxide Forecasting in Palermo, Italy

Accurate forecasting of nitrogen dioxide (NO2) in high-traffic urban environments is a critical challenge for public health management and environmental policy. This study presents a hybrid physics-informed machine learning pipeline for NO2 dispersion modeling and short-term forecasting along Viale Regione Siciliana in Palermo, Italy, one of the highest traffic-density corridors in Europe, over six full years (2020–2025) of hourly data resolved at approximately 11 m. Station measurements and weather data are harmonized hourly over the OpenStreetMap road network, converted into emissions with COPERT-Italy (COmputer Program to calculate Emissions from Road Transport) factors, and dispersed with the SIRANE street-network model; the chain is also inverted to recover corridor traffic from the observed NO2. The resulting field is joined with the measured predictors in the machine learning stage, and the output is mapped in GIS Geographic Information System. The hourly scatter between observations and the SIRANE model built on 44,752 h is centered on the 1:1 line, with a Pearson correlation of r=0.83. Of two “memory-less” ensembles on an identical predictor set, XGBoost (eXtreme Gradient Boosting) returns the better R2, mean absolute error, and RMSE Root Mean Square Error at all twelve recursive lead times, declining from R2=0.79 at t+1 h to a plateau of 0.66 at t+12 h, compared to 0.76 to 0.59 for random forest; both reproduce the mean field almost exactly (r≥0.97), with the XGBoost residual staying between 1.0 and 1.7 μg m−3 at every horizon and that of random forest growing to 4.3 μg m−3 by t+12 h. The proposed pipeline offers a transferable methodology for urban air quality management in traffic-dominated environments.