Nowcasting and forecasting urban air pollution with meteorological and wind profile data: A framework for Smart City Digital Twins

This study presents a framework for nowcasting and forecasting urban air pollution using meteorological and wind profile data. It is designed to support urban air-quality management and to monitor the durability of the built environment within smart city digital twins. The proposed framework employs the Series-cOre Fused Time Series forecaster (SOFTS) as the primary predictive model. Its performance is evaluated using data from Lamezia Terme, Italy, under both fully informed and missing/reconstructed data scenarios. For benchmarking purposes, two widely used machine-learning approaches — extreme gradient boosting (XGBoost) and long short-term memory (LSTM) — are also considered. Numerical results demonstrate the accuracy and robustness of the framework. In real-time nowcasting scenarios at Lamezia, the SOFTS model achieved coefficients of determination (R2) exceeding 0.91 and weighted mean absolute percentage errors (WMAPE) below 21% for all target pollutants, outperforming the XGBoost benchmark. For short-term forecasting up to 6 h, the model maintained strong predictive skill, particularly for particulate matter, with R2 values above 0.70 at the sixth hour. Results from the independent Cabauw site in the Netherlands further support the generalizability of the framework, with R2>0.97 for nowcasting and R2>0.62 at the 6-hour forecasting horizon. Model interpretability is also examined, and the results are discussed in relation to the existing literature. Overall, the framework advances short-term urban air-pollution prediction for smart city digital twin applications.