Forecasting Air Pollution at Construction Sites: An Extended Comparison of Machine Learning Approaches

In this paper, we introduce a new approach designed to forecast up to 24 hours in advance five air pollutants at construction sites given the temperature as an exogenous variable. The considered pollutants are two particulate matter (PM2.5, PM10) and three gaseous pollutants (NO_2, CO and O_3). The proposed approach employs and compares three well-known predictive model types on forecasting each air pollutant: random forest, feed-forward artificial neural network and transformer architecture. Their performances are evaluated in terms of Root Mean Squared Error on the single pollutants with single-output models. Experimental results, based on data collected from AirQino sensor stations at a real construction site, demonstrate that random forest continues to be a reliable and frequently top-performing choice for long-term forecasting in this context. However, the transformer architecture emerges as a strong alternative, especially well-suited for short-term prediction tasks and scenarios with limited data availability.