Station-specific variability and short-term forecasting of a particulate air-quality index using machine learning in Charleroi, Belgium

Air-quality management in urban–industrial environments require reliable assessment of particulate pollution and realistic evaluation of forecasting performance. This study examined temporal variability and short-term predictability of a particulate air-quality index at Marchienne-au-Pont, Charleroi and Lodelinsart in Belgium. The analysis included 206,184 half-hourly observations collected during 2020–2023. Quality control identified a probable inconsistency in the recorded PM 2.5 and PM₁₀ fields, which was addressed through a reproducible provisional reconstruction. k-nearest neighbours, Random Forest, Decision Tree, Gradient Boosting, AdaBoost and a Neural Network—were evaluated using expanding-window validation and an independent 2023 test period. The four-year mean particulate index was highest at Marchienne-au-Pont (35.13), followed by Lodelinsart (31.62) and Charleroi (28.60). PM 2.5 determined the index in 67.63–74.84% of observations, whereas PM 10 determined 25.16–32.37%. Tree-based models produced contemporaneous testing R 2 values above 0.999 because the target was mathematically derived from the simultaneous particulate predictors. Separate future-forecast experiments therefore provided the principal operational assessment. One-hour forecasts achieved R 2 values of 0.857–0.920, decreasing to 0.551–0.641 at 6 h and 0.270–0.335 at 24 h. Compared with persistence, Gradient Boosting reduced RMSE by 5.80–6.29% at 1 h, 15.31–17.19% at 6 h and 12.41–12.96% at 24 h. These findings demonstrate the importance of data-quality verification, chronological validation and explicit separation of contemporaneous index estimation from genuine future forecasting. Because monitoring-agency confirmation of pollutant identity and processing history was unavailable, the reconstructed concentrations and derived results remain provisional.