Machine Learning Approaches for Ozone Forecasting in Urban and Rural Areas of Greece: A Comparative Study Using an IoT Monitoring Network

Ground-level ozone (O3) is a secondary photochemical pollutant whose formation under intense Mediterranean solar radiation makes Greece prone to elevated concentrations, with well-documented respiratory and cardiovascular health effects. Forecasting of ozone in Greece has nonetheless remained limited: prior work has been largely statistical, focused on one or two sites in Athens, and restricted to a single horizon, and no study has compared machine learning forecasts across the urban-to-rural gradient of the national monitoring network. This study addresses that gap with, to our knowledge, the first multi-site comparison of machine learning ozone forecasts spanning the full urban-to-rural typology gradient of a Greek monitoring network, introducing a typology × model accuracy cross-matrix that links site character directly to forecasting performance and offers concrete, horizon-specific operational guidance for early warning deployment. We develop a single, reproducible pipeline and compare three model families, Support Vector Regression (SVR), Random Forest (RF) and gradient boosting (GBM), against a naïve persistence baseline, for next-hour and next-day-maximum O3 prediction at eight urban, suburban and rural stations of the National Air Pollution Monitoring Network (2021–2024). All predictors are derived from the ozone series itself (lagged, rolling and cyclical calendar features); models were tuned and evaluated using a strictly chronological 70/15/15 train/validation/test split per station, with hyperparameters selected on the validation partition and the test partition evaluated only once to prevent temporal leakage. At the next-hour horizon, all models outperformed persistence at every site, with the single exception of SVR at the rural THR site; gradient boosting was best or joint-best (test R2 from 0.83 at the heavily titrated central-urban site to 0.94 at a rural site), and a Wilcoxon signed-rank test on paired per-sample errors confirmed that GBM significantly outperformed Random Forest at all seven evaluated stations (p<0.05, and p<0.001 at six of them). The next-day-maximum task was substantially harder (R2 between 0.41 and 0.79), and the ranking reversed, with SVR being the best model almost everywhere. Crucially, forecast accuracy declined systematically from rural and suburban toward central-urban sites: the same nitric oxide titration that suppresses urban ozone is also associated with degraded predictability. These results, summarised in the typology × model accuracy cross-matrix, indicate that an operational early warning system for Greece should adopt horizon-specific models, gradient boosting for next-hour nowcasting and SVR for day-ahead maxima, with explicitly higher uncertainty in the urban core, and can be deployed on the network’s real-time (IoT) data stream using ozone observations alone.