Hourly ozone prediction using ensemble forecasting to accurately capture peak concentration: a case study of Tehran, Iran.

Accurate short-term ozone (O3) forecasting is critical for mitigating respiratory and ecological impacts, protecting public health, and guiding urban air quality management. This study presents the Residual-Aware Meta Ensemble (RAME), a hybrid deep learning framework that integrates bidirectional recurrent units with dilated convolutions to deliver high-resolution hourly surface ozone predictions, with a particular focus on peak concentration events and regulatory exceedances. The model first generates foundational forecasts through an ensemble of base learners trained on 24-h historical pollutant data and real-time meteorological parameters. These predictions are then refined via a residual-correction mechanism, where a meta-learner is trained to compensate for the discrepancy between the weighted ensemble forecast and actual concentrations. Using observations from eight monitoring stations across Tehran, RAME achieved city-wide mean 24-h averaged R2 of 0.87, RMSE of 8.12 ppb, and MAE of 5.30 ppb, consistently outperforming all individual baselines across heterogeneous urban sites. The model accurately identified 96% of critical ozone days and demonstrated superior skill in capturing diurnal variability, particularly during photochemical peak periods when conventional systems tend to significantly underestimate concentrations. These findings highlight RAME as a robust, scalable, and methodologically adaptable solution for urban ozone forecasting, enabling timely interventions to protect public health and support regulatory compliance.