Regime-aware hybrid ensemble learning for adaptive [Formula: see text] forecasting in urban environments.

Accurate forecasting of fine particulate matter PM2.5 is critical for mitigating adverse health and environmental impacts, yet existing models often fail to adapt to dynamic meteorological regimes. This study introduces a novel hybrid ensemble framework that adaptively switches between CatBoost and TabNet based on meteorological dispersion parameters, enabling context-specific model dominance. The approach was evaluated using the OpenAQ dataset, comprising hourly PM2.5 concentrations and corresponding meteorological variables across Delhi, India. The adaptive switching mechanism, driven by ventilation coefficient and relative humidity thresholds, resulted in CatBoost dominance for 58% of days, TabNet dominance for 35%, and balanced contributions for the remainder. Experimental results demonstrated that the proposed ensemble achieved an RMSE of 15.54, MAE of 11.02, MAPE of 12.2%, and [Formula: see text] of 0.86, outperforming established baselines such as LSTM, XGBoost, and LightGBM, with statistically significant improvements ([Formula: see text]). Efficiency analysis revealed only a 14% increase in inference time compared to CatBoost alone, while offering substantial gains in predictive accuracy. Hyperparameter sensitivity analysis further identified an optimal configuration (learning rate = 0.01, batch size = 64) balancing convergence stability and generalization. By integrating model interpretability, adaptive learning, and computational efficiency, this research presents a scalable forecasting solution for urban air quality management. The methodology's design allows seamless extension to other pollutants, geographic regions, and ensemble combinations, positioning it as a practical tool for data-driven environmental policy and public health protection.