A Robust and Scalable Deep Learning Framework for Short- and Long-Horizon Air Quality Forecasting and AQI Classification

Accurate and reliable air quality forecasting is essential for environmental monitoring, public health planning, and real-time decision support systems. This study presents a comprehensive evaluation of multiple machine learning and deep learning models for short-term (1-hour) and long-term (24-hour) air pollutant forecasting and Air Quality Index (AQI) category classification. Performance is assessed across key pollutants, including PM₂.₅, PM₁₀, and NO₂, using regression metrics such as RMSE, MAE, MAPE, and R², alongside categorical evaluation through confusion matrices and class-wise F1-scores. In addition to predictive accuracy, the study systematically examines computational efficiency, scalability under concurrency, and robustness under real-world stress scenarios including data missingness, sensor drift, noise injection, and seasonal shifts. A hybrid CNN–LSTM model demonstrates superior robustness and generalization, maintaining RMSE degradation within the predefined 10% tolerance threshold under most stress conditions, while also achieving strong AQI classification performance across all categories. Ablation studies further highlight the critical contributions of spatial modeling, attention mechanisms, and exogenous features to forecasting accuracy. Statistical validation using ANOVA, DMRT, and effect size analysis confirms that observed performance improvements are statistically significant with moderate to large effects.