Explainable One-Hour-Ahead Air Quality Index Forecasting Across Saudi Cities Using Pollutant and Meteorological Data

Air quality forecasting supports public-health protection, yet prior Saudi studies have mostly addressed single pollutants or single-city air quality index (AQI) classification. This study evaluates one-hour-ahead AQI forecasting across eight climatically diverse Saudi cities (Makkah, Riyadh, Jubail, Abha, Taif, Madinah, Tabuk, and Yanbu) using attention-based bidirectional recurrent networks, trained with and without meteorological inputs and compared against statistical and machine-learning baselines. Against a stringent persistence baseline that quantifies the predictability ceiling imposed by the strong autocorrelation of hourly AQI, the deep models matched it in RMSE and R2 (cross-city mean R2 ≈ 0.90) while lowering the mean absolute error by roughly a fifth, an advantage most valuable in the more variable cities where warnings matter most. In an event-based evaluation, the models flagged 84 to 91% of high-AQI threshold exceedances, including in the most volatile city, so the forecasts retain warning value even where the largest peaks are compressed. Surface meteorological inputs contributed little on average and their value was city-specific, and at Riyadh also architecture-dependent. SHAP and LIME confirmed the models rely on physically expected drivers, chiefly particulate matter, with feature rankings stable across cities, methods, and architectures. Evaluating one common framework across climatically diverse cities, we find that forecast skill varies more with local regime and input configuration than with model architecture, which offered no consistent advantage. Per-city input configuration within a common framework accordingly offers a practical basis for deployment.