The escalating concentration of atmospheric pollutants in India poses severe complications for public health and daily life. This paper presents a comprehensive system integrating real-time Air Quality Index (AQI) monitoring from 453 stations spanning 28 states and 3 union territories of India with advanced PM \(_{2.5}\) forecasting models. Using 13 years of historical air quality data (2010–2023), we conducted a rigorous comparative evaluation of ensemble machine learning algorithms (XGBoost, LightGBM, CatBoost, Decision Tree, BART) and deep learning architectures (Vanilla LSTM, Bi-LSTM, Stacked LSTM, Stacked Bi-LSTM, CNN-LSTM) trained on the complete multi-city dataset. The Bidirectional Stacked LSTM model achieved superior performance with mean absolute error (MAE) of 0.062, root mean squared error (RMSE) of 59, and coefficient of determination (R \(^{2}\) ) of 0.55 for Delhi, outperforming ensemble methods by 25–40% and simpler LSTM variants by 15–20%. Feature engineering substantially enhanced accuracy: incorporating nitrogen oxide (NOx) pollutants improved MAE by approximately 15%, while 2-year lagged PM \(_{2.5}\) values captured crucial seasonal patterns. Multi-step forecasting maintained robust accuracy up to 6 hours (MAE: 0.17), demonstrating practical utility for early warning systems. Cross-regional validation across nine cities representing severely polluted Indo-Gangetic Plain regions, moderately polluted inland cities, and cleaner coastal and southern cities confirmed model generalizability, with MAE ranging from 0.062 in highly polluted areas to 0.031 in cleaner regions. These findings demonstrate that bidirectional deep learning architectures with strategic feature engineering provide reliable short-term air quality forecasts across India’s diverse urban environments, with practical deployment through a mobile application serving public health decision support.
