Temporal forecasting of PM2.5, PM₁₀, NO₂ using deep learning models on Noida Sector 125 and Noida Sector 62

Air pollution has emerged as a critical environmental and public health issue in rapidly urbanising regions such as Noida, where pollutants including PM 2.5 , PM ₁₀ , and NO ₂ significantly deteriorate air quality and contribute to respiratory and cardiovascular diseases. Accurate forecasting of these pollutants is therefore essential for effective environmental monitoring and sustainable urban management. This study presents a comparative deep learning-based framework for multi-pollutant air quality forecasting at two major urban locations, Noida Sector 125 and Sector 62, representing institutional, commercial, and industrial environments with varying pollution characteristics. Time-series air quality datasets were collected from official monitoring agencies, including CPCB and SPCB, and subjected to preprocessing techniques such as missing value imputation, outlier handling, and Min-Max normalisation to ensure data reliability and consistency. Several deep learning architectures, including CNN, RNN, LSTM, GRU, and BiLSTM, were implemented to capture nonlinear temporal dependencies in pollutant concentration patterns. Furthermore, a hybrid BiLSTM-GRU model was proposed to improve forecasting performance by integrating bidirectional sequence learning with efficient gated recurrent mechanisms. The predictive performance of all models was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and coefficient of determination ( R² ). Experimental results demonstrated that the proposed Hybrid_BiLSTM_GRU model consistently outperformed the standalone models for all pollutants at both monitoring locations. At Noida Sector 62, the hybrid model achieved RMSE values of 41.01, 64.41, and 7.54 for PM 2.5 , PM ₁₀ , and NO ₂ , respectively, along with corresponding R² scores of 0.4846, 0.4280, and 0.6391. Similarly, at Noida Sector 125, the model obtained RMSE values of 63.37, 90.05, and 16.90, with R² values of 0.2155, 0.3665, and 0.4748 for PM 2.5 , PM ₁₀ , and NO ₂ , respectively. The hybrid framework also demonstrated significant percentage improvements over conventional deep learning models, achieving up to 43.08% reduction in MSE and 24.61% reduction in MAE for PM 2.5 prediction. Overall, the findings confirm that the proposed hybrid deep learning architecture effectively captures complex temporal pollution dynamics and provides reliable multi-pollutant forecasting for urban air quality management and environmental decision support systems.