Hybrid Deep Learning Models for Multivariate Time-Series Forecasting of Air Quality Index toward Urban Sustainability and Public Health

Escalating air pollution and the limitations of conventional forecasting approaches necessitate advanced models for accurate Air Quality Index (AQI) prediction. This study proposes four hybrid deep learning architectures—Model-1: Hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM), Model-2: Hybrid LSTM– Deep Neural Network (DNN), Model-3: Hybrid GRU–CNN, and Model-4: Hybrid Conv1D–DNN, that integrate convolutional feature extraction, recurrent temporal modeling, dense representation learning, and attention mechanisms to capture complex temporal dependencies and nonlinear interactions among multiple pollutant and meteorological variables. The four-year daily data from Delhi, consisting of major pollutants such as PM 2.5 , PM 10 , NO 2 , NOx, CO, benzene, and meteorological variables like temperature, relative humidity, and atmospheric pressure, is employed. Optimized models have been developed by considering different loss functions and learning rates. Low prediction error (RMSE ≈ 36.29) and high correlation (R ≈ 0.9) have been achieved. Model-1 has shown superior performance compared to other models. Significant reduction in error has been observed. The performance of the model has been tested for different seasons, showing robustness in high variability during winter. The performance of the model has also been tested on different city data, such as Ghaziabad and Faridabad. Although the accuracy of the model can be increased by considering more complex models, the results have shown a good trade-off in accuracy and complexity, thus establishing the robustness of the framework. Moreover, accurate AQI forecasting can support public health by enabling timely air quality warnings and informed environmental decision-making.