Air pollution ( \(\text {PM}_{2.5}\) ) can be considered a critical environmental and public health challenge. This work presents a hybrid deep ensemble framework for forecasting daily \(\text {PM}_{2.5}\) concentrations across four major urban cities of Uttar Pradesh (Agra, Kanpur, Noida, and Varanasi) using data from Jan 2018 to Aug 2025. The methodology integrates advanced feature engineering, including spectral decomposition via Fast Fourier Transform, lag-based temporal variables, and statistical descriptors, with multiple deep learning models (LSTM, Bi-LSTM, GRU, Bi-GRU, RNN, and CNN). To capture complicated nonlinear and temporal dependencies in the data, these base learners are stacked in an ensemble with XGBoost as a meta-learner. The designed framework consistently beats individual deep learning models in every city, according to research findings. In terms of MAE, the ensemble reduces prediction errors to as low as 3.64–5.35, compared to baseline values exceeding 11–20. Similarly, the MSE is reduced by nearly an order of magnitude (e.g., 37.82 in Agra versus 362–428 for conventional models), while the RMSE is lowered to 4.80–7.31 compared to 18–30 from baselines. Furthermore, the \(R^2\) values approach 0.98–0.99, significantly surpassing the 0.69–0.85 range of traditional models. These results demonstrate the suggested ensemble’s accuracy, generalisation potential, and resilience, making it a potent forecasting tool for \(\text {PM}_{2.5}\) dynamics in heavily polluted urban settings.
