Time series forecasting is essential in do-mains such as environmental monitoring, particularly for predicting air pollutant levels like PM \(_{2.5}\) . This study introduces a deep ensemble framework, termed Feature Engineered LSTM-BiLSTM-GRU-RNN-CNN (FE-LBGRC), that integrates multiple deep learning architectures, including Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN), and Convolutional Neural Network (CNN). Advanced feature engineering techniques, such as rolling window statistics and Savitzky-Golay smoothing, are applied to enhance input quality. The ensemble leverages XGBoost as a meta-learner to combine individual model predictions, improving accuracy and robustness. The framework is evaluated using PM \(_{2.5}\) data collected from 20 monitoring stations across major Indian cities (Bengaluru, Chennai, Mumbai, Pune, Lucknow, Delhi, and Hyderabad) between 2016 and January 2025. The proposed FE-LBGRC framework achieved the maximum predictive accuracy among all examined models, yielding an MAE of 2.87, MSE of 23.23, RMSE of 4.42, and an R \(^2\) of 0.9906. These results highlight the superiority of the proposed ensemble architecture in understanding complex temporal and spatial patterns, demonstrating that the integration of deep learning models with feature engineering and ensemble learning leads to more accurate and generalized PM \(_{2.5}\) forecasting.
