A Hybrid Ensemble Learning Framework for Accurate and Reliable Air Quality Index Forecasting

Abstract

Forecasting of the Air Quality Index (AQI) is an important task due to nonlinear interaction among pollutants in the atmosphere and time changes. In this case, the most commonly used techniques depend on using sequential deep learning models such as Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Although these techniques prove to have the ability to model sequential data, their limitations include high sensitivity to parameters and inability to work properly when the data is heterogeneous. To overcome those limitations, we propose to use a more advanced forecasting framework by combining multiple machine learning models, such as HistGradient Boosting, Gradient Boosting, Extra Trees, Random Forest, and K-Nearest Neighbors. Stacking and weighted blending were applied for modeling different types of predictive patterns. Also, conformal prediction was used to get uncertainty estimation. The results show a high effectiveness of the proposed approach as demonstrated in the metrics, R² score = 0.9696, RMSE = 20.521, MAE = 14.559, and MAPE = 7.51%. The metrics for agreement and reliability, including WI (0.9921), KGE (0.9638), and NSE (0.9696), also validate the stable nature of the model performance. The developed method offers better prediction reliability, smaller errors, and accurate estimation of uncertainties, which makes it applicable for AQI predictions.