AQIStack: An Interpretable Ensemble Learning Framework for Reliable Air Quality Index Prediction

Abstract

Air pollution in Delhi has reached dangerous levels, necessitating accurate and dependable Air Quality Index (AQI) forecasts for timely intervention. However, traditional monitoring methods and stand-alone machine learning algorithms frequently fail to capture the complex, nonlinear relationships between contaminants and temporal aspects. To address these limitations, AQIStack a hybrid stacking-based ensemble model was developed. AQIStack is an integration of XGBoost, CatBoost, Gradient Boosting, Random Forest, and Extra Trees with Bayesian hyperparameter optimization as base model and Linear Regression as meta-learner uses base predictions to improve generalization and predictive capability. The proposed AQIStack model outperformed the baseline models with $\mathbf{R}^{\mathbf{2}}$ of 0.9490, MAE of 17.7134, and RMSE of 25.6117. In addition, the interpretability was ensured by LIME prediction transparency and stability. The proposed system offers an accurate, explainable, and scalable approach for realtime AQI predictions, assisting officials with active pollution control and public health protection. AQIStack could also assist municipal authorities in framing evidence-based air quality management strategies and timely intervention policies.