The correct prediction of Air Quality Index (AQI) is essential in the management of the health of the population and making decisions about the environment in a timely manner. Yet, most of the existing models are very sensitive to raw pollutant characteristics, and are restrictive to the interpretability and strength of prediction. This paper presents a feature-engineering-based ensemble model enabling the integration of statistical feature selection with domain sensitive hybrid features, such as pollutant ratios, interaction products, and multi-pollutant dependency factors. The first step was to filter out redundant features through correlation filtering followed by the construction of hybrid features based on predictive relationships of atmospheric chemistry variables between PM \(_{2.5}\) , PM \(_{10}\) , NO \(_2\) , NO, CO, SO \(_2\) , and volatile organic compounds. It was evaluated with a series of supervised machine learning models, which included SVM, KNN, Logistic Regression, Random Forest, Gradient Boosting, Extra Trees, LightGBM, XGBoost, and CatBoost. Among them, the proposed Voting Ensemble (RF + XGBoost + CatBoost) exhibited the best performance with the accuracy of 97.53, precision of 97.50, and F1-score of 97.48, which is superior to such strong baselines as XGBoost (94%), Random Forest (93%), and LightGBM (91%). The stability and generalizability of the model in the case of different levels of pollution were also supported by the regression metrics, namely RMSE = 2.47, MAE = 1.98, and R \(^2\) = 0.9929. The results highlight that hybrid feature engineering has a considerable enhancement of the representation of pollutant-interaction, a stronger interpretability of the models, and a lower uncertainty in prediction. The suggested ensemble framework provides a scalable, data-efficient, and domain-congruent framework to be applied in real-time AQI forecasting applications, policy evaluation, and urban air-quality monitoring systems.
