Evaluating machine learning models and imputation strategies for Air Quality Index forecasting in urban India

Accurate Air Quality Index (AQI) prediction is essential for timely health risk management in urban environments, yet challenges such as missing data and complex pollutant interactions limit the performance of traditional approaches. This study investigates AQI prediction for three years from six stations in Chennai, a South Indian coastal city, by coupling fourteen imputation techniques with five machine learning (ML) models to identify the most effective framework. Among the tested combinations, the Multi-Layer Perceptron (MLP) model with k-Nearest Neighbor imputation (kNNI-MLP) achieved the best performance, with a coefficient of determination of 0.9999, a root mean squared error of 0.4920, a mean absolute error of 0.2723, a symmetric mean absolute percentage error of 0.4522%, and a mean absolute scaled error of 0.0069%. Residual and calibration analyses confirmed unbiased and well-calibrated predictions, while trend analysis showed strong alignment between actual and predicted AQI values. Seasonal evaluation revealed consistent fluctuations, with AQI peaking in winter and post-monsoon and stabilizing during summer and monsoon. Station-wise patterns further highlighted site-specific pollution drivers such as traffic density, industrial activity, and waste burning. The findings establish kNNI-MLP as a robust AQI prediction framework and provide evidence for targeted interventions, including improved traffic regulation, waste management, and emission controls. Future research will focus on external validation to confirm the model's generalizability across diverse urban contexts, as well as exploring interpretability techniques such as SHAP or variable importance analysis to enhance understanding of predictor contributions.