TimeGAN-based generative modelling for air quality index forecasting and missing data imputation in Indian cities

Abstract Air pollution remains a pressing public-health issue in Indian metropolitan cities, where elevated levels of particulate matter (PM2.5 and PM10), carbon monoxide, nitrogen dioxide, sulphur dioxide and ozone regularly exceed safe thresholds. Air Quality Index offers a convenient single-number summary of these hazards, building accurate forecasts with this data not straightforward. The monitoring stations produce records riddled with missing values, high-pollution events are rare enough to create severe class imbalance, and the temporal dependencies. This paper proposes a framework built around a Time-series Generative Adversarial Network (TimeGAN) that simultaneously tackles AQI forecasting, missing-data imputation, and synthetic scenario generation. The pipeline combines Synthetic Minority Oversampling (SMOTE) for class balancing with TimeGAN's adversarial and supervised objectives, and is trained on hourly observations from the Central Pollution Control Board covering six major Indian cities over 2018–2023. Forecasting is framed primarily as a regression problem, while predicted AQI values are additionally mapped to discrete categories (Good, Moderate, Poor, Severe) for a secondary classification evaluation. The proposed model is compared against benchmark ARIMA, Support Vector Machine, Random Forest, and Long Short-Term Memory. The experimental results revealed that the proposed model reported better performance on both regression metrics (MAE, MSE, R 2 ) and classification metrics (Accuracy, Precision, Recall, F1-score). It achieved R 2 = 0.94, MAE = 3.72, and MSE = 4.25 alongside a secondary classification accuracy of 95.8% and F1-score of 0.946. A paired Wilcoxon signed-rank test confirms statistically significant improvement over LSTM ( p < 0.01), and fivefold cross-validation yields R 2 = 0.93 ± 0.012, indicating stable generalisation across temporal partitions.