AQIntel: Formula Approximation and AQI Estimation under Incomplete Pollutant Measurements
Year: 2026
Type: preprint
Abstract: AQIntel is an experimental machine learning study examining two questions: how accurately standard machine learning models can approximate a deterministic AQI calculation, and whether learned models can estimate AQI when pollutant measurements are unavailable.The formula-approximation study compares linear regression, ridge regression, decision trees, random forests, gradient boosting, and XGBoost under interpolation, breakpoint, dominant-pollutant switching, and out-of-distribution conditions. Random Forest achieved an interpolation MAE of 0.62 AQI units with 16,384 training examples, while stress testing showed substantially larger errors under distribution shift.The missing-measurement study uses 420,768 hourly observations from the UCI Beijing Multi-Site Air Quality dataset, producing 13,320 eligible station-days. A Random Forest pollutant-imputation pipeline reduced equal-site, equal-mask MAE from 13.40 to 4.65 on later dates and from 13.50 to 4.71 across independently held-out stations. Compared with available-pollutant-only estimation, the learned pipeline improved MAE by 6.27 AQI units on later dates and 6.22 on held-out stations.The study uses a bounded six-pollutant CPCB-derived research convention and does not claim official AQI equivalence, future forecasting, deployment safety, or global generalization. Reproducibility materials include code, saved predictions, model-selection records, bootstrap uncertainty estimates, processed data, and experiment logs.
Author: Armaan Kumar
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221462545
Cites:
Cited by:
Related to: