Background Reliable daily Air Quality Index (AQI) prediction can support timely public-health advisories in Indian cities. However, AQI forecasting studies can report over-optimistic performance if time ordering is not strictly respected during feature construction and evaluation. Methodology We propose a transparent, time-aware machine learning pipeline for daily AQI prediction using publicly available CPCB measurements (Rao dataset) for multiple Indian cities. The workflow includes consistent data typing, removal of duplicate city-date records, within-city imputation for missing entries, and engineered predictors including calendar variables, pollutant interaction features (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {PM}_{2.5}\)</EquationSource> </InlineEquation>/<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\hbox {PM}_{10}\)</EquationSource> </InlineEquation> ratio and difference), and time-history statistics. Lag features (1, 3, and 7 days) and rolling mean/standard deviation (3, 7, and 14 days) are computed on shifted time series to prevent data leakage. We evaluate Linear Regression, Ridge Regression, Random Forest, and Histogram-based Gradient Boosting using a strictly chronological split, reserving the last 20% of observations as a forward holdout test set. Multi-city validation setting Evaluation is performed on a pooled multi-city panel in which each row corresponds to a City<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation>Date observation. The strictly chronological forward holdout split is applied to the full panel; as a result, the holdout test set contains observations from all cities included in the dataset. Reported metrics therefore summarize aggregate performance across cities for future dates under a unified protocol. City-held-out validation (e.g., leave-one-city-out) is not implemented in the present baseline and is identified as a relevant extension for evaluating spatial transferability. Major findings Random Forest achieved the best holdout performance with MAE = 12.7742, RMSE = 24.8427, R2 = 0.9320, and MAPE = 11.3123%, with gradient boosting performing comparably. Conclusions Leakage-safe time-series features combined with tree-based models provide a reproducible baseline for daily AQI prediction across Indian cities, while remaining peak-error cases motivate future extensions with additional exogenous drivers.

Machine learning approaches for predicting air pollution levels: a transparent, time-aware pipeline for daily AQI in Indian cities

期刊论文
发表日期：

2026年2月21日

查看全文

Philipp Goetzinger,
Sebastian Noy,
Karsten Huffstadt

摘要

Background

Reliable daily Air Quality Index (AQI) prediction can support timely public-health advisories in Indian cities. However, AQI forecasting studies can report over-optimistic performance if time ordering is not strictly respected during feature construction and evaluation.

Methodology

We propose a transparent, time-aware machine learning pipeline for daily AQI prediction using publicly available CPCB measurements (Rao dataset) for multiple Indian cities. The workflow includes consistent data typing, removal of duplicate city-date records, within-city imputation for missing entries, and engineered predictors including calendar variables, pollutant interaction features ( \(\hbox {PM}_{2.5}\) / \(\hbox {PM}_{10}\) ratio and difference), and time-history statistics. Lag features (1, 3, and 7 days) and rolling mean/standard deviation (3, 7, and 14 days) are computed on shifted time series to prevent data leakage. We evaluate Linear Regression, Ridge Regression, Random Forest, and Histogram-based Gradient Boosting using a strictly chronological split, reserving the last 20% of observations as a forward holdout test set.

Multi-city validation setting

Evaluation is performed on a pooled multi-city panel in which each row corresponds to a City \(\times \) Date observation. The strictly chronological forward holdout split is applied to the full panel; as a result, the holdout test set contains observations from all cities included in the dataset. Reported metrics therefore summarize aggregate performance across cities for future dates under a unified protocol. City-held-out validation (e.g., leave-one-city-out) is not implemented in the present baseline and is identified as a relevant extension for evaluating spatial transferability.

Major findings

Random Forest achieved the best holdout performance with MAE = 12.7742, RMSE = 24.8427, R² = 0.9320, and MAPE = 11.3123%, with gradient boosting performing comparably.

Conclusions

Leakage-safe time-series features combined with tree-based models provide a reproducible baseline for daily AQI prediction across Indian cities, while remaining peak-error cases motivate future extensions with additional exogenous drivers.