Multivariate Statistical Techniques for Air Quality Index Prediction: Machine Learning and Time Series Analysis of Indian Metropolitan Cities

Introduction

Urban air pollution constitutes pressing environmental health challenge affecting over 1.4 billion people residing in Indian cities experiencing rapid industrialization, vehicular growth, and population density increases [1] . Ambient particulate matter exposure causes an estimated 1.7 million premature deaths annually in India, with economic costs exceeding 3% of GDP through healthcare expenditures, productivity losses, and premature mortality [2] . Indian metropolitan cities including Delhi, Mumbai, Kolkata, and Bangalore consistently record air quality index values exceeding World Health Organization guidelines, with Delhi experiencing annual average PM2.5 concentrations three times above safe levels [3] . The Air Quality Index provides standardized metric translating complex multi-pollutant measurements into single numeric value enabling public comprehension of health risks [4] . Indian AQI incorporates six criteria pollutants: particulate matter (PM2.5 and PM10), nitrogen dioxide (NO2), sulfur dioxide (SO2), carbon monoxide (CO), and ozone (O3). AQI calculation identifies dominant pollutant determining overall index value, with categories ranging from good (0-50) to severe (401-500) triggering corresponding health advisories and emergency response protocols [5] . Accurate AQI forecasting enables proactive public health interventions including pollution advisories warning vulnerable populations, traffic management implementing odd-even vehicle restrictions, industrial emissions curtailment during pollution episodes, and school closure decisions protecting children from exposure [6] . However, air quality prediction presents substantial methodological challenges due to complex non-linear relationships among pollutants, meteorological variables, and emission sources exhibiting diurnal, weekly, and seasonal patterns. Traditional deterministic models require comprehensive emissions inventories and computational resources limiting operational deployment, while statistical approaches offer computationally efficient alternatives leveraging historical patterns for predictive inference [7] . Previous research established statistical relationships between meteorological conditions and air quality, identifying temperature inversions, wind patterns, and atmospheric stability as key drivers of pollution concentrations [8] . Machine learning algorithms including random forests, support vector machines, and neural networks demonstrated superior predictive accuracy compared to linear regression through automatic feature engineering and non-linear relationship modeling [9] . However, limited research comprehensively compared multiple statistical techniques across Indian cities or evaluated temporal forecasting incorporating lagged variables and seasonal patterns. Additionally, model interpretability remains critical for policy applications requiring understanding of driving factors rather than black-box predictions [10] . This research addresses these gaps by systematically developing and evaluating multivariate statistical techniques for AQI prediction across five Indian metropolitan cities. The objectives encompass: (i) characterizing spatial and temporal air quality patterns; (ii) quantifying relationships among pollutants and meteorological variables; (iii) comparing predictive performance of statistical and machine learning models; (iv) identifying dominant predictors through feature importance analysis; and (v) demonstrating operational forecasting applications supporting pollution control policy.

Data Quality Assessment and Preprocessing: Air quality data quality assessment constitutes critical prerequisite ensuring reliable predictive models. Monitoring station data underwent rigorous quality control procedures following Central Pollution Control Board protocols. Completeness analysis identified 87.3% valid measurements across five cities during 2017-2019 period, with missing data primarily attributable to instrument calibration, power failures, and maintenance activities. Missing value treatment employed multiple imputation using chained equations preserving temporal autocorrelation and cross-pollutant relationships. Outlier detection utilized Tukey's method identifying extreme values exceeding 1.5 times interquartile range; suspected outliers underwent manual validation against meteorological events, emission episodes, and measurement errors. Valid extreme values including Diwali firework pollution and crop burning episodes were retained. Data consistency checks verified temporal continuity, eliminated duplicate records, and ensured measurement unit standardization. Meteorological data from India Meteorological Department exhibited 96.2% completeness with minimal missing values imputed through spatial interpolation from neighboring stations. The preprocessing pipeline transformed raw measurements into analysisready dataset facilitating reproducible model development and evaluation [11] .

Materials and Methods

Materials

Research utilized daily air quality measurements from Central Pollution Control Board continuous ambient air quality monitoring stations in Delhi, Mumbai, Bangalore, Kolkata, and Chennai. Pollutant concentrations included PM2.5, PM10, NO2, SO2, CO, and O3 measured through automated analyzers following standard protocols. AQI calculated using sub-index methodology identifying dominant pollutant. Meteorological data from India Meteorological Department encompassed temperature, humidity, wind speed, wind direction, and atmospheric pressure. The dataset covered January 1, 2017 to December 31, 2019, comprising 1,095 days per city yielding 5,475 city-day observations after quality control and complete case selection.

Methods

Model Validation Procedures and Performance Metrics

Model validation employed temporal hold-out strategy reserving entire 2019 calendar year for independent testing, ensuring predictions utilized only historical information available at forecast time. Walk-forward validation assessed one-day-ahead forecasting accuracy mimicking operational deployment wherein previous day observations inform next-day predictions. Directional accuracy measured percentage of correctly predicted AQI increases or decreases, relevant for early warning systems triggering interventions based on pollution trend detection. Categorical accuracy evaluated correct AQI category classification, critical for public health advisory systems communicating qualitative risk levels. Residual analysis examined prediction errors for systematic patterns, heteroscedasticity, and temporal autocorrelation indicating model inadequacy. Quantile-quantile plots assessed residual normality assumptions. Model robustness evaluated performance across pollution regimes including low, moderate, and high AQI episodes, ensuring reliable predictions across operational range. Cross-city generalization tested models trained on combined fourcity data and validated on held-out fifth city, assessing transferability across geographic contexts. These validation procedures ensured comprehensive assessment of predictive accuracy, calibration, and operational utility supporting evidence-based deployment recommendations [12] . The implementation provides template for national-scale air quality forecasting infrastructure supporting National Clean Air Programme goals [13] . extension to longer forecast horizons enabling multi-day planning; and investigation of causal inference methods establishing pollutionhealth relationships supporting intervention impact evaluation. Scaling operational implementation requires: investment in monitoring infrastructure ensuring data availability and quality; development of automated quality control procedures minimizing manual oversight; establishment of institutional frameworks integrating forecasts into emergency response protocols; and public engagement campaigns promoting forecast awareness and protective behavior adoption. The outcomes demonstrate that multivariate statistical techniques provide accurate, interpretable, and operationally viable approach for air quality prediction in developing country contexts. The methodology addresses data limitations, computational constraints, and policy requirements characteristic of emerging economy environmental management. Successful deployment requires sustained commitment to monitoring infrastructure, analytical capacity development, and science-policy integration enabling evidence-based environmental governance supporting public health protection and sustainable urban development across India's rapidly growing metropolitan regions.

Results

Conclusion

References

1. Balakrishnan K, Dey S, Gupta T, Dhaliwal RS, Brauer M, et al. The impact of air pollution on deaths. Lancet. 2019;394(10216):2489-2497.
2. Pandey A, Brauer M, Cropper ML, Balakrishnan K, Mathur P, et al. Health and economic impact of air pollution in India. Lancet Planet Health. 2021;5(1):e25e33.
3. Gurjar BR, Jain A, Sharma A, Agarwal A, Gupta P, et al. Human health risks in megacities due to air pollution. Atmos Environ. 2010;44(36):4606-4613.
4. Central Pollution Control Board. National Air Quality Index. New Delhi (IN): Ministry of Environment, Forest and Climate Change; 2014.
5. Kumar P, Morawska L, Martani C, Biskos G, Neophytou M, et al. The rise of low-cost sensing for managing air pollution. Environ Int. 2015;75:199-205.
6. Sharma S, Chatani S, Mahtta R, Goel A, Kumar A. Sensitivity analysis of ground level ozone in India. Atmos Environ. 2016;131:29-40.
7. Stieb DM, Burnett RT, Smith-Doiron M, Brion O, Shin HH, et al. A new multipollutant index for air quality. Environ Int. 2008;34(8):1085-1092.
8. Perez P, Reyes J. An integrated neural network model for PM10 forecasting. Atmos Environ. 2006;40(16):28452851.
9. Biancofiore F, Busilacchio M, Verdecchia M, Tomassetti B, Aruffo E, et al. Recursive neural network model for analysis and forecast of PM10. Atmos Environ. 2017;155:62-73.
10. Jiang D, Zhang Y, Hu X, Zeng Y, Tan J, et al. Progress in developing an ANN model for air pollution index forecasting. Atmos Environ. 2004;38(40):7055-7064.
11. Jian L, Zhao Y, Zhu YP, Zhang MB, Bertolatti D. An application of ARIMA model to predict submicron particle concentrations. Atmos Environ. 2012;62:403409.
12. Sayegh AS, Munir S, Habeebullah TM. Comparing the performance of statistical models for predicting PM10. Aerosol Air Qual Res. 2014;14(3):653-665.
13. Silibello C, D'Allura A, Finardi S, Bolignano A, Sozzi R. Application of bias adjustment techniques to improve air quality forecasts. Atmos Pollut Res. 2015;6(6):928-938.
14. Donnelly A, Misstear B, Broderick B. Real time air quality forecasting using integrated parametric and nonparametric regression techniques. Atmos Environ. 2015;103:53-65.
15. Arhami M, Kamali N, Rajabi MM. Predicting hourly air pollutant levels using artificial neural networks. Environ Model Softw. 2013;48:193-203.