Pollutant Adaptive Forecasting of India’s Greenhouse Gas and Air Pollutant Emissions Using Bayesian Dynamic Regression and Conformal Prediction

Abstract

Reliable multi-pollutant emission forecasts are needed to coordinate climate mitigation and air quality management. This study developed a Pollutant Adaptive Bayesian Dynamic Regression and Conformal Prediction framework (PA-BDR-CP-X) for India’s methane (CH 4 ), carbon dioxide (CO 2 ), nitrous oxide (N 2 O), oxides of nitrogen (NO x ) and particulate matter 2.5 micrometers (PM 2.5 ) emissions using annual EDGAR emissions, version EDGAR_2025_GHG, and socioeconomic, energy and agricultural covariates for 1970–2024. Its contribution is a workflow that selects a forecasting structure separately for each pollutant from trend-only or covariate augmented Bayesian dynamic regression, autoregressive integrated moving average (ARIMA) and exponential smoothing (ETS), and then applies conformal uncertainty calibration. Models were assessed using a fixed 2020–2024 test and 90 expanding window rolling origin forecasts per pollutant. In the fixed test, covariate augmented PA-BDR-CP-X reduced root mean square error (RMSE) relative to the best ARIMA/ETS benchmark by 35.07% for CH 4 and 18.56% for PM 2.5 . Rolling origin validation selected trend-only PA-BDR-CP-X for CH 4 and PM 2.5 , ARIMA for CO 2 and N 2 O, and ETS for NO x , demonstrating why no single model family should be imposed across pollutants. After full sample refitting, projected increases from 2025 to 2035 ranged from 12.34% for CH 4 to 37.96% for CO 2 . The 95% conformal intervals covered all 35 rolling origin evaluation forecasts per pollutant, but their width indicates conservative uncertainty bounds rather than precise or guaranteed future coverage. The forecasts provide statistical baselines for coal and energy transition planning, fertilizer and nitrogen use management, methane control in livestock, rice and waste systems, and combustion-emission controls for NO x and PM 2.5 .

Bibliographic Information