Scenario aware multi pollutant emissions forecasting for India using hybrid ridge-VARX and gradient boosting with inequality aware reconciliation

Abstract

Abstract Air quality planning in India needs forecasts that stay interpretable, probabilistically reliable, and physically coherent across pollutants. This study develops a novel scenario-aware multi-pollutant forecasting methodology with growth-sensitive ridge vector autoregression with exogeneous drivers and gradient boosting model with inequality-aware reconciliation (RVRX-GBM-Rec) to project annual pollutants of BC, OC, CO, NO x , NMVOC, NH 3 , SO 2 , PM 2.5 , and PM 10 . The methodology models all series in Δlog space over 1970–2022, conditioning dynamics on exogenous macro energy growth (GDP, population, and primary energy). Forecast accuracy is assessed with an expanding window backtest over 2013–2022 and benchmarking against a random walk, AR(1) in Δlog, and fixed parameter vector autoregression with exogeneous drivers. Across horizons, RVRX-GBM-Rec reduced MAE by 52% versus random walk, 41% versus AR(1), and 36% versus fixed vector autoregression with exogeneous drivers. Probabilistic forecasts for 2023–2035 are generated under baseline, accelerated control, and rebound macro energy scenarios. Inequality-aware reconciliation enforces non-negativity, particulate identities (PM 2.5 ≤PM 10 ; BC + OC≤PM), and uncertainty was quantified via dependence-preserving residual bootstrap medians and 95% prediction intervals with coverage and sensitivity checks. Accelerated control bends median downward earlier and yields narrow absolute bands by the early 2030s, while rebound showed an early uptick with wider near-term uncertainty. Physical coherence was enforced throughout PM 2.5 ≤PM 10 , BC + OC≤PM 2.5 , and non-negativity with no median path violations post-reconciliation over 2023–2035. Scenario separation was primarily driven by energy growth, implying that clean power, industrial fuel switching, and sulfur/NO x controls are central levers for shifting both forecast medians and tails through 2035.