Vehicular emission trajectories in Sri Lanka : a scenario-based analysis

Abstract

This study presents a novel, data-driven framework for estimating and forecasting vehicular emissions in Sri Lanka, addressing a critical gap in disaggregated emissions modelling on a national scale. While current estimations rely on a top-down methodology, this research introduces a bottom-up approach utilising vehicle-level data from the national Vehicle Emissions Test (VET) programme, spanning 2016– 2024. The study focuses on internal combustion engine (ICE) passenger cars, and processes raw VET data through cleaning, mileage estimation, and integration of vehicle technology/category classifications and emission factors as per EMEP/EEA Tier 02 guidelines. A Long Short-Term Memory (LSTM) neural network is developed to forecast average annual mileage using time-series data, including fleet composition, economic indicators, and disruption events. The model achieved a MAE of 121.81 km and an R-squared value of 0.85 during validation, showing strong predictive accuracy and a good fit. The trained model is used to predict two scenarios, with Scenario 01 being the continuation of the national vehicle import ban and Scenario 02 being the lifting of the ban, allowing for modernisation with Euro 6-compliant vehicles. Using the predicted annual spanning mileage, annual spanning emissions of CO, NMVOC, NOₓ, N₂O, NH₃, and PM are estimated by multiplying by matched emission factors. Results indicate a divergence with Scenario 02 having higher total annual spanning emissions with an increase of 1.34% (CO), 3.68% (NMVOC), 1.62% (NOₓ), 10.5% (N₂O), 6.33% (NH₃) and 6.92% (PM), at the end of the prediction horizon, reflecting the impact of increased vehicle activity despite cleaner technologies. This study demonstrates the potential of regulatory datasets for predictive modelling in low- resource environments and supports evidence-based air quality policy planning through a scalable methodology.