A machine learning approach capturing the effects of driving behaviour and driver characteristics on trip-level emissions

Abstract

This study investigates the effects of different variables including meteorology, trip characteristics (such as time of day), driving characteristics (such as the frequency of extended idling), and driver characteristics (such as driving experience) on trip-level emission factors (EFs). Drivers in the Greater Toronto and Hamilton Area (GTHA) were recruited to collect in-vehicle GPS data over a one-week study period from March to July 2018. Data from 1113 driving trips were collected, including characteristics of the trips and the drivers (51 independent variables). Trip emissions were estimated in addition to a driving eco-score indicator (on a hundred point scale) based on log-transformed emissions of greenhouse gases (GHG) in CO2eq and fine particulate matter (PM2.5). A machine learning approach, the Extreme Gradient Boosting (XGBoost), was used to develop prediction models for CO2eq and PM2.5 emissions at a trip level. The coefficient of determination (R2) and root-mean-square-error (RMSE) of eco-score models were respectively 0.84 (std. dev. 0.05), and 10.26 (std. dev. 1.24) for CO2eq, and 0.85 (std. dev. 0.03), and 10.64 (std. dev. 0.79) for PM2.5. The novel Shapley additive explanation (SHAP) measures were employed to reveal the importance of various features affecting trip emissions. For CO2eq, driving behavior such as the frequency of extended idling was found to have the most significant impact on the trip emission intensity. Additionally, driving experience was the most significant discrete feature affecting the eco-score. For PM2.5, the most significant feature was driver age, which was highly correlated with vehicle model year. Finally, commuter drivers were found to have lower CO2eq and PM2.5 emission intensities, owing to their familiarity with route and traffic conditions.