Effects of Driving Behavior on Fuel Consumption with Explainable Gradient Boosting Decision Trees

Fuel consumption modeling using real-world collected driving data has been at the center of recent research, due to the emergence of low-cost driving data collection sensors. In this paper, Gradient Boosting Decision Trees for the estimation of fuel consumption per trip are developed, trained and evaluated based on data related to the type of vehicle, the driving and the route characteristics, in order to depict the strategic, tactical and operational level of the decision-making process related to driving. The training exploits a dataset of naturalistic driving data collected by smartphone sensors, enriched with fuel consumption measured through OBD devices. The model achieves a very satisfactory performance, better than most similar models of recent literature, with a mean absolute percentage error of 9.8%. The further ex-post analysis of the explainability of the developed model using Shapley Additive explanations (SHAP) values reveals that average speed, deceleration, acceleration, as well as idling percentage, road inclination and the type of vehicle have the highest significant influence on fuel consumption. The paper ends with a discussion on the implications of such models for future research and their use in real-world transportation problems.