Eco-Driving Level Evaluation Model for Electric Buses Entering and Leaving Stops

Abstract

Sustainable driving is a crucial driver to optimise energy use and emissions in electric public transport. But the current evaluation approaches mostly depend on human assessment or poor statistical parameters which do not adequately reflect complex driving activities around bus stops. The purpose of this effort is to evaluate an electric bus' ecological driving level during stop entry and exit via data-driven intelligence. We developed natural driving datasets for both phases with operational, kinematic and behavioral parameters of electric buses during their normal service time. In this study, the preprocessing of the data included removal of missing data, duplicate data, feature engineering from the Pearson correlation analysis with a p-value, multicollinearity removal via variance inflation factor assessment, and stepwise forward regression for selecting representative variables of the data. Several machine-learning classifiers such as RF, Gradient Boosting and LightGBM were introduced and improved pipelines such as recursive feature elimination and Min-Max scaling were utilized. We also developed sophisticated combination/boosting models such as XGBoost and Voting Classifier. The accuracy, precision, recall, F1-score and ROC–AUC were used to assess the performance of the performance. Performance of Experimental models that were tested on both datasets, shows that XGBoost surpassed all the baseline models with an accuracy of 89.0% and ROC–AUC of 0.944 while Voting Classifier excelled in terms of discrimination potential. Furthermore, we implemented explainability through LIME and SHAP, and the real-time prediction and visualization have been achieved by a flask-based framework. The proposed framework improves the accuracy of eco-driving evaluation and facilitates the intelligently interpretable decision making for intelligent bus operation system.