A Data-Centric Model for Analyzing Driving Efficiency of Electric Buses Around Stops

Abstract

Eco-driving practices are essential for minimizing energy usage and emissions in electric public transportation systems. Traditional evaluation methods predominantly depend on human criteria or constrained statistical indicators, which inadequately represent intricate driving behaviors in proximity to bus stops. This study examines the ecological driving performance of electric buses during stop entry and exit situations through data-driven analysis. Natural driving datasets were developed for both stages, encompassing operational, kinematic, and behavioral attributes captured from electric buses during standard service intervals. Data preprocessing entailed the elimination of missing values and duplicates, succeeded by feature engineering via Pearson correlation analysis with associated p-values, multicollinearity mitigation through variance inflation factor evaluation, and stepwise forward regression for the selection of representative variables. Various machine-learning classifiers, such as Random Forest, Gradient Boosting, and LightGBM, were employed, with optimized pipelines featuring recursive feature reduction and Min–Max scaling. Advanced ensemble and boosting models, including XGBoost and a Voting Classifier, were also created. Performance was assessed utilizing accuracy, precision, recall, F1-score, and ROC–AUC metrics. Experimental results indicate that XGBoost attains the highest overall performance, achieving an accuracy of 89.0% and a ROC–AUC of 0.944, surpassing all baseline models across both datasets, but the Voting Classifier exhibits enhanced discrimination capabilities. Additionally, explainability was achieved with LIME and SHAP, while a Flask-based framework was utilized to facilitate real-time prediction and visualization. The suggested framework enhances the accuracy of eco-driving assessments and facilitates interpretable decision-making for intelligent bus operation systems.