A COMPARATIVE MATHEMATICAL FRAMEWORK FOR BINARY CLASSIFICATION OF ELECTRIC-VEHICLE RANGE ANXIETY

Source abstract

Range anxiety has limited the spread of electric vehicles to some extent due to concerns about travel and charging. This paper uses 50,000 observations to explore how well people with range-anxiety scores above 5 can be identified and how nine predictors related to mobility, charging, cognition, experience and demographics affect the fitted boundary. Missing values are filled with medians; otherwise, the IQR rule is used for capping extreme values, and continuous variables are Z-standardised. Regularised logistic regression, random forest and XGBoost are compared via a stratified 70:30 split and five-fold cross-validation; probability thresholds are searched and 100 Bayesian trials are used if necessary. XGBoost has achieved the highest accuracy (85.47%), precision (88.61%) and F1-score (84.56%) on the independent test set. Random Forest has the largest AUC of 0.9006, and XGBoost is only slightly lower at 0.8995. Electric-vehicle knowledge, environmental awareness and technology affinity account for 97.28% of the XGBoost importance and 90.92% of the random-forest importance. The above results separate the performance at a selected threshold from ranking ability across thresholds and also show the effect of non-linear relationships and cases near the score cutoff.