Leakage-Aware Ensemble Learning for Electric Vehicle Range Prediction and BEV/PHEV Classification

Abstract

This study presents a supervised regression model for predicting the EPA-certified driving range of battery electric vehicles and a binary classification of battery electric vehicles from plug-in hybrid electric vehicles that utilizes the complete 2025 Washington State Department of Licensing EV Population Dataset. On the data quality assessment process, it reveals that 98.69% of the base Manufacturer’s Suggested Retail Price is absent. In the battery electric vehicle, the Electric Range value is zero for $75.9 \%$ of the records due to Clean Alternative Fuel Vehicle eligibility. With these finding, an ablation study is done to model the regression and classification performance. For the regression results, XGBoost achieves the highest test performance with an $\mathbf{R}^{\mathbf{2}}$ of 0.978, and a root mean squared error of 10.68 miles. Random Forest demonstrates the most consistent results in 5-fold cross-validation, with an $\mathbf{R}^{\mathbf{2}}$ of $0.987 \pm 0.001$. On the classification results, near-perfect accuracy in the full-feature scenario is achieved using CAFV Eligibility encoding, which accounts for 85.7% of XGBoost’s feature importance. In a more realistic deployment condition utilising only specification features, the Stacking Ensemble attains 93.09% accuracy and an AUC-ROC of 0.980, indicating a leakage gap of 6.9 to 10.4 percentage points. The results of the ablation study demonstrate that temporal features and vehicle identity are both essential when employing leave-one-out encoding, as its removal increases RMSE by 200.9%. The findings further indicate that CAFV Eligibility, rather than Electric Range, is the principal source of classification leakage.