Scholarlink

<p>Accurate prediction of electric vehicle (EV) driving range is important for consumers, manufacturers, and transportation planners. Among the available range standards, the Environmental Protection Agency (EPA) range is widely used as a practical benchmark, yet its estimation from vehicle specifications remains challenging because of nonlinear relationships, feature redundancy, and limited transferability across unseen groups. This study develops an interpretable machine learning framework for EPA range prediction using EV specification data, with the primary objective of determining whether a compact subset of EV specification variables can deliver competitive predictive performance while improving interpretability and reducing reliance on unnecessarily large feature sets. A split-first workflow was adopted to prevent information leakage, followed by training-only feature screening, repeated cross-validation, and held-out test evaluation. Multiple regression models were compared across compact numeric, battery-centered, and ablation-based feature sets. Tree-based ensemble models delivered the strongest overall performance, with Extra Trees selected as the final champion. Battery capacity emerged as the dominant predictor, while top speed, number of cells, and weight-related variables provided additional predictive value. SHAP analysis, partial dependence plots, and leave-one-feature-out ablation confirmed the interpretability of the final model and clarified the contribution of the strongest predictors. Out-of-group validation across body style, brand, and market produced substantially weaker performance than conventional random-split testing, highlighting the importance of stricter generalization assessment in EV range modeling. Overall, the findings show that accurate and interpretable EPA range prediction is achievable using compact EV specification sets, although transferability across unseen groups remains a significant challenge.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable machine learning for electric vehicle EPA range prediction: compact feature design, ablation analysis, and out-of-group validation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月14日

查看全文
  • Muhammad Zeeshan Younas

摘要

Accurate prediction of electric vehicle (EV) driving range is important for consumers, manufacturers, and transportation planners. Among the available range standards, the Environmental Protection Agency (EPA) range is widely used as a practical benchmark, yet its estimation from vehicle specifications remains challenging because of nonlinear relationships, feature redundancy, and limited transferability across unseen groups. This study develops an interpretable machine learning framework for EPA range prediction using EV specification data, with the primary objective of determining whether a compact subset of EV specification variables can deliver competitive predictive performance while improving interpretability and reducing reliance on unnecessarily large feature sets. A split-first workflow was adopted to prevent information leakage, followed by training-only feature screening, repeated cross-validation, and held-out test evaluation. Multiple regression models were compared across compact numeric, battery-centered, and ablation-based feature sets. Tree-based ensemble models delivered the strongest overall performance, with Extra Trees selected as the final champion. Battery capacity emerged as the dominant predictor, while top speed, number of cells, and weight-related variables provided additional predictive value. SHAP analysis, partial dependence plots, and leave-one-feature-out ablation confirmed the interpretability of the final model and clarified the contribution of the strongest predictors. Out-of-group validation across body style, brand, and market produced substantially weaker performance than conventional random-split testing, highlighting the importance of stricter generalization assessment in EV range modeling. Overall, the findings show that accurate and interpretable EPA range prediction is achievable using compact EV specification sets, although transferability across unseen groups remains a significant challenge.