Scholarlink

<p>The sudden growth in electric vehicle (EV) usage has highlighted the need for advanced and efficient charging infrastructure. This research applies sophisticated ensemble learning models to actual charging session data and proposes a data-based approach to estimating the EV charging duration. The dataset comprises features like battery specs, charging session parameters, arrival and departure times, energy consumed, and temporal variables derived from departure time (hour, day of the week, month). Further extensive preprocessing was undertaken to treat the outliers, especially those from heavy-duty EVs, and to derive more helpful features such as the State of Charge (SOC) differential and charging ratio. The goal of this paper was to predict the charging duration, measured in minutes, using ensembles of algorithm frameworks: XGBoost, LightGBM, CatBoost, and Random Forest. Feature importance analysis was undertaken to evaluate the contribution of each variable to the prediction. In its best run, with an R2 of 0.90, an RMSE of 5.49, and a MAPE of 11.05%, XGBoost achieved the best performance, beating all the other models. This demonstrates the capability of ensemble methods in learning and simulating the EV charging behavior and thereby benefiting the management of charging stations. The study also provides some valuable insights into consumption patterns and possible congestion timings that can aid in better infrastructure planning and user experience.</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

EV charging duration prediction using advanced ensemble learning techniques and feature importance analysis

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Fatma M. Talaat,
  • Mohamed Salem,
  • Alyaa A. Hamza

摘要

The sudden growth in electric vehicle (EV) usage has highlighted the need for advanced and efficient charging infrastructure. This research applies sophisticated ensemble learning models to actual charging session data and proposes a data-based approach to estimating the EV charging duration. The dataset comprises features like battery specs, charging session parameters, arrival and departure times, energy consumed, and temporal variables derived from departure time (hour, day of the week, month). Further extensive preprocessing was undertaken to treat the outliers, especially those from heavy-duty EVs, and to derive more helpful features such as the State of Charge (SOC) differential and charging ratio. The goal of this paper was to predict the charging duration, measured in minutes, using ensembles of algorithm frameworks: XGBoost, LightGBM, CatBoost, and Random Forest. Feature importance analysis was undertaken to evaluate the contribution of each variable to the prediction. In its best run, with an R2 of 0.90, an RMSE of 5.49, and a MAPE of 11.05%, XGBoost achieved the best performance, beating all the other models. This demonstrates the capability of ensemble methods in learning and simulating the EV charging behavior and thereby benefiting the management of charging stations. The study also provides some valuable insights into consumption patterns and possible congestion timings that can aid in better infrastructure planning and user experience.