Vehicle-wise validated ensemble learning framework for robust electric vehicle energy consumption prediction using physics-guided features and statistical benchmarking.

Journal: Scientific reports
Published Date:

Abstract

Predicting EV energy consumption accurately is essential for solving range anxiety and helping with better route planning and energy management in real conditions. Many current studies use random data splitting and few evaluation measures, which leads to overly optimistic results. Moreover, this approach does not properly test how well the models work on new data. According to the research requirements, this study presents a systematic ensemble learning framework using vehicle-wise data splitting to prevent information leakage and ensure realistic evaluation. Five machine learning models-Random Forest, ExtraTrees, XGBoost, LightGBM, and CatBoost-are tested using real electric vehicle data that includes only physics-based features for vehicle movement and environmental conditions. The model performance is evaluated using multiple metrics like MAE, RMSE, MAPE, R2, and EVS, and statistical significance is further checked through Friedman and Wilcoxon tests. The results show that the ExtraTrees model gives the best performance with R2 = 0.9407, RMSE = 0.6113 kWh/km, and MAPE = 7.62%. Additionally, this model shows good accuracy across all driving conditions. The results indicate that ExtraTrees works significantly better than LightGBM, while it gives similar results to other ensemble models. Furthermore, the results demonstrate that ensemble methods work well for different types of EV data by reducing errors. The suggested framework provides a robust and dependable solution for predicting EV energy use and range in real situations.

Authors

Keywords

No keywords available for this article.