Vehicle-wise validated ensemble learning framework for robust electric vehicle energy consumption prediction using...

Area of Science:

• Energy Systems Engineering
• Machine Learning Applications
• Automotive Engineering

Background:

• Accurate electric vehicle (EV) energy consumption prediction is vital for addressing range anxiety and optimizing route planning.
• Current prediction methods often suffer from overly optimistic results due to random data splitting and limited evaluation metrics.
• A realistic evaluation is necessary to assess model performance on unseen data.

Purpose of the Study:

• To develop a systematic ensemble learning framework for accurate EV energy consumption prediction.
• To ensure realistic model evaluation by employing vehicle-wise data splitting, preventing information leakage.
• To compare the performance of five distinct machine learning models using physics-based features.

Main Methods:

• Implemented an ensemble learning framework with vehicle-wise data splitting.
• Tested five machine learning models: Random Forest, ExtraTrees, XGBoost, LightGBM, and CatBoost.
• Utilized real-world EV data incorporating physics-based features for vehicle movement and environmental conditions.
• Evaluated model performance using MAE, RMSE, MAPE, R², and EVS metrics.
• Performed statistical significance testing with Friedman and Wilcoxon tests.

Main Results:

• The ExtraTrees model achieved the highest performance, with R² = 0.9407, RMSE = 0.6113 kWh/km, and MAPE = 7.62%.
• ExtraTrees demonstrated consistent accuracy across diverse driving conditions.
• Statistical analysis confirmed ExtraTrees significantly outperformed LightGBM, with comparable results to other ensemble models.
• Ensemble methods effectively reduced prediction errors for various EV data types.

Conclusions:

• The proposed framework offers a dependable solution for predicting EV energy consumption and range.
• The ExtraTrees model, within the ensemble framework, provides a robust approach for real-world applications.
• Vehicle-wise data splitting is critical for realistic performance evaluation in EV energy consumption studies.