Machine Learning-Based Prediction of Electric Vehicle Energy Consumption Using Real-World Field Data

Abstract

Accurately predicting electric vehicle (EV) energy consumption in real-world conditions enables better trip planning, charging optimization, and fleet cost control. This work proposes a data-driven framework that learns from field data (CAN-bus logs, GPS, weather, road grade, traffic) to estimate per-trip and per-link energy usage. We benchmark feature-based regressors (Linear/Elastic Net, Random Forest, Gradient Boosting, XGBoost), sequence models (LSTM/GRU, Temporal CNN), and hybrid graph–temporal models on multi-vehicle datasets. Our pipeline addresses data quality (synchronization, denoising, imputation), driver/vehicle heterogeneity (meta-features, mixed-effects), and domain shift (seasonality, geography) via calibration and transfer learning. Results show >10–25% MAE reduction over physics-only baselines and robust generalization to unseen routes, while providing interpretable importance scores for operational use.