Machine Learning-Based Electric Vehicle Remaining Range Prediction: Assessing the Relative Impact of Battery Status, Driving Context, and Environmental Conditions
- 1Department of Industrial Engineering, Çanakkale Onsekiz Mart University, Çanakkale 17010, Türkiye
- 2Department of Energy Management, Çanakkale Onsekiz Mart University, Çanakkale 17010, Türkiye
Abstract
This study presents a comprehensive machine learning (ML) evaluation methodology, grounded in physical principles, for range estimation in battery-electric vehicles (BEVs). In a simulation environment consisting of 50 driving and 12 virtual vehicles, 2000 observations were generated by classifying 23 input variables into three main categories: battery and energy status (SOC, voltage, current, battery power, rolling energy consumption, SOH, battery temperature, regenerative power), driving and route context (speed, acceleration, cumulative distance, route gradient, traffic density, driving style, route type, load), and environmental and thermal context (ambient temperature, relative humidity, wind speed, headwind component, precipitation intensity, solar radiation, HVAC power). The Gradient Boosting (GB) model, selected from among eight algorithms, achieved an R2 = 0.8949, a mean absolute error (MAE) of 12.58 km, and a root mean square error (RMSE) of 16.83 km in the independent test set; In 79.25% of the test observations, the error remained below 20 km. Permutation significance analysis revealed the overwhelming dominance of battery charge state (SOC) and instantaneous energy consumption (R2 reductions of 1.435 and 1.084, respectively). Meanwhile, group-based permutation tests showed that the battery/energy group provided an R2 reduction of 2.0815, compared to reductions of only 0.0419 and 0.0413 for the environmental/thermal and driving/route groups. Ablation tests showed that the model created with only the battery/energy variables achieved R2 = 0.8472 and MAE = 14.17 km, performing close to the full model (R2 = 0.8949, MAE = 12.58 km); however, the model created with only SOC proved unable to generalize in independent driving (R2 = −0.0458). These findings quantitatively confirm that remaining energy and current consumption are the primary determinants of range estimation, while driving, route, and climate variables provide only minor, corrective contributions, and demonstrate that the proposed synthetic framework offers a methodological basis for validation studies using real-world data.
Related
- Remaining Useful Life Prediction of Lithium-Ion Batteries Based on Artificial Neural Network-Supported Wiener Process with Random Effects
- A Study on Lithium-Ion Battery Health Estimation and Remaining Life Prediction Based on Real-World Operational Data
- Physics-Informed Explainable Machine Learning for Hybrid State of Health and Remaining Useful Life Estimation of Lithium-Ion Batteries
- Morphological adaptation of the root system of silver fir responding to changes of local environmental conditions
- Assessing the state of the art in ADMET prediction: Lessons from the OpenADMET-ExpansionRx community blind challenge
BibTeX
@article{atalan2026machine,
title = {{Machine Learning-Based Electric Vehicle Remaining Range Prediction: Assessing the Relative Impact of Battery Status, Driving Context, and Environmental Conditions}},
author = {Abdulkadir Atalan and Yasemin Ayaz Atalan},
journal = {Batteries},
year = {2026},
month = oct,
doi = {10.3390/batteries12100399},
url = {https://www.mdpi.com/2313-0105/12/10/399},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Machine Learning-Based Electric Vehicle Remaining Range Prediction: Assessing the Relative Impact of Battery Status, Driving Context, and Environmental Conditions
AU - Abdulkadir Atalan
AU - Yasemin Ayaz Atalan
PY - 2026
DA - 2026/10/05
JO - Batteries
DO - 10.3390/batteries12100399
UR - https://www.mdpi.com/2313-0105/12/10/399
AB - This study presents a comprehensive machine learning (ML) evaluation methodology, grounded in physical principles, for range estimation in battery-electric vehicles (BEVs). In a simulation environment consisting of 50 driving and 12 virtual vehicles, 2000 observations were generated by classifying 23 input variables into three main categories: battery and energy status (SOC, voltage, current, battery power, rolling energy consumption, SOH, battery temperature, regenerative power), driving and route context (speed, acceleration, cumulative distance, route gradient, traffic density, driving style, route type, load), and environmental and thermal context (ambient temperature, relative humidity, wind speed, headwind component, precipitation intensity, solar radiation, HVAC power). The Gradient Boosting (GB) model, selected from among eight algorithms, achieved an R2 = 0.8949, a mean absolute error (MAE) of 12.58 km, and a root mean square error (RMSE) of 16.83 km in the independent test set; In 79.25% of the test observations, the error remained below 20 km. Permutation significance analysis revealed the overwhelming dominance of battery charge state (SOC) and instantaneous energy consumption (R2 reductions of 1.435 and 1.084, respectively). Meanwhile, group-based permutation tests showed that the battery/energy group provided an R2 reduction of 2.0815, compared to reductions of only 0.0419 and 0.0413 for the environmental/thermal and driving/route groups. Ablation tests showed that the model created with only the battery/energy variables achieved R2 = 0.8472 and MAE = 14.17 km, performing close to the full model (R2 = 0.8949, MAE = 12.58 km); however, the model created with only SOC proved unable to generalize in independent driving (R2 = −0.0458). These findings quantitatively confirm that remaining energy and current consumption are the primary determinants of range estimation, while driving, route, and climate variables provide only minor, corrective contributions, and demonstrate that the proposed synthetic framework offers a methodological basis for validation studies using real-world data.
SN - 2313-0105
ER -