Open Access

ARTICLE

Explainable Machine Learning for Electric Bus Energy Prediction under Tropical Urban Conditions: A Physics-Informed Parametric Framework

Mohammed Almajed¹, Mahbub Hassan², Turjoy Das Turjo³, Md Ashequl Islam^4,*, Md Ehtesamul Haque¹, M. M. Hafizur Rahman⁵
1 Department of Computer Science, College of Computer Sciences and Information Technology (CCSIT), King Faisal University, Al-Ahsa, Saudi Arabia
2 Department of Civil Engineering, Faculty of Engineering, Chulalongkorn University, Pathumwan, Bangkok, Thailand
3 Department of Civil Engineering, Ahsanullah University of Science and Technology, Dhaka, Bangladesh
4 School of Electrical and Mechanical Engineering, College of Engineering and Information Technology, Adelaide University, Adelaide, Australia
5 Department of Computer Networks & Communications, College of Computer Sciences and Information Technology (CCSIT), King Faisal University, Al-Ahsa, Saudi Arabia
* Corresponding Author: Md Ashequl Islam. Email:
(This article belongs to the Special Issue: Sustainable Transport Technologies and Strategies: Impacts on Energy and Environment)

Energy Engineering https://doi.org/10.32604/ee.2026.084024

Received 15 April 2026; Accepted 20 May 2026; Published online 16 June 2026

Abstract

Accurate per-kilometer energy consumption prediction is essential for effective electric bus (EB) fleet management in urban transit networks. Existing studies report mean absolute percentage errors of 5%–8% and rarely cross-validate explainability attributions or quantify prediction uncertainty. This study presents a physics-informed, interpretable machine learning framework for EB energy consumption prediction under Bangkok-like tropical urban conditions. Due to the institutional inaccessibility of operational telemetry, a parametric dataset of 4000 trip-level scenarios was constructed across 29 input features anchored to published primary sources. Four algorithms were benchmarked under Bayesian hyperparameter optimization: Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Random Forest, and Support Vector Regression. XGBoost achieved the best performance (R2=0.9439, MAPE =3.51%), improving on the 5%–8% range of comparable prior studies within the defined parametric scenario space. SHapley Additive exPlanations (SHAP) identified average speed, HVAC load, and road gradient as the dominant energy drivers. Cross-method validation against Local Interpretable Model-Agnostic Explanations (LIME) provided near-threshold support for the dominant attribution hierarchy (ρ¯=0.686), with both methods agreeing on the top-three predictors. An ablation study across four feature groups supported the non-redundant predictive contribution of speed, congestion, thermal, and infrastructure variables within the parametric scenario space. Quantile regression-based prediction intervals achieved 79.4% empirical coverage, indicating useful but under-calibrated exploratory uncertainty estimates. A companion XGBoost classifier assigned trips to ordinal efficiency categories with 88.25% accuracy and macro-F1 of 0.8423, with no misclassifications across extreme class boundaries. These findings provide a transparent, quantitatively grounded framework that may support future fleet energy management, route optimization, and driver training after validation on live fleet data, while advancing interpretable EB energy modeling for tropical urban contexts.

---

Keywords

Electric bus; energy consumption prediction; energy efficiency modeling; explainable machine learning; sustainable transit

---