A Hybrid Machine Learning and Physics-based Approach for Accurate Energy Consumption Modeling of Electric Buses in Public Transport

ABSTRACT: Public transport organizations are increasingly concerned about reducing air pollution, leading many to transition their fleets into electric vehicles (EVs). In this context, limited battery range and charging times remain significant hurdles. Precise modeling of electric bus energy consumption is crucial. Still, existing methods often face difficulties due to the complexities of real-world conditions, such as diverse driving patterns and external factors. To tackle this, the study proposes a hybrid model combining physical principles and machine learning using real-world data from 30 buses across 130 routes over one year. Key variables like passenger load, weather, and route characteristics are incorporated. Several machine learning models, including MLP, KAN, and XGBoost, are compared using Mean Absolute Percentage Error (MAPE). The hybrid model outperforms others, achieving a low MAPE of 5.59 % on test data and 5.79 % on validation data with a low Standard Deviation. Additionally, models incorporating operational factors, such as bus lines and time of day, enhance prediction accuracy. The study concludes that integrating physical laws with machine learning offers a more accurate and stable approach to energy consumption modeling, providing a promising framework for fleet management and energy efficiency in public transport systems.