EBus EcoPredict: Machine Learning for Electric Bus Energy Consumption Forecasting

Abstract

Accurate energy consumption forecasting is a critical enabler for the large-scale adoption of electric buses (ebuses) in urban public transit systems. Unpredictable energy usage leads to range anxiety, inefficient charging schedules, fleet underutilisation, and elevated operational costs. This paper presents EBus EcoPredict, a machine learningbased energy consumption forecasting system that integrates an XGBoost gradient-boosted tree model with a stacked LSTM neural network in a weighted ensemble architecture to predict trip-level and segment-level e-bus energy consumption. The model ingests real-time and historical data streams including GPS route telemetry, CAN bus signals (speed, state-of-charge, motor current), passenger load, ambient temperature, topographic gradient, and HVAC utilisation. Trained on a curated dataset of 38,600 annotated trip records collected from a 45-bus urban fleet over 18 months, EBus EcoPredict achieves an overall classification accuracy of 96.1%, precision of 95.2%, recall of 94.8%, and F1-score of 95.0%, outperforming standalone baselines including Linear Regression, Decision Tree, Random Forest, XGBoost, and LSTM. Field deployment across a 45-bus fleet demonstrated a 75% reduction in energy forecast mean absolute error (MAE from 4.8 to 1.2 kWh/trip), a 25% reduction in fleet energy cost, and an 83.3% reduction in range anxiety incidents, validating EBus EcoPredict as a scalable and operationally effective solution for smart electric bus fleet management