Lightweight Machine Learning for Electric-Vehicle Energy Consumption Prediction: A Simulation Study with Edge-Deployable Models

Accurate short-horizon prediction of electric-vehicle (EV) energy consumption enables better range estimation, adaptive eco-driving, and charging planning. While deep models can be accurate, their computational cost and data requirements limit on-board deployment, especially in data-scarce scenarios. This paper presents a compact, edge-deployable pipeline for predicting per-kilometer energy consumption using readily available telematics-vehicle speed, longitudinal acceleration, road grade, ambient temperature, and state-of-charge (SOC). We synthesize a physically consistent dataset that emulates traction power demand under real driving cycles and evaluate three models of increasing complexity: linear regression (LR), gradient boosting regression (GBR), and a small gated recurrent unit (GRU) network. Physics-informed features (e.g., drag and rolling components, smoothed tractive power) are incorporated to improve sample efficiency. In ۵-fold cross-validation, GBR reduces mean absolute error (MAE) by ۲۲-۳۵% relative to LR across urban, suburban, and highway profiles, approaching GRU accuracy while maintaining an order-of-magnitude lower inference latency and memory footprint suitable for microcontroller-class hardware. Ablation shows road-grade and speed-variance are dominant predictors, and model performance gracefully degrades under sensor noise. Results indicate that carefully engineered, shallow ML models can provide reliable energy forecasts for on-board use, offering a practical alternative to heavyweight deep networks. We release the simulation configuration and code to facilitate reproducibility and extensions to real fleets.

Keywords:

vehicles , energy consumption prediction , edge AI , gradient boosting , physics-informed features , telematics