Personalized Prediction of Vehicle Velocity and Energy Consumption Using Road Features Information

Abstract

Electric Vehicles (EVs) represent a compelling alternative for ground transportation, yet range anxiety—caused by limited battery capacity and sparse charging infrastructure—remains a key adoption barrier. A reliable pre-trip prediction of EV battery energy consumption can substantially mitigate this concern. This paper presents a two-fold prediction framework that bridges behavioural forecasting and deterministic energy optimisation. The framework couples a Hybrid LSTM-Transformer architecture for pre-trip trajectory generation with a physics-based cruise-speed optimizer that defines an energy-minimal \"ideal\" driving profile. Experiments on the University of Michigan Vehicle Energy Dataset (VED), comprising 83 real-world EV trips from three Nissan Leaf vehicles (VehIds 10, 455, 541), demonstrate that the Hybrid model achieves an R² of 0.805 ± 0.166 versus ?0.176 ± 5.621 for the baseline ANN—a decisive advantage in structured sequence learning on real, noisy data. The per-trip slope-aware optimizer identifies energy-optimal cruise speeds ranging from 30 to 120 km/h (mean 56 km/h) conditioned on actual road gradients extracted from battery-power back-calculation. Ideal driving profiles reduce predicted energy consumption by 15–65% and extend range estimates by 40–250% (1.4× to 3.5×) relative to realistic pre-trip predictions, with ideal ranges spanning 80–258 km against predicted ranges of 69–107 km across test trips. These results validate the framework on real-world data and demonstrate its potential applicability to departure-time eco-driving advisory systems.

Introduction

The text describes a research approach to improving electric vehicle (EV) range prediction and driving efficiency planning before a trip begins, addressing “range anxiety” in battery electric vehicles.

Current EV range estimators react to past driving data, so they cannot warn drivers in advance about energy-heavy routes. This work proposes a pre-trip forecasting system that uses known factors—road slope, speed limits, traffic, weather, and driver behavior—to predict total trip energy consumption immediately after a destination is entered. It also suggests an optimal cruise speed that minimizes energy use.

The study uses real-world data from the University of Michigan Vehicle Energy Dataset (VED), which contains high-frequency telemetry (speed, GPS, battery state, HVAC load, etc.) from Nissan Leaf vehicles in real driving conditions. This real data captures realistic driving patterns like stop-and-go traffic, climate control usage, and road gradients, which synthetic datasets often miss.

A key challenge is that machine learning models are usually trained on real speed data but must operate pre-trip without knowing actual speed, causing a mismatch. The work addresses this by replacing real speed with a synthetic pre-trip driving proxy and aligning all features to avoid distribution shifts.

Methodologically, the system combines:

• A vehicle physics model to estimate power use and road slope from battery data
• Engineered features like temperature, HVAC load, SOC, acceleration, and estimated slope
• A hybrid LSTM–Transformer neural network to predict speed/energy patterns
• A baseline neural network for comparison
• Monte Carlo dropout to estimate prediction uncertainty
• A physics-based optimizer that finds the most energy-efficient constant cruising speed per trip using slope-aware energy modeling

The optimizer improves prior work by correctly separating auxiliary power from propulsion energy, avoiding bias that would incorrectly favor unrealistically slow speeds.

Results show that real-world trips produce asymmetric “energy vs. speed” curves, with an optimal efficiency zone typically around moderate speeds (e.g., ~30 km/h in a sample trip). Overall, the system demonstrates that reliable pre-trip EV range estimation and energy-efficient speed recommendations are possible using real telemetry data combined with physics-informed machine learning.

Conclusion

A pre-trip BEV energy prediction and eco-driving advisory pipeline has been presented and validated on the University of Michigan Vehicle Energy Dataset, comprising 83 real-world Nissan Leaf trips. The Hybrid LSTM-Transformer achieves R² = 0.805 ± 0.166 with MAE = 2.017 ± 1.276 m/s and calibrated MC Dropout uncertainty on real urban driving data, substantially outperforming a baseline ANN (R² = ?0.176) that fails to exploit sequential stop-and-go structure. The per-trip slope-aware cruise-speed optimizer, which excludes auxiliary power from its efficiency sweep to find the true aerodynamic-versus-stop-and-go valley, identifies energy-optimal cruise speeds of 30–120 km/h across the VED test trips a range that reflects real road-gradient diversity absent from synthetic studies. Ideal driving profiles reduce predicted Wh/km by 15–65% and extend range estimates by 40–250% (1.4× to 3.5×) relative to realistic pre-trip predictions, providing a physically meaningful upper bound for driver guidance rather than a directly actionable target. The framework operates from metadata available at departure (initial SOC, OAT, and a mapped route with elevation data); the current implementation derives road slope from the driven GPS trace via battery-power back-calculation, which is a post-hoc step that would need to be replaced by map-based elevation profiling in a production system. Future work should incorporate historical trip-pattern embeddings, multi-vehicle fleet learning, direct map-based slope profiles, battery degradation modelling, and segment-level speed advisory profiles constrained by posted speed limits.

References

[1] S. Oh et al., \"Vehicle Energy Dataset (VED), A Large-scale Dataset for Vehicle Energy Consumption Research,\" IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 4, pp. 3302–3312, Apr. 2022. [2] W. Kong et al., \"Hybrid deep learning for EV energy consumption forecasting,\" Applied Energy, vol. 340, p. 121042, 2023. [3] H. Shen, X. Zhou, H. Ahn, and J. Wang, \"Personalized Velocity and Energy Prediction for Electric Vehicles With Road Features in Consideration,\" IEEE Transactions on Transportation Electrification, vol. 9, no. 3, Sep. 2023. [4] H. Shen, Z. Wang, X. Zhou, K. Yang, P. Chen, J. Wang, and M. Lamantia, \"Electric Vehicle Velocity and Energy Consumption Predictions Using Transformer and Markov-Chain Monte Carlo,\" IEEE Transactions on Transportation Electrification, vol. 8, no. 3, Sep. 2022. [5] A. Upadhya and C. Mahanta, \"Improving Velocity Prediction in Electric Vehicles using Hybrid Artificial Neural Network (ANN),\" 2022 IEEE 10th Conference on Systems, Process & Control (ICSPC), Dec. 2022. [6] A. Kadechkar and X. Llauradó, \"AI-Enhanced Velocity Prediction for Efficient EV Energy Management with Hybrid Storage,\" 12th International Conference on Smart Grid, May 2024. [7] M. A. Eissa and P. Chen, \"An Efficient Hybrid Deep Learning Approach for Accurate Remaining EV Range Prediction,\" 2023 IEEE/ASME AIM, Jun. 2023. [8] L. Qu, W. Zhuang, and N. Chen, \"Instantaneous Velocity Optimization Strategy of Electric Vehicle Considering Varying Road Slope,\" 31st Chinese Control and Decision Conference, 2019. [9] A. Vaswani et al., \"Attention is all you need,\" Proc. NeurIPS, pp. 5998–6008, 2017. [10] S. Hochreiter and J. Schmidhuber, \"Long short-term memory,\" Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997. [11] Y. Gal and Z. Ghahramani, \"Dropout as a Bayesian approximation,\" Proc. ICML, pp. 1050–1059, 2016. [12] C. Fiori et al., \"Power-based EV energy consumption model,\" Applied Energy, vol. 168, pp. 257–268, 2016. [13] Y. Liu et al., \"Driving cycle prediction for EV energy management,\" Energy, vol. 235, p. 121257, 2021.

Copyright

Copyright © 2026 Y. Udaykiran, Dr. K. Krishnaveni. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.