Abstract
Accurately predicting vehicle energy consumption is critical for improving electric vehicle (EV) range estimation and developing green routing algorithms. This paper introduces the Extended Vehicle Energy Dataset (eVED), a significant enhancement of the original Vehicle Energy Dataset (VED). By applying map-matching techniques to resolve noisy GPS data and integrating external environmental context—such as road elevation, speed limits, and traffic signals—the authors provide a high-fidelity, "deep-learning-ready" resource for the automotive research community.
Introduction
While internal vehicle sensors provide a wealth of data regarding speed, RPM, and temperature, they often lack the environmental context necessary to explain complex energy consumption patterns. The original VED, while massive in scale, suffered from jittery GPS coordinates and a lack of information regarding the physical road environment. The eVED addresses these gaps by transforming raw sensor logs into a context-rich dataset that accounts for road grade, speed limits, and discrete road events like intersections and stop signs.
Methodology: From Raw Logs to eVED
The authors developed a multi-step pipeline to enrich the raw VED data into the enhanced eVED format.
Map Matching
To correct the noisy GPS data produced by on-board diagnostic (OBD) sensors, the team utilized a Hidden Markov Model (HMM) via the Valhalla routing engine. This process "snaps" raw, jittery coordinates to the most probable road segments based on temporal sequences, ensuring that every data point is accurately localized on the map.
Data Enrichment and Feature Engineering
Once spatial accuracy was achieved, the dataset was cross-referenced with external sources:
- OpenStreetMap (OSM): Used to append speed limits and road classifications.
- Google Maps Elevation API: Used to add altitude data, which was subsequently smoothed to calculate accurate road gradients.
- Focus Points: The authors identified nearly 30,000 stop signs and 200,000 traffic signals, recognizing that these discrete infrastructure elements often dictate energy spikes more than steady-state driving.
Dataset Overview
The resulting eVED dataset contains over 20 million records from 383 vehicles operating in Ann Arbor, Michigan. As shown in the comparison below, eVED (built upon the VED foundation) offers a superior combination of record volume and feature richness compared to other common research datasets.
Key Results and Insights
The utility of eVED was validated through three primary use cases:
- Energy Estimation: A statistical approach based on speed distributions estimated trip energy consumption with a Root Mean Squared Error (RMSE) of 0.0781 liters of gas, showing high alignment with ground truth.
- Traffic Tension Analysis: Spatial heat maps revealed clear urban "tension" zones during peak rush hours, validating the dataset's ability to capture real-world city dynamics.
- Speed Prediction via LSTM: A Long Short-Term Memory (LSTM) network trained using only environmental features (time of day, gradient, road type) achieved an RMSE of 8.9078 km/h for predicting median trip speeds. This proves that environmental context alone is a powerful predictor of vehicle behavior.
Conclusion
The eVED dataset represents a shift toward context-aware automotive modeling. By providing accurate spatial data and environmental features, it enables more sophisticated deep learning applications in eco-routing and carbon footprint analysis. While currently focused on Ann Arbor, the methodologies used to create eVED are universal and can be applied to future large-scale vehicle data collection efforts.
Summary updated