A Machine Learning Method for EV Range Prediction with Updates on Route Information and Traffic Conditions

Dohee Kim, Hong Gi Shim, Jeong Soo EoAAAI 2022 · DOI

Introduction: Addressing Range Anxiety

Despite the increasing capacity of battery packs, "range anxiety"—the psychological fear that an electric vehicle (EV) lacks sufficient energy to reach its destination—remains a significant hurdle for widespread EV adoption. Current "Distance to Empty" (DTE) metrics often rely on simplistic physics-based models or historical averages that fail to account for the dynamic nature of real-world driving.

The relationship between State of Charge (SoC) and driver anxiety as travel time increases.

Traditional methods struggle with three primary variables: volatile traffic conditions (stop-and-go vs. highway), non-driving loads (HVAC consumption in extreme weather), and route topography. To address these challenges, researchers from Hyundai Motor Group have developed a hybrid machine learning model that integrates real-time traffic data and navigation-based "preview" information to provide a far-sighted and reliable range estimate.

Methodology: The LSTM-DNN Hybrid Model

The proposed solution utilizes a two-stage LSTM-DNN mixture model designed to capture both temporal vehicle dynamics and static environmental states.

Flowchart of the machine learning model integrating route, traffic, and vehicle data for energy prediction.
  1. Stage 1: Long Short-Term Memory (LSTM) Network. This stage processes time-series profiles of predicted speed and acceleration derived from Intelligent Transportation System (ITS) data. LSTMs are uniquely suited for this task as they can learn patterns in sequential data, effectively "looking ahead" at the upcoming route.
  2. Stage 2: Deep Neural Network (DNN). The features extracted by the LSTM are combined with "static" or slowly changing variables, including the current State of Charge (SOC), outdoor temperature, and auxiliary electric loads (HVAC).

A critical innovation in this research is the Distance per Battery energy consumption (DPB) index. By categorizing driving data into "Normal Load" (NL) for mild seasons and "High Load" (HL) for extreme weather, the researchers trained specialized models that account for the erratic nature of HVAC energy consumption.

Experimental Setup and Datasets

The model was trained and validated using a substantial real-world dataset collected from a Hyundai Kona EV. Over the course of one year, approximately 160,000 km of driving data (roughly 1,000 hours) was recorded. This data was processed using a 40-minute sliding window with a 3-minute shift, ensuring the model captured the correlations between adjacent driving segments and varying traffic environments.

Key Results and Performance

The hybrid model demonstrated significant improvements over traditional DTE estimation methods:

  • High Precision: The model achieved a range prediction accuracy of 2–3 km within a 40-minute driving window.
  • Seasonal Specialization: The specialized "Normal Load" model showed a 5.2% improvement over a general-purpose model, while the "High Load" model (handling extreme winter/summer conditions) improved accuracy by 6.8%.
  • Robustness: The LSTM-DNN architecture proved stable and converged successfully across over 40,000 data windows, handling diverse map and traffic conditions effectively.

Discussion and Future Directions

By treating traffic "preview" information as a primary feature, the model can adjust range predictions before a vehicle even enters a congested area. This proactive approach is a significant step forward from reactive, physics-based models.

However, several areas for future work remain. Current models do not explicitly account for battery State of Health (SOH) or long-term degradation. Furthermore, deploying these complex ML models in production vehicles requires optimization for embedded control boards with limited computational resources. Future iterations may also include "continual learning" frameworks, allowing the vehicle to personalize its predictions based on an individual's unique driving style over time.

Summary updated

BibTeX
@inproceedings{kim2022machine,
  author = {Dohee Kim and Hong Gi Shim and Jeong Soo Eo},
  title = {{A Machine Learning Method for EV Range Prediction with Updates on Route Information and Traffic Conditions}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21525},
}
Figure