Personalized electric vehicle energy consumption estimation framework that integrates driver behavior with map data

1. Problem Statement

The transition to Battery Electric Vehicles (BEVs) is hindered by "range anxiety," largely caused by inaccurate predictions of remaining driving range and State-of-Charge (SOC). Existing estimation methods typically rely on:

• Recent consumption averages.
• Simplified driving assumptions.
• Limitations: These approaches fail to account for critical contextual factors such as road geometry (grade, curvature), traffic controls, and, most importantly, individual driver behavior. Aggressive driving styles can reduce EV range by 20–40% compared to conservative drivers. Current literature lacks comprehensive frameworks that jointly encode driver-specific behavior and future road geometry for long-horizon prediction.

2. Methodology

The authors propose a modular, seven-stage framework that integrates map-derived context, personalized driver modeling, and physics-based energy consumption. The system operates as an end-to-end pipeline:

A. Data Acquisition and Preprocessing

1. Map Display Handler: Uses Selenium to interact with OpenStreetMap (OSM) to retrieve route geometry (GeoJSON) based on user-selected start and end points.
2. Route Processor: Interfaces with the Valhalla routing engine via RESTful APIs to enrich the raw GPS coordinates with detailed road features, including speed limits, curvature, slope (elevation), intersection density, and traffic control signals.

B. Velocity Prediction Pipeline

3. Velocity Profile Generator: Creates a rule-based reference velocity trajectory. It enforces legal speed limits, reduces speed for high-curvature segments, and sets velocity to zero at traffic signals and stop signs.
4. PID Controller: Simulates realistic vehicle dynamics (acceleration, jerk limits) by tracking the reference profile. This generates a physically plausible "past driving" dataset (velocity, acceleration, jerk) that serves as the input context for the machine learning model.
5. Driver LSTM Module (Core Innovation):
   • Architecture: A Bidirectional LSTM (BiLSTM) encoder-decoder hybrid.
   • Inputs:
     • Past Context: Vehicle states (velocity, acceleration) and past road features.
     • Future Context: Future road features (elevation, speed limits, curvature, traffic signals) derived from the map data.
   • Mechanism: The model learns the temporal dependencies between a driver's historical behavior and upcoming road conditions. It predicts future velocity profiles that reflect individual driving styles (e.g., anticipation of curves, reaction to traffic lights).
   • Training: Trained on the IEEE Vehicle Speed Dataset (5973 rides) with enhanced road features.
   • Output: A personalized, noise-filtered velocity profile.

C. Energy Estimation

6. BEV Energy Consumption Model: A quasi-steady backward physics-based model (based on the VT-CPEM framework).
   • Calculations: Computes tractive power by summing aerodynamic drag, rolling resistance, grade forces, and inertial forces.
   • Regenerative Braking: Models efficiency as a function of deceleration.
   • SOC Evolution: Integrates battery power (accounting for motor/driveline efficiency and auxiliary loads) over time to predict SOC depletion.
7. Visualization Module: Generates comparative plots of velocity, power, and SOC trajectories against map features for validation.

3. Key Contributions

• Personalized Driver Modeling: Unlike previous works that aggregate data across drivers, this framework explicitly encodes driver identity and style using a BiLSTM, capturing stochastic behaviors like micro-adjustments in speed and anticipation of road geometry.
• Map-Context Integration: The system uniquely combines future road features (look-ahead) with driver history, allowing the model to predict how a specific driver will react to upcoming hills, curves, or intersections.
• Hybrid Architecture: It bridges the gap between data-driven prediction (LSTM) and physics-based simulation (PID controller and energy model), ensuring that predictions are both behaviorally realistic and physically consistent.
• End-to-End Pipeline: A fully automated system from route selection to SOC estimation, utilizing open-source tools (OSM, Valhalla) and standard machine learning libraries.

4. Results and Evaluation

The framework was tested on three distinct route types: Urban, Freeway, and Hilly Terrain.

• Urban Routes: The model successfully captured deceleration trends near intersections and traffic signals. While it did not always predict a full stop (0 km/h) due to training data nuances and limited look-ahead, it accurately modeled the deceleration behavior and speed variations typical of real-world driving, unlike the rigid PID reference.
• Freeway Routes: The model predicted velocity oscillations around the speed limit, reflecting individual driver micro-adjustments, whereas the reference PID profile remained perfectly smooth. This resulted in more realistic power fluctuation profiles.
• Hilly Terrain: The system accurately modeled grade-dependent responses. It predicted higher energy consumption and faster SOC depletion during climbs and significant regenerative braking (negative power) during descents.
• Accuracy: The resulting SOC trajectories aligned with physical expectations, demonstrating that the combination of learned behavior and map context significantly improves prediction realism compared to standard average-based methods.

5. Significance

• Mitigating Range Anxiety: By providing accurate, personalized range estimates that account for how a specific driver drives on specific roads, the system reduces uncertainty for EV owners.
• Eco-Routing and Navigation: The framework enables the development of navigation systems that can suggest routes based on energy efficiency tailored to the driver's specific style, rather than just distance or time.
• Driver-Adaptive Management: The ability to predict energy consumption based on individual behavior allows for better battery management strategies and charging planning.
• Scalability: The modular design allows for future integration of advanced architectures (e.g., Transformers) and more granular efficiency maps, making it a robust foundation for next-generation EV energy management systems.