kevinkyi/EV_Battery_Consumption_XGBoost

EcoRoute: Universal EV Energy Predictor

Model Description

EcoRoute is a physics-based machine learning model designed to predict the Instantaneous Energy Consumption (kW) of any electric vehicle based on dynamic environmental conditions.

Unlike traditional range estimators that rely on simple averages, this model uses XGBoost to learn the non-linear physics of driving—specifically how Gravity (Slope), Aerodynamics (Speed), and Thermodynamics (Temperature) impact battery drain in real-time.

• Model Type: XGBoost Regressor
• Task: Tabular Regression (Physics Modeling)
• Input Features: Vehicle Speed, Road Slope (Gradient), Ambient Temperature, Vehicle Weight, Acceleration.
• Target: Instantaneous Power (kW)

Intended Use

This model is the backend brain for the EcoRoute Optimizer Application. It is designed to:

• Predict energy usage for precise route segments (e.g., every 10 meters).
• Account for Regenerative Braking on downhill slopes.
• Adjust predictions based on the specific Vehicle Weight (making it universal to any EV).
• Provide "Explainable AI" insights (e.g., "High consumption due to steep 5% grade").

Training Data

The model was trained on the Extended Vehicle Energy Dataset (eVED), a rigorous enhancement of the IEEE-standard VED dataset.

• Source: Real-world OBD-II sensor telemetry from electric vehicles in Ann Arbor, Michigan.
• Enhancements: The raw GPS data was map-matched to OpenStreetMap to append verifiable Road Gradient (Slope) and Speed Limit data.
• Data Volume: Trained on ~62,000 verified moments (second-by-second snapshots) of active EV driving.
• Filtering: Filtered to exclude idling (speed < 1 km/h) and sensor noise to ensure the model learns pure driving physics.

Performance Metrics

The model was evaluated on a held-out test set (20% split) of real-world driving logs.

Metric	Score	Interpretation
R² Score	0.731	Explains 73.1% of the variance in instantaneous energy usage (High for raw sensor data).
MAE	4.71 kW	On average, predictions are within ~4.7 kW of the actual sensor reading.

Note: While 0.73 is the per-second accuracy, the aggregate accuracy over a full trip (summing thousands of seconds) typically exceeds 90% due to error cancellation.

Input Feature Dictionary

To use this model, your input dataframe must contain these exact columns in this order:

1. Speed_Smooth (km/h): Rolling average speed (removes GPS jitter).
2. Road_Slope_pct (%): The gradient of the road (positive = uphill, negative = downhill).
3. Ambient_Temp_C (°C): Outside air temperature.
4. Weight_kg (kg): Curb weight of the vehicle + passengers.
5. Acceleration_m_s2 (m/s²): Change in speed over time.

Limitations

• Vehicle Year: Trained on data from 2017-2018. While physics (gravity/drag) doesn't change, modern EVs may have slightly more efficient motors.
• Extreme Weather: The training data has fewer samples in extreme sub-zero (< -10°C) temperatures.

Citations

• eVED Dataset: Zhang et al., "Extended vehicle energy dataset (eVED): an enhanced large-scale dataset for deep learning on vehicle trip energy consumption," arXiv:2203.08630, 2022.
• VED Dataset: Oh et al., "Vehicle Energy Dataset (VED), A Large-scale Dataset for Vehicle Energy Consumption Research," IEEE Transactions on Intelligent Transportation Systems, 2020.