Vehicle CO₂ Emissions Prediction

Author: Mohammad Taha — November 2025

Project Overview

This project develops a high-precision regression model to predict vehicle CO₂ emissions based on technical specifications such as engine size, fuel consumption, and vehicle class. Using data from Transport Canada, the final Random Forest Regressor achieves an exceptional R² of 0.9972 on the test set, with an average prediction error of less than 2 g/km.

The model provides actionable insights for environmental policy, vehicle purchasing decisions, and automotive engineering, demonstrating how machine learning can support sustainability goals.

Environmental Context

Transportation is a major contributor to greenhouse gas emissions. Accurate prediction of vehicle CO₂ emissions enables:

• Informed consumer choices: Help buyers select environmentally friendly vehicles
• Regulatory compliance: Support emission standards enforcement
• Automotive R&D: Guide manufacturers in designing cleaner vehicles
• Carbon footprint analysis: Enable accurate environmental impact assessments
• Policy evaluation: Test effectiveness of emission reduction strategies

Dataset

Source: Transport Canada Vehicle Emissions Database (via Kaggle)
Size: 7,385 vehicles
Coverage: Multiple model years and manufacturers
Quality: Complete data with no missing values

Key Features

Feature	Type	Description
Make	Categorical	Vehicle manufacturer (e.g., Toyota, Ford, BMW)
Model	Categorical	Specific vehicle model
Vehicle Class	Categorical	Body type (SUV, Sedan, Pickup, Compact, etc.)
Engine Size(L)	Numeric	Engine displacement in liters
Cylinders	Numeric	Number of engine cylinders
Transmission	Categorical	Transmission type and number of gears
Fuel Type	Categorical	Gasoline, Diesel, E85, etc.
Fuel Consumption City (L/100 km)	Numeric	City driving fuel consumption
Fuel Consumption Hwy (L/100 km)	Numeric	Highway driving fuel consumption
Fuel Consumption Comb (L/100 km)	Numeric	Combined fuel consumption
Fuel Consumption Comb (mpg)	Numeric	Combined fuel economy (miles per gallon)
CO2 Emissions(g/km)	Numeric (Target)	Carbon dioxide emissions

Technologies and Tools

• Python 3.x: Core programming language
• scikit-learn: Machine learning models and pipelines
• pandas: Data manipulation and analysis
• NumPy: Numerical computing
• Matplotlib & Seaborn: Data visualization
• KaggleHub: Dataset integration
• Jupyter Notebook: Interactive development

Methodology

1. Data Exploration and Cleaning

• Completeness check: Verified no missing values
• Duplicate handling: Preserved intentionally (reflect market distribution)
• Outlier analysis: Identified but retained (represent real extreme cases)
• Distribution analysis: Examined target variable and feature relationships

2. Feature Engineering

Created three powerful engineered features:

• engine_per_cylinder: Engine size divided by number of cylinders

  • Captures engine design efficiency
  • Larger values indicate bigger cylinders → higher emissions

• city_hwy_ratio: City fuel consumption / Highway fuel consumption

  • Measures urban inefficiency
  • Higher values indicate vehicles poorly optimized for city driving

• total_consumption: Sum of city and highway consumption

  • Aggregate fuel usage metric

3. Data Preprocessing Pipeline

• Categorical encoding: One-hot encoding for Make, Model, Vehicle Class, Transmission, Fuel Type
• Feature scaling: StandardScaler for numeric features (essential for SVR)
• Train/Validation/Test split: 60% / 20% / 20%
• Pipeline construction: Integrated preprocessing and modeling

4. Model Training and Comparison

Four regression algorithms were evaluated:

Linear Regression

• Baseline model
• Surprisingly effective (R² = 0.9906)
• Indicates near-linear relationship in transformed feature space

Random Forest Regressor (Best Model)

• Ensemble of 100 decision trees
• Handles non-linear relationships
• Robust to outliers and feature interactions

Support Vector Regressor (SVR) with RBF Kernel

• Non-linear mapping via kernel trick
• Requires careful feature scaling

Gradient Boosting Regressor

• Sequential ensemble learning
• Strong performance but slightly behind Random Forest

Results

Model Performance Comparison

Model	R² (Train)	RMSE (Train)	MAE (Train)	R² (Val)	RMSE (Val)	MAE (Val)
Linear Regression	0.9966	3.43	1.83	0.9906	5.64	3.24
Random Forest	0.9995	1.32	0.80	0.9960	3.69	1.99
SVR (RBF)	0.9891	6.14	1.80	0.9901	5.82	3.33
Gradient Boosting	0.9975	2.92	2.29	0.9954	3.97	2.47

Final Test Set Performance

The Random Forest Regressor achieved:

• R² (Test): 0.9972
• RMSE (Test): 3.03 g/km
• MAE (Test): 1.94 g/km

Key Insights

• Exceptional predictive accuracy: R² > 0.99 indicates CO₂ emissions are highly predictable from vehicle specifications
• Low prediction error: Average error under 2 g/km is excellent for practical applications
• Strong generalization: Test performance matches validation performance
• Feature importance: Combined fuel consumption is the dominant predictor, followed by engine design metrics

Feature Importance Analysis

The Random Forest model identified the most influential predictors:

1. Fuel Consumption Comb (L/100 km) — Dominant factor (directly proportional to CO₂)
2. engine_per_cylinder — Engine design efficiency metric
3. city_hwy_ratio — Urban driving inefficiency indicator
4. Vehicle Class — Body type and weight category
5. Fuel Type — Different fuels have different carbon content per liter

These findings align perfectly with physical and engineering principles, increasing confidence in the model's reliability.

Project Structure

vehicle-co2-emissions-prediction/
├── vehicle_co2_emissions_prediction.ipynb  # Main analysis notebook
├── README.md                               # This file
├── LICENSE                                 # MIT License
├── requirements.txt                        # Python dependencies
└── .gitignore                              # Git ignore rules

How to Run

Prerequisites

• Python 3.8+
• Minimum 4GB RAM

Installation

1. Clone the repository:

   git clone https://github.com/tahamohmadf19-dev/vehicle-co2-emissions-prediction.git
   cd vehicle-co2-emissions-prediction

2. Install dependencies:

   pip install -r requirements.txt

3. Run the notebook:

   jupyter notebook vehicle_co2_emissions_prediction.ipynb

Running on Google Colab

1. Upload the notebook to Google Colab
2. The dataset will be automatically downloaded via KaggleHub
3. Run all cells sequentially

Real-World Applications

For Consumers

• Compare emissions of different vehicles before purchase
• Estimate environmental impact of driving habits
• Calculate carbon footprint for personal sustainability goals

For Policymakers

• Evaluate effectiveness of emission standards
• Identify vehicle categories requiring stricter regulations
• Design incentive programs for low-emission vehicles

For Manufacturers

• Benchmark new designs against market standards
• Optimize engine and transmission configurations
• Predict regulatory compliance before production

Future Enhancements

• Temporal analysis: Incorporate model year trends to track emission improvements over time
• Electric vehicle integration: Extend model to handle EVs and plug-in hybrids
• Uncertainty quantification: Implement prediction intervals using quantile regression
• Interactive web application: Deploy as a user-friendly emissions calculator
• API development: Create REST API for integration with automotive databases
• Multi-pollutant modeling: Extend to NOx, particulate matter, and other emissions
• Causal inference: Use causal models to identify design changes that reduce emissions

Educational Value

This project demonstrates:

• End-to-end regression modeling workflow
• Feature engineering for domain-specific problems
• Model selection and hyperparameter tuning
• Interpretability and validation of ML models
• Application of machine learning to environmental challenges
• Best practices in data science project documentation

Limitations

• Geographic scope: Data represents Canadian market; may not generalize to other regions
• Temporal scope: Does not account for technological evolution over time
• Correlation vs causation: Model captures empirical patterns, not causal mechanisms
• Real-world driving: Test cycle data may differ from actual on-road emissions

License

This project is licensed under the MIT License - see the LICENSE file for details.

References

• Transport Canada Vehicle Emissions Database
• Kaggle Dataset: CO2 Emission by Vehicles
• scikit-learn Documentation
• IPCC Guidelines for Greenhouse Gas Inventories

Contact

Mohammad Taha
📧 Email: tahamohmadf19@gmail.com
🔗 LinkedIn: Mohmad Taha Alhmad
🐙 GitHub: @tahamohmadf19-dev

---

This project contributes to environmental sustainability by making vehicle emissions transparent and predictable through data science.