CO₂ Emissions Prediction from Vehicle Features

Authors: Shashvat Jain
Affiliation: Integrated M.Tech. in Mathematics & Computing, IIT Dhanbad
GitHub: https://github.com/Shashvat-Jain/CO2-predictions-using-Automotive-Features

co2emissionsml

CO₂ Emissions Prediction from Vehicle Features
End-to-end Python package for analyzing and predicting on-road vehicle CO₂ emissions (g/km) via machine learning.

Features

• Preprocessing & Feature Engineering: scaling, one-hot encoding, target transformation
• Baseline Models: linear, polynomial, ridge/lasso, random forest, XGBoost, LightGBM, CatBoost
• Stacked Ensemble: LightGBM + XGBoost + CatBoost → MLP meta-learner → Ridge residual correction
• Bayesian Hyperparameter Tuning: Optuna pruners, early stopping
• Diagnostics & Explainability: parity plots, residual analysis, learning curves, permutation importance, SHAP

Key result:

Test set: (R^2 = 0.9830), MAE ≈ 3.08 g/km, RMSE ≈ 8.64 g/km

📦 Repository Structure

bash . ├── README.md ├── LICENSE ├── CITATION.cff ├── CODE_OF_CONDUCT.md ├── CONTRIBUTING.md ├── DATA_DICTIONARY.md ├── .gitignore ├── environment.yml ├── requirements.txt ├── setup.py ├── Dockerfile │ ├── data/ │ └── New Dataset.csv │ ├── notebooks/ │ └── co2-emissions-predict.ipynb │ ├── src/ │ ├──models │ └──co2_emissions_ml │ ├── __init__.py │ ├── preprocessing.py │ ├── models.py │ ├── evaluation.py │ └── pipeline.py │ ├── tests/ │ └── test_pipeline.py │ ├── scripts/ │ └── train_and_save.py │ ├── Figures/ │ ├── parity_plot.png │ ├── residual_hist.png │ ├── qq_plot.png │ ├── residuals_vs_pred.png │ ├── mae_decile.png │ ├── learning_curve.png │ ├── perm_importance.png │ ├── shap_summary.png │ ├── shap_dependence.png │ └── pipeline_diagram.png │ ├── Slides/ │ └── End Evaluation.pdf │ └── Reports/ ├── Split Report └── Final Report with plag report.pdf

⚙️ Installation

```bash

From PyPI

pip install co2emissionsml

Or install latest from GitHub

pip install git+https://github.com/Shashvat-Jain/CO2-predictions-using-Automotive-Features.git ```

Quickstart

1. Predict via CLI

bash run_co2 \ --data path/to/your_new_data.csv \ --model path/to/pretrained_bundle.pkl \ --output path/to/predictions.csv

• --data (required): input CSV with vehicle features

• --model (optional): path to serialized bundle.pkl (default: models/bundle.pkl)

• --output (optional): CSV path for predictions

• --target (optional): dependent variable name in input CSV

1. Programmatic API

```python import pandas as pd import joblib from co2emissionsml.models import predict_bundle

Load pre-trained bundle

bundle = joblib.load("models/bundle.pkl")

Prepare new data

dfnew = pd.readcsv("yournewdata.csv") Xnew = dfnew.copy()

Predict

dfnew["predictedCO2"] = predictbundle(bundle, Xnew) dfnew.tocsv("predictions.csv", index=False) ```

🚀 Usage of GitHub Repository

1. Prepare data Place New Dataset.csv under data/.

2. Run notebook Open and execute notebooks/co2emissionspredict.ipynb to reproduce EDA, model training, and evaluation.

3. Diagnostics & plots Generated in figures/:

   • Parity plot
   • Residual histogram & Q-Q plot
   • Learning curve
   • Permutation & SHAP importance charts

Note: The notebook co2emissionspredict.ipynb contains the complete code for the thesis whereas the src folder only contains the code for the new pipeline presented in this research.

📊 Results Snapshot

Figure: Figure:

📚 References

• Smith A., Jones B., Lee C. (2020). Random Forest–Based Prediction of Vehicle CO₂ Emissions. Int. J. Automotive Technol.

• Gupta R., Ramesh S. (2021). XGBoost Regression for Estimating Vehicle Emissions. IEEE Trans. Intelligent Vehicles.

• Tansini A., Pavlović I., Fontaras G. (2022). Forecasting CO₂ Emissions Using Ensemble, ML & DL. PeerJ.

• Zhao P., Zhang X., Li Y. (2023). Global Fuel- and Vehicle-Type-Specific CO₂ Emissions. Earth Syst. Sci. Data.

• Government of Canada (2024). Fuel Consumption Ratings. Open Gov. Portal.

• U.S. EPA (2022). 2022 EPA Automotive Trends Report. EPA-420-S-22-001.

• (See full bibliography in reports/.)

📄 License

This project is licensed under the MIT License. See LICENSE for details.