Using ML and data science to analyze emission factors and pinpointing technical characteristics of cars which contribute to pollution.
==============================
==============================
This project aims to analyze vehicle CO2 emissions using machine learning and data science techniques. It includes exploratory data analysis, modeling, and visualization techniques to derive insights from the data and predict emission levels based on vehicle characteristics.
- Car emissions remain the highest compared to other sectors in the EU.
- Using ML and data science to analyze emission factors.
- Pinpointing technical characteristics of cars which contribute to pollution.
- Preventing the introduction & development of high-pollution vehicle types
- Avoiding penalties for non-compliance with CO₂ regulations
- Reducing redesign costs by optimizing vehicle development
- Efficient production strategies
- Enabling competitive pricing
- Boosting brand reputation with eco-friendly recognition
- Data mining, engineering, select features, train, and evaluate models
- Identifying emission-related data patterns
- Focus on CO₂ emissions during operation
- Correlate vehicle features with CO₂ emissions
- Predict emissions from technical characteristics
==============================
├── LICENSE
├── README.md <- The top-level README for developers using this project.
├── data <- Should be in your computer but not on Github (only in .gitignore)
│ ├── processed <- The final, canonical data sets for modeling.
│ └── raw <- The original, immutable data dump.
│
├── models <- Trained and serialized models, model predictions, or model summaries
│
├── notebooks <- Jupyter notebooks. Naming convention is a number (for ordering),
│ the creator's name, and a short `-` delimited description, e.g.
│ `1.0-alban-data-exploration`.
│
├── references <- Data dictionaries, manuals, links, and all other explanatory materials.
│
├── reports <- The reports which summarize the project
│ └── figures <- Generated graphics and figures to be used in reporting
│
├── requirements.txt <- The requirements file for reproducing the analysis environment, e.g.
│ generated with `pip freeze > requirements.txt`
│
├── src <- Source code for use in this project.
│ ├── __init__.py <- Makes src a Python module
│ │
│ ├── features <- Scripts to turn raw data into features for modeling
│ │ └── build_features.py
│ │
│ ├── models <- Scripts to train models and then use trained models to make
│ │ │ predictions
│ │ ├── predict_model.py
│ │ └── train_model.py
│ │
│ ├── visualization <- Scripts to create exploratory and results oriented visualizations
│ │ └── visualize.py
==============================
The project uses the EU (EEA) dataset "CO2 emissions from new passenger cars" from 2010-2023, covering 30 countries. Originally over 16 GB, it was reduced via data transformation. Source
==============================
-
Clone the repository:
git clone https://github.com/TilloStralka/CO2_Emission_Predictor.git cd CO2_Emission_Predictor -
Create a virtual environment and activate it:
python -m venv env source env/bin/activate # On Windows: env\Scripts\activate
-
Install dependencies:
pip install -r requirements.txt
==============================
The dataset required for this project is not included in the repository because it is listed in the .gitignore file.
Please download the original dataset from the specified source and save it in the data/raw/ folder. Ensure the file is named as required for seamless processing.
Use the Jupyter Notebook located in the notebooks/ folder to process the data:
- Notebook:
DataPipeline.ipynb - Default Settings: The callable functions in this notebook have default parameter values, but these can be adjusted as needed.
- Further Details: For more in-depth information about the processing pipeline, refer to Report #3 in the
reports/folder.
After processing, the finalized dataset will have the following properties:
- Name:
minimal_withoutfc_dupesdropped_frequencies_area_removedskew_outliers3_0_NoW_tn20_mcp00.10.parquet(cryptic title) - Size: 2,000,450 rows, 56 columns
- Memory Usage: 282.4 MB
To perform the modeling, use the notebook:
- Notebook:
ModelPipeline.ipynb - This notebook contains all the necessary steps, fully commented for clarity.
The functions called in the notebooks are implemented in the utils_co2.py file. You can review this file for additional details about the underlying implementation.
After training, the trained models are saved in the models/ folder for future use.
For any modifications to the processing or modeling pipeline, ensure that changes are consistent across all related files. For further assistance, consult the documentation and comments within the notebooks.
==============================
The following metrics were evaluated across three models:
-
XG Boost (Best Overall Performance):
- Test MSE (scaled): 0.02
- Test MSE: 25.1
- Test RMSE: 5.0
- Test MAE: 3.7
- Test R²: 0.97
- CV R²: 0.97
- Training time: 1.41 mins
-
Dense Neural Network:
- Test MSE (scaled): 0.06
- Test MSE: 64.0
- Test RMSE: 8.0
- Test MAE: 6.2
- Test R²: 0.93
- Training time: 0.26 mins
-
Linear Regression:
- Test MSE (scaled): 0.12
- Test MSE: 120.5
- Test RMSE: 11.0
- Test MAE: 8.3
- Test R²: 0.87
- CV R²: 0.87
- Training time (mins): 7.07
-
XGBoost emerged as the optimal choice, demonstrating:
- Superior predictive accuracy with lowest error metrics
- Excellent handling of non-linear relationships
- Strong balance between performance and computational efficiency
-
Dense Neural Network showed promise with:
- Fast training speed
- Strong R² score of 0.93
- Potential for improvement through overfitting prevention
-
Linear Regression, while most interpretable:
- Provided baseline performance
- Limited by linear assumptions
- Required longest training time
-
XGBoost:
- Achieved the best predictive accuracy.
- Lowest Mean Squared Error (MSE) and highest R-squared.
- Effectively captures non-linear relationships.
- Provides an optimal balance of performance and efficiency.
-
Dense Neural Network (DNN):
- A strong contender with fast training and high R-squared.
- Requires further fine-tuning to prevent overfitting.
-
Linear Regression:
- Offers simplicity and interpretability.
- Underperforms relative to XGBoost and DNN.
Recommendation: XGBoost
- Ensemble Learning: Combining Linear Regression, XGBoost, and Dense Neural Networks for improved predictions.
- Hyperparameter Tuning: Leveraging techniques like Bayesian Optimization for XGBoost and DNN.
- Advanced Architectures: Experimenting with CNNs or Residual Connections to boost model performance.
- Data Augmentation: Using methods like SMOTE to address class imbalances.
==============================
This project was developed as part of the Data Scientist certification program at DataScientest.com, certified by Université Paris Panthéon-Sorbonne. The following team members contributed equally to this project:
==============================
This project is licensed under the MIT License.