Project for RASIO 2025 Competition Data Science Category by Padjadjaran University View the Project Presentation Here
As the industrial sector continues to grow rapidly, it becomes increasingly important to accurately predict daily CO₂ emission levels. Reliable forecasts are essential for developing effective environmental policies and public health strategies. This project was developed as part of a data science competition, leveraging advanced data science methodologies to build a robust predictive model for daily CO₂ emissions.
Comprehensive data preprocessing was performed to clean and prepare the dataset for modeling. A key component of our approach was feature engineering, where we extracted and created various time-based features. These features were critical in allowing the models to capture underlying temporal patterns and seasonal trends in the emission data.
To maximize predictive accuracy and ensure strong generalization, we implemented an out-of-fold stacking ensemble approach.
We initially experimented with nine different algorithms: CatBoost, XGBoost, Gradient Boosting, LightGBM, Random Forest, Extra Trees, Ridge, Lasso, and Elastic Net. Based on their individual performances, we excluded Lasso and Elastic Net, moving forward with the top 7 models as our base learners:
- CatBoost
- XGBoost
- Gradient Boosting
- LightGBM
- Random Forest
- Extra Trees
- Ridge Regression
The predictions from the 7 base models were combined using a meta-learner. We tested several algorithms for this role, including ElasticNet, Ridge, Lasso, and Multi-Layer Perceptron (MLP). Ridge Regression achieved the best performance as the meta-model and was selected for the final architecture.
Our stacked ensemble model demonstrated strong generalization performance and highly reliable predictive capability, earning a top-tier placement in the competition.
- Training RMSE: 1.6933
- Private Leaderboard RMSE: 1.953