| Samia Akter Erin, Firuja Tasneem, Mim Mony, Mst. Umma Nourin Sawon, Mohammad Nyme Uddin CO2 Concentration Prediction in a Multi-Functional Semi-Open University Auditorium Using Regression-Based Machine Learning Models |
|---|
| Abstract. Carbon dioxide (CO2) concentration serves as a critical metric for evaluating air quality in densely populated urbanized regions like Dhaka, Bangladesh. In educational settings, where complex human activities occur in semi-open auditoriums, interactions can elevate CO2 levels. Prolonged high CO2 concentrations degrade air quality, leading to respiratory ailments, exacerbating climate change, and causing discomfort. While existing studies focus on outdoor or indoor air quality using Machine Learning (ML) models, research on semi-open educational spaces remains limited. This study aims to develop regression-based ML models to forecast CO2 levels in a university's semi-open auditorium in Dhaka, Bangladesh, featuring partial natural ventilation. 502 data samples from November 2025 to April 2026, encompassing 29 parameters (e.g., environmental, behavioral, demographic), were collected to predict CO2 concentrations. Three ML models, such as Decision Tree (DT), eXtreme Gradient Boosting (XGBoost), and Random Forest (RF), were deployed to forecast CO2 levels. GridSearchCV facilitated performance optimization and hyperparameter tuning, using metrics like R2 (coefficient of determination), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). SHapley Additive exPlanations (SHAP) analysis assessed feature importance and model transparency. The XGBoost model exhibited the highest R2 value (95%), followed by DT (93%) and RF (74%), with K-fold average after cross-validation. SHAP analysis identified HCHO and Light Intensity as key factors in CO2 prediction. This study contributes as a context-specific empirical and interpretable analysis, rather than a methodological innovation, offering insights into CO2 dynamics in semi-open environments. Future research should incorporate seasonal variations, real-time sensing, and robustness for broader applicability. |
| Keywords: CO2 Concentration, Semi-Open, University Auditorium, Machine Learning, Regression Models |
 Download PDF |
| DOI: https://doi.org/10.54381/itta2026.1.04 |