Feature analysis and predictive modeling for enhancing occupant thermal comfort prediction in educational buildings using artificial neural networks and statistical analysis

Abstract

This study presents a comprehensive analysis of occupant thermal comfort prediction within educational buildings in Dhaka, Bangladesh, a city characterized by a tropical savanna (Aw) climate with high temperatures, humidity, and distinct wet and dry seasons. It integrates Artificial Neural Networks (ANNs) and statistical techniques, including one-way ANOVA, correlation analysis (bivariate), and linear regression. The research investigates 1670 field survey datasets to identify critical factors influencing occupants' thermal comfort levels. Employing Principal Component Analysis (PCA), Random Forest (RF), Lasso Regularization (LR), and Recursive Feature Elimination (RFE), the study evaluates feature importance and rankings. SMOTE-Tomek and SMOTE-ENN methods are utilized to address classification imbalances and enhance model performance. The analysis consistently emphasizes the significance of specific features such as floor level, lighting level, age, number of windows, room orientation, and gender in predicting occupant comfort levels. Furthermore, Temperature, Humidity, Room Orientation, Number of Fans, and Number of Lights exhibit varying importance in the predictive process. The study concludes by developing an accurate ANN-based classification model that predicts thermal comfort levels based on key features. The model performances stand out with notable accuracies: PCA achieved an accuracy of 0.98, RF demonstrated 0.97, LR yielded 0.98, and RFE exhibited an accuracy of 0.92. These high accuracies attest to the models' robustness in predicting thermal comfort levels within the tropical climate context, providing valuable insights for stakeholders seeking to optimize indoor comfort conditions in educational buildings and empowering them to make informed decisions about building design and management.