Optimizing energy efficiency and occupant comfort in smart buildings using SMOTE-augmented deep learning approaches

The architecture and construction sector is a major consumer of energy and a significant contributor to greenhouse gas emissions, emphasizing the need for intelligent, energy-efficient, and sustainable building systems. Although deep learning techniques have shown strong potential for thermal comfort prediction and building energy management, existing studies remain limited by localized experimental settings, class imbalance, and insufficient comparative evaluation of advanced deep learning models, restricting their practical applicability.

This study aims to develop a robust deep learning framework for optimizing energy efficiency and occupant thermal comfort by addressing class imbalance and improving prediction accuracy.

A unified framework incorporating data visualization, missing value treatment, label encoding, feature normalization, and the Synthetic Minority Over-sampling Technique (SMOTE) was developed. Multiple deep learning architectures, including Deep Neural Networks (DNN), Deep Flatten DNN, Bi-LSTM, and Attention-based LSTM, were comparatively evaluated. The novelty of the proposed framework lies in integrating SMOTE-based class balancing with the comparative evaluation of multiple deep learning architectures within a unified prediction framework.

The Attention-based LSTM model achieved the highest thermal comfort prediction accuracy of 91%, outperforming the Bi-LSTM (89%), Deep Flatten DNN (88%), and DNN (87%). It also achieved superior Precision, Recall, and F1-score, with an 8-percentage-point improvement over the previously reported GNN model on the same dataset.

The proposed SMOTE-augmented framework improves the accuracy and robustness of thermal comfort prediction, providing an effective solution for intelligent, energy-efficient, and sustainable building energy management with strong potential for real-world applications.