Thermal Comfort Prediction Using Machine Learning: A Comparative Study of Algorithms and Features

Predicting thermal comfort with precision is vital for enhancing indoor environments and promoting occupant well-being. As the focus on energy-efficient and occupant-focused buildings intensifies, the demand for accurate comfort modeling has grown. However, many existing models struggle to accommodate diverse environmental conditions and individual preferences. This study deploys predictive models using deep learning (DL) algorithms to solve challenges. For this purpose, the research utilizes a Multi-Input LSTM-Attention Based Deep Neural Network (MI-LSTM-ATTN) DL algorithm. The method begins by collecting global thermal comfort data. It performs preprocessing to remove missing or duplicate elements and compresses the dataset from 70 columns to 12 columns. Following this, wavelet scattering-based feature extraction is applied to enrich the dataset by capturing vital temporal and frequency-based attributes of thermal comfort parameters. Two feature selection methods are then used to determine the optimal feature sets, enhancing the accuracy of thermal preference predictions. The MI-LSTM-ATTN model is ultimately employed for classification, capitalizing on its ability to handle sequential data and highlight significant time-dependent patterns through the attention mechanism. D.The algorithm's performance is evaluated using statistical error measures such as the Root Mean Square error (RMSE), the Mean Square Error (MSE), the Mean Absolute Error (MAE), and the constant error (R²). According to the research results, the proposed method has excellent performance, with a maximum R² score of 0.982, MAE of 0.986, MSE of 0.048, and RMSE of 0.154. We achieved a maximum error of 1.398 using competitive error metrics.