Climate-specific machine learning for thermal sensation prediction of buildings in hot desert climates

Abstract

Thermal comfort is essential for the well-being, productivity, and energy efficiency of building occupants. The predictability of PMV does not always ensure accurate TSV, particularly in arid desert climates, where generalized models show limited precision. Many models were developed using globally aggregated data and frequently overlook climate-specific factors, potentially diminishing their relevance in various climatic locations. Differently, this study introduces novel preprocessing, climate-specific machine learning models, and ablation analyses for predicting TSV in hot desert areas, derived from the ASHRAE Database, which includes 6283 observations. A two-stage elimination method that integrates FIS and thermophysiological suitability was utilized for variable selection, while KNN-based imputation addressed missing values to preserve the multivariate structure. RF, XGBoost, and LightGBM models were trained for both 7-point including cold, cool, slightly cool, neutral, slightly warm, warm and hot, and 3-point including hot, neutral, and cold TSV classes. The results indicated that the RF model excelled in generalization performance across 7-point classes, achieving a Macro-F1 value of 0.6307 and an accuracy of 70.80%, with a Cohen's kappa of 0.572. Upon reducing the target variable to 3 classes, predictive reliability improved significantly, with the RF reaching a Macro-F1 of 0.7927 and an accuracy of 87.83%, alongside a kappa of 0.6843. The RF model demonstrated about a twofold increase in predictability compared to the traditional PMV for 7-point classes. Incorporating contextual variables such as city, season, and building type improved the Macro-F1 value from 0.520 to 0.605, highlighting the necessity of local context in TSV within the examined climate.