Predicting Passivhaus certification of dwellings using machine learning: A comparative analysis of logistic regression and gradient boosting decision trees

Abstract

Passivhaus has gained recognition as a reliable solution for low-energy residential housing. However, existing studies lack comprehensive quantitative analysis of real-life cases, restricting the availability of evidence and experience from certified buildings. This study addresses this gap by examining the relationship between characteristics of certified passive houses and their energy performance. The objective is to establish dependable statistical models and best practices for early design stages of passive houses. The study analyzes a significant number of certified cases (n = 785) in residential settings situated in two climate zones: Temperate, no dry season, warm summer (n = 328, 41.78%) and Cold, no dry season, warm summer (n = 457, 58.22%). To assess whether residential buildings meet passive house energy standards, models using logistic regression and gradient boosting decision trees were developed employing three boosting methods: XGBoost, LightGBM, and CatBoost. Results reveal that CatBoost (F1 = 0.73, Logloss = 0.63, AUC = 0.78, Brier = 0.22) outperforms logistic regression (F1 = 0.72, Logloss = 0.62, AUC = 0.72, Brier = 0.22) in terms of accuracy. Additionally, LightGBM (F1 = 0.79, Logloss = 0.58, AUC = 0.75, Brier = 0.20) exhibits superior predictive calibration. This study indicates that external wall, floor slab, and roof thermal performance are crucial factors for meeting passive house standards (P-value = 0.0046). The research also discusses innovative practices and technologies derived from the certified cases. Overall, this study provides valuable insights for certification systems and early design strategies.