| Title |
Prediction of Annual Energy Loads Based on Machine Learning for Aging Houses with Green Remodeling Focused on Impact Analysis by Variable Indexes of Responsive Facades |
| Authors |
김재향(Kim, Jae-Hyang) ; 한승훈(Han, Seung-Hoon) |
| DOI |
https://doi.org/10.5659/JAIK.2026.42.4.305 |
| Keywords |
Aging Houses; Responsive Facade; Machine Learning; Regression Model; Green Remodeling |
| Abstract |
Aging detached houses exhibit non-uniform envelope performance due to deteriorated insulation and airtightness and construction-related
variability, and the energy performance of retrofit alternatives varies substantially depending on combinations of key design variables such as
house type, window Type, insulation level, and building orientation. Heat-balance-based simulation (EnergyPlus) can reflect these multivariable
conditions; however, as the number of variable combinations increases, the computational burden grows, which limits repeated review in the
early-stage decision-making process in practice. Accordingly, based on simulation data for combinations of envelope-retrofit variables for aging
houses?including a responsive envelope proposed in a related prior study (under review)?this study developed a machine-learning-based
high-speed prediction model that can rapidly estimate annual cooling load, heating load, and total energy load (kWh/㎡·yr) and validated its
performance. Annual EnergyPlus simulations were conducted for six detached-house cases from the 1960s to the 1990s, producing a training
dataset of 15,336 samples. The input variables comprised house type, window type, wall and roof insulation thickness, and building
orientation, while the output variables were defined as annual cooling load, heating load, and total load. A tree-ensemble LSBoost regression
model was trained, and performance was evaluated using 5-fold cross-validation (out-of-fold predictions). The RMSE values were 0.3073 for
cooling, 0.6423 for heating, and 0.6913 kWh/㎡·yr for total load (R2 = 0.9995, 0.9982, and 0.9990). Variable-importance analysis indicated
that Window Type had the greatest influence on cooling and total loads, whereas wall insulation thickness was most influential for heating
load, confirming differences in dominant factors by load type. The proposed model demonstrates the potential of machine-learning-based
predictors as a rapid alternative-evaluation tool that complements repeated simulations during the early stages of envelope retrofit planning for
aging detached houses. |