ZEMCH 2019 International Conference Proceedings April.2020 | Page 24
Figure 2. Decision tree model for EUI for cooling (Nodes with bold outline are terminal nodes, and mean of
each terminal node represents the prediction. Separation point is set to 0.05)
3.2.2 Determinants of EUI for cooling
The decision tree result for the EUI for cooling is given in Figure 2. Similar with the result of MLR
model, number of children under 8 is adopted as the most influential variable. The presence of children
under 8 almost doubles the EUI for cooling. The COP of air conditioner appears to be another influential
factor, unlike the result of MLR model. However, the influence depends on the presence of children
under 8. When there are children under 8 in a household, the EUI for cooling is high (4.948kWh/(㎡∙y))
regardless of the COP of air conditioner. Otherwise, air conditioners with high efficiency (COP > 3.460)
can further reduce the EUI for cooling. In addition, occupant density has significant positive
relationship with the EUI for cooling: the higher occupant density, the higher EUI for cooling.
Unlike the MLR model, number of employed residents is not significant in the DT model. This can
be explained by one of the limitations it has. In DT model, the upper subset split into lower subsets by
a significant explanatory variable. It means that whether the tree will grow further or not is highly
dependent on the number of samples of the upper subset. Hence, even there is still a significant factor
left, it does not appear when the sample size is not large enough. Since the terminal nodes of the model
have less than 10 samples each, further separation does not occur by any other variables.
4. Conclusion
This study selected 53 apartment units in Seoul and collected real‐metered heating and cooling
energy consumption data. Using statistical methods, characteristics that have significant effect on the
heating and cooling energy consumption were identified. Unlike prior studies, this study introduces a
novel way to discover hidden determinants, using both linear and nonlinear statistical models: multiple
linear regression (MLR) and decision tree (DT).
Table 2 shows the comparison of two models. For heating, all of the building, system, and occupant
characteristics appeared to be significant in both MLR and DT models. Surface area, use of auxiliary
heating devices, and heating set temperature appeared to be significant in both MLR and DT models.
However, year of building permit, which represents insulation level of the building, is significant only
in DT model. It is because, for the dataset used in this study, the EUI is in nonlinear relationship with
the year of building permit. For cooling, while occupant characteristics are dominant, none of the
13
ZEMCH 2019 International Conference l Seoul, Korea