Exploring the potential use of building facade information to estimate energy performance

Accepted Manuscript Title: Exploring the potential use of building facade information to estimate energy performance Authors: Andrea Martinez, Joon-Ho Choi PII: DOI: Reference:

S2210-6707(17)30569-3 http://dx.doi.org/10.1016/j.scs.2017.07.022 SCS 716

To appear in: Received date: Revised date: Accepted date:

31-3-2017 20-7-2017 30-7-2017

Please cite this article as: Martinez, Andrea., & Choi, Joon-Ho., Exploring the potential use of building facade information to estimate energy performance.Sustainable Cities and Society http://dx.doi.org/10.1016/j.scs.2017.07.022 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Exploring the potential use of building facade information to estimate energy performance Andrea Martinez a, Joon-Ho Choi a* a

Building Science, School of Architecture, University of Southern California, Los Angeles, CA 90089,

United States *

Corresponding Author: 850 West 37th St. Watt Hall #318, Los Angeles, California, 90089, United States

Highlights 1. This research investigated the potential use of building facade information to estimate buildings’ energy efficiency. 2. This study was conducted based on the data obtained from 92 existing non-residential buildings in the U.S. 3. Multiple data mining techniques were adopted to investigate significant building and facade features. 4. Facade symmetry, facade area to South (all climates), facade-to-roof ratio (hot climate), window-to-wall ratio (cold) were found as significant building parameters. 5. This study revealed the potential use of facade features to estimate building energy performance.

Abstract In spite of prolific research on the energy performance of buildings in the last decades, and the growing focus on reducing their operational energy, buildings still prevail as the main end users of energy in the U.S. The goal of this research is to investigate the potential use of building 1

facade information to estimate its energy performance, and to find significant facade attributes depending on different climate conditions in the U.S. This study adopted Energy Use Intensity (EUI) for total consumption and described building information, including window-wall ratio, orientation, aspect ratio, and other building components. Concentration was given to achieve a balanced data collection from best practices and green certified non-residential projects located in different climate conditions in the U.S. Data mining techniques, such as classification tree and statistical tools that included Analysis of Variance (ANOVA), a 2-sample T-test, and regression, were adopted for analysis of this group of buildings. It was found that there were common functional and technical features, as well as similar performances of existing buildings linked to these buildings’ energy consumption. These findings could not only inform the new design of facades, but facade retrofits could be strategically established based on lessons learned from real practice. Keywords: Building energy performance, data mining, evidence-based design, best-practice buildings.

2

1 Introduction Buildings have become a major target for energy reduction in the 21st century. In the US, energy used on existing buildings represent nearly half of all energy consumed nationally (Architecture 2030 2017). Despite being built as strong structures, a significant portion of these buildings have inefficient envelopes, on account of the relatively low-cost, abundant energy that was available to provide heating, ventilation, and air conditioning by mechanical means. With dwindling resources and an increase in energy prices, the performance of the building facade plays a major role in the building’s total energy efficiency, while fulfilling its loadbearing, and appearance functions. From load bearing walls, the facade has evolved in complexity due to complex geometry and active components for adaptive responses to external conditions (Compagno 1995, Wigginton and Harris 2002). These evolved systems are intended to reestablish the foundational role of a facade as the main protector layer of the building, creating spaces that are not only comfortable, but also efficient. Indeed, a high-performance facade design is incorporated in an interdisciplinary process that accounts for the specializations of each component in a building’s components (Knaack 2007). Integration is highly relevant in early design, particularly if decisions are to be cost effective. Towards the pursuit of making informed decisions on facade systems, evidence-based research is fundamental when facing new designs and renovation of existing buildings; but, unfortunately, the sources of information about actual performance are limited. The Building Performance Database (BPD), an on-line platform of the U.S. Department of Energy, contains more than 180,000 building records with an interface for statistical analysis and filtering capabilities by building type, size, age, occupancy, location, and systems (U.S. Department of Energy 2014). However, the data analysis findings derived from these datasets is more useful as a threshold than for identifying design methodologies. Other databases include the Low 3

Energy Building (a comprehensive compilation of European best-practices that exposes energy intensities of individual buildings (AECB/TSB) 2014). These types of platforms are informative; however, they are still limited to displaying specific facade features and derived design recommendations. Computer simulations are currently one of the most used tools for energy prediction, and the facade is included as a relevant predictor of a whole-building level. With different capabilities and limitations (Crawley, Hand et al. 2008), energy simulation is continually improving to allow the exploration of facade features and their influence on energy performance. Though several studies have used simulations of conceptual buildings (Cheung, Fuller et al. 2005), other studies have used the actual performance of existing buildings as a baseline. Among these simulation studies, different facade configurations have been explored for retrofitting existing buildings in subtropical conditions (Bojić and Yik 2005, Pan, Yin et al. 2008), mild climates (Martinez, Noble et al. 2012), and cold climates (Aksoy and Inalli 2006, Yildiz, Ozbalta et al. 2014). Other facade studies covered the effects of double-glazed facades (Hien, Liping et al. 2005, Kim, Schaefer et al. 2013). Also, predictive models using data mining techniques have estimated building performance in terms of thermal comfort and schedules (Ahmed, Korres et al. 2011). Multiple regression models have been used in groups of buildings with strategies for total energy reduction (Aranda, Ferreira et al. 2012), and more specifically such as heating energy demand reduction (Catalina, Iordache et al. 2013) However, fewer studies have been identified exploring the relationship between facade and energy performance by using actual data. The main objective of this study is to explore energy performance of existing buildings through predictive models as a function of facade features that could provide insights into design decisions. With an increasing number of energy disclosure laws, reporting of data will be more open and more information will be available. 4

This study values these efforts and is confident that actual data concerning energy consumption constitutes a solid foundation for making practical design decisions that could contribute to higher building performance.

2 Methodology The main objective of this study was to explore the energy performance of existing buildings through predictive models as a function of their facade features. The first step toward this goal was the collection of best practice building projects with available energy performance and enough information about constructed features, especially of the facade. The dataset was built with a balance of buildings located in hot and cold climate areas, so that weather influence could be considered in the analysis. To identify commonalities, the main design features and patterns of the dataset were observed, and then organized on a series of parameters at a building and facade level. Finally, the analysis used the Energy Use Intensity (EUI) of the buildings as an indicator of the performance in a series of models that were created using statistical tools (correlation, ANOVA, stepwise regression) and data mining techniques (J-48 classification tree). These models were set up for the two levels of analysis (building and facade).

2.1 Database Individual data obtained from 92 existing non-residential buildings in the U.S. was collected for this study. The dataset included an equal number of buildings from hot and cold climates, which are listed in Table 1 according to ASHRAE’s climate classification (ASHRAE/IESNA 2007). Along with weather criteria, buildings with at least three stories above ground were preferred, as they were expected to have a greater predominance of facade surface than lowrise constructed buildings with larger roof surfaces. A significant number of LEED-certified

5

and EnergyStar-rated buildings were integrated into the dataset, as these buildings were better described in available technical reports and/or pertinent information was obtained from knowledgeable parties. These best-practice projects constituted a valuable repository with detailed building architectural characteristics, facade features, and energy consumption records. Table 1: Number of buildings collected by climate zone classification for the U.S. (ASHRAE/IESNA 2007)

Hot climate

Cold climate

Climate zone 1A 1B 2A 2B 3A 3B 4C 5A 5B 5C 6A 6B 7

# bldgs. 2 0 18 0 14 12 1 0 0 0 37 8 0

type

CDD

Very Hot - Humid Very Hot - Dry Hot - Humid Hot - Dry Warm - Humid Warm - Dry Mixed-Marine Cool-Humid Cool-Dry Cool-Marine Cold-Humid Cold-Dry Very cold

5000 < CDD 10ºC 5000 < CDD 10ºC 3500 < CDD 10ºC<5000 3500 < CDD 10ºC<5000 2500 < CDD 10ºC < 3500 2500 < CDD 10ºC < 3500 -

HDD 2000 < HDD18ºC ≤ 3000 3000 < HDD18ºC ≤ 4000 3000 < HDD18ºC ≤ 4000 3000 < HDD18ºC ≤ 4000 4000 < HDD18ºC ≤ 5000 4000 < HDD18ºC ≤ 5000 5000 < HDD18ºC ≤ 7000

The collected information allowed for establishing a series of numeric, nominal, and ordinal data to be used as parameters, as detailed in Table 2. General building characteristics were comprised of the building location, shape, and size, while facade features were presented in considerable detail. Facade dimensions were traced, measured, and classified when drawings were available; otherwise, estimations were based on observation. When drawings were not available, dimensions were estimated from online sources and scaled with a certain degree of accuracy. Facade features such as exterior wall areas, orientation, glazing properties, and sunshades were collected to examine their significance in building performance. The axis and orientation of the main building were referred to the main four cardinal points (N, S, E, and W). Based on the dimensions and the orientation of the building, four ratios were estimated: Aspect Ratio (AR), Surface-to-Volume Ratio (SVR), Facade-to-Roof Ratio (FRR), and the 6

Window-to-Wall ratio (WWR). For AR, the South facade was placed as the nominator. Due to the importance of the envelope in heat transmission in buildings, the SVR served to identify those buildings with larger exposed surfaces of the envelope. For a more accurate distinction between the influence of the facade in relation to the exposed envelope, the FRR represented the relation of the facade surface regarding the roof surface. WWR was detailed for the major four orientations as a coefficient in respect to the facade area. Table 2: Details of parameters described for buildings included in the study

Parameters

description

Level-1 (Building) HDD

Heating degree days

CDD

Cooling degree days

Climate zone

Based on ASHRAE classification

Climate subtype

a= Moist, b= Dry, c= Marine

Year Built

Year of original construction

Area

Gross building useful area (m2)

Stories

Number of stories above ground

Building Purpose

Height

CO= Commercial Office, CM=Commercial Manufacturing, GO= Government Office, CH= Courthouse, EK= Education K-12, EH= Education Higher, EL= Education Library Measured to roof level (meters)

Roof area

Footprint area (m2)

Footprint Length (N, S, E, W) Footprint type

Length of footprint in each orientation

Building volume

R=Rectangular, RL=Rectangular L-shape, RU=Rectangular U-shape, RC= Rectangular Courtyard, T= T-shape, SC= Semicircular Measured in m3

SVR

Surface-to-volume ratio

SVR classification

[1] less than 0,25 [2] 0,25-0.5 [3] 0.5-0.75, [4] 1-2, (6) 2+

AR

Aspect ratio

AR classification

[1] less than 1, [2] 1-1.5, [3] 1.5-2, [4] 2-3 [5] 3+

Level-2 (Facade) Facade areas (N, S, E, W)

Total facade surface facing ±45º from true cardinal point

Facade type

Glass U-factor

CP=concrete or masonry walls with punched windows, CW=curtainwalls (highly glazed), DS=Double Skin facade, RC=Rain screen cladding, PC=Precast Concrete, MC=Metal cladding Expressed as W/m2/°C

Glass SHGC

Expressed from 0-1

Glass Tvis

Expressed from 0-1

WWR (N, S, E, W)

Window-to-wall ratio by orientation

WWR Average

Window-to-wall average value of all windows

7

FRR

Facade-to-roof ratio

FRR classification

[1] less than 1, [2] 1-2, [3] 2-4, [4] 4+

Indicators EUI

Energy Use Intensity (kWh/m2yr)

2.2 Methods adopted for data analysis This study used statistical tools, such as correlation, a two-sample T-test, Analysis of Variance, and stepwise regression assisted by Minitab (Minitab Inc. 2013). Correlation, a linear association of two parameters, was described by the Pearson correlation coefficient r, which describes both the strength and direction of the relationship with ranges from -1 to 1. The pvalue indicates the statistical significance of the correlation coefficient. ANOVA and the twosample T-test were used for an analysis of the correspondence between numeric and nominal categorical parameters. Similar building design features of the two defined groups were evaluated based on hot or cold climate conditions. Since the set of parameters was large, stepwise regression was used to help identify a subset of significant predictors of energy performance. Stepwise regression analysis adds and removes predictors one at a time, until it finds all variables with significance levels greater than the acceptable range while removing the least significant ones. A statistical significance level of 95% was adopted throughout the study. Any uncertainty of the results was addressed by using a couple of indicators. In the first instance, R-sq was used as the indicator of any statistical significance of stepwise regressions. R-sq defines the percentage of the response variations that can be explained by these regression models. Values of R-sq of at least 85 were commonly accepted. In addition, the Mean Bias Deviation (MBE) was adopted later for evaluating the accuracy of the models, using the 5% suggested by ASHRAE standards (ASHRAE 2002). In addition to those statistical methods, a machine learning tool-Weka 3.6.10 (Machine Learning Group 2014)- was used to explore a model with predictive algorithms. Data mining 8

tasks, such as a classification tree and clustering, were adopted for data analyses to identify significant building features. Pre-process capabilities allowed cleaning and filtering of data, while classification functions allowed estimation of the major impact of parameters on energy use reduction. A classification algorithm, J48 decision tree, was used to aid in identifying the significance of facade parameters in predicting EUIs. This tool, a top-down induction abstract classifier, was used to generate a series of unpruned trees using predictive analytics. As a characteristic of this method, a root node is selected to create branches for each possible attribute value and to split instances into subsets. After selecting only instances that reach the branch through repetition, the process stops when all the instances have achieved the same class. A 10-fold cross-validation test model was used. The two levels of analysis also followed this method, obtaining a building-level tree and a facade-level tree.

3 Results

3.1 Building energy performance by facade elements

3.1.1 Building-level features Some general relations in the dataset with regards to location, size, and vintage were observed. Buildings in the cold-climate group were located at latitudes of more than 40º N, while buildings in the hot-climate latitudes were in the lower 30º. A significant relationship was observed between HDD with the building area (r=0.175, p-value=0.029) and the building height (r=0.208, p-value=0.009). That correlation was not observed for CDD. The year of construction of buildings in the dataset ranged from 1934 to 2014, with 1990 as the mean year. Buildings constructed post-1950 tended to be in higher latitudes than older buildings. According to building size, building area ranged from 938 m2 to 102,193 m2, with a 9

mean size of 2,567 m2. Newer buildings were smaller (r=-0.269, p-value=0.010) and tended to be in colder climates (p-value= 0.041). Building’s heights ranged from 3 to 62 stories, with a mean of 7.9 stories. However, some distinctions were observed considering their location. In a hot climate, buildings averaged 10.3 stories, while in cold climate zones 5.5 stories. The predominant orientation of the building was recorded as a major building feature. High latitude was correlated with buildings with a predominance of a South exposure (r=0.248, pvalue=0.017) and a North (r=0.217, p-value=0.038). Moreover, the buildings’ footprints were described in their lengths and were measured and classified. Single rectangular footprints were predominant with 76% of the dataset, whereas 15% were L-shaped, and the rest were some derived type (central courtyard, U-shaped, semi-circular).

3.1.2 Facade-level features A series of facade types were identified in the dataset. The dominant types were concrete/ masonry walls with punched windows (CP), representing 44% of all buildings; and curtain walls (CW) comprising 40% of the dataset. The concept of facade symmetry was used in this study to represent buildings with a similar facade solution in all orientations, and lacking facade design in response to solar, wind, and other context conditions, such as views. Figure 1 illustrates that this facade symmetry increased with area (r=0.447, p-value=0.000) and number of stories (r=0.354, p-value=0.001). Moreover, buildings with symmetric facades had a higher predominance of HDD (r=0.305, p-value=0.003) and CDD (r=0.274, p-value=0.008). This facade symmetry was particularly present in buildings constructed from 1950 to 1990. A decrease in this symmetry might show that buildings built in the past 20 years tend to have more consideration for climate or contextual conditions.

10

95% CI for the Mean 20.0 17.5

Stories

15.0 12.5 10.0 7.5 5.0 asymmetry

symmetry Façade symmetry

Figure 1: Representation of facade symmetry and its relation to building height

Glass performance parameters, such as U-factor, SHGC, and Visible transmittance (Tvis), were analyzed in relation to location and window areas. A significant relation was found between the glass U-factor and climate zones, as illustrated in Figure 2. Values for U-factors ranged from 1.41W/m2/°C to 2.78 W/m2/°C. As illustrated, lower U-factor values (less heat transfer) were significant in relation to climate zone (p-value=0.012). Higher glass SHGC values were shown to have some significance for buildings with larger WWR in the East facade (r=0.549, p-value=0.052). However, it is important to highlight the consequent and statistical significance of higher SHGC in buildings with sunshade elements such as screens (r=0.718, pvalue=0.006), interior light shelves (r=0.586, p-value=0.022), or exterior light shelves (r=0.720, p-value=0.019). This showed that glass performance has some degree of synergy in 11

working with these elements for sun protection. Nonetheless, no significant findings were observed in these relationships with regard to Tvis.

95% CI for the Mean 2.0

Glass u-factor

1.9

1.8

1.7

1.6

1.5 cold

hot hot/cold

Figure 2: Interval plot for glass U-factor according to hot/cold climate conditions

Several functional elements, such as different types of sunshades were also identified in the study. It was found that horizontal sunshades were mostly implemented in buildings with differentiations in facades, but curiously decreased in buildings with a larger aspect ratio (r= 0.216, p-value=0.030), which were Southern-exposed buildings. Horizontal sunshades were also implemented in conjunction with vertical sunshade elements, or fins. Sun protection from interior blinds was more likely to be incorporated in buildings with larger WWR in a Southern facade (r=0.292, p-value=0.012), but less in buildings with larger WWR in an Eastern facade (r= -0.234, p-value=0.045).

3.1.3 Building and Facade Ratios The aspect ratio (AR) was calculated using the main axis lengths from the building footprints, with Southern exposure as a reference value. Building aspect ratios ranged from 0.7 to 2.8, 12

with a mean value of 1.39 and a standard deviation value of 0.697. More than 67% of the buildings contained aspect ratios over 1, which represented a larger South/ North exposure of the building mass. These buildings usually had up to eight-story structures. The surface-to-volume ratio (SVR) is the main indicator of envelope predominance. More energy efficient buildings tend to have lower ratios, as they are more compact, minimizing their envelope surface for energy transfer. Of all cuboid building shapes, the prevalence of the single cube was the most practical shape, but derivative forms, such as L-shape, U-shape, or central courtyard, increased that ratio. The facade-to-roof ratio (FRR) was calculated to corroborate the selection criteria of buildings with a minimum of three stories. It rated exposed vertical areas (facades) against horizontal areas (roof) of the building envelope. The FRR ranged from 0.7 to 9.8, with a mean value of 3.2 and a standard deviation value of 2.2. Eighty-eight percent of the dataset fulfilled the minimum facade predominance ratio. Buildings with FRR between 1 and 3 corresponded to mid-rise buildings of around 4 to 7 stories, while FRR values over 4 corresponded to high-rise buildings with at least 10 stories. The window-to-wall ratio (WWR) was considered a key parameter affecting building energy performance related to climate. The dataset showed a significant relation between WWR and building vintage (r=0.361, p-value= 0.000), with modern buildings having more openness in all the orientations, as illustrated in Figure 3. However, the average WWR was observed to decrease in relation to latitude (r= -0.214, p-value=0.042).

13

80 70

WWR Average

60 50 40 30 20 10 (1) before 1950

(2) 1950-1990 YearBuilt

(3) 1990+

Figure 3: Boxplot of WWR average values for vintage buildings

3.1.4 Energy use intensity by building facade condition Common aspects that defined energy performance were observed in the dataset. EUI was linked to certain building conditions, such as building location and vintage. Even though newer buildings tended to have better energy performance than older buildings, an exploration of EUIs in three main categories of building vintage showed that this was not always true. Figure 4 indicates that the so-called postwar buildings presented a higher range of energy use intensities as compared to pre-war buildings and those built within the past 20 years.

350

EUI (KWh/m2/yr)

300 250 200 150 100 50 (1) before 1950

(2) 1950-1990 YearBuilt

(3) 1990+

Figure 4: Boxplot of Energy Use Intensity (EUI) and vintage categories

14

According to location, buildings showed an increase in the EUI in colder climates (p-value= 0.007), which can be explained, in part, by an observed growing use of energy for lighting (r= 0.591, p-value=0.026). The mean value of EUI in cold-climate buildings was 224.1 kWh/m2yr, while in hot climates it was 156.1 kWh/m2yr. A significant number of buildings had a level of LEED certification. However, buildings with higher LEED certification did not necessarily have a better energy performance. Despite a decrease in heating (r= -0.522, p-value=0.004) and lighting (r=-0.663, p-value=0.003) as the LEED level increased, increased energy was used in cooling these certified buildings (r= 0.529, p-value=0.004). Additionally, the presumption that a lack of facade differentiation (symmetry) would lead to higher energy use was observed in this dataset. Buildings with the same facade solutions were shown to tend to have EUIs of over 200 kWh/m2yr more energy use than buildings with facade differentiation. Even though the distribution of the two building groups seemed to overlap, the difference between the confidence intervals of these two groups of buildings was shown as significant in a 2-Sample T-test (p-value=0.021). This result supports the difference in energy use of symmetric and asymmetric buildings. As illustrated in Figure 5, the average EUI of symmetric buildings is higher than that of asymmetric buildings by around 64%. As discussed in Section 3.1.1, a symmetric facade was prevalent in buildings constructed between 1950 and 1990, while an asymmetric facade was constructed in buildings in recent decades, depending on climate or contextual conditions. This architectural trend might contribute to a large EUI difference in the symmetric and asymmetric building groups. Additionally, buildings with larger eastern facade areas showed higher EUI values (r=0.224, p-value=0.036) than other orientations. Among these characteristics of non-symmetric buildings, increased EUI were observed on buildings with larger WWR in Northern and

15

Eastern facades. Nevertheless, certain facade features such as horizontal sunshades helped to decrease the EUI (r= -0.234, p-value=0.029). 95% CI for the Mean

EUI (KWh/m2/yr)

350

300

250

200

150 non-symmetric

symmertric Façade symmetry

Figure 5: Boxplot of EUI in relation to buildings with and without facade symmetry

Finally, the relation to building and facade features were explored in terms of energy end-use intensities. For heating, the EUI was related to vintage and building sizes. This relation of heating energy usage was statistically significant at a 5% level for total area (r= -0.368, pvalue=0.049) and number of stories (r= -0.497, p-value=0.006). Two specific facade features were influential in decreasing heating: horizontal sunshades (r= -0.730, p-value=0.000) and interior blinds (r= -0.761, p-value=0.000). Additionally, heating need decreased in buildings with larger WWR in Southern (r= -0.608, p-value=0.000) and Western (r= -0.795, p-value= 0.000) facades. However, heating EUI increased with larger SVR (r=0.493, p-value=0.007) and for buildings with screens as a sunshade strategy (r=0.424, p-value=0.022). Cooling EUI was shown to decrease in buildings with larger WWR in Northern facades. As discussed above, multiple significant facade parameters, including WWR, stories, and building geometry ratios, showed a significant relationship to energy performance. Based on these physical performance principles, this study adopted regression and data mining strategies for an in-depth analysis, which are discussed in the following sections. 16

3.2 Selection of significant facade elements for EUI performance by climate condition

3.2.1 Stepwise regression with hot climate zone data Level-1 (Building level): The analysis resulted in the surface-to-volume ratio (SVR) being selected as the most important predictor of EUI, and length of the footprint (North) and the year of construction as the next significant contributors. As shown in Table 3, the coefficients for the year built and the area are negative, whereas the other two are positive. All predictors were efficient and significant with p-values smaller than 0.05. These coefficients represented modern and smaller buildings that have lower energy intensity. The resulting equation from the stepwise regression is as follows: 𝐻𝑜𝑡 𝑐𝑙𝑖𝑚𝑎𝑡𝑒 𝐿𝑒𝑣𝑒𝑙1 𝐸𝑈𝐼 (𝑘𝑊ℎ/𝑚2/𝑦𝑟 = = 2032 + 49 𝑆𝑉𝑅 + 0.63 𝐿𝑒𝑛𝑔𝑡ℎ (𝑁𝑜𝑟𝑡ℎ) − 0.98 𝑌𝑒𝑎𝑟 𝑏𝑢𝑖𝑙𝑡

The stepwise regression resulted in a P-value of 0.031 and R-sq of 34.55%. The selected five parameters explained the variations in the total EUI by showing that 34% of the dataset had statistical significance. Level-2 (Facade-level): The analysis resulted in the selection of eight parameters, with FRR as the more significant predictor. As shown in Table 3, larger FRR, WWR (South and West), and total facade area would decrease the EUI, whereas larger facade area (North and East) and WWR (North), and higher U-factors would increase the EUI in hot climates.

17

𝐻𝑜𝑡 𝑐𝑙𝑖𝑚𝑎𝑡𝑒 𝐿𝑒𝑣𝑒𝑙2 𝐸𝑈𝐼 (𝑘𝑊ℎ/𝑚2/𝑦𝑟) = 150 − 5.78 𝐹𝑅𝑅 + 4.53 𝑊𝑊𝑅 (𝑁𝑜𝑟𝑡ℎ) − 3.43 𝑊𝑊𝑅 (𝑆𝑜𝑢𝑡ℎ) – 3.49870 𝑊𝑊𝑅 (𝑊𝑒𝑠𝑡) + 0.20280 𝐹𝑎𝑐𝑎𝑑𝑒 𝐴𝑟𝑒𝑎 (𝑁𝑜𝑟𝑡ℎ) + 1.42816 𝑊𝑊𝑅 (𝐸𝑎𝑠𝑡) – 0.07972 𝐹𝑎𝑐𝑎𝑑𝑒 𝐴𝑟𝑒𝑎 (𝑇𝑜𝑡𝑎𝑙) + 0.12978 𝐹𝑎𝑐𝑎𝑑𝑒 𝑎𝑟𝑒𝑎 (𝐸𝑎𝑠𝑡) The stepwise regression resulted in an R-sq of 99.9%, which represents the significance of presenting the total data. The model also showed a p-value of 0.000. Table 3: Results of Level-1 and Level-2 stepwise regressions for the hot-climate dataset for total EUI

Level-1 (building level) Step Constant

1 127.75

2 80.51

3 2032.45

SVR T-Value P-Value

40 2.63 0.012

46 3.17 0.003

49 3.49 0.001

0.61 2.42 0.02

0.63 2.65 0.012

Length North T-Value P-Value Year built T-Value P-Value

-0.98 -2.24 0.031

S R-Sq

69.8 14.76

65.9 25.9

62.7 34.55

7 167

8 154.9

Level -2 (façade level) Step Constant FRR T-Value P-Value

-

WWR North T-Value P-Value WWR South

7.77677 -5.67533 -2.86 -50.22 0.014 0 4.9743 19.62 0

-

4.53712 394.77 0

3.83181 -3.42397

18

T-Value P-Value WWR West T-Value P-Value

-14.82 0 -

-296.96 0

3.83984 -3.50682 -15.22 -320.1 0 0

Façade Area North T-Value P-Value

0.06731 7.72 0

0.20226 124.79 0

WWR East T-Value P-Value

1.38194 6.27 0

1.42817 159.36 0

Facade Area Total T-Value P-Value

0.01238 -0.07962 -5.11 -100.25 0 0

Facade Area T-Value P-Value S R-Sq

East

0.12978 85.32 0 13.7 97.71

0.555 99.99

3.2.2 Stepwise regression with cold climate zone data Level-1 (Building level): A different set of parameters was shown to be relevant for this group of buildings. Six parameters were selected, with latitude as the most significant parameter; whereas AR, stories, area, and west/east length of footprint, as the next significant contributors to the total EUIs. Table 4 illustrates the coefficients for latitude, AR, area, and footprint length (East) are negative. This suggests that buildings located in colder latitudes tend to increase the EUI, as those are taller or have larger western facades. On the contrary, smaller buildings and those with larger exposure in the southern and northern facades tend to show decreased EUI. The resulting equation from the stepwise regression is as follows:

19

𝐶𝑜𝑙𝑑 𝑐𝑙𝑖𝑚𝑎𝑡𝑒 𝐿𝑒𝑣𝑒𝑙1_𝐸𝑈𝐼 (𝑘𝑊ℎ/𝑚2/𝑦𝑟) = 668.8 – 13.5 𝐿𝑎𝑡𝑖𝑡𝑢𝑑𝑒 – 27.1 𝐴𝑅 + 29.6 𝑆𝑡𝑜𝑟𝑖𝑒𝑠 – 0.00402 𝐴𝑟𝑒𝑎 + 73.7 𝐿𝑒𝑛𝑔𝑡ℎ (𝑊𝑒𝑠𝑡) – 72 𝐿𝑒𝑛𝑔𝑡ℎ (𝐸𝑎𝑠𝑡) The 6-step regression model resulted in R-sq of 95.96%. Therefore, these parameters explained the variations in almost the whole subset of buildings in the cold climate, with a highly significant p-value of 0.000. Level-2 (Facade-level).: Also, as shown in Table 4, the use of only facade parameters resulted in a regression with a selection of only two significant parameters, which were relevant in their Eastern orientation: WWR and facade area. Buildings with larger Eastern regions and with larger windows on the East increased the EUI as they compared to other orientations. 𝐶𝑜𝑙𝑑 𝑐𝑙𝑖𝑚𝑎𝑡𝑒 𝐿𝑒𝑣𝑒𝑙2 𝐸𝑈𝐼 (𝑘𝑊ℎ/𝑚2/𝑦𝑟) = −44.21 + 7.08 𝑊𝑊𝑅 (𝐸𝑎𝑠𝑡) + 0.0248 𝐹𝑎𝑐𝑎𝑑𝑒 𝐴𝑟𝑒𝑎 (𝐸𝑎𝑠𝑡)

The stepwise regression resulted in an R-s of 95.58%; however, the p-value obtained was beyond the 5% range (p-value=0.116).

Table 4: Results of Level-1 and Level-2 stepwise regressions for the cold-climate group of buildings for total EUI

Level-1 (Building level) Step Constant Latitude T-Value P-Value AR T-Value P-Value

1 3105.4

2 3062.9

3 2745.2

4 2212.1

5 1904.5

6 668.8

-65.5 -8.1 0

-60.8 -10.82 0

-55.8 -11.93 0

-44.3 -8.48 0

-41.7 -8.73 0

-13.5 -2.17 0.036

-91.8 -7.04 0

-74.7 -6.69 0

-76.2 -7.74 0

-40.4 -2.83 0.007

-27.1 -2.45 0.019

20

Stories T-Value P-Value

12.4 4.82 0

Area T-Value P-Value

24.5 6.07 0

34.6 7.17 0

29.6 7.87 0

0.00221 -3.62 0.001

0.00586 -4.63 0

0.00402 -3.97 0

3.3 3.2 0.003

73.7 5.82 0

Length West T-Value P-Value Length East T-Value P-Value S R-Sq R-Sq(adj)

-72 -5.57 0 102 59.84 58.92

70 81.34 80.48

56.9 87.99 87.13

1 -47.42

2 -44.21

8.23 19.4 0

7.08 8.73 0

50.1 90.89 90

Level-2 (Facade level) Step Constant WWR East T-Value P-Value Facade area East T-Value P-Value S R-Sq R-Sq(adj)

0.0248 1.65 0.116 50 94.95 94.7

48 95.58 95.12

21

45.3 92.75 91.84

34.2 95.96 95.34

3.3 Regression dataset based on the whole dataset In order to explore the interrelation between building-level and facade-level impacts in all climates, a new regression analysis was performed for the entire dataset (hot and cold climate groups together). Level-1 (Building level): This model resulted in the selection of five parameters as significant predictors. The year of construction was selected as the most important contributor to EUI. As shown in Table 5, year built, latitude, and height showed negative coefficients. All predictors were significant with p-values smaller than 0.05, except for height. Table 5: Stepwise regression of the whole dataset for Level-1 (building parameters)

Level-1 (building parameters) Step Constant

1 5386

2 5574

3 4567

4 4410

5 4220

Year built T-Value P-Value

-2.6 -3.27 0.002

-2.76 -3.59 0.001

-1.94 -2.66 0.01

-1.9 -2.69 0.009

-1.8 -2.57 0.013

24.4 2.53 0.014

105.6 4.56 0

106.3 4.74 0

100.3 4.46 0

-25.2 -3.79 0

-23.9 -3.69 0

-22.7 -3.53 0.001

3.9 2.24 0.029

20 1.94 0.057

Climate number T-Value P-Value Latitude T-Value P-Value Stories T-Value P-Value Height T-Value P-Value S R-Sq

-4.4 -1.59 0.118 132 14.28

127 22.2

115 36.82

22

111 41.6

110 43.96

R-Sq(adj) Mallows Cp

12.95 37.9

19.73 30.7

33.76 15.7

7 146

8 158.4

37.77 12.1

39.28 11.3

Level-2 (facade parameters) Step Constant WWR East T-Value P-Value

0.3 0.83 0.412

FRR T-Value P-Value Facade area West T-Value P-Value

-0.0832 11.29 0

0.0858 -12.86 0

0.0851 10.43 0

0.0874 11.38 0

Facade area North T-Value P-Value

0.109 -9.4 0

-0.113 -10.32 0

Facade area South T-Value P-Value

0.113 7.27 0

0.117 7.95 0

S R-Sq R-Sq(adj)

59.4 83.7 82.34

59.2 83.52 82.44

Facade area East T-Value P-Value

-

𝐴𝑙𝑙 𝑐𝑙𝑖𝑚𝑎𝑡𝑒 𝐿𝑒𝑣𝑒𝑙 − 1 𝐸𝑈𝐼 (𝑘𝑊ℎ/𝑚2/𝑦𝑟) = 4220 – 1.80 𝑌𝑒𝑎𝑟 𝑏𝑢𝑖𝑙𝑡 + 100.3 𝐶𝑙𝑖𝑚𝑎𝑡𝑒 − 22.7 𝐿𝑎𝑡𝑖𝑡𝑢𝑑𝑒 + 20 𝑆𝑡𝑜𝑟𝑖𝑒𝑠 – 4.4 ℎ𝑒𝑖𝑔ℎ𝑡 The stepwise regression using the eight predictors resulted in R-sq of 43.96% and a p-value of 0.118.

23

Level-2 (Facade level): This model resulted in the selection of four significant predictors. Facade area (West) was identified as the more important parameter, followed by facade areas for the other three orientations. Larger facade areas to the East and South would increase the EUI, whereas larger facade areas to North and West would decrease the EUI. The resulting equation is as follows: 𝐴𝑙𝑙 𝑐𝑙𝑖𝑚𝑎𝑡𝑒 𝐿𝑒𝑣𝑒𝑙 − 2 𝐸𝑈𝐼 (𝑘𝑊ℎ/𝑚2/𝑦𝑟) = 158.4 – 0.0858𝐹𝑎𝑐𝑎𝑑𝑒 𝑎𝑟𝑒𝑎 (𝑊𝑒𝑠𝑡) + 0.0874 𝐹𝑎𝑐𝑎𝑑𝑒 𝑎𝑟𝑒𝑎 (𝐸𝑎𝑠𝑡) – 0.113 𝐹𝑎𝑐𝑎𝑑𝑒 𝑎𝑟𝑒𝑎 (𝑁𝑜𝑟𝑡ℎ) + 0.117 𝐹𝑎𝑐𝑎𝑑𝑒 𝑎𝑟𝑒𝑎 (𝑆𝑜𝑢𝑡ℎ) The stepwise regression resulted in an R-sq of 83.52%, which would explain a large portion of the cases used in the regression, all parameters with p-values=0.000.

3.4 Prediction model of EUI as a function of facade features A J48 decision tree was used to help in the identification of the significance of facade parameters in the prediction of EUI. The buildings’ EUI were organized as follows to be used as response: less than 100 kBtu/sf-yr, 100-150 kBtu/sf-yr, 150-200 kBtu/sf-yr, 200-250 kBtu/sf-yr, 250-300 kBtu/sf-yr, 300-350 kBtu/sf-yr and 300+ kBtu/sf-yr. Level-1 Tree (building + facade parameters)-Using the whole set of parameters of all buildings, including building parameters, building/facade ratios, and facade parameters the model resulted in a tree with 38 parts and 25 leaves, with 82% of correctly classified instances. As illustrated in Table 6, the North length of the footprint was selected as the root node. If the North length of the footprint is larger than 45.72 m., and if the WWR in the South facade is higher than 0.7, the model classifies the building according to its use. Based on the previous features, commercial offices (CO), ranged between 150-200 kWh/m2yr. All other building 24

purposes classified in the 250-300 kWh/m2yr range. A second branch from the root was defined for buildings with North length less than 45.72 m., on which the footprint type defined the next level of classification. Buildings with footprint other than rectangular (such as rectangular Lshape (RL), Rectangular U-shape (RU), Rectangular with Courtyard (RC), Semi-Circular shapes (SC), Rectangular T-shape (RT)), were classified in the lower EUI category (less than 100 kWh/m2yr).

25

Figure 6: Visualization of unpruned Level 1 (building parameters) decision tree

26

Level-2 Tree (facade parameters only) – Corresponded to only facade parameters, the result was a model of 31 parts and 18 leaves, illustrated in Figure 7. The classification model resulted with 91% of the instances correctly classified.

Figure 7: Visualization of unpruned Level 2 (facade parameters) decision tree

27

As observed, the South facade area was selected as the root node in the model, which relates to the predominance of buildings with larger orientations to south and north observed in the building-level tree. Facade areas East and North were defined as the main class for the two main branches. If the South facade area was smaller than 1,505 m2 and the East facade area smaller than 247 m2, the building was in the EUI 100-150 kWh/m2yr range. On the other section of the tree, if the South facade area of the building was larger than 1,505 m2, and the North facade area was greater than 485 m2, the WWR South was a determinant parameter for classification. For buildings with WWR South lower than 70% that had a facade of concrete wall with punched windows (CP), they resulted in a 100-150 kWh/m2yr EUI category. A summary of the two models is illustrated in Table 6. All models achieved significant representative levels, independently of the size of the attributes included.

Table 6: Summary of the two J48 unpruned tree for all (92) building instances

Tree level Level 1

Level 2

Parameters Building level + Facade level Facade level

Attributes number 66

Size of the tree 38

Number of leaves 25

Correctly classified

27

31

18

91%

82%

3. Discussion As described in the results section, all models presented in this study confirm the relevance of facade parameters in building energy consumption. The results from regression analysis by climate showed that different set of building and facade parameters were sensitive predictors for these conditions. Though, facade area and WWR for East orientation were significant in both climates. 28

Even though both methods are different in nature, similar parameters were selected as significant predictors of EUI when analyzing the whole set. The decision tree helped to confirm parameters that were previously selected in the regression analysis. Table 7 summarizes the parameters considered as significant predictors of EUI along all the models covered under the study. Table 7: Summary table of all selected parameters in analyses

x

Building level

x

Stories

x

x

Height

x

x

Latitude

x

Length East

x

x

x

x

Length West

x

Footprint shape

x x

AR

x

Facade area North Facade area South

x

Facade area East

x

x

x

Facade area West

Facade Level

All climates Decision Tree- Level 2

x

Area

SVR

All climates Decision Tree- Level 1

All climates Level 2

x

Climate number

Length North

Decision Tree

All climates Level 1

Cold climate Level 2

Cold climate Level 1

Hot -Climate Level 2

Parameters Year built

Hot-Climate Level 1

Stepwise Regression

x x

x x

x

x

x

Facade area Total

x

Glass U-factor WWR North

x

WWR South

x

WWR East

x

WWR West FRR

x x

x

x

x

x

As can be observed, some parameters were commonly selected in multiple climate conditions and algorithms, stepwise regression and decision tree. On the other hand, some other 29

parameters were not selected or were estimated as insignificant components. Since design features perform differently depending on climate conditions, this finding seems very natural. The decision tree algorithms showed some parameters commonly adopted in the regression models, and the higher estimation accuracy rates in both models confirmed the potential of facade features to estimate the building EUIs. Therefore, these selected parameters per climate condition and algorithm illustrated their significances to the building EUI performance, which would be utilized as design solutions for a construction project by understanding the sensitivity of individual facade features to the EUIs. A comparison of both levels used in the regression analyses showed that facade levels achieved higher representation of the data. Level-2, which used only facade parameters, resulted in a higher R-sq value than those from including building-level parameters. This reveals the relevance of facade components in addition to some traditionally and well-known building design strategies. All R-sq values for level-2 models were over 80%. The previous accuracy estimated using R-sq was corroborated according to the Normalized Mean Bias Error (NMBE) and the Coefficient of Variance of Root Mean Square Error (CV RMSE). As a measurement of error for these regression formulas, 5%-10% and 15%-30% were adopted as acceptable rate ranges, for NMBE and CV (RMSE), respectively, based on the ASHRAE Standards 90.1. The error values were then normalized based on the differential between actual and predicted EUI resulting from the regression equations.

2

NMBE =

̂ⅈ) ∑(yj −y (n−p)xy ̅

̂ 𝑖) √∑(𝑦𝑖 −𝑦

× 100

𝐶𝑉𝑅𝑀𝑆𝐸 = 100𝑥

𝑛−1

𝑦̃

Where 𝑦𝑖 = actual data, 𝑦̂= predicted data, 𝑦̃=mean actual data (ASHRAE 2002)

30

Table 8 includes the NMBE and CV (RMSE) values for each level of regression analyses. Most of the predicted values were slightly higher than the actual energy used in the buildings. Equations resulting from level-1 (building level) overpassed the maximum error rate defined. However, equations resulting from level-2 (facade parameters) were more accurate, within the maximum rates established, especially in Hot Climate and All Climate scopes. The model of All-Climate Level -2 generated the smallest error rates compared to other models. These variations of NMBE and CV (RMSE) may be because the individual regression models and selected parameters were different depending on the climate datasets (i.e., heating, cooling, and all climates). Table 8: Summary of results from all stepwise regression equations, including NMBE and CV(RMSE) values

Stepwise Regression model

R-sq

p-value

Hot climate Level-1 Hot climate Level-2 Cold climate Level-1 Cold climate Level-2 All climate Level-1 All climate Level-2

34.55% 99.99% 92.75% 99.95% 43.96% 83.52%

0.031 0.000 0.000 0.116 0.118 0.000

Number selected parameters 3 8 6 4 5 4

NMBE

CV(RMSE)

-7.8% -1.7% 11.1% -4.5% -4.7% -0.1%

34 7.2 66.7 27.3 38.2 0.5

4. Conclusion Based on the many real datasets of building performances collected in the U.S., this study provides a better understanding of optimally combined facade features, including a distinction for hot and/or cold climate conditions. As summarized in Results and Discussion, building design features and facade characteristics have significant potential to accurately predict energy use intensity in each construction as a function of facade features. Both stepwise regressions and decision tree approaches showed higher prediction performances when we used facade parameters only. Since the study mainly dealt with overall building performance based on various data sets, the results that distinguished among facade parameters or combined factors 31

with energy end-use intensities were analyzed in their higher sensitivity to the building energy efficiency, i.e., EUIs. The stepwise regression tool generated higher R-sq values for the analysis by including only facade parameters, and estimation accuracies were in an acceptable range suggested by the current industry standard. A machine learning classification, J48 decision tree algorithm, also confirmed this study approach within a reasonable accuracy range. These two developed computation models helped in the identification of significant parameters for estimating energy performance as a function of suggested facade features. Therefore, the analysis of the relationship between the selected building facade characteristics and the energy efficiency indicated that energy consumption could be greatly reduced by identifying the best combination of facade resolutions and other building features. By better understanding the energy impact of design variables, it is possible to focus design efforts and resources on issues with the largest potential energy benefit. These findings are informative and would be used as potential design guidelines or to improve design decision support methods by providing useful references for the decision process, based on existing best practices. This study started with a large, real building performance and facade information, but the sample size became smaller when we established sub-groups by climate. This reduced sample size might weaken the results of the data analysis and the developed models’ general applications. In addition, there may be more advanced data mining / analysis tools, such as artificial neural networks, available, which could be well matched with the collected dataset. Such an advanced algorithm may be too complicated to use in this study, but it could be considered for future research. Thus, additional real building performance and facade data should be considered in further study in order to incorporate additional advanced computational algorithms to establish an accurate data-driven building energy efficiency model as an effective design decision tool.

32

References o

Low Energy Building Database. Association for Environment conscious Building (AECB) and the Technology Strategy Board (TSB) Retrofit for the Future, Association for Environment conscious Building (AECB) and the Technology Strategy Board (TSB) Retrofit for the Future,. 2014.

o

Ahmed, A., N. E. Korres, J. Ploennigs, H. Elhadi and K. Menzel (2011). "Mining building performance data for energy-efficient operation." Advanced Engineering Informatics 25(2): 341-354.

o

Aksoy, U. T. and M. Inalli (2006). "Impacts of some building passive design parameters on heating demand for a cold region." Building and Environment 41(12): 1742-1754.

o

Aranda, A., G. Ferreira, M. D. Mainar-Toledo, S. Scarpellini and E. Llera Sastresa (2012). "Multiple regression models to predict the annual energy consumption in the Spanish banking sector." Energy and Buildings 49: 380-387.

o

Architecture 2030. (2017). "Why the building sector?"

Retrieved 20 February 2017, 2017,

from http://architecture2030.org/buildings_problem_why/. o

ASHRAE (2002). Guideline-14. Measurement of Energy and Demand Savings. Atlanta, GA, ASHRAE

o

ASHRAE/IESNA (2007). ASHRAE Standard 90.1 -- Energy Standard for Buildings Except Low-Rise Residential Buildings. Table B-1. Atlanta.

o

Bojić, M. and F. Yik (2005). "Cooling energy evaluation for high-rise residential buildings in Hong Kong." Energy and Buildings 37(4): 345-351.

o

Catalina, T., V. Iordache and B. Caracaleanu (2013). "Multiple regression model for fast prediction of the heating energy demand." Energy and Buildings 57: 302-312.

o

Cheung, C. K., R. J. Fuller and M. B. Luther (2005). "Energy-efficient envelope design for high-rise apartments." Energy and Buildings 37(1): 37-48.

o

Compagno, A. (1995). Intelligent Glass Facades: Material, Practice, Design. Zürich, Artemis Verlag,Switzerland.

33

o

Crawley, D. B., J. W. Hand, M. Kummert and B. T. Griffith (2008). "Contrasting the capabilities of building energy performance simulation programs." Building and Environment 43(4): 661-673.

o

Hien, W. N., W. Liping, A. N. Chandra, A. R. Pandey and W. Xiaolin (2005). "Effects of double glazed facade on energy consumption, thermal comfort and condensation for a typical office building in Singapore." Energy and Buildings 37(6): 563-572.

o

Kim, G., L. Schaefer and J. T. Kim (2013). "Development of a Double-Skin Facade for Sustainable Renovation of Old Residential Buildings." Indoor and Built Environment 22(1): 180-190.

o

Knaack, U. (2007). Facades: Principles of Construction. Basel, Birkhäuser Architecture.

o

Machine Learning Group, U. o. W. (2014). "Weka 3: Data Mining Software in Java."

o

Martinez, A., D. Noble, M. Schiler and M. Patterson (2012). Facade Retrofit: Strategies for energy reduction in an office building in a mild climate. PLEA, Lima.

o

Minitab Inc. (2013). "Minitab ".

o

Pan, Y., R. Yin and Z. Huang (2008). "Energy modeling of two office buildings with data center for green building design." Energy and Buildings 40(7): 1145-1152.

o

U.S. Department of Energy, E. E. R. E. (2014). "Building Database."

o

Wigginton, M. and J. Harris (2002). Intelligent skins. Oxford, Butterworth-Heinemann.

o

Yildiz, Y., T. Ozbalta and A. Eltez (2014). "Energy-saving retrofitting of houses in cold climates." Isi Bilimi Ve Teknigi Dergisi-Journal of thermal science and technology 34(1): 5361.

34