Iw

CTT

a

ARRA

KBOOP

1

o(Rwaataocbdp

ea

f

0h

Energy and Buildings 68 (2014) 99–107

Contents lists available at ScienceDirect

Energy and Buildings

j ourna l ho me page: www.elsev ier .com/ locate /enbui ld

ntegrating probabilistic methods for describing occupant presenceith building energy simulation models

hristopher M. Stoppel ∗, Fernanda Leite1

he University of Texas at Austin, Department of Civil, Architectural and Environmental Engineering, 301 E. Dean Keeton Street, Stop C1752, Austin,X 78712-1094, United States

r t i c l e i n f o

rticle history:eceived 15 April 2013eceived in revised form 19 August 2013ccepted 22 August 2013

eywords:uilding energy modelsccupant presenceccupant behavior

a b s t r a c t

This paper presents a method for developing a probabilistic-based occupancy model that focuses on occu-pants’ long vacancy activities (greater than 1 week) and other potential building underutilization that isfurther integrated with a building energy simulation model. The combined model is then applied towardan existing Leadership in Energy and Environmental Design (LEED) certified military dormitory and latercompared with corresponding values from the energy model’s original prediction as well as actual build-ing energy data. The occupancy model simulates annual building occupancy rates comprised of weeklyvalues based on the frequency, duration, and seasonality of occupants’ long vacancy activities. The energymodel uses the simulated occupancy rates to yield the building’s predicted range of energy performance.

robabilistic methods Applying the combined model to the existing LEED building resulted in an improved, predicted EnergyUse Intensity (EUI) mean value of 612 MJ/m2 as compared to the original model and actual EUI values of691 and 590, respectively. While the model also demonstrated its utility in describing the change in pre-dicted performance over a range of probabilities associated with certain long vacancy activities, effortsto incorporate other occupant behavior-related aspects such occupant schedules and thermal set pointscould further improve modeling efforts.

. Introduction

Building energy modeling remains a key element in the designf energy efficient and sustainable buildings. The United StatesU.S.) Department of Energy’s (DoE) Office of Energy Efficiency andenewable Energy cites nearly 400 building energy simulation soft-are tools available for use in simulating various building design

lternatives’ impact on energy consumption [1]. Buildings consume significant amount of energy as compared to other sectors, such asransportation and industrial. According to the DoE, the residentialnd commercial building sector accounted for approximately 41%f the nation’s total energy consumption in 2010 [2]. As buildingsontinue to use a significant portion of the total energy consumed,uilding energy models will continue to play an intricate role in theesign of new buildings and renovation of existing ones for both therivate and public building sectors.

The literature contains numerous studies that evaluate theffectiveness of energy modeling efforts by comparing predicted toctual energy consumption, particularly buildings certified through

∗ Corresponding author. Tel.: +1 719 433 3396; fax: +1 512 471 3191.E-mail addresses: stoppel@utexas.edu (C.M. Stoppel),

ernanda.leite@austin.utexas.edu (F. Leite).1 Tel.: +1 512 471 5957; fax: +1 512 471 3191.

378-7788/$ – see front matter. Published by Elsevier B.V.ttp://dx.doi.org/10.1016/j.enbuild.2013.08.042

Published by Elsevier B.V.

the Leadership in Energy and Environmental Design (LEED) pro-gram [3–9]. Turner and Frankel’s [3] large study of 121 LEEDbuildings cited that from a programmatic standpoint, the energymodels predict building energy consumption rather well. How-ever, they also concluded that significant variance existed betweenpredicted and actual energy consumption values among individualbuildings and that further research was necessary to further explainthese sources of model error.

Sources for building energy model error can normally beattributed to one of four different areas: differences betweenthe energy model parameter values and the as-built building,its mechanical systems, or presumed activities within designatedspaces within the building not accounted for in the model; build-ing systems operating at suboptimal performance levels, perhapsdue to insufficient commissioning or maintenance activities; dif-ferences in climate conditions affecting building performance; andoccupant influence. Torcellini’s [9] study of six high-performancebuildings revealed inadequate building controls to enable efficientintegration of building systems, less than expected savings fromdaylighting and photovoltaic (PV) systems, higher than expectedplug loads, over estimating the building’s effective insulation val-

ues, and an overly optimistic estimation of occupant acceptance ofbuilding systems as actual sources of model error.

The latter of these actual sources of model error fall under thebroader category of occupant influence. As building energy models

100 C.M. Stoppel, F. Leite / Energy and

0

10

20

30

40

50

60

70

80

90

100

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

1-Feb1-Jan1-Dec1-Nov1-Oct1-Sep1-Aug1-Jul1-Jun1-May1-Apr1-Mar1-Feb1-Jan

Water C

onsumption (m

3)

Rep

orte

d O

ccup

ancy

(%)

Dat e

Building A Occupancy Building B Occupancy Building A Water Building B Water

Ft

bfwipiaiitttbvrrpb

idvrTitsbtostvtpfoDMowssi

of occupant responses to adjusting indoor lighting levels. He inte-

ig. 1. Reported daily occupancy rates and average weekly building water consump-ion.

ecome more sophisticated in predicting energy consumption,uture models will need to incorporate the vast variation that lieithin building occupant influence. Occupant influence on build-

ng performance can be further divided into occupant behavior andresence. Most energy modeling efforts describe occupant behav-

or and presence through predetermined occupant schedules andssumed plug load factors that deterministically describe occupantnfluence in the building’s energy consumption. For example, build-ng energy models typically use weekday and weekend scheduleso model occupant presence throughout the week. However, whilehese model assumptions may suffice under certain circumstances,hey do not adequately describe the range of variation that cane observed among building occupants. They do not account forariation within individual work schedules, energy consumptionelated behavior, or periods of intermediate and long vacancy. Pastesearch studies have tried various methods for describing occu-ant behavior and to a lesser extent, occupant presence, and wille described in greater detail in the literature review section.

Fig. 1 illustrates the reported, actual occupancy for two,dentically-constructed, and co-located LEED-certified militaryormitory buildings and their associated water usage. The primaryertical axis represents the building’s daily reported occupancy aseported by the military installation’s housing management office.he secondary vertical axis represents the building’s daily waterntake in cubic meters (m3) which is used as an indirect measure ofhe building’s actual occupancy within the building. The illustrationerves two purposes. First, it illustrates that unlike the respectiveuilding energy model’s assumption of fully-utilized dorm roomshroughout the year, the number of unassigned rooms varies notnly with time, but also among buildings. Second, the noticeablepikes and troughs in building water consumption illustrate thathe reported occupancy data does not account for tenants’ longacancies, but do in fact occur. The increase in water consump-ion observed in both buildings from May through September mayartly be explained by the building’s usage of additional wateror outdoor irrigation purposes. But it does not account for otherccurrences, such as the larger troughs observed in November andecember, and Building A’s noticeable decrease from February toay. A similar graph illustrating energy consumption and reported

ccupancy was constructed in a previous paper [10]. The buildings’ater data used in this study was used to improve our under-

tanding of the buildings’ actual occupancy by detecting large andudden changes in the buildings’ actual occupancy. While the build-ng energy model assumed a consistently fully-occupied building,

Buildings 68 (2014) 99–107

Fig. 1 clearly shows significant differences exist. Further researchis necessary to integrate building occupant presence and morespecifically, long periods of vacancy, in order to modeling effortsassociated with building energy consumption.

The purpose of this study is to propose a method that accountsfor the potential variance in occupant presence, specifically longvacancies, through probabilistic methods that are in turn integratedwith a building energy model. For demonstrative and validationpurposes, this method is applied toward the aforementioned mil-itary dormitory building and compared these results with theoriginal building energy model’s prediction and actual energy con-sumption results.

2. Literature review

Integrating user impact with building performance throughoccupants’ related behavior and presence patterns are importantelements in any whole building energy simulation analysis. Occu-pants influence building energy consumption through a variety ofactivities such as emitting heat and water vapor through mere pres-ence, making changes to the building’s indoor environment such asopening/closing window shades and adjusting thermostat controls,and engaging in work and leisure-related tasks conducted withinthe building. Degelman [11] noted that the building’s operationalcharacteristics, implicitly implying occupant behavior, can havean even greater impact on building energy performance than thebuilding’s thermal envelope. However, much less has been done onmodeling building occupant behavior as compared to building sys-tems. The American Society of Civil Engineers (ASCE) Visualization,Information Modeling and Simulation (VIMS) technical commit-tee states that accurately modeling building occupant behavior isa challenge that demands attention [12].

Integrating behavior models with whole building energy mod-els largely began with incorporating behavior models with energymodels associated with indoor artificial lighting. Hunt [13] beganthis area of research by translating detailed observations of howpeople used indoor lighting under varying circumstances into aprediction model for the likely use of manually-operated light-ing systems. Unfortunately, many building energy models typicallyaddress user presence and behavior through static and rigidmethods such as occupant profiles and consumption factors. Theoccupant profiles generally consist of 24 hourly values representingthe percentage of an assumed peak load. Building energy modelsmay also use separate schedules for weekdays, weekends, and hol-idays to account for the assumed differences in usage based onbuilding type during these time periods. Page et al. [14] listed mul-tiple shortcomings of this approach that included over simplifyingthe variety of occupancy patterns among building occupants dur-ing weekdays or weekends and excluding atypical behaviors suchas intense presence or long periods of vacancy, all of which canbe accounted for when observing actual energy data. This agreeswith Degelman [11] who noted that energy models tend to betteralign with reality when building operations are more constant androutine (e.g. when the occupant exhibits less control over indoorenvironmental conditions). Kwok and Lee [15] further illustratedthis shortcoming in their study of relating occupant behavior tobuilding energy consumption. Their study consisted of evaluatinga large office building comprised of numerous multi-national firmsthat utilized the building during various sets of business hours.

Bourgeois et al. [16] addressed these shortcomings by develop-ing a behavior model that accounted for the variety and frequency

grated a sub-hourly occupancy-based control (SHOCC) model withthe Lightswitch2002 behavior model, developed by the contribu-tions from Newsham et al. [17] and Reinhart [18], into the whole

y and Buildings 68 (2014) 99–107 101

boriDetcdaboc

waraieipb

ifpsfdofisa

eoeebeVoHmoofrd

pmmocctbm[oawmi

Long activity identification and characterization

Develop activity pdf profiles

Generate simulated tenant occupancy profiles

Calculate predicted energy values based onreported and simulated tenant occupancies

Analysis

Identify building occupant groups

C.M. Stoppel, F. Leite / Energ

uilding energy simulation program ESP-r so that the related effectsn heating and cooling requirements could be better realized. ESP-represents one of numerous building energy software tools listedn the U.S. Department of Energy’s Building Energy Software Toolsirectory and is used by building design professionals to evaluatenergy consumption among competing design alternatives [1]. Inhis study, he demonstrated a savings of greater than 40% in energyonsumption resulting from building occupants actively seekingaylighting solutions as opposed to relying on artificial means. Theuthor also cited however, that this approach while adequate foruildings containing single occupancy patterns like a classroomr single-person office may be somewhat unsuitable for buildingsontaining more complex occupancy patterns.

While integrating human behavior into energy models beganith activities associated with controlling the indoor environment,

dditional research has focused on greater levels of complexityegarding occupant activities affecting energy consumption. Tabaknd de Vries [19] studied intermediate activities in office build-ngs that require users to move through the building’s spaces. Hoest al. [20] recognized the growing range of complexity in integrat-ng occupant behavior with building energy simulation models andrescribed a methodology for determining the appropriate level ofehavior resolution in a modeling effort.

The study of occupant presence and absence in a buildings another aspect of occupants’ impact on building energy per-ormance. Wang et al. [21] focused on intermediate periods ofresence and absence among building occupants of single per-on offices in a large office building. From the data collectedrom individual office motion detectors, she developed a pre-iction model that could simulate the building’s overall hourlyccupancy, thereby generating multiple daily occupancy pro-les based on the data collected. While the method focused onhort periods of absence, it did not account for occupants’ longbsences.

Page et al. [14] recognized this shortcoming involving the mod-ling of occupant presence and further elaborated that excludingccupant long absences actually overestimates total occupant pres-nce and the related, yearly energy consumption prediction. Paget al. also developed an occupancy model for simulating presencey using a Markov chain to create random occupancy profiles thatxhibited similar statistical properties to actual occupancy data.irote and Nueves-Silva [22] also used Markov chains to simulateccupant presence and their related behavior in a building space.owever, Page et al. discovered during calibration that while theodel worked well in replicating short and intermediate periods

f daily presence and absence, it did not account for long periodsf absence (i.e. time periods greater than 24 h). In order to accountor these occurrences, the authors added an algorithm that createdandom long periods of vacancy based on probabilistic parametersescribing the frequency and duration.

Page et al. [14] and Wang et al.’s [21] studies focused on occu-ancy presence on a timescale of hours. However, circumstancesay exist where focusing on a greater timescale of days and weeksay be more appropriate. The significant and prolonged changes

bserved in Fig. 1 water consumption patterns due to likely relatedhanges in occupancy appear to suggest focusing on long vacancy-ausing activities of days and even weeks. The characteristics ofhese activities can then, in turn, be integrated into the wholeuilding energy model to simulate building performance underore realistic conditions with respect to occupancy. Kwok and Lee

15] also used building operation-related metrics to mimic buildingccupancy. Their study utilized the building’s fresh air supply rate

s an indirect indicator of building occupancy as the air supply rateas related to the building’s indoor CO2 rate. This study presents aethod for integrating the characteristics of known activities caus-

ng long periods of vacancy for building occupants. The method

Fig. 2. Research methodology for constructing simulated energy consumption val-ues based on occupants’ long vacancies.

is subsequently applied to a LEED-certified (version 2.2) militarydormitory building.

3. Research method

Fig. 2 illustrates the steps describing the research methodused for this study. The process begins with first identifying thebuilding’s major occupant groups as determined by common longvacancy activities which were defined as any activity resulting ina physical absence from the building of greater than or equal to 1week. Shorter term absences (<7 days) such as those due to train-ing, 3-day weekend, illness, and vacation were excluded from thisstudy. It was also assumed that the building’s occupant group pro-portions would be constant throughout the analysis period (1 year).In reality, this could change slightly or dramatically depending ona number of factors such as military personnel summer rotationcycles, deployments, or changes in leadership philosophy regardingunit cohesion in the dormitories. In this research study, the occu-pant groups were established based on the occupant’s respectiveassigned military unit. The installation’s Enterprise Military Hous-ing (eMH) database supplied the data to determine the building’soccupant group proportions.

The next step involved identifying the long vacancy activitiesassociated with each occupant group based on input from theinstallation’s military housing office, public works department, anddata collected from building occupants that resulted in three dis-tinct activity groups: training activities, deployments, and vacationperiods. Characterizing the activities not only included knowledgeof the activity frequency and duration, but also how the activi-ties related to one another. For example, some activities may besequence-driven and exhibit tendencies as to when they occur withrespect to other activities. In our study building, the occupants’long vacancy activities were sequence-driven and it was thereforenecessary to incorporate their related interaction rules in order tocreate more realistic occupancy profiles. Fig. 3 illustrates the deci-sion flow diagram used to randomly select the annual activities andtheir associated start date and duration.

In this study, three training activity types were identified.The first training activity represents the significant drop in waterconsumption observed in Fig. 1 October–November timeframe.This activity lasts approximately 4–6 weeks, affects all building

102 C.M. Stoppel, F. Leite / Energy and Buildings 68 (2014) 99–107

struc

oocia

twvTa

tofribwnieaesddaulnpoegera

Fig. 3. Decision flow diagram for con

ccupant groups, and was considered as a prerequisite to anccupant group’s deployment activity. The next training activityorresponded to individual occupant groups, while the third train-ng activity also reflects training within the occupant group, but at

lower organizational level.The deployment and vacation activities were relatively simpler

o characterize. While the deployment activity was characterizedith an unlimited range of start date and prescribed durations, the

acation period was always defined as the year’s last two weeks.he drop in water consumption in Fig. 1 December timeframe isscribed to this activity.

The next step consisted of integrating the characteristics ofhe long vacancy activities into a method for simulating annualccupancy profiles. This entailed constructing probability densityunctions (pdf’s) for the activities’ duration (weeks) as well as theange of occupant group members (percentage) directly partic-pating in the activity and therefore physically absent from theuilding. For simplicity, each occupancy profile consisted of 52eekly values and all simulated activity durations’ were wholeumbers. Activities resulting in partial week absences were not

ncluded. The decision analysis program @RISK was used to gen-rate random iterations of annual, long vacancy-related activitiesnd their associated start dates, durations, and % participating forach occupant group. The program @RISK enables simulation andensitivity analysis by allowing users to assign various values andistributions among variables [23]. The program also features a ran-om number generator function used to generate values from anssigned probability distribution. The following stipulations weresed in constructing the long vacancy-related activities. First, the

ong vacancy activities within the same occupant group couldot overlap one another, which would dampen the overall occu-ancy rate. Second, long vacancy activities involving all buildingccupant groups took precedence over individual occupant groupvents. In other words, training activities for individual occupant

roups could not be scheduled during periods which required thentire military organization’s participation. This stipulation rep-esents Fig. 1 observed long vacancy activity in November whichffected all building occupant groups. Third, occupant groups’

ting occupant group long vacancies.

deployment-related activities could be simulated to occur at anytime during the analysis period, while training-related activitiescould only be scheduled around deployment and vacation-relatedactivities. For each iteration, the building occupancy profile wascalculated for each occupant group by subtracting out the per-centage of the occupant group’s participation (P) in each simulatedactivity in accordance with its simulated start time and duration.Therefore, each group’s occupancy profile consisted of 52 weeklyoccupancy values that accounted for each simulated, long absence.The building’s combined occupancy profile was then calculated asthe weighted summation of each of the corresponding occupantgroup profiles.

Note that the above calculation implicitly assumes a year-round,fully-occupied building. In order to account for periods of varyingbuilding sub-utilization, as observed in Fig. 1, a reported occupancy(RO) variable is created and defined as the percentage of assigneddormitory rooms. The calculation for determining the building’ssimulated occupancy profile may now be defined as the series ofweekly values corresponding to the sum of n occupant group’sweighted simulated occupancy (wj), as shown in Eq. (1).

Simulated building occupancyi =∑

wj(ROi)(1 − Pj) (1)

The final step in the simulation process entailed pairing thevalues from the simulated building occupancy series with thecorresponding building energy model results. To accomplish this,multiple building energy model (TRACE 700, version 6.2.8.3)simulations were conducted by adjusting the occupancy-relatedschedules (lighting, dormitory rooms, and common use areas) by1% in each simulation and using the TMY3 weather dataset for thebuilding’s corresponding geographical location.

Fig. 4 illustrates these simulation results in a sensitivity anal-ysis graph that depicts predicted annual electricity and naturalgas consumption based on varying occupancy values. Both elec-

tricity and natural gas values were converted to kWh’s usingthe thermal conversion factors prescribed by the U.S. Environ-mental Protection Agency’s Portfolio Manager Program [24]. Thegraph’s solid line represents the Energy Use Intensity (EUI) value

C.M. Stoppel, F. Leite / Energy and Buildings 68 (2014) 99–107 103

0

100

200

300

400

500

600

700

800

0

1,000 ,00 0

2,000 ,00 0

3,000 ,00 0

4,000 ,00 0

5,000 ,00 0

6,000 ,00 0

0% 10 % 20 % 30% 40 % 50% 60 % 70 % 80% 90 % 100 %

(MJ/m

2) EU

I

Sim

ulat

ed C

onsu

mpt

ion

(MJ)

Occ upancy Rate (%)

Ele ctricity Natural Gas EUI

Ft

cc(ufigasutwblaera

saupctabmDCs

3

idtgawti

Table 1Characteristics of occupant group long vacancy activities.

Ai Activity type Duration % affected

1 Deployment 5, 12 mo 90–952 TE-1 4–6 wks 90–953 TE-2 2–3 wks 30–40

90–954 TE-3 10–15 wks 15–205 TE-2 2–3 wks 30–40

scenarios, the probability of the deployment activity occurring

ig. 4. Sensitivity analysis of occupancy (%) on simulated building energy consump-ion.

alculated as the sum of the building’s predicted total annual energyonsumption divided by the building’s gross area in square metersMJ/m2). The EUI values corresponding to varying occupancy val-es are annotated on the graph’s secondary vertical axis and rangerom approximately 400–680 MJ/m2. The predicted annual electric-ty consumption increases at a more consistent rate and exhibits areater range of values as compared to natural gas. The predictednnual natural gas values showed a somewhat different trend,harply spiking from 5 to 6% occupancy rate, nearly constant val-es between 6 and 43% occupancy, and then a noticeable increasehereafter. The predicted annual natural gas values showed a some-hat different trend, sharply spiking at 5% occupancy (due to the

uilding energy software’s treatment of building spaces with 5% oress occupancy as unoccupied), nearly constant values between 6nd 43% occupancy, and then a noticeable increase thereafter. Thenergy model’s treatment of occupancy values of less than 5% likelyesulted in slightly underestimated simulation values in this rangend was a limitation in this study.

The results from this series of simulations produced a table ofimulated energy consumption values corresponding to any day in

calendar year and level of occupancy. From this table, the sim-lated sets of daily building occupancy values were converted toredicted energy consumption values. The final step in the processreated a simulated Energy Use Intensity (EUI) metric by dividinghe sum of simulated daily energy values by the building’s grossrea (MJ/m2). We selected the EUI metric as it is used as a commonuilding energy performance metric to compare building perfor-ance within its respective building type and is used by the U.S.epartment of Energy’s Energy Information Agency in creating theommercial Buildings Energy Consumption Survey (CBECS) dataet [2].

.1. Research method validation

In the methodology validation step, the occupancy-based build-ng energy model was applied to the previously mentioned militaryormitory building for five various occupant group configura-ions under three different scenarios. We identified seven occupantroups based on the data collected from the eMH database as wells six potential long vacancy activities based on earlier discussions

ith the installation’s military and civilian personnel. Table 1 lists

he long vacancy activities and their associated characteristics usedn the simulation.

90–956 Vacation 2 wks 90–95

In total, five long vacancy activities were included in the simu-lation with one activity occurring twice a year. Prior to simulation,all duration values within each activity as well as the %-affected byeach activity were assigned equal probabilities of occurrence. Thedeployment activity affects individual occupant groups once a yeargiven the simulation’s assigned probability of occurrence. For alloccupant groups affected by this activity, an equal start time andduration were assumed. For example, one possible simulated setof long vacancy activities could include occupant groups 1, 2, and4 all participating in a 5-month deployment beginning on Week30. The Training Event-1 (TE-1) activity represents a single train-ing event participated in by all occupant groups (e.g. same starttime and duration) unless the event occurred during a time whena particular occupant group was already engaged in a deploymentactivity. The TE-2 activity affects all occupant groups and occurstwice a year, but was not restricted to occurring simultaneouslywith other occupant groups’ TE-2 start dates. We observed two dif-ferent %-affected ranges for the set of occupant groups based on thecharacter of their primary mission. The operations-focused occu-pant groups were assigned the higher %-affected values (90–95%)while the support-focused occupant groups contained the lesser(30–40%). TE-3 represents smaller echelon training and thereforeexhibits a longer duration, but lesser %-affected, as compared tothe TE-1 and TE-2 activities. The vacation activity represents thecalendar year’s final two weeks reserved for military members totake annual leave for the holiday season. In our simulation, weassumed no training events occurred during the last two weeksof the year, but deployment activities could still occur during thistime.

We used the occupancy model based on long vacancy activi-ties developed with @RISK to generate random annual occupancyprofiles for five buildings each containing different occupant groupconfigurations. The five buildings consisted of a Base Case buildingwhich contained equally-portioned occupant groups and Build-ings A–D which comprised of actual occupant group percentagescollected on four of the identically-constructed, LEED dormitorybuildings from the eMH database. The fifth LEED dormitory buildingcontained different occupant groups from the other four build-ings and was therefore not included in this study. Occupantgroups (OG) 1–3 represent support-focused groups while occu-pant groups 4–7 comprise of the operations-focused groups. Thegiven set of occupant group configurations represents a wide rangeof possible scenarios and therefore deemed useful in quantify-ing how much, if any, the building’s energy performance couldchange based on differences within the building’s occupant groupmakeup.

We implemented the model in three basic scenarios. The firstscenario consisted of running the model for each building assum-ing continuous 100% occupancy. In the second scenario, we usedthe building’s actual reported building occupancy. In both of these

was held to 0, so that only the long vacancy activities describedin Table 1 affected building energy performance. The third sce-nario comprised of a series of simulations at various probability of

104 C.M. Stoppel, F. Leite / Energy and Buildings 68 (2014) 99–107

% and

daa

3

ipwirat

Fig. 5. Buildings A–D comparison at 100

eployment values (0.25, 0.5, 0.75, and 1.0) performed on Base Casend Building A. Section 4 presents the results and related discussionnd analysis.

.2. Building description

The subject building is a 4-story, military dormitory constructedn 2008 and has been in operation and utilized since 2009. Com-rising of approximately14,200 m2, it contains 368 bed spaces andas certified LEED-Gold (version 2.2) in 2009. The subject build-

ng is one of five identically-constructed and co-located buildings,eferred to in this paper as Building A. The building uses electricitynd natural gas to operate its mechanical heating and cooling sys-ems. Twelve months of Building A’s actual building energy data

actual occupancy to base case scenarios.

collected from 2012 were compared with the original buildingenergy model’s prediction as well as the modeling results per-formed in this study.

4. Results

In order to illustrate the impacts on energy consumptionfrom occupants’ long vacancy activities and the building’s under-utilization (<100% occupancy rates), we first conducted simulationsunder two different scenarios for Buildings A–D. In each simula-

tion, we compared the results to the Base Case building definedas being continuously 100%-occupied and comprised of equally-sized occupant groups. The first set of simulations comprised ofcontinuously 100%-occupied buildings containing the same ratio

y and Buildings 68 (2014) 99–107 105

aesgob(2tfupal

BtwBsswv1trBbyCdCagtAdt2

FodmgtoocltoeopB

btTPodbPo

P(D) as well as increasing the range of likely values due to increas-ing P(D). Toward the upper end of P(D) values, the distribution oflikely EUI values appears to decrease slightly noted by the relativelysmaller peak located at approximately 580 MJ/m2.

C.M. Stoppel, F. Leite / Energ

mong occupant groups as recorded during the data collection ofach buildings’ occupant group makeup. The purpose of this set ofimulations was to determine the impact of different occupantroup configurations on building energy consumption. In the sec-nd set of simulations, the buildings retained their respectiveuilding occupant group ratios, but the Reported Occupancy valueRO variable in Eq. (1)) now reflected the buildings’ calendar year012 reported daily occupancy values. The purpose of these simula-ions was to determine the additional effect on energy consumptionrom building under-utilization (i.e. reported daily occupancy val-es less than 100%). In both of these initial scenarios, we held therobability of a deployment activity to zero so that only trainingnd vacation activities described in Table 1 contributed to occupantong vacancies.

Fig. 5 illustrates a set of box plot simulation results for thease Case and Buildings A–D conducted at 100% Occupancy andhe actual Reported Occupancy values. In this set of simulations,e observed virtually no difference between the Base Case anduilding A’s simulation conducted at 100% occupancy. This wasomewhat expected as the Base Case and Building A possess veryimilar occupant group configurations. The simulation conductedith Building A’s actual reported occupancy yielded a lesser mean

alue of approximately 626 MJ/m2 as compared to Building A’s00% occupancy simulated mean value of nearly 637 MJ/m2. Similaro Building A, Building B’s simulation conducted at 100% occupancyesulted in a distribution of EUI values slightly greater than thease Case. Buildings C and D, whose occupant group configurationsoth exhibit higher percentages of operations-focused groups, bothielded sets of simulated EUI values slightly less than the Basease. In all cases, the first set of simulations (100% occupancy) pro-uced minor deviations from −5 to +5 EUI units from the Basease. The second set of simulations (reported occupancy), nowccounting for vacant and unassigned rooms, produced relativelyreater shifts in the simulation mean values (approximately −11o +12 EUI units), respective of the Base Case simulation. Building

exhibited the greatest difference among simulation results con-ucted at 100% occupancy and actual occupancy which correspondso the lower observed occupancy rates during the first half of012 (Fig. 1).

There are three additional points associated with Fig. 5.irst, the original energy model calculated a predicted EUI valuef 691 MJ/m2. Incorporating the long vacancy activities aloneecreased the mean value of the model estimate by approxi-ately 45–55 EUI units depending on the building’s occupant

roup configuration. This range of simulated values correspondso the figure’s second point that despite a relatively wide rangef occupant group configurations, the range of means varied bynly approximately 10 EUI units indicating the occupant grouponfigurations in this case did not significantly contribute towardarge changes in energy consumption, as compared to the rela-ively larger range of simulated mean values using the reportedccupancy data Finally, its worth noting that the distance betweenach building’s respective box plots reflects the building’s reportedccupancy rate. Building C exhibited the highest reported occu-ancy rates (small change in 100% and Reported box plots) whileuilding A demonstrated the least.

The next set of simulations consisted of varying the probabilityy which a deployment activity, P(D), would occur. Fig. 6 illustrateshe simulation results for P(D) values of 0, 0.25 0.50, 0.75 and 1.0.here are three main observations taken from this figure. First, as(D) increases, the range of possible EUI values also increases. Thisbservation results from the simple fact that as P(D) increases, so

oes the likelihood of multiple combinations of occupant groupseing affected by this long vacancy activity. Second, as soon as(D) > 0, the distributions begin to exhibit a negative skew. Theutliers are likely due to iterations when the simulation selected

Fig. 6. Comparison between building A and base case (reported occupancy) at var-ious probabilities of deployment activity.

a 12-month deployment activity to begin early in the year thataffected a large percentage of occupant groups. Finally, we superim-posed the building energy model’s predicted EUI value (691) as wellas Building A’s actual EUI (590) value calculated for 2012 on the BaseCase and Building A’s respective box plots at P(D) = 0. Superimpos-ing these points on the P(D) = 0 boxplots was deemed appropriateas the occupant groups did not participate in this activity dur-ing the subject year. Including these points illustrate a noticeableimprovement in accurately predicting the building’s energy perfor-mance by incorporating occupants’ known long vacancy activitiesas a method of modeling occupant presence. As such, BuildingA’s simulation at P(D) = 0 yielded a mean value of approximately612 MJ/m2.

Histograms and box plots are useful for displaying a dataset’sdistribution shape and range, but overlaying numerous datasets ona single plot can quickly become too cumbersome. To better illus-trate the effect on the shape and distribution of EUI values at variousP(D) values, we constructed a surface plot for Building A, illustratedin Fig. 7. Similar to the histograms, the horizontal and vertical axis’srepresent EUI and percentage values, respectively, and the depthaxis now represents P(D) values. The surface plot clearly illustratesthe higher peak representing building performance at low values of

Fig. 7. Surface plot depicting simulations for building a conducted at various P(D)values.

1 y and

eitbuupiolm

imFchubnuaftvpFkeo

5

bovwivmiEatmauab

biEsirs

iwcad

[

[

[

[

[

[

[

manual and automated lighting control, Energy and Buildings 38 (7) (2006)814–823.

[17] G.R. Newsham, A. Mahdavi, I. Beausoleil-Morrison, Lightswitch: a stochastic

06 C.M. Stoppel, F. Leite / Energ

In this surface plot, we focused on the deployment activity’sffect on building performance since it represented the activity pos-ng the greatest influence on building energy consumption amonghe pre-identified long vacancy activities. In fact this method coulde applied toward other variables exhibiting greater degrees ofncertainty that are difficult to model with deterministic val-es, such as occupant behavior. The method could be used as aotential approach in addressing areas of uncertainty or manag-

ng expectations with respect to building performance for buildingwners and/or decision makers as opposed to current practices thatargely use deterministic methods for calculating building perfor-

ance.Although the study results suggest a potentially modest

mprovement to the overall model’s accuracy on building perfor-ance, it also demonstrates the study’s limitations in two key areas.

irst, the energy model values for various occupancy rates were allalculated assuming constant thermal set points for cooling andeating seasons. In reality, as occupancy decreases due to under-tilization or other long vacancy activities, room thermostats coulde adjusted accordingly to save energy. Second, the model doesot consider the potentially wide range of occupant presence orsage patterns that can easily exist during actual building operationnd instead uses a single schedule for hourly occupant presenceor weekday and weekend time periods. Further discussion withhe building’s operations and maintenance personnel could pro-ide greater clarification of more realistic outcomes on buildingerformance as a result of large decreases in building occupancy.uture research could include integrating the occupant groups’nown long vacancy activities with other human behavior mod-ls that could provide greater understanding of occupant impactn building performance.

. Conclusions

In this paper, we presented a method for integrating proba-ilistic methods to describe building energy performance basedn occupants known long vacancy activities. We included longacancy activities related to training, vacation, and deployments asell as known building underutilization as reported in the build-

ngs daily occupancy rates. The characteristics of the known longacancy activities were further incorporated into an occupancyodel using @RISK that generated random sets of simulated build-

ng occupancy profiles that were later translated into simulatedUI values. The simulated buildings included a Base Case buildings well as four other buildings using actual occupant data to charac-erize each building’s occupant group configuration. We performed

ultiple simulations by varying the probability of a deploymentctivity occurring in 1% increments and later displayed the resultssing histograms, box plots, and finally a surface plot to provide

single representation of all simulations performed for a singleuilding.

Comparing the simulation results (612 MJ/m2) to a singleuilding’s actual annual energy consumption value (590 MJ/m2)

ndicated a modest improvement from the original energy model’sUI prediction of 691 MJ/m2. However, the results also pointed toignificant limitations within the approach that included assum-ng constant thermal set points for heating and cooling seasonsegardless of occupancy rates as well as deterministic occupantchedules.

Nevertheless, the method discussed in this paper could bencorporated into future energy modeling efforts of new buildings

ith relative ease and offer additional benefits. First, the methodould facilitate discussion regarding the building’s operationmong building owners, maintainers, and designers early in theesign phase. Most other building types may not exhibit the

[

Buildings 68 (2014) 99–107

prolonged vacancy patterns on a timescale of weeks as observed inthe subject military dormitory. However, knowledge of occupantpresence and behavior in similar building types can be used tofacilitate discussions of building energy controls early in thedesign process and subsequently included into the building energymodel. Knowing how state variables such as occupant behaviorcan potentially affect building performance can lead to moreinformed decisions. The method could also facilitate discussion inthe building’s operation and maintenance phase. In this example,the method identified a relatively wide range of possible buildingoccupancy scenarios that could potentially impact subsequentdecisions regarding building space utilization. Applying thismethod to other building types could generate similar discussionson building operation and utilization.

Acknowledgements

The authors would like to thank Ms. Paula Loomis and Mr.Matthew Ellis from the U.S. Army Corps of Engineers, membersfrom the subject building’s Department of Public Works staff andHousing Office. The views expressed in this article are those of theauthors and do not reflect the official policy or position of the U.S.Air Force, Department of Defense, or the U.S. Government.

References

[1] U.S.D. Energy o. Building Energy Software Tools Directory, 2013.[2] E.I.A. Commercial Buildings Energy Consumption Survey (CBECS), U.S. Energy

Information Agency, 2011.[3] C. Turner, M. Frankel, Energy Performance of LEED for New Construction Build-

ings, U.S. Green Building Council, 2008, pp. 46.[4] C. Turner, LEED Building Performance in the Cascadia Region: A Post Occu-

pancy Evaluation Report, Cascadia Region Green Building Council, 2006,pp. 18.

[5] K.M. Fowler, E.M. Rauch, Assessing Green Building Performance: A Post Occu-pancy Evaluation of 12 GSA Buildings, 2008, pp. 171.

[6] K. Fowler, E. Rauch, J. Henderson, A. Kora, Re-Assessing Green Building Perfor-mance A Post Occupancy Evaluation of 22 GSA Buildings, U.S. Department ofEnergy, 2011.

[7] G.R. Newsham, S. Mancini, B.J. Birt, Do LEED-certified buildings save energy?Yes, but, Energy and Buildings 41 (8) (2009) 9.

[8] J.H. Scofield, Do LEED-certified buildings save energy? Not really, Energy andBuildings 41 (12) (2009) 5.

[9] P. Torcellini, S. Pless, M. Deru, B. Griffith, N. Long, R. Judkoff, Lessons Learnedfrom Case Studies of Six High-Performance Buildings, National RenewableEnergy Laboratory, 2006, pp. 151.

10] C.M. Stoppel, F. Leite, Evaluating building energy model performance of LEEDbuildings: identifying potential sources of error through aggregate analysis,Energy and Buildings (2013).

11] L.O. Degelman, A Model for simulation of daylighting and occupancy sensorsas an energy control strategy for office buildings, in: IBPSA Building SimulationConference, International Building Performance Simulation Association, Kyoto,Japan, 1999, p. 9.

12] F. Leite, C. Bogen, J. Gong, Grand challenges in information modeling forthe architecture, engineering, construction facility management industry, in:2013 A.S.C.E. Computing in Civil Engineering Conference, Los Angeles, CA,2013.

13] D.R.G. Hunt, Predicting artificial lighting use – a method based upon observedpatterns of behaviour, Lighting Research and Technology 12 (1) (1980)8.

14] J. Page, D. Robinson, N. Morel, J.L. Scartezzini, A generalised stochastic modelfor the simulation of occupant presence, Energy and Buildings 40 (2) (2008)83–98.

15] S.S.K. Kwok, E.W.M. Lee, A study of the importance of occupancy to build-ing cooling load in prediction by intelligent approach, Energy Conversion andManagement 52 (7) (2011) 2555–2564.

16] D. Bourgeois, C. Reinhart, I. Macdonald, Adding advanced behavioural modelsin whole building energy simulation: a study on the total energy impact of

model for predicting office lighting energy consumption, in: Right Light Three:Third European Conference on Energy-Efficient Lighting, Newcastle Upon Tyne,UK, 1995, p. 7.

18] C.F. Reinhart, Lightswitch 2002: a model for manual control of electric lightingand blinds, Solar Energy 77 (1) (2004) 14.

y and

[

[

[

[

C.M. Stoppel, F. Leite / Energ

19] V. Tabak, B. de Vries, Methods for the prediction of intermediate activities by

office occupants, Building and Environment 45 (6) (2010) 1366–1372.

20] P. Hoes, J.L.M. Hensen, M.G.L.C. Loomans, B. de Vries, D. Bourgeois, User behav-ior in whole building simulation, Energy and Buildings 41 (3) (2009) 295–302.

21] D. Wang, C.C. Federspiel, F. Rubinstein, Modeling occupancy in single personoffices, Energy and Buildings 37 (2) (2005) 121–126.

[

[

Buildings 68 (2014) 99–107 107

22] J. Virote, R. Neves-Silva, Stochastic models for building energy prediction based

on occupant behavior assessment, Energy and Buildings 53 (0) (2012) 183–193.

23] S.C. Albright, W.L. Winston, C.J. Zappe, Data Analysis and Decision Making, 4thed., Joe Sabatino, 2011.

24] U.S. E.P.A. Thermal Conversions in EPA’s Energy Star Portfolio Manager, U.S.Environmental Protection Agency, 2011.