You are on page 1of 14

Energy & Buildings 255 (2022) 111644

Contents lists available at ScienceDirect

Energy & Buildings


journal homepage: www.elsevier.com/locate/enb

Load shifting and energy conservation using smart thermostats in


contemporary high-rise residential buildings: Estimation of runtime
changes using field data
Helen Stopps a,⇑, Marianne F Touchie a,b
a
Department of Mechanical and Industrial Engineering, University of Toronto, Canada
b
Department of Civil and Mineral Engineering University of Toronto, Canada

a r t i c l e i n f o a b s t r a c t

Article history: As the proportion of the population who resides in urban centers grows, the number of dwellings in high-
Received 16 August 2021 rise residential buildings is also increasing. In Canada, heating, ventilation, and air conditioning (HVAC)
Revised 25 October 2021 typically accounts for over half of all energy use in this building type and, as such, consideration of these
Accepted 28 October 2021
loads is key to effective energy demand-side management. Smart thermostats present a new opportunity
Available online 2 November 2021
conserve energy and shift electricity load to off-peak periods in these buildings through modifications to
HVAC control strategies. In this study, three thermostat control strategies were tested in 49 suites across
Keywords:
two contemporary high-rise residential buildings: 1) baseline operation (thermostat operating as a stan-
Demand management
Energy flexibility
dard programmable thermostat); 2) occupancy-based control (thermostat operating using an occupancy-
HVAC reactive control strategy); 3) load shifting control (thermostat operating using a pre-conditioning strat-
Buildings egy). HVAC runtime data was collected from smart thermostats installed in the suites and was used to
Residential develop data-driven regression models to estimate baseline suite HVAC runtime based on weather con-
Connected Thermostats ditions and suite characteristics. Thermostat-measured HVAC runtime data for periods in which
occupancy-based or load shifting control strategies were in use in the suite was then compared with esti-
mated baseline runtime for similar outdoor conditions. Occupancy-based control was found to reduce
HVAC system runtime during the cooling seasonby 5.9% ± 46% on average. On aggregate, the load shifting
control strategy was not effective at reducing peak period HVAC load during the cooling season. The strat-
egy was able to reduce peak period loads in a subset of suites substantially, which may be leveraged to
reduce electricity demand during capacity-constrained periods of grid operation, however, based on mar-
ginal emissions factors for Ontario’s electricity grid, even in these suites, the strategy was not effective in
reducing GHG emissions overall.
Ó 2021 Elsevier B.V. All rights reserved.

1. Introduction including conservation and load shifting. In addition to managing


transmission and generation capacities, in regional electricity grids
In 2018, 55% of the global population lived in urban regions, with GHG emissions intensities which vary based on the time of
with this proportion expected to grow to 68% by the year 2050 day, shifting electricity demand away from on-peak periods may
[1]. The resulting increase in energy use drives resource consump- result in a net decrease in GHG emissions even if it results in a sub-
tion, increases greenhouse gas (GHG) emissions, and places strain stantial increase in total energy use. The electricity grid in Ontario,
on energy infrastructure. Due to electricity storage limitations, Canada, where the study was conducted, is one example of this
electricity infrastructure is particularly sensitive to this growth in type of grid: Average hourly GHG intensities associated with elec-
population. To combat the increases in transmission and genera- tricity production during on-peak periods (generally considered to
tion capacity needs, effective demand-side management is be between 11:00 and 17:00 in Ontario) are almost four times
required, which can be achieved using a variety of measures, higher than typical GHG intensities overnight (13 g CO2eq/kWh)
[2] in the region.
Another consequence of population growth in urban regions, is
⇑ Corresponding author at: University of Toronto, Department of Mechanical and an increase in the proportion of dwellings in high-rise residential
Industrial Engineering, 35 St. George St., Toronto, ON M5S 1A4, Canada. buildings. In Canada, for example, construction of dwellings in
E-mail address: helen.stopps@mail.utoronto.ca (H. Stopps).

https://doi.org/10.1016/j.enbuild.2021.111644
0378-7788/Ó 2021 Elsevier B.V. All rights reserved.
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

high-rise residential buildings outpaced the construction of dwell- vided through the forced-air HVAC supply system. The tempera-
ings in all other building types in 2018, accounting for 53% of all ture change in the structural mass from over-conditioning during
new dwelling construction starts [3]. Heating, ventilation and air off-peak periods is used to ‘‘store” cooling or heating energy for
conditioning (HVAC) typically accounts for 50% of all energy use use during on-peak periods when the thermal mass is discharged
in high-rise residential buildings in Canada [4] and managing and to the indoor air, providing passive space conditioning.
reducing energy used by HVAC systems is one way to reduce the This pre-cooling approach has been widely studied in commer-
growing strain placed on energy infrastructure systems. Newly cial and office buildings using surface activation of thermal mass
constructed Canadian high-rise residential buildings typically [13–24]. These studies have found peak period electricity load
include non-programmable thermostats in each suite. As many of reductions ranging from 0 to 100% reduction with an average peak
these buildings are constructed by developers with suites then sold load reduction of around 30%. A wide range of set-back depths
to individual owners, we assume this thermostat selection is dic- (1 °C to 11 °C, average 3.7 °C) and set-back durations (2 h to
tated by lowest initial cost. In the rare case that programmable 12 h, average 9 h) were used in the studies, although the load
thermostats are installed in suites, residents frequently do not pro- reduction performance did not consistently track with set-back
gram them or fail to take full advantage of energy-saving opportu- depth or duration. Likely, differences in internal building structure,
nities [5]. Connected, smart thermostats, which typically integrate finishes, and building envelopes greatly impact the performance of
occupant sensing technology and Wi-Fi connectivity, can be used the pre-cooling, along with the baseline building operating
to conserve energy and reduce the need for occupants to program conditions.
and update thermostat schedules through the use of temperature Investigation of pre-conditioning strategies in residential set-
set-backs when occupant presence is not detected. Furthermore, tings, however, has been limited with no high-rise residential
smart thermostats also present an opportunity to optimize space building-based studies and only one simulation study in single-
conditioning load profiles through the use of load shifting control family homes. Turner, Walker, and Roux simulated the use of
strategies which can be deployed remotely over the thermostat’s pre-cooling in a high-performance single-family home for 12 U.S.
Wi-Fi connection climate zones, using surface activation of thermal mass through a
forced-air system [25]. A range of pre-cooling strategies were
tested with a set-back depth of either 3.4 °C or 4.4 °C and set-
2. Background
back durations of 3, 5, and 8 h. Peak period loads were reduced
by 50% to 75% and a set-back depth of 3.4 °C combined with a
Previous studies on occupancy-reactive temperature control
set-back duration of 8 h was found to be the most effective. It is
and load shifting through thermal mass activation were reviewed
difficult to assess whether the findings of this study may apply in
in order to understand the state-of-the-art in the high-rise residen-
high-rise residential buildings: while high-rise residential building
tial building sector.
occupants likely have similar comfort expectations and tempera-
ture schedules to occupants of single-family homes, the higher
2.1. Occupancy-reactive temperature control mass associated with high-rise residential buildings will likely
impact pre-cooling effects and control strategy performance.
Occupancy-reactive temperature control is a common feature in While past literature has investigated the impacts of alternative
commercially-available smart thermostats which use presence thermostat control strategies in commercial buildings and single-
sensors to detect when a zone is occupied and applies a tempera- family homes, the efficacy of these strategies has not yet been
ture set-back when unoccupied. The efficacy of occupancy-reactive investigated in high-rise residential buildings, which differ sub-
thermostat control has been previously demonstrated in single- stantially in terms of HVAC system configuration, thermal mass,
family homes. Field studies in the U.S. have found average heating and occupant schedules. This study is the first to investigate the
and cooling load energy savings ranging from 6% to 14% [6–10], impacts of occupancy-reactive control and load shifting in high-
although these savings have been shown to differ slightly between rise residential buildings. Smart thermostat data, collected over a
smart thermostat makes and models [7,8]. Despite the promise one-year period in 49 suites, were used to assess the change in
demonstrated in single-family homes, limited work has been com- suite-based HVAC unit runtime resulting from the use of
pleted in a high-rise residential setting with typical apartment or occupancy-reactive control and a novel loading-shifting control
condominium building occupants. A 2016 study by Pritoni, Wool- strategy during the cooling season in southern Ontario, Canada.
ley, and Modera examined the effect of smart thermostats on
energy use in university residence halls [11]. However, as univer-
sity students and typical condominium or apartment residents 3. Methodology
likely have significantly different behaviours and expectations as
occupants, the results of this study may not be applicable in In this section, the studied buildings, participant recruitment
high-rise residential buildings. Previous work on thermostat pro- methods, data collection methods, tested thermostat control
gramming behaviours and occupancy-patterns in contemporary strategies, and analysis methods are outlined.
high-rise residential buildings showed promise for the efficacy of
occupancy-based control in these suites [12], however changes in 3.1. Study buildings and participant recruitment
terminal HVAC unit operation were not quantified.
Two contemporary, highly glazed high-rise residential buildings
2.2. Load shifting using thermal mass activation in downtown Toronto, Canada, were recruited for the study. The
facades of the study buildings are shown in Fig. 1 and key charac-
Surface activation of building thermal mass is one strategy to teristics of the studied buildings are outlined in Table 1.
reduce HVAC energy consumption during periods of peak electric In both buildings, occupants control their suite temperature
grid demand by shifting these loads to off-peak periods. This load setpoint through a thermostat which controls the in-suite terminal
shift is accomplished through temperature set-backs (temperature HVAC unit (fan coil in Building A or heat pump in Building B). A
reductions during the cooling season and increases during the pressurized corridor ventilation system provides additional suite
heating season) prior to peak periods. These set-backs pre- space conditioning through heated or cooled hallway air that, in
condition building structural mass via conditioned indoor air pro- theory, is forced into suites through door undercuts. Windows in
2
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

Fig. 1. Facades of studied buildings. Images from Google Maps [43].

Table 1 heating degree days and 470 cooling degree days (°Cd, base tem-
Key Characteristics of Studied Buildings. perature of 18 °C) [26]. For the purpose of this analysis, only the
Building A Building B cooling season is considered, as heating and shoulder season in-
Year Construction 2012 2014 suite HVAC loads are relatively low due to excessive conditioning
Completed of building common areas in the heating season and the mild
Number of Floors 17 22 weather conditions during the shoulder season (see Fig. 3b for
Number of Suites 343 357 heating season).
Number of Participating 23 (6.7%) 26 (7.3%)
Suites
Suite Size 35 m2 – 85 m2 48 m2- 129 m2 3.2. Data collection
Suite Type 1 and 2 Bedroom 1, 2 and 3 Bedroom
Suite-based HVAC Two-Pipe Fan Coil Unit Water-loop Heat Pump
System (FCU) (WLHP)
HVAC unit runtime data for the study (measured from the time
the thermostat signaled the HVAC system to turn on until the time
the thermostat signaled the HVAC system to turn off) were col-
lected remotely from smart thermostats installed in participants’
both buildings are operable, however, due to building code restric-
suites. Previous work has shown that, on an hourly basis,
tions, the opening area of windows is limited to 100 mm (4 in.) in
thermostat-signaled runtime is well correlated with actual FCU
diameter.
runtimes in Building A, however, the relationship between
Three main types of suite layouts exist in the two buildings:
thermostat-signaled runtime in Building B has not been investi-
Layout 1 (one bedroom apartment, n = 14), Layout 2 (one bedroom
gated [27]. Thermostat data from 54 suites were collected between
apartment with set-back bedroom, n = 9), and Layout 3 (two- or
June 2018 and August 2019 at five-minute intervals, however ther-
three- bedroom corner suite, n = 16). Characteristic suite floor
mostat data were missing for short periods of time during the
plans for each layout type are presented in Fig. 2.
study due to intermittent connectivity issues and, in some cases,
Residents of the buildings were informed of the study by their
for longer periods of time in certain suites due to longer-term con-
condominium boards and given the option to opt-in. Participants
nectivity issues. Additionally, two participants left the study after
were required to allow researchers to install an ecobee3 smart
selling their suite and data were no longer collected after they
thermostat in their suites, remotely collect data, and change ther-
moved out. Ultimately, data were available for analysis from 49
mostat control strategies over the one and a half year monitoring
suites during the cooling season.
period which commenced in February 2018 and ended in August
Historical weather data (outdoor air temperature and cloud
2019. Residents who owned their suite were compensated for their
cover) were collected at a government weather station in Toronto
time through the retention of the thermostat after study comple-
and daily solar irradiance data from NASA’s Prediction of World-
tion and residents who were renters received a total of $50 CAD
wide Energy Resources (POWER) databases were downloaded for
in gift cards for participation. This study was approved by the
the period between June 2018 to August 2019 [26,28]. Data were
University of Toronto Research Ethics Board (protocol number
aggregated or disaggregated to hourly intervals for the runtime
38815).
comparison. Collected data measurement errors, intervals, aggre-
Annual weather in Toronto, Canada can be largely grouped into
gation methods, and sources are summarized in Table 2.
three categories: heating season (typically October though March),
In addition to the collected thermostat and weather data, the
cooling season (June through August), and shoulder season
solar azimuth angle corresponding to each thermostat data reading
(September, April, and May), during which space conditioning
was calculated at five-minute intervals using the SolarPy package
demand is inconsistent and may comprise both heating and cool-
[30] and aggregated to hourly intervals by taking the average
ing loads (which building HVAC equipment may or may not be
across the five-minute data.
configured to provide). Toronto’s climate is generally heating-
Dummy variables were created to indicate if each observation
dominated, however, cooling loads are not insignificant - for the
occurred during the cooling, heating, or shoulder seasons. The
one-year period beginning August 1st, 2018, Toronto had 3707
value of the dummy variable was based on typical trends of high
3
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

Fig. 2. Characteristic suite layouts in both Building A and Building B. Layout 1 (one bedroom apartment), Layout 2 (one bedroom apartment with set-back bedroom) and
Layout 3 (two-bedroom corner suite).

Fig. 3. Average hourly terminal HVAC unit runtime in Buildings A and B.

Table 2
Measurement intervals, errors, and aggregation/disaggregation methods.

Measurement Measurement Measurement Aggregation/Disaggregation Method Source


Error Interval
Suite Temperature Setpoint (°C) N/A 5-minutes Average Thermostat
Thermostat Mode (Heat, Cool, Off) N/A 5-minutes ‘Heat’ or ‘Cool’ if thermostat reported being in that
mode at any point in the interval, otherwise ‘Off’.
Indoor Air Temperature (°C) ± 0.5 °C 5-minutes Average
Indoor Relative Humidity (%) ± 5% RH 5-minutes Average
Thermostat-Signaled Runtime (seconds) N/A 5-minutes Summation of the seconds of thermostat-reported
runtime for each 5-min interval
Outdoor Air Temperature (°C) Not Available 1-hour N/A Historical Weather
Outdoor Relative Humidity (%) Not Available 1-hour N/A Data [26]
Cloud Cover (Cloudy, Mostly Cloudy, Not Available 3-hours Linear Interpolation
Mostly Clear, or Clear)
Shortwave Solar Radiation ± 16.8% of measurement 1-day Daily value used for all time intervals NASA POWER [29]
(MJ/m2/day)
Longwave Solar Radiation (MJ/m2/day) ± 5.5% of measurement 1-day Daily value used for all time intervals

4
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

and moderate heating and cooling demand in southern Ontario, 3.3.3. Load shifting control
Canada and central building system heating/cooling changeover In the load shifting control scenario, default set-points changes
dates. Specifically, an observation was deemed to take place in were set by the researchers for the overnight period during the
the cooling season if it occurred between June 1 and September cooling season (pre-cooling) as detailed in Table 4. Participants
30, in the heating season if it occurred between November 1 and did not have the option to change the depth of the pre-
February 28, and in the shoulder season if it occurred in any other conditioning set-back but could cancel the hourly scheduled tem-
month. perature changes all-together and revert to their originally sched-
Additionally, suite characteristic data were collected from uled temperature set-point at any time (‘‘strategy override”). The
building floors plans. These data are summarized in Table 3. All aim of load shifting control was to reduce HVAC load between
data were linearly scaled such that all model inputs were between 11:00 and 17:00.
0 and 1. The depth and length of these temperature setbacks were deter-
mined through the use of a simple spreadsheet model of the ther-
3.3. Tested control strategies mal mass of a representative suite layout. For simplicity, the model
considered only the concrete mass in the suite (i.e., concrete sheer
Thermostat data were collected under three scenarios to estab- walls, floor, and ceiling slabs) as the thermal storage potential of
lish HVAC runtime profiles for each suite during: 1) baseline con- other materials in the suite is comparatively negligible. A descrip-
ditions (thermostat operating as a standard programmable tion of this model is included in the Supplemental Information (SI)
thermostat); 2) occupancy-based control conditions (thermostat document.
operating using an occupancy-reactive control strategy); 3) load
shifting control conditions (thermostat operating using a pre- 3.4. Analysis procedure
conditioning strategy).
Changes in suite-based HVAC runtime resulting from the use of
occupancy-based thermostat control and load shifting control
3.3.1. Baseline conditions
were evaluated with respect to the estimated baseline runtime
In the baseline scenario, all thermostat smart features were
which would have occurred if the thermostat was operating in
turned off and it was operated as a standard-programmable ther-
standard-programmable mode. To enable a direct comparison of
mostat. Temperature setpoints were controlled by participants’
a suite’s HVAC runtime using occupancy-based or load shifting
set daily schedule (if any) or as a constant temperature (if partici-
control against the HVAC runtime in the baseline scenario under
pants had not set a schedule).
the same weather conditions, a baseline runtime prediction model
was developed for each suite using baseline runtime and weather
3.3.2. Occupancy-Based control condition data. This prediction model was then used to estimate
In the occupancy-based control scenario, the smart thermostat baseline runtime for the same set of weather conditions observed
‘‘Auto-Away” feature was turned on. This meant that the tempera- during periods in which occupancy-based or load shifting control
ture set-points would follow the participant’s set daily schedule, strategies were used. The estimated baseline control runtime and
when occupancy was detected. However, when occupancy was measured occupancy-based or load shifting control runtime were
not detected for two hours during a scheduled home period, an then compared to assess the change in runtime between the differ-
away temperature set-back was applied in the suite, the depth of ent control scenarios. This change in HVAC runtime from the use of
which was chosen by the suite occupant. The average temperature the occupancy-based control strategy is calculated using Equation
set-back depth was + 2.9 °C in the cooling season. If occupancy was (1):
detected during an auto-away period, the thermostat reverted to
the participant’s set schedule. DRuntimeOccupancyPredicted ¼ RuntimeOccupancyActual  RuntimeBaselinePredicted
ð1Þ
Table 3 Likewise, the difference between the actual runtime resulting
Suite characteristic features used in the regression models. from the load shifting control strategy and the predicted baseline
Suite Characteristic Feature Description runtime from the prediction model, is calculated using Equation
Suite Identifier Unique, numeric identifier which differentiates data (2):
from different suites.
Upper Floor Dummy variable indicating if the suite is located in
DRuntimeLoadShiftingPredicted ¼ RuntimeLoadShiftingActual  RuntimeBaselinePredicted
the top one-third of the building. ð2Þ
Lower Floor Dummy variable indicating if the suite is located in
the bottom one-third of the building. Note that the formulations of Equations (1) and (2) yield reduc-
Building A Dummy variable indicating if the suite is located in tions in runtime as negative values and increases in runtime as
Building A.
positive values.
Corner Suite Dummy variable indicating if the suite is a corner
suite (i.e., the suite has a façade exposed in two Runtime predictions were made for hourly periods using a
orientations). suite’s regression model, however, these predictions were aggre-
North/South/East/ Dummy variables indicating the orientation of the gated over a larger timeframe to increase prediction accuracy (fur-
West suite façade. ther discussed in 3.4.1.3).
Suite Footprint Numeric variable corresponding to the suite’s
estimated floor area based on floor plan drawings.
Shaded Window Fractional variable indicating the linear portion of 3.4.1. Prediction model development and Testing:
the suite’s glazed façade which is shaded by an In the study buildings, which have short terminal HVAC unit
overhang. runtimes on average (as seen in Fig. 3, Section 3.1), accurately pre-
Window-to-Wall Fractional variable indicating the linear portion of
Ratio the suite’s façade which is glazed.
dicting thermostat-signaled HVAC runtime under different
Window-to-Suite Numeric variable representing the ratio of the suite’s weather conditions is a significant challenge. In prior work, we
Depth Ratio window horizontal span relative to suite depth. tested different regression modelling approaches on the dataset
Heating Capacity/ Numeric variable representing the suite-level HVAC and showed that different modelling approaches are needed in dif-
Cooling Capacity unit nameplate capacity for heating or cooling.
ferent suites to improve runtime prediction accuracy [27]. Further,
5
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

Table 4
Pre-cooling and pre-heating temperature set-back schedule used in the summer and winter months for the load shifting scenario. All temperatures are relative to participants’
temperature setpoints.

Cumulative Temperature Set-back


19:00 20:00 21:00 22:00 22:30 23:00 23:30 00:00 02:00 07:00 11:00
Pre-Cooling 0.5 °C 1 °C 1.5 °C 2 °C 2.5 °C 3 °C 3.5 °C 4 °C 5 °C 2 °C 0 °C
Pre-Heating +0.5 °C +1 °C +1.5 °C +2 °C 0 °C

even when a large variety of model types (e.g., random forest, neu- performance metrics were evaluated using the testing dataset,
ral networks, etc.), training methods (e.g., using different subsets of which was never seen by the model during training.
data to train the model), and model tuning processes are consid- An adapted mean absolute percentage error calculation (A-
ered, sufficient predictive accuracy may not be achieved for many MAPE) was used to guide the model development process. The
suites. As such, a mixed-approach was used to develop HVAC run- adapted MAPE formula used to calculate error is shown in Equation
time prediction models for each suite. Three main model develop- (3), where n is the number of test values, At is the actual runtime
ment approaches were tested for each suite: value, A is the average of the test set runtime values, and Ft is the
Approach (1): Develop individual models for each suite which predicted runtime value. Note that the A-MAPE formula was used
are trained using only data collected from that suite (note that due to the existence of numerous actual zero-runtime values in the
suite characteristic features were not included in this modelling dataset, the error for which cannot be evaluated using standard
process). MAPE calculations.

n  
Approach (2): Develop individual models for each suite which 1X 
At  F t 
A  MAPE ¼ ð3Þ
are trained using data collected from all suites for all seasonal n t¼1  A 
periods.
After the models in Approaches 1, 2, and 3 were developed for
Approach (3): Simple stacked model with one Stage 1 Model
each suite, the best-performing model in terms of A-MAPE for each
and one Stage 2 Model.
suite was selected for use as the final prediction model. However,
for most suites, the predictive accuracy of these final models for
For Approaches (1) and (2), models were developed for each
individual hourly periods was insufficient to use for evaluation of
suite using both random forest (RF) and extreme gradient boosting
runtime changes due to occupancy-based or load shifting control.
(XGBoost) [31] regression algorithms, which have been shown to
To improve the accuracy of evaluating changes in runtime due to
have good performance in building energy management applica-
different control strategies, several aggregation methods were
tions [32–37]. Further information on the random forest and
applied to the data (Section 3.4.1.3).
XGBoost models and the rational for their selection can be found
in the SI document. For each model type tested in Approaches 1
and 2, three different training data subsets were tested for each 3.4.1.1. Feature selection. Features were selected out of the avail-
suite: only cooling season data; cooling and heating season data; able data, as outlined in Section 3.2. First, each model was trained
and all available data (i.e., from cooling, heating, and shoulder sea- using only outdoor air temperature as the independent variable.
sons). For each of the Approach 1 and 2 models, the same develop- Next, features were added one at a time to the model. If an individ-
ment process was followed. First, feature selection was performed ual feature improved (i.e., reduced) the A-MAPE of the model, it
(Section 3.4.1.1) followed by hyperparameter tuning (Sec- was selected to remain in the model. For models developed using
tion 3.4.1.2). For Approach (3), the best model from Approach (2) the RF algorithm, feature selection was completed using the
was used as the Stage 1 Model. For the Stage 2 Model, the type default hyperparameters, with the exception of the number of esti-
was selected based on the model type used in the Stage 1 Model. mators, which was set to 1500. For the XGBoost algorithm, the
If the Stage 1 Model was an RF model, the Stage 2 Model was an learning rate was set to 0.1, the maximum depth to 6, the mini-
XGBoost model (as shown in Fig. 4) and vice-versa. The Stage 2 mum child weight to 5, and the maximum number of estimators
Model used the same feature set as the Stage 1 Model, however, to 10,000 with early stopping after 100 rounds and all other
the input dataset for the Stage 2 Model was limited to cooling sea- parameters left as the default values.
son data collected form the individual suite for which the model
was developed.
All tested models were implemented using the Scikit-learn 3.4.1.2. Hyperparameter tuning. Hyperparameter tuning was com-
package in Python [38]. During model development, available pleted using a Bayesian search for the random forest model and
baseline-control data were divided into separate training and test- an exhaustive grid search for the XGBoost model to test different
ing sets (training set = 75% of data; testing set = 25% of data). Model permutations of selected hyperparameter values using two-fold

Fig. 4. Approach 3 process flow for the tested stacked model approach when a random forest regression model is used as the Stage 1 Model.

6
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

cross-validation. Further details on the hyperparameter selection 4.2. Occupancy control strategy performance results
process for individual model classes is included in the SI document.
An average of 1303 occupancy-based control hourly periods for
3.4.1.3. Data Aggregation for improved accuracy. Two different each thermostat were available for the cooling season analysis
aggregation methods were used to generate higher accuracy run- (min: 172 hourly periods, max: 2567 hourly periods). In the cool-
time predictions: 1) aggregation of individual suite data over the ing season under baseline thermostat operation, the average
entire cooling season (‘‘Total Aggregation”); and 2) aggregation of suite-based HVAC runtime was 287 min/day (range: 16.6 min/day
individual suite runtime changes for hourly intervals with all suite to 763 min/day). When occupancy-based control was used, a wide
data in each building (‘‘Building Aggregation”). range of runtime changes as compared to the baseline control sce-
nario were observed in the suites (see Fig. 5). Runtime reductions
as high as 142 ± 13 min/day were observed, along with runtime
4. Results and discussion increases as high as 118 min ± 47 min/day (using Total Aggrega-
tion). On average, suite-based HVAC system runtime during the
In this section, first, the accuracy of the baseline runtime pre- cooling season decreased by 5.9% (15 min/day, standard deviation
diction models is reviewed. Next, the performance of the of ± 117 min/day) over baseline scenario. We hypothesize that, in
occupancy-based and load shifting control strategies are analyzed many suites, runtime reductions were limited by building opera-
and the potential impacts of the control strategies on thermal com- tional patterns, which were previously shown to result in low
fort and dehumidification are discussed. suite-based HVAC runtimes and frequent over-conditioning of
suites, despite available thermostat temperature controls [39].
4.1. Prediction model accuracy However, when a t-test was used to evaluate if there was any sta-
tistically significant difference in runtime changes between the
The A-MAPE for each thermostat model for the cooling season two groups, suites with lower than average daily HVAC runtime
data was between 9% and 83% (average error of 31%), when indi- had higher percent change in runtime as compared to suites with
vidual hourly periods were considered for each suite. As the run- higher than average daily HVAC runtimes (p = 0.24).
time prediction errors remained high for individual hourly A difference between occupancy-based control strategy perfor-
periods, Total Aggregation was primarily used for evaluating mance in the two study buildings was observed. Higher runtime
changes in runtime between the baseline control scenario and reductions were achieved through the use of the occupancy-
the occupancy-based control scenario in aggregate over the entire based control strategy in Building B as compared to Building A.
testing period for each suite (Section 4.1.1) and, when hourly data On average, in Building B, runtime decreased by 8.3% (19 min/day)
was required, Building Aggregation was used for evaluation of both and, in Building A, runtime decreased by 3.1% (9.6 min/day) using
the occupancy-based and load shifting control scenarios occupancy-based control. It is unclear why this difference
(Section 4.1.2). occurred, as the average runtime in Building B was lower in the
baseline scenario (average = 250 min/day) as compared to Building
A (average = 382 min/day) and participant demographics were
4.1.1. Total Aggregation similar in the two buildings. Further, in previous work, suites in
The average error with Total Aggregation was 10% (range: 0.4% Building B were observed to have more frequent periods of over-
to 115%). Nine suites had HVAC runtime prediction errors over 15% conditioning in the cooling season as compared to Building A [39].
when Total Aggregation was used, which were excluded from fur- The hourly percent change in runtime for Buildings A and B
ther cooling season analysis. Seven of these nine suites had very when Building Aggregation was used is shown in Fig. 6. Runtime
low HVAC runtimes during the baseline scenario (average of decreases due to the use of occupancy-based control were predom-
2.8 min/hour) when compared to the average across all suites inately observed during the midday period, when suites are fre-
(11.5 min/hour), which likely explains the high percentage error. quently unoccupied, between the hours of 08:00 and 18:00
Further, due to the minimal existing suite loads, HVAC runtime (average runtime reduction of 11 ± 8%), while runtime increases
reductions in these seven low-runtime suites would have minimal were predominately observed during the overnight period,
impact on building energy use. The reason for the high prediction between the hours of 23:00 and 05:00 (average runtime increase
error in the remaining suites was not easy to discern, however, it of 4 ± 5%). Notably, the estimated average increase in runtime dur-
may be a result of a combination of lower than average HVAC run- ing the overnight hours was within the bounds of the prediction
times (7 min/hour), high variation in runtime due to frequent occu- error for that time period and may not actually represent an actual
pant control overrides, and/or changes in occupant behaviour increase in runtime due to the control strategy. However, this gen-
throughout the study period. When the runtime prediction error eral pattern (runtime decreases during midday hours and small
under the Total Aggregation scenario for the cooling season was runtime increases during overnight hours) is amenable with elec-
considered with the seven suites omitted, the average prediction tricity grid operations in regions similar to Ontario Canada, where
error across suites was reduced to 4.6%. Ultimately, data from 41 on-peak hours (and associated higher marginal electricity genera-
suites (84% of suites) were available for analysis for the cooling tion costs and emissions) occur during midday periods and off-
season when Total Aggregation is used, based on model predictive peak hours occur overnight.
performance.
4.2.1. Remarks on efficacy of occupancy-based control in high-rise
4.1.2. Building Aggregation residential buildings
When Building Aggregation was used, the average baseline run- While the observed suite-based HVAC runtime reductions
time prediction accuracy was 4.6% in Building A and 6.1% in Build- resulting from the use of occupancy-based control were moderate
ing B for all hourly periods. In Building A, the minimum average in the cooling season, there may be an opportunity to increase run-
hourly error occurred at 21:00 (0.8% prediction error) and the max- time reductions (and associated building energy use reductions) by
imum average hourly error occurred at 13:00 (11.7% prediction changing how the central space conditioning systems in these two
error). In Building B, the minimum average hourly error occurred buildings are operated. Over-conditioning of the suites was
at 19:00 (1.6% prediction error) and the maximum average hourly observed in both buildings during the heating and cooling seasons
error occurred at 07:00 (10.9% prediction error). [39], which resulted in low suite-based HVAC runtimes and mini-
7
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

Fig. 5. Runtime changes from occupancy-based control in each of the 49 studied suites. Each bar shows the result for one of the 49 suites.

Fig. 6. Percent runtime change due to occupancy-based control in the cooling season for Building A (nsuites = 23) and Building B (nsuites = 26) by hour using Building
Aggregation. Error bars indicate range of uncertainty in hourly runtime change based on model prediction accuracy using Building Aggregation.

mal opportunity for runtime reductions through occupancy-based to reduce building HVAC loads by applying this strategy in contem-
control mechanisms. Reducing space conditioning provided to porary high-rise residential buildings is limited, given the frequent
building common areas (i.e., through conditioned corridor ventila- lack of on-going commissioning in this building type.
tion) could shift the conditioning burden to the suites and reduce
the energy use by central building systems. This approach would
4.3. Load shifting strategy performance
allow individual suite-based HVAC systems to be more responsive
to demand and only provide conditioned air where and when it is
Load shifting HVAC control may be used to reduce HVAC loads
needed. After this building tuning, occupancy-based thermostat
during electricity grid peak periods by over-conditioning spaces
controls in suites would likely be more effective in reducing overall
overnight. When the performance of the load shifting strategy
building HVAC loads. However, without this tuning, the potential
was assessed on average across all suites, regardless of whether
8
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

Fig. 7. Percent runtime change due to load shifting-based control in the cooling season for Building A (nsuites = 23) and Building B (nsuites = 26) by hour using Building
Aggregation. Error bars indicate range of uncertainty in hourly runtime change based on model prediction accuracy using Building Aggregation.

occupants chose to cancel the strategy setpoint changes at any As the magnitude and time distribution of suite HVAC loads
point, a 6% decrease in HVAC load between the hours of 11:00 vary based on the cardinal direction the suite faces, it is worth-
and 17:00 was observed (Fig. 7). To achieve this increase, however, while to consider the impacts of the load shifting strategy on suites
an average daily load increase of 36% was required. facing each cardinal direction for test days when few (less than two
cumulative hours) strategy overrides occurred (Fig. 8). Data from

Fig. 8. Average suite HVAC runtime in the baseline scenario and using the load shifting control strategy during cooling season for suites facing each cardinal direction for days
when no strategy overrides occurred: North (nsuites = 3), South (nsuites = 2), East (nsuites = 3), West (nsuites = 10).

9
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

18 suites were available for analysis (three north-facing suites, two and the west-facing suite showed a 44% reduction during the same
south-facing suites, three east-facing suites, and ten west-facing time period. The east-facing suite did not show a reduction in
suites). The timing of the periods of load reductions and increases HVAC load between the hours of 11:00 and 17:00 overall, however,
varied substantially between the groups due to differences in when HVAC loads were reduced by 100% between the hours of 12:00 and
direct solar radiation is incident on the suites’ exterior facades. In 13:00. To achieve this, overall daily HVAC loads in the west- and
this scenario, pre-existing short HVAC runtimes limited the oppor- east-facing suites were increased by approximately 80% each and
tunities for absolute reduction in runtime in south-facing suites. Of the HVAC load in the north-facing suite increased by 39%. Of the
the suites facing the other cardinal directions, north-facing suites suites with data from periods with few strategy overrides, two of
exhibit the most useful shifted load profile, in which HVAC load the three north-facing suites, one of the three east-facing suites,
between the hours of 11:00 and 15:00 was decreased by 51% with four of the ten west-facing suites, and both of the two south-
an overnight (18:00 to 10:00) load increase of 185%. Better man- facing suites, showed substantial reductions for some portion of
agement of solar gains (e.g. through improved window shade use the on-peak hours between 11:00 and 17:00. It is not clear why
or solar films), would likely improve the effectiveness of this strat- the load shifting strategy was effective in some suites and not
egy and allow for improvements to mid-day load reductions in others. There did not appear to be a relationship between average
suites facing the other three cardinal directions. While on average, baseline suite HVAC runtime, suite direction, or which building the
across all of the studied suites, the load shifting strategy was lar- suite was in. Usage patterns and type of window coverings
gely ineffective at reducing peak period loading, north-facing installed in the suite and occupant use of auxillary devices are
suites in general as well as some individual suites facing other potential reasons for this difference.
directions demonstrated significant load reductions during the
daytime peak periods. As an example, three individual suites facing
north, east, and west are shown in Fig. 9 (note that the south-facing 4.3.1. Potential for GHG emissions reductions from load shifting
suites are excluded as the orginal and shifted load profiles do not While GHG emissions reductions from load shifting control will
substantially vary from the aggregate profiles shown in Fig. 8). vary significantly based on suite load profiles and the jurisdiction
In the suites shown in Fig. 9, the north-facing suite showed a the electricity is consumed in, it is still worthwhile to consider
66% HVAC load reduction between the hours of 11:00 and 17:00 an approximation of the GHG impacts of this strategy. For the pur-
pose of this estimate, the HVAC runtime changes over a one-day

Fig. 9. Suite HVAC runtime in the baseline and load shifting scenarios during cooling season for a north-facing suite in Building B and an east-facing and west-facing suite in
Building A.

10
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

period from the north-, west-, and east -facing suites shown in GHG emissions factors associated with marginal electricity con-
Fig. 9 are considered and marginal emissions factors based on sumptions will substantially how this control strategy impacts
the Ontario electricity grid are used. Note that the energy and total GHG emissions. Therefore, detailed consideration of grid-
GHG emission calculations for the north-facing suite (located in specific MEFs which take into variation due to changing grid
Building B) assume the use of a fan coil and chiller system, as used demand and generation conditions should be used to assess when
in Building A, for consistency of the example amongst the four and where use of the load shifting control strategy is needed.
suites. Marginal emissions factors (MEFs) represent the GHG emis- While load shifting control does not make sense as an emissions
sions (increase or decrease) associated with any increase or reduction strategy for the Ontario electricity grid based on the
decrease in electricity consumption during a specific period. MEFs studied populations, this control strategy has the potential reduce
are dependent the operating conditions of the electricity grid and peak loading during periods when the grid capacity is constrained
vary with respect to the time of day and season, as well as year- or when building HVAC systems do not have sufficient capacity to
to-year based on generation infrastructure and demand changes. meet peak daytime conditioning loads (e.g., during an extreme
As MEFs must be estimated based on grid operation data and can heat event). However, careful consideration of suite characteristics
vary significantly, several different hourly MEFs were tested. These which drive baseline and shifted load profiles and strategy efficacy
include marginal emissions factors estimated by two different (e.g., suite orientation) is needed. Further study of a larger sample
sources: 1) the summer seasonal average hourly MEF for 2018 by size of suites is needed to better understand the key mechanisms
The Atmospheric Fund [2]and 2) monthly-average hourly MEFs which influence strategy performance. The use of night-time, pas-
based on 2017 data estimated by Gai et al. [40]. To calculate the sive free-cooling via increased outdoor air ventilation may also
energy changes from strategy use in the example suite, the manu- provide opportunities for emissions reductions. However, the nat-
facturer nameplate cooling capacity of the suite terminal HVAC ure of ventilation systems in this building type regionally (i.e., the
unit (2.7 kW in the north-facing suite, 5.7 kW in the west-facing use of pressurized corridor ventilation systems with no provision
suite, and 6.0 kW in the east-facing suite) was used to calculate of direct ventilation to suites) and the typically high relative
the hourly energy use within the suite based on an assumed pri- humidity of outdoor air during summer nights in the region may
mary energy use efficiency of 65% (accounting for fan coil opera- limit the effectiveness of this approach and necessitates further
tion and central system energy use by the chiller, pumps, etc.). investigation.
Equation (4) was used to calculate hourly energy use:

EnergyUseHour ¼ RuntimeHour xNameplateCoolingCapacityxPrimaryEnergyUseEfficiency ð4Þ


4.4. Thermal comfort and dehumidification impacts
The hourly change in GHG emissions between the baseline sce-
nario and the load shifting scenario was calculated using Equation While not discussed in detail this paper, the impacts of the
(5): occupancy-based and load shifting control strategies on occupant
thermal comfort in the studied suites was assessed in previous
DGHGEmissionsHour ¼ MEF Hour ðEnergyUseLoadShifting;Hour work by the authors [41]. Based on analysis of survey data and
occupant-initiated control strategy overrides (e.g., the use of ther-
 EnergyUseBaseline;Hour Þ ð5Þ
mostat ‘‘holds”), no significant change in occupant thermal comfort
Changes in GHG emissions from the use of the load shifting in the suites was observed between periods of occupancy-based
strategy in the example suites are shown for the different tested control and periods of baseline thermal control. While no survey
MEFs in Table 5. The changes in emissions estimated vary signifi- data were available to evaluate any change in thermal sensation
cantly based on the MEFs used, however, for the example west- when the load shifting control strategy was used during the cool-
and east-facing suites, all scenarios resulted in an overall increase ing season, participants frequently overrode the load shifting con-
in GHG emissions. For example, in the north-facing suite, when the trol strategy, which indicates that they experienced some degree of
seasonal MEFs for summer 2018 were used, a net increase in GHG thermal discomfort beyond their tolerance level. Further discus-
emissions of 412% was found, however when the monthly MEFs for sion of the thermal comfort impacts of the control strategies can
July 2017 were used, a GHG emissions reduction of 11% was found. be found in [41].
Based on this, the load shifting HVAC control strategy does not The effects of the tested control strategies on the dehumidifica-
appear to have the potential to reduce GHGs emissions in the tion of indoor air relative to outdoor air dew point temperatures
Ontario electricity market, however, in the future, as MEFs change during the cooling season was also considered. Fig. 10 shows the
in the Ontario market, or in different jurisdictions with different difference in suite dew point temperatures and outdoor air dew
electricity grid operations, very different results may be found. point temperatures for periods in which the thermostats were
The timing of on-peak periods and the ratio of on– to off-peak operated using each of the three control strategies. Note that the

Table 5
Estimated suite HVAC energy use changes and GHG emissions changes for a one day period resulting from the use of load shifting strategy in the example suite.

Summer 2018 June 2017 July 2017 August 2017 September 2017
Total One-day Change in Energy Use North-Facing: +39 kWhWest-Facing: +52
from Load-Shifting kWhEast-Facing: +53 kWh
Change in GHG Emissions North-Facing
+7722 g CO2 eq (+412%)* +231 g CO2 eq 201 g CO2 eq (- +2419 g CO2 eq +2133 g CO2 eq
(+12%)* 11%)* (+129%)* (+114%)*
West-Facing
+10,353 g CO2 eq (+269%)* +454 g CO2 eq +950 g CO2 eq +3828 g CO2 eq +4749 g CO2 eq
(+12%)* (+25%)* (+100%)* (+123%)*
East-Facing
+11,147 g CO2 eq (+500%)* +1837 g CO2 eq +4863 g CO2 eq +5341 g CO2 eq +6819 g CO2 eq
(+82%)* (+218%)* (+240%)* (+306%)*
MEF Source [2] [40]

*Based on estimated baseline energy use and average hourly emissions factors for 2018 from [2].

11
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

Fig. 10. Difference between outdoor and in-suite dew point temperatures during different thermostat control periods in Building A and B.

dehumidification data for the load shifting control strategy the indoor air: modelled surface temperature data suggest a lower
excludes test periods in which an individual suite overrode the limit on wall surface temperature of around 22 °C, while indoor
control strategy for more than two-hours. dewpoint temperatures reached a maximum of 21 °C and only
exceeded 20 °C 10 % of the time during load shifting control peri-
4.4.1. Occupancy-based control ods. That being said, given the high indoor dew point temperatures
A significant difference in dehumidification between periods of in study suites, surface condensation may occur during some time
occupancy-based control and baseline control was observed in periods due to variation in thermal mass surface temperatures
both study buildings (Building A p-value < 0.001, Building B p- between suites and differences in outdoor conditions. However,
value < 0.001). In Building A, occupancy-based control was associ- even during baseline operation, high humidity is present in both
ated with increased dehumidification as compared to the baseline of the study buildings. The 2019 revisions of ASHRAE Standard
scenario, with a small effect size (Hedge’s g = 0.19). In Building B, 62.1 suggests that a maximum dew point of 15 °C be maintained
periods in which occupancy-based control was used were associ- in mechanically cooled buildings [42]. During baseline control
ated with less dehumidification on average, however, in practical- periods, this threshold was exceeded on average 48% of the time
ity, due to the small effect size of this change (Hedge’s g = 0.06), no in Building A and 29% of the time in Building B during cooling sea-
significant impact was observed. son. As such, indicate that additional dehumidification is needed in
both study buildings to keep indoor dew points below the recom-
mended 15 °C maximum, which, when implemented, would
4.4.2. Load shifting based control reduce the likelihood of surface condensation due to load shifting
To understand the impacts of the load shifting control strategy control.
as designed, only data from test periods during which a suite did
not override the strategy for at least 22 out of 24 h were considered
in this portion of the analysis. In both buildings, the load shifting 5. Conclusions
control strategy was associated with increased dehumidification
as compared to the baseline scenario (Building A p-value < 0.001, In this study, the energy performance of occupancy-based and
Building B p-value < 0.001). In Building A, the load shifting control load shifting thermostat control strategies was tested using smart
strategy resulted in a large increase dehumidification (Hedge’s thermostats in 49 suites across two contemporary high-rise resi-
g = 1.11), corresponding to an incremental decrease in indoor air dential buildings in Toronto, Canada during the cooling season.
dew point of approximately 1.8 °C typically. In Building B, the Data were collected on HVAC terminal unit runtime and weather
increase in dehumidification was smaller (Hedge’s g = 0.44), corre- conditions under three thermostat control scenarios: 1) baseline
sponding to an incremental decrease in indoor dew point of control, in which the thermostat was operated as a standard pro-
approximately 1.5 °C on average. gramable thermostat; 2) occupancy-based control using the ther-
While the load shifting control strategy resulted in improved mostat’s built-in occupancy detection features; and 3) a novel
dehumidification overall, surface condensation on cooled thermal load shifting control strategy, in which pre-conditioning was used
mass may be of concern under some outdoor conditions. In gen- to shift HVAC load from peak day time periods to overnight peri-
eral, the surface temperatures of suite thermal mass during load ods. Using black-box regression models, measured suite HVAC run-
shifting control are expected to remain above the dew point of time during occupancy-based and load shifting control periods
12
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

were compared with prediction runtime for the baseline control References
scenario during the same outdoor and in-suite conditions. We
found: [1] The Population Division of the Department of Economic and Social Affairs of
the United Nations, ‘‘World Urbanization Prospects 2018,” 2018. https://
population.un.org/wup/ (accessed Nov. 21, 2019).
 Substantial variation in control strategy performance was [2] The Atmospheric Fund, ‘‘A Clearer View on Ontario’s Emissions: Electricity
observed for different suites, which is likely driven by suite emissions factors and guidelines,” 2019. [Online]. Available: https://taf.ca/wp-
content/uploads/2019/06/A-Clearer-View-on-Ontarios-Emissions-June-2019.
usage patterns and occupant behaviours (including the use of pdf.
window coverings); [3] Canadian Mortgage and Housing Corporation, ‘‘Canada Mortgage and Housing
 On average, across the study population, occupancy-based con- Corporation, housing starts, under construction and completions, all areas,
quarterly.” https://www150.statcan.gc.ca/t1/tbl1/en/tv.action?
trol reduced suite HVAC runtime by 5.9% (standard
pid=3410013501&pickMembers%5B0%5D=1.1&pickMembers%5B1%5D=4.1
deviation = 46%); (accessed Nov. 13, 2019).
 The largest change in suite HVAC runtime occurred during the [4] K. Seemann, ‘‘Energy Management Strategies for MURB & Commercial
daytime period (08:00 to 18:00), during which an average Buildings.” [Online]. Available: https://www.enercare.ca/sites/default/files/
assets/library/pdfs/
reduction in runtime of 11% was observed; energy_management_strategies_for_murb_commercial_buildings.pdf.
 This suggests that installation of smart thermostats may be a [5] A. Meier, C. Aragon, T. Peffer, D. Perry, M. Pritoni, Usability of residential
cost-effective energy conservation measure in this type of thermostats: Preliminary investigations, Build. Environ. 46 (10) (2011) 1891–
1898, https://doi.org/10.1016/j.buildenv.2011.03.009.
building, however, building re- or retro-commissioning should [6] The Cadmus Group, ‘‘Evaluation of the 2013–2014 Programmable and Smart
be completed first to ensure space conditioning loads are being Thermostat Program,” Evansville, Indiana (US), 2015. [Online]. Available:
met by in-suite HVAC systems as opposed to common area con- http://www.cadmusgroup.com/wp-content/uploads/2015/06/
Cadmus_Vectren_Nest_Report_Jan2015.pdf?submissionGuid=7cbc76e9-41bf-
ditioning systems; 459a-94f5-2b13f74c4e52.
 On average, load shifting-based control did not substantially [7] N. Lieb and S. Dimetrosky, ‘‘Thriller in Asilomar: Battle of the Smart
reduce HVAC runtime during on-peak hours (11:00 to 17:00) Thermostats,” pp. 1–13, 2016.
[8] Apex Analytics LLC, ‘‘Energy Trust of Oregon Smart Thermostat Pilot
in the studied suites; Evaluation,” Oregon (US), 2016. [Online]. Available: https://energytrust.org/
 In 10 of the studied suites, substantial reductions in HVAC run- wp-content/uploads/2016/12/Smart_Thermostat_Pilot_Evaluation-Final_wSR.
time for portions of on-peak periods were observed, however, to pdf.
[9] D. Parker, K. Sutherland, D. Chasar, ‘‘Evaluation of the Space Heating and
achieve these reductions large increases in overnight HVAC run-
Cooling Energy Savings of Smart Thermostats in a Hot-Humid Climate using
time was required; Long-term Data”, no, Nevius 2000 (2016) 1–15.
 Based on marginal emissions factors for the Ontario electricity [10] Southern California Edison, Nest Labs, ‘‘Working Paper: Residential Smart
grid, even in suites with substantial on-peak HVAC runtime Thermostat,” 2017. [Online]. Available: https://www.peakload.org/assets/
SCE17HC054.0_Residential_Sma.pdf.
reductions, the load shifting control strategy was shown to [11] M. Pritoni, J.M. Woolley, M.P. Modera, Do occupancy-responsive learning
increase GHG emissions in the majority of scenarios; and thermostats save energy? A field study in university residence halls, Energy
 While load shifting control does not make sense as an emissions Build. 127 (2016) 469–478, https://doi.org/10.1016/j.enbuild.2016.05.024.
[12] H. Stopps and M. F. Touchie, ‘‘Residential smart thermostat use : An
reduction strategy based on the studied populations, this con- exploration of thermostat programming , environmental attitudes , and the
trol strategy has the potential to be employed to reduce peak influence of smart controls on energy savings,” Energy Build., vol. 238, no.
loading during periods in which grid capacity is constrained 110834, 2021, doi: 10.1016/j.enbuild.2021.110834.
[13] A. Rabl, L.K. Norford, Peak load reduction by preconditioning buildings at
or when building HVAC systems do not have sufficient capacity night, Int. J. Energy Res. 15 (9) (1991) 781–798, https://doi.org/10.1002/(ISSN)
to meet peak daytime conditioning loads (e.g., during an 1099-114X10.1002/er.v15:910.1002/er.4440150909.
extreme heat events). [14] W. Andrews, M. Piraino, J. Strasser, Laboratory testing of control strategies to
reduce peak air-conditioning loads, ASHRAE Trans. 99 (pt 1) (1993) 518–528.
[15] R. Becker, M. Paciuk, Inter-related effects of cooling strategies and building
features on energy performance of office buildings, Energy Build. 34 (1) (2002)
25–31, https://doi.org/10.1016/S0378-7788(01)00081-0.
Declaration of Competing Interest [16] F.B. Morris, J.E. Braun, S.J. Treado, Experimental and simulated performance of
optimal control of building thermal storage, ASHRAE Trans. 100 (1) (1994)
The authors declare that they have no known competing finan- 402–414, https://doi.org/10.1016/0140-6701(95)93503-7.
[17] J.E. Braun, Load control using building thermal mass, J. Sol. Energy Eng. Trans.
cial interests or personal relationships that could have appeared
ASME 125 (3) (2003) 292–301, https://doi.org/10.1115/1.1592184.
to influence the work reported in this paper. [18] P. Blondeau, M. Spérandio, F. Allard, Night ventilation for building cooling in
summer, Sol. Energy 61 (5) (1997) 327–335, https://doi.org/10.1016/S0038-
092X(97)00076-5.
Acknowledgements [19] R. Yin, P. Xu, M.A. Piette, S. Kiliccote, Study on Auto-DR and pre-cooling of
commercial buildings with thermal mass in California, Energy Build. 42 (7)
The authors would like to the participating condominium (2010) 967–975, https://doi.org/10.1016/j.enbuild.2010.01.008.
[20] P. E. Peng Xu, ‘‘Case study of demand shifting with thermal mass in two large
boards, suite owners and renters, and property management per- commercial buildings,” ASHRAE Trans., vol. 115 PART 2, pp. 586–598, 2009.
sonnel for their assistance in acquiring data for the study. The [21] V. Geros, M. Santamouris, A. Tsangrasoulis, G. Guarracino, Experimental
authors would also like to thank ecobee for their support, funding evaluation of night ventilation phenomena, Energy Build. 29 (2) (1999) 141–
154, https://doi.org/10.1016/S0378-7788(98)00056-5.
and hardware contributions throughout the project and Enbridge [22] M. Bojić, D. Cvetković, V. Marjanović, M. Blagojević, Z. Djordjević,
and Toronto Hydro for their financial and in-kind support. His Performances of low temperature radiant heating systems, Energy Build. 61
work was funded in part by the Natural Sciences and Engineering (2013) 233–238, https://doi.org/10.1016/j.enbuild.2013.02.033.
[23] M. Kintner-Meyer, A.F. Emery, Optimal control of an HVAC system using cold
Research Council of Canada (NSERC) (RGPIN-2016-06325), The storage and building thermal capacitance, Energy Build. 23 (1) (1995) 19–31,
Atmospheric Fund (TAF), the Independent Electricity System Oper- https://doi.org/10.1016/0378-7788(95)00917-M.
ator (IESO), and the Ontario Centres of Excellence (OCE) Voucher in [24] K. Lee, J.E. Braun, A Data-Driven Method for Determining Zone Temperature
Trajectories that Minimize Peak Electrical Demand, ASHRAE Trans. 114 (2008)
Innovation Program (VIP). This work has benefitted from the
1–15.
thoughtful discussions with our colleagues in IEA EBC Annex 79. [25] W.J.N. Turner, I.S. Walker, J. Roux, Peak load reductions: Electric load
shifting with mechanical pre-cooling of residential buildings with low
thermal mass, Energy 82 (2015) 1057–1067, https://doi.org/10.1016/j.
Appendix A. Supplementary data energy.2015.02.011.
[26] Government of Canada, ‘‘Past weather and climate: Historical Data.”
https://climate.weather.gc.ca/historical_data/search_historic_data_e.html
Supplementary data to this article can be found online at (accessed Nov. 11, 2019).
https://doi.org/10.1016/j.enbuild.2021.111644.
13
H. Stopps and Marianne F Touchie Energy & Buildings 255 (2022) 111644

[27] H. Stopps, M.F. Touchie, Smart Choice or Flawed Approach? An Exploration of [36] F. Divina M. García Torres F.A. Goméz Vela J.L. Vázquez Noguera 12 10 1934
Connected Thermostat Data Fidelity and use in Data-Driven Modelling in 10.3390/en12101934
High-Rise Residential Buildings, Under Rev. (2021). [37] R. Bekkerman, ‘‘The present and the future of the kdd cup competition: an
[28] P. W. Stackhouse et al., ‘‘POWER Release 8 . 0 . 1 (with GIS Applications) outsider’s perspective.,” 2015. .
Methodology (Data Parameters , Sources , & Validation),” 2018. [38] F. Pedregosa et al., Scikit-learn : Machine Learning in Python, J. Mach. Learn.
[29] P.W. Stackhouse et al., POWER Release 8.0.1 (with GIS Applications) Res. 12 (2011) 2825–2830.
Methodology, NASA Langley Research Center (2018). [39] H. Stopps M.F. Touchie Managing thermal comfort in contemporary high-rise
[30] ‘‘SolarPy.” 2019. residential buildings: Using smart thermostats and surveys to identify energy
[31] T. Chen, C. Guestrin, ‘‘XGBoost, A Scalable Tree Boosting System” (2016). efficiency and comfort opportunities Build. Environ. 173 November 2020 2019,
[32] S. Lu, Q. Li, L.i. Bai, R. Wang, Performance predictions of ground source heat p. 106748 10.1016/j.buildenv.2020.106748
pump system based on random forest and back propagation neural network [40] Y. Gai, A.n. Wang, L. Pereira, M. Hatzopoulou, I.D. Posen, Marginal Greenhouse
models, Energy Convers. Manag. 197 (2019) 111864, https://doi.org/10.1016/j. Gas Emissions of Ontario ’ s Electricity System and the Implications of Electric
enconman.2019.111864. Vehicle Charging, Environ. Sci. Technol. 53 (13) (2019) 7903–7912, https://doi.
[33] M.W. Ahmad, M. Mourshed, Y. Rezgui, Trees vs Neurons : Comparison between org/10.1021/acs.est.9b0151910.1021/acs.est.9b01519.s001.
random forest and ANN for high-resolution prediction of building energy [41] H. Stopps, M.F. Touchie, Perceptions of thermal conditions in contemporary
consumption, Energy Build. 147 (2017) 77–89, https://doi.org/10.1016/j. high-rise apartment buildings under different temperature control strategies
enbuild.2017.04.038. Perceptions of thermal conditions in contemporary high-rise apartment
[34] T. Ahmad, H. Chen, Energy & Buildings Short and medium-term forecasting of buildings, Sci. Technol. Built Environ. 27 (10) (2021) 1492–1504, https://doi.
cooling and heating load demand in building environment with data-mining org/10.1080/23744731.2021.1929465.
based approaches, Energy Build. 166 (2018) 460–476, https://doi.org/10.1016/ [42] S. T. Taylor et al., ‘‘Ventilation for Acceptable Indoor Air Quality,” 2019.
j.enbuild.2018.01.066. [43] Google Maps, ‘‘Google Maps,” 2017.
[35] L.M. Candanedo, V. Feldheim, D. Deramaix, Data driven prediction models of
energy use of appliances in a low-energy house, Energy Build. 140 (2017) 81–
97, https://doi.org/10.1016/j.enbuild.2017.01.083.

14

You might also like