Applied Energy 215 (2018) 31–40

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Evaluating approaches for district-wide energy model calibration T

considering the Urban Heat Island effect
⁎
Luis Guilherme Resende Santosa, Afshin Afsharia, , Leslie K. Norfordb, Jiachen Maob
a
Masdar Institute of Science and Technology, Abu Dhabi 54224, United Arab Emirates
b
Massachusetts Institute of Technology, Cambridge, MA 02139, United States

H I G H L I G H T S

• Evaluation of the impact of the urban context via district-scale modelling and inverse parameter estimation (calibration).
• Automated calibration process for 56 buildings in a representative district in Abu Dhabi.
• AInforms
straightforward methodology to calibrate a large number of buildings, considering the urban microclimate condition.
• city planners, energy engineers, and governmental agencies on building energy efficiency and UHI intensity.

A R T I C L E I N F O A B S T R A C T

Keywords: Over the past decade, the building energy research community has increasingly focused on urban-scale models.
Energy modelling The shortcomings of analyzing isolated buildings in an urban context are well-known and far from negligible,
Urban energy model mainly due to the inability to account for Urban Heat Island (UHI), shading from neighboring obstructions, and
EnergyPlus obstructed wind flow. The aim of this paper is to evaluate the impact of the urban context via urban-scale
District-wide model calibration
modelling and inverse parameter estimation (calibration) using metered building energy consumption. We de-
Genetic algorithm
scribe an automated calibration process for modelling 56 buildings in a representative district in downtown Abu
Urban heat Island
Dhabi (UAE), where a detailed energy audit was conducted with data from 2008 to 2010. Since the urban
ambient air temperature could differ significantly from the reference rural air temperature used in most building
simulations, the calibration procedure will also consider this UHI effect. Two main approaches of district-wide
energy model calibration are proposed using a genetic algorithm and compared to a baseline case where UHI is
not considered. The first approach estimates seven building-related parameters together with four microclimate-
related variables (describing annual average and seasonal variation of the UHI effect on both air temperature
and humidity). The second one uses the Urban Weather Generator (UWG) to pre-process the urban EPW file,
thereby reducing the number of the parameters to be estimated. In addition, two approaches are investigated for
the calculation of the ASHRAE Guideline 14 calibration error metrics (CvRMSE and NMBE). One approach is to
look at the whole district as one aggregate building, while the other, introduced for the first time herein, consists
in deriving the weighted-average of the error of each building. The main contribution of this study is to provide
simultaneous calibration for multiple buildings in the same district—and subject to the same UHI intensity.
Hence, the UHI intensity (urban-rural temperature/humidity differential) is estimated alongside other calibra-
tion parameters. For the weighted average approach, CvRMSE found is between 19.09% and 19.40%, while
NMBE is between 16.24% and 16.39%. For the aggregated building, CvRMSE is between 2.71% and 4.04%,
while NMBE is between 1.95% and 2.35%.

1. Introduction quantified throughout the world in cities with various morphologies

and climates. This phenomenon is a major determinant of the urban
The rapid development and growth of urban areas has caused sig- microclimate and influences, sometimes significantly, the indoor space
nificant impacts on the built environment. One of the consequences is conditioning energy use in buildings.
the Urban Heat Island (UHI) phenomenon, which has been studied and Physics-based building energy models (BEM) have been used

⁎
Corresponding author.
E-mail address: aafshari@masdar.ac.ae (A. Afshari).

https://doi.org/10.1016/j.apenergy.2018.01.089
Received 15 November 2017; Received in revised form 24 January 2018; Accepted 28 January 2018
Available online 20 February 2018
0306-2619/ © 2018 Elsevier Ltd. All rights reserved.
L.G.R. Santos et al.
extensively in research and engineering for the estimation of heating/
cooling energy demand. BEMs take into account various heat exchanges
that may occur within the building and between the building and its
immediate environment. BEM parameters are often derived from design
specifications. Commonly used BEM software engines include DOE-2,
EnergyPlus, and TRNSYS.
In the case of existing buildings with a recorded history of opera-
tion, a calibration process can be conducted to better estimate uncertain
BEM parameters. Calibration can be either manual or automated. The
goal of calibration is to reduce the discrepancy between model outputs
(e.g. simulated building energy consumption) and measurements ex-
tracted from real buildings (e.g., utility meters). Once the model is
properly calibrated, it could potentially allow for building performance
evaluations under different conditions, enabling the comparison of al-
ternative design options and retrofit scenarios.
However, one recognized obstacle is that the current BEM practice
rarely incorporates the specific urban microclimatic characteristics. In
the quasi totality of cases, buildings are simulated in isolation without
any surrounding buildings and ‘rural’ environment is assumed. The
weather files used in simulations are usually extracted from local air-
port weather stations. This study proposes several adaptations to the
common practice in order to model and calibrate BEMs considering the
urban microclimate at a district level.
We modelled a district in downtown Abu Dhabi (UAE) composed of
67 buildings for which relatively detailed information was gathered in
the course of an energy audit conducted in 2011. We investigate two
novel district-wide calibration approaches that incorporate the UHI
effect and compare to a more conventional calibration with rural
weather. The first proposed approach uses an urban energy model
wherein the urban-rural temperature/humidity differential due to UHI
will be estimated during the calibration process together with the
building-related properties. The second proposed approach involves a
pre-processing of the reference rural weather data using the Urban
Weather Generator (UWG) [1] before the calibration process. The
latter, then, only tunes the building-related parameters.
The main novelty of this study, therefore, is that it proposes a dis-
trict-wide energy model calibration procedure accounting for the UHI.
Through the introduction of new error metrics similar to those set forth
in ASHRAE Guideline 14, the calibration attempts to simultaneously
minimize the global district level modelling error and the building level
modelling error.
2. Literature review
2.1. Urban Heat Island
The Urban Heat Island (UHI) effect has been observed in many cities
worldwide. It is a complex phenomenon that results in urban tem-
peratures that are, in average, higher than surrounding rural areas. The
UHI is caused by several factors, including increased absorption of the
short-wave radiation (from reduced solar reflectance of urban surfaces),
increased sensible heat storage (in urban structures), generation of
anthropogenic heat (mainly from air-conditioning system heat rejection
and motorized vehicles), reduced evapotranspiration (due to the lack of
vegetation in urban areas), reduced sensible heat losses (due to reduced
wind speeds in urban canyons), etc. [2]. In general, there is a growing
concern about the UHI effect, especially during the summer when
outdoor air temperatures are considerably higher and heat waves may
endanger public health and aggravate air pollution. Furthermore, the
UHI impacts the building indoor space conditioning load [3]. To
quantify these impacts, one simple indicator is the Urban Heat Island
(UHI) intensity, usually defined as the temperature difference between
the urban area and surrounding rural area. Distinct climates and levels
of urbanization tend to present different UHI intensities [4].
In order to mitigate the UHI effect, Rosenfeld, et al. [5] explored
solutions at building and city scales, including cool (reflective) surfaces,
32
Applied Energy 215 (2018) 31–40
tree shading, and district cooling. The study highlighted the potential
for energy savings, especially with the use of cool surfaces that can be
implemented at low cost. Furthermore, specific attention was drawn to
policies and standards to motivate the implementation of new tech-
nologies and reduce CO2 emissions. Rizwan et al. [4] compiled different
works of mitigation strategies and reported the resulting temperature
reduction and savings. The studies mainly include increase of vegeta-
tion, reduction of anthropogenic heat, increase of shading, cool roofs,
and humidification. Rossi et al. [6] investigated the use of reflective
surfaces and concluded that it could reduce radiative effects, especially
in areas with buildings in close proximity. Santamouris [7] studied
reflective and green roofs as mitigation levers for heat islands, con-
sidering the limitation of this solution for high-rise buildings, but
highlighting the potential for latent heat losses. The author concluded
that the reflective roofs gain advantage in sunny climates, while ve-
getation presents higher benefits in colder climates.
2.2. Energy modelling
2.2.1. Building energy model (BEM)
A typical building energy model (BEM) is a white-box engineering
model used for estimating real or proposed building thermal behavior.
A series of detailed calculations is conducted by the model, based on a
large amount of information such as building characteristics, climate
conditions, and activity schedules [8]. The high input uncertainty, quite
common given the level of detail required, often leads to lower accu-
racy of the simulated output. Furthermore, developing a reliable model
is a long-term process that requires a high level of expertise. That is why
simplified models are often used and it is up to an expert modeler to
decide which assumptions and simplifications should be made [9].
Sensitivity analysis is thus recommended to understand how certain
variables affect the building performance [10]. Applications of BEM
include:
• Testing of new technologies, estimating savings and potential for
retrofits, such as green roofs [11]or thermochromic glazing [12],
• Parametric analysis to test feasibility, cost-benefit, and design op-
tion for a real project [13],
• Verifying compliance for certifications, such as the LEED, BREEAM
(UK), Green Star (Australia), Estidama (Abu Dhabi, UAE), etc. [14],
• Estimating more suitable solutions for different regions and climates
[15],
• Research or consultancy work for design, retrofits, operation, and
commissioning [14],
• Generation of life cycle cost (LCC) and life cycle analysis (LCA), as
well as measurement and verification (M&V) [13].
2.2.2. Urban energy model (UEM)
Energy system modelling has been well explored at the building
level. Over the past four decades, we have achieved significant progress
in modelling the heat and mass flows entering or leaving buildings to
the point where, given reasonably comprehensive input data, most si-
mulation engines can predict future energy use of standard construc-
tions with acceptable accuracy. However, the study of the UHI requires
the use of neighborhood-to-city scale urban energy models (UEMs).
The BEMs usually run simulations under isolated conditions (no
shading effect from adjacent buildings) and consider that rural weather
conditions, uniformly distributed around the envelope, prevail. The
latest UEMs have been created as adaptations to traditional BEM soft-
ware and are able to model, to varying degrees, the interactions be-
tween the building, occupant, and environment [16].
Pisello et al. [17] considered inter-building effect to study how they
influence the accuracy of building energy performance evaluations.
They concluded that, when evaluating the energy performance of a
single building, the special relationship with surrounding buildings
should be taken into account. Martin et al. [18] coupled an urban
L.G.R. Santos et al.
canopy model with EnergyPlus using the Building Controls Visual Test
Bed (BCVTB) to estimate temperature and humidity for urban canyon in
Masdar city, Abu Dhabi. Reinhart et al. [19] reviewed emerging UEM
methods and gave an overlook for future development. Frayssinet et al.
[20] studied the impacts of adaptations of building envelopes for dis-
trict energy simulations. It was observed that the mean differences
between modelled and measured (under 4%) are significantly lower
than maximum differences, indicated by the peak loads for heating and
cooling.
Another example is the Urban Weather Generator (UWG) by Bueno
et al. [1]. The UWG calculates the urban air temperatures via four
coupled models based on the rural meteorological data and mesoscale
information to account for various properties in the target neighbor-
hood (e.g. building, street, vegetation, etc.). First, the Rural Station
Model calculates the sensible heat fluxes at the (rural) weather station
at the soil surface. Next, the Vertical Diffusion Model calculates profiles
of air temperature above the rural site. Then, the Urban Boundary Layer
Model calculates the air temperature above the urban canopy player.
Finally, the Urban Canopy and Building Energy Model finishes the
process by calculating the urban sensible heat fluxes and estimating
urban canyon air temperature and humidity. Therefore, the UWG
output is the urban microclimate condition as a result of the rural
weather condition and various properties in the target neighborhood
(e.g. building, street, vegetation, etc.). With continuous development
and update, it has been validated in Toulouse (France), Basel (Swit-
zerland) [1], Singapore [21], and Abu Dhabi (UAE) [22]. The UWG was
generally able to represent the urban microclimate condition with
reasonable accuracy for all those cases. In addition, Mao et al. [22]
performed a global sensitivity analysis of urban microclimate using the
UWG tool with 30 candidate inputs, including meteorological factors,
urban characteristics, vegetation variables, and building systems. For
the case of Abu Dhabi, the most critical parameters are also the para-
meters with high uncertainty, such as the reference height of the ver-
tical diffusion model, the urban canopy model boundary layer exchange
coefficient, the fraction of waste heat into urban canyon, and the night-
time urban boundary layer height.
2.3. Model calibration
Calibration methods are used to reduce discrepancies between si-
mulated and real buildings, usually based on measured data. These
differences in the models are due mainly to different sources of un-
certainty, either resulting from lack of information or modelling lim-
itations a priori. The least uncertain parameters should be used in early
simulation stages. The estimated baseline model after calibration is
considered more suitable for building design, operation, diagnosis, and
commissioning [23]. Amongst the benefits of calibration, Raftery et al.
[24] highlighted the potential to estimate savings from retrofit options
and to acquire information to enhance future design and develop best-
practice recommendations. Monetti et al. [23] suggested a workflow for
calibration that is summarized as follows:
(1) Carry out a building energy audit and define a model based on
design data (uncalibrated);
(2) Use metered data to refine the modelling inputs. If possible, un-
certainty and sensitivity analysis should be conducted to determine
inputs that significantly influence the building energy performance;
(3) Conduct an optimization process, altering uncertain parameters
based on certain constrains. This process aims to minimize the
difference between measured and simulated data. The optimization
finishes when the global or near-global minimum is found, and it
should give results that match closely the monitored data;
(4) Validate the optimized parameter set. Mean bias error (MBE) or its
normalized value (NMBE) and root-mean-square error (RMSE) or its
coefficient of variation (CvRMSE) are calculated in accordance with
33
Applied Energy 215 (2018) 31–40
such standards as ASHRAE Guideline 14 [25] to confirm its com-
pliance. For calibration with monthly data, an absolute MBE
smaller than 5% and a CvRMSE smaller than 15% are deemed ac-
ceptable.
Calibration can be conducted via a manual or automated method.
Manual calibration, although more common and requiring less pre-
paration, might need more labor time to test many possibilities and is
prone to errors. In this process, an expert manually changes a set of
parameters, trying to achieve satisfactory results. Automated methods,
on the other hand, usually make use of an algorithm to change the
parameters within boundaries and evaluate the simulation result, di-
recting the test of parameters towards an optimal result [26]. In par-
ticular, a genetic algorithm (GA) is often used for simulation-based
optimization and automated model calibration. Although GA does not
indicate a single optimal solution, it is able to solve convex problems
and take into account multiple objectives, allowing the user to choose
the best fit depending on the situation according to a Pareto curve
distribution. Any automated calibration is bound to be slow because it
executes the relatively complex BEMs many times before it converges to
an optimal/sub-optimal solution.
Extensive research has been conducted in this field and compiled in
literature in the past years. Nguyen et al. [27] reviewed optimization
methods, observing that the GAs are by far the most used method given
their ability to handle continuous and discontinuous variables, to solve
multi-objective problems, to avoid local minima, and to allow multi-
processor simulation. The most popular optimization engines are Ma-
tlab and GenOpt, while EnergyPlus and TRNSYS are the most common
BEM tools used in the calibration process. Royapoor and Roskilly [28]
coupled an EnergyPlus model with two environmental data sensors,
highlighting the importance of including local weather files in the ca-
libration process. Ruiz et al. [29] used GA for building envelope cali-
bration, considering a parametric and sensitivity analysis for the ele-
ments of the façade and using jEPlus + EA for the optimization, proving
its feasibility and showing the advantages of a precise automated pro-
cess. Chaudhary et al. [30] compared a whole-building calibration by
manual and automated approaches. Using automated optimization, the
calibration results were similar to the results from a manual process but
around 6 to 7 times faster. Sun et al. [31] used a pattern-based ap-
proach for an automated Building Energy Model calibration, comparing
two patterns: (1) simulated monthly electricity consumption compared
to monthly electricity bill and (2) simulated monthly gas consumption
compared to monthly natural gas bill. The process identifies and
eliminates universal and seasonal bias, respectively, using the Com-
mercial Building Energy Saver (CBES) toolkit. The tuning steps for the
calibration included changes in lighting power intensity, occupancy
density, average outdoor airflow per person and cooling COP with
monthly values measurements and simulations. It was concluded that,
although some limitations of calibration are maintained and the tool
does not offer a wider range of building types and model parameters, it
is a feasible process capable of identifying and adjusting bias. Lara et al.
[32] conducted a process of calibration in a large school building in
Treviso (Italy) using jEPlus and jEPlus + EA. The tool jEPlus couples
with EnergyPlus for parametric analysis and allows an implementation
of GA to conduct a calibration process. Comparing the results of the
parametric analysis with the genetic optimization, both reached the
best monthly NMBE and CvRMSE. The GA presents an advantage of
shorter simulation time (90 h compared to 4746 h from the parametric
analysis) and fewer simulations (1368 compared to 72,030 from the
parametric analysis). Hong et al. [33] coupled the EnergyPlus model
with a Python optimization script aiming to minimize the CvRMSE.
After establishing the building energy simulation model, a CvRMSE of
18.1% was calculated and the design variables were set. Then, a GA was
initiated to adjust the uncertain parameters, reducing the CvRMSE to
12.62%.
L.G.R. Santos et al.
Fig. 1. Workflow of information/data in the proposed calibration procedure.
3. Methodology
In this study, we propose two main approaches for energy model
calibration at the district level, which will be compared with a baseline
case in order to see how the building-related parameters are tuned
when simulations incorporate the impact of the urban microclimate.
The baseline case is a standard district simulation, where only the
building parameters are calibrated and the reference rural weather file
is assumed to prevail (the UHI effect is neglected). The first proposed
approach (Method 1 or M1) intends to simultaneously tune the un-
certain parameters related to the building energy and the coefficients
describing the UHI intensity. The second proposed approach (Method 2
or M2) is similar to the first one, except that the weather file is pre-
processed by the UWG. The latter incorporates information related to
urban morphology and building/road properties in order to create an
urban EPW file at the target street level. Consequently, only the
building-related parameters are calibrated in Method 2, given that the
weather variables are assumed to have already been adapted to the
specific urban microclimate.
As shown in Fig. 1, the calibration process starts by analyzing the
data and information obtained from an energy audit. At this stage, the
major building design parameters are extracted, together with the
measured monthly electricity consumption data from utility meters.
The unknown input parameters for the subsequent calibration are se-
lected at the end of this stage. Finally, the unknown parameters are
tuned in order to identify the plausible set of values that can minimize
the model/measurement discrepancy. In the present study, this cali-
bration process is automatically conducted using heuristic-based opti-
mization (i.e. GA).
3.1. Data analysis
The selected case study is District E3-02 in downtown Abu Dhabi
(UAE), where an energy audit was conducted in 2011 for the previous
three years (i.e. 2008–2010). The district is composed of 67 mixed-use
buildings. This study will consider only 56 buildings that have reliable
meter data. The remaining buildings were removed in our analysis
based on the enforcement of two statistical criteria. The first criterion
assumes a normal distribution for the annual energy use intensity (EUI)
of each building and eliminates buildings that have an EUI deviating
from the mean by more than two standard deviations (2σ or 95%
confidence interval). The second elimination criterion looks at the ratio
of peak monthly energy use (corresponding to the month of August)
over annual energy use for each building and retains, once again, the
buildings that are within ± 2σof the mean. As a result, the first filter
removes seven buildings while the second filter removes an additional
four buildings, leaving 56 buildings for the calibration analysis. The
histogram for each criterion (before elimination of outliers) is depicted
in Fig. 2, together with its corresponding mean (μ) and standard de-
viation (σ).
The possible reasons for a building to be out of range might include:
information missing from some of the building’s utility meters
34
Applied Energy 215 (2018) 31–40
(including the one dedicated to the chiller plant), disruption in the
periodic measurements of the utility meter data, large portions of the
building not being used, and large positive bias that indicates an unu-
sual consumption that is much higher than the average. All the build-
ings (including the ones removed from the calibration) will still be
considered as adjacent shading obstacles for neighboring buildings.
The District E3-02 is representative of downtown Abu Dhabi, with
high-rise buildings on the periphery (along the main roads) and low/
mid-rise buildings in the inner part. Mixed-use buildings are quite
common in downtown Abu Dhabi. Since it is impractical to fully
characterize all building properties at the urban scale in an audit, the
buildings were assumed to be subdivided into three main use types:
residential, commercial, and retail. The gross floor area (GFA) for built-
up was 552,190 m2, with 97% of the area (537,740 m2) air-conditioned.
The energy audit provided the following inputs for each building:
• Building height;
• Window-to-wall ratio (glazing percentage);
• U-values – including values for the wall, roof, and window;
• Typology distribution – percentage of the buildings that corresponds
to retail, office, and residential usage;
• Electricity consumption (meter readings).
3.2. Baseline model
The baseline BEMs are developed in EnergyPlus, a well-known and
reliable tool for building-level dynamic simulation [34] that can
quickly simulate the performance for multiple buildings during a full
year by automating the process with the use of a single script. We use
the Rhino 3-D modeler (with the Grasshopper graphical algorithm
editor and Honeybee, and Ladybug plug-ins) to describe the building-
related inputs. The 56 buildings were developed separately, but in each
case, adjacent buildings were included as shading obstacles (Fig. 3).
The HVAC system was considered to be ideal with constant COP
(Coefficient of Performance) of 2.5. Since we are considering several
buildings as case study and chiller performance was not assessed at part
of the previously conducted energy audit, we opted for a constant COP
value to represent the seasonal efficiency. The selected seasonal COP
value is inferred from several pilot projects in the downtown area.
Several parameters, not included in the energy audit, will be esti-
mated during the calibration procedure. They are assumed to be the
same for all buildings or, in some cases, differ only on the basis of use
type. These calibration parameters are:
• Albedo: Average solar reflectance of the walls and roofs (one para-
meter);
• Internal gains: One value per use type: residential, retail, and office
(three parameters). The schedule was considered to be “always on”,
therefore the gains are estimated as average annual values;
• Infiltration: Infiltration rate in air changes per hour (ACH) is esti-
mated with two different values: one for the residential use type
where operable windows are common and the other for the retail/
office use type with predominantly non-operable windows (two
parameters);
• Window SHGC: The average solar heat gain coefficient of glazing
(one parameter);
• UHI intensity parameters (for Method 1 only): Daily and seasonal
profiles of urban-rural temperature and humidity differentials are
parameterized using periodic functions (four parameters). This
parameterization procedure will be explained in detail in the fol-
lowing subsections.
3.3. Model calibration
A Matlab script, dynamically coupled with EnergyPlus, is used for
all calibration approaches. Calibration processes include a standard
L.G.R. Santos et al. Applied Energy 215 (2018) 31–40

14 9

8
12
7
10
6

Number of buildings
Number of buildings

8 5

4
6

3
4
2

2
1

0 0
50 100 150 200 250 300 350 400 450 500 0.09 0.095 0.1 0.105 0.11 0.115 0.12 0.125 0.13 0.135
Building's annual energy intensity [kWh/m²] peak monthly energy use / building annual consumption ratio

Fig. 2. Building’s annual energy intensity with mean (μ) = 278.58 and standard-deviation (σ) = 72.25 (left), and peak monthly energy ratio with mean (μ) = 0.115 and standard-
deviation (σ) = 0.008364 (right).

one of the simulated buildings. Significantly, each IDF file incorporates

Fig. 3. Treatment of shading obstacles (red) for an individual building (yellow with
glazing area in blue). (For interpretation of the references to color in this figure legend,
the reader is referred to the web version of this article.)
district calibration with rural weather (baseline) and two methods to
account for the urban microclimate and specifically the UHI intensity.
Method 1 calibrates both a detailed building energy model and an ap-
proximate parameterization of the UHI intensity, while Method 2 ca-
librates only the building energy model and derives a temperature-only
assessment of the UHI intensity from the UWG module. The UWG cal-
culation is performed offline and results in the adjustment of the re-
ference weather file used in subsequent online simulations.
3.3.1. Standard energy model calibration (baseline)
3.3.1.1. Weather. The reference weather file contains hourly data
measured during calendar year 2010 at a rural weather station
located in Masdar City. Available data include air temperature,
relative humidity, wind speed, wind direction, and solar radiation
(Global Horizontal Irradiance, Direct Normal Irradiance and Diffuse
Horizontal Irradiance). In the baseline case, as highlighted before, the
reference weather file is used, unaltered, during the calibration process
and only the building-related parameters are adjusted (seven
parameters).
the shading effect of adjacent structures. Based on prior information,
parameter lower/upper bounds are defined and enforced during the
calibration process.
Calibration was performed using a genetic algorithm (GA), known
for being robust to local minima and allowing straightforward in-
corporation of multiple objectives. GA has been widely used in building
energy model calibration, as described in Section 2.3. The algorithm
iteratively modifies the parameters leading to an optimal or near-op-
timal solution. A population size of 25 was applied for the baseline and
Method 2 and increased to 30 for Method 1, since the latter has four
more parameters. The maximum number of generations is set to 150.
The default Matlab settings were kept for the crossover (scattered) and
mutation operator (Gaussian).
Our calibration is based on a bi-objective optimization, as the al-
gorithm aims to achieve an optimal solution minimizing two model
accuracy criteria: CvRMSE and NMBE. The latter are determined as
weighted averages. We defined two approaches to quantify model ac-
curacy: the weighted-average building approach (Eqs. (1) and (2)) and
the grouped building approach (Eqs. (3) and (4)). For the weighted-
average building, the GA finds the set of parameters that will provide
the best model fit for each building. For the grouped building, we
consider the aggregate building energy consumption and try to find the
optimal model corresponding to a fictitious building that would re-
present the combination of all the selected buildings in the district.
Given the incremental challenge of minimizing the CvRMSE and NMBE
for each building, the error for the weighted average building approach is
preferred as the target of the bi-objective calibration. The error corre-
sponding to the grouped building approach will still be reported for the
optimal solution in order to illustrate discussion in subsequent sub-
sections.
For the weighted average building approach:
12
N ∑ j = 1 (deij )2 1
yCvRMSE = 100∗ ∑ ∗ [%]
i=1
11 E (1)

3.3.1.2. Automated multi-objective coupled calibration. For each iteration 12

of the calibration, a Matlab script runs an EnergyPlus simulation for N

∑ deij
each building using a selected set of parameters. This loop automates j=1 1
yNMBE = 100∗ ∑ ∗ [%]
the calibration, enabling the optimization algorithm to change the i=1
11 E
parameters at each iteration. The discrepancy between simulated and (2)
measured electricity consumption is the basis for calibration. All of the
EnergyPlus IDF files are stored together, with each file corresponding to For the grouped building approach:

35
L.G.R. Santos et al. Applied Energy 215 (2018) 31–40

Baseline Model Method 1 (Weather calibration) Method 2 (UWG)

16.83 16.35 16.41

16.82 16.4
16.3
16.39
16.81 16.25
16.38
16.8 16.2 16.37
16.79 16.15 16.36
19.74 19.76 19.78 19.8 19.25 19.35 19.45 19.55 19.65 19.05 19.1 19.15 19.2
Fig. 4. Pareto front for the three calibrations observed with solution chosen (red) x-axis: CvRMSE and y-axis: NMBE. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)

12 N Table 1
∑ j = 1 ( ∑i = 1 deij )2 1
YCvRMSE = 100∗ ∗ [%] Summary of the calibrated parameters.
11 E (3)
Parameter Calibrated parameters
12 N
∑ j = 1 (∑i = 1 deij )
YNMBE = 100∗ [%] Baseline Method 1 (M1) Method 2
11 (4) Weather (M2)
calibration UWG
where deij is the monthly electricity use difference for building i, j is the
month of the year (1 to 12), E is the mean monthly energy use over all Wall and roof albedo 0.46 0.69 0.65
months and all buildings and N is the total number of the buildings. Internal gain (retail) [W/m2] 10.65 5.15 6.54
Internal gain (office) [W/m2] 10.23 8.77 10.05
Internal gain (residential) [W/m2] 12.46 12.67 12.92
3.3.2. Detailed energy and weather model calibration (Method 1) Infiltration (retail and office) [ACH] 0.49 0.50 0.42
Since one of the main goals of this paper is to estimate the UHI Infiltration (residential) [ACH] 0.66 0.65 0.67
intensity, the weather file used in the baseline case (2010) is now ad- Window SHGC 0.51 0.46 0.61
justed at each iteration as a part of the calibration process. Based on the Average annual urban/rural – 0.54 1.08a
temperature bias [°C]
average diurnal profile of the UHI intensity for Abu Dhabi [2] as well as
Urban/rural seasonal temperature – 1.09 –
the actual measurement at the urban/rural sites in Abu Dhabi, it was variation [°C]
determined that the diurnal profile of the urban/rural temperature and Average annual humidity bias [g/g] – 0.16 –
humidity differential can be conveniently fitted into periodic (diurnal) Urban/rural seasonal humidity – −0.08 –
functions. variation [g/g]

The temperature differential profile can be expressed as: a

Temperature daily variation for Method 2 was not calibrated, but calculated from the
2πt π 4πt π ⎤ UWG (output from pre-processing stage).
UHITemp = A + 2.5 ⎡cos ⎛ − ⎞−0.3cos ⎛ − ⎞
⎣ ⎝ 24 4 ⎠ ⎝ 24 15 ⎠ ⎦
2πm π observation, sporadic measurement or anecdotic evidence: no vegeta-
+ B ⎡sin ⎛ + ⎞⎤
⎣ ⎝ 12 2 ⎠⎦ (5) tion, initial temperature of 20 °C, glazing ratio of 50%, window U-value
of 3.88 W/m2, window SHGC of 0.75, envelope infiltration of 0.75
where A is the overall annual average of UHI intensity (to be cali-
ACH, indoor air set point of 22 °C, HVAC waste fraction1 of 50%, and
brated), B is the amplitude of seasonal variation (to be calibrated), t is
constant cooling COP of 2.5. The other parameters, especially the ones
the time of day in hours (1 to 24), and m is the corresponding month of
related to the urban boundary layer and the rural site, were set in ac-
the year (1 to 12).The specific humidity differential profile can be ex-
cordance with one of our previous studies (cf. [22]).
pressed as:
The process of calibration is exactly the same as the one applied in
2
1 (t − 12)
2πm π the baseline scenario, except that the reference weather is pre-pro-
× ⎛⎜C . e + ⎞ ⎞⎟
−0.5.
UHIHum = 2.52 + D . sin ⎛
1000 ⎝ ⎝ 12 2 ⎠⎠ cessed off-line and adapted to account for the UHI in District E3-02.
(6)
where C is the maximum daily UHI intensity at noon (to be calibrated), 4. Results and discussion
D is the amplitude of seasonal variation (to be calibrated), t is the time
of day in hours (1 to 24), and m is the corresponding month of the year As each calibration performed by the GA gives a set of optimal or
(1 to 12). These profiles were derived based on the observation of the near-optimal results represented on a Pareto front, we selected as op-
actual UHI profiles in Abu Dhabi. timal the knee point, which is closest to the ‘utopia point’. The Pareto
front for the three calibrations can be observed in Fig. 4.
3.3.3. Detailed energy model calibration with UWG (Method 2)
The UWG converts a reference rural weather file to the estimated
4.1. Parameter
urban conditions at a neighborhood scale based on specific urban area
characteristics. Starting with the rural weather data in an EPW format
The selected parameters were calibrated using the bi-objective GA
and the urban characteristics, the UWG outputs the simulated urban
to find the optimal or near-optimal solution that could minimize the
weather data (over an entire calendar year or a subset thereof). In this
combination of CvRMSE and NMBE for weighted average buildings. A
study, these calculations are executed using a Matlab script (.m). The
summary of the calibrated parameters is presented in Table 1.
main output is an adjusted (urban) weather file in EPW format [35].
For the envelope’s albedo, a value of 0.65 (M1) or 0.69 (M2) was
UWG constructs a simplified mono-zone energy model of the
average building and relates it to the rural weather via an urban
boundary layer model. Some of the parameters required for the speci- 1
The “HVAC waste heat fraction” is the fraction of the chiller waste heat that is re-
fication of UWG’s urban canopy model were estimated based on direct jected directly into the urban canyon.

36
L.G.R. Santos et al. Applied Energy 215 (2018) 31–40

0.8
Method 1 0.8
Method 2

0.7 0.7

0.6 0.6

0.5 0.5

0.4 Albedo 0.4 Window SHGC

Window SHGC Albedo
0.3 0.3

1000
1111
1222
1333
1444
1555
1666
1777
1888
1999
2110
2221
2332
2443
2554
2665
1
112
223
334
445
556
667
778
889
1171
1366
1561
1756
1951
2146
2341
2536
2731
2926
3121
3316
3511
3706
3901
4096
4291
4486
4681
196
391
586
781
976
1

Fig. 5. Convergence of envelope’s albedo and window SHGC (y-axis) by number of iterations (x-axis).

Method 1 Method 2
30 30
I.G. (Retail) I.G. (Retail)
25 25
I.G. (Office) I.G. (Office)
20 I.G. (Residential) 20 I.G. (Residential)

15 15

10 10

5 5

0 0
1087
1268
1449
1630
1811
1992
2173
2354
2535
2716
2897
3078
3259
3440
3621
3802
3983
4164
4345
4526
1
182
363
544
725
906

1
112
223
334
445
556
667
778
889
1000
1111
1222
1333
1444
1555
1666
1777
1888
1999
2110
2221
2332
2443
2554
2665
Fig. 6. Convergence of internal gains [W/m2] parameters (y-axis) by number of iterations (x-axis).

Method 1 Method 2
0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3
Inf. (Retail and Office) Inf. (Residential) Inf. (Retail and Office) Inf. (Residential)
0.2 0.2
1
189
377
565
753
941
1129
1317
1505
1693
1881
2069
2257
2445
2633
2821
3009
3197
3385
3573
3761
3949
4137
4325
4513

1000
1111
1222
1333
1444
1555
1666
1777
1888
1999
2110
2221
2332
2443
2554
2665
112
223
334
445
556
667
778
889
1

Fig. 7. Convergence of infiltration [ACH] parameters (y-axis) by number of iterations (x-axis).

obtained, which corresponds to a light grey/brown. These are the
commonly observed façade features in District E3-02. Meanwhile, the
baseline calibration provided a darker color (0.46).
The internal gains were found to be 5.15 or 6.54 W/m2 for retail
buildings, 8.77 or 10.05 for office buildings, and 12.67 or 12.92 for
residential buildings (M1 or M2, respectively). The higher internal
gains for residential buildings can possibly be a compensation for the
fact that indoor spaces corresponding to this use type usually have less
efficient cooling systems with lower COPs. Furthermore, since the in-
ternal gains are calculated as an average for an “always on” schedule, it
is expected that the retail and office buildings would present lower
gains, given that they operate in reduced mode when the businesses are
closed, whereas the residential buildings, in our current setting, tend to
be occupied continuously. The baseline provided overall higher internal
gains when compared to the other two methods.
For the infiltration, calibration identified the values of 0.50 (M1)
and 0.42 (M2) ACH for retail and office buildings as well as 0.65 (M1)
to 0.67 (M2) ACH for residential buildings. In residential spaces, the
area of operable windows and the time during which they stay open are
usually higher, so the increase in infiltration aims to compensate for the
increase in natural ventilation. For the baseline case, results are very
similar, with 0.49 ACH for retail and office spaces and 0.66 for re-
sidential.
The window SHGC indicates the percent of solar thermal energy
incident on the glass that is transferred indoors. For the case study, the
optimization obtained a value of 0.46 (M1) to 0.61 (M2), which is an
average value between common clear glass and a mid-performance one.
These are glazing types that could be commonly observed in the
buildings of this district. The baseline value is between the two cali-
brated values, with a SHGC of 0.51.
It can be observed that the albedo and internal gains were the
parameters that differed the most from the baseline and, Furthermore,
the two proposed methods are closer to each other than to the baseline.
This shows the relevance to include some type of urban microclimate
model (either on-line as in Method 1 or off-line as in Method 2) in the
calibration to estimate more accurate building properties—not to
mention the fact that the UHI intensity estimate is valuable in itself.
Figs. 5–7 depict the convergence, after ordering, of building-related
parameters for Method 1 (4680 iterations) and Method 2 (2650 itera-
tions). Faster convergence is achieved in Method 2 compared to Method

37
 L.G.R. Santos et al. Applied Energy 215 (2018) 31–40

4 Fig. 8. Daily variation of the temperature (°C) for Method

1 (red) and Method 2 (blue). (For interpretation of the
references to color in this figure legend, the reader is re-
3 ferred to the web version of this article.)
Urban Heat island intensity (°C )

Calibrated weather
UWG
2

-1

-2

-3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hour of the day (h)

1. weather), which resulted in CvRMSE of 19.76% for NMBE of 16.8%.

For Method 1, the annual average daily UHI intensity is 0.54 °C
while the seasonal average is 1.09 °C. These values are inserted into the
Eq. (5) as parameters A and B respectively, and the resulting average
annual daily profile is then compared to the average annual daily
profile of UHI calculated by UWG in Method 2 (Fig. 8). The observed
outcome of the proposed calibration methods provides a validation of
the UWG model, given that both daily profiles follow a similar patterns
and intensities, albeit with a notable mid-day difference. In Method 1,
the monthly seasonal variation parameters indicate a peak of tem-
perature-related UHI intensity in winter. On the other hand, according
to Method 1, the urban/rural humidity differential is negligible. For
Method 2, the weather is pre-processed by the UWG to estimate urban
microclimate condition in the case study. This approach aims to reduce
the number of parameters required for the calibration in order to
achieve a faster calibration process. It is important to note that the
current version of UWG only adjusts dry-bulb temperature. The average
annual UHI intensity estimated for the year with UWG is 1.08. How-
ever, UWG suggests a summer-time peak of the UHI intensity at odds
with the results of Method 1.
4.2. Error
Together with the calibrated parameters, the discrepancies between
the predicted and measured building electricity consumption were
calculated. For a district-wide calibration, we calculated the CvRMSE
and NMBE for both a weighted-average building approach and a
grouped building approach (Table 2).
Using the weighted-average building approach, the CvRMSE values
were found to be 19.09% (M1) and 19.40% (M2), while the NMBE
values were found to be 16.24% (M1) and 16.39% (M2). Both methods
that consider the UHI reveal lower errors than the baseline case (rural
Table 2
Summary of errors for each calibration approach.
Scenario Error Calibration approach
Baseline Method 1 (M1)
Weather
calibration
Although these numbers are not compliant with the single building
calibration thresholds set by the ASHRAE Guideline 14 (15% and 5%,
respectively), they seem satisfactory for the calibration at a district
level, especially given the very demanding nature of the weighted-
average building approach and the large diversity of constructions and
use types in the case study at hand. The weighted-average buildings
approach estimates average properties of buildings and UHI intensity
while ensuring that each building taken separated is accurately cali-
brated to the greatest possible extent. In other words, this approach can
be considered a combination of the grouped building approach (dis-
cussed below) and individual BEM calibration.
In the grouped building approach, monthly building energy con-
sumption is added up prior to calibration and the parameters found are
applied to a fictitious aggregate building that represents the sum of all
real buildings. Therefore, it is closer to a standard single-building ca-
libration. The CvRMSE was 4.04% (M1) and 2.71% (M2), while the
NMBE was 2.35% (M1) and 1.95% (M2). This approach is therefore
capable of generating smaller errors, since in this case the calibration is
performed for a fictitious aggregate building without considering the
specificities of individual buildings within the district. On the other
hand, the results are probably less reliable than the previous approach
in terms of the individual building performance. The baseline presents a
higher error for the grouped building’s CvRMSE (4.30%), but a lower
one for the NMBE (1.52%).
To see the calibration performance at the building level, the cali-
brated parameters were used as inputs for each single BEM, calculating
the CvRMSE and NMBE for each one of the 56 buildings. Table 3 shows
the number of buildings (out of 56) that could meet each criteria
CvRMSE < 15% and NMBE < 5%, respectively. The buildings that
meet both criteria are considered to be fully calibrated.
We can see that the best result is obtained from Method 2 (the UWG
Table 3
Number of buildings compliant with ASHRAE Guideline 14.
Criteria Number of buildings compliant
Method 2 Baseline Method 1 (M1) Method 2
(M2) Weather (M2)
UWG calibration UWG

Weighted-average CvRMSE 19.76% 19.40% 19.09% CvRMSE (< 15%) 30 28 29

building approach NMBE 16.80% 16.24% 16.39% NMBE (< 5%) 12 15 16
Grouped building CvRMSE 4.30% 4.04% 2.71% Fully calibrated (CvRMSE < 15% 12 13 15
approach NMBE 1.52% 2.35% 1.95% and NMBE < 5%)

38
L.G.R. Santos et al.
simulations), where 29 buildings (52% of all buildings) are calibrated
for CvRMSE, 16 buildings (29% of total) are calibrated for NMBE, and
15 buildings (27% of total) are compliant with both CvRMSE and NMBE
criteria (i.e. fully calibrated at the building level according to current
standards).
These results reveal the challenge to have compliance at the
building level by performing a district-wide model calibration, espe-
cially considering the diversity of buildings that could be observed in a
single urban district. In this case study, the diversity of number of
floors, typologies, schedules and equipment technology requires drastic
simplifying assumptions in model settings. In order to increase the
number of individually calibrated buildings, a subdivision of the dis-
trict/city in terms of building archetypes (i.e. typologies, buildings age,
buildings height etc.) would be recommended, given that similar
buildings tend to give more reliable individual parameters for one ar-
chetype.
5. Conclusion
This study evaluates two main approaches for an automated district-
wide calibration via genetic algorithm, coupling building energy
models with the Urban Heat Island (UHI) effect to generate a more
accurate energy model. The development of an urban energy model
calibration aims to reduce the computational time for the case with a
large number of neighboring buildings. Furthermore, it seems not
possible to incorporate the UHI effect in model calibration for single
buildings, given that each calibration would provide a different value
for temperature difference. Thus, the averaging effect of calibrating
numerous building provides a robust estimate of the UHI intensity.
The proposed calibration methods aim to better estimate parameters
that are uncertain but have influence on building energy performance.
Preferred ranges for each parameter are defined to guide the algorithm
to reach an optimal or near-optimal solution with reasonable parameter
setting. Regarding the methods proposed, the UWG revealed a good
capability of estimating the urban microclimate, increasing the accu-
racy of urban energy modelling with a faster calibration process com-
pared to the case where the UHI is included as a to-be-tuned parameter
(Method 1). Incorporating the UWG into the calibration leads to the
best performance for the calibration errors (CvRMSE and NMBE) and
the highest number of individually calibrated buildings. As previously
noted, some of the parameters (mainly the albedo and the internal gains
in this study) are likely to compensate for the case where UHI is ne-
glected. Finally, the outcome of the two methods effectively provides a
validation of the UWG model.
This study could be applied for urban energy studies where there is
interest in discovering, via calibration, average properties of the built
structures or use types within the sector without focusing on individual
buildings. In addition, Method 1 enables us to estimate the UHI in-
tensity concurrently. In the governmental field, district-wide model
calibration reveals important information for the area of interest and
allows a better estimation of the peak load, which is important to de-
termine the capacity of the energy supply. For energy management
consultants, the calibration generates a more realistic district/city-scale
model, which could potentially allow for more accurate assessment of
district- or urban-scale energy retrofits.
For residential villas, for example, buildings usually have similar
properties and activities and both methods presented herein can be
applied quite effectively. In more mixed-use districts or where the
buildings have high variability in their construction, as is the case here,
it can be challenging to reach sufficiently low calibration errors.
Therefore, a subdivision of the domain is recommended, depending on
the desired level of detail.
The proposed work can assist in the creation of new standards for
calibration, as currently standards exist only for a single BEM where the
urban microclimate is not considered. It is recommended to have
standards that cover different approaches such as weighted-average
39
Applied Energy 215 (2018) 31–40
buildings and grouped buildings. The percentage of individually cali-
brated buildings also can assist in revealing the accuracy of the cali-
brated model.
5.1. Limitation
Limitations include data inaccuracy for some of the buildings,
especially with regard to utility meter data. Pre-processing data from
the energy audit (to eliminate outliers) is necessary but does not
guarantee that all the data used is completely reliable, which might
cause slight variations in the expected result of a calibration process.
Given that the energy measurements provided by the audit were only
monthly and not hourly, it was not possible to exactly determine peak
load. The weather data may also present some limitations, depending
on the years observed and the functions and software used to adapt it
from rural to urban.
Regarding the model limitations, the assumptions that unknown
building parameters are the same for all buildings or for given use
types, may influence the outcome. Furthermore, the UWG tool (M2)
and the simplified UHI parameterization (M1) used in this study present
their own shortcomings.
Although, at each iteration, the individual BMES were executed in
parallel to save time (using a powerful multi-processor desktop), the GA
procedure has a very high computational cost. This fact hindered our
ability to run a large number of GA optimizations to ensure achieve-
ment of a global minimum.
Acknowledgment
The authors hereby acknowledge the support of the Abu Dhabi
Municipality, and Mr. Khaled Aljunadi specifically, for facilitating ac-
cess to the results of the energy audit of sector E3-02.
References
[1] Bueno Bruno, Norford Leslie, Hidalgo Julia, Pigeon Gregorie. The urban weather
generator. J Build Perform Simul 2013:269–81.
[2] Afshari A, Liu N. Inverse modelling of the urban energy system using hourly elec-
tricity demand and weather measurements. Part 2: Gray-box model. Energy Build
2017.
[3] Mirzaei Parham A, Haghighat Fariborz. Approaches to study Urban Heat Island –
abilities and limitations. Build Environ 2010:2192–201.
[4] Rizwan Ahmed Memon, Leung Dennis YC, Liu Chunho. A review on the generation,
determination and mitigation of Urban Heat Island. J Environ Sci 2008:120–8.
[5] Rosenfeld Arthur H, Akbari Hashem, Bretz Sarah, Fishman Beth L, Kurn Dan M,
Sailor David, et al. Mitigation of urban heat islands: materials, utility programs,
updates. Energy Build 1995:255–65.
[6] Rossi Federico, Pisello Ana Laura, Nicolini Andrea, Filipponi Mirko, Palombo
Massimo. Analysis of retro-reflective surfaces for urban heat island mitigation: a
new analytical model. Appl Energy 2014:621–31.
[7] Santamouris M. Cooling the cities – a review of reflective and green roof mitigation
technologies to fight heat island and improve comfort in urban environments. Sol
Energy 2014:682–703.
[8] Foucquier Aurelie, Robert Sylvain, Suard Frederic, Stephan Louis, Jay Arnaud. State
of the art in building modelling and energy performances prediction: a review.
Renew Sustain Energy Rev 2013:272–88.
[9] Zhao Hai-xiang, Magoules Frederic. A review on the prediction of building energy
consumption. Renew Sustain Energy Rev 2012:3586–92.
[10] Tian Wei. A review of sensitivity analysis methods in building energy analysis.
Renew Sustain Energy Rev 2013:411–9.
[11] Castleton HF, Stovin V, Beck SBM, Davison JB. Green roofs; building energy savings
and the potential for retrofit. Energy Build 2010:1582–91.
[12] Saeli Manfredi, Piccirillo Clara, Parkin Ivan P, Binions Russel, Ridley Ian. Energy
modelling studies of thermochromic glazing. Energy Build 2010:1666–73.
[13] Colorado Governor's Energy Office. Energy Modeling: A Guide for the Building
Professional; 2011.
[14] Eisenhower Bryan, O’Neill Zheng, Narayanan Satish, Fonoberov Vladimir A, Mezic
Igor. A methodology for meta-model based optimization in building energy models.
Energy Build 2012:292–301.
[15] Pedrini A, Westphal FS, Lamberts R. A methodology for building energy modelling
and calibration in warm climates. Build Environ 2002:903–12.
[16] Martin Miguel, Wong Nyuk Hien, Hii Daniel Jun Chung, Ignatius Marcel.
Comparison between simplified and detailed EnergyPlus modelscoupled with an
urban canopy model. Energy Build 2017.
[17] Pisello Anna Laura, Taylor John E, Xiaoqi Cotana, Franco Xu. Inter-building effect:
L.G.R. Santos et al.
Simulating the impact of a network of buildings on the accuracy of building energy
performance predictions. Build Environ 2012:37–45.
[18] Martin M, Afshari A, Armstrong PR, Norford LK. Estimation of urban temperature
and humidity using a lumped parameter model coupled with an EnergyPlus model.
Energy Build 2015:221–35.
[19] Reinhart CF, Cerezo Davila C. Urban building energy modeling – a review of a
nascent field. Build Environ 2016:196–202.
[20] Frayssinet Loic, Kuznik Frédéric, Jean-Luc Milliez, Hubert Maya, Roux Jean-
Jacques. Adaptation of building envelope models for energy simulation at district
scale. Energy Procedia 2017:307–12.
[21] Bueno Bruno, Roth Matthias, Norford Leslie, Li Reuben. Computationally efficient
prediction of canopy level urban air temperature at the neighbourhood scale. Urban
Clim 2014:35–53.
[22] Mao Jiachen, Yang Joseph H, Afshari Afshin, Norford Leslie K. Global sensitivity
analysis of an urban microclimate system under uncertainty: design and case study.
Build Environ 2017:153–70.
[23] Monetti Valentina, Davin Elisabeth, Fabrizio Enrico, André Philippe, Filippi Marco.
Calibration of building energy simulation models based on optimization: a case
study. Energy Procedia 2015:2971–6.
[24] Raftery Paul, Keane Marcus, O’Donnell James. Calibrating whole building energy
models: an evidence-based methodology. Energy Build 2011:2356–64.
[25] American Society of Heating. Refrigerating and Air-Conditioning Engineers
(ASHRAE). 2002. ASHRAE Guideline 14-2002: Measurement of Energy Demand
and Savings. Atlanta.
[26] Sanyal, Jibonananda, Joshua New, Richard E. Edwards, Lynne Parker. Calibrating
building energy models using supercomputer trained machine learning agents.
40
Applied Energy 215 (2018) 31–40
Wiley Online Library; 2014.
[27] Nguyen Anh-Tuan, Reiter Sigrid, Rigo Philippe. A review on simulation-based op-
timization methods applied to building performance analysis. Appl Energy
2014:1043–58.
[28] Royapoor Mohammad, Roskilly Tony. Building model calibration using energy and
environmental data. Energy Build 2015:109–20.
[29] Ruiz, Ramos Germán, Bandera Carlos Fernández, Temes Tomás Gómez-Acebo,
Gutierrez Ana Sánchez-Ostiz. Genetic algorithm for building envelope calibration.
Appl Energy 2016:691–705.
[30] Chaudhary Gaurav, New Joshua, Sanyal Jibonananda, Im Piljae, O'Neill Zheng,
Garg Vishal. Evaluation of “Autotune” clibration against manual calibration of
building energy models. Appl Energy 2016:115–34.
[31] Sun, Kaiyu, Hong Tianzhen, Taylor-Lange Sarah C, Piette Mary Ann. A pattern-
based automated approach to building energy model calibration. Appl Energy
2016:214–24.
[32] Lara, Arambula Rigoberto, Naboni Emanuele, Pernigotto Giovanni, Cappelletti
Francesca, Zhang Yi, Barzon Furio, Gasparella Andrea, Romagnoni Piercarlo.
Optimization tools for building energy model calibration. Energy Procedia
2017:1060–9.
[33] Hong Taehoon, Kim Jimin, Jeong Jaemin, Lee Myeonghwi, Ji Changyoon.
Automatic calibration model of a building energy simulation using optimization
algorithm. Energy Procedia 2017:3698–704.
[34] Allegrini J, Orehounig K, Mavromatidis G, Ruesch F, Dorer V, Evins R. A review of
modelling approaches and tools for the simulation of district-scale energy systems.
Renew Sustain Energy Rev 2015:1391–404.
[35] Yang Joseph H. The Curious Case of Urban Heat Island: A Systems Analysis; 2016.