Renewable Energy 118 (2018) 343e350

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

Development of robust meteorological year weather data

Sleiman Farah a, *, Wasim Saman a, John Boland b
a
University of South Australia, School of Engineering, Barbara Hardy Institute, Adelaide, Australia
b
University of South Australia, School of Information Technology & Mathematical Sciences, Barbara Hardy Institute, Adelaide, Australia

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

Article history: Building energy performance simulations are limited to typical meteorological weather conditions
Received 8 June 2017 available in simulation software. Such simulations are insufficient for analysing energy performance
Received in revised form sensitivity to a range of probable weather conditions. This research presents a method for developing
20 September 2017
robust meteorological weather data that can be used for energy performance sensitivity analysis without
Accepted 12 November 2017
Available online 14 November 2017
the need to access historical weather data. The method decomposes dry bulb temperature (DBT) and
global horizontal solar radiation (H) into deterministic and stochastic components. For the typical
weather data of the City of Adelaide, the deterministic component for each of DBT and H consists of a
Keywords:
Typical meteorological year (TM2)
single frequency Fourier series. The stochastic components consist of 1-lag and 2-lags autoregressive
Robust synthetic data models for DBT and H respectively. The stochastic components also include randomly selected values
Fourier series from the residuals of the autoregressive models. Based on this method, the coldest and hottest weather
Energy simulation conditions were selected to simulate the energy performance of a single space. The results revealed 39%
Levene's test more cooling and 15% less heating in the hottest year, and 14% more heating and 64% less cooling in the
Kolmogorov-smirnov two-sample test coldest year. The results indicate that simulations based on typical weather conditions only are insuf-
ficient for assessing buildings' energy performance.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction not guarantee the expected performance when weather conditions

deviate from that in the simulation. The uncertainties in weather
Energy usage in the building sector accounts for approximately conditions are often overcome by applying safety factors to oversize
20% of the total end-use of energy worldwide [1]. Reducing this different components of the system. However, without proper
high energy usage requires proper designs to improve the perfor- analysis, this approach may unnecessarily oversize the system and
mance of buildings and buildings' energy systems. Typically, the increase the capital cost or undersize the system and reduce ther-
designs are based on building energy simulation results to identify mal comfort. In either of these two cases, the operating conditions
necessary size and capacity of different components. For instance, of the system will not match those of optimal performance, which
the results of such simulations can reveal whether additional unnecessarily increases energy usage. Therefore, a proper analysis
thermal insulation for walls or roof is required. Such simulations requires simulating the building's energy performance under
consider ambient conditions and weather variables, such as dry multiple weather conditions which are as likely to happen as the
bulb temperature (DBT), relative humidity and global solar radia- typical conditions.
tion (H). Weather data, such as TM2 weather data, available in In addition, performance analysis of renewable energy and en-
simulation software are compiled from historical and estimated ergy storage systems, as in Refs. [3,4], can be improved by using
weather data based on a defined statistical method [2]. Typically, a multiple weather conditions. Similarly, when the performance of a
simulation software provides a single weather data file for a specific renewable energy system is optimised for a period more than a year
location which allows analysing a building's energy performance [5,6], instead of repeatedly using the same weather conditions for
under typical long term weather conditions. Although designs the required number of years, a variation of likely weather condi-
based on typical long term weather conditions may seem suitable tions can be applied for different years. The different yearly
to achieve typical expected long term performance, such designs do weather conditions are likely to influence the optimisation results.
To analyse the performance sensitivity in relation to weather
conditions, researchers can simulate the energy performance using
* Corresponding author. historical weather data. However, this data is often unavailable and
E-mail address: sleiman.farah@mymail.unisa.edu.au (S. Farah). incomplete [7], and detailed processing is required to organise the

https://doi.org/10.1016/j.renene.2017.11.033
0960-1481/© 2017 Elsevier Ltd. All rights reserved.
344 S. Farah et al. / Renewable Energy 118 (2018) 343e350

data in a format, such as TM2, TM3 or EPW format, that is initial data. This selection eliminates the need for detecting outliers
compatible with simulation software [8]. These limitations deny in the hourly synthetic data.
most researchers the ability to analyse building's performance The method presented in this research uses the TM2 weather
sensitivity to weather conditions. data for the City of Adelaide developed by Meteonorm. The focus of
To overcome these limitations researchers can use weather data the method is on DBT and H as the two main weather conditions
generators such as RUNEOLE [9]. Weather data generators are affecting building energy performance.
useful when weather data are unavailable. Using historical data,
and physical and statistical models, weather data generators can 2. Data deterministic component e Fourier series models
generate typical and extreme weather conditions [10]. However,
weather data generators are unnecessary when typical weather 2.1. Dry bulb temperature
data are available, especially when the weather data generators
may be unavailable to researchers. The seasonality of the daily average dry bulb temperature (DBT)
When typical weather data are available, researchers can is clearly shown in Fig. 1, with the average DBT in summer being
develop synthetic data that have similar statistical characteristics higher than that in winter. The variation of DBT can be modelled by
as the typical weather data. Synthetic weather data were generated a deterministic function; a Fourier series (FS) which has a frequency
in a top-down approach using CLIMED software. Using algorithmic equal to one cycle per year.
chains, the software generated monthly weather data which were Higher frequencies that may represent quarterly (frequency
used as input to generate daily weather data, and the daily weather equals 4) or monthly (frequency equals 12) variations could also be
data were used to produce hourly data [7]. However, this process significant to represent the data with a FS. However, the frequency
required accessing weather data similar to the considered site to power spectrum shown in Fig. 2, reveals that the power corre-
adjust the model parameters used in the software. The generation sponding to frequencies higher than the fundamental frequency
of synthetic data were also explored by using the “smooth” function (frequency equals 1) is negligible compared to the power of the
in MATLAB software to identify a trend in the weather data for each fundamental frequency. This result indicates that using the
month, and the residuals between the identified trend and the fundamental frequency is sufficient to capture most of the periodic
initial data were then randomly resampled. The synthetic data from variation in the data.
this process would be the sum of the monthly trends and the Consequently, the FS model (TFS ) of the average daily DBT can be
randomly resampled data. However, the generated synthetic data represented by Equation (1)
using this process were unsatisfactory as excessive fluctuations
were observed in the trend for the entire year [11]. Based on a TFS ¼ T þ a cosðutÞ þ b sinðutÞ (1)
previous work of Boland [12], an improvement in identifying the
trend was the use of Fourier series analysis [13]. Additional im- where T is the average temperature calculated as in Equation (2)
provements were also included in the work of Rastogi and Ander-
sen [13] for developing synthetic data based on typical weather P365
i¼1 Ti
data. These improvements included fitting a seasonal autore- T¼ (2)
365
gressive moving average (SARMA) model to the Fourier series re-
siduals and performing 3-days blocks of random sampling within where Ti is the average daily DBT.
each month to maintain the intrinsic weather inertia in the data as The value of u is calculated as shown in Equation (3)
suggested by Magnano, Boland and Hyndman [14].
While these improvements produce synthetic data statistically 2p
u¼ (3)
similar to the original data, some of the adopted procedures 365
complicate the process of synthetic data generation. For instance,
instead of fitting a SARMA model to the Fourier series residuals, an and the remaining unknown coefficients a and b in Equation (1) are
autoregressive moving average (ARMA) model should be sufficient calculated to minimise the sum of the squared errors (SSE) between
as all the important frequencies can be detected and detrended the data and the FS. The SSE is calculated as shown in Equation (4)
using the Fourier series analysis. In addition, using 3-days blocks for
X
365  2
random sampling seems unnecessary as an ARMA model of the SSE ¼ TFSi  Ti (4)
Fourier series residuals is meant to model the intrinsic inertia of i¼1
weather data.
This research presents a method for developing a robust mete- The minimum value of SSE is achieved for a and b equal to 5.122
orological year (RMY), without the need of accessing historical and 2.075 respectively, and the FS model can by written as in
weather data, based only on typical data available in simulation
software. The method decomposes the data into deterministic
(Fourier series) and stochastic (ARMA þ residuals) components.
The deterministic component is maintained the same throughout
the process of developing the RMY data, while the stochastic
component is modified by random sampling. This method has
three main differences compared with other methods. First, this
method is based on average daily values for the modelling of both
deterministic and stochastic components. The average values
smooth the data and simplify the modelling of deterministic and
stochastic components. Second, instead of resampling each month
separately, the method allows mixing the errors from different
months. The mixing allows creating a wider range of variations in
the robust data. Third, the generation of hourly data from average
daily synthetic data is based on selecting hourly data from the Fig. 1. Average dry bulb temperature ( C).
 S. Farah et al. / Renewable Energy 118 (2018) 343e350 345

Fig. 4. H-FS model versus H data.

Fig. 2. Frequency power spectrum.

The scatter plot in Fig. 5 shows no significant correlation be-

Equation (5). The plot of the FS model against the DBT data shows tween the residuals as the plotted data seem randomly scattered
clearly that the FS model can represent the seasonal temperature with no identifiable linear trend or curvature.
change as shown in Fig. 3. Although the residuals are not correlated, testing the cross-
    correlation is still important. For instance, higher solar radiation
2pt 2pt
TFS ¼ 16:651 þ 5:122 cos þ 2:075 sin (5) at time t may lead to a higher temperature at a later time t þ i. As
365 365
the analysed data are based on daily values, testing the cross-
correlation between the residuals of TFS model at day t and the
2.2. Total solar radiation residuals of HFS model at one step back (day t-1) seems sufficient.
Similar to the correlation scatter plot, the cross-correlation plot in
An analysis of the daily total solar radiation (H) similar to that of Fig. 6 shows no cross-correlation between the residuals as the
the DBT shows that the seasonal variation of solar radiation can be plotted data seem randomly scattered with no identifiable linear
represented by the FS model (HFS ) as shown in Equation (6). Fig. 4 trend or curvature.
shows that the FS model can represent the solar radiation seasonal A cross-correlation between the residuals of the HFS model at
variation. day t and the residuals of TFS model at day t-1 does not have a
    physical interpretation and therefore this cross-correlation should
2pt 2pt
HFS ¼ 204:1 þ 115:6 cos  16:1 sin (6) not be considered.
365 365 Based on this analysis of FS residuals, the absence of both cor-
relation and cross-correlation indicate that the DBT and H can be
2.3. Analysis of FS residuals considered independent variables for the generation of synthetic
data.
Although the TFS and HFS models represent the seasonal varia-
tions of the DBT and H respectively, significant differences (re-
3. Data stochastic component e auto regression models
siduals) still exist between the initial data and the FS models. These
residuals are considered to have a stochastic nature; they may be
3.1. Dry bulb temperature
random, but more probably also contain some autoregressive
characteristics. The aim of this section is to analyse the correlation
Removing seasonality from DBT data provides the FS residuals
and cross-correlation between the stochastic residuals of the two
shown in Fig. 7. The residuals seem centred around zero with no
Fourier series. The existence of correlation between the residuals of
apparent trend. If a trend existed in the residuals, the DBT data
the two variables would indicate that modifying one of the vari-
would have needed to be detrended before the development of a
ables requires also modifying the other variable by an amount that
stochastic model. The residuals also show a higher variance in
corresponds to the strength of the correlation. In contrast, the
summer compared to that in winter. This variance difference is
inexistence of correlation would indicate that DBT and H are un-
inherited from the DBT data and does indicate an unsuitable FS
correlated and can be treated separately in the generation of syn-
model.
thetic years.
To be able to fit a stochastic model to the FS residuals, an analysis

Fig. 3. DBT-FS model versus DBT data. Fig. 5. Scatter plot of DBT-FS residuals versus H-FS residuals.
346 S. Farah et al. / Renewable Energy 118 (2018) 343e350

Fig. 6. Scatter Plot of DBT-FS Residuals(t) vs H-FS Residuals (t-1).

Fig. 9. Partial autocorrelation function for DBT-FS residuals.

Using the FS residuals, the minimum value of the AR model SSE is

obtained for a equal to 0.711.

3.2. Total solar radiation

Similar to the DBT AR model, the plots of autocorrelation and

partial autocorrelation, shown in Figs. 10 and 11 respectively, reveal
that an AR (2) is a suitable model for the FS residuals of total solar
radiation. By minimizing the SSE, the H-AR (2) model can be
written as shown in Equation (8)
Fig. 7. DBT-FS residuals ( C).
t t1 t2
HFSR ¼ 0:345  HFSR  0:126  HFSR þ RtH (8)

of the autocorrelation and partial autocorrelation is required. This t , H t1 , H t2 and Rt are the FS residuals at time t, the FS
where HFSR FSR FSR H
analysis allows identifying the model form which can be an residuals at time t-1, the FS residuals at time t-2 and the AR model
autoregressive (AR) model, a moving average (MA) model, or a residuals at time t respectively.
combination of these two model forms; an ARMA model. This
analysis also allows identifying the degree of the required model;
4. Development of synthetic data
the number of lags that should be considered. The results of
autocorrelation and partial autocorrelation, shown in Figs. 8 and 9,
In a generic form, synthetic data (S) for either DBT or H are the
indicate that an AR (1) model is suitable; a rapidly decaying auto-
sum of the FS and AR models plus a random variation (RV) as shown
correlation with a peak for lag 1 in the partial autocorrelation.
in Equation (9).
The DBT-AR (1) model can be written as shown in Equation (7)
S ¼ FS þ AR þ RV (9)
t
TFSR ¼ aTFSR
t1
þ RtT (7)
where the only remaining unknown is RV.
where a, RtT , TFSR
t t1 are the auto regression coefficient, the AR
and TFSR
model residuals at time t, the FS residuals at time t and the FS re- 4.1. Grouping residuals
siduals at time t-1 respectively. The value of the regression coeffi-
cient a is obtained by minimizing the SSE of the AR model residuals. The values of RV are obtained by random sampling of the

Fig. 8. Autocorrelation function for DBT-FS residuals. Fig. 10. Autocorrelation function for TSR-FS residuals.
 S. Farah et al. / Renewable Energy 118 (2018) 343e350 347

Fig. 13. Total solar radiation residuals.

Fig. 11. Partial autocorrelation function for TSR-FS residuals.

Each group contains residuals from other months that have similar
variation to the residuals of the month that this group is assigned
residuals of the AR models RtT and RtH . The larger the sampling to. For instance, the group assigned to January would contain the
group the larger is the variation of the synthetic data. This desirable residuals in January and all the residuals from other months that
feature of having a larger variation of synthetic data may suggest are similar in distribution to the residuals in January. This grouping
using all the residuals as one sampling group. However, the re- method allows increasing data in each group as monthly residuals
siduals' variance is unstable throughout the entire year; the vari- can be in more than one group and maintaining similar residual
ance in winter is significantly smaller than that during the rest of variations in each of the 12 groups. This grouping method requires
the year as can be seen in Figs. 12 and 13. Therefore, the use of one systematic pairwise testing of residuals variance in each month.
sampling group of the residuals will produce extreme values that The test of variance adopted in this study is Levene's test which has
do not represent the initial data. Consequently, to reduce these a null hypothesis that the variances of the tested monthly residuals
extreme values, the residuals can be split into multiple sampling are similar. The tests results based on a 5% significance level for DBT
groups, with each group being formed from the residuals with and H are summarised in Tables 1 and 2 respectively.
similar variation in different months. Each of these groups can then The “Month” columns indicate the residuals of the corre-
be assigned to the corresponding months for sampling. sponding months that have been tested, and the values (0 or 1) in
For example, two groups could be formed with each group column “Hypothesis” refer to Levene's test result. The value
containing 6 months of residuals. The first group contains the 6 0 means that the null hypothesis could not be rejected and the
months which have the highest variations of residuals, and the
second group contains the 6 months which have the lowest vari-
Table 1
ations of residuals. The sampling of residuals for any month
Pairwise Levene's test results for dry bulb temperature residuals.
included in the first group will be from the first group, and the
sampling of residuals for any month included in the second group Month Month Hypothesis p-Value Month Month Hypothesis p-Value

will be from the second group. While this grouping may seem Jan Feb 0 0.537 Apr Aug 0 0.435
suitable in reducing extreme synthetic data, variance homogeneity Jan Mar 0 0.204 Apr Sep 0 0.764
Jan Apr 1 0.011 Apr Oct 0 0.410
within each of the groups is not guaranteed. Achieving a homog-
Jan May 1 0.003 Apr Nov 0 0.098
enous variance within each group would require increasing the Jan Jun 1 0.000 Apr Dec 1 0.008
number of groups, with each group containing monthly residuals Jan Jul 1 0.001 May Jun 0 0.262
with similar variance. However, increasing the number of groups Jan Aug 1 0.001 May Jul 0 0.681
would reduce the number of data in each group which may lead to Jan Sep 1 0.005 May Aug 0 0.684
Jan Oct 0 0.051 May Sep 0 0.939
limited variation in the synthetic data. Jan Nov 0 0.449 May Oct 0 0.206
To overcome the variation stability while maintaining a large Jan Dec 0 0.959 May Nov 1 0.045
number of data in each group, a new grouping method of residuals Feb Mar 0 0.062 May Dec 1 0.002
is proposed in this research. The method is based on creating 12 Feb Apr 1 0.002 Jun Jul 0 0.436
Feb May 1 0.000 Jun Aug 0 0.362
groups with each of the groups being assigned to a specific month.
Feb Jun 1 0.000 Jun Sep 0 0.244
Feb Jul 1 0.000 Jun Oct 1 0.025
Feb Aug 1 0.000 Jun Nov 1 0.008
Feb Sep 1 0.001 Jun Dec 1 0.000
Feb Oct 1 0.010 Jul Aug 0 0.963
Feb Nov 0 0.184 Jul Sep 0 0.632
Feb Dec 0 0.487 Jul Oct 0 0.095
Mar Apr 0 0.194 Jul Nov 1 0.022
Mar May 0 0.093 Jul Dec 1 0.001
Mar Jun 1 0.015 Aug Sep 0 0.631
Mar Jul 1 0.045 Aug Oct 0 0.082
Mar Aug 1 0.040 Aug Nov 1 0.019
Mar Sep 0 0.113 Aug Dec 1 0.000
Mar Oct 0 0.540 Sep Oct 0 0.244
Mar Nov 0 0.654 Sep Nov 0 0.055
Mar Dec 0 0.201 Sep Dec 1 0.003
Apr May 0 0.705 Oct Nov 0 0.295
Apr Jun 0 0.168 Oct Dec 1 0.043
Apr Jul 0 0.446 Nov Dec 0 0.458
Fig. 12. Dry bulb temperature residuals.
348 S. Farah et al. / Renewable Energy 118 (2018) 343e350

Table 2 Table 4
Pairwise Levene's test results for solar radiation residuals. Sampling groups for solar radiation.

Month Month Hypothesis p-Value Month Month Hypothesis p-Value Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Jan Feb 0 0.619 Apr Aug 0 0.264 Jan Jan Jan Jan Jan Jan Jan Jan Jan Jan Oct Jan
Jan Mar 0 0.681 Apr Sep 0 0.892 Feb Feb Feb Feb Feb Feb Feb Feb Feb Feb Nov Feb
Jan Apr 0 0.509 Apr Oct 0 0.356 Mar Mar Mar Mar Mar Mar Mar Mar Mar Mar Dec Mar
Jan May 0 0.345 Apr Nov 1 0.021 Apr Apr Apr Apr Apr May May Apr Apr Apr Apr
Jan Jun 0 0.282 Apr Dec 0 0.388 May May May May May Jun Jun May May Aug Aug
Jan Jul 0 0.194 May Jun 0 0.887 Jun Jun Jun Aug Jun Jul Jul Jun Jun Sep Sep
Jan Aug 0 0.735 May Jul 0 0.664 Jul Jul Jul Sep Jul Aug Aug Jul Jul Oct Oct
Jan Sep 0 0.692 May Aug 0 0.503 Aug Aug Aug Oct Aug Sep Sep Aug Aug Nov Nov
Jan Oct 0 0.154 May Sep 0 0.222 Sep Sep Sep Dec Sep Sep Sep Dec Dec
Jan Nov 1 0.007 May Oct 1 0.013 Oct Oct Oct Oct Oct
Jan Dec 0 0.178 May Nov 1 0.000 Dec Dec Dec Dec Dec
Feb Mar 0 0.938 May Dec 1 0.021
Feb Apr 0 0.924 Jun Jul 0 0.741
Feb May 0 0.132 Jun Aug 0 0.409 However, such sampling does not eliminate the necessity to verify
Feb Jun 0 0.098 Jun Sep 0 0.186
Feb Jul 0 0.059 Jun Oct 1 0.008
that the synthetic data conform to reality. For instance, any nega-
Feb Aug 0 0.383 Jun Nov 1 0.000 tive value of solar radiation is unreasonable and should not be
Feb Sep 0 0.962 Jun Dec 1 0.015 included in the synthetic data. Similar discretion should also be
Feb Oct 0 0.364 Jul Aug 0 0.278 used for DBT by adopting a maximum and minimum acceptable
Feb Nov 1 0.030 Jul Sep 0 0.132
temperature values, such as the historical maximum and minimum
Feb Dec 0 0.391 Jul Oct 1 0.004
Mar Apr 0 0.857 Jul Nov 1 0.000 temperature values. Such maximum and minimum values do not
Mar May 0 0.175 Jul Dec 1 0.008 require access to historical weather data as they are typically
Mar Jun 0 0.138 Aug Sep 0 0.478 available on weather websites, such as on the Bureau of Meteo-
Mar Jul 0 0.090 Aug Oct 0 0.065 rology for Australia [15]. Depending on the application of the syn-
Mar Aug 0 0.445 Aug Nov 1 0.002
Mar Sep 0 0.982 Aug Dec 0 0.084
thetic data, the maximum and minimum values could also be
Mar Oct 0 0.326 Sep Oct 0 0.381 slightly modified within reasonable and justified limits. In this
Mar Nov 1 0.025 Sep Nov 1 0.040 research, the adopted maximum and minimum values for both
Mar Dec 0 0.351 Sep Dec 0 0.402 synthetic DBT and H are equal to the maximum and minimum
Apr May 0 0.057 Oct Nov 0 0.180
values in the initial TM2 data.
Apr Jun 1 0.034 Oct Dec 0 0.997
Apr Jul 1 0.015 Nov Dec 0 0.202
4.2. Synthetic data verification

variance of residuals in the corresponding months are considered Based on the method presented, one thousand synthetic years of
similar. In contrast, the value 1 means that the null hypothesis has
DBT and H have been developed. The synthetic DBT and H data have
been rejected and the variance of residuals in the corresponding a significant variation between the different years as the sample (5
months are significantly different. The results of Levene's test are
years) synthetic data show in Figs. 14 and 15 respectively. These
based on the comparison of values, shown in the “p-Value” column, figures also show that the synthetic data are indistinguishable from
to the level of significance. For example, the first test in Table 1
the initial TM2 data (shown in black colour in Figs. 14 and 15).
shows that the residuals in January and February can be consid- Although the synthetic and original data are pictorially indis-
ered similar as the p-value (0.537) is greater than 5% and the null
tinguishable, a comparison between the two data sets on monthly
hypothesis could not be rejected. In contrast, the third test shows
that the null hypothesis has been rejected as the p-value (0.011) is
less than 5%; indicating that the residuals in January and April are
significantly different.
Based on the results in Tables 1 and 2, the sampling groups for
each month for DBT and H are formed as shown in Table 3 and 4
respectively. For instance, the first column in Table 3 indicates
that the sampling group for January contains the residuals of the
listed months, namely, January, February, March, October,
November and December.
Sampling from these groups reduces excessive-extreme values
while maintaining a wide variation of values in the synthetic data.
Fig. 14. Sample of synthetic dry bulb temperature.

Table 3
Sampling groups for dry bulb temperature.

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Jan Jan Jan Mar Mar Apr Apr Apr Mar Jan Jan Jan
Feb Feb Feb Apr Apr May May May Apr Mar Feb Feb
Mar Mar Mar May May Jun Jun Jun May Apr Mar Mar
Oct Nov Apr Jun Jun Jul Jul Jul Jun May Apr Nov
Nov Dec May Jul Jul Aug Aug Aug Jul Jul Sep Dec
Dec Sep Aug Aug Sep Sep Sep Aug Aug Oct
Oct Sep Sep Oct Oct Sep Sep Nov
Nov Oct Oct Oct Oct Dec
Dec Nov Nov Nov
Fig. 15. Sample of synthetic total solar radiation.
 S. Farah et al. / Renewable Energy 118 (2018) 343e350 349

basis is necessary to identify and eliminate the distributions of

synthetic data which are significantly different from the distribution
of the original data. The comparison between the distributions is
achieved by a two-sample KolmogoroveSmirnov statistical test
with a 5% significance level. When the test results indicate that the
null hypothesis should be rejected, monthly synthetic data used in
the test is substituted by another set of monthly synthetic data. This
process is repeated until the null hypothesis could not be rejected;
indicating that the distributions of synthetic and original data can
be considered to originate from the same distribution. This pro-
cedure of substitution is eminently reasonable since, as stated
previously, each set of the synthetic data is equally likely to occur.
Therefore, synthetic data that does not fit the required criterion can
be replaced with an equally probable data that satisfies the criterion. Fig. 17. Synthetic and TM2 global solar radiation.

4.3. Development of hourly data

perspective view of the simulated space is shown in Fig. 18, and the
The synthetic data are based on daily average values for both dimensions and orientation are shown in the front and side views
DBT and H. Therefore, to be able to implement the synthetic data in in Fig. 19.
a weather file for thermal performance evaluation, hourly data for The simulations are performed in TRNSYS using the envelope
DBT and H should be calculated. The method adopted to calculate thermal properties shown in Table 5, with 18  C and 25  C set points
the hourly data consists of selecting the day with hourly TM2 data for heating and cooling respectively. The results of these simula-
which has the closest daily average to that of the synthetic data. The tions are compared to the energy performance under the TM2
hourly synthetic DBT data also required smoothing the tempera- weather conditions of the City of Adelaide.
ture values between days as significant temperature difference may The results reveal significant differences of energy demand be-
occur between the last hour of the day and the first hour of the next tween the considered weather conditions for both heating and
day. The synthetic DBT values before and after midnight are cooling as shown in Fig. 20. As the simulation considered a simple
adjusted to limit the difference to the maximum temperature in- space, the importance of the results is not the exact demand values;
crease or decrease in the TM2 data. Based on the proposed method instead, the importance is the difference between the demand
to calculate hourly synthetic data, two samples of DBT and H required for each of the considered years. For cooling, the required
showing the differences between synthetic and original data are demand in the hottest year is 39% more than that in the TM2 year,
presented in Figs. 16 and 17 respectively. while the required demand in the coldest year is 64% less than that
in the TM2 year. Similarly, the required heating demand in the
coldest year is 14% more than that in the TM2 year, while the
5. Application of synthetic data required heating demand in the hottest year is 15% less than that in
the TM2 year.
The usefulness of the generated synthetic data can be shown by These results indicate that TM2 weather conditions are insuffi-
simulating the energy performance of a house using all the cient for analysing heating and cooling demand for the range of
developed synthetic data. Nevertheless, this approach is compu- probable weather conditions. This analysis may be less important
tationally expensive and generates an extensive amount of results, for grid-connected homes as additional energy required can be
which may complicate the resulting analysis. Typically, the interest imported from the electricity grid. However, this analysis becomes
from such energy simulation is to analyse the performance for critical for off-grid homes which have access to limited capacity of
probable extreme weather conditions, such as the coldest year, the energy supply and storage. An off-grid home designed based only
hottest year, the coldest winter and the hottest summer. Unlike the on TM2 weather conditions can experience significant excess or
historical extreme weather conditions, the extreme weather con- shortage of energy.
ditions from the synthetic data are as likely to happen as the typical
weather conditions. In this research, the usefulness of the gener-
ated synthetic data is demonstrated by simulating the energy 6. Summary and conclusions
performance of a single space for the coldest and hottest years. A
This research presented a method for using typical weather data

Fig. 16. Synthetic and TM2 dry bulb temperature. Fig. 18. Perspective view of the simulated space.
350 S. Farah et al. / Renewable Energy 118 (2018) 343e350

Fig. 19. Schematic front and side views of the simulated space.

Table 5 weather conditions is important, especially for houses with access

Envelope thermal properties. to limited energy supply and storage.
U-Value W/(m2.K) Solar Absorptance Emissivity
Acknowledgments
External Walls 0.510 0.6 0.9
Ground Floor 0.039 0.6 0.9
Roof 0.316 0.6 0.9 This research is supported by an Australian Government
Window 2.890 Varies 0.84 Research Training Program (RTP) Scholarship. This research is
funded by the CRC for Low Carbon Living Ltd, supported by the
Cooperative Research Centres program, an Australian Government
initiative.

References

[1] EIA, International Energy Outlook 2016, US Energy Information Administra-

tion (EIA), 2016, p. 4.
[2] S.A. Kalogirou, Generation of typical meteorological year (TMY-2) for Nicosia,
Cyprus, Renew. Energy 28 (2003) 2317e2334.
[3] F.M. Vieira, P.S. Moura, A.T. de Almeida, Energy storage system for self-
consumption of photovoltaic energy in residential zero energy buildings,
Renew. Energy 103 (2017) 308e320.
[4] S. Farah, W. Saman, M. Belusko, Integrating solar heating and PV cooling into
the building envelope, in: A. Hakansson, M. Ho € jer, R.J. Howlett, L.C. Jain (Eds.),
Sustainability in Energy and Buildings, Springer Berlin Heidelberg, 2013,
pp. 887e901.
nez-Fern
[5] S. Jime andez, S. Salcedo-Sanz, D. Gallo-Marazuela, G. Go  mez-Prada,
Fig. 20. Cooling and heating energy demand for different weather conditions.
J. Maellas, A. Portilla-Figueras, Sizing and maintenance visits optimization of a
hybrid photovoltaic-hydrogen stand-alone facility using evolutionary algo-
rithms, Renew. Energy 66 (2014) 402e413.
for developing robust weather data suitable for sensitivity analysis [6] H. Yousefi, M.H. Ghodusinejad, A. Kasaeian, Multi-objective optimal compo-
of building energy performance. The robust data contained a nent sizing of a hybrid ICEþPV/T driven CCHP microgrid, Appl. Therm. Eng.
122 (2017) 126e138.
Fourier series component and a stochastic autoregressive with [7] R. Aguiar, S. Camelo, H. Gonçalves, Assessing the value of typical meteoro-
random sampling component. The presented method maximised logical years built from observed and from synthetic data for building thermal
the variation of random sampling by grouping the sampling re- simulation, in: Proceedings of the 6th International IBPSA Conference ob
Building Simulation'99 in Kyoto, 1999, pp. 627e634.
siduals into twelve different groups with each group being dedi- [8] National Renewable Energy Laboratory, Data Elements in the TMY2 Format,
cated to a specific month. The grouping procedure was based on February 2017. http://rredc.nrel.gov/solar/pubs/tmy2/tab3-2.html.
Levene's test to maintain similar variation within each of the twelve [9] L. Adelard, H. Boyer, F. Garde, J.C. Gatina, A detailed weather data generator for
building simulations, Energy Build. 31 (2000) 75e88.
groups. The developed data were tested against the initial data [10] L. Adelard, T.A. Mara, H. Boyer, J.C. Gatina, Elaboration of a New Tool for
using Kolmogorov-Smirnov two-sample test to confirm that the Weather Data Sequences Generation, 2012. ArXiv Preprint arXiv:1212.5599.
distribution of the developed data did not significantly deviate [11] P. Rastogi, M. Andersen, Generation of weather files using resampling tech-
niques: an exploratory study, in: Building Simulation 2013: 13th International
from the distribution of the initial data. Conference of the International Building Performance Simulation Association,
Based on the coldest and hottest weather conditions from the International Building Performance Simulation Association, 2013.
developed data, the energy performance of a single space was [12] J. Boland, Time-series analysis of climatic variables, Sol. Energy 55 (1995)
377e388.
simulated in TRNSYS and compared to the energy performance
[13] P. Rastogi, M. Andersen, Embedding stochasticity in building simulation
under typical conditions. The results indicated significant differ- through synthetic weather files, in: Proceedings of BS 2015, IBPSA, 2015.
ences of energy performance for different weather conditions; 39% [14] L. Magnano, J. Boland, R. Hyndman, Generation of synthetic sequences of half-
more cooling and 15% less heating in the hottest year, and 14% more hourly temperature, Environmetrics 19 (2008) 818e835.
[15] Australian Bureau of Meteorology, Climate Statistics for Australian Locations,
heating and 64% less cooling in the coldest year. These results April 2017. http://www.bom.gov.au/climate/averages/tables/cw_023090_All.
suggest that energy performance sensitivity analysis for different shtml.