*3.2. Energy Analysis Results*

S

After analyzing the variations in the different weather parameters between the *on-site* and *third-party* weather data for the three weathers, the following analysis consisted of the study of the impact produced by these variations in the test sites' building energy demands using detailed BEMs. Figure 5 shows the four test sites analyzed in this study showing a real image from the building and an image of the performed model in EnergyPlus (in colors for the different thermal zones of each building).

**Figure 5.** The test sites analyzed. From left to right: office building (University of Navarre, Spain); Gedved school (Denmark); H2SusBuildand administration building in Lavrio (Greece). On top, the real buildings and, on the bottom, the building energy models (SketchUp thermal zone representation, OpenStudio plugin [67]).

The first test site was the office building attached to the Architecture School at the University of Navarre in Pamplona (Spain). It hosts administration uses and classrooms for the postgraduate students. This building is a 755 square meter single-story building built in 1974. It has a concrete structure; the outdoor walls are built of red brick fabric (*U* value = 0.3 W/m2K); the flat roof has the insulation above the deck (*U* value = 0.2 W/m2K); and aluminum window frames were installed in situ with an air chamber. The Gedved public school (Denmark) consists of six buildings and was built in 1979 and then renewed in 2007. The library of one of the school buildings was selected as the test site. It is a one-story building with a total surface area of 1138 m2, with a big central space—the library—and nine classrooms and serving spaces around it. The building walls consist of two brick layers with 150 mm insulation in between (*U* value = 0.27 W/m2K). The windows are two-layer double-glazed windows with cold frames. The ceiling is insulated with 200–250 mm mineral wool for the sloping and flat ceiling, respectively (*U* value = 0.07 W/m2K and 0.16 W/m2K, respectively), and the floor is made of concrete and contains 150 mm insulation under it (*U* value = 0.21 W/m2K).

In Lavrion (Greece), two buildings from the Lavrion Technological and Cultural Park (LTCP) were used as test sites: H2SusBuild and the administration building. H2SusBuild has a ground floor and an attic floor with a total surface area of approximately 505 m2 . The ground floor hosts a small kitchen, toilets, the control room, and the main area. The attic also hosts two offices and a meeting room. Its envelope consists of a concrete structure with double concrete block walls and single-glazed windows with aluminum frames. It also has external masonry consisting of double brick walls with 10 cm expanded polystyrene (EPS) insulation (*U* value = 0.25 W/m2K). The roof consists of metallic panels with a 2.5 cm polyurethane insulation layer in the middle (*U* value = 0.75 W/m2K). The administration building hosts the LTCP managing authority and administrative services. It is a two-story renovated neoclassic building with a surface area of about 644 m2. The building envelope is made of stone approximately 70 cm thick (*U* value = 1.85 W/m2K) with wooden-framed double-glazed large windows. The roof consists of a wooden frame with gutter tiles placed on top (*U* value = 0.49 W/m2K).

For each building, an individual pattern of use corresponding to the actual use of the building was implemented in the simulation model. Each building had its own calendar of use, occupancy, and internal loads of electric equipment and lighting. Regarding the HVAC systems, setpoints and usage hours were defined for each. The office building in Pamplona and H2SusBuild and administration building in Lavrion implemented both heating and cooling systems in the models, and the Gedved school had only a heating system. Table 2 shows the input data of the four models.



Wd: weekdays. Sat: Saturdays.

The results of the statistical analysis for the energy study are presented in Table 3. The table is divided into four sub-tables, one for each test site. They show the three uncertainty metrics calculated for the energy demand obtained from simulations using the *on-site* and *third-party* weather files (*TPW*), with *on-site* as the reference. The first column of each test site's table, designated as *TPW*, shows the difference in percentage (*MADP*) of the energy demand when the third-party weather file is used in the simulation with respect the the reference simulated with *on-site* weather data. With the inputs shown in Table 2, the models provide the following annual energy demand: 21.9 kWh/m2 for the office building, 91.5 kWh/m<sup>2</sup> for the Gedved school, 142.2 kWh/m2 for H2SusBuild, and 91.6 kWh/m2 for the administration building.

The table allows performing two different analyses depending on how it is read. The vertical interpretation of the table shows the percentage results as a function of the time resolutions (from annual to hourly criteria) employed for the analysis. On the other hand, horizontally, the variations in the energy demand for the sensitivity analysis changing only one weather parameter at a time are presented (*DHI*, *DNI*, *RH*, *Temp*, *WD*, and *WS*).

**Table 3.** Uncertainty metrics (*MADP*, *CV*(*RMSE*), and *R*2) were used in the energy analysis for the four test sites. *TPW*: the results using the weather file with all the parameters from the *third-party* weather data source. The rest changed only one parameter at a time: *DHI* (diffuse horizontal irradiation), *DNI* (direct normal irradiation), *RH* (relative humidity), *Temp* (temperature), *WD* (wind direction), and *WS* (wind speed).


The first analysis obtained from Table 3 was the influence of the time resolution used in the study of the energy demand variation. In this case, the analysis was done in the vertical from the annual to hourly criteria. The percentage metrics *MADP* and *CV*(*RMSE*) allowed us to compare the results for each time grain and study its influence in the results. Both indexes were closely related; however, *CV*(*RMSE*) gave a relatively high weight to large variations. It is remarkable that the differences between *CV*(*RMSE*) and *MADP* decreased as the time grain increased (from hourly to annual criteria) as, when the energy demand was accumulated, the outliers were minimized. The results for the hourly basis showed that the *CV*(*RMSE*) values were around twice the *MADP* values for the four sites and all the weather parameters. This indicates that significant outliers were present in the energy demand results when both simulations based on the *on-site* and *third-party* weather datasets were compared.

On the other hand, both the *MADP* and *CV*(*RMSE*) metrics showed how, in the four test sites, the variations in the energy demand grew as the time grain decreased, which matches with ASHRAE's statement about the energy data granularity [44]. If the results were analyzed with the accumulated energy demand for a period of time (i.e., monthly, annual, etc.), the energy variation was minimized with respect to the hourly analysis. For example, differences of *MADP* up to ±38% between the annual and hourly criteria are seen in the results for the administration building. In this case, the *MADP* for the accumulated data for the year was only 1.29%; thus, the annual building energy demand simulated for the *third-party* weather file was very similar to the reference, simulated with the *on-site* weather file.

However, for the hourly basis, this variation grew up to 39.45%, which is a significant deviation. The reason is because, alternately, in some cases, the model simulation overestimated the energy demand needed by the building (the model demanded more energy than the reference), and in other cases, the model underestimated it. When the data were accumulated from the hourly basis to longer periods of time (daily, weekly, monthly, seasonal, and annual), a compensation effect occurred by canceling each other out, which resulted in the minimization of the energy demand variation. As the length of the periods increased, so did the compensation effect and, therefore, also the minimization of the variations.

It is also remarkable that for all the test sites, the *CV*(*RMSE*) results showed high values for the monthly and hourly resolutions, which are the time criteria commonly used by the energy analysis standards.

The second analysis was the study of the influence of each weather parameter in the energy demand variation. In this case, the interpretation of the tables from 3 was done horizontally: the first column presents the results for the simulation with the *third-party* weather file (*TPW*), which had all the weather parameters changed, and the following columns show the results for the different weather parameters.

Comparing the results of the four test sites using the *MADP* and *CV*(*RMSE*) indexes, some common observations can be made. In all of them, the weather parameter that generated less impact in the simulated energy demand was *WD*, even though it was the weather parameter that worst fit the *on-site* weather data, as was shown in the Taylor diagram (Figure 3). The reason is because the mechanical ventilation and infiltration EnergyPlus objects used in these models did not account for *WD* in the simulations [68].

On the other hand, in the four test sites, when *WS* was analyzed, it showed an important impact in the energy demand simulations' outputs. This was mainly due to two causes: The first was the use of dynamic infiltrations introduced by using the EnergyPlus object *ZoneInfiltration:EffectiveLeakageArea*, which took into account the *WS* parameter in the calculations [68]. The leakage area in cm<sup>2</sup> was calculated by the calibration process previously developed by the authors [69–71]. The second was because the differences between the *third-party WS* data and *on-site* data (see the Taylor diagram in Figure 3 and the wind roses from Figure 4) were significant.

In both the Gedved school and H2SusBuild, the *third-party* wind speed provided faster values, which generated a significant increase in the energy demand during almost all the year, but there were a few moments with a decrease in the energy demand. Therefore, the compensation effect between the overestimated and underestimated energy demand was reduced. This explains why the variation due to *WS* was high for all the time grains for these test sites. This effect was especially clear in the Gedved school, which did not have a cooling system. In this case, all the time grains for *WS* provided the same *MADP* because the higher *WS* of the *third-party* data always meant a higher heating energy demand and no energy demand compensation existed.

Regarding the *Temp* parameter, in the weather data analysis (Section 3.1), based on an hourly time grain, it was the parameter with less variation between *on-site* and *third-party* data for the three sites. However, the *MADP* and *CV*(*RMSE*) results, especially for the hourly criteria, showed that it had a significant impact on the energy demand in the four test sites. It was the second parameter of influence for the Gedved school, H2SusBuild, and administration building after wind speed and the first one in the office building with an *MADP* of ±26%. In relation to the solar irradiation, the Taylor diagram (Figure 3) showed that *DHI* from the *third-party* weather data provided a better correlation than *DNI* for the three locations, and this was reflected in the sensitivity energy analysis. For the Gedved school, these two parameters had less impact on the energy demand than for the other three models. The reason is because the school lacked a cooling system; therefore, in summer, when more solar access was available, no energy demand was taken into account.

To conclude the explanation of Table 3, factor *R*<sup>2</sup> was analyzed. It compared the shape of the energy demand curves from the different simulations and showed that the energy demand simulated with the *third-party* weather data fit quite well with the energy demand simulated from the *on-site* data for the four test sites (with *<sup>R</sup>*<sup>2</sup> between ±82% and ±95%). Regarding the different weather parameters, the results for each parameter matched with the analysis of the hourly percentage indexes. The parameters with lower hourly *MADP* and *CV*(*RMSE*) values had higher *R*<sup>2</sup> values.

Finally, to show in a visual way the previous analysis of the influence of the time resolution employed in the study and the sensitivity of each weather parameter, the *MADP* index results are plotted in Figure 6. Each graph presents the *MADP* result for each test site. In the graphs, the six time resolutions are shown on the *x* axis, and the dashed line presents the results for the simulation with all the *third-party* weather parameters (*TPW*). Each color represents the results for each weather parameter of the sensitivity analysis. The graphs show how the variations in the energy demand grew as the time grain decreased, especially *Temp*. They also show that *WS* was the most sensitive weather parameter for the Gedved school, H2SusBuild, and administration building. Only in the case of the office building in Pamplona was *WS* the most sensitive weather parameter taking into account an annual criterion; however, per hour, it changed to *Temp*.

Previous analyses showed the significant variations in the energy demand when using different *actual* weather datasets. In order to study if these differences in the energy demand were mostly due to the building architectures or to the weather, a complementary theoretical analysis was performed, and this is presented in the following section.
