**4. Discussion**

This paper shows how to study the impact of using two different *actual* weather datasets on building energy model simulations (weather data for the year 2019): one weather dataset with data measured in the building's surroundings (*on-site*), which was considered the reference weather data; and the other supplied by a weather service provider (*third-party*). Four test sites with different uses and architecture characteristics, located in three different locations, were employed in the studMor diagram, (2) an energy demand comparison, and (3) an indoor temperature comparison. In the case of the energy approach, a sensitivity analysis of the main weather parameters was also performed to study the influence of each parameter in the energy simulation. The energy results were provided with a different temporal resolution from the annual to hourly criteria in order to highlight the differences in the results.

The results for the energy analysis showed that as the time grain decreased, the impact of using different weather datasets grew, which agrees with ASHRAE [44]. The differences between the annual and hourly *MADP* until 38% are shown in the results. This was because when the energy demand was accumulated in periods of time longer than an hour, the variation in the results was minimized due to the compensation effect of the underestimated and overestimated energy use. This must be taken into account when a weather data source is chosen according to its purpose. For instance, if the weather data will be used for model calibration purposes, it is important to take into consideration the monthly or hourly criteria, as the most used standards (*ASHRAE* [44], *FEMP* [58], and *IPMVP* [59]) employ these time grains for their recommendations, and as the study showed, the use of different weather datasets had a significant impact on the *CV*(*RMSE*) results. Another application of BEM where the time grain of the analysis is relevant is model predictive control, where the hourly criteria are required.

The sensitivity analysis of the main weather parameters showed the different influence that each parameter had on the energy demand variation of each test site. In this regard, the relative humidity and wind direction had little influence on the models. In the case of the wind direction, the low influence was due to these test sites using mechanical ventilation instead of natural. On the other hand, the results also showed that the two parameters that produced a higher impact in the energy use were the wind speed and temperature. The high influence in the energy demand due to wind speed was explained by the *third-party* wind speed data having a low correspondence to the *on-site* data and because the models employed in the study used dynamic infiltrations that took into account the wind speed, instead of other models with constant values in the infiltration parameters. Therefore, the energy results for the wind speed showed that particular attention should be paid to this parameter when BEMs use dynamic infiltrations, as it has a great influence on the model's energy performance.

With the available data, the results obtained in this study suggested that for this models, some of the weather parameter data could be obtained from *third-party* weather sources, avoiding the installation of on-site sensors, as they had a low influence on the simulation results. This is the case of the relative humidity, wind direction, and even diffuse horizontal irradiation, the sensor being very expensive. On the other hand, to obtain the wind speed and outdoor temperature data, which are the weather parameters that were shown to be the most influential in the models' energy performance, we recommend the installation of an anemometer and a temperature sensor near the building. Having both *on-site* and *third-party* weather data sources would allow the verification of the data. An *on-site* sensor would also provide information regarding the micro-climate generated due to the surrounding characteristics of the building, which could be difficult to see reflected in the calculated weather data from a *third-party*.

The energy study showed that the weather dataset selected for the dynamic energy simulations had a great impact on the buildings energy performance, especially for short temporal resolutions. To emphasize the impact of the weather datasets in the building energy models, a theoretical study was performed, simulating all the test sites with the Pamplona weather file. The results showed that all the weather parameters produced similar variations in the energy demand and also a similar trend of the curve for the different time grains, independently of the model. This demonstrated the significant role played by the weather data and the importance of their correct selection when performing the building energy simulations.

In the case of the indoor temperature study, the significant impact of using different weather datasets was also shown. Although the high *R*<sup>2</sup> results for the four test sites showed that the shape of the indoor temperature curves was similar when both the *on-site* and *third-party* weather files were used, the quantitative metrics demonstrated a significant influence on the indoor temperature of the test sites with a *MAE* higher than 1.5 ◦C in some cases.

This paper showed the variation in the simulation results when two different *actual* weather datasets (*on-site* and *third-party*) were employed. In future research, it would be interesting to collect the empirical energy and temperature data from the test sites to study which of the weather datasets is closer to reality when comparing the simulation results using both weather datasets and the actual energy and temperature measurements.
