**1. Introduction**

The sustainable development goals report of 2019 highlighted the concern of the United Nations toward a more sustainable world where people can live peacefully on a healthy planet. One of the most important areas for the protection of our planet is the actions to mitigate climate change. "*If we do not cut record-high greenhouse gas emissions now, global warming is projected to reach 1.5* ◦*C in the coming decades*" [1]. This concern was also endorsed by 186 parties in the Paris agreement on climate change in 2015 [2]. One of the strategies for tackling climate change is to reduce energy consumption (by increasing the system efficiency) and increase the use of clean energy so that greenhouse gas emissions are reduced. In this process of decarbonization, the buildings and the construction sector are critical elements, since as the Global Status Report for Buildings and Construction highlighted, they are responsible "*for 36% of final energy use and 39% of energy and process-related carbon dioxide (CO2) emissions in 2018, 11% of which resulted from manufacturing building materials and products such as steel, cement and glass.*" [3].

For this reason, building energy models (BEMs) play an important role in the understanding of how to reduce the energy consumed by buildings (lighting, equipment, heating, ventilation, and air-conditioning (HVAC) systems, etc.) and how to optimize their use. As Nguyen et al. highlighted, there is a huge variety of building performance simulation tools [4], and EnergyPlus is one of the most used [5]. In any simulation program, to obtain reliable results, the model must not only be adjusted to the behavior of the building, but all the files on which it depends must be reliable and suitable for the use that will be given to the model. Of all these files, the weather file is, perhaps, one of the most important [6].

Bhandari et al. highlighted that there are three kinds of weather data files, *typical*, *actual*, and *future*, which should be selected according to the use of the energy model [6]. The first corresponds to a representation of the weather pattern of a specific place taking into account a set of years (commonly 20–30 years). For each month, the data are selected from the year that was considered most "typical" for that month so that it represents the most moderate weather conditions, excludes weather extremities, and reflects long-term average conditions for a location. In general, they are used to design and study the behavior of the building under standard conditions, to obtain Energy Performance Certificates, to study the feasibility of a building's refurbishment, etc. The typical files are not recommended in extreme conditions, such as designing HVAC systems for the worst case scenario.

There are two main sources of typical weather files: those developed by the National Renewable Energy Laboratory (NREL), which come from stations in the United States and its territories, where the different versions (typical meteorological year (TMY), TMY2, and TMY3) take into account different numbers of stations, time periods, solar radiation considerations, etc. [7,8]; and those developed by the American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) as a result of the ASHRAE Research Project 1015 [9,10]: the International Weather for Energy Calculations (IWEC), which takes into account weather stations outside the United States and Canada. The latest version (IWEC2) covers 3012 worldwide locations [11] taking into account a 25 year period (1984–2008) from the Integrated Surface Hourly (ISH) weather database.

The second kind of weather data file, *actual*, corresponds to a specific location and time period. It could be obtained from an on-site weather station or by processing data from several nearby stations. The latter option is commonly used by external data providers. This type of file is usually used to carry out building energy model calibrations, calculate energy bills and utility costs, study the specific behavior of HVAC systems, etc. As can be seen, these files take into account extraordinary situations, such as heat waves, adverse or extraordinary atmospheric phenomena, etc. [12].

Finally, the last kind of weather data file, *future*, is mainly used to simulate how it is possible to adapt the building energy demand to an external energy requirement (demand response [13]) or to obtain better use strategies of HVAC systems in certain situations (model predictive control [14,15]). As Lazos et al. highlighted, there are three common groups of forecasting techniques: statistical, machine learning, and physical and numerical methods [16].

Therefore, each kind of weather data file has a certain purpose; therefore, the reliability of the results will largely depend on the accuracy and suitability of the file selected [17,18].

Many studies highlighted the importance of the weather file when performing building energy analysis, as its accuracy is generally assumed, although it is out of the control of the person in charge of the simulation, its difference being significant. The following are some examples of studies that use *typical* weather files for different applications where the weather data are relevant. It is meaningful in retrofitting scenarios [19], or when quantifying renewable energy as in the cases when sizing photovoltaic solar panels [20], or when used to obtain ground temperatures [21], even when studying climate change; the accuracy of the weather files is also very important, as they are the baseline files in the process of morphing the data [22–24]. Therefore, when using such files, it is important to try to use the most recent [25].

There are other studies that analyze the building energy performance using *actual* weather files, either from commercial vendors [26], or developed using nearby weather stations, not located in the building [6,27–29], or the lesser ones, from weather stations placed in the building or in its surroundings [30]. Finally, regarding the *future* weather files, the uncertainty of the forecast files is closely related to the accuracy of the files on which they are based [31–33]. There are many studies that highlight the importance of the weather files, measuring, for example, their impact on passive buildings [34], on micro-grids [35,36], etc., calculating the loads of district energy systems [37], the electricity consumption with demand response strategies [13], or evaluating the effect on comfort conditions [38]. Some analyzed the effect that certain parameters of the weather file have, emphasizing the temperature as the most sensitive value for load forecasting [39,40].

In terms of temporal resolutions, most research focused on annual results when comparing different sets of weather data. Wang et al. analyzed the uncertainties in annual energy consumption due to weather variations and operation parameters for a reference office prototype, concluding that uncertainties caused by operation parameters were much more significant than weather variations [26]. Crawley et al. analyzed the energy results using measured weather data for 30 years and several weather datasets for a set of five locations in the USA, and the variation in annual energy consumption was on average ±5% [17]. Seo et al. conducted a similar study, also analyzing the peak electrical demand with similar results: a maximum difference of 5% [41].

In terms of research that focused on monthly criteria, Bhandari et al. found that, when using different weather datasets, the annual energy consumption could vary around ±7%, but up to ±40% when monthly analysis was performed [6]. Radhi compared the building's energy performance of using past and recent (up-to-date) weather data with annual and monthly criteria. This showed a difference of 14.5% between the annual electricity consumption simulated with past data and actual consumption, while this difference grew up to 21% for one month when monthly criteria were used [42]. Finally, there were a few studies where the temporal resolution was less than one month. With weekly criteria, Silvero et al. compared five different weather data sources with the observed meteorological data, showing that, for the annual criteria, the results were similar, but for the hottest and coldest weeks of the year, the inaccuracies increased [28].

The aim of this work is to show how to evaluate the impact of using two different *actual* weather datasets on the building energy model simulations, measuring both their effect on energy demand and indoor temperature. The purpose is to analyze if the data provided by the *on-site* weather stations (with a high economic cost and maintenance) could be replaced by the *actual* data provided by a *third-party*. For this study, four test sites based on real buildings were used to compare the existing variations when weather files with data obtained from real stations and external provider were used in the simulations. These test sites are part of an EU funded H2020 research and innovation project SABINA (SmArt BI-directional multi eNergy gAteway) [43], which aims to develop new technology and financial models to connect, control, and actively manage the generation and storage of assets to exploit synergies between electrical flexibility and the thermal inertia of buildings. The energy demand variation analysis was measured by grouping it into different periods (annual, seasonal, monthly, weekly, daily, and hourly) since as explained by ASHRAE [44], ". . .*the aggregated data will have a reduced scatter and associated CV(RMSE), favoring a model with less granular data."* The objective of the paper is to highlight these differences in the results when using different granular criteria since, depending on the use of the building energy model, their influence can be significant, for example for calibration purposes, where the monthly or hourly criteria are required. A sensitivity analysis was also performed to evaluate the influence of each weather parameter on the energy demand variation when using the two different *actual* weather datasets.

The main contributions of this research are: (1) four real test sites, with different uses and architectural characteristics, located in three different climates were employed in the study; (2) while most of studies that analyze weather data influence in building energy simulation use the typical meteorological year (TMY) [19–24], this study performed a comparison of two *actual* datasets: *on-site* and *third-party* weather data; (3) when the studies used *actual* weather data, they usually lacked a local weather station due to its expensive installation [6,27–29]; instead, the observed weather data from this study were provided by three weather stations installed on the building roofs or in their surroundings, providing on-site measurements; and (4) the energy results are shown with different temporal resolutions (from annual to hourly) in order to highlight the differences in the variations when the data are accumulated.

The paper is structured as follows. In Section 2, we describe the methodology used to analyze both the differences between the *on-site* and *third-party* weather datasets and the variation produced by these weather files in the results of the simulations in terms of the energy demand and in terms of the indoor temperature. In Section 3, we show the results obtained from the different approaches: the weather datasets comparison (Section 3.1), energy demand (Section 3.2), and indoor temperature (Section 3.3). Finally, in Section 4, we discuss the results obtained in the study, and in Section 5, the conclusions are presented.
