NV refers to the process of supplying air to and removing air from an indoor space without using mechanical systems. Taking advantage of pressure differences arising from natural forces, for example, wind-driven force and buoyancy-driven force, the external air flows into an indoor space, while the internal air flows out (Asfour, 2015). With a proper design, NV can replace cooling systems in the milder months of the year, reducing ventilation- and cooling-related energy demand (Dutton et al., 2013). The NV strategy can also be integrated with the mechanical HVAC system, forming a mixed-mode strategy that allows alternatives between the HVAC and NV throughout the day or the year depending on weather conditions (Luo et al., 2015). The advantages of applying the NV strategy in buildings include but are not limited to 1) reducing operation costs, 2) increasing occupants’ thermal comfort, and 3) improving air quality due to more fresh air.

While many studies emphasize the advantages of NV, the actual NV use can be affected by many external factors such as the outdoor climate and pollutant levels (Zhou et al., 2015; Costanzo et al., 2019). The results of the study by Martins and Carrilho da Graça (2017) showed that using NV in moments when the outside weather is favorable can result in HVAC energy savings of 25–80%. However, limiting NV use to moments with outdoor particle levels below 12 μg/m³ decreases this energy-saving potential to 20–60%. In the majority of the cities analyzed in the study by Martins and Carrilho da Graça (2017), the use of NV led to an increase in indoor exposure to PM_2.5 of outdoor origin of 400–500%. Additionally, building occupants’ habits and altitudes in interacting with building openings like the windows will also affect the NV performance by a large margin. Therefore, understanding how these factors would limit the NV usage and how to consider these factors in real-building NV operation are of great value.

On the one hand, many studies have investigated the correlations between window status and the outdoor climate conditions. Raja et al. (Raja et al., 2001) studied the relationship of windows, doors, blinds, fans, etc. with indoor and outdoor temperatures in 15 naturally ventilated office buildings in Oxford and Aberdeen, in the UK, during a summer period. It is found that proportion of open windows increases with an increase in indoor and outdoor temperature. Only few windows are open when the outdoor temperature is below 15°C, whereas most windows are open when the temperature exceeds about 25°C. Nicol and Humphrey (2001) surveyed the window usage in naturally ventilated buildings in the UK, Sweden, France, Portugal, Greece, and Pakistan. They found that occupants started to open windows at a temperature above 10°C, and as the temperature rises, there is an increased probability that a window will be open. Based on the observations in the literature, the percentage of open windows, opening hours, and the frequency of opening or closing windows depend on seasons, outdoor temperature, indoor temperature, time of the day, and the presence of the occupants. For a well-designed building with low internal gains, user-controlled windows may be opened for outdoor temperatures for as low as 10°C (Raja et al., 2001). The typical maximum outdoor temperature for NV use in an office is 28°C (de Dear and Brager, 2002); above this temperature, the indoor environment tends to become uncomfortably warm.

On the other hand, few studies have considered outdoor particle levels as limiting factors of NV use. When the building operates in the NV mode, windows will be opened to promote the large outdoor airflows that are required for ventilated cooling so that indoor exposure to outdoor particulate matter (PM) can be significant, with I/O ratios that are close to one (Martins and Carrilho da Graça, 2017). To avoid this problem, a building connected to an outdoor PM_2.5 sensor network must close the NV openings and revert to conventional HVAC during periods of high PM_2.5. This requirement reduces the number of hours when NV can be used and requires HVAC energy consumption during these periods. It is likely that the magnitude of this impact will depend on the local weather condition and particle source patterns, as particles suspended in the outdoor air are in an unstable state and their concentration can be changed by meteorological conditions, such as precipitation and wind sweeping. In the majority of urban environments, outdoor air is a source of pollutants that have a detrimental impact on indoor air quality (IAQ). There is strong evidence of adverse health effects from exposure to airborne particles that are small enough to be inhaled (diameter below 10 µm (Fenger, 2009; Talbott et al., 2015)). Limiting airborne particle exposure has long been a priority of the World Health Organization (WHO), leading to continuously updated guidelines for maximum short-term and annual mean exposures to airborne particles (WHO Regional Office for Europe, 2013). Continued exposure to PM_2.5 in amounts that are just above the natural background concentration of 3–5 μg/m³ can cause adverse health effects (Nicol and Humphrey, 2001). The combination of a mostly anthropogenic origin and an increased exposure risk makes PM_2.5 the preferred indicator for assessing health impacts from outdoor particles.

Internet of Things (IoT) environmental sensing platforms for the measurement of various environmental parameters are deployed on the urban and building scale. Low-cost sensing platforms provide higher measurement granularity than the government-owned and operated air quality station on an urban scale (Morawska et al., 2018). On a building scale, a number of different environmental sensing platforms are deployed to enable visibility of indoor conditions to building managers and potentially occupants (Parkinson et al., 2019). These two fields of science and engineering are currently kept entirely separate, with a single publication showing the potential of integrating indoor and outdoor data for effective building operations during wildfire (Pantelic et al., 2019). There is a clear gap in understanding of how to use available information to better operate buildings or to understand key aspects of how buildings work, especially naturally ventilated buildings that are dependent on outdoor conditions, indoor conditions, and the behavior of occupants.

This study aimed to demonstrate how a combination of an IoT environmental sensing network implemented locally outdoors and indoors can help to determine NV potential and actual utilization in the building throughout the year. NV potential is evaluated by considering outdoor climate variance and air pollution levels, and the actual NV use was evaluated using environmental data and occupant behavior with respect to the use of windows. In doing this, we analyzed indoor and outdoor PM_2.5, indoor CO₂, and outdoor air temperature throughout the year at different spatial (from room scale to building level and local weather stations) and temporal scales (instantaneous and long-term) to demonstrate the application of NV assessment tools. We used an NV case building located in Berkeley (California) during the period of August 2018 to the end of 2019. By reliance on an NV potential index considering outdoor PM_2.5, temperature, and wind speed, the study sought to quantify the NV availability throughout the year.

The IoT sensing network was deployed in Wurster Hall, a 9-floor mixed-mode operated building in Berkeley, CA. During the spring, summer, and autumn, the building is operated in the natural ventilation mode, while in winter there is mechanical heating. When the building operates in the NV mode, it relies on operable windows for ventilation and cooling. The building has a high level of infiltration, as shown by typical CO₂ levels below 550 ppm during normal operation. The building has multiple function rooms, such as a classroom, conference room, private, and open offices, with approximately 300 full-time occupants.

The IoT sensing network was deployed on July 19, 2018, and the study ended in December 2019. During this time period, the town of Berkeley was affected by the Chico Camp wildfire from November 18, 2018 to November 25, 2018. This offered a great opportunity to investigate how high outdoor particulate pollutants would affect the NV use and how building occupants would respond to these episodes. When doing the test, the building level of outdoor temperature and PM_2.5 concentration, room level of indoor CO₂ and PM_2.5 concentrations, and window status were monitored. The indoor sensors were installed in 15 rooms (the room number and room functions are listed in Table 1), including classrooms, meeting rooms, and office rooms. Figure 1 shows the locations of installed sensors on the third floor. The window closure sensors were installed at the bottom or on the side of the frame of openable windows. The CO₂ and PM_2.5 sensors were installed on the walls ∼1.2 m from the floor.

The study also contained surveys of building occupants about their motivations or reasons for open windows. Detailed survey questions can be found in the Appendix. Questions could be classified into several categories of which there are three main aspects related to thermal comfort, air quality, and psychological motivations.

The study utilized two types of indoor PM_2.5 sensors (Clarity Inc., and Senseware) and outdoor PM_2.5 sensor (Clarity Inc.). The outdoor sensors were installed at the top of Wurster Hall, a 9-floor building. Indoor sensors were placed in different spaces in Wurster Hall (see Figure 1 as examples). Both Clarity and Senseware PM_2.5 nodes count the particle number using the principle of light scattering. The accuracy of all the sensors was the same—within ± 10 μg/m³ in the range of 0–100 μg/m³ and ±10% in the range of 100–1,000 μg/m³. Data were collected at 15 min intervals for Clarity nodes and 1 min intervals for Senseware nodes. Clarity and Senseware PM_2.5 nodes are factory-calibrated with Arizona Test Dust (ATD). To accurately measure PM_2.5 concentrations, the Clarity nodes apply colocation of nodes and post-processing correction factors based on referencing government air quality stations. We colocated all sensors and adjusted readings. Indoor CO₂ levels were measured by Senseware CO₂ sensors with an accuracy of ±50 ppm from 400 ppm to 2000 ppm. The CO₂ sensors were paired with PM_2.5 measurements.

Window status was monitored with Senseware contact sensors (Model COZIR-LP) placed on the windows. The detection was binary (i.e., open/closed) and did not provide information on the opening area or window angle. The on/off signals from the sensors enable us to know generally if the window was open or closed, but not to which extent the window was open. All the windows in the buildings were awning type, opening outward for up to 45°. Friction hinges on the windows were able to maintain the angle after they were open. The contact sensors sensed status information through Senseware IoT platform every 5 min.

Weather conditions including outdoor temperature and wind speed were collected from the Weather Underground webpage from the Oakland-9925 International Boulevard, which is 16 km from Wurster Hall. Outdoor PM_2.5 concentrations were collected from California Air Quality Board webpage from the Oakland-9925 International Boulevard (station 1), Berkeley Aquatic Part (station 2), and Oakland-Laney College stations (station 3), which are 10, 16, and 9 km away from Wurster Hall, respectively. PM_2.5 and CO₂ concentrations were also measured at the edge of the roof of Wurster Hall.

Figure 2 shows the overall distribution of hourly average indoor and outdoor PM_2.5, indoor CO₂, and outdoor temperature during the period from July 19, 2018 to the end of 2019. During this period, the median outdoor PM_2.5 concentration was 3.5°μg/m³, slightly higher than the 1.4 μg/m³ of indoor PM_2.5. A majority of the PM_2.5 concentration values, 97.3% for outdoor and 95.5% for indoor, were lower than the recommended value of 25 μg/m³. With a median value of 437.2°ppm, the measured indoor CO₂ concentration was low, 94.0% of the time was lower than 500°ppm and 72.3% of the time was lower than 450°ppm. The outdoor temperature ranged from 11.5°C (lower quartile) to 17.4°C (upper quartile), with a median value of 14.0°C. But there were also few extreme weather conditions with outdoor temperatures lower than 10°C or higher than 30°C.

To see how these parameters varied with time, Figure 3 presents them in the time series view. Most of the time, both indoor and outdoor PM_2.5 maintained at a level lower than 25 μg/m³. But there were some short periods, for example, August 15 to August 28 and November 8 to November 25 in 2018, and the PM_2.5 concentrations climbed up to as high as over 200 μg/m³. This sudden change in the pollutant level was caused by wildfire, which will be discussed in the later analysis. Different from the PM_2.5, the CO₂ concentration fluctuated over the study period, while the temperature was mainly affected by the season.

Figure 4 shows the E index and I/O ratio of the PM_2.5 and CO₂ concentrations, respectively, for each room. When there was no wildfire, the indoor PM_2.5 concentrations were usually lower than 25 μg/m³ so that the E indexes were smaller than 1, and only 1.9% of E indexes were larger than 1. When there was a wildfire, the indoor PM_2.5 increased significantly, resulting in higher E indexes, and 93% of those are larger than 1. The observation can be validated by the I/O ratio which compares the indoor and outdoor PM_2.5 concentrations. For non-wildfire periods, the room level I/O ratios were usually below 0.5, with only 0.6% of spikes higher than 0.5. During the wildfire period, the I/O ratios climbed to 0.5–1 range. Different from PM_2.5, the indoor CO₂ concentration in different rooms was not significantly affected by extreme events like the wildfire.

To investigate how the building occupants interacted with the windows, Figure 5 shows the use of windows at three levels. First, the building-level chart compares the night-time and working-time window usage during the wildfire period and that in the non-fire period. 21% of windows were open during the non-fire period, while all of the windows were closed if there was a wildfire. The window usage were similar in nighttime and working time. 24.3% windows were open during working hours, close to the 21.5% of night hours. Second, different room functions may have different window usage rates. The classroom with 63.1% of window open time used the windows more frequently than the meeting room and office room. Third, the window usages for each room varied significantly during the non-wildfire period. Some rooms opened the windows for over 50% of the time, while some others kept the window closed all the time. This observation suggests that usage pattern of the windows depends on the occupants’ behavior and also on the number of occupants in the room. Some occupants opened and closed the window actively, some did not.

Figure 6 shows window usage during different outdoor conditions. The visible outdoor pollutants like the PM_2.5 significantly affect the window usage. Occupants tended to close windows when the outdoor pollutant level was “heavy” (PM_2.5 > 100 μg/m³) and open windows more when the outdoor pollutant level was “low” (PM_2.5 < 25 μg/m³). At the same time, the outdoor temperature can also affect the window usage. As the outdoor temperature increased from “cold” (<10°C) to “hot” (>26°C), there was an increasing percentage of opening windows.

The survey depicted in Figure 7, asked occupants about their motivations and reasons for opening windows and shows that the primary reason was to improve their thermal comfort and indoor air quality. A significant body of knowledge already exists on the use of NV for thermal comfort and indoor air quality. These reasons are intuitive, and often adopt the perspective of occupants. People open windows if they want to feel cooler or they feel that air is stuffy and needs to be refreshed. Psychological reasons like feeling connected to the outdoors ranked third in this survey. This is a very important finding that points out the third most important group of drivers that cause people to use windows. These groups of reasons are responsible for the significant amount of time windows are actually used when compared to the total amount of hours windows are open.

The results depicted in Figure 7 also show that energy conservation or energy savings ranked as fourth in this survey with 5% responses. This suggests that energy saving does not play a very important role in the occupant’s decision-making process when it comes to keeping windows open or closed. Details of the survey can be found in the Appendix.

Figure 8 shows the effects of window status on indoor PM_2.5 and CO₂ concentrations. When outdoor air was polluted, the median of I/O and E-index were much higher. Although the I/O ratio is higher due to equilibrium between indoor and outdoor conditions, E-index is important to show if those levels have impact on occupant’s health. Desired building operation would be to have the I/O ratio ∼1, indicating that windows are open and outdoor air is entering indoor environment but keeping E-index < 1 suggesting that air is clean. During the wildfire period, although the windows were kept closed, the median I/O (about 0.48) and median E-index (about 1.1) were much higher than those of non-wildfire periods with open and closed windows. The median I/O ratio and E-index were both close to 0.1 during the non-wildfire period. CO₂ concentration varied with the occupancy of the rooms but did not change significantly with the window status. During the non-wildfire period, CO₂ was 424 and 428°ppm when windows were closed and open, while it grew slightly to 453°ppm when there were wildfires outside the building. During the wildfires, although the windows were closed, only partial occupancy was in the building. To this end, the CO₂ concentrations were similar to those during the open-window periods.

Figure 9 shows the hourly deviation between the room-level PM_2.5 and CO₂ and the building-level median values. Both during the wildfire period and the non-fire period, the PM_2.5 and CO₂ deviations were around 0, and they all have very small quartile numbers, indicating that the difference between the median of each room and the median of the whole building was very small. Given this, every room can be used as a partial study of the whole.

Figure 10 shows the comparison of outdoor temperature, wind speed, and PM_2.5 concentration at different weather stations from January 2019 to January 2020. For outdoor temperature, the variation and the distribution are similar from three weather stations. For wind speed, the violin plot shows that the three weather stations had approximately the same temperature distribution with similar medians. The p values from Wilcoxon tests show no significant difference among weather stations. For outdoor PM_2.5 concentration, data from three weather stations are close to each other, and the outdoor PM_2.5 concentration measured at the top of Wurster Hall (the case building) was almost the same as those from the three weather stations, all in a very low level. Overall, it can be seen that outdoor temperature, wind speed, and PM_2.5 concentration from different weather stations show little difference.

Figure 11 shows the NV potential index throughout the study period. The red color represents periods when outdoor PM_2.5 pollution was above 25 μg/m³, periods caused by episodic events like the wildfire. The blue and brown colors represent cold and hot outdoor conditions, respectively. The gray color marks periods when mean outdoor wind speed is higher than the upper limit calculated using Eq. 5. The green color depicts periods when outdoor conditions are favorable and NV potential is available. Results show that the high outdoor PM_2.5 concentrations mostly occurred during the hot and dry seasons, usually from August to November. The hot outdoor temperatures mainly occurred from August to October, while the cold outdoor conditions mostly happened from November to March of the next calendar year. Throughout the year of 2019, 5,075 h out of 8,287 h (with valid data), resulting in 61.2% of time, were favorable for the use on NV.

Table 2 compares the potential NV availability and the actual window opening hours in typical rooms. There is a huge gap between the actual NV usage and the potential availability. Even using the more conservative method (Eq. 3), the actual usage rate was less than 34.5%. This significant gap can be attributed to the occupant’s behavior. In an earlier study, Gao et al. (2014) showed that indoor conditions are better when manually opening windows is replaced with automatic window opening. A breakdown of the multiple motives for opening the windows is shown in Figure 7. Opening the windows or keeping them closed was always based on the occupant’s perception of conditions and without knowing if conditions outside are suitable for utilization of NV or if indoor conditions can be improved with the use of NV. Informing occupants that NV is available and should be used can potentially reduce the gap between actual NV use and available potential. A previous work that focused on CO₂ and classroom pointed out that visual signals were effective means of increasing NV use (Wargocki, 2015).

Figure 12 takes room 272 (a classroom) as an example to show the actual NV usage. It can be seen that occupants in that room frequently interacted (open/close) with the windows. From February to May, the windows were frequently opened to take advantage of the NV when the room was occupied. The indoor CO₂ and PM_2.5 were very low during the measured period. CO₂ was typically below 600 ppm, while the PM_2.5 was below 25 μg/m³.

This study shows how to integrate indoor and outdoor information to better describe building operation. Previously, the study by Pantelic et al. (2019) demonstrated how IoT sensing can be used to describe building resilience to episodical pollution events. The current study builds on that knowledge and extends it to demonstrate how environmental and occupant behavioral data can be combined and quantify available natural potential and level of utilization of the potential. Results in this study indicate that IoT sensing information needs to be communicated with the building occupants to improve their use of NV. This is a new and largely unexplored field. When doing this, there are some limitations that could be noteworthy for future studies. First, the percentage of missing data in this study (Table 1) is relatively high for quantifying the window usage and indoor air quality throughout the year. Second, if the outdoor weather conditions, including outdoor temperature, relative humidity, wind speed, and direction can be monitored, they could provide useful information when comparing with nearby weather stations. The current study used information from the weather station 5 km away from the buildings and omitted the microclimate effects which may induce errors. Third, there are other important factors that can influence the NV availability. For example, humidity can affect the building occupant’s perception, changing their sense of a comfortable temperature range (Zhang et al., 2014). In addition to PM_2.5, nitrogen dioxide (NO₂) can be a major air pollution source that affects the NV usage, particularly in areas that are close to major roads (Zhang and Batterman, 2013).