Variability in Consumption and End Uses of Water for Residential Users in Logan and Providence, Utah, US

This study combined data from multiple sources (Table 1). The area of study comprised the cities of Logan and Providence in northern Utah, which have about 7,500 and 2,100 single family residential (SFR) connections, respectively. Monthly water use data was provided by the municipalities, and we collected high temporal resolution (4 s) data for 31 SFR properties, 19 in Logan and 12 in Providence. These data were a superset of the data used by Attallah et al. (2022), who used data from five of the 31 properties in the development and testing of the end use disaggregation and classification algorithm we used for our analyses (see “End Use Classification”).

Meter readings provided by the cities of Logan and Providence were collected on different days of the month, depending on the utility’s working schedule. Therefore, the volume of water used within a given month had to be estimated from two meter readings. We calculated standardized monthly water use, from the first to the last day of each month, as follows:where $V_n$ = volume of water used for a month $n$; $V_{MR1}$ = water volume from the first meter reading ($MR1$) that contains water use for month $n$; $D_{MR1}$ = number of days covered by $MR1$ (i.e., the number of days since the previous meter reading); $D_{n-MR1}$ = number of days within month $n$ to which $MR1$ applies; and $V_{MR2}$, $D_{MR2}$, and $D_{n-MR2}$ = same information for the second meter reading ($MR2$) that contains water use for month $n$. User ranking, monthly variation, and annual averages used when selecting participants to enroll in the study were derived from these standardized monthly values.

Participants were enrolled in this study using multiple methodologies. First, four households were enrolled by word of mouth to deploy and test data collection hardware and software (Bastidas Pacheco et al. 2020, 2021b). The longest data records were from these users. Second, we invited users based on their annual average water use (computed from monthly records) in an attempt to create a sample with participants from different water consumption levels so we would have representatives ranking in the lower, middle, and higher end of water consumption. Prospective participants were sent a letter in the mail inviting them to participate. Of 200 letters sent, 11 participants responded positively and enrolled. Given the low response rate to the mailed letters, an additional 16 participants were recruited and enrolled by word of mouth and targeted invitations. Originally, we intended to enroll 50 participants, but due to public health conditions associated with the COVID-19 pandemic, additional participation was limited.

Participation in this study was voluntary. Residents agreed to participate but did not necessarily know when data were being collected. In all cases, we conducted multiple data collection periods for each household (referred to as sites in this study). During enrollment, information was collected on the water fixtures in each home, the age of appliances, common times of the day for irrigation, and typical timing for the use of clothes washers. High temporal resolution data were collected, and a small set of short duration events were registered (toilet flushes, opened and closed showerheads, faucets, bathtub faucets). The information and event data collected during enrollment were collected to support and evaluate the accuracy of the end use events generated. Study sample household characteristics ($n=31$) are reported in Table 2, including length of the data record, number of occupants, irrigable area, building area, irrigation mode, volumetric pulse resolution of the meter, and average annual water use. The information for each site was obtained through different sources: (1) the survey conducted during enrollment, (2) publicly available data from the county, (3) analysis of the monthly water use records provided by each city, and (4) geographic information systems (GIS) analysis of high resolution imagery for Utah available from the Utah Geospatial Resource Center (UGRC 2021).

High temporal resolution water use data for all sites were collected using the Cyberinfrastructure for Intelligent Water Supply (CIWS)-Datalogger (Bastidas Pacheco et al. 2020) or the CIWS-Node Datalogger (Attallah et al. 2021), which were attached to the existing meters at each site. These external dataloggers measured the magnetic field around magnetically-driven residential water meters and counted peaks in the magnetic field associated with the movement of the measurement element within the meter. They registered peaks as pulses that represented a fixed volume of water passing through the meter. The volumetric pulse resolution ($\mathrm{L}/\mathrm{pulse}$) used in this study was determined in the laboratory (Bastidas Pacheco et al. 2020) and was used for all meters of the same size and brand found in the field deployments. Other studies self-calibrated this parameter for each meter when volumes (from the meter and datalogger) did not match (DeOreo et al. 2016), resulting in accuracies that are not directly comparable with ours. While we did not calibrate pulse resolution in this study, instead choosing to use only data that passed the quality control (QC) procedure described subsequently, our field data logs provided the volumes recorded by the meters’ registers and the raw pulse data we collected so that calibration methods could be applied to this data, if warranted, for other studies.

Data were collected over a period of three years before and during the COVID-19 pandemic. We recorded the number of residents in each home during enrollment but did not collect any information related to the participants’ schedules or employment status, nor did we assess any changes in these parameters due to the COVID-19 pandemic. We did not anticipate the COVID-19 pandemic; therefore, collecting data that would allow us to fully assess the impact of the COVID-19 pandemic on participants’ schedules was not part of the study design. We were limited in our ability to modify the institutional review board (IRB) protocol governing this study and were also constrained by a complete pause of all human subject research implemented by our institution early in the pandemic. While we did not assess the impact of the COVID-19 pandemic on participants of this study, recent studies evaluating the impact of the COVID-19 pandemic on water demand suggested that residential water use increased, and some nonresidential use (e.g., bars, restaurants, hotels, schools) decreased when stay-at-home orders were issued (Cooley et al. 2020; Menneer et al. 2021; Cahill et al. 2021; Lüdtke et al. 2021). Other studies found changes in residential diurnal water use patterns resulting from COVID-19 lockdowns (Alvisi et al. 2021; Balacco et al. 2020). The COVID-19 pandemic changed the way water was consumed (Berglund et al. 2021), and it is likely that these changes were evident in some of the data we collected, but such changes were not specifically analyzed.

We collected at least two weeks of data during months when irrigation was expected to occur (referred to as summer months, including May through October), and two weeks during the rest of the months (referred to as winter months) at each site. In December, January, and February, access to meter pits was restricted by the municipalities due to cold temperatures, resulting in shorter records for those months. Log files included within the LogFiles folder in the HydroShare data repository (Bastidas Pacheco et al. 2021a) contain information about the data collection periods at each site, including the exact start and end time of each period, the volume registered by the meter’s register and by our datalogger for each period, the percent error in volume for each period, the number of expected data values (computed using the start and end times of each data collection period, assuming one value was collected every 4 s), the number of recorded data values, the percent error in the number of values logged, and an indicator of whether outdoor water use was expected or not. The HydroShare resource also contains the high temporal resolution data for each site. The data were managed using cyberinfrastructure described in Bastidas Pacheco et al. (2021b). The data management process involved collection and processing of raw 4-s resolution data, QC to ensure validity of the data, and storage in a centralized database for analysis. QC was conducted using the following procedure:

Analysts considered all the aforementioned elements in accepting or rejecting the raw data collected. Only data that passed these QC checks were used.

In many past studies that have analyzed end uses of water, a single device measured total water use for a site, and the data were later disaggregated and classified (Al-Kofahi et al. 2012; Beal and Stewart 2011; DeOreo et al. 2011, 2016; Mayer et al. 1999, 2020; Otaki et al. 2011; Roberts 2005). Less commonly, water use for each individual fixture was measured (Kofinas et al. 2018; Di Mauro et al. 2019). We adopted the first approach and used a single device (Attallah et al. 2021; Bastidas Pacheco et al. 2020) to measure total water use; data were disaggregated using an open source algorithm described in Attallah et al. (2022). In summary, the disaggregation and classification process worked in the following way: (1) an algorithm filtered the raw data using a low-pass filter, facilitating the identification of single and overlapping events; (2) overlapping events were separated into single events using an iterative splitting process; (3) several features (e.g., average flow rate, mode flow rate, duration) were calculated for each event; and (4) the events were classified using a combination of clustering to identify atypical or outlier events that were later labeled as unknown and a semisupervised machine learning methodology to assign labels to the remaining events (Attallah et al. 2022). The machine learning model used a random forest classifier (Liaw and Wiener 2002) trained using a set of events manually labeled by a resident at one of our data collection sites to classify new events for individual residential homes.

Using the disaggregation and classification tool described by Attallah et al. (2022), water use was classified among the following end uses—irrigation, faucet, shower, toilet, clothes washer, bathtub, and unknown—using features of each event (mode, average, root-mean square, and peak flow rate, duration, and volume). A detailed description of the end use classification process, included to ensure this paper stands alone, is presented in Appendix I.

The time period for which data were collected at a site influenced the amount of outdoor water use captured. In order to obtain an estimate of outdoor irrigation efficiency comparable across different time periods of the year, the landscape irrigation ratio (LIR) (Glenn et al. 2015) was calculated at weekly intervals for each site. The LIR is defined as the ratio between landscape water use and landscape water needs [Eq. (2)]

Landscape water needs were determined for each site based on a water budget [Eq. (3)], similar to what was used in Glenn et al. (2015):where $K_c$ = crop coefficient; and $ET{o}_i$ and $P_i$ = reference evapotranspiration and precipitation for a given week $i$, respectively (in mm). Daily rainfall data and estimates of evapotranspiration from the Utah State University (USU) Environmental Observatory weather station were used to estimate the landscape water need for all properties in Logan, and data from the Evans Farm weather station was used for properties in Providence (Bastidas Pacheco and Horsburgh 2022). The crop coefficient represents the ratio between reference evapotranspiration and actual crop evapotranspiration (Doorenbos and Pruitt 1977). To determine this value, we assumed a uniform turfgrass surface for all sites and used $K_c=0.8$, similar to what was done in Endter-Wada et al. (2008). Typically, residential landscapes are composed of turfgrass and trees immersed in a turfgrass landscape (Kjelgren et al. 2000). However, there is limited information about crop coefficients for turfgrass (Romero and Dukes 2015) or landscapes with multiple plant species (White et al. 2004). Therefore, the actual $K_c$ value for each site was most likely lower than 0.8.

Landscape water use for a given week $i$ (in m) was computed for each site using Eq. (4):where $\mathrm{weekly}\mathrm{outdoor}{\mathrm{volume}}_i$ = total volume of water used outdoors for week $i$ (in ${\mathrm{m}}^3$); and landscape area = area being irrigated (in ${\mathrm{m}}^2$) at each site. Landscape areas were identified and manually digitized from high resolution aerial imagery for each site, and the areas were calculated using GIS. Using LIR helps classify outdoor water use (Table 3).

Of the 31 sites enrolled in this study, participants at two sites (Sites 21 and 22) moved into newly built homes between the times when we collected winter and summer data. These sites had not yet developed their landscape when summer data were collected and so outdoor water use was not assessed for those sites. All outdoor analyses presented in this paper were for the remaining 29 sites.

To assess the performance of indoor water fixtures, we compared the characteristics (showerhead and faucet flow rates, toilet volume used per flush) of existing fixtures at the 31 sites enrolled with the federal standard defined in the US Energy Policy Act of 1992 (DOE 1992) and the USEPA’s WaterSense efficient fixtures (EPA 2021) (Table 4). The Energy Policy Act of 1992 (DOE 1992) set national water efficiency standards for toilets, faucets, and showerheads and has been in effect since 1994. The EPA WaterSense program labels water fixtures using higher efficiency standards than the 1992 Energy Policy Act, aiming to achieve 20% more efficiency than average products in the same category (EPA 2021).

We divided faucet, toilet, and shower events into three categories: efficient (flow rate or volume per flush less than or equal to WaterSense specifications), compliant (flow rate or volume per flush greater than WaterSense specifications but less than or equal to federal standards), and inefficient (flow rate or volume per flush greater than federal standards). Events with small frequencies ($<5\%$) were not accounted for in the final assessment in order to reduce the impact of double toilet flushes, errors in classification, or unintended use. We did not assess the performance of clothes washers, because it would have required information about load sizes. We also did not assess bathtub events, because there was no defined criterion for an efficient bathtub event.

We examined indoor water use for sites with records longer than four weeks (18 sites, including five sites with record lengths varying between 11 and 19 weeks). For this analysis, we used data collected during summer and winter months and removed irrigation events to focus on indoor water use. We quantified the total volume used for indoor water use, the distribution across end uses, and hourly aggregated water use estimations at the weekly level for each site. In addition, we quantified differences in the mean volumes (for faucet, shower, and bathtub events) and frequency of end use events (for all end uses) in winter versus summer months using Student’s $t$-test (Student 1908) for sites that had at least 4 weeks of data during summer and winter months (10 sites).

Our analysis of indoor water use frequency and end use distribution showed the following:

To address our first research question, we used the per capita daily average indoor water consumption to rank sites as low ($<33$rd percentile), medium (33rd–66th percentile), or high ($>66$th percentile) water users depending on their percentile ranking of per capita daily average indoor water consumption. Fig. 2 shows the average volume per event, the number of events per capita, and the per capita average water use for all groups across end uses. There was less than a 13% difference in the average volume used for all events across the three groups. A Kruskal-Wallis test (Kruskal and Wallis 1952) with $p>0.05$ showed that these differences were not statistically significant. Fig. 2 shows that high consumption sites had larger numbers of events per capita per day across all end uses. A Kruskal-Wallis test (Kruskal and Wallis 1952) with $p<0.05$ showed that the differences in frequency across all end uses were significant. These results indicate that event frequency, which is an indicator of behavior, had the largest influence on per capita daily water consumption at the group level.

The distribution among end uses remained relatively constant across consumption levels. As the largest indoor water use category, shower events accounted for 30.3%, 30.8%, or 31.6% of total indoor per capita daily water use (for low, medium, and high consumption sites, respectively). Toilet events accounted for 21.8%, 20.8%, or 23.8% of total indoor volume across the three levels (low, medium, and high, respectively). Faucets accounted for 17% of total indoor water use across all categories.

The analysis of indoor water use timing found the following:

Hourly and daily aggregated values were considered to further address our first research question and assess indoor water use timing and its variations across users of different consumption levels. Indoor water use timing is behavioral and is determined by personal preferences and schedules. Fig. 3 shows the hourly distribution of indoor water use across all users. Some sites had distinctive periods of water consumption during the day, while other sites had relatively similar water use throughout most of the day. Water use was larger in the morning than in the afternoon at most sites. The bimodal distribution with peaks in the morning and in the evening observed in the past (Buchberger and Wells 1996) was visible only at a few sites (e.g., Sites 6, 20, and 29). Timing patterns were independent of water consumption levels. Sites within the low, medium, and high consumption categories followed different patterns indistinctively.

Participants consumed, on average, 21% more water on weekend days (Saturday and Sunday) than on weekdays (Monday to Friday). High consumption sites used 15.7% more water on weekend days compared with weekdays, while medium and low users used 19.6% and 28.9% more water on weekend days than on weekdays, respectively. Smaller increases in weekend versus weekday average per capita daily volume likely indicate longer participant presence at home on weekdays. This partially explains the results observed in Fig. 2. In general, the differences in the observed hourly and daily patterns were most likely dictated to a large degree by the heterogeneity in the schedules of the occupants.

The analysis of outdoor water use found the following:

To address outdoor water use for our first research question, we used a combined 278 weeks of data collected between May and October at the 29 sites where outdoor water use was analyzed, recording 4,533,939 L of water use during these months. Approximately 83% of this volume was used for outdoor irrigation. Outdoor water use is largely driven by personal preferences, but in some instances may be required by homeowner’s associations. The volume used may also be impacted by the type of system used for irrigation (i.e., hose, sprinkler system, automated timer, smart weather controller, soil sensors). While the level of technology used for irrigation is a personal preference, each type of system has a potential technological impact related to device performance. In our sample, the eight sites that irrigated using a hose (Sites 3, 6, 7, 8, 10, 16, 17, and 18—some with an automated timer, others manually) ranked below the 40th percentile for annual water use in Logan City (see Appendix II). The rest of the participants used sprinkler systems with automated controllers, and 88% of those sites ranked above the 40th percentile. These results are similar to results from past studies, which have found a strong correlation between automated sprinkler systems and higher water use (DeOreo et al. 2016; Endter-Wada et al. 2008; Mayer et al. 1999).

Fig. 4(b) shows the average weekly outdoor volume for each site. Sites 27 and 5 had the largest outdoor water use, consuming, on average, more than 80,000 L per week, and had the largest and the third largest landscape areas ($\mathrm{3,843}{\mathrm{m}}^2$ and $\mathrm{3,118}{\mathrm{m}}^2$, respectively). During six weeks (five in 2019 and one in 2020) the landscape irrigation needs were zero (rainfall supplied all the water the landscapes needed), and any outdoor water use that occurred was unnecessary. Given that the LIR had an undefined value during these weeks [Eq. (2)], these six weeks were not included in Fig. 4 but are addressed in the following paragraph. All hose irrigators used less than $8{\mathrm{m}}^3$ of water per week, while 80% of sites with automated sprinkler systems used more than this value (Fig. 4). The landscape areas for hose irrigators were not all smaller than the sprinkler irrigated sites (ranking 1st, 2nd, 4th, 9th, 17th, 21st, 22nd, and 24th among the 28 sites presented in Fig. 4). Using the LIR values determined for each week of irrigation [Fig. 4(c)], irrigation was excessive at Sites 2, 9, 11, and 14 during week 36 of 2019, and at Site 14 during week 38 of 2019. During week 36 of 2019, a rainfall event that supplied 97% of landscape water needs was registered by the USU Environmental Observatory station, making outdoor water use inefficient during that time period at those sites. In summary, outdoor water use was either efficient (62%) or acceptable (27%) during most the weeks in which data were collected and was excessive (5%) or inefficient (6%) the rest of the time. Given the large percentage of time that irrigation fell in the efficient or acceptable categories, it is likely that larger reductions in outdoor water use would require reductions in irrigated landscape area.

During the six weeks in which landscape water needs were zero (the LIR was undefined), we collected 33 full weeks of data across 20 sites. Fig. 5 shows the number of full weeks of data collected at each of these 20 sites and the volume (average when more than one week was available) used during weeks in which landscape irrigation needs were zero. Most sites (80%) reduced their outdoor water use between 11% and 90% in response to precipitation compared to the remainder of the weeks. Nevertheless, precipitation can occur at the end of a week after all outdoor water has been applied, and the regular weekly intervals we used for our analysis did not attempt to account for this. Furthermore, homeowners would have needed additional information (landscape water needs, rainfall data, usage from their irrigation system) to accurately respond to precipitation events. Even with the reduction in outdoor water use observed, the volume used for outdoor irrigation during weeks when landscape water needs were met by precipitation ($366{\mathrm{m}}^3$) represented a large water conservation potential without changing existing irrigated landscape area at participant sites.

Separating indoor from outdoor water use was more accurate when automated sprinkler systems were present, because the flow rates for automated irrigation events can be as much as twice the values observed for indoor events. In addition, irrigation events produced by automated irrigation controllers have similar timing, flow rates, and durations. The flow rate of irrigation events was the highest among all end uses. At five sites (Sites 2, 5, 9, 19, and 27), flow rates of irrigation events exceeded 70 litres per minute (Lpm). Irrigation events also had the longest duration among all end uses, with an average of 42.1 min across all sites. Participants with automated sprinkler systems irrigated in the early morning or late evening, which was within the recommended irrigation timing for reducing losses from evaporation; the exceptions were Sites 25 and 26, at which a few irrigation events were detected close to noon.

Individual sites were classified as low, medium, or high according to their monthly outdoor water use ranking, dividing at the 33rd and 66th percentile (computed for all SFR users by city, using the entire record of monthly data available, shown in Table 1). Monthly outdoor water use was computed as the difference between the average monthly water use during months when irrigation occurs (May through October) and the average monthly water use during months when irrigation was not expected (November through March). Using this procedure, eight sites were ranked as low, 11 sites were ranked as medium, and the remaining 10 sites were ranked as high (Fig. 6). Fig. 6(a) shows average monthly outdoor water use per landscape area, in mm. We collected two full weeks of data at Site 13 during the summer months, shown in Figs. 5 and 6, in which outdoor water use was inefficient and unnecessary. Fig. 6(b) shows average outdoor monthly water use. Hose irrigators generally ranked lower than irrigators who used sprinkler systems, and they applied less water per unit area. Broadly, sites classified as medium and high applied water at similar rates per unit of area; this result was similar to results of previous studies (DeOreo et al. 2016; Mayer et al. 1999; Endter-Wada et al. 2008). This indicates that the differences in outdoor water use observed were, in most cases, the result of the irrigation method used or the landscape area irrigated. Fig. 6 shows that Providence residents consumed more water than Logan residents (the 33rd and 66th percentiles of outdoor water use were higher).

Our review of the performance of indoor water use fixtures found the following:

To analyze the performance of water fixtures among participant sites—our second research question—we examined the features of shower, toilet, and faucet events at each of the participant sites. Our analyses focused primarily on the technological performance of fixtures (e.g., flow rates of showers and faucets and volume per flush for toilets) rather than on behavior (e.g., frequency or duration of events). As an exception, we analyzed shower durations to highlight potential opportunities for conservation related to behavior. Most of our participant sites (28) had more than one bathroom. The performance analysis was conducted on individual events and not on average characteristics in order to observe differences in the performance of different fixtures at the individual level. The results are presented in Fig. 7.

For showerhead flow rates, we found inefficient shower events at 14 sites, compliant shower events at 29 sites, and efficient shower events at all sites. Two sites (Sites 14 and 17) had only efficient events, indicating all showerheads at those sites operated below the WaterSense standard. While most sites had a mix of shower events across two or all three of the efficiency categories defined, the percentage of inefficient shower events was below 25% at all sites with three exceptions (Sites 12, 29, and 31). Shower durations are related to social norms, and, because of this, there is no consistent standard or guidance as to what shower duration is considered efficient. In these data, the distribution of shower durations was as follows: 25% of showers lasted less than 3.2 min, 25% lasted between 3.2 min and 5.87 min, 25% lasted between 5.87 min and 10 min, and the top 25% were longer than 10 min. The average duration across all sites was 7.5 min, which was similar to the 7.8 min found in the 2016 REUWS (DeOreo et al. 2016). However, the distribution of shorter showers was different, for example. The first quartile of shower durations in the 2016 REUWS was 6 min (compared to 3.2 min in our study). The different methods used for analysis, the varied households characteristics, or geographical factors could explain these differences.

The EPA allows some flexibility ($\pm 0.38\mathrm{L}=0.1\mathrm{gallon}$) in the values used for certifying toilets (EPA 2014). We added this 0.38 L to the threshold definitions shown in Table 4. We found inefficient toilet events at all sites, compliant toilet events at 26 sites, and efficient toilet events at only 8 sites. The percentage of inefficient toilet events was greater than 50% at 23 of the 31 sites (Fig. 7). The 2016 REUWS classified individual toilet events as efficient if they used less than 8.3 L. Using this higher threshold, they found that 48% of toilet events were in this efficient category, compared to 16% in 1999 (DeOreo et al. 2016). When comparing toilet events in our sample with the 8.3-L threshold, we found that about 58% of toilet events were in this range, suggesting that there has been continuous improvement in toilet performance in the last few decades. However, the percentage of efficient toilets in the 2016 REUWS varied from 38% to 61% across participant cities (DeOreo et al. 2016), indicating that geographical factors influenced the values observed. Further studies across different cities are needed to discern between temporal and geographical effects on toilet performance.

The disaggregation and classification algorithm we used was unable to separate kitchen and bathroom faucets, and, while it is possible that higher flow rate faucet events occur in the kitchen, this cannot be guaranteed. A total of 90% of the faucet events identified across all sites lasted less than 48 s and had a flow rate less than $5\mathrm{L}/\min$. Only 0.14% (1,136 occurrences) of faucet events had a flow rate larger than $8.3\mathrm{L}/\min$, indicating that faucet events exceeding the federal standard for maximum flow rate were rare. Faucets were the highest-performance category among those analyzed (Fig. 7). Faucets are most likely replaced more frequently than other water-using appliances in a home, and the growing presence of water efficient faucets may explain why this category exhibited higher performance. The 2016 REUWS found similar results in terms of faucet event flow rates, with 99% of faucet events in that study having a flow rate less than $8.7\mathrm{L}/\min$.

When analyzing water use at sites with more than four weeks of data, we observed the following:

To assess how estimates of volume, distribution across end uses, and timing of indoor use may change with longer data collection periods—our third research question—we focused on sites for which we had four or more weeks of data. Indoor water use was relatively constant across weeks at some sites (e.g., Sites 7, 12, and 22), whereas relatively drastic differences in week-to-week volumes were observed at others (e.g., Sites 2, 9, 18, and 19) (Fig. 8). In some cases, these variations occurred within the same period of the year; this can be observed by analyzing the separation of points of similar shape across sites in Fig. 8. Some of these variations may have been the result of changes in occupancy. For example, there were weeks with minimal water use at Sites 5, 14, and 15, which indicates that the occupants were likely not at home during these weeks. In addition, changes in weekly schedules and personal preferences may have affected the amount of water used.

The daily timing of water use also varied significantly between weeks. Fig. 9 shows total hourly water use week by week for Site 19, that is, the total water used within each hour summed across all the days of each week. The time of occurrence and the magnitude of peaks in water use varied significantly from week to week, and this lack of patterns in hourly and weekly water use data was observed at most of the study sites. There were weeks at this site that had one or more distinctive periods of high water use, but in other weeks water use was relatively constant across the day, indicating that the distribution of use varied over time. Fig. 10 shows weekly variations in indoor water use across end uses for the same site and the same weeks. The percent of indoor volume used for shower events varied the most from week to week, from over 40% to 16% of total water use; it ranked as the largest end use in some weeks but ranked 4th in others. The percent of indoor water use dedicated to bathtub events also exhibited changes, ranging from 6% to 20% and ranking from 2nd highest to 5th highest. The percentage of indoor water use dedicated to toilets varied between 25% and 36%. While we have generally reported use in terms of percentage of volume, the frequency with which end use events occurred also varied depending on the week.

Of the 10 sites selected for indoor winter and summer comparison, four did not have bathtub events. We compared eight (or six, when bathtub events were not present) parameters at each site between winter and summer months. Of the 72 parameters analyzed across these 10 sites, there were significant differences ($p\mathrm{value}<0.05$) in the mean of 19 (26% of the cases) of them, according to $t$-test results. In 11 cases, we observed changes in the mean frequency, and in eight cases we observed differences in the mean volume. The direction of these differences by event type is reported in Table 5 (for frequencies) and Table 6 (for volumes). The inconsistency in the differences between frequency and volume of end use events suggests that the observed differences are not generalizable. In some cases, the number of events or the volume was larger in the summer, and the opposite occurred at other sites. In addition, with the exception of faucets, no significant changes were observed at the majority of the sites analyzed.

As the number of weeks of data increased for a site, we observed only small variations in the average, mode, and peak flow rates for showerheads and faucets and in the volume per flush used in toilets. Therefore, it appears that the technological performance of indoor fixtures can be accurately assessed with short data collection periods unless a fixture is replaced. However, capturing behavioral changes in indoor water use volume and timing and developing a comprehensive representation of the distribution of indoor water use across end uses that were evident in our data requires longer data collection periods or a different study design than has been typically used. The lack of consistency in indoor water use patterns cannot be characterized using coarser resolution (e.g., monthly) data or when analyzing indoor and outdoor water use together. Other studies have pointed to similar results. For example, Rathnayaka et al. (2015) found differences in shower durations and frequencies in summer versus winter months in 117 houses across two municipalities in Australia. Suero et al. (2012), who analyzed two weeks of data before and four weeks of data after retrofitting with efficient appliances in 96 homes in the US, found differences in the frequency of use of toilets and clothes washers between the pre- and post-retrofit data sets. The seasonal differences we observed and those observed in prior studies indicate that there are seasonal and shorter-term changes in the frequency, timing, and distribution of end uses. Longer, and continuous, periods of data collection are required to characterize these types of temporal variations (Fig. 9), including changes in the distribution of water use across end uses (Fig. 10) and the seasonal component of indoor water use (Rathnayaka et al. 2015). In addition, we observed that indoor water use varies differently across sites (Fig. 8), suggesting that the record length needed to characterize indoor water use variability may be different across sites.

Collecting indoor water use data for short periods of time can generate parameters (volume, timing, and distribution across end uses) that may not be representative of water consumption at a site given that water use depends on behavioral factors in addition to fixture performance. End use level data provide a basis for evaluating and designing water demand strategies (Beal and Stewart 2014), demand and infrastructure modeling (Blokker et al. 2010), and general planning (Willis et al. 2013). Using water use estimations resulting from data that do not capture a representative sample of water use may impact the accuracy of such applications and lead to the implementation of ineffective water management strategies, under- or over-dimensioning of infrastructure, and other issues. However, defining a fixed record length that secures a complete characterization of indoor water use across multiple residential properties is infeasible using currently available data, most of which are short duration data.

Logan City reads meters once per month, and Providence City reads meters once per month from April through September. We calculated volumes for October through March for Providence by dividing the total winter volume measured (calculated using the September and May meter readings) by six, resulting in the constant values shown in Fig. 11 during those months.

Single family residential water use varied throughout the year in Logan and Providence, peaking in July with average monthly values, across all connections, per household, close to 125,000 L in Logan and 220,000 L in Providence (as shown in Fig. 11). During winter months, SFR water use remained relatively constant in Logan, with per household monthly averages just below 20,000 L and below 30,000 L in Providence. Sociodemographic variables like differences in household and landscape sizes could account for the differences observed in water use between the two cities during winter months. However, we did not collect data at the city level, and the monthly data provided by the cities did not contain information that would have allowed us to further assess these differences. Outdoor water use drives the increase in residential water use observed during summer months, constituting the largest component of residential water use. Total annual water use did not vary significantly from one year to the next during the period of data available for each city. Winter water use for Logan (Fig. 11) showed some variations that were likely due to differences in indoor water use; however, when compared with the magnitude of the annual variations, changes during winter months appear to be minimal.

In order to place our sample of households in the context of single family residential water use in their cities and of other residential water use studies, we report brief general statistics about ranking and water use. Participant sites ranked between the 4th and 95th percentiles of annual SFR water use in each city (computed from monthly meter records). Figs. 12(a and b) show the percentile rankings of annual water use for each participating site, computed for the last two years of data available in each city—2017 and 2018 for Logan, and 2018 and 2019 for Providence. Despite the combined user sampling approach used (targeted invitations, word of mouth) and the relatively small number of participating homes, the samples contained a broad range of percentile rankings and annual water use volumes. Percentile rankings were not consistent from year to year, indicating that there is significant interannual variability in water use that is not determined solely by the climatic conditions that drive outdoor use. Fig. 12(c) presents the participating sites’ per capita daily average water use for the same two years. Occupancy was registered during enrollment (2019–2021), and monthly water use data were recorded during previous years; therefore, changes in occupancy during this period were not accounted for.

Fig. 12(c) shows that our sample included users that differed from the per capita average values presented here. The average per capita daily water use among participants in this study, computed from monthly water meter data, was 695 L per capita per day (Lpcd), and the same figure computed from the high temporal resolution data we collected was 754 Lpcd. A total of 73% of our high temporal resolution data were collected in summer months, which could explain the differences between the estimations from monthly records and the high temporal resolution water use data collected. One recent estimate places per capita daily average residential water consumption in Utah at approximately 640 Lpcd (Dieter et al. 2018). However, this value was calculated by compiling data from different agencies, using coefficients in areas of the state where supply is not measured, and using population estimates that may have impacted the accuracy of this estimation (Milligan 2018)—these factors may have hidden a lot of variability within Utah. The Utah Division of Water Resources (DWR) estimated that in 2015, residential water use in Cache County, where Logan and Providence are located, was 784 Lpcd and that there were differences in water use across counties in the state (Utah DWR 2020 ).

It is well known that per capita averages, while useful for estimating total water demand at aggregated scales, provide little information about water use patterns or behavior within individual households—in particular, because outdoor water use is not dependent on the number of occupants of a house. A household with a small number of occupants and a large landscape will have a larger per capita consumption. The average daily per capita indoor water use among participants in this study was 174 Lpcd.

Table 7 shows the average per capita daily volume used for each indoor category and the percentage of indoor water use that each category represents. Short events lasting less than 4 s with a single recorded pulse (unclassified) were the most common indoor event (79.2% of all indoor events were in this category) but represent only 4.73% of average indoor water consumption. This category included leaks, very short duration events (e.g., faucets and refrigerators with ice makers), and other events that we were not able to separate or identify because they all had the same volume and duration. Unknown events included outlier events identified during clustering that we were unable to classify and represented approximately 1.51% of total indoor volume.

Showers were the largest indoor water use in our study, representing 31.2% of total indoor water use. Toilets were the second largest end use across all sites, representing 25.6% of total indoor use, although toilets were the largest water end use in 13 of the 31 homes. In contrast, the 2016 REUWS (DeOreo et al. 2016) found that toilets were the largest indoor end use, consuming 24% of indoor volume, followed by showers (19%). The South East Queensland Residential End-Use Study (SEQREUS) (Beal and Stewart 2011) conducted in Australia found that showers consumed 29.5% of indoor volume and toilets 16.5%. These results indicated that the distribution of water use across end uses is different across individual residential homes and regionally. The distribution among end uses remained relatively constant across the three consumption levels described in the main article. As the largest indoor water use category, shower events accounted for 30.3%, 30.8%, or 31.6% of total indoor per capita daily water use (for low, medium, and high consumption sites, respectively). Toilet events accounted for 21.8%, 20.8%, or 23.8% of total indoor volume across low, medium, and high consumption sites, respectively. Faucets accounted for 18% of total indoor water use across all categories.

The number of per capita faucet events, showers, and toilet flushes in our study was 23.2, 0.97, and 5, respectively. The 2016 REUWS found similar results in terms of per capita daily frequency of toilet flushes (5) and faucet events (20) but lower shower frequency (0.69) per capita per day. The 2016 REUWS used a much larger sample of homes (763 homes across nine cities in the US versus 31 homes in two neighboring cities in this study), but the average household occupancy was 2.7, which was much lower than the 3.8 in our study. The number of bathtub events per capita per day in our study was 0.12, higher than the 0.05 encountered in the 2016 REUWS. The frequency of clothes washer events in our study was considerably less than that of the 2016 REUWS. Assuming each load had 2 cycles (wash and rinse), we estimated 0.19 clothes washer events per capita per day versus 0.3 in the 2016 REUWS.

High use associated with showers can be the result of personal preferences (longer and/or more frequent showers) or the presence of less efficient fixtures (showerheads operating at higher flow rates). Fig. 13 shows the average flow rate and duration of shower events for each site. Site 31 had the largest average shower flow rate and the largest per capita daily average shower use, despite having shorter shower durations. Site 18 had the second largest daily per capita shower consumption. This site had a much lower shower flow rate but higher shower durations. Site 17 had lower durations and shower flow rate but a higher number of showers per day, ranking in third place for daily shower volume. At Site 27, median shower duration was 15 min. The average shown in Fig. 13 was increased by three events that had durations longer than 80 min at flow rates in the same range as all showers. Without additional information, it was not possible to identify if these were erroneous events (i.e., incorrectly labeled as showers), and they remained labeled as showers.

Fig. 14 shows the volume distribution of toilet flushes for all sites. Toilets are a mechanical end use (i.e., the flow rate and duration do not depend on user preferences and are expected to be similar for each flush). We observed some variability in the volumes shown in Fig. 14. Some sites had a multimodal distribution (e.g., Sites 16, 19, and 30) that was the result of having multiple toilets with different characteristics. For example, Site 19 had toilets that used approximately 8.3 and $13.2\mathrm{L}/\mathrm{flush}$. The average flush at this site used 10.8 L, but this value ranged lower or higher depending on which toilet was used more frequently. This was true for every site with a multimodal distribution.

The values at the extremes of each violin plot are typically events with flow rates similar to most toilet events but with different durations. These may be half flushes, double flushes, a toilet valve remaining open longer than normal due to flapper valve malfunction, or other misclassified uses. At some sites with multiple toilets, the distribution showed a single mode with higher variability. This was the case at Sites 20 and 23, which had 3 and 4 toilets, respectively. In these cases, it was likely that the toilets performed similarly enough that the volumes mixed, giving the appearance of a single mode distribution.

Toilet and faucet events that happened simultaneously (i.e., washing hands before the toilet tank is done refilling) were common and can be identified when examining the raw data. However, given the low flow rate of faucets, attempting to automatically separate such events tends to make the algorithm too sensitive toward classifying single events as overlapping. For the purposes of this study, we decided to not separate toilet events from simultaneous short duration faucet events, which meant that these events were lumped in with toilet events.

Faucet events were the third largest category of indoor water use by volume. This category included kitchen and bathroom faucets, hose bibs, and other short duration and low flow rate events that did not fit other categories. Faucet events were the second most common events behind the unclassified category, which can also include very short duration faucet events. The characteristics of dishwasher cycles were indistinguishable from faucet events. Therefore, they were labeled as and lumped in with faucet events. Future improvements of our classification method could include identification of cycles for dishwashers and clothes washer events, as has been described in other methods (Nguyen et al. 2018). Most faucet events were short (93% lasted less than 1 min) and low volume (80% used less than 2 L). Fig. 15 shows the flow rates and duration of faucet events across all sites. Sites 5, 18, and 23 had the largest faucet event durations. Site 9 had the smallest flow rate variability for faucet events; 91% of the faucet events at this site had a flow rate less than $2.2\mathrm{L}/\min$. Sites 8 and 31 had the highest median faucet flow rates among all participants.

Despite clothes washers being a mechanical end use, the identification and classification of clothes washer events is not straightforward. Clothes washers can have different configurations such that the volume of water used can vary depending on the load size and cycle selected. In addition, the flow rate can vary depending on the temperature of water used—hot, cold, or both. According to DeOreo et al. (2016), the average per load volume decreased from 155 L in 1996 to 117 L in 2016, and this change was attributed to the adoption of more efficient appliances. However, it is not clear how clothes washer cycles were grouped in these studies. Other methods used to classify end uses of water use a time span of two hours to aggregate and identify clothes washing cycles (Nguyen et al. 2018), adding all events with clothes washer characteristics in this time span to a single load. How clothes washer events are aggregated to loads has a significant impact on the statistics reported.

For this study, we did not aggregate clothes washer cycles (i.e., we identified individual clothes washer cycles but did not aggregate them into multicycle loads). The average water consumption per clothes washer cycle was 60 L. If we assume a load consists of one wash and one rinse cycle, the average (120 L) is close to the 117 L reported by DeOreo et al. (2016). Fig. 16 shows the volume distribution of clothes washer events for all sites. We observed large variability in the volume used in clothes washer events at most sites, which we attributed to different load sizes and clothes washer settings. Site 24 had two clothes washers; all other sites had a single clothes washer. Sites with large numbers of events with similar volumes (e.g., Sites 19, 20, and 29) were likely doing laundry without constantly modifying appliance settings. Site 32 had the highest volume per clothes washer event among all sites (Fig. 16) and ranked fifth overall for average per capita daily clothes washer use.

Bathtub events can use up to 265 L of water, and the time at which the drain is plugged (before or after the temperature is adjusted) can increase this volume (EPA 2021). Bathtub events were not found at 11 sites (35% of participating homes). The average volume used in bathtub events among the remaining participants was 77 L, similar to the 76 L per bath found in the 2016 REUWS study (DeOreo et al. 2016). The average flow rate at which bathtubs were filled was $14.5\mathrm{L}/\min$, and the average duration of these events was 5.6 min.