Water End-Use Disaggregation for Six Nonresidential Facilities in Logan, Utah

We used MySQL Workbench version 6.3.7 to design a database schema to store the monthly water-use, business licensing, and stormwater data provided by City of Logan. We then imported the three data sets into the relational database (Atallah and Rosenberg 2020). The three data sets were linked by the physical address attribute, which was the common attribute among the three data sets.

To identify patterns in the monthly commercial and industrial (CI) data, we applied Shapiro–Wilk normality and ANOVA tests to characterize the frequency distribution of the monthly billing data and identify which facility types significantly explained Logan water use (Atallah 2018) (Table 2). Because many parametric statistical tests rely upon the assumption of normality, we used the Shapiro–Wilk normality test to verify or refute the null hypothesis that the Logan CII data are normally distributed. The frequency distribution of water-use data showed that the majority of users consumed less than $\mathrm{7,600}\mathrm{L}/\mathrm{day}$. The $p$-value for the Shapiro–Wilk test was $2.2\times {10}^{-16}$, which is less than 0.05 and indicates that the water-use data were not normally distributed (Atallah 2018). This skewed distribution of these CII water-use records followed similar skewed findings in several residential water-use studies (Abdallah and Rosenberg 2014; Suero et al. 2012; Rosenberg et al. 2008). Because the distribution of water use did not fit a normal distribution, we log-transformed the data and again performed the Shapiro–Wilk test. The test $p$-value was 0.5, which was high enough ($p\text{-}\text{value}>0.05$) to accept the hypothesis of normality of the log-transformed data.

We also identified which lot size, location, and facility-type factors explained water use. NAICS codes classify facility types based on the particular product or service they supply and place them into the appropriate group among 20 different facility types (manufacturing, assisted-care, construction, education, and so forth). Because facility type is categorical, we developed a generalized linear model for the Logan city data. The $p$-values for the manufacturing and the health-care and social-assistance sectors were less than 0.05, which suggests that these two sectors explain period-average water use (Table 2). For other CII categories such as retail trade, wholesale trade, and information, $p$-values were greater than 0.05, which suggests these facility types do not explain water use. Lot size and location did not explain water use.

We also estimated the average daily water use for each CII facility as the total water use for a facility divided by the number of billing-period days. We then ranked facilities according to their average daily water use. Rankings indicated that the top CII categorized water users in the City of Logan were manufacturing and health-care facilities. Therefore, we focused recruitment for further high-frequency monitoring on those facilities.

From the 472 CII facilities, we identified the top 30 as mostly manufacturing and assisted-care facilities, and recruited 6 facilities to participate in monitoring their water use at high frequency (every 5 s). Four manufacturing facilities, ranked 1, 2, 3, and 5 in daily water use, produced circuit boards, metals, or printed materials. They had unique features, including $24/7$ working hours, up to four different water meters serving a single facility, and up to $\mathrm{18,600}{\mathrm{m}}^2$ of irrigated turf grass. The four manufacturing facilities provided replication within this sector. The two assisted-care facilities ranked 9 and 10 in average daily water use, and each had more than 70 residents and over 300 different fixtures in each facility (e.g., faucets, toilets, urinals, sprayers, showers, storage tanks, and commercial cloth washers). The assisted-care facilities had very different types of water use than the manufacturing facilities, and their water use somewhat resembled that of the multifamily residential use sector (except for large industrial-scale kitchens and laundromats). Several components of assisted-care facility water use—toilets, showers, and faucets—might compare directly with that of prior residential studies (DeOreo et al. 2016) or manufacturing facilities. To our knowledge, each facility’s sole water source was metered city water.

A representative for each facility signed an informational letter that described the study and the data-sharing provisions. These provisions included permission to place an anonymized version of the high-frequency data we collected within a data repository for permanent publication and potential reuse (Atallah and Rosenberg 2020). To respect the anonymity of the facilities as required in the data-sharing agreement, we refer to different facilities by an ID (Manufacturer 1, Manufacturer 2, Assisted-Care 5, and so forth).

The 10 water meters serving the 6 participating facilities varied in size from 25 to 102 mm (1.5 to 4 in.). Two participating facilities had 102 mm (4-in.) compound meters that required two registers (to measure low and high flows). On the low end, all meters can measure flow rates as low as $2\mathrm{L}/\min$ regardless of meter size. Compound meters consist of a combination of an AWWA Class II turbine meter for measuring high rates ($>12\mathrm{L}/\min$) of flow and a nutating disc–type positive displacement meter for measuring low rates ($<2\mathrm{L}/\min$). The two meters are enclosed in a single main case. Any value between 2 and $3\mathrm{L}/\min$ can be measured on either head. On the high end, 25-mm (1.5-in.), 2-, and 102 mm (4-in.)  meters—both single and compound—can measure up to 600, 1,700, and $\mathrm{3,800}\mathrm{L}/\min$, respectively. An automatic valve directs flows through the disc meter at low flow rates and through the turbine meter at high flow rates. At high flow rates, the automatic valve also serves to restrict the flow through the disc meter to minimize wear (Neptune 2016). The meters have an accuracy of $\pm 1.5\%$. The 25-mm (1.5-in.), 2-, and 102 mm (4-in.) service lines suppling the facilities were larger than the 0.75 or 1-in. lines that typically supply residential properties.

The 5-s data were stored in the internal memory of the pulse counters and collected on site. At 5-s frequency, the storage capacity was 31 days. When the memory filled, the pulse counter stopped recording. To protect the MagdeTech pulse (Warner, New Hampshire) counters from moisture damage, we placed them inside weatherproof enclosures that we then zip-tied to the underside of the manhole that provided access to the water meter and meter pit. For data collection, approximately monthly site visits were conducted to retrieve the data logs. We used MadgeTech software to export the 5-s data from the pulse counters to a Microsoft Excel workbook (Fig. 2). Visits depended on winter weather conditions and the schedule of city staff who provided access to manholes. Despite the weather-proof enclosures, on several visits we found that moisture had entered the datalogger compartment and we could not download data.

Because the dataloggers produced counts of pulses by each meter head every 5 s along with the respective timestamp, we calculated the volume of water in liters by multiplying the count of pulses by the volume of water per pulse ($3.785\mathrm{L}/\mathrm{pulse}$). Finally, we designed a simple MySQL database to hold the high-frequency records. The database had four tables: (1) Participant, which contained personal information of the participating facilities, such as the name and the address; (2) Meter, which contained attributes of the meters connected to different participating facilities, such as the size and the type of the meter; (3) Pulse counter, which contained attributes of the pulse counters, such as the serial number of each device and its status; and (4) WaterUse, which contained the data value table that held the high-frequency records. Foreign keys were used to connect the four tables. The WaterUse table had more than 25 million records of water use measured at 5-s intervals.

To check the compatibility of the meter, register, and pulse-counter components when assembled together, we tested the metering components at the Utah Water Research Laboratory (UWRL) in May 2016 using different pipe sizes, meter sizes, and variable flow rates (Fig. 3). We also used the tests to determine what pulse resolution (i.e., volume per pulse) to use with the pulse counter. The register options were 0.38, 3.8, 38, 380 L/pulse (0.1, 1, 10, 100, and so forth gal./pulse). The Innov8 register’s sensor transmits the actual turns of a magnet inside the meter to a microcontroller, which displays the magnitude of water use on the register’s display and generates output pulses according to a user-configurable pulse resolution. When the flow rate is too high, we suspected that the meter’s sensor and/or pulse output module might become overwhelmed by the number of rotations of the magnet and underreport the actual volume of water flowing through the meter. Based on the monthly billing data records of the participating facilities, we estimated that flow rates varied from 76 to $265\mathrm{L}/\min$. In our lab tests we increased flow rates to $380\mathrm{L}/\min$.

Laboratory test results showed that, at flow rates below $95\mathrm{L}/\min$, the pulse output modules and pulse counters could accurately record flow using a pulse resolution of $0.3785\mathrm{L}/\mathrm{pulse}$. However, for flow rates above $95\mathrm{L}/\min$, we discovered that a pulse resolution setting of $0.3785\mathrm{L}/\mathrm{pulse}$ massively underestimated the flow rate through the meter. Because of this limitation, and because we did not expect flow rates for participating facilities to exceed $265\mathrm{L}/\min$, we used a resolution of $3.785\mathrm{L}/\mathrm{pulse}$ (1 gal.) for the study. Thus, we recorded data using the finest pulse resolution that could be captured accurately using the pulse output modules and pulse counters across the flow rates we expected.

We conducted a walk-through of each participating facility with an employee. During the walk-through, we identified the technology, demographic, and behavioral factors that affected water use at each facility, such as size of the facility (number of employees/residents and landscape area), different types of water uses inside the facility, irrigation behavior, working routines, and questions facility staff had about water use and conservation that the study could help answer.

The event disaggregation algorithm makes use of unique data stream features created by the 5-s data collection system. The system works like a tipping-bucket rain gauge in which one electronic pulse is recorded in the 5-s period when the meter’s register fills and finally measures 3.875 L of water use [Fig. 4(a)]. Thus, we separated high-flow (1) industrial, (2) outdoor, and (3) humidifying events from lower-flow (4) indoor shower and (5) indoor toilet events. The high-flow events were characterized by (1) a start time when pulse counts per 5-s period followed a zero count, (2) sustained pulse counts of 1 or more for each 5-s period, and (3) an end time when pulse counts transitioned back to 0. A sustained count of $1\mathrm{pulse}/5\mathrm{s}$ translates to a flowrate of $45\mathrm{L}/\min$. Lower-flow-rate shower events were characterized by a pulse count of 1 followed by multiple 5-s periods with zero pulse counts, then another 5-s period with 1 pulse count. This pattern of several periods of zero pulse counts followed by 1 period of 1 pulse count repeated for the duration of the shower event [Fig. 4(b)]. The lag time between pulse counts depended on the flow rate: shorter lag times indicated a higher shower flow rate, whereas a longer lag time indicated a lower flow rate. For example, if the lag time between single pulse counts is 30 s (six 5-s periods), then it takes 30 s for 3.875 L to accumulate in the register. Thus, one pulse (3.785 L) every 30 s corresponds to a flow rate of $7.8\mathrm{L}/\min$ even though the data system records all the water use in the final 5 s of the 30-s interval. For shower events, lag times between periods with 1 pulse count varied from 10 to 45 s. Toilet events were characterized by a count of 1 pulse for one or possibly two consecutive 5-s periods.

To further characterize and differentiate event types, we identified the event start time, end time, duration, peak flow (calculated from the 5-s period with largest number of pulses), and event volume (sum of all pulse counts between start and end time multiplied by the pulse resolution). The summation of pulses between the start and end times also served as a low-pass data filter. For high-flow-rate events, this summation helps to smooth portions of the 5-s data stream in which the number of pulses per 5 s varies between two adjacent integer values [e.g., 4 and 5 pulses, corresponding to 16 and 20 L in the irrigation event in Fig. 4(b)]. For events with lower flow rates, the summation of pulses also serves to filter the data stream when the lag time between 5-s periods with single pulse counts varies. The classification for the larger-flow event types is described further subsequently.

Outdoor events were identified byime of use (most occurred after midnight and in the early morning), volumes of at least $\mathrm{3,785}\mathrm{L}/\mathrm{event}$ (1,000 gal.), and durations of at least 20 min. For these events, we also investigated the ratio of applied water to landscape water need for two businesses for which we monitored water use during the irrigation season. This landscape irrigation ratio (LIR) was the water volume applied to the landscape [$V_{\mathrm{event}}$ ($\mathrm{L}/\mathrm{day}$)] divided by the water volume needed by the landscape (Utah Climate Center 2018)where ${\mathit{ET}}_o$ = reference evapotranspiration ($\mathrm{in.}/\mathrm{day}$) values for which were retrieved from the Utah Climate Center (2016); PF = plant factor, i.e., fraction of ${\mathit{ET}}_o$ needed by plants (turfgrass = 1.0); SM = area to be irrigated (${\mathrm{m}}^2$), which was retrieved from the Cache County parcel and zoning interactive (Official Site of Cache County, Utah – Home 2016); and 0.136 is a conversion factor to report output in metric units (liters). Based on LIR, irrigation efficiency can be inferred. An irrigation system is considered efficient if the calculated LIR is less than 1. If LIR is more than 3, water is applied excessively in irrigation. Acceptable irrigation efficiencies have LIR values between 1 and 3.

Industrial events were identified by volumes of at least 645 L and durations of at least 12 min (values were identified during walk-throughs). One facility used a model water distribution network to test the durability of valves, pumps, fittings, and sensors they made. They pumped water into the network at variable flow rates for about 1 h. Network test events had volumes of more than 3,785 L and flow rates of about $114\mathrm{L}/\min$. Two manufacturing facilities printed circuit boards. They used water in pressurized jets to cut the boards into different shapes and sizes, rinse and wash boards between application steps, and cool boards after processing. During the walk-through, one manufacturing facility did not report any industrial water uses.

Humidifying events in two facilities were defined by the day of use and the flow rate. Humidification systems force moisture in the form of mist into the facility’s indoor air to maintain the humidity in the facility. Humidifying events occurred on days when relative humidity values were less than 35%. The water used in each humidification event varied from 114 to $227\mathrm{L}/\mathrm{event}$. The humidifier at Manufacturing Facility 3 was automated, whereas the humidifier at Manufacturing Facility 1 was manual.

The end-use disaggregation algorithm outputs a comma-separated values (CSV) file with the classified events. The attributes of the generated CSV file are the volume (liters), duration (minutes), flow rate (liters per minute), start time, and end time. The algorithm was coded in Visual Studio version 16.8.4 (C#) by Atallah and Rosenberg (2020).

It was difficult to classify faucet events because their short duration and small volume often were less than the 3.8-L volumetric resolution of the pulse count system. Water use for faucet events typically was assigned to the next event. It also was difficult to characterize individual fixtures because each facility had a large number of different fixtures inside. Water uses were different than in residential homes, and measurements were made on the bulk flow through one or more meters serving each facility. It also was difficult to classify some events with multiple subevents (e.g., some industrial processes, and industrial clothes washers and dishwashers) and simultaneous flows at multiple fixtures (overlapped events from fixtures of different types). We were able to identify overlapped events produced from the same fixture type (e.g., two concurrent showers). For overlapped shower events, the lag time decreased between 5-s periods with a single pulse.

The event analysis was divided into two components to answer the three major research questions listed in section “Introduction.” The first component involved the identification of the average water use from each water use category per facility. Average water use per facility for a category was calculated as the total volume of water used by fixtures of each category divided by the monitoring period for each business. The average water use per category estimates the similarities and differences between facilities with comparable fixtures. The second component of our analysis consisted of calculating a facility per capita water use and comparing that use with the standard residential per capita water use of the City of Logan. Our facility contacts provided a daily patient census (assisted-care facilities) and number of employees (manufacturing facilities) for the months of December 2017 and January 2018. Facility per capita water use was calculated as the total volume of indoor water use divided by the number of people working or residing within the facility.

In the six facilities that we studied, outdoor use events varied from 3,785 to $\mathrm{22,700}\mathrm{L}/\mathrm{event}$. The LIR identified for each facility ranged between 1.44 and 5. These ratios indicate that facilities used at least 1.5 times more water than the amount actually needed for their outdoor usage. Outdoor water use at Assisted-Care Facility 2 was 5 times more than actually needed in the month of June. Results also showed some outdoor water use activity for Manufacturing Facility 3 in October, even though it is recommended that the irrigation system be turned off by the end of September because the Utah growing season ends by October.

This high use mainly resulted from facilities irrigating landscaping for long durations, which exceeded 20 h at high flow rates for some events. Proper management of sprinkler irrigation systems can boost irrigation efficiencies greatly and reduce total water use. This can be maintained by matching the application rate and duration to the actual water needs of the landscape. Irrigation results showed that at least 75% of irrigation events lasted less than 120 min. Manufacturing Facility 3 had a couple of irrigation events that lasted more than 17 h. Furthermore, irrigation events for all the facilities except Manufacturing Facility 3 had a median of $\mathrm{56,780}\mathrm{L}/\mathrm{event}$. Manufacturing Facility 3 had volumes that exceeded $\mathrm{75,708}\mathrm{L}/\mathrm{event}$. The long durations and the large volumes of water associated with irrigation for Manufacturing Facility 3 were because the facility has more than $\mathrm{18,600}{\mathrm{m}}^2$ of irrigated area. Flow-rate values for irrigation events at all the facilities varied between 75 and $265\mathrm{L}/\min$.

Industrial water use for Manufacturing Facilities 2 and 4 varied from 1,135 to $\mathrm{1,514}\mathrm{L}/\mathrm{event}$. Pressure-wash processes were the main industrial water use inside those facilities. In contrast, water volumes for the humidifying events for the two manufacturing facilities that had humidifiers depended on the size of the facility, the humidification method used, and the capacity of the steam humidification pipelines inside each facility.

One hundred thirty-one shower events were compiled from one manufacturing facility over a period of 6 months, and 5,845 shower events were identified from the two assisted-living homes over a 4-month period. For the manufacturing facility, 75% of shower events lasted less than 8 min, whereas for the assisted-care facilities, 75% of their shower events lasted less than 11 min, with many outliers at durations of more than 15 min. The longest shower duration of 24 min was recorded at one of the assisted-care facilities. According to the data, a 6-min shower used about 75 L of water, whereas an 11-min shower used approximately 125 L of water. According to DeOreo et al. (2016), an average shower in the US uses approximately 67 L. In the manufacturing facility, 50% of the showers had flow rates that varied from 10 to $14\mathrm{L}/\min$, which exceeds the maximum flow rate of $9.5\mathrm{L}/\min$ set by the US Energy Policy Act. On the other hand, 50% of shower events at the assisted-care facilities had flow rates that complied with the standards, varying from 7.5 to $11.4\mathrm{L}/\min$. A few shower events in the assisted-care facilities had flow rates of more than $21\mathrm{L}/\min$. For the assisted-care facilities, shower events were distributed throughout the day, whereas manufacturing facility shower events were either before 8:00 a.m. or after 3:00 p.m.

The remaining events were unclassified and had flow rates, volumes, and durations that varied between 45 and $60\mathrm{L}/\min$, 26.5 and 110 L, and 2 and 7.5 min, respectively. These unclassified events could include cycles of industrial clothes washers and dishwashers (e.g., in the assisted-care facilities) or overlapped events that we were unable to separate. Overall, we were not able to classify approximately 10% of all events. These amounted to approximately 454,250 L across all facilities and all times.

We estimated indoor per capita water use in December and January for the two assisted-care facilities and two manufacturing facilities to compare water use among the facilities and with residential users (Fig. 5). Liter per capita day (LPCD) estimates from the assisted-living facilities can be directly compared with Utah’s indoor residential rates because both users share many of the same water-use characteristics. Indoor per capita water use in the assisted-living facilities varied from 48 to 105 LPCD, which ranges from 14 LPCD less than Utah’s overall rate to more than 2 times Utah’s overall LCPD (Fig. 5). For the manufacturing facilities, indoor per capita water use varied from 0 to 151 LPCD and averaged 30 and 56.8 LPCD for Manufacturing Facilities 2 and 3, respectively. These values are commensurate with Utah’s industrial indoor use of 41.6 LPCD. These results provide an estimate of daily indoor water use for workers outside their homes. Zero values in industrial facilities reflected weekend days when there was no activity or water use. The overall LPCD values for the assisted-living homes were much higher than those of the manufacturing facilities, and more resembled residential use. Thus, indoor per capita water use is very contextual (Fig. 5). A person’s total indoor water use per day could be up to 35% higher than prior published values for residential users.

Although the LPCD method is simple, it has limitations. The method focuses on average water use rates in each category of use. This simplification ignores trends, variability among users or by a single user, changes in water use due to conservation, user type, or working hours. The accuracy of the method depends on the water use, the underlying activity assumed to drive water use, and the estimate of the number of persons.

Using the 5-min data, we investigated the temporal water usage patterns for the six participating facilities by quantifying the average hourly water use over the entire period of data collection. Manufacturing facilities showed no variation in winter water use throughout different hours of the day, whereas assisted-living homes had two peaks in their hourly water use (Atallah 2018). One peak was in the early morning, and the other was in the evening, both of which were driven by shower events. All businesses had summer water peak demands from midnight to early morning. These peaks were up to 3 times the daytime use, and were driven by outdoor water-use activities.

The peak hour demand for each participating facility ranged from 114 to $\mathrm{7,950}\mathrm{L}/\mathrm{h}$. Facilities exhibited diurnal patterns with peak use in dawn hours and slight variation during other times of the day. The peak demand was driven by either irrigation water use or industrial water use. Daytime/nighttime and indoor/outdoor patterns of water use for the six facilities we studied were similar to residential patterns. However, all six facilities used much larger volumes than did residential users.

Based on expected flow rates, the finest volumetric pulse resolution we could use for accurate flow rate estimates was $3.785\mathrm{L}/\mathrm{pulse}$. This volumetric resolution prevented us from identifying short-duration and/or low-volume faucet events. These uses typically were assigned to the next event.

Maintaining regular site visits to collect 5-s water-use data was a challenge due to the city’s tight schedule. The irregular data collection resulted in gaps within the 11-month data collection window. Several months of data from some facilities and meters also were lost when moisture infiltrated some pulse counters, despite the fact that they were placed inside weatherproof enclosures.

The inherent difficulties in collecting, training, and testing data from each event type meant that we were not able to identify some events. In the four manufacturing and two assisted-care facilities, multiple water end uses often are operating at the same time. When two different types of events overlapped, we counted the smaller-volume end use as part of the larger volume end use. For example, a toilet flush during an irrigation event was classified as part of the irrigation event and would add approximately 3.875 L of use to the irrigation event. We classified 90% of all events, with the unclassified events comprising approximately 454,250 L across all participating facilities.

Scaling the information from the small number of facilities we monitored at high temporal frequency to a much larger number of facilities across the city also presents a challenge. Additionally, the results for the six facilities in the City of Logan may or may not be applicable to other cities in Utah or elsewhere. However, this scaling would provide valuable information for water managers interested in characterizing CII water use at the city level and encouraging water conservation.

To provide some insight into potential scaling, we analyzed pervious area, business licensing, and monthly water-use data provided by the City of Logan pertaining to all its commercial and industrial customers. Fig. 6(a) compares the existing water use (circles) and estimated water savings (stars) of the six participating facilities with the existing water use of the other 466 facilities (crosses). Several of these facilities with large water use and a range of pervious areas may benefit from the same water conservation actions. Fig. 6(b) groups the water use of all City of Logan commercial and industrial facilities by NAICS classification. Fig. 6(b) shows that additional facilities in the manufacturing and assisted-care sectors may benefit from conservation actions. It is hard to infer anything for the other sectors. Many sectors such as administrative, arts, education, public administration, and real estate rental already have very low water use and/or few facilities. Other sectors such as accommodation, construction, finance, information, services, professional service, retail trade, and transportation have some facilities with daily water use as high as that of the study facilities. However, we do not know the end uses in these facilities or which end uses present opportunities for water conservation. It also is difficult at present to scale results for these six commercial and industrial facilities to other cities. Facilities—even within the same sector—are heterogenous, and we required a walk-through of each facility to identify the end-uses and interpret the high-frequency data. In contrast, researchers have monitored thousands of residential households at high frequency and identified and characterized relatively more-homogeneous end uses (DeOreo et al. 2016; Mayer et al. 1999). These circumstances motivate the need to expand data collection for more commercial and industrial sectors and facilities to better quantify end uses and identify water conservation opportunities. To support this goal, we have published our raw and processed high-frequency data for others to use in future work (Atallah and Rosenberg 2020).