Capturing high-resolution water demand data in commercial buildings

Water demand measurements have historically been conducted manually, from meter readings less than once per month. Leading water service providers have begun to deploy smart meters to collect high-resolution data. A low-cost flush counter was developed and connected to a real-time monitoring platform for 119 ultra-low flush toilets in 7 buildings on a university campus to explore how building users influence water demand. Toilet use followed a typical weekly pattern in which weekday use was 92%± 4 higher than weekend use. Toilet demand was higher during term time and showed a strong, positive relationship with the number of building occupants. Mixed-use buildings tended to have greater variation in toilet use between term time and holidays than office-use buildings. The findings suggest that the flush sensor methodology is a reliable method for further consideration. Supplementary data from the study’s datasets will enable practitioners to use captured data for (i) forecast models to inform water resource plans; (ii) alarm systems to automate maintenance scheduling; (iii) dynamic cleaning schedules; (iv) monitoring of building usage rates; (v) design of smart rainwater harvesting to meet demand from real-time data; and (vi) exploring dynamic water pricing models, to incentivise optimal on-site water storage strategies.


INTRODUCTION
Early studies on residential and commercial water use were driven by the need to quantify network demand, improve the design of water distribution systems, develop short-and long-term demand forecasts, and establish appropriate pricing structures (Buchberger & Wells ). Water demand at commercial premises such as offices and educational facilities has historically been measured using analogue water meters. The majority of these meters are read monthly or quarterly to coincide with the local water service provider's billing cycle (Thames Water a).
Where more granular water use data is needed, sub-metering systems are implemented to enable the costs of water to be suitably allocated to internal clients and other on-site users. Even where additional data are captured through sub-metering, the analogue and sparse temporal nature of the data collection prevents supply patterns from being derived. Although monthly usage data is typically sufficient for billing purposes, it has little value in terms of insight into leakage; user practices; seasonality; and responses to extreme events. In an ideal world, water resources managers might hope to access novel tools based on big data solutions (Chen & Han ) supported by high-performance computing (Morales-Hernández et al. ) and receive highly granular, near-real-time forecasts for water demands that could, in turn, help them to optimally manage the operation of water distribution networks upstream of a customer.
Water demand models can be built from statistical information on users, end-uses, frequency, intensity, and duration of use, as well as time of water use (Blokker et al. , ).
These models rely on statistical information of water-using appliances, and those that use them, instead of actual water use measurements (Blokker et al. ) and can enable a realistic estimation of demand where costs of a conventional metering approach are prohibitive. Simulations, based on a variety of non-residential buildings, have shown that water demand is a function of the way in which water is used within that building type. Water use in office buildings is primarily determined by the flushing of toilets, while water use in hotels is primarily determined by the total use within customers' rooms (e.g. WC, shower, bathroom tap, bathtub, etc.) (Blokker et al. ). Therefore, for an office building, reliable information on toilet use is of primary concern; while for a hotel, reliable information is required both on the number and occupancy of the rooms and the water-using appliances in the rooms (Blokker et al. ). Furthermore, users are the key to understanding and predicting water demand in buildings. The number of people present in a non-residential building, such as an office, will vary throughout the day, and across the week (Blokker et al. ). Pieterse-Quirijns et al. () suggest that with a proper estimation of the number of users and appliances within a building, such models can create a realistic diurnal demand pattern. However, while some nonresidential buildings, for example, offices, hotels, etc., have a defined number of users at a given time, that can be reliably estimated from the number of employees or the number of guests, other buildings, such as those on a university campus are less static and contain a large number of transient users.
In the past two decades, data-logging, storage, and lowpower wide-area network (LPWAN) communications technologies have developed apace, resulting in the wide-scale application of (IoT) technologies (Gurung et  With the progress in IoT technologies, smart water meter installation programmes have been implemented by many of the UK's Water Service Providers (WSPs).
One leading effort was initiated by Thames Water in January 2016 (ThamesWater b). Within 2 years, they achieved significant demand reduction benefits, for example in the identification of customer-side leakage (WWT ). By June 2016, Thames Water had installed 240,000 smart meters providing 5.8 million data points per day at an hourly time-step. This helped them to identify 11 million litres of leakage, but perhaps more importantly, is enabling Thames Water to undertake more accurate analysis of water user behaviour within each district metered area.
A typical practice adopted by WSPs in the UK sees their smart water meter networks target hourly data capture (Hackett ; Thames Water b). This is the result of an economic trade off wherein WSPs must balance the need for water meters to have adequate battery life (i.e. 5-10 years); an appropriate near-real-time data uplink frequency; while securing a manageable total number of data points; against the desire to have the highest resolution dataset possible. Higher-resolution, 15-min data can be captured from Anglian Water's systems (Hackett  showed how 1-min resolution data from smart water meters can add value to the development of predictive models and called for practitioners to make use of the finest resolution possible when undertaking data acquisition. As cost-effective cloud data storage has become available, capturing higher-resolution water demand data has increased. For example, micro-component analyses have been performed to identify how water meter data can be linked to each household appliance. Wills et al. () demonstrated that demand signals for various appliances, such as dishwashers, can be successfully identified from high-resolution pulse data on smart meters.
Recent developments in reliable, low-cost telemetry systems associated with Industry 4.0 technologies have enabled local data collection to be achieved using a wider set of novel sensors. In the future, reliable battery-powered IoT technologies are forecast to provide fully connected smart cities, with data being available from any and all infrastructure within the urban realm (Lom et al. ; Rathore et al. ).
Repurposing or upgrading traditional water auditing methods represents a field of opportunity wherein analogue monitoring methods can be upgraded using IoT technologies.
Analogue flush counters can be deployed to measure the frequency of use at toilets by measuring the frequency of cistern refills. These simple technologies can help estates managers understand toilet usage and thus explore in-cistern leakage rates when compared with water meter data (Waterwise ). They can also be deployed for short-term monitoring studies to demonstrate the economic value of undertaking appliance upgrades, for example, by switching to a waterless urinal or upgrading old toilets with oversized cisterns. Analogue devices such as these are used during water audits, however, they provide coarse resolution data (i.e. daily reads) and although they are low cost to purchase, they are expensive to operate, as the process of taking manual readings is highly laborious. Float switch sensors which include reed switches have been successfully deployed in rainwater harvesting tanks to support the efficient operation of pump controls. Studies in that field have also seen them deployed as event counters in low cost, data-logging pilots for a series of alternative water resource studies (Melville-Shreeve et al.

).
Such flush counters can be easily mounted (and removed) to capture the events associated with a change in water level in a toilet cistern (i.e. flushes). Combining these event counting technologies with contemporary microcomputers and associated bespoke software permits data to be captured from many water closets (WC) and shared in raw format for further analysis.
Digital turbine meters have increased in quality and reduced in price ($ < 50). Again, when combined with a micro-computer, these sensors can enable practitioners and building designers to incorporate a network of water demand monitoring within commercial premises and thus derive accurate demand management datasets. Implementation of such approaches is likely to become a new normal in the years ahead as IoT technologies fall in price and sensors can be incorporated within infrastructure (or appliances themselves) at with very low capital cost. Opportunities to correlate and link high-resolution water demand data (e.g. from washrooms within a campus) to other building data such as floor space and occupancy could pose value to building designers who, for the first time will be able to design new water systems while getting real-time feedback once the site is operational.
At a city-scale, high-resolution data from meters within water networks have been used to train forecasting algorithms, such as artificial neural networks, to predict future water demands. There remains, however, a gap in the literature As part of a wider study providing a living laboratory for water demand management techniques, a research opportunity was identified to develop a novel dataset of high-resolution water demand across a zone of a university campus. In this paper, we investigate data from a highresolution smart water metering platform installed as part of a large-scale water demand management programme at the University of Exeter, UK. We seek to generate a set of high-resolution water demand profiles that can be made available to researchers to enable further investigations and to support practitioners when designing commercial water supply systems. The research described herein explores the deployment, management, and interpretation of data from a series of washrooms within a campus setting.
High-resolution flush counters were installed to 119 WCs in 7 buildings. In addition, digital turbine flow meters were deployed within one building to capture micro-component washroom data. Floor plans, building user data, and local water meter data were obtained from the estates team to enable patterns in water usage to be explored. Outputs from the study could present value to the global water systems design community as design tables and heuristic methods deployed at the design stage are infrequently retro-tested against operational datasets.
To inform campus managers, building designers, and water resource planners, the authors have sought to explore to following research questions associated with the dataset: were fitted to monitor flows on the pipework feeding the toilets, enabling verification of the flush count data. In principle, these high-grade meters provide excellent accuracy, however, as with any retrofit plumbing, it was not always practical for the plumbers to ensure straight inlet and outlet runs to comply with the installation specifications. A standard volumetric water meter was installed on the final leg of pipework feeding each set of ULFTs within Harrison Building. These analogue meters were not connected to the realtime monitoring system and can be read manually to identify gaps in data/provide long-term records of performance as a fall back asset.
Throughout the trial, where data were lost or corrupted, a maintenance visit was conducted by a member of the project team and the data recording system reset. Data completeness throughout the study is described in the Results section.
Data quality assurance: flush count data Data collection for the 6-month period was completed and analysed for quality issues against a series of metrics. Data quality was assessed at the washroom level (Supplementary Annex B) and the building level (Table 1) 2. Data integrity was calculated as the amount of valid data in the dataset, expressed as a percentage of the available dataset. In some circumstances, the data-logging system may be functioning and recording data, but this data may be blank or erroneous (e.g. if a sensor has a loose connection, etc.) Data integrity was assessed by comparing the sum of invalid data points (null or erroneous) against the total data points in the dataset.
3. Data viability was, therefore, derived by multiplying data integrity by data completeness, when compared with the maximum number of potential data points.
4. The flush count data was compared with the digital and volumetric flow meters to explore the validity of the flush count data.

Data capture and manipulation
Data were captured at a micro-computer in each washroom in a time-series format. These raw data files were shared to a remote server to enable access to the time-series data. Data manipulation tasks were coded in python to provide demand data for each washroom as plain text files with flushes and flow events captured at 1-min, 15-min, and 1-day time series. These were further manipulated to enable analysis by building and time period to facilitate the presentation of results.
Following an initial review of the available data, two 8-week subsets were selected to enable a comparison between term-time water demand and non-term-time water  Figure A2). The mean washroom flush count over a week was calcu-  Figure 3), such that the 7 buildings could be directly compared (Table 2). Data from digital flow meters was automatically converted to a minutely time-step for washrooms throughout the Harrison Building (Figures 1 and 2).
A 1-week period following the two 8-week subsets was evaluated to assess how well the mean demand profiles represented a typical week in low-demand and high-demand windows ( Figure 5). The first week of September 2019 was chosen to evaluate the window of low-demand averages.
The first week of December 2019 was chosen to evaluate the high-demand averages. The mean value for each building ( Figure 3) is shown as a solid line; the mean flush count during the 1-week period following the 2-month average is shown as a bar chart with error bars (Figure 6).

RESULTS AND DISCUSSION
Data quality The quality of the flush count data was evaluated for completeness, integrity, and viability. Data integrity was 100% across all buildings during the data collection period (

Data validation
Digital water meters in the Harrison Building recorded total inflow to the ULFTs in each washroom (Figure 1). The seven Harrison building washrooms also had a volumetric water meter that was read intermittently by the maintenance team. Evidence from these seven washrooms showed that the flush count monitoring system was a robust method for recording water demand. However, a total water demand recorded on the water meters was found to be far higher than the flush count data suggested. In each case, these were found to be associated with maintenance issues such as leaky fittings, or one-off continuous flow events.
Discounting these washrooms, flow meter data for the remaining five washrooms was used to verify that the average flush volume flush count data. In each case, the fit was within 1.5-1.7 litres per flush ( Figure 2). However, flush counters alone fail to capture evidence of certain maintenance or high water usage events.  (Table 1).
Both buildings are used by staff who operate on normal working hours contracts. Neither building has laboratories or teaching spaces, and thus, they are more representative of commercial office space. A Pearson's partial correlation was run to assess the relationship between flush count and building after adjusting for the total number of WCs monitored, given that only two WCs were monitored in Kay, while 41 WCs were monitored in Amory. Pearson's partial correlation showed that there was no statistically significant relationship when the number of WCs monitored was controlled for r (9) ¼ 0.169, p ¼ 0.619.

Diurnal variation in toilet usage
Across all buildings, the flush count on weekdays was considerably higher than the weekend flush count (Figure 3).
The weekend flush count was 92.4% ± 4.6 lower than the weekday flush count during the low-demand period and 92.1% ± 3.9 lower than the weekday flush count during the high-demand period (Figure 3). There was little difference between the diurnal pattern of toilet use between buildings, with highest WC demand exhibited Monday to Friday and markedly reduced demand on Saturday and Sunday. In the high-demand window, Wednesday was found to have the lowest flush weekday count at the two large teaching buildings (Amory and Harrison). This could be associated with student sports activity on Wednesday afternoons and could be a useful indicator that footfall is lower. Such evidence warrants further exploration and could be useful to cleaning coordinators or café managers to understand when high footfall was experienced.

Temporal variations in water demand
Variation in water demand between the two observed periods (low and high demand) was not uniform between buildings. The Amory building, which is a mixture of  Following initial analyses herein, we anticipate that the dataset can be explored further to understand how it can generate further value for end-users. We anticipate that the data could meet the needs of a variety of users including facilities and estates managers, potable water service, and wastewater network providers, as well as informing security and lone working protocols and maintenance management.
It is important to note that there may be a much larger pool of users, than those described here, that may benefit from using such data. We anticipate that the supplementary data is well suited for use in future hackathon-style events.
The following explores a series of initial concepts for further investigation: can be optimally sized and operated using novel control regimes. This, in turn, enables dynamic water pricing models to be explored. For example, a water service provider might conceivably provide water during periods of low demand (e.g. winter/overnight) at a lower charge rate than during periods of high demand (e.g. diurnal peaks and summer). With accurate data captured for the buildings studied, models can be developed and tested which explore novel economic incentives as a way to reduce peak water demand and more optimally incentivize good practice within customer's water behaviours. Further development of the sensor network and data capture systems at the case study site offers the potential to investigate a wider set of benchmarks and metrics related to high-resolution water demand data.

CONCLUSIONS
A systematic regime for monitoring toilet usage at a 1-min time-step was conceived and implemented alongside the retrofit of 119 ULFTs at University of Exeter's Streatham Campus. The water consumption data gathered at 119 WCs show that water demand has significant variation in term time versus non-term time periods. The results support the following conclusions: 1. Toilet use events within buildings in a campus setting can be monitored using low-cost sensors (∼$1) and were found to follow a typical weekly pattern in which weekday use is significantly higher than weekend use.
2. Toilet demand is considerably higher during term time than during holiday periods, with flush counts 46% lower on weekdays (and 48% lower on weekends) in holidays.
3. Generally, toilet use showed a strong, positive relationship (r ¼ 0.931 low demand; r ¼ 0.877 high demand) with building metrics such as the number of occupants in office and teaching spaces. This could offer a means of reflection on building use rates and inform management practices or the scheduling of events.
4. Mixed-use buildings, such as those used for teaching and laboratories, tended to have greater variation in WC use between term time and holidays than office-use buildings. This is likely due to the high ratio of students to staff using these buildings. Office-use buildings had a more steady WC water demand throughout the year.
5. Real-time toilet flush data can provide accurate high-resolution water demand patterns. The data captured has been provided in Supplementary Annex C to enable research practitioners to explore concepts and benefits associated with the data captured in this study, for example through hackathon-style events or data science projects.
Additional benefits were identified for further explora-