Evaluation of Free-Chlorine Data from Online Sensors in a Water Supply Network ^†

Abstract

The use of data from reagent-free water quality sensors in water supply networks, which monitor at a high spatiotemporal resolution, is limited by variations in data quality and sensor sensitivity. This study examines a dataset from state-of the-art sensors installed in a UK water distribution network, providing unprecedented spatiotemporal resolution. By comparing continuous free chlorine data with monthly grab samples using Bland–Altman plots, we quantify the uncertainties of sensors. The results indicate that unlike the grab samples, data from the online sensors offer significant insights into the fluctuations in water quality dynamics. An analysis of sensor performance and limitations identifies sources of uncertainty.

1. Introduction

The Drinking Water Inspectorate (DWI) is advocating the use of risk-based drinking water safety plans, which necessitates the development and implementation of accurate water quality models and enhanced sampling techniques with high spatiotemporal frequencies and accuracy. Free chlorine is a commonly employed method for disinfecting water networks. Currently, water utilities rely on customer reports to identify water quality incidents [1], and free chlorine monitoring from the treatment plant to the consumers tap consists of periodic discrete (grab) sampling [2]. This approach is becoming increasingly insufficient, and the latest Surface Water Treatment Rule recommends regulatory reporting of free chlorine on distributed water systems serving more than 3300 people [3]. This indicates a need for more proactive and systematic monitoring methods.

Reagent-free water quality sensors in water supply networks, which continuously monitor chlorine and other water quality parameters at sampling rates of 15 min or less, are becoming increasingly available [4]. However, electrochemical chlorine sensors rely on membrane technology, are susceptible to contamination, and require regular calibration and servicing. They are also sensitive to flow, pressure, pH, and temperature, and their performance in water networks can significantly vary. Additional research is required to analyse the continuous chlorine data collected from online sensors and assess their ability to meet the requirements for modelling and managing water quality in water networks.

The USEPA and some states have set a minimum free chlorine compliance level of 0.2 mg/L throughout the system for effective microbial protection. In the UK, while no official regulation exists, utilities set internal targets (e.g., Bristol Water, serving 1.2 million aims for 0.2 mg/L). These specifications highlight the need for free chlorine data to achieve a minimum accuracy of 0.1 mg/L, i.e., half the lowest compliance level.

This paper presents an exploratory analysis of a water quality dataset acquired with state-of the-art water quality sensors at high spatiotemporal resolution in a real-world operational water distribution network in Bristol, UK, for 2.5 years. We focus on free chlorine data, discuss whether it meets the criteria for water quality modelling, and assess the sensors’ accuracy, repeatability, and uncertainty.

2. Case Study Description

2.1. Experimental Setup

The Bristol Water “Field Lab” is an operational water distribution network, supplying approximately 8000 customer connections in the city of Bristol, UK. Figure 1 shows the locations of the nine MetriNet-Q52 sensors developed by Analytical Technology (ATi) (Manchester, UK), measuring chlorine, pH, temperature, conductivity and colour at rates ranging from 1 sample/sec to 1 sample/15 min, for 2.5 years. There are the following two main flow directions in the distribution system: (i) BW1 to BW3 and (ii) BW4 to BW9. The other locations experience flow mixing. Two sources, with different chlorine concentrations, supply BW1. The network is also equipped with continuous hydraulic sensing.

2.2. Chlorine Sensors Setup and Maintenance

Figure 1 shows the three types of ATi installation. One device is placed in a telemetry bollard specially designed to guard against subterranean risks. Five devices are placed in telemetry kiosks and three in fire hydrants. All sensors are mounted within flow cells with constant flow, 200 mL/min. The chlorine electrochemical sensor generates a current proportional to the free chlorine. Sensors are calibrated with a one-point calibration, using DPD (N,N-Diethyl-p-Phenylenediamine) samples. Sensors with slopes 30–300% of a new sensor’s slope are considered to provide adequate sensitivity [5].

Monthly visits were conducted, during which the following maintenance procedures were implemented: (i) DPD samples were taken for validation and calibration; (ii) calibration was executed when the chlorine difference between the methods exceeded 10% of DPD; (iii) rebuilding of the sensor (electrolyte and membrane replacement), was conducted when the slope fell below 30%; and (iv) continuous flow through the flow cell was ensured by mitigating blockage risks.

3. Results and Discussion

3.1. Continuous Chlorine Data

Figure 2 shows chlorine data over 6 months at BW1 and BW12. The online sensors detect large chlorine fluctuations. This is because BW1 is supplied from sources with varying chlorine concentrations; one with nearly zero chlorine. Mixing water with different free chlorine concentrations is common, and the grab samples do not adequately capture the implications of such mixing. These rapid changes in chlorine concentration propagate through the pipe and are observed at sites BW12 and BW3. Assuming plug flow, the travel time from BW1 to BW12 (800 m, 250 mm diameter pipe) is approximately 2 h, matching the observed data (Figure 2c). This verifies the sensors’ fast response to chlorine changes.

3.2. Impact of Installation on the Online Sensors—Fire Hydrants vs. Telemetry Kiosks

Figure 3a,b present the scattered chlorine data pairs collected at BW12 (in a kiosk) and BW7 (in a fire hydrant) with the two measuring methods. The sensor in the fire hydrant encountered low or no-flow conditions, leading to extended resident times and decreased chlorine concentrations. The online sensor successfully captured these low concentrations. However, it experienced repeated failures stemming from flooding or no-flow conditions, leading to the drying out of the electrochemical sensor’s membrane. Installation conditions significantly affect electrochemical sensor performance, stressing the importance of selecting suitable installation points.

3.3. Chlorine Data Uncertainty Analysis

Our aim is to evaluate the uncertainty of the sensor and the qualitative agreement between the two datasets. To accomplish this, we utilized the Bland–Altman plot, which identifies any systematic bias between two measuring methods and the consistency of agreement [6]. This plot determines the limits of agreement (LOA) between the datasets. A scatterplot is drawn in which the x-axis represents the average chlorine values collected from the DPD and the ATi sensor, and the y-axis represents the difference of the two measurements. The mean bias and the 95% LOA for each comparison are plotted as follows:

$$Upperlimit=\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}-1.96\times \sigma d$$

(1)

$$Lowerlimit=mean+1.96\times \sigma d$$

(2)

where $\sigma d$ is the standard deviation of the differences. These lines represent the range within which 95% of the differences between the two methods are expected to lie.

The Bland–Altman plots show the following: (i) the mean difference between the two methods is negligible (order of 10⁻²), suggesting that there is no systematic bias between the two methods; (ii) uncertainties are not correlated to chlorine; (iii) no trend is observed with time, indicating that our maintenance strategy is efficient; and (iv) the largest difference between the measurements occurs during instances of blockages within the flow cell, highlighting the importance of regulating flow conditions within the cell. The LOA, which tells us how far apart measurements by the two methods are more likely to be, are 0.14 and −0.2 for BW12 location and 0.2 and −0.2 for BW7. These findings suggest that while the ATi data and DPD grab samples are not currently interchangeable, such water quality sensors could potentially provide the necessary accuracy for water quality modelling (0.1 mg/L). Future works aim to modify and extend the experimental program to quantify and minimize the sources of uncertainty, such as the installation conditions, the effect of pH value, temperature fluctuations, and uncertainties in DPD measurements.