Bounds on water quality sensor network performance from design choices and practical considerations

Since the start of the 21st century, events in society and technological advances have pushed the development of techniques for online water quality monitoring in drinking water distribution systems, to protect customers from incidental and/or intentional drinking water contamination (e.g. Kroll & King 2010). Budgetary constraints always limit the number of online monitoring sensors that can be placed in any system, so methods have been developed to determine optimal sensor placement (Ostfeld et al. 2008; Berry et al. 2010) within a drinking water distribution network. Optimality is, however, a matter of definitions and requirements. Many objectives have been presented in the literature. They can be classified roughly into 3 categories, aimed at obtaining information, facilitating utility response, and mitigating the effects of contamination (Table 1). The literature is focused mostly on the first (Ostfeld et al. 2008) and the last (Berry et al. 2010).

Many studies have been presented on the optimal placement of water quality sensors. Very few water companies, however, have actually deployed a water quality sensor network in their distribution system. Vitens Water Company (Vitens) is one of them, having a sensor network in their Vitens Innovation Playground (VIP) (De Graaf et al. 2012). This provides a unique opportunity to study sensor placement in practice.

The three objective classes are compared here in the context of this live network. Theoretical and practical considerations concerning the application of optimal sensor placement strategies in a real distribution network are discussed, as well as the response oriented approach.

In order to optimize sensor placement and apply the strategies described in Table 1, a custom tool called the Contamination Source Toolkit (Van Thienen 2014) was created. This implements several information and utility response oriented strategies. It uses EPANET-MSX (Shang et al. 2008) to calculate contaminant transport and genetic algorithms implemented in the Inspyred library (Garrett 2015), to perform optimizations. In addition, TEVA-SPOT (Berry et al. 2010) was used for performing optimizations aimed at effect mitigation.

Mean detection likelihood and mean time to first detection are common objectives for sensor placement optimization, and are included here for comparison. Ostfeld et al. (2008) note that detection likelihood and time to first detection are in conflict with one another with respect to sensor location optimization. This assertion is closely related to their choice to exclude undetected events from their analysis. When, for example, a substantial penalty time would be given to each undetected event, a detection-time optimized sensor configuration would tend to be optimized for detection likelihood. In this work, the water's maximum residence time in the injection point when it is injected is used as the penalty time.

Network coverage is defined as that fraction from which water is sampled during a predefined observation window. It is closely related to detection likelihood, but allows the user to choose how the network fraction is expressed, e.g. as network length, number of connections, number of connected costumers, etc. Redundancy is introduced by demanding concurrent observation of a network segment by at least n sensors. Redundancy allows water companies to, for example, start preparatory actions at the first detection and escalate when confirmation is obtained from a second sensor.

When using a (partially) skeletonized network model – see below – it is important to express network coverage in terms of a parameter conserved during skeletonization, such as the number of connections, rather than something like pipe length or volume, which are not.

The most important tool for determining the source area of a contaminant is an accurate hydraulic model of the distribution network, in which a contaminant can be traced back in time from its point of observation to all places where it might have originated in the network. Several approaches to such back-tracing or backtracking are presented in the literature (e.g. Shang et al. 2002; Laird et al. 2006). The approach used in this study was based on a combination of forward traces (Van Thienen 2014), but any alternative back-tracing algorithm that also takes the dynamic flow field into account will give equally good results.

A dedicated area of Vitens' distribution network, the VIP, is used to test innovative added-value technologies. One of these is the application of sensors (sensor hardware, data transfer, algorithms, etc.). The VIP is in the north Netherlands, and includes both rural and urban areas. For the study discussed here, part of the VIP, with about 600 km of pipes, was used. Some skeletonization for performance purposes resulted in a model with about 300 km of pipes. The VIP distribution network can be supplied from either a single water source or the three sources supplying the area. These degrees of freedom are important as they can be used to help in both understanding sensor response and modelling the distribution network.

Of the three classes of sensor placement optimization objectives, those that are information-oriented are the simplest to compute (Ostfeld et al. 2008), requiring only a network model (hydraulics and material transport). The more complex effect-oriented approach has been implemented in TEVA-SPOT (Berry et al. 2010) and applied to a network model of the VIP. Many simulations were performed to study the relationship between the assumed and actual response times of the utility (after which water consumption is assumed to cease), on the one hand, and the performance of the sensor network (reduction in the number of people affected) on the other. Some results are shown in Figure 1. A number of observations can be made:

• sensors are useless for event detection if the utility's response time is too long;

• every additional sensor contributes less to the objective than its predecessors (the law of diminishing returns);

• optimizing the sensor network for a long response time leads to reduced performance if the actual response time is shorter.

With a custom-made optimization tool Contamination Source Toolkit (Van Thienen 2014), sensor placement optimization calculations were also performed for information- and response- oriented objectives. Two response-oriented objectives were considered. The first, redundant coverage, is response-oriented in the sense that a utility will not (fully) act upon a single sensor signal, preferring/requiring further confirmation first by observing the same event at a second sensor. Therefore, this objective maximizes the observability of individual contamination events by at least two sensors. The second, contamination source identifiability, minimizes the mean potential source area size that can be reconstructed from back-traces of contamination reports from multiple sensors (Van Thienen 2014). Figure 2 shows the relative performance of several optimized sensor networks in the VIP for these objectives and the current installation configuration of water quality sensors, based on experience and practical considerations. Different sensor network configurations optimized for different objectives and the actual configuration at that time are organized in rows in the figure. For each configuration, the performance is indicated for three information-oriented and two response-oriented objectives (colors indicating the best to worst performance for each column). As can be seen:

• 1.

  the sensor networks perform best on the objectives they have been designed for, and

• 2.

  the original configuration based on insight into network hydraulics, experience, and practical considerations performs relatively poorly compared to numerically optimized configurations.

In this diagram, there is a positive correlation between optimization towards and performance on the information-oriented objectives detection likelihood and time to first detection, meaning that configurations optimized for detection likelihood also perform well on time to first detection and vice versa. This is contrary to earlier results reported by Ostfeld et al. (2008) and arises from the different treatment of non-detections in the approach used here.

On the other hand, the response-oriented objectives perform relatively poorly on sensor networks optimized for information-oriented objectives and vice versa. The main message from Figure 2 is, therefore, that the objective of a sensor network must be well defined before design is undertaken. Also, the figure clearly shows that such optimization is worthwhile.

Even though it is tempting to use all network model nodes as potential sensor locations, practical considerations limit the actual number that are suitable. Some examples are shown in Figure 3. The problems encountered include accessibility, power availability, the possibility of data transmission, etc. In the example in Figure 3(a), the optimal location is in a residential street.

Although (some) other locations on the same pipe in the same street would probably be equally suitable, in any case either a manhole should be created in the street, and fitted with electricity and a data connection, or a sensor should be installed in a home in the street, using the owner's electricity and, if possible, broadband connection. Also, in order to ensure that the sensor is exposed to a sufficiently large and continuous flow of water, permanent flow conditions must be created at this home, either by wasting water or by bypassing the pipe in the street. Similar considerations apply to the second (Figure 3(b), similar situation) and third (Figure 3(c), no buildings) examples.

When the water utility elects to consider only practically suitable locations for sensor placement, this may have a significant effect on the network's performance – see Figure 4. This figure shows how optimizing a sensor network configuration based on practically available locations does not only result in performance loss compared to networks in which sensor placement is not restricted. The figure comprises a comparison between the performance of optimizations for different sets of uniformly distributed nodes as potential locations (300, and all 2,700, respectively). For very small and very large numbers of sensors, the results vary significantly between the practical set and the limited set of 300 potential locations, while for moderate numbers the differences are small. In some cases, the network based on the practical set of potential locations performs better than one based on a larger number of uniformly distributed nodes. The practical set may include, by chance, suitable locations that are absent from the uniformly distributed sets. However, when all 2,700 network nodes are considered as potential locations, a significantly better performance can be achieved (grey curve in Figure 4).

Figure 4 also shows that numerical optimization of sensor locations is expected to result in better performance than the current design. Based on some of the results reported here and other calculations, a number of sensors have been relocated by Vitens. Analysis of the effects of these translocations on sensor network effectiveness are ongoing.

Nevertheless, even though the sensor positions in the network (generic water quality) are not optimal, all known events (change in water hardness, failure of color removal at the treatment plant, turbidity increases associated with bursts) have been detected during its evaluation. Thus it has been demonstrated that this sensor network works, but questions remain about optimal sensor density with respect to investment and returns – e.g., in relation to safety and/or security.

As response time is a critical parameter in the actual effect on the population of a contamination event, sensor placement strategies aimed at optimizing the response strategy – e.g., verification through redundancy or contamination source identifiability – have the potential to contribute strongly to the success of a water quality sensor network in a drinking water distribution system. Practical limitations, such as the availability of power and data connections (or the utility's willingness to invest in these at every sensor location) determine the degree of realization of an optimal sensor network. Nevertheless, the trial sensor network picked up many water quality events. Combining practical considerations of sensor placement with numerical evaluation and optimization of the sensor network, enables realistic expectations towards these sensor networks before installation, and ultimately the best consumer protection.

The work was performed in the SmartWater4Europe project. This project received funding from the European Union's Seventh Programme for research, technological development and demonstration under grant agreement number 619024.