Optimal Deployment of the Water Quality Sensors in Urban Drainage Systems ^†

Abstract

In the water sector, the problem of pollution-source identification was mainly investigated regarding pressurized distribution networks, with respect to sewers. Even if the Water Framework Directive 2000/60/EC and equivalent law-making bodies in many countries introduce the principle that the polluter pays, it is asking the water manager to detect the most pollutant discharges in sewers. In previous studies, a probabilistic approach to positioning water quality sensors in urban drainage networks shows the progressive increase in identification probability obtained through the Bayesian approach. Following previous literature, the present work aims to improve it by inserting new information beyond network topology. The methodology is applied to the real test case represented by the sub-catchment of the sewer system Palermo (Italy).

1. Introduction

The problem of water quality has always aroused considerable interest within the scientific community, as it is closely connected with the risks associated with public health [1,2,3] and environmental problems [4]. For this reason, many researchers have applied different methodologies to monitor the water quality parameters (chemicals and biologicals) of water distribution networks (WDN) and urban drainage systems (UDS) [5,6,7]. The latter can be a helpful tool to check the input values of the sewage treatment plants to monitor the efficiency of the removal process. To monitor the numerous chemical and biological parameters that can come into play in contamination episodes, different types of technologies and sensors have been studied and developed in the literature, which find application in most cases within WDN. According to the review, from the electronic point of view, sensors work with the same standards both in WDN and UDS. This feature allows for using different types of sensors on the same monitoring frame, enabling a more extensive choice on the market [8]. From the measurement point of view, on the other hand, they differ according to the parameters studied, timing, and accuracy of measures. Some parameters (temperature or pH) can be measured directly using specific technologies. In contrast, other parameters, such as salinity or dissolved matter concentration, are measured through indirect measurements obtained from redox potential, conductivity measurement, turbidity, or a combination of them [9].

The development of new computational algorithms, technological improvement of new measuring instruments, and the creation of reliable and economical microcontrollers allow the researchers to develop new monitoring techniques. Passing from discontinuous monitoring, in which the measurements are carried through levies of a sample at specific points of the networks, to continuous real-time tracking at strategic points of the networks through the development of algorithms able to identify any contamination within networks [10], a critical issue that allowed technological innovation was the introduction of a new science, called the Internet of Things (IoT), allowing sensors to be continuously accessible and data to be retrieved on demand. IoT can be defined as the application of simple devices connected via the Internet to improve system quality [11]. This science has developed in recent years due to the development of simple and inexpensive hardware. Through microcomputers, such as Raspberry Pi, and microcontrollers, such as Arduino, it was possible to start developing IT solutions aimed at direct interaction between people and computers.

IoT has evolved so much that its definition has changed compared to the original. We mean all the technologies developed on the “things” for IoT. This technology aims to improve the classical use of things by providing an internet connection. This technology can enhance the potential of any object, instrument, or sensor connected to a telecommunications network. Therefore, this science aims to simplify everything, from the simple ordinary life tool to the most complex sensor. It has been demonstrated [12] how this approach can help reach the concept of digitalised industry.

2. Materials and Methods

2.1. The Adopted Optimisation Approach

Using the Bayesian approach, new information from the analysis is incorporated, allowing the operator to gain insight into the system once recent contamination events (infiltration of brackish waters in the present case) are detected, identified, and monitored. For the solution to this problem, two main components are required: a calibrated model for hydraulic and water quality simulations in sewer systems, and a Bayesian Artificial Intelligence (AI) solver for likelihood estimation and probability update. In this case, the EPA SWMM model was used to perform the hydraulic and water quality simulations. Decision-making support of the Bayesian Decision Network (BDN) type was implemented to position the water quality sensors. A Bayesian network (BN) is a graphical structure that allows us to represent an uncertain domain. The BN is a very robust and advantageous method for assessing risk and uncertainty, providing a complete framework for analysing all cause-effect relationships. The standard formulation of Bayesian posterior probability can be found in the literature [13].

The problem of sensor location for identifying the illicit intrusion for the considered case study has been already investigated in [14] but using genetic algorithms. In particular, different single and multi-objective optimisation procedures to optimally locate sensors in the sewer have been compared. In [7], without pre-conditioning, initially, all nodes had an equal probability of being the candidate nodes for sensor placement, and all nodes had an equal chance of being a brackish water intrusion source.

In the analysis, the likelihood function was slightly adapted from that presented in [15] to comply with sewer networks instead of water distribution networks. According to [15], the isolation likelihood F₁ is expressed by the following equations:

$$one{F}_1=\frac{1}{S}{\displaystyle \sum }_{i=1}^S{d}_r$$

(1)

where S is the total number of analysed contamination events, d_r is one of the contamination that was identified by the sensor network and is 0 otherwise. The indicator F₁ provides information on the ability of the sensor network to locate the infiltration source.

Therefore, the infiltration events are randomly simulated to evaluate each sensor’s probability of identifying the source of contamination. The analytical approach was based on three phases:

◾

Based on available data, random simulations are used to explain available data with the position of the potential source of contamination; such degree provides the share of probabilities that infiltration may be located in different parts of the network.

◾

According to such probabilities, a possible sensor network is designed employing BN to maximise the ability of the network to detect the position of the contamination source (based on the maximisation of the probability to detect the proper position of an unknown contamination source).

◾

An experimental search of the contamination source using the designed sensor network.

2.2. The Experimental Campaign

The present analysis was carried out on a real case study, the network of Mondello, a touristic seaside village adjacent to Palermo city centre (Italy), using the SWMM model as a simulation tool for analysing contamination events propagation. As that drainage system is affected by seawater infiltration inflow, the approach was used to identify potential infiltration locations among 1786 nodes. Seawater, in principle, has the same dynamic of dissolved conservative contaminants with the only difference being that the contamination source is not determined by human activities, but by the altimetric proximity of the drainage system to the sea level, and by the presence of damages in the drainage systems that may be the point of access for infiltrated seawater. In the first simulation phase, each contamination event is simulated by a random mass of contaminant (ranging from 10 g to 500 g) constantly injected in a node for a random time (ranging from 15 min to 3 h). In this sensor location exercise, two of the previous limiting hypotheses were removed:

◾

Contamination probability is no longer equal in all network nodes, but some have a higher probability of being origins of seawater infiltration; such probability is based on the altimetric distance between pipes and groundwater level.

◾

Pipes were divided into three categories: pipes laying under average groundwater level, pipes laying in the average capillary fringe, and pipes lying above (Figure 1).

◾

The hypothesis was made that nodes in the first category (173) share 60% of the probabilities of hosting groundwater infiltration; nodes in the second category (211) share only 30%, and the nodes in the third category (1402) share only 10%.

3. Discussion of Results

The application of the BN approach allowed us to identify 16 nodes of the network that may be used to locate salinity sensors for which the likelihood function F1 equals 84%. Adding another sensor would produce less than a 1% increment in the likelihood function. The number of selected nodes is less than 1% of the network’s total number of nodes (manholes). Eight sensors are located in the red zone of the network (where urban drainage is under groundwater level), six are in the capillary fringe, and two are located in the other parts of the network (corresponding to more than 78% of the network). After the installation of sensors, the mathematical model and the BN were used in a predictive way to locate the most probable locations of seawater infiltration. Six potential sites were identified in two different areas (each with a search radius of less than 200 m). Two locations were positively validated as infiltration sources by visual inspection (Figure 2).

4. Conclusions

The proposed approach showed an effective integration of experimental campaigns, mathematical modelling, and AI to solve an urban drainage management problem: seawater infiltration in sewers. The model and AI were combined to analyse the system and propose possible locations of a limited number of sensors (based on external constraints). The AI applications’ main objective was to maximise the overall probability of locating contaminant sources (seawater infiltration points). After the deployment of the monitoring network, the AI system and the model were again used to predict the most probable location of sources and to guide an experimental campaign to validate the prediction. The application to the Mondello network successfully identified two source areas containing two infiltration locations.