Welcome to NOAH
Water Security of a Water Distrubition System, Funded by United States Environmental Protection
Agency with American Water Participation

Problem Definition

This research funded by the United States Environmental Protection Agency, with American Water participation (largest privately owned water supplier in United States), was conducted to assess the feasibility of using artificial neural networks (ANNs) in conjunction with mathematical optimization for developing a “smart” water distribution security system. The system could be used to detect potential terrorist attacks against a water distribution system, and, in the event of such an attack, identify the optimal operational responses to mitigate it to the extent possible. Using hydraulic and water quality data provided by a mid-sized city’s water distribution system, hundreds of different ANNs were developed and tested for predicting hydraulic states and water quality conditions at points of measurement within the system. The predictive accuracy of the different ANNs assessed under a variety of operational and data set conditions, and data requirements and other critical modeling issues were identified. In addition, several hydraulic ANN models were combined with mathematical optimization to quickly (i.e. seconds) identify the optimal operational controls for a given management objective (e.g. maximizing tank water levels for fire protection).



Source water originates from both a surface water system and a groundwater via four wellfields located in different parts of the city. For security purposes, the name of the city or system nor potentially distinguishing features is omitted. A rough schematic of the principle areas of the system used in this study is shown in Figure 1. The system has two main service areas, a higher pressure zone, located at higher topographic elevations, and a second lower pressure zone. A variety of groundwater and surface water sources supply water to the distribution system. In order to facilitate delivery of this water through the system, particularly to higher elevation areas, booster station pumps are used as necessary. There are five water storage tanks within the service area of interest that help augment water supplies during periods of high demand. Flow rates from the water sources and at the booster stations are measured and stored at high frequencies (i.e. 30 seconds) by the SCADA system, as well as hydraulic pressures at select locations within the system and tank water levels.


Figure 2 is a general schematic of the ANNs input variable types used. For this particular modeling problem, select initial system states, water source and booster station flow rates were used as inputs to the ANNs, with the single final state variable(s) of interest constituting the output. Data collected continuously at five minute intervals were used to generate the training and testing patterns used for ANN development and assessment.

The ANNs accurately predicted hydraulic conditions over both shorter (i.e. 5 and 15 minutes) and longer (i.e. 8 hour) prediction periods. Because water levels in tanks change little over 5 and 15 minute periods, high predictive accuracy with the ANN models was expected. For the longer prediction periods (i.e. 4 and 8 hours), the ANNs also accurately predicted final tank water levels, as shown by Figure 3 for a representative tank.


The mean absolute prediction errors were significantly smaller than the mean absolute changes. For example, for Tanks 1 and 2, the mean absolute prediction errors for the eight-hour prediction periods were 0.54 and 0.73 feet, respectively, while the corresponding mean absolute changes in water levels in these tanks were 5.6 and 7.5 feet. The ANNs also accurately predicted highly variable and dynamic pressure states in the water distribution system as well.

The water quality models accurately predicted quasi-steady-state chlorine levels, but were unable to predict sudden significant changes (e.g. 0.5 ppm in 5 minutes). This is attributable to inadequate characterization of highly variable chlorine levels at critical water source/sink (i.e. tank) locations, and relatively sparse sensor coverage. One potential benefit of ANN models is they can help improve monitoring of the system, and can even QA/QC data; for example, the ANNs in this research identified likely spurious hydraulic data for one of the months provided.

For conducting mathematical optimization, a variety of objective functions were either maximized or minimized subject to imposed management constraints. The ANN-derived state-transition equations were coupled with LINDO, proprietary optimization software, and one hundred and forty different optimization problems were solved. Because of the complexity of the water distribution system, not just in terms of physics, but in terms of the water balance constraints, minimum and maximum constrained flow rates, minimum and maximum constrained tank water levels, and minimum and maximum constrained pressures, the solutions were evaluated for accuracy using a number of heuristic criteria. First, the computed optimal solutions for conflicting objectives, such as maximizing a certain tank’s water level and minimizing the same tank’s water level, given the same initial state, were compared against each, in terms of values of the decision variables and final states. Second, it was verified that the final values of the state and decision variables were all within the imposed minimum and maximum constraint values. Third, it was verified that the two water balance constraints imposed for the higher and lower pressure zones were met. In all cases, the optimal solutions were within the bounds of the constraint set. In addition, the computed decision variables within the context of the objectives displayed numerical consistency and conformed to physical intuition.

This work demonstrates the potential for developing multi-state ANN-based water security system that continuously assess both hydraulic and water quality conditions in real-time for identification of a possible terrorist attack (or accident). In addition, by coupling the technology with mathematical optimization software, appropriate crisis responses can be identified quickly and accurately. The research results to date, particularly with regard to hydraulics, demonstrate strong promise for protecting water distributions systems with this system as well as performing other important functions, such as minimizing energy consumption.