Sensor Placement Guidance in Small Water Distribution Systems

Transcription

1 Sensor Placement Guidance in Small Water Distribution Systems Developed by the University of Kentucky and KYPIPE LLC Prepared for the National Institute for Hometown Security 368 N. Hwy 27 Somerset, KY June 31, 2013 This research was funded through funds provided by the U.S. Department of Homeland Security, administered by the National Institute for Hometown Security, under an Other Transactions Agreement, OTA #HSHQDC , Subcontract #02-10-UK.

2 TABLE OF CONTENTS TABLE OF CONTENTS... i LIST OF TABLES... iv LIST OF FIGURES... v CHAPTER Introduction Background Objectives of Study Content of Report... 4 CHAPTER Sensor Placement in Water Distribution Systems TEVA-SPOT TEVA-SPOT Methodology TEVA-SPOT Case Study KYPIPE KYPIPE Methodology KYPIPE Case Study Current Trends in Sensor Placement Betweenness Centrality and Receivability Rule-Based Expert System Rule-Based Decision Support System Demand and Reachability CHAPTER Water Distribution System Models System Configurations General Procedures of Model Development Pipe Roughness Coefficients Model Demand Input Final Adjustments to Model i

3 3.3 Description of Models used in Study Steady State and EPS Simulations CHAPTER KYPIPE Tool Sensor Placement Analysis Theory Procedure for Execution Performance Evaluation Contamination Scenarios Time to Detection Comparison Comparison of Identical Sensor Placement Results for All Contamination Scenarios CHAPTER Conclusion KYPIPE Sensor Placement Tool Conclusion Comparison of Results between KYPIPE and TEVA-SPOT Conclusion CHAPTER Acknowledgements CHAPTER References Appendix A Appendix B B.1 Data Acquisition in GIS B.2 Data Input to KYPIPE B.3 Addition of Elevation Data B.4 Final Adjustments to Model Appendix C Appendix D D.1 Execution of Tool D.2 Sensor Placement Tool in Progress D.3 Results Provided by Tool Appendix E ii

4 E.1 Results for Placement of One Sensor E.2 Results for Placement of Two Sensors iii

5 LIST OF TABLES Table 1: System Characteristics Table 2: Contamination Scenarios Table 3: Comparison between KYPIPE and TEVA-SPOT for Baseline Conditions Table 4: Analysis of Faster Times to Detection between KYPIPE and TEVA-SPOT Table 5: Identical Sensor Selection between KYPIPE and TEVA-SPOT (1 sensor) Table 6: Identical Sensor Selection between KYPIPE and TEVA-SPOT (2 sensors) Table 7: Sensor Placement Results for KY 1 (1 sensor) Table 8: Sensor Placement Results for KY 1 (2 sensors) Table 9: Attribute Matching between GIS and KYPIPE Table 10: Hazen-Williams Roughness Coefficients Table 11: KYPIPE and TEVA-SPOT Sensor Placement Results (1 sensor) Table 12: KYPIPE and TEVA-SPOT Sensor Placement Results (2 sensors) iv

6 LIST OF FIGURES Figure 1: Flow Chart of TEVA-SPOT Software (Murray et al., 2010)... 7 Figure 2: Sensor Placement at BWSN Network 1 (Ostfeld et al., 2008) Figure 3: Objective Comparisons Network 1 (Ostfeld et al., 2008) Figure 4: Objective Comparisons Network 2 (Ostfeld et al., 2008) Figure 5: Clustering of Nodes with High Betweenness Centrality (Xu et al., 2008) Figure 6: Community Divide in a WDS (Xu et al., 2008) Figure 7: Selected Nodes within each Community (Xu et al., 2008) Figure 8: Sensor Placement based on Receivability (Xu et al., 2008) Figure 9: Inner Nodes and Path Nodes (Chang et al., 2011) Figure 10: General Procedure for RBDSS (Chang et al., 2012a) Figure 11: System Configurations (a) Loop, (b) Grid, (c) Branch (taken Gagliardi and Liberatore) Figure 12: System in Branch Configuration Figure 13: System in Loop Configuration Figure 14: System in Grid Configuration Figure 15: Water Line Shapefile Figure 16: Model Development Procedure Figure 17: Systems in Loop Configuration: (A) KY1; (B)KY2; (C) KY3; (D) KY4; (E) KY13 47 Figure 18: Systems in Grid Configuration: (A) KY5; (B) KY6; (C) KY8; (D) KY14; (E) KY7 48 Figure 19: Systems in Branch Configuration: (A) KY9; (B) KY10; (C) KY11; (D) KY12; (E) KY Figure 20: Sensor Placement Tool Theory (Minimum Travel Time) Figure 21: Sensor Placement Tool Theory (Average Travel Time) Figure 22: Sensor Placement Tool Flowchart Figure 23: Comparison between KYPIPE and TEVA-SPOT - Baseline Conditions (1 sensor).. 60 Figure 24: Comparison between KYPIPE and TEVA-SPOT - Baseline Conditions (2 sensors) 61 Figure 25: Identical Sensor Selection between KYPIPE and TEVA-SPOT (1 sensor) Figure 26: Example of Differing Sensor Placement in Close Proximity Figure 27: Example of Differing Sensor Placement not in Close Proximity Figure 28: Identical Sensor Selection between KYPIPE and TEVA-SPOT (2 sensors) Figure 29: Water Utility Poster (Determining Water Distribution System Configuration) Figure 30: Water Utility Poster (Procedure for Executing KYPIPE Sensor Placement Tool) Figure 31: Distribution System Component Shapefiles in GIS Figure 32: Select by Attributes Figure 33: Calculating Length of Water Lines Figure 34: Data Export in GIS Figure 35: Matching Attributes between GIS and KYPIPE (Pipes) Figure 36: Matching Attributes between GIS and KYPIPE (Pumps) Figure 37: NRCS Geospatial Data Gateway v

7 Figure 38: KYPIPE Nodes Shapefile Export Figure 39: Nodes Shapefile in GIS Figure 40: Digital Elevation Model Figure 41: Combined DEM Process Figure 42: Elevation Extraction Process Figure 43: Opening Elevation File in Excel Figure 44: Elevation Data Sorting in Excel Figure 45: Nodal Elevation Data in Excel Figure 46: Pipe Connection Errors Figure 47: Changing Roughness Values of Pipes in KYPIPE Figure 48: Demand Allocation in KYPIPE Figure 49: Demand Patterns in KYPIPE Figure 50: Control Switches in KYPIPE Figure 51: KY 1 System Layout Figure 52: KY 2 System Layout Figure 53: KY 3 System Layout Figure 54: KY 4 System Layout Figure 55: KY 5 System Layout Figure 56: KY 6 System Layout Figure 57: KY 7 System Layout Figure 58: KY 8 System Layout Figure 59: KY 9 System Layout Figure 60: KY 10 System Layout Figure 61: KY 11 System Layout Figure 62: KY 12 System Layout Figure 63: KY 13 System Layout Figure 64: KY 14 System Layout Figure 65: KY 15 System Layout Figure 66: EPS Setup in KYPIPE Figure 67: Execution of EPS Analysis Figure 68: Starting Sensor Placement Tool Figure 69: Setting Parameters in Sensor Placement Tool Figure 70: Initiating Sensor Placement Run Figure 71: Sensor Placement Tool in Progress Figure 72: Sensor Selection Process Figure 73: Completion of Sensor Placement Simulation Figure 74: Completed Sensor Placement Simulation Figure 75: Completed Sensor Placement Tool Display Figure 76: Turning on Node Labels Figure 77: Displaying Selected Nodes on the Map vi

8 Figure 78: Sensor Placement Summary Report Figure 79: WQC File Figure 80: Time Matrix Excel File vii

9 CHAPTER 1 1 Introduction 1.1 Background Water distribution systems are an integral part of society, and the availability of a clean and dependable supply of water influences both the socioeconomic status and health of a populace. In recent years, protecting the nation s critical infrastructure from terrorist attacks has become a priority, and efforts to protect the water infrastructure are included in this goal. Water distribution systems are considered to be vulnerable to intentional, along with accidental, contamination because they have a large spatial distribution and multiple points of access. Many systems lack monitoring and security systems, greatly increasing the risk and potential danger associated with an attack (Hart and Murray, 2010). Public awareness of this threat has also increased from media coverage of two international terrorist plots against drinking water systems. One attack planned to introduce a cyanide compound into water lines near a U.S. Embassy in Italy, and another was a direct threat from an Al-Qaeda operative to American water supplies (Murray et al., 2010). Threats to the nation s water supply are concerning because they can cause a significant negative impact to public health and the economy in a short amount of time. Possible terrorist attacks to water supplies include sabotage of Supervisory Control and Data Acquisition (SCADA) systems, the physical destruction of facilities, airborne release of hazardous chemicals onsite, and the injection of chemical, biological, or radiological contaminants into the water supply. The threat of contaminant injection is perhaps the most dangerous because of the major public health, economic, and psychological impacts that could result (Murray et al., 2010). Intentional contamination of water distribution systems has become an increasing concern in recent years, but the accidental contamination of drinking water is also possible. Humans can unintentionally contamination distribution systems with pesticides, toxic industrial chemicals, or other materials. Systems can also be contaminated if metals, organic contaminants, or asbestos in pipe materials and linings are able to leach into the network. The risk of soil and groundwater contaminants permeating through plastic pipes is also present. Pesticides, insecticides, or other chemicals are able to enter the system 1

10 through accidental backflow occurrences or breaks in pipes/leaky joints (Murray et al., 2010). In an effort to mitigate the risks from intentional or accidental contamination of the water supply, contamination warning systems have been proposed as a cost-effective and reliable strategy. Contamination warning systems (CWS) are proactive strategies to lessen the effects of a contamination event in a water distribution system. The goal of a CWS is to deliver an early indication of intentional or accidental contamination in order to reduce public health impacts and economic loss, and it also works to improve the water utility s capability for a quick response (Janke et al., 2006). A CWS includes deployment and operation of online sensors, other surveillance systems, fast communication technology, and data analysis procedures to provide early alert of a contamination event (Murray et al., 2010). Arguably the most critical component of CWS, classified as online quality monitoring, involves the network of sensors that can assess the quality of water in the distribution system and alert an operator of a potential contamination event. Utilities developing these water quality monitoring systems are faced with the decision of what locations are best suited for deployment of these sensors; the location of these sensors is a critical component of a CWS. These water quality sensors must be placed in locations that maximize their ability to detect contamination events and provide the greatest protection of human health (McKenna et al., 2006). 1.2 Objectives of Study To date, there is no applicable federal or state guidance to assist utilities in the deployment of water quality sensors. Distribution systems are complex, dynamic infrastructures that differ greatly for individual utilities. This creates difficulties in the development of general guidance for sensor placement that are applicable to all distribution systems. Technological advancements in sensor placement optimization software may help solve the problem of sensor placement issues for some utilities. The TEVA-SPOT software (Threat Ensemble Vulnerability Assessment Sensor Placement Optimization Tool) has been developed to analyze the vulnerability of drinking water distribution networks and aid utilities in the design of sensor networks. A hydraulic model is setup in EPANET, and this is used as input for TEVA-SPOT to recommend sensor 2

11 placement based on a variety of user defined objectives. However, this software does not have the ease of use and simplicity that is needed to be beneficial to small utilities. It requires the use of complex water quality models along with sophisticated optimization methods to perform sensor placement guidance. Many small, or even medium sized, utilities might not have the technical or financial resources that are needed to effectively use TEVA-SPOT. Because of deficiencies in the current resources, the following research objectives were established: Develop sensor placement tool in KYPIPE as a simple tool to aid small utilities in sensor placement Execute new KYPIPE sensor placement tool on model database of 12 small water distribution systems for a variety of contamination scenarios Use TEVA-SPOT to run sensor placement simulations on the same models and contamination scenarios Compare results given by KYPIPE and TEVA-SPOT to verify the effectiveness of the new sensor placement tool The KYPIPE software is already useful in allowing utility managers to gain a better understanding of the flow dynamics and overall behavior of their distribution system. The program is able to complete a hydraulic analysis for any configuration of pipes including hydraulic components such as pumps, valves, fittings, and storage tanks. The program can also execute an extended period simulation (EPS) to account for variation over time such as changes in storage tanks levels and varying pump schedules. The Water Quality (WQ) sensor placement tool can be used within the KYPIPE program to offer helpful information and recommendations for water quality sensor placement. The tool will recommend optimal locations for online sensors based on simple water quality analyses and heuristic methods that require very little or no added input from utilities. The tool has been developed to work with the existing KYPIPE graphical user interface. The goal is to provide a simple tool to aid utility managers in the optimal placement of sensors within their distribution systems. Optimal placement of water quality sensors will allow contamination events to be quickly detected, minimizing the negative events of a contamination event. 3

12 The newly developed sensor placement tool is tested on a database of 12 distribution system models that are considered small utilities. The tool is executed with a variety of contamination scenarios for all systems, placing a number of sensors reasonable for the budget of a small utility. The same scenarios are also executed using the TEVA-SPOT software, both programs operating with the objective of minimizing the time required to detect the contaminant based on the given contaminations scenario. The results from both programs for all executions are compared in order to verify the effectiveness of the new KYPIPE sensor placement tool. This research provides the foundation for future work in developing sensor placement guidance. The recommended sensor locations from the KYPIPE sensor placement tool can be analyzed to determine if patterns exist based on system characteristics. If trends in the optimal sensor locations can be observed, guidance can be developed to offer small utilities assistance in placing water quality sensors. Sensor placement guidance without the need for a calibrated hydraulic model can be beneficial to the limited resources and budget of a small utility. 1.3 Content of Report Chapter 2 presents a technical background on topics to support the contents of this research. These topics include the methodology behind TEVA-SPOT and KYPIPE and current research in the area of sensor placement guidance. Chapter 3 presents the database of water distribution system models created to support this research. A general overview of model development is presented, along with details of all systems used in the model database. Chapter 4 presents the KYPIPE sensor placement tool created to provide simple sensor placement guidance to small utilities. The chapter outlines the methodology behind the sensor placement tool and results from the execution of both TEVA-SPOT and KYPIPE on the systems in the model database. The sensor placement results using the two programs are compared to verify the effectiveness of the new tool in KYPIPE. Chapter 5 presents conclusions of this research, including an overall summary of the sensor placement tool and results found using the tool in KYPIPE. Several appendix sections are also included in this report. Appendix A presents visual tools to aid utilities in identifying the system configuration of their water distribution network 4

13 and in executing the sensor placement tool. Appendix B outlines a procedure for developing models of distribution systems in KYPIPE. Appendix C presents a detailed layout of each system in the model database. Appendix D outlines the procedure for executing the sensor placement tool in KYPIPE, and Appendix E contains results for all simulations run with the sensor placement tool. 5

14 CHAPTER 2 2 Sensor Placement in Water Distribution Systems 2.1 TEVA-SPOT TEVA-SPOT Methodology With the increased risk of contamination of water distribution systems through intentional terrorist activities or accidental occurrences, a methodology was needed that was able to effectively reflect the vulnerabilities of such systems to all forms of contamination. To meet these needs, the Threat Ensemble Vulnerability Assessment Sensor Placement Optimization Tool (TEVA-SPOT) Program was developed as a probabilistic framework for analyzing the vulnerability of drinking water distribution networks (Murray et al., 2004). This collection of software tools to aid utilities in the design of sensor networks was developed by researchers from the Environmental Protection Agency (EPA), Sandia National Laboratories, the University of Cincinnati, and Argonne National Laboratory (Murray et al., 2010). TEVA-SPOT creates a threat ensemble, consisting of a set of contamination scenarios, and the vulnerability of the network is assessed using the entire threat ensemble (Murray et al., 2004). TEVA-SPOT contains three main software modules. The first module simulates the set of incidents in the threat ensemble, the second module calculates the potential consequences of each incident in the threat ensemble, and the third module optimizes for sensor placement. The design basis threat consists of the set of incidents for the sensor network to detect. The consequences are calculated based on one or more of the performance objectives (people made ill, length of pipe contaminated, etc.). When TEVA places sensors, the mean consequence for a given objective is minimized. The mean consequences are averaged over the ensemble of contamination incidents. Minimizing the value is equivalent to assuming that each contamination scenario is equally likely to occur and that each is important when selecting sensor locations. The user is able to specify weights to put more importance on locations with a higher likelihood of being contaminated (Murray et al., 2010). A flow chart of the TEVA-SPOT software is shown in Figure 1. 6

15 Figure 1: Flow Chart of TEVA-SPOT Software (Murray et al., 2010) TEVA uses simulation and optimization models to select optimal placement of sensors for a contamination warning system by implementing two steps: modeling and decisionmaking. The modeling process first involves creating a network model for a hydraulic and water quality analysis. The modeling process also must include the following steps: describing sensor characteristics, defining the design basis threat, setting up performance measures, defining utility response to sensor detection events, and finally identifying potential sensor locations (Murray et al., 2006). The decision making process uses an optimization method and evaluates sensor placement; this step is performed by analyzing trade-offs and comparing a set of designs to account for modeling and data uncertainties (Murray et al., 2008). The first step in the modeling process, developing a network model as input to a hydraulic and water quality modeling software, is critical. In many cases, system models are developed by utilities to aid in planning, designing new components, and fixing water quality or hydraulic problems. Using models for the purpose of contamination warning systems requires a high degree of accuracy; models should be up-to-date and include all network components, also ensuring they are accurately represented. Characteristics of the 7

16 sensor behavior are also needed to measure performance of a CWS, so the sensor type, detection limit, and accuracy should be included (Murray et al., 2008). An assumption commonly used is an ideal or perfect sensor that is 100 percent reliable. This assumption in unrealistic, but it is useful in determining an upper bound on performance. A more realistic input is to assume a detection limit for sensors. The sensor is 100 percent reliable detecting a contaminant above a certain concentration, and the sensor will fail to detect the contaminant if it is below the concentration (Murray et al., 2006). The design basis threat describes the type of threat that the utility wants to protect against when designing a contamination warning system. Contamination incidents are described by the specific contaminant, the quantity and duration of injection of the contaminant, and the locations where the contaminant is introduced. The program understands that these conditions cannot be known before an incident occurs, so the modeling process takes this uncertainty into account. The program automatically assumes that each possible injection location is equally likely to be used to inject the contaminant. An ensemble of incidents is then simulated, and sensor network designs are chosen based off how they perform for the entire ensemble of incidents (Murray et al., 2008). TEVA measures performance of sensor network designs based on minimizing certain performance objectives such as the number of people who become ill from exposure to a contaminant, percentage of incidents detected, time to detection, or length of pipe contaminated. Other objectives like costs or economic impacts can also be used. If a utility has several important priorities for the performance of their sensor design, multiple objectives can be considered by assigning a relative importance, or weight, to each objective (Murray et al., 2010). Modeling the utility response to contamination events is another important aspect of the modeling process. Response time is defined as the time between initial detection of the contaminant and effective warning of the population. This time includes credibility of the detection, verification of the contaminant presence, and public warning, and this time period is usually considered to be between 0 and 48 hours (Murray et al., 2010). 8

17 When selecting nodes for potential sensor locations, certain requirements are needed such as accessibility, security, and protection from the environment. Obvious locations that satisfy all requirements are utility owned locations such as pumping stations, tanks, valve stations, etc. However, other locations could be easily adapted to meet requirements, such as fire/police stations, schools, city buildings, etc. Even consumer connections could be adapted to meet sensor location needs, although securing access to private homes or businesses could be problematic. However, a longer list of feasible sensor sites results in a sensor design that is more likely to perform well. So the benefits of using sites that need some adaptation to meet requirements may be worth the additional costs (Murray et al., 2008). The second main step in the TEVA sensor placement framework is the decision process. The goal of this step is to aid utilities in understanding the public health and cost tradeoffs between different sensor placement designs and ultimately help them choose the sensor design that will best meet their needs. This is accomplished by using an incremental approach for applying optimization to generate a set of sensor placement designs. The first sensor placement design is found under ideal conditions with simplifying conditions. The assumptions are then removed one at a time to make the designs more realistic. After every iteration, the performance of the new sensor design is compared with the previous designs to understand what has been gained or lost with each assumption (Murray et al., 2010). The decision making step uses the contamination warning system model to evaluate a series of sensor network designs in a systematic way (Murray et al., 2008). The TEVA-SPOT software allows a utility to achieve objectives important to their needs by optimizing the contamination warning system. Sensor locations are chosen based on the given performance characteristics, likely utility response times, and performance measures important to the utility. The sensor locations will be designed to protect against a variety of contamination threats, and the locations can be restricted to locations preferred by the utility (Murray et al., 2008) TEVA-SPOT Case Study The Battle of the Water Sensors Network (BWSN) was held at the Eighth Annual Water Distribution Systems Analysis Symposium in Cincinnati, OH on August 27-29, The 9

18 goal of the BWSN was to compare the performance of 15 sensor network designs that were applied to two water distribution systems. The sensor network designs were evaluated with four design objectives: expected time of detection, population affected, consumption of contaminated water, and detection likelihood. The expected time to detection, the elapsed time from the start of contamination to the first identified presence of a nonzero contaminant concentration, will be the focus of this case study (Ostfeld et al., 2008). Participants were asked to provide designs for selecting the optimal placement of five sensors and 20 sensors for a base case and three derivative cases. This case study will strictly focus on the placement of five sensors for the base case. Some characteristics of the base case include an injection flowrate of 125 L/h, contaminant concentration of 230,000 mg/l, and injection duration of 2 hours. Each contamination scenario utilized a single injection location, occurring at any injection node with equal probability. Sensors were able to instantly detect nonzero contaminant concentrations and actions were taken to eliminate further exposure with no delay (Ostfeld et al., 2008). 15 different sensor designs were examined at the BWSN. Berry et al. submitted a p-median formulation to define the sensor placement problem, and this was further solved using a heuristic method. This method reflects the procedure executed in TEVA-SPOT. This sensor placement algorithm, along with the 14 other methods submitted to the BWSN, was tested with two water distribution systems. Network 1 consisted of 126 nodes, two tanks, 168 pipes, two pumps, and one constant head source. Network 2 was comprised of 12,523 nodes, two tanks, 14,822 pipes, four pumps, and two constant head sources. During the BWSN, it was determined that the sensor placement problem was actually multi-objective, so a unique single optimal solution could not be identified (Ostfeld et al., 2008). However, each algorithm was still evaluated with the four objective functions, and the results are outlined. Figure 2 displays the selected sensor locations using all 15 methods for Network 1, Case A (time to detection), and placement of five sensors. The method used in TEVA- SPOT is labeled using the red triangle (also labeled Berry et al.). Figure 2 shows that three of the five sensors selected by Berry et al. were common to at least three other methods. 10

19 Figure 2: Sensor Placement at BWSN Network 1 (Ostfeld et al., 2008) It was desired to minimize the time to detection while selecting optimal sensor locations, so the time to detection for the chosen sensors were compared among all 15 sensor designs. Figure 3 shows the first three objectives (all compared to the detection likelihood) from the 15 sensor designs for the Network 1, five sensor scenario. The objective labeled Z1 represents the time to detection objective. The method used in TEVA-SPOT, labeled as point 1, is in the lower half of all 15 sensor designs in terms of the goal to minimize the first three objectives (time to detection, population affected, and consumption of contaminated water). 11

20 Figure 3: Objective Comparisons Network 1 (Ostfeld et al., 2008) The same procedure was also executed to compare sensor designs for Network 2. Optimal sensor locations were chosen using all 15 design algorithms, and the results were compared in terms of the design objectives. Figure 4 shows the comparisons of the first three objectives (also in relation to the detection likelihood objective) for Network 2 and the placement of five sensors. Because it is desired to minimize the first three objectives, Figure 4 shows that the algorithm developed by Berry et al. performs very favorably. The method used in TEVA-SPOT resulted in the lowest values for the first three objectives, although it is equivalent to results using the method proposed by Krause et al. These results prove that the method developed by Berry et al. is effective in selecting sensor locations that minimize time to detection, population affected, and consumption of contaminated water. 12

21 Figure 4: Objective Comparisons Network 2 (Ostfeld et al., 2008) Although these results show the effectiveness of the sensor placement algorithm proposed by Berry et al., an important point should also be noted. In a water distribution system, there are many injection points on the exterior of the system that lead to small pipes that do not propagate throughout the system. If a contaminant is injected at these nodes, it may never be detected. If optimization is performed to minimize time to detection taking into account non-detections, it is desired to avoid non-detections as much as possible. This results in placement of sensors far from the center of the network in order to maximize detection, which is opposite the intuitive approach for minimizing time to detection. To avoid this non-intuitive behavior, it was decided at the BWSN not to include nondetections when looking at the time to detection objective. 13

22 2.2 KYPIPE KYPIPE Methodology KYPIPE was first developed to calculate steady state flows and pressures in a water distribution system. The program is able to complete an analysis for any configuration of pipes including hydraulic components such as pumps, valves, fittings with significant head losses, flow meters, and storage tanks. The program also has the capabilities to execute an extended period simulation (EPS). This extended simulation can account for the variation in storage tanks levels over time, even controlling the open-closed status of a pipe or pump based on the water level in a tank or hydraulic grade at another location. In addition to calculating the flow in each pipe and pressure at each node for a certain operating condition, KYPIPE can also calculate certain design, operation, and calibration parameters to meet specified pressure requirements. Some of these parameters include pump speed/power, HGL setting for a storage tank, control valve settings, pipe diameters, etc. (Wood, 2010). In order to execute a hydraulic simulation of a system, the user must enter data to describe the pipes, pumps, minor loss components, valves, storage tanks, pressure switches, etc. For example, parameters that must be input for all pipes include length, inside diameter, and roughness coefficient. KYPIPE then analyzes the distribution system by solving the set of mass continuity and energy equations using linearization methods to handle the nonlinear terms. After extensive testing of different algorithms for analysis of pipe systems, the creators of KYPIPE determined that the approach used in the program is the most powerful and has the best convergence of all approaches (Wood, 2010). The methods used by KYPIPE to solve water distribution systems are outlined in this section. The relationship between the number of pipes, junction nodes, fixed-grade nodes, and loops can be helpful in understanding the methods of solving water distribution systems. Equation 1 is held true for a distribution system, and it can be related to the basic hydraulic equations to describe steady state flow in a system (Wood, 1981). p j l f 1 (1) 14

23 where p represents the number of pipes, j is the number of junction nodes, l signifies the number of primary loops, and f is the number of fixed-grade nodes. Nodes are divided into two different classifications, junction nodes and fixed-grade nodes. Junction nodes are defined as connections of two or more pipes or a location where flow is removed or input to the system. A fixed-grade node is a node where the values for pressure and elevation are fixed. Reservoirs, tanks, and large constant pressure mains are classified as fixed-grade nodes (Mays and Tung, 2002). Equations used in a hydraulic analysis of a water distribution system are expressed in two different forms. Loop equations show mass continuity and energy conservation in relation to the flow in each pipe section. The other form, node equations, expresses mass continuity in terms of grades at junction nodes. The KYPIPE program utilizes loop equations when performing hydraulic simulations; it has been shown that loop equation have superior convergence behavior over node equations (Wood, 1981). Therefore, the loop equations will be focused on in this section. The information provided in Equation 1 aids in creating continuity and energy conservation equations. A continuity equation is created for each junction node in this system, resulting in j continuity equations. These equations, shown in Equation 2, represent the conservation of mass at a junction node (Wood, 1981). where in Qout Q in represents the flow into the junction, Q Q (2) e Q out is the flow out of the node, and Q e represents the external inflow or demand at the junction node. The conservation of mass concept requires that the sum of mass flows in all pipes entering a junction must equal the sum of mass flows leaving the junction (Panguluri et al., 2005). Conservation of energy equations are also created for each independent loop and pathway between fixed-grade nodes. If the number of pipes, junctions, and fixed-grade nodes are known, Equation 1 can be used to determine the number of loops in the system. A loop is an independent closed path in the system, and these can also be identified by visual inspection for a simple network. For each loop, the conservation of energy must be true, meaning the energy 15

24 gained from pumps subtracted from the sum of energy or head losses around the loop must equal zero (Mays and Tung, 2002). The concept is shown in Equation 3. where L E p h L is the energy loss in each pipe and h 0 (3) E P is the added energy from pumps. If no pumps are present in the loop, the energy equation simply states that the sum of head losses around the loop equals zero. For a system with f fixed-grade nodes, there will be f-1 energy conservation equations for pathways between two fixed-grade nodes. The energy conservation equation for paths between fixed-grade nodes is shown in Equation 4 (Mays and Tung, 2002). E h L E (4) where ΔE represents the difference in total grade between the two fixed-grade nodes. To select the path between the fixed-grade nodes, any connected path of pipes between the fixed-grade nodes can be chosen. However, it is important to avoid redundant paths. The continuity and energy equations together create a set of p (number of pipes) loop equations that are simultaneous nonlinear algebraic equations. An analysis based on the loop equations requires a set of equations for the flow in all pipes. Therefore, the loop equations are expressed in terms of the flow in each pipe (Wood, 1981). The head losses in the pipes, minor losses, h LM p h L, represent the sum of losses in the pipe, h LP, along with. The head loss in the pipes is primarily attributed to friction along the pipe walls and turbulence (Panguluri et al., 2005). The head loss for flow in the pipes is shown in Equation 5 (Wood, 1981). where n hlp K PQ (5) K P represents a loss constant that is a function of pipe length, pipe diameter, and roughness. Several empirical equations have been developed to represent these losses in a pipe, with the Darcy-Weisbach, Hazen-Williams, and Manning equations being common forms. All three of these equations relate head or friction loss to the velocity, length of pipe, pipe diameter, and pipe roughness (Panguluri et al., 2005). Although KYPIPE has the 16

25 capabilities to perform calculations using the Darcy-Weisbach equations, the Hazen- Williams equation is most commonly used and will be outlined. The energy loss for flow in a pipe based on the Hazen-Williams equation using English units is shown in Equation 6 and Equation 7 (Mays and Tung, 2002). h LP 4.73LQ C D K Q P (6) 4.73L K P (7) C D where L is the pipe length (in feet), Q is the flow rate (in cubic feet per second), C is the Hazen-Williams roughness coefficient, and D is the pipe diameter (in feet). The roughness coefficient is a function of the pipe material and age of the pipe. Because the total head loss will be dependent on the flow in the pipe, it is desired to first determine the loss constant, K P, for each pipe. The minor loss in a pipe, h LM, takes into account losses at fittings, valves, expansions, contractions, meters, and other components that may disturb flow in the pipe. Like the equation for friction head loss in a pipe, the minor losses experienced in the system are dependent on the flow in a pipe. Therefore, the minor loss coefficient, K M, is first calculated to quantify minor losses independent of flowrates. These equations are shown in Equation 8 and Equation 9 (Mays and Tung, 2002). h 2 LM K MQ (8) K M M 2 (9) 2gA where the M term represents the sum of minor loss coefficients for fittings present in the pipe section, and A is the cross-sectional area of the pipe in square feet. Pumps in water distribution systems can be described with a constant power input, pump curve, etc. For purposes of this study, the pump energy, E, will be expressed by the simple expressions shown in Equation 10 (Wood, 1981). P 17

26 E p Z P( Q) (10) Q where the energy provided by a pump, E p, is a function of the flowrate and a pump constant, Z. The variable Z is calculated by dividing the useful horsepower of a constant power pump (multiplied by 550) by the specific weight of water. After developing equations for the head losses in a pipe and energy added by a pump, the energy equation shown in Equation 4 can be expressed in terms of loss coefficients and the flowrate in the pipe. This new equation is shown in Equation E ( K P Q K MQ ) P( Q) (11) The continuity equation and energy equations create a set of p simultaneous equations that are collectively called the loop equations. These equations are expressed in terms of the unknown flowrates in pipes. Because these equations represent nonlinear algebraic relationships, a direct solution is not possible (Wood, 1981). An iterative solution must be utilized to solve these equations for the unknown flowrates, and the algorithm utilized by KYPIPE is outlined below. A gradient method is used to deal with the nonlinear flowrate terms in the energy equation. The right hand side of Equation 11 represents the grade difference across a pipe section with a given flowrate Q. The grade difference, expressed at an approximate value of the flowrate Q Qi, is shown in Equation 12 (Wood, 2010). H i K Q K Q P( Q ) (12) p i m i i where H i is the grade difference. The gradient function evaluated at the approximate flow Qi is shown in Equation 13 (Wood, 2010). G i K Q 2K Q P'( Q ) (13) P i M i i where G i is the gradient function. The linear theory method, the methodology used in KYPIPE, is based on a simultaneous solution of basic hydraulic equations for the distribution system. Because the energy equations are nonlinear, the first step in the linear 18

27 theory method is to linearize the equations with respect to the approximate flow, 19 Q i, in each pipe. This step is accomplished by taking the derivative of the variables in the equation for the grade difference (Equation 12) with respect to the flow and evaluating the results at Q Qi. The following linearized equation shown in Equation 14 is the result when this relationship is used with the energy equation (Wood, 2010). iq GiQi G H ) (14) ( i This equation represents each pipe in the given loop, and it is used to create l f 1 energy equations that are combined with j continuity equations to create a set of p simultaneous linear equations that are in terms of the flow in each pipe. An initial value is estimated ( 4 ft / s ) for the velocity in each pipe, and the linearized equations are solved utilizing matrix procedures. The new set of flowrates is used to linearize the equation again, producing an updated second solution. This procedure is repeated until the change in flowrates between trials is very small (Wood, 1981). Convergence is expected since all flowrates are computed simultaneously, and a final solution should occur quickly compared to other methods. Even for an analysis of a large distribution system, convergence should occur for a high degree of accuracy after only four to eight trials (Wood, 2010) KYPIPE Case Study A study by Wood and Rayes (1981) compared five different methods for solving for the unknown flowrates in water distribution systems, and the linear method used in KYPIPE was included in this study. Hardy Cross developed the original method for solving systems, which solves the loop equations based on adjusting flowrates to individually balance each of the energy equations (Wood and Rayes, 1981). The flow adjustment factor calculated for the pipes in each energy equation is applied to all pipes in each path, and improved solutions are used for each trial until the correction factor is within a desired limit. Although it is not as widely used, Hardy Cross also proposed a method to solve the node equations by adjusting the head at each node to balance the continuity equation. These methods are known as the Hardy Cross method, or the single path adjustment method (using loop equations) and single node adjustment method (utilizing node

28 equations). The Hardy Cross method requires an initial set of flowrates that balance the continuity equation at every node, and convergence is partially dependent on how close the initial set of flowrates is to the actual solution (Wood, 1981). It was noted that this method sometimes led to convergence problems. This issue, along with the limitations discussed, resulted in researchers developing new methods for solving distribution systems. The simultaneous path adjustment method simultaneously adjusts the flowrate in each path of pipes for an energy equation. An initial set of flowrates that satisfy continuity equations is first required. A flow adjustment factor is simultaneously calculated for each path that satisfies the energy equation without disrupting the continuity constraint. Improved solutions are then used until the flow adjustment factor is within a desired limit. The simultaneous node adjustment method was also proposed. This method utilizes a simultaneous solution of the node equations and requires a linearization of the node equations in respect to approximate values of the head. These simultaneous linear equations are solved by first assuming a set of junction node heads. The heads are used to linearize the node equations, and this is repeated until convergence is satisfactory (Wood and Rayes, 1981). The last method included in the study by Wood and Rayes (1981) is the linear method. This method is used in the KYPIPE program, and was described in detail in Section These five algorithms were tested on 30 water distribution systems with less than 100 pipes and 21 systems with over 100 pipes (maximum of 509 pipes). Different situations were examined for each system that included minor changes in system characteristics; 60 situations were examined for systems with less than 100 pipes, and 31 situations were analyzed for the systems with over 100 pipes. Initial assumptions for flowrates were assumed where necessary, and trials for each method were executed until the average change in flowrates between successive trials was less than 0.5 percent. For the loop equations, accuracy was measured by the unbalanced heads for the energy equations. The unbalance in continuity was used to quantify solution accuracy for methods using the node equations (Wood and Rayes, 1981). All five methods were compared to exact solutions for the systems under 100 pipes. The exact solution was found by executing the linear method one additional trial after the 20

29 relative accuracy of was achieved. The flowrates and heads at junction nodes were compared to the exact solutions for each method in the study. After observing results from the study, the linear method proved to be very reliable. It converged for every situation and system, and the number of trials required to reach the solution averaged at 6.4 trials (for the 31 systems with over 100 pipes). The simultaneous path method also had good convergence characteristics, and only one failure occurred. Failure occurs when the relative accuracy is reached but the average percent deviation based on flowrate and head range from the actual solution exceeded 10 percent and the maximum percent deviations exceeded 30 percent, or the maximum number of trials is executed without reaching the desired accuracy. This method requires an average of 8.5 trials to reach the desired accuracy (Wood and Rayes, 1981). The other three methods tested did show problems with convergence. The frequency of problems increased with the larger systems, but only the results for the systems less than 100 pipes were presented. Eight failures were observed for the single path adjustment (Hardy Cross) method. All eight cases reached the specified accuracy, and six had a relatively low unbalanced head. Five of the eight failures improved with additional trials, but three failures still occurred. Two of these failures recorded large unbalanced heads, but one failure seemed to reach a good solution with a high degree of relative accuracy and a low unbalanced head. However, the flowrates had significant error from the actual results. Wood and Rayes believe this error occurred because some pipes with significantly high or low head losses were included in the same energy equations (Wood and Rayes, 1981). The simultaneous node adjustment method resulted in a total of 18 failures, and the patterns and characteristics of the failures varied greatly. Some failures had an accurate calculation of heads, but large errors in the flowrates. Other failures never reached the desired accuracy regardless of the number of trials executed. The single node adjustment (Hardy Cross) method was the least reliable of all five algorithms investigated. The specified accuracy was achieved in many cases, but the solutions had large errors in the flowrates. A total of 51 failures occurred. When the specified accuracy was changed from to , a satisfactory solution was found for all but two situations. Even though 21

30 most solutions did improve, convergence was slow and it was difficult to ensure that the solution was satisfactory. This Wood and Rayes study (1981) showed that significant convergence problems occurred using the single path and node adjustment (Hardy Cross) methods and simultaneous node adjustment method. The single adjustment methods must be executed to a greater accuracy to increase the likelihood of finding a satisfactory solution. Strict convergence requirements for unbalanced head (path method) and unbalanced flow (node method) may also improve reliability, but this does not guarantee accuracy in the final results. The study also showed that the simultaneous node method proved to be unreliable. Results showed that if convergence was not met in a reasonable number of trials (40 in this particular study), additional trials usually did not improve the results. These three methods that displayed issues with convergence all require an initial set of flowrates or heads. The probability of failure can be reduced if initial guesses for flow and head are close to actual results. However, a good set of initial values still does not guarantee convergence (Wood and Rayes, 1981). The simultaneous path and linear methods both proved to have good convergence characteristics. A relative flow accuracy of is adequate for finding accurate flows and heads. Because gradient methods are used to deal with the nonlinear terms in the energy equations, convergence issues are always possible. However, convergence problems are unlikely while using the linear or simultaneous path methods. Although both methods are desired when solving distribution systems, this study showed that the linear method is slightly more accurate because it resulted in zero failures while the simultaneous path method led to one failure (Wood and Rayes, 1981). Because the KYPIPE program utilizes the linear method to solve for the unknown flowrates in the pipes (and pressure at the nodes), it can be assumed that the program is reliable for providing an accurate hydraulic analysis of a system. A reliable hydraulic analysis is the first step in providing an accurate tool for determining optimal sensor placement. 2.3 Current Trends in Sensor Placement As previously discussed, the placement of sensors in water distribution systems has become a critical component of contamination warning systems. Many optimization 22

31 methods for placement of sensors require an understanding of flow dynamics and general behavior in the system. This behavior can be estimated by using a simulation-based analysis utilizing calibrated hydraulic and water quality models of the network. However, water quality and hydraulic models require significant expertise and calibration to produce an effective model. Many water utilities, especially utilities serving small populations, do not have the resources to build effective models of their system. Even if a model exists, the execution of sensor placement software, along with the computational requirements, can be problematic. Placing sensors based only on hydraulic models negates the need for information about contaminant behavior, but still requires the need for a hydraulic model (Xu et al., 2008). Not only is the development of accurate system models potentially problematic, but optimization methods used to determine sensor placement also have limitations. A number of different optimization methods have been used to address the sensor placement issue, including integer programming (IP) solvers, genetic algorithms, and local search. Other accepted heuristic optimization methods have also been utilized. However, optimization methods can be limited by the performance guarantee for the final solution, the available computer memory, and the runtime required for performing the optimization. IP solvers can guarantee the best sensor placement that minimizes contamination risk, but they often have difficulties solving large applications. Computational runtime and memory space can be problematic using IP solvers. Heuristic optimization methods, such as genetic algorithms and local search methods, usually cannot guarantee that the final solution is optimal. However, these methods can typically find near optimal solutions in a short time (Hart and Murray, 2010). It is desired to develop a simple method of sensor placement guidance to aid small utilities with limited resources in sensor placement. Ideally, sensor placement guidance would avoid the limitations experienced with optimization methods and accomplish guidance without the need for water quality modeling, or even hydraulic models of the distribution system. This section presents various research studies aimed at new methods in optimal sensor placement. 23

32 2.3.1 Betweenness Centrality and Receivability A study by Xu et al. (2008) simplifies the sensor placement problem by applying a graphtheoretic, or network analysis, approach, which eliminates the need for a calibrated water quality model. An undirected graph represents the physical structure of a water distribution network, and it does not require hydraulic information about the system. This helps shed light on identifying structurally important nodes, which may have implications on the optimal placement of sensors. A parameter called betweenness centrality is used to define the centrality of a node in terms of the degree to which the node is located on the shortest path between other sets of nodes. Nodes with high betweenness centrality lie on the path of many pairs of other nodes, and these nodes would also be between many potential upstream contamination events and downstream receptor populations. Therefore, the authors argue that nodes with high betweenness centrality would be potential locations for sensors. After observing a set of water distribution networks, it was noted that nodes with high betweenness centrality tend to cluster in the network. This concept of clustering is shown in Figure 5. Figure 5: Clustering of Nodes with High Betweenness Centrality (Xu et al., 2008) Selecting several sensor locations based solely on this parameter would result in clustered sensors and redundant information. Therefore, the concept of dividing the system into a set of exhaustive and mutually exclusive communities, and then selecting the node with the highest betweenness centrality within each community, was introduced. The sensor 24

33 locations would also be biased towards the downstream nodes to increase the detection likelihood. The concept of dividing the system into communities is displayed in Figure 6. Figure 6: Community Divide in a WDS (Xu et al., 2008) The general procedure of using this methodology to predict optimal sensor placement is as follows. An undirected graph is created to describe the system, and an adjacency matrix (N x N) is created based on the physical structure of the network (N is the number of nodes in the network). For every node in the system, the graph distance between the node and each water source is found, and then the shortest distance is chosen. Next, the number of communities created in the system is set to the number of sensors to be placed. Within each community, a node with a high betweennness centrality and a long graph distance from water sources is selected as the potential sensor location. This is shown in Figure 7. 25

34 Figure 7: Selected Nodes within each Community (Xu et al., 2008) Xu et al. (2008) also utilized the concept of receivability, used to describe the set and number of nodes that have paths to the measured node in a graph. This concept is developed from reachability. The reachability concept says that if there is one or more paths from node i to node j, then node j is reachable from node i and node i is receivable to node j. Receivability is able to measure the capability of a node to detect contamination events; sensors located at nodes with high receivability should detect more contamination events. To maximize the detection likelihood, it is desired to maximize the coverage of the sensors. Therefore, the set of nodes with the highest receivability would maximize coverage in the system. The general procedure of selecting a set of nodes based off non-time constrained receivability is outlined. A dynamic directed graph, which represents the system including flow direction, is developed for the network (again creating an N x N adjacency matrix). The receivability of each node is calculated utilizing a breadth-first search algorithm. The first sensor is placed at the node with the highest receivability, and the nodes covered by this sensor are then removed. The second sensor is placed at the node with the highest receivability among the remaining nodes. This process continues until all desired sensors are placed, or until all nodes in the system are covered. For placing sensors based on time constrained receivability, the process is similar. The values in the adjacency matrix are water travel time instead of a binary 1 or 0 based on whether water 26

35 flows between two nodes during the study period. Sensor placement based on the concept of receivability is shown in Figure 8. Figure 8: Sensor Placement based on Receivability (Xu et al., 2008) Xu et al. (2008) tested their theory by placing sensors in a system based on betweenness centrality, non-time constrained receivability, time constrained receivability, and the exhaustive simulation analysis. Sensor placed was measured on detection time, population at risk, volume of water contaminated, and detection likelihood. Results showed that the exhaustive simulation based approach performed better than timeconstrained receivability, which also performed better than the betweenness centrality approach. However, the differences in performance between the methods were not significant. The comparison between non-time constrained receivability and the simulation showed very similar results for detection likelihood. Therefore, when a utility is not able to develop and calibrate a model of their system, they can effectively use information about the physical structure of the system and the methods presented to identify key nodes for sensor placement. These methods will be close in effectiveness to an exhaustive simulation. Xu et al. (2008) also points out that receivability metrics are useful in educating the utility, formulating a mitigation plan for a contamination event, and developing preventative measures. 27

36 2.3.2 Rule-Based Expert System Another study by Chang et al. (2011) worked to develop a rule-based expert system (RBES) to generate sensor deployment methods without the computational burden typically encountered with optimization methods. The system was constructed using a combination of EPANET and Excel with a goal of addressing the complexity of the system and reducing the computer runtime while achieving the same level of robustness. This RBES utilizes the accessibility rule and complexity rule to achieve these goals. Each rule is used to analyze the system independently, and then the results are used to find a common set of nodes for final sensor locations. The accessibility rule utilizes results from a hydraulic simulation performed in EPANET to determine the flow fraction for nodes in the network. The flow fraction is found with the flow from the main pipeline, a pipe with a larger diameter at each node, and the flow in a secondary pipeline, a pipe with a smaller diameter than the main pipe. A higher flow fraction means that the population density downstream of the node is higher because of the higher baseline demand in the downstream nodes (Chang et al., 2011). Because flow in a pipe is driven by the downstream water demand, the flow fraction can also be assumed as an index used to estimate the percentage of population that could be affected in the case of an unexpected contamination event (Chang et al., 2012a). The flow fraction is calculated for every node with at least one or more secondary pipes connected to it. The accessibility rule is used to rank the nodes from highest to lowest flow fraction in the system, and the design objective of this rule is to maximize flow fraction. The authors argue that deploying sensors upstream will prevent the highly populated downstream area from exposure to the contaminants, and this maximizes the total population being protected while also meeting budget constraints by using a small number of sensors. The complexity rule classifies nodes in the distribution system as inner nodes or path nodes. A path node has one or more pipes connected to the main pipe (junction with three or more pipes connected to it), and an inner node is located between two path nodes (maximum of two pipes connected at the junction). Figure 9 illustrates this concept. 28

37 Figure 9: Inner Nodes and Path Nodes (Chang et al., 2011) The complexity rule proceeds to determine the number of inner nodes with a hydraulic connection to the path node systematically (Chang et al., 2011). The complexity rule works to deconstruct the node structure configuration to account for a larger population that could possibly be affected by a contamination event, eliminating the need to consider temporal variability (Chang et al., 2012a). An effective radius for each path node is calculated by finding the summation of all pipe distances from a path node to each inner node, and then this value is divided by the number of inner nodes for each path node in a system. The paths nodes are then ranked from the highest number of inner nodes to the lowest, and optimal sensor locations are selected as path nodes with the highest number of inner nodes. In other words, the goal of the complexity rule is to determine the number of path nodes with the maximum combined number of inner nodes based on the path nodes impact zone. The impact zone is found by averaging the distance from all inner nodes with a hydraulic connection to the path node. The path nodes are ranked based on the number of inner nodes located in the impact zone, where the path nodes with the highest number of inner nodes are considered as sensor locations. The optimal sensor location is ideally found by using both the complexity and accessibility rules simultaneously. To measure the performance of the RBES, the strategies were implemented on models used in The Battle of the Water Sensor Network (BWSN): A Design Challenge for Engineers and Algorithms, held as part of the Annual Water Distribution Systems Analysis (WDSA) in Cincinnati, OH in August The RBES was compared to the 14 optimization and heuristic models presented at the BWSN for four design objectives. The 29

38 RBES outperforms over half of the optimization and heuristic models in terms of the time to detection, population affected, and consumption of contaminant objective. The two rules developed with RBES, derived to address effectiveness and efficiency required for sensor deployment, can compete with most of the optimization models (Chang et al., 2011) Rule-Based Decision Support System Chang et al. (2012a) also expanded this concept to a rule-based decision support system (RBDSS), which utilizes the same complexity and accessibility rules. Each rule follows the same concepts previously outlined, and each rule is applied independently in sequence to produce two sets of sensors. Further selection is carried out according to half of the number of sensors from each independent set. Using an equally weighted optimization method instead of using both rules simultaneously to generate a single set of sensor locations will improve the probability of detection. If a system can afford to place six sensors, the accessibility rule should be used to generate three locations, and the complexity rule should be implemented to place the remaining three sensors. The RBDSS expands the node classification concept to derive an effective radius. This improved complexity rule was developed to adjust for a large-scale network with a large number of inner nodes, it can also be used to improve analysis of small systems. The improved complexity rule will cause sensor locations to be closer to highly populated areas and improve performance with design objectives. To find the effective radius for each node in the system, the distances from the pipe connecting the node of interest to its hydraulically connected neighbors in all directions were calculated. The number of nodes within the effective radius is counted. After finding all combined inner nodes and path nodes, the nodes are ranked in descending order based on the inner nodes and path nodes counted (Chang et al., 2012a). The RBDSS was tested with Network 2 used in the BWSN, and it was compared against 11 models and algorithms for four objectives. The results of RBDSS proved to be competitive in comparison; it outperformed several other models for all four objectives. RBDSS is also advantageous because it produces one set of sensor locations that 30

39 represent an optimal solution and it can be executed with inexpensive software such as Excel (Chang et al., 2012b). Further work by Chang et al. (2012b) expanded the rule-based decision support system (RBDSS) to include the intensity rule, along with the accessibility and complexity rules. The improved RBDSS aims to generate near-optimal sensor deployment strategies with low computational burden with the addition of the intensity rule to accompany the two rules previously discussed. The system was also designed to minimize the total number of sensors needed and maximize the monitoring coverage in order to increase the cost effectiveness of a sensor deployment system. The intensity rule focuses on the concentration of contaminants in the system. Because the maximum contaminant levels (MCLs) in a network are regulated by the U.S. Environmental Protection Agency (EPA), the intensity rule was analyzed before the accessibility or complexity rules. The intensity rule uses information of population exposure, so EPANET was used to complete the vulnerability assessment. The main goal of this rule is to ensure that the concentration of potential contaminants such as microorganisms, disinfection by-products, disinfectants, inorganic chemicals, etc. remain under MCLs. The intensity rule can be utilized with many chemical species of concern to prevent public harm relating to accidental or intentional contamination events. The first-order decay was used in the chlorine residual analysis in EPANET. Nodes are ranked from highest to lowest based on how much they exceed the MCLs at any point during the day. Nodes that exceed the MCLs are ranked highest, and the top ranked nodes are chosen as sensor locations. For chemicals that have to meet minimum concentration requirements, the goal is to minimize the summation of total concentrations at the nodes that violate minimum standards (Chang et al., 2012b). Applying the intensity rule, along with the previously discusses complexity and accessibility rules, results in near-optimal sensor placement guidance. The general procedure for execution of RBDSS is shown in Figure

40 Figure 10: General Procedure for RBDSS (Chang et al., 2012a). The three rules were tested on the Hardin County No. 1 water distribution system in Elizabethtown, Kentucky. The system is considered a small water distribution network; the capacity of the plant is 2 million gallons per day (MGD). The data was first analyzed by the intensity rule to choose more than 10 possible nodes. These nodes were then evaluated by the accessibility rule to narrow down the possible sensor nodes, and the final sensor nodes were chosen after eliminating more possibilities using the complexity rule. Based on the intensity rule, the location with the highest population density is selected as a sensor location more since higher exposure levels occur along the main pipe and tanks. This was consistent with results of the accessibility and complexity rules, because flow fractions in these areas should be higher and the number of inner nodes should be picked up more often. The authors state that RBDSS is an effective tool that eliminates the impact of changing flow direction in pipes, and it can also be applied to distribution systems elsewhere with any scale (Chang et al., 2012a) Demand and Reachability A study by Isovitsch and VanBriesen (2008) looked at the spatial trends in sensor placement determined by optimization methods. The relationship between sensor location and water demand was also analyzed. The authors caution that they believe sensor placement is most likely dependent on network hydraulics, but the goal of their spatial 32

41 analysis is to improve understanding of sensor network design criteria that would hopefully lead to simplification of design methods. In this study, Geographic Information Systems (GIS) is used to develop a visual inspection of the frequency of sensor placement and analyze the spatial relationships among the sensor locations determined in the BWSN. Further analysis was performed in GIS utilizing the average nearest neighbor and spatial autocorrelation tools. GIS is also used to aid in a chi-square analysis to investigate dependence of sensor location on node attributes such as demand or reachability. The average nearest neighbor (ANN) tool is used to determine the degree of clustering among nodes by measuring the extent to which the spatial distribution of nodes differs from a randomly distributed set. An average nearest neighbor ratio (R) was generated. This value is found by comparing the average distance to the mean nearest neighbor distance for a random distribution. The value used for average distance is found by measuring the Euclidean distance (length of a straight line between two nodes which is not always equal to pipe distance) between each node and its nearest neighboring node and then averaging these measurements. R values that are closer to zero mean the nodes are more clustered. The spatial autocorrelation tool aims to measure the underlying pattern between nodes based on their location. It provides information about how clustered, random, or dispersed the data are. In the study, an analysis was performed on a distribution system using four scenarios and five objectives. Sensor placement was determined using an optimization method that accounted for time to detection, population affected, contaminated water demand, and detection likelihood. Results from the average nearest neighbor analysis showed that sensor locations were clustered (with a less than 1 percent likelihood that the pattern could be the result of random chance), and the first sensors placed were more intensely clustered. The authors hypothesized that average demand may be an effective indicator of optimal sensor placement because population affected and contaminated water consumed were design objectives. EPANET was used to find a value for average demand at every node 33

42 over the 48 hour simulation. The frequency of the networks with 20 sensors placed was compared with average demands, and no relationships were obvious. The frequency analysis was not able to make conclusions for a correlation between sensor placement and average demand. Reachability and reachable average demand was also investigated. Reachability is the number of nodes in the network to which water can flow from the node in question. Reachable average demand represents the total demand for all nodes that are reachable from the node in question. A visual inspection in GIS was used to investigate the relationship between these parameters and sensor placement. There was not an obvious relationship present when looking at all cases and scenarios together. However, when the systems were divided according to objective, some patterns were observed. When examining reachability of selected sensor nodes, the optimal nodes had high reachability for the objectives of expected population affected and contaminated water demand. When sensor selection was based off expected time to detection and detection likelihood, the selected sensor nodes had low reachability. Similar results were observed for average reachable demand. These results make sense because population affected and contaminated water demand are functions of average demand, so optimal sensor locations should have high reachable demand and high reachability. When observing results generated from the detection likelihood objective, sensor locations will be most likely to detect contaminants, disregarding the time to detect these contaminants. Therefore, this would result in sensor placement on the exterior of the system at nodes with low reachability and low average reachable demand. A chi-square analysis was performed to further investigate the relationship between sensor placement and average demand, reachable average demand, and reachability. The chi-square analysis eliminates the effects of overlapping sensor locations. Results of the analysis showed that more sensor nodes than expected had high average demand, although this relationship was not strong. The analysis performed on all cases for all objectives showed no significant dependence of sensor placement on reachability and reachable average demand. However, some dependencies were observed when each different objective function was examined independently. 34

43 A statistically significant dependency was found between sensor placement and high average demand for the objective functions time to detection and detection likelihood. When the objectives for population affected and contaminated water demand were used, a dependency between sensor nodes with high reachability and high reachable demand was noted. Sensors were more likely to be placed at nodes with low reachability and low reachable average demand when using the detection likelihood objective. The results of this study show that using a system attribute like average demand for certain design objectives, specifically time to detection, could be practical for water utilities (Isovitsch and VanBriesen, 2008). Designing sensor placement in a system based solely on one design criteria can be dangerous. If sensor placement design is executed on a system by emphasizing any one design criteria over others, various trials using different objectives could result in a different sensor placement pattern (Aral et al., 2010). The goal of contamination warning systems is to reduce the exposed population to the contaminant and reduce contaminated water volume. It makes sense that this goal would be accomplished by minimizing the time to detection with high reliability (Aral et al., 2010). Therefore, further work in this area will utilize data provided by sensor placement software that recommends sensor placement based on low time to detection. Sensor placement determined by this objective function for numerous contaminant injection scenarios will be used to explore possible trends that could exist between optimal sensor nodes and system parameters. Parameters previously studied in the literature, such as demand, accessibility, complexity, etc. will be further examined. Other network parameters, such as proximity to storage tanks, pipe capacities, and connectivity to surrounding nodes will also be investigated. Further work in this area will focus on developing sensor placement guidance based on easily measurable network characteristics, with the hope that this guidance will be a helpful tool for sensor placement for small utilities. 35

44 CHAPTER 3 3 Water Distribution System Models The main objective of this research is to develop guidance for optimal sensor placement that would be applicable to small utilities. Sensor placement software is used to evaluate optimal sensor placement for a database of water distribution system models. This database consists of 15 models that can be classified as small systems based on the typical service demand of the system. All the systems used in this research average a daily demand between one and three million gallons per day (MGD). The models used in this study were also selected based on their spatial configuration and diversity in general system characteristics. All 15 models can be characterized as grid, loop, or branch spatial configuration. The systems also provide variation in their basic system characteristics such as number and size of pumps, tanks, and reservoirs. The variation in system components was meant to avoid apparent trends in sensor placement of models within a configuration that were actually just caused by similarities in system components. The goal of this study is to develop trends for optimal sensor placement within system configuration that would be applicable to any small utility classified as the given configuration. Using distribution system models that have a variety of characteristics for each configuration will allow the trends to be applicable to a range of systems. 3.1 System Configurations Each model used in this research can be classified as one of the three basic system configurations for water distribution networks: branch, loop, or grid. Figure 11 shows a diagram displaying the basic setup of each system configuration. 36

45 Figure 11: System Configurations (a) Loop, (b) Grid, (c) Branch (taken Gagliardi and Liberatore). A branch system is named for its similarities to a tree branch. Smaller pipes branch off more centralized, larger pipes so that water can theoretically only take one path from the source to customers (National Research Council, 2006). This type of system is frequently used in rural areas where the service area is fairly large, but some consumers in the far branches are spaced far apart from each other. High flows are experienced in the large transmission lines running through the center of the system, and lower flows are present in distribution mains as pipes become smaller farther away from the center of the system. These systems contained more pumps, tanks, and a greater total length of water lines because the systems are more spread out. Even though these systems typically contain a greater total length of pipeline than other configurations, the average diameter of pipes are usually smaller. An example of a system in branch configuration is shown in Figure

46 Figure 12: System in Branch Configuration The branch system is typically easy to distinguish, but the loop and grid systems have similar characteristics and it is sometimes difficult to classify systems into these configurations. Both systems consist of connected loops of pipelines, allowing several pathways that the water can flow from the source to customers. These system configurations are more widely used in large municipal areas or densely populated systems (U.S. Environmental Protection Agency, 2008). Loop and grid systems are considered very reliable because line breaks can be easily isolated, allowing only a small portion of the system to be affected (National Research Council, 2006). Looping is not only advantageous because it provides continuous service even if a portion of the system is shut down, but it also provides flow from multiple directions for reliable fire flow and reduces the number of dead-ends that potentially cause water quality problems (McGhee, 1991). 38

47 In loop systems, there is typically a large, centralized transmission line that feeds smaller lines. The purpose of the central lines is to supply high flows from the source through the middle of the system, and the system then transitions to lower flows as the lines move outward from the central area. These smaller lines connect at each end into the main loop (Gagliardi and Liberatore). An example of a real distribution system classified as a loop configuration is shown in Figure 13. Figure 13: System in Loop Configuration In grid configured systems, the water lines are laid out to look similar to a checkerboard. The main water line infrastructure, that are typically the larger pipes in the system, loop around the outside of the network. The system then transitions to smaller pipes in the interior of the system. Pipe sizes usually decrease as the distance away from the supply source increases (Gagliardi and Liberatore). An example of a distribution system in grid configuration is shown in Figure

48 Figure 14: System in Grid Configuration Many systems are a combination of different configurations (systems containing both looped and branch configurations are common). However, for the purposes of this research, all systems were classified strictly as one configuration based on which configuration characteristics were most prominent. 3.2 General Procedures of Model Development The model database used in this research consisted of 15 models, all representing real distribution systems located in Kentucky. 12 models were used with the sensor placement software, and the remaining three models will be used in the future for verification of the developed trends. The Kentucky Infrastructure Authority (KIA) sponsored the Water Resources Information System (WRIS). This system contains shapefiles representing water 40

49 distribution system components for all utilities established in Kentucky (Kentucky Infrastructure Authority, 2010). The shapefiles for water lines, pumps, tanks, and water treatment plants were downloaded and imported to ArcGIS (Geographic Information System). The files for meters, surface sources, well sources, and purchase sources were also available, but these components were not necessary for the purpose of this project. Each shapefile contained data about a system component for the entire state, and the Owner attribute was utilized to isolate the system components for each individual utility used in the study. Figure 15 displays the entire water line shapefile for Kentucky, along with the water lines of one utility isolated. Figure 15: Water Line Shapefile. The shapefiles acquired from the KIA database do not have elevation data associated with them, which is necessary to perform a hydraulic analysis in KYPIPE. Therefore, digital elevation models (DEM) were necessary to assign elevations to system components. This data was acquired from the National Resources Conservation Service (United States Department of Agriculture). Once shapefiles containing system components and elevation data was acquired, a series of imports and exports of data between GIS and KYPIPE was executed to create a working hydraulic model. The general process of model creation is outlined in Figure 16. A step-by-step procedure for creation of models in KYPIPE is outlined in Appendix B. 41

50 Figure 16: Model Development Procedure Pipe Roughness Coefficients In order to make the system models as representative to the actual distribution system as possible, roughness values were added to all pipes in the system. This allowed the model to account for head loss in the pipes due to friction, and the KYPIPE model uses the Hazen- Williams equation as the basis for head loss. The Hazen-Williams equation is widely used to relate the physical properties and flow parameters of a pipe to the resulting head loss or pressure drop that will occur. A widely used version of the equation in English units is shown in Equation 15 (Mays, 2005). h L L Q (15) C D where h L is the head loss (ft), L is length of pipe (ft), Q represents the flow rate (cfs), C is the Hazen-Williams C-Factor (also known as roughness factor), and D is the diameter of the pipe (ft). The roughness coefficient, C, used in the Hazen-Williams equation varies for pipes based on pipe material and age of the pipe. Different pipe materials will result in varying roughness factors because pipe roughness is dependent on pipe material. Steel and 42

51 PVC pipes tend to be smoother and result in less friction loss than cast iron pipes (AWWA, 2005). The roughness coefficient is also dependent on the age of the pipe. New pipes are typically very smooth and have not yet undergone a great deal of corrosion and deposition, resulting in minimal head loss. After time, the pipes will accumulate deposits and experience tuberculation on the interior of the pipe. This reduces the actual inside diameter of the pipe, causing the actual inside diameter to be less than the expected nominal diameter, which allows less water than expected to flow through the pipe. The accumulation of deposits also causes greater frictional head loss from the increased roughness in the pipe (Walski et al., 2003). In terms of the roughness coefficient, C, used in the Hazen-Williams equation, the frictional head loss experienced in the pipe will increase as the coefficient decreases. Therefore, pipes made out of smoother material, such as PVC, will have higher C coefficients than materials with greater roughness values like cast iron. Similarly, older pipes of the same material that have experienced significant corrosion and deposition will have lower coefficients than new pipes of the same material (AWWA, 2005). If the flow rate remains constant, a smaller roughness coefficient will result in a larger pressure drop in a segment of pipe. Hazen-Williams coefficients were applied to all pipes in the model to estimate head loss and increase accuracy in the model. A reference roughness value based on pipe material was entered to represent the roughness factor of a new pipe. An aged roughness (10 yr) value was also entered to represent the reduced roughness coefficient of the pipe after 10 years. These values were also used to estimate roughness of pipes that were greater than 10 years old. The procedure for applying these values to pipes is outlined in Appendix B Model Demand Input To create a model that is a close reflection of an actual water distribution system, water demand data also needed to be incorporated into the model. The most accurate method of demand distribution would be acquiring meter data from the utility and applying this actual 43

52 demand data to nodes throughout the system. This process was not feasible for creating a database of 15 models, so an estimation of the demand allocation was used. In order to add demand data to the distribution system model, data was acquired for the total average daily demand in the given system. This data was acquired from the WRIS database (Kentucky Infrastructure Authority, 2010), and demand data was provided for total water usage in million gallons per year. This data was converted to million gallons per day, then to gallons per minute (GPM) to match the units in KYPIPE. The Automatic Demand Distribution tool in KYPIPE was used to allocate the total demand to nodes throughout the system. This tool distributes the total demand to nodes in the system based on the diameters of adjacent pipes. It assigns greater values of demand to larger pipes, modeling higher flows in large pipes and lower flows in small pipes. This is fairly representative of how a real system operates, except for the case of large transmission lines. The main purpose of these larger pipes is to transmit water to smaller arterial and distribution mains, which then deliver water to customers. Transmission lines usually do not directly service a high amount of demand (Mays, 2000). However, this process does meet the goal of distributing the total average daily demand throughout the system in the general pattern that smaller pipes will service lower demands. This demand allocation was accurate enough for the purposes of this research. Because water usage in a typical water distribution system has varying water demand patterns throughout the day, it was necessary to investigate demand patterns over a 24 hour period. For example, residential areas will typically have higher demand in the early morning/evening and lower demand during the day when residents are at work. Demand in areas mainly consisting of businesses and industrial plants will reflect the operating hours of the facilities, usually during the day between 8 a.m. and 5 p.m. Although it would be more accurate to develop demand patterns based on water usage types (residential, commercial, or industrial), this data would be very time consuming to acquire so implementing a demand pattern that applied to all parts of the system was sufficient for the purposes of this research. KYPIPE calculated nodal demand by multiplying the stated demand by a demand factor that is defined for every time period in the hydraulic analysis (1 hour). This caused the hourly demand to change from the 44

53 average hourly demand in the system, either by increasing or decreasing the demand at a given hour based on the time of day. The demand factors applied to the models were developed by the American Water Works Association. The factors were less than one during the night when demand would be low and above one during the day when demand is the highest. The highest demand factors were set during the evenings when customers are cooking dinner and showering before bed. This allowed the model to simulate the demand spikes that are experienced during the day in a real distribution system. The procedure for applying demands and demand factors to the models are outlined in Appendix B Final Adjustments to Model Some of the system characteristics used to describe distribution systems in the model database were further modified to simulate the behavior of real systems. Changes were made to create models that operated under reasonable pressure ranges, and this was determined to be between 40 and 150 psi. Problems with low pressures in systems were typically solved by raising initial tank levels (raising minimum and maximum tank levels in some cases), increasing roughness value of pipes, and increasing the power of pumps. In cases of extreme low pressure, pumps were added to the system. To correct high pressures experienced in some systems, initial tank levels were lowered (along with minimum and maximum tank levels in some cases), roughness values for pipes were decreased, the power of some pumps was decreased, and pressure regulators were added in some areas. In some cases of extreme high pressure, pumps were removed completely. This is reasonable because some pumps were necessary in systems to help transmit flows to nearby systems, but these connecting systems nearby were not included in the distribution system model of concern. Control switches were also added to the model to simulate the pump schedules controlled by the utility. Control switches are able to turn pumps on and off based on the level (pressure, head, or HGL) of a certain node in the system. In a typical system, pumps are turned on to fill tanks when they get to a low water level, and this typically occurs at a low demand time. Control switches were applied to certain pumps in the system that had a primary purpose of filling tanks. Pumps present in the system to provide pressure usually 45

54 did not have control switches. These control switches caused the pumps to turn off when the tanks reached a high water level (usually close to the maximum tank level) and turn back on when the tank reached a low level (close to the minimum tank level). These control switches help simulate the real behavior of water distribution systems that is typically controlled and monitored by the utility. The procedure used to implement these changes in KYPIPE is outlined in Appendix B. Ideally, a field calibration would be executed for each water distribution system to create models that accurately represent the actual system. Elevation data for all system components would be verified with surveying, and hydraulic field testing would be executed to determine actual roughness coefficients for the pipes. However, the model calibration process is time consuming and requires a great deal of labor and data collection. For the purposes of this research, a full scale calibration is not necessary. The small changes made to the system to simulate realistic flow conditions are adequate for this study. All changes made to the model database were reasonable alterations that were not unrealistic conditions for small water distribution systems. For example, pipe roughness coefficients were not changed outside of the reasonable range for a given pipe material. This ensures results realistic of an actual distribution system, even if data was slightly altered from that of the actual system. The necessity to make minor changes to the model occurs for two main reasons. First, some characteristics of systems are determined based on the presence of connections to nearby systems. Because any connections to neighboring systems were not included, some aspects were altered to make them suitable for use solely in the system of concern. Also, some data acquired from the KIA database for certain systems was outdated or incorrect. In one case, the values for maximum elevation of water in the tanks were lower than the ground elevation of the tanks. Because this data was clearly incorrect, steps were taken to develop more realistic estimates for these values. 3.3 Description of Models used in Study The model database used in this research consists of 15 hydraulic models. However, only 12 models were used in this phase of the research in testing of the sensor placement tool. The remaining three models will be used in the future for verification of sensor placement 46

55 guidance. The models represent real distribution systems in Kentucky, but all models were given a name in the form KY #. All identifying information for the actual systems represented by the models was removed, such as names of pumps and tanks, to protect the security of the utilities. Model names were grouped by configuration type. The first four models, KY 1 KY 4, along with KY 13 are in the loop configuration. The first four models (KY 1 KY 4) were used to develop sensor placement guidance, while the last model KY 13 will be used to verify the newly developed guidance. The layout of each system in the loop configuration is displayed in Figure 17. A detailed layout of each distribution system model is shown in Appendix C. The schematic of each loop system shows a centralized transmission line running through the center of the system, and smaller water lines deliver water to the exteriors of the system. These smaller lines form looping patterns that provide water multiple pathways to reaching most parts of the system. Figure 17: Systems in Loop Configuration: (A) KY1; (B)KY2; (C) KY3; (D) KY4; (E) KY13 The models KY 5 KY 8, along with KY 14, are classified as models in grid configuration. Similar to the loop configuration models, the first four models (KY 5 KY 47

56 8) were used in the development of sensor placement trends, while KY 14 will only be used in the verification process. The general layout of each system classified in the grid configuration is displayed in Figure 18. The system schematics show main transmission lines looping around the exterior of the system, along with the same looping patterns on smaller distribution lines similar to that of loop configured systems. Figure 18: Systems in Grid Configuration: (A) KY5; (B) KY6; (C) KY8; (D) KY14; (E) KY7 The remaining models, KY 9 KY 12 and KY 15, can be classified as branch configuration systems. KY 9 KY 12 were used with sensor placement software to develop guidelines for sensor placement, and the final model KY 15 will be used to verify sensor placement trends in the future. Figure 19 displays the layout of each model in branch configuration. These systems layouts display how branch systems resemble a tree in that water lines branch out from the main transmission lines. The branch systems are more spread out than other systems, and they contain more tanks and pumps as a 48

57 result of the large area covered. These systems also contain more dead ends as the system branches out, varying from the looping pattern of pipes in the grid and loop systems. Figure 19: Systems in Branch Configuration: (A) KY9; (B) KY10; (C) KY11; (D) KY12; (E) KY15 Even though all system models are classified into three configurations, each system has varying characteristics that distinguish them from other systems within each configuration. A detailed layout of each distribution system model, including all system components, is shown in Appendix C. The data displayed in Table 1 also shows differences in system characteristics among the systems. The data shows a variation in characteristics, including number of tanks, number of pumps, number of reservoirs, total length of water lines, number of nodes, and total system demand. However, there are still trends present based on system configuration classification. For example, the branch systems all have a greater number of tanks and pumps, and they also have greater total length of water lines than the other systems. 49

58 System Name Configuration Numbe r of Tanks Table 1: System Characteristics Number of Pumps Number of Reservoirs Total Length of Pipes (miles) Number of Nodes Total System Demand (MGD) KY 1 Loop KY 2 Loop KY 3 Loop KY 4 Loop KY 5 Grid KY 6 Grid KY 7 Grid KY 8 Grid KY 9 Branch KY 10 Branch KY 11 Branch KY 12 Branch KY 13 Loop KY 14 Grid KY 15 Branch Steady State and EPS Simulations During the model development phase of this research, the models were first run under a steady state simulation. This feature in KYPIPE runs the simulation at time t=0 and does not continue the simulation over an extended period of time. The program uses all initial settings to execute the hydraulic analysis. For example, the initial tank levels are used as the head in the tanks. Alterations were first made to the models to ensure they ran successfully, and within the desired pressure range, for the steady state simulation. An extended period simulation (EPS) was then set up in KYPIPE. The total time of the simulation, computational period, and report period can be specified in the program. For all EPS simulations in this study, the total times were set as 24 hours, and both the computational and report periods were set as one hour. The starting time was set as hour 0. These parameters will run the simulation for 24 hours and output all hydraulic results 50

59 every hour during the 24 hours. The control switches, used to regulate the operation of pumps in the systems based on tank levels, were also added at this point. The control switches caused pumps to turn on when tank levels were low and turn off when tank levels reached high levels. The purpose and settings for control switches are discussed in further detail in Section The EPS calculates hydraulic conditions in the system over the entire computational period at the specified computational periods. The program displays all data provided in the steady state simulation (pressure, HGL, head, demand at all nodes and flow, velocity, and loss at all pipes) at every report period. Hydraulic conditions are also provided at times when a tank is emptied or filled. The KYPIPE program displays data at all nodes and pipes in both tabular and graphical form. The program also provides data for the pressure/hgl/head/flow at all tanks, reservoirs, and pumps in the system. 51

60 CHAPTER 4 4 KYPIPE Tool Sensor Placement Analysis 4.1 Theory The Water Quality sensor placement tool has been developed to work with the existing KYPIPE graphical user interface. The goal is to provide a simple tool to aid utility managers in the optimal placement of sensors within their distribution systems. The tool will recommend optimal locations for online sensors based on simple water quality analyses and methods that require very little or no added input from utilities. The simplicity and ease of use of the sensor placement tool makes it attractive for the use in small utilities. Its enumeration methods used to determine optimal sensor placement also ensure accurate and useful results. The sensor placement tool recommends optimal sensor placement, regardless of how many sensors are implemented, based on minimizing time to detection. The tool considers detection events at nodes throughout the entire system, and recommends optimal sensor placement based on the locations that can detect contamination events the fastest. The enumeration process executed by the sensor placement tool is explained in detail. The tool first reads the INP file (hydraulic network model data file in EPANET format) and makes the necessary changes in order to perform a simple water quality analysis of conservative constituents. The tool then performs water quality simulations, where the injection of a contaminant is dependent on the user defined input. The WQ sensor placement tool is able to recommend optimal sensor placement for up to five sensors, but the explanation on theory will use placement of two sensors. The tool will first consider the first possible combination of two sensors. Next, the first possible injection site in the system will be selected. The contaminant is injected at the injection site, and the travel time for the contaminant to reach each of the sensors is determined. Because the contaminant is considered to be detected when it reaches the first sensor, the minimum of the two travel times is taken. This is used as the travel time for this particular set of possible sensor nodes and injection site (it is not necessary for the contaminant to be detected at both sensors). Detection is based on the detection limit entered in the default parameters option. For this study, a detection limit of 0.01 mg/l was used. When the 52

61 concentration of the contaminant reaches 0.01 mg/l at the particular sensor node, the contaminant is considered to be detected. The tool considers 24 hours as the maximum travel time. Any travel time past 24 hours will be considered 24 hours for calculation purposes. This theory is illustrated Figure 20. The values for T1 and T2 represent the travel times from the injection node to sensor 1 and sensor 2, respectively. The travel time assigned to this particular set of sensor locations and injection location will be 360 minutes, because it is the minimum of the two travel times. The minimum travel time is now considered the travel time from that injection site for the set of two sensor locations. Figure 20: Sensor Placement Tool Theory (Minimum Travel Time) This process is repeated for all possible injection nodes in the system. The minimum of the two travel times for all possible injection nodes is taken for the same set of possible sensors. The average travel time for the particular set of sensor locations is calculated by averaging the minimum travel times from all injection sites. The sum of travel times from all injection sites is calculated and divided by the total number of injection nodes to determine the average travel time for that set of sensor locations. This concept is illustrated in Figure

62 Figure 21: Sensor Placement Tool Theory (Average Travel Time) This process is then repeated for every possible set of sensor locations (in this case, every possible set of two sensors), resulting in an average travel time for every possible combination of two sensors in the system. The sensor combination with the lowest average travel time will be considered the optimal sensor location. When selecting nodes for potential sensor locations, certain requirements are needed such as accessibility, security, and protection from the environment. Obvious locations that satisfy all requirements are utility owned locations such pumping stations, tanks, valve stations, etc. However, other locations could be easily adapted to meet requirements, such as fire/police stations, schools, city buildings, etc. Even customer connections could be adapted to meet sensor location needs, although securing access to private homes or businesses could be problematic. However, a longer list of feasible sensor sites results in a sensor design that is more likely to perform well. So the benefits of using sites that need some adaptation to meet requirements may be worth the additional costs (Murray et al., 2008). Because considering many nodes for potential sensor locations is ideal, the KYPIPE tool considers possible sensor locations to be all nodes (including tanks, pumps, reservoirs, and junctions) except dead-end nodes. The average travel time to dead-end nodes will generally be much higher, skewing the average times to detection. Possible injection sites are considered to be all non-zero demand nodes, excluding dead-end nodes. Dead-end nodes are considered to be consumption nodes, so any contaminant injected at these nodes will be consumed immediately and the contaminant will not be able to travel further in the 54

63 system. The reality of this concept may be slightly different, but this assumption is used in the sensor placement tool. 4.2 Procedure for Execution In order to obtain optimal sensor placement in a water distribution system using the sensor placement tool, a model of the system must first be developed in KYPIPE. The model should accurately depict the structure of the system, and it should include characteristics of all system components such as tanks, pumps, reservoirs, and pipes. The process of model development is outlined in Section 3.2. Once the system runs effectively for a steady state simulation, an extended period simulation (EPS) is set up for the model. For this study, the total time was set to 24 hours, and both the computational period and report period were set to one hour. The EPS needs to be executed on the model. After the analysis is complete, the sensor placement tool is started by using the shortcut Shift + F7. The sensor placement tool open will window, and parameters are entered, starting with the number of sensors to place. Under default parameters, the total simulation time, WQ computational time, mass injection rate, injection start time, injection end time, and detection limit values are also entered. The INP file is generated, and then the sensor placement tool is run. When the run is completed, the tool will output the name of the optimal sensor locations, along with the average travel time for the chosen sensors. The general procedure for execution of the sensor placement tool is outlined in Figure 22. A detailed step-by-step procedure for using the sensor placement tool, along with all parameters used in this study, is outlined in Appendix D. A poster that can be used as a tool to aid utilities in executing the sensor placement tool is included in Appendix A. 55

64 Figure 22: Sensor Placement Tool Flowchart. When the sensor placement tool is executed, an Excel file with the file name systemnametimematrix.csv will be generated. The file includes all data used to determine the sensor location with the lowest time to detection. The first column shows all possible sensor nodes, and the first row of nodes represents the injection nodes. The values show the travel times between the injection and sensor nodes (in minutes). If the cell shows 0 for a travel time, this means that the sensor nodes are too far away from the injection nodes and the contaminant will not reach the sensor within 24 hours. Therefore, the travel time is considered to be 24 hours for calculation purposes. A report of the simulation (text document) can also be accessed. The report provides information about parameters used in the simulation, along with the selected sensors and average travel time in hours. This file, along with the time matrix, can be accessed in the systemname.kyp folder. 4.3 Performance Evaluation Contamination Scenarios The sensor placement tool was executed on the 12 models in the model database for 15 different contamination scenarios. The contamination scenario is determined by both the rate of injection of the contaminant (in mg/min) and the total injection time (in hours). 56

65 Contamination scenarios were created for three different general scenarios: fixed amount, fixed rate, and fixed time. Each general scenario is comprised of five specific sets of an injection rate with a total injection time. The theory behind the three different contamination scenarios is explained. For the fixed amount scenarios, the scenario simulates a drum of contaminant to be injected, and it is desired to inject the entire drum. The pump speed used to inject the contaminant can be varied and unlimited time is available. The fixed rate scenarios simulate an injection pump with a constant speed, so injection rate cannot be varied. However, unlimited time and materials (contaminant) is available. The fixed time scenarios model a limited amount of time available to inject the contaminants, but the pump speed can be varied and supplies are unlimited. The 15 contamination scenarios performed on each model in the model database are displayed in Table 2. Table 2: Contamination Scenarios Amount (Vary Time) Rate (Vary Amount) Time (Vary Rate) Injection Rate (mg/min) Injection Time (hours) Total Contaminant Injected (mg) All 15 contamination scenarios were executed on all water distribution system models using both TEVA-SPOT and KYPIPE. It was desired to compare the sensor placement 57

66 results, both sensor placement and times to detection, between KYPIPE and TEVA-SPOT for a variety of scenarios. To be able to directly compare results from the two sensor placement programs, it was ensured that all parameters matched between the programs. First, the models used in each were identical. The TEVA-SPOT program uses a model input from EPANET. Even though minor differences exist between KYPIPE and EPANET, all major system components and characteristics of these components matched between the two programs. This included pipe roughness values, grade of reservoirs, pump power, pipe diameters, etc. An example of a difference between KYPIPE and EPANET is that KYPIPE allows tanks to be measured as a total volume or fixed diameter, while EPANET only allows a fixed diameter as input for tank size. To make the models as similar as possible, all tanks in both KYPIPE and EPANET were set as fixed diameters. Parameters used in the sensor placement tool in KYPIPE and TEVA-SPOT were also standardized. The WQ computational time (labeled as hydraulic timestep in TEVA-SPOT) were both set to 60 seconds (or one minute in TEVA-SPOT), and the total simulation time was set to 24 hours. The detection limit for both programs was also set to 0.01 mg/l. This ensured one program would not detect the contaminant faster than the other simply because it had a lower detection limit. A study investigating the impact of sensor detection limit on performance showed that a sensor detection limit of 0.01 of the average source concentration was adequate for maximum protection for the example system examined (McKenna et al., 2006) Time to Detection Comparison The baseline contamination scenario (a contaminant injected at 1000 mg/min for four hours) was considered the baseline case because it was present in all three general contamination scenarios. The comparisons between the two programs for the baselines conditions for all models are shown in this section, and the results for all remaining contamination scenarios are included in Appendix E. Table 3 displays the sensor placement simulation results for both KYPIPE and TEVA-SPOT for the baseline contamination scenario. The table displays the time to detection calculated by both programs (KYPIPE outputs time in hours and these were converted to minutes). The column labeled Fastest Time to Detection shows which program resulted in the lowest 58

67 time to detection for that particular system. The last column displays the difference between the higher and lower detection times (in minutes). Table 3: Comparison between KYPIPE and TEVA-SPOT for Baseline Conditions 1 sensor 2 sensor System Time to Detection (min) KYPIPE TEVA- SPOT Fastest Time to Detection Difference (Higher Time - Lower Time) in min KY KYPIPE KY TEVA-SPOT 3.59 KY KYPIPE KY KYPIPE KY TEVA-SPOT 3.75 KY KYPIPE KY TEVA-SPOT KY TEVA-SPOT KY KYPIPE KY KYPIPE 6.22 KY KYPIPE KY KYPIPE KY KYPIPE KY KYPIPE KY KYPIPE KY KYPIPE KY TEVA-SPOT 9.53 KY KYPIPE KY TEVA-SPOT KY TEVA-SPOT 6.87 KY KYPIPE KY KYPIPE KY KYPIPE KY KYPIPE The results in Table 3 show that the sensors selected by KYPIPE led to lower times to detection for most, but not all, system models. For the one sensor scenario, KYPIPE had lower times to detection for eight out of the 12 models. When placing two sensors, KYPIPE produced lower times to detection for nine out of the 12 models. TEVA-SPOT 59

68 calculated lower times to detection for KY 5, KY 7, and KY 8 for both one and two sensors. These results can also be seen in Figure 23 and Figure 24. Figure 23 displays the time to detection comparison for one sensor placement, and Figure 24 shows the same comparisons for placement of two sensors Baseline Conditions (1000 mg/min x 4 hr) - 1 sensor 1200 TEVA-SPOT KYPIPE Time to Detection (min) KY1 KY 2 KY 3 KY 4 KY 5 KY 6 KY 7 KY 8 KY 9 KY 10 KY 11 KY 12 System Figure 23: Comparison between KYPIPE and TEVA-SPOT - Baseline Conditions (1 sensor) 60

69 Baseline Conditions (1000 mg/min x 4 hr) - 2 sensors TEVA-SPOT KYPIPE Time to Detection (min) KY1 KY 2 KY 3 KY 4 KY 5 KY 6 KY 7 KY 8 KY 9 KY 10 KY 11 KY 12 System Figure 24: Comparison between KYPIPE and TEVA-SPOT - Baseline Conditions (2 sensors) The results were also investigated to determine which program resulted in faster times to detection for all 15 contamination scenarios performed on each system. The data showing the number (out of a total of 15 scenarios) and percentage of contamination scenarios that led to lower times to detection using each program is displayed in Table 4. 61

70 Table 4: Analysis of Faster Times to Detection between KYPIPE and TEVA-SPOT 1 sensor 2 sensors System Number of scenarios that resulted in lower time to detection TEVA-SPOT Percentage of scenarios that resulted in lower time to detection Number of scenarios that resulted in lower time to detection KYPIPE Percentage of scenarios that resulted in lower time to detection KY 1 0 0% % KY % 6 40% KY 3 0 0% % KY 4 0 0% % KY % 3 20% KY 6 0 0% % KY % 0 0% KY % 0 0% KY 9 0 0% % KY % % KY % % KY % % KY 1 0 0% % KY 2 0 0% % KY 3 0 0% % KY 4 0 0% % KY % 0 0% KY 6 0 0% % KY % 0 0% KY % 2 13% KY 9 0 0% % KY % % KY % % KY % % Examining the results shown in Table 4, TEVA-SPOT had lower times to detection for the majority of contamination scenarios for only four out of the 12 systems, when placing one sensor. KYPIPE produced faster times to detection for the majority of scenarios for the remaining eight systems for the placement of one sensor. When placing two sensors, TEVA-SPOT had faster times for the majority of contamination scenarios for three systems, and KYPIPE produced faster times for the majority of scenarios in nine systems. This analysis reflects the same results presented for the baseline conditions in Table 3. 62

71 These results outlined above prove the effectiveness of the KYPIPE sensor placement tool. For the majority of system models, just observing results from the baseline contamination scenario, KYPIPE selected sensors that had lower times to detection. This is effective in accomplishing the goal of online quality monitoring, to assess water quality and alert operators of a contamination event. Faster detection times of a contamination event will result in quicker notification to those affected and fewer negative effects. The KYPIPE sensor placement tool also has the advantage of its simplicity and ease of use. After running a successful extended period simulation, the sensor placement tool is easily activated by using the Shift + F7 keys. Basic parameters are entered into the tool (number of sensors, total simulation time, WQ computational time, injection rate and time, injection start and end time, and detection limit). The tool will then run with a fairly short computation time. The optimal sensor locations are output along with detection time, and the sensors can be easily displayed on the system map. This is advantageous to utility managers, and the simplicity and ease of use of the sensor placement tool will allow it to be an effective tool for utilities Comparison of Identical Sensor Placement Along with comparing the times to detection from KYPIPE and TEVA-SPOT, the sensor locations chosen as optimal sensors by both programs were also compared. Some contamination scenarios for the same system model resulted in TEVA-SPOT and KYPIPE selecting the same sensor nodes as the optimal locations. Other scenarios lead to different locations chosen as the optimal sensor locations between KYPIPE and TEVA-SPOT. Even in cases where different sensors were chosen by the programs, the times to detection for the chosen sensors were typically very similar. For each system model, the selected sensors for all 15 contamination scenarios were investigated. The number of contamination scenarios that resulted in identical sensor selection between KYPIPE and TEVA-SPOT, out of the 15 total scenarios, was recorded. This data (for placement of one sensor) is shown in Table 5. The percentage of contamination scenarios resulting in identical sensor selection between KYPIPE and TEVA-SPOT is also illustrated in Figure

72 Percentage with Identical Sensor Table 5: Identical Sensor Selection between KYPIPE and TEVA-SPOT (1 sensor) System Scenarios with Matching Sensor Selection Percentage of Scenarios with Matching Sensor Selection KY % KY % KY % KY % KY % KY % KY % KY % KY % KY % KY % KY % 100% 90% Percentage of Contamination Scenarios with Identical Sensor Selection (KYPIPE and TEVA-SPOT) - 1 sensor 80% 70% 60% 50% 40% 30% 20% 10% 0% KY1 KY 2 KY 3 KY 4 KY 5 KY 6 KY 7 KY 8 KY 9 KY 10 KY 11 KY 12 System Figure 25: Identical Sensor Selection between KYPIPE and TEVA-SPOT (1 sensor) 64

73 After observing the data in Table 5 and Figure 25, it is clear that some system models had matching sensor selection between KYPIPE and TEVA-SPOT for all 15 contamination scenarios, while other systems did not have any matching sensor placement among contamination scenarios. Three of the 12 systems had matching optimal sensor nodes for all 15 scenarios, and six of the 12 models had at least 80 percent matching sensor nodes for the placement of one sensor. On average, 7.7 out of the 15 scenarios (51 percent) resulted in identical placement of sensors. There were three systems that did not have any identical sensor nodes between KYPIPE and TEVA-SPOT. Even in these systems with no matching sensors, further investigation revealed that the vast majority of these sensors were still in close proximity to each other. Only two contamination scenarios (out of the 15 scenarios performed for 12 systems for a total of 180 simulations) led to sensor locations that were considered to be far away from each other in the distribution system. Both of these cases were in the KY 7 system. Figure 26 displays an example of different optimal sensor nodes between KYPIPE and TEVA-SPOT in KY 2. For three out of the 15 contamination scenarios, TEVA-SPOT selected J-138 and KYPIPE recommended J-485 as the optimal sensor location (both are shown in the KYPIPE program for comparison purposes). However, these sensors are in very close proximity. The sensor placement recommendations between KYPIPE and TEVA-SPOT are similar. Figure 26: Example of Differing Sensor Placement in Close Proximity 65

74 Figure 27 shows an example of selected sensor nodes with KYPIPE and TEVA-SPOT that do vary considerably in location. For two out of 15 scenarios performed on KY 7, TEVA-SPOT recommended J-249 and KYPIPE selected J-271. These sensors are not close in proximity to each other. Even though the times to detection may be similar, a few cases did lead to recommended sensors between the two programs that were not close to each other. However, the scenario displayed in Figure 27 was rare, even in the two sensor placement scenarios. Figure 27: Example of Differing Sensor Placement not in Close Proximity The location of selected sensors between KYPIPE and TEVA-SPOT was also investigated for the placement of two sensors. For the same 15 contamination scenarios performed on all 12 models, the number of scenarios with one out of two matching sensors between KYPIPE and TEVA-SPOT was found. The number of scenarios that resulted in two out of two identical sensors between the two programs was also recorded. This data is shown in Table 6 and Figure

75 Percentage with Identical Sensors Table 6: Identical Sensor Selection between KYPIPE and TEVA-SPOT (2 sensors) System Scenarios with 1 (out of 2) Matching Sensors Percentage of Scenarios with 1 (out of 2) Matching Sensors Scenarios with 2 Matching Sensors Percentage of Scenarios with 2 Matching Sensors KY % % KY % % KY % % KY % % KY % 0 0.0% KY % % KY % 0 0.0% KY % 1 6.7% KY % 0 0.0% KY % 0 0.0% KY % 0 0.0% KY % % 100% Percentage of Contamination Scenarios with Identical Sensors Selection (KYPIPE and TEVA-SPOT) - 2 sensors 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 1 of 2 sensors 2 of 2 sensors KY1 KY 2 KY 3 KY 4 KY 5 KY 6 KY 7 KY 8 KY 9 KY 10 KY 11 KY 12 System Figure 28: Identical Sensor Selection between KYPIPE and TEVA-SPOT (2 sensors) 67

76 The data presented in Table 6 and Figure 28 shows similar trends in identical sensor placement as with the one sensor scenario. Some of the system models resulted in both sensors matching between KYPIPE and TEVA-SPOT for many of the 15 contamination scenario, while other systems did not have any identical sensors placed for any of the 15 scenarios. KY 3 was the only system to have identical sensor placement for both sensors between KYPIPE and TEVA-SPOT for all 15 contamination scenarios, and KY 11 was the only system without any matching sensors between the two programs. All other systems ranged in the percentage of scenarios that matched one out of two or two out of two sensors. KY 1, KY 2, KY 3, KY 4, and KY 6 all resulted in two out of two identical selected sensors for over 50 percent of the contamination scenarios, showing that KYPIPE and TEVA-SPOT produced very similar results in these systems. The cases where KYPIPE and TEVA-SPOT recommended different sensor nodes were investigated. As with the placement of only one sensor, the vast majority of these cases resulted in placement of sensors that were in close proximity to each other. Only a few cases produced results where the sensors recommended by KYPIPE and TEVA-SPOT were significantly far away from each other. Specifically, 14 cases (out of the 15 simulations run on all 12 systems) led to differing sensor selection between the two programs that varied considerably spatially. In all of these cases, only one of the two sensors placed showed significant spatial variation between the two programs Results for All Contamination Scenarios All 12 system models were run for all 15 contamination scenarios using both KYPIPE and TEVA-SPOT. For each contamination scenario, the program that resulted in the lowest time to detection was identified. The difference between the times to detection between the two programs was also calculated. These results for the system KY 1, for the placement of one sensor, are shown in Table 7. 68

77 Table 7: Sensor Placement Results for KY 1 (1 sensor) System KY 1 Amount Rate Time Injection Rate (mg/min) Injection Time (hours) TEVA-SPOT Sensor Node Time to Detection (min) Sensor Node KYPIPE Time to Detection (hr) Time to Detection (min) Fastest Time to Detection Difference in min (Higher - Lower Time) J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE J J KYPIPE 19.4 The results for the placement of two sensors in KY 1 are shown in Table 8. The sensor placement results for all 12 models using both KYPIPE and TEVA-SPOT are included in Appendix E. 69

78 Table 8: Sensor Placement Results for KY 1 (2 sensors) TEVA-SPOT KYPIPE System Injectio n Rate (mg/mi n) Injectio n Time (hours) Sensor Node #1 Sensor Node #2 Time to Detectio n (min) Sensor Node #1 Sensor Node #2 Time to Detectio n (hr) Time to Detectio n (min) Fastest Time to Detectio n Differen ce in min (Higher - Lower Time) J-235 J J-235 J KYPIPE 18.7 Amou nt J-235 J J-235 J KYPIPE J-235 J J-235 J KYPIPE J-245 J J-245 J KYPIPE 23.0 KY 1 Rate Time J-244 J J-244 J KYPIPE J-245 J J-235 J KYPIPE J-245 J J-235 J KYPIPE J-235 J J-235 J KYPIPE J-235 J J-235 J KYPIPE J-235 J J-235 J KYPIPE J-245 J J-235 J KYPIPE J-244 J J-235 J KYPIPE J-235 J J-235 J KYPIPE J-235 J J-235 J KYPIPE J-235 J J-235 J KYPIPE

79 CHAPTER 5 5 Conclusion 5.1 KYPIPE Sensor Placement Tool Conclusion TEVA-SPOT has been developed to analyze the vulnerability of drinking water distribution networks and recommend locations for the deployment of water quality sensors as a component of a contamination warning system. However, the software is not appropriate for small utilities in terms of the simplicity and ease of use. The water quality sensor placement tool was developed with KYPIPE as a simple tool to aid utility managers in the optimal placement of sensors within their distribution systems. The new sensor placement tool has been developed to work with the existing KYPIPE graphical user interface. The sensor placement tool recommends optimal sensor placement based on minimizing time to detection. The tool considers detection events at nodes throughout the entire system, and recommends optimal sensor placement based on the locations that can detect contamination events the fastest. The WQ sensor placement tool is able to provide sensor placement for up to five sensors. The sensor placement tool recommends optimal sensor placement based on an enumeration process. The process considers injection of the contaminant at every possible injection node for every plausible combination of sensor locations. The travel time is calculated from each injection node to every possible sensor node, resulting in an average travel time for each set of possible sensor locations. The optimal sensor locations are chosen as the nodes with the fastest time to detection, resulting in a sensor design that can quickly notify the utility of a contamination event. The development of the sensor placement tool accomplishes the objective of providing a simple tool to aid small utilities in sensor placement. To obtain optimal sensor placement in a drinking water system using the sensor placement tool, a model of the system must first be developed in KYPIPE. The model should accurately depict the structure of the system and should include characteristics of all system components. An EPS is set up and executed for the model. To execute the sensor placement tool, the number of sensors to place, total simulation time, WQ computational time, mass injection rate, injection start time, injection end time, and detection limit values 71

80 must be entered. When the run is completed, the tool will output the optimal sensor locations, along with the average travel time for the selected sensors. 5.2 Comparison of Results between KYPIPE and TEVA-SPOT Conclusion The KYPIPE sensor placement tool was executed on the 12 distribution system models in the model database for 15 contamination scenarios. The contamination scenario was defined by both the rate of injection of the contaminant (in mg/min) and the total injection time (in hours). Contamination scenarios were created for three different general scenarios: fixed amount, fixed rate, and fixed time. All 15 contamination scenarios were executed on all system models using both TEVA-SPOT and KYPIPE (for one and two sensors). The goal was to compare the sensor placement results, both sensor placement and times to detection, between KYPIPE and TEVA-SPOT. To be able to compare results from the two programs, it was important that all parameters matched between the programs. This included the models used in each along with the contamination scenarios, possible injection nodes and sensor nodes, etc. The WQ computational times were both set to 60 seconds, the total simulation time was set to 24 hours, and the detection limit for both programs was also set to 0.01 mg/l. Examining only the baseline contamination scenario (1000 mg/min for 4 hours) for each system produced results that were closely reflective of examining all 15 contamination scenarios. When placing one sensor, KYPIPE resulted in lower times to detection for eight of the 12 distribution systems. TEVA-SPOT led to faster times to detection for the remaining four systems. For the placement of two water quality sensors, KYPIPE produced faster times for nine systems, and TEVA-SPOT led to faster times for three of the 12 systems. It was also noted that the times to detection between TEVA-SPOT and KYPIPE for the same system were similar; there were no dramatic differences in times between the two programs. The sensor locations chosen as optimal sensor nodes by both programs were also compared. Some contamination scenarios for the same model resulted in selection of the same sensor nodes using KYPIPE and TEVA-SPOT, while other scenarios lead to different locations chosen as the optimal sensor nodes between the programs. For the 72

81 placement of one sensor, three of the 12 systems had matching sensor locations for all 15 scenarios, and six of the 12 models had at least 80 percent matching sensor nodes for the placement of one sensor. On average, 7.7 out of the 15 scenarios (51 percent) resulted in identical placement of sensors between the two programs. There were three systems that had zero identical sensor nodes between KYPIPE and TEVA-SPOT, but further examination showed that the vast majority of the different sensors selected by the two programs were still in close proximity to each other. Only two contamination scenarios (out of the 15 scenarios performed for 12 systems) resulted in optimal sensor locations that were considered to be far away from each other in the network. The optimal sensor locations between KYPIPE and TEVA-SPOT were also investigated for the placement of two sensors. The data showed similar trends in identical sensor placement found with the one sensor placement scenarios. KY 3 was the only network to have identical sensor placement for both sensors between KYPIPE and TEVA-SPOT for all 15 contamination scenarios, and KY 11 was the only system without any matching sensors between the programs. KY 1, KY 2, KY 3, KY 4, and KY 6 all resulted in two out of two identical selected sensors for over 50 percent of the contamination scenarios. Therefore, the data showed that KYPIPE and TEVA-SPOT produced fairly similar results in these systems. When KYPIPE and TEVA-SPOT selected different sensor nodes, the majority of these cases resulted in sensor nodes that were in close proximity to each other. Only 14 cases (out of the 15 scenarios run on all 12 systems) led to results where the sensors chosen by KYPIPE and TEVA-SPOT were significantly far away from each other. In all of these cases, only one of the two sensors placed had significant spatial variation. The results accomplished the objective of proving the effectiveness of the KYPIPE sensor placement tool. For the majority of system models, KYPIPE selected sensors that had lower times to detection. This accomplishes the goal of online quality monitoring to assess water quality and alert operators of a contamination event. The KYPIPE sensor placement tool is also beneficial because it is simple and easy to use. Basic parameters are entered into the tool, and it will then run with a short computational time. The optimal sensor locations are output and displayed on the map along with detection times. The simplicity and ease of use of the sensor placement tool will allow it to be an effective tool for utilities. 73

82 This results acquired with this research provides the foundation for future work on developing sensor placement guidance. The recommended sensor locations from the KYPIPE sensor placement tool can be examined to determine if patterns exist based on system characteristics. The differences in sensor location based on the variation in each system or general network configuration can also be analyzed. If trends in the placement of sensor nodes are observed, guidance can be developed to assist small utilities in placing water quality sensors. Development of sensor placement guidance, without the need for a costly calibrated hydraulic model, would be greatly beneficial to a small utility in protecting their water supply. 74

83 CHAPTER 6 6 Acknowledgements Funding for this research was provided by the U. S. Department of Homeland Security, Science & Technology Directorate, through a technology development and deployment program managed by The National Institute for Hometown Security, under Other Transactions Agreement (OTA) #HSHQDC We are very grateful for these contributions. In addition, we would like to acknowledge the assistance of Dr. Srini Lingireddy of KYPIPE, LLC. 75

84 CHAPTER 7 7 References Aral, M. M., Guan, J., & Maslia, M. L. (2010). Optimal Design of Sensor Placement in Water Distribution Networks. Journal of Water Resources Planning and Management, AWWA. (2005). Computer Modeling of Water Distribution Systems. Manual of Water Supply Practices- M32. Denver: American Water Works Association. Chang, N.-B., Pongsanone, N. P., & Ernest, A. (2011). Comparisons between a rule-based expert system and optimization models for sensor deployment in a small drinking water network. Expert Systems with Applications, Chang, N.-B., Pongsanone, N. P., & Ernest, A. (2012a). A rule-based decision support system for sensor deployment in small drinking water systems. Journal of Cleaner Production, Chang, N.-B., Prapinpongsanone, N., & Ernest, A. (2012b). Optimal sensor deployment in a large-scale complex drinking water network: Comparisons between a rule-based decision support system and optimization models. Computers and Chemical Engineering, Gagliardi, M. C., & Liberatore, L. J. (n.d.). Water Systems Piping. Lyndhurst, NJ. Hart, W. E., & Murray, R. (2010). Review of Sensor Placement Strategies for Contamination Warning Systems in Drinking Water Distribution Systems. Journal of Water Resources Planning and Management, Isovitsch, S. L., & VanBriesen, J. M. (2008). Sensor Placement and Optimization Criteria Dependencies in a Water Distribution System. Journal of Water Resources Planning and Management, Janke, R., Murray, R., Uber, J., & Taxon, T. (2006). Comparison of Physical Sampling and Real- Time Monitoring Strategies for Designing a Contamination Warning System in a Drinking Water Distribution System. Journal of Water Resources Planning and Management, Kentucky Infrastructure Authority. (2010, February 25). Water Resources Information System. Retrieved 2012, from Mays, L. W. (2000). Water Distribution Systems Handbook. New York, NY: McGraw-Hill. Mays, L. W. (2005). Water Resources Engineering. John Wiley & Sons, Inc. 76

85 Mays, L. W., & Tung, Y.-K. (2002). Hydrosystems Engineering and Management. Highlands Ranch, CO: Water Resources Publications, LLC. McGhee, T. J. (1991). Water Supply and Sewage. Hightstown, NJ: McGraw-Hill, Inc. McKenna, S. A., Hart, D. B., & Yarrington, L. (2006). Impact of Sensor Detection Limits on Protecting Water Distribution Systems from Contamination Events. Journal of Water Resources Planning and Management, Murray, R., Hart, W., & Berry, J. (2006). Sensor Network Design for Contamination Warning Systems: Tools and Applications. American Water Works Association (p. 20). Washington, D.C.: Water Security Congress Conference. Murray, R., Haxton, T., & Janke, R. (2010). Sensor Network Design for Drinking Water Contamination Warning Systems. Cincinnati, OH: National Homeland Security Research Center, Office of Research and Development, U.S. Environmental Protection Agency. Murray, R., Janke, R., & Uber, J. (2004). The Threat Ensemble Vulnerability Assessment (TEVA) Program for Drinking Water Distribution System Security. Critical Transitions in Water and Environmental Resources Management (p. 8). Salt Lake City, UT: World Water Congress. Murray, R., Janke, R., Hart, W. E., Berry, J. W., Taxon, T., & Uber, J. (2008). Sensor network design of contamination warning systems: A decision framework. American Water Works Association, National Research Council. (2006). Drinking Water Distribution Systems: Assessing and Reducing Risks. Washington, DC: The National Academic Press. Ostfeld, A., Uber, J. G., Salomons, E., Berry, J. W., Hart, W. E., Phillips, C. A., et al. (2008). The Battle of the Water Sensor Networks (BWSN): A Design Challenge for Engineers and Algorithms. Journal of Water Resources Planning and Management, Panguluri, S., Grayman, W. M., & Clark, R. M. (2005). Water Distribution System Analysis: Field Studies, Modeling and Management. Cincinnati, OH: United States Environmental Protection Agency. U.S. Environmental Protection Agency. (2008). Water Quality in Small Community Distribution Systems - A Reference Guide for Operators. Cincinnati, OH: Office of Research and Development. United States Department of Agriculture. (n.d.). Geospatial Data Gateway. Retrieved 2012, from National Resources Conservation Service: 77

86 Walski, T. M., Chase, D. V., Savic, D. A., Grayman, W., Beckwith, S., & Koelle, E. (2003). Advanced Water Distribution Modeling and Management. Bentley Institute Press. Wood, D. J. (1981). Algorithms for Pipe Network Analysis and Their Reliability. Lexington, KY: University of Kentucky Water Resources Research Institute. Wood, D. J. (2010). KYPipe Reference Manual. KYPipe, LLC. Wood, D. J., & Rayes, A. (1981). Reliability of Algorithms for Pipe Network Analysis. Journal of the Hydraulics Division, Xu, J., Fischbeck, P. S., Small, M. J., VanBriesen, J. M., & Casman, E. (2008). Identifying Sets of Key Nodes for Placing Sensors in Dynamic Water Distribution Networks. Journal of Water Resources Planning and Management,

87 Appendix A Tools for Water Utilities 79

88 Visual tools were developed to aid utility managers in understanding their distribution system and deciding on optimal placement of water quality sensors. The posters shown in this section were intended to be placed on the wall of a utility office or some other location where utility workers can quickly and easily reference the information presented. The poster depicted in Figure 29 will aid a utility in determining which classification of general system configuration (loop, grid, or branch) represents their water distribution system. The poster includes a figure at the top portion of the poster defining the general configuration of each classification based on pipe geometry and size. The lists of bullet points further define the varying characteristics of each configuration. The three small figures on the right-hand side of the poster represent three real distribution systems that represent the different system configurations. Figure 30 shows a poster to help aid a utility in executing the KYPIPE sensor placement tool. The flowchart described the general steps involved in the procedure. Above (and below) each step in the flowchart, a screenshot of the KYPIPE interface is included. These screenshots will further simplify the process of executing the KYPIPE sensor placement tool. 80

89 Figure 29: Water Utility Poster (Determining Water Distribution System Configuration) 81

90 Figure 30: Water Utility Poster (Procedure for Executing KYPIPE Sensor Placement Tool) 82

91 Appendix B Model Development Procedure 83

92 B.1 Data Acquisition in GIS The model database used in this research was created by following a procedure that utilized the Geographic Information Systems (GIS) software. The data for all distribution system components was first acquired from the Water Resources Information System (Kentucky Infrastructure Authority, 2010). The Geospatial Data link was selected from the WRIS homepage, displaying zip files containing shapefiles for all systems components for the state of Kentucky. The shapefiles for water lines, pumps, tanks, and water treatment plants were downloaded and unzipped (all to the same folder). The files for meters, surface sources, well sources, and purchase sources were also available, but these components were not necessary for the purpose of this project. A blank ArcGIS document was opened, and the Add Data bottom was used to add shapefiles for water lines, tanks, pumps, and water treatment plants. The data for the entire state of Kentucky is shown in Figure 31. Figure 31: Distribution System Component Shapefiles in GIS In the process of model creation, only the data for the specific utility of concern is necessary. The data for one utility was isolated by right clicking on the layer name for each component in the Table of Contents and selecting Open Attribute Table. The Table Options icon in the top left-hand corner was selected and then the Select by Attributes option. A clause was then 84

93 formulated to isolate parts of the shapefile based on an attribute. The Owner attribute was double-clicked, followed by the equal sign and the Get Unique Values button. The desired utility name was double clicked to complete the clause, and the Apply bottom was clicked. This action selected all the shapes associated with that utility from the data for the entire state, and this process is displayed in Figure 32. Figure 32: Select by Attributes The shapes for the utility were selected (and shown as highlighted in the table). Next, the attribute table was closed, the component layer was right-clicked, and the Create Layer from Selected Features option was selected under the Selection menu. This created a new layer for the water lines, pumps, tanks, and water treatment plants for the system of interest. Next, it was advantageous to calculate the lengths of the water lines. The new water lines layer was right clicked, and the Open Attribute Table was again selected. Under the Table Options icon, the Add Field option was chosen. A name was specified for the new field (LENGTH_FT), and Double precision under type was chosen. To calculate the length of the pipes with the new field, the gray box containing the name of the new field was right-clicked, and Calculate Geometry was selected. The option Length was selected under Property, the coordinate 85

94 system was left as the default option, and Feet US [ft] was selected as the unit. This process calculated the length of all water lines, and the steps are displayed in Figure 33. Figure 33: Calculating Length of Water Lines Next, the shapefiles of all system components were exported by right-clicking on each layer, clicking on Data, and then selecting Export Data. The folder where data was stored was found, and the data was saved as Shapefile under the Save as Type option. It was important to give the file a name that did not exceed eight characters. This process is shown in Figure

95 Figure 34: Data Export in GIS This step was followed for the other system components, making sure to save files for all components in the same folder. This portion of the model creation process completes the first phase of data manipulation in GIS. B.2 Data Input to KYPIPE Next, the data exported from GIS needed to be input to KYPIPE. To start this process, a new KYPIPE file was opened. Under the File menu, the Pipe2000 Utilities option was selected, followed by Import ArcView File. When the Shape File Import Utility opened, the Select shape file folder box was selected, and the pipes shapefile that was exported from GIS was found and opened. The utility showed a list of fields in KYPIPE on the left side of the screen under the drop-down menu, and the right portion of the screen displayed attributes present in the shapefile from GIS. Characteristics of each system component were able to be transferred from the attribute table in GIS to the KYPIPE file. Attributes were matched by clicking on an attribute on the left list, clicking on the corresponding attribute on the right list, and then selecting the Match Selection box. Once all the desired attributes were matched, the Fix Connectivity Errors and Check for Crisscross Lines boxes were checked, and then the Read Pipe Shape 87

96 File box was clicked followed by the Process Pipe Data box. This procedure is displayed in Figure 35. Figure 35: Matching Attributes between GIS and KYPIPE (Pipes) After the pipe data was processed, the other system components were processed by selecting the component (by the name given when the layer was exported from GIS) from the drop-down box on the left of the utility and selecting the bullet for the corresponding component in KYPIPE from the list on the far-left side of the utility. Attributes were matched for each component, similar to the process for pipes, and the Process Data button was clicked to process each component. The process for importing the data for pumps is shown in Figure

97 Figure 36: Matching Attributes between GIS and KYPIPE (Pumps) Once all system components were added, the Save P2K File button was selected to save the file. Although it is possible to match numerous attributes between the GIS shapefile and corresponding KYPIPE data, experience executing this procedure has shown that matching numerous attributes can lead to error in the model development process. For example, the shapefiles from GIS for the pumps contain data about capacity and horsepower that would be useful data in KYPIPE. However, attempts to match these attributes often lead to errors that resulted in all pumps being absent from the KYPIPE model. It is recommended to only match a few attributes for each component to be able to distinguish them, and then match all other necessary attributes by hand. Table 9 shows the recommended attributes to match between GIS and KYPIE for each system component. The GIS shapefile for Water Treatment Plants (WTP) corresponds to reservoirs in the KYPIPE Program. 89

98 Table 9: Attribute Matching between GIS and KYPIPE Component KYPIPE Attribute (left side) GIS Attribute (right side) Diameter SIZE Pipes MaterialRating MATERIAL Length LENGTH_FT Reference Year YEARCON Pumps Name WP_ID Tanks Name WT_ID Reservoirs (WTP) Name WTPNAME B.3 Addition of Elevation Data Completion of the procedure thus far resulted in a KYPIPE model containing all of the necessary system components (pipes, tanks, pumps, and reservoirs). All components had accurate (X,Y) coordinates and system components were connected to each other (the check for this assumption is discussed later). However, the current model did not have any elevation data associated with system components. Elevation data is necessary to carry out a hydraulic analysis in the KYPIPE program. Therefore, the next steps in model development involve acquiring elevation data for the area encompassing the utility and assigning values of elevation to components in the model. A Digital Elevation Model (DEM) was acquired from the National Resources Conservation Service Geospatial Data Gateway (United States Department of Agriculture). After navigating to the website, the green circle labeled Get Data was selected, followed by the desired state and county. Once the selected county was submitted, the Elevation data was found from the list of data available to download. The site presented a list of National Elevation Datasets, and it listed options for 30 meter, 10 meter, or 3 meter datasets depending on the location. These values indicated the grid cells (a 10 meter DEM will consist of 10m x 10m grid cells with a corresponding value for elevation), so the smaller grid cells result in more accurate elevation data. The smallest available grid cell dataset was selected, and the FTP delivery method was selected (delivery by ). The Place Order button was selected to order the desired dataset. The DEM selection process is shown in Figure

99 Figure 37: NRCS Geospatial Data Gateway Next, a shapefile of all nodes in the KYPIPE model was generated to add to GIS and later combine with the DEM. The newly created KYPIPE model was opened, and the Export ArcView File was selected in the Pipe2000 Utilities menu. When the utility opened, the Nodes bullet was selected in the top left-hand corner of the utility. The box for Name in the window below was checked, the right arrow icon was clicked, and then the Generate Shape Files box was selected. This created a shapefile named Nodes in the same location as the.kyp folder, and this process is shown in Figure

100 Figure 38: KYPIPE Nodes Shapefile Export The new nodes shapefile was then added to the ArcMap document (located in the KYPIPE folder), and the nodes shapefile lined up with the water lines layer. This step of the procedure is displayed in Figure

101 Figure 39: Nodes Shapefile in GIS The next step of the model creation process was adding the DEM to the GIS file. The DEM files were downloaded from the sent by NRCS Geospatial Data Gateway and unzipped. There were numerous files for the specified county, so the elevation files were added until all of the nodes in the system were covered by the DEM. In the example shown, three of the 19 DEM raster provided were needed to cover the area of the distribution system. The newly added DEMs are shown with the nodes and water line shapefiles in Figure

102 Figure 40: Digital Elevation Model Since more than one DEM raster files was required to cover the entire area of the distribution system, it was necessary to combine all DEM files. This was accomplished by opening ArcToolbox, then selecting Data Management Tools, Raster, Raster Dataset and Mosaic to New Raster. When the Mosaic to New Raster window appeared, each DEM was selected from the drop-down menu, a folder location was selected for the new combined DEM, a name for the DEM was chosen, and the number of bands was set as 1. In order to define the spatial reference of the DEM, the icon next to the Spatial Reference for Raster box was clicked. In the Spatial reference window, the Import box was selected, and the water line shapefile for the system was found and added as the spatial reference. This step combined all of the necessary DEM rasters into one combined raster file, and this process is displayed in Figure

103 Figure 41: Combined DEM Process Once the individual DEMs were combined into one, the next step was to extract the elevation data to each node. First, the Spatial Analyst box was checked under the Customize then Extensions menu. ArcToolbox was then opened, and the Spatial Analyst option was selected, followed by Extraction and Extract Values to Points. In the Extract Value to Points tool, the Nodes feature was first selected from the drop-down menu as the input point feature. The combined DEM was chosen as the input raster, and a location and name was entered for the new point features. The checkbox labeled Interpolate values at point locations was also checked to interpolate elevations in the DEM data. The process created a new shapefile containing the nodes from KYPIPE with assigned elevations from the DEM, and the steps are illustrated in Figure

104 Figure 42: Elevation Extraction Process The previous step created a new shapefile containing elevation data for each node in the system. The elevations needed to be added to the KYPIPE model, and Microsoft Excel was utilized in this process. A blank Excel document was opened, the Open icon was selected, and the folder where the shapefile was saved in the previous step was located. On the drop-down menu labeled Files of Type, the option for All Files was chosen. The file of nodes with assigned elevations (created in the previous step) with the file extension.dbf was opened. This step is shown in Figure

105 Figure 43: Opening Elevation File in Excel The Excel file showed the node names in the first column, and the elevation of the node was contained in the second column. These elevations were in meters, so the elevations were converted to feet by entering the equation =B2* into the C2 grid. This equation was applied to all elevation by double clicking on the lower right-hand corner of the C2 cell. Once all nodal elevations were calculated (in feet), the data needed to be copied to the KYPIPE model. The model was opened in the KYPIPE program, and the data tables were accessed by selecting Data Tables in the Edit menu. The Nodes box was clicked in the upper left-hand side of the screen to access data for all nodes in the system. The column for elevation (labeled Elv. ) was located. It was then necessary to sort the data in Excel so the order of node names matched the order in KYPIPE; this would allow a simple copy and paste of the entire column to elevation data. To sort the data in Excel, the cells for node name and elevation were highlighted, and the Sort icon was selected under the Data menu. The option NAME was selected in the Sort by drop-down box, Values was entered in the Sort on box, and A to Z was chosen in the Order box. This process changed the order of node names in Excel to match the order already present in the Node table of the KYPIPE model, and this step is shown in Figure

106 Figure 44: Elevation Data Sorting in Excel To transfer values of elevation from the Excel document to the KYPIPE model, the entire column for elevations in feet (starting at cell C2) was highlighted. The column was copied by hitting Ctrl + C and then copied into KYPIPE by hitting Edit and Paste after the first cell in the Elevation column was selected in KYPIPE. This step assigned an elevation value to every node in the KYPIPE model, and the results of this step are displayed in Figure

107 Figure 45: Nodal Elevation Data in Excel B.4 Final Adjustments to Model Completion of the outlined steps resulted in a model containing all distribution system components with elevations assigned to each component along with nodes throughout the system. However, other alterations were required to fix errors that could have occurred in the model creation process. First, data for tanks including the size, minimum water level, maximum level, and initial level were entered manually for each tank. The horsepower of each pump was also entered manually into the model, along with grade of the reservoirs and any other necessary data. Various tools in KYPIPE were also utilized to help detect possible errors. Under the Analyze menu, the Connectivity Check option was selected (followed by clicking on any pipe in the system). This tool highlighted pipes that were not connected to the rest of the system. To fix pipes that were disconnected from the system, the pipe was manually extended in the same direction to a node in a nearby pipe. If a node was not present nearby, an intermediate node was added to the nearby pipe and the elevation of this node was interpolated using elevations for the closest nodes. In all cases, it was very obvious where the pipe should extend and connect to the system. It was also true in all cases that the pipe appeared to be connected when observing the 99

108 system at a normal zoom level, and the disconnection was only noticeable if the portion was zoomed in at high levels. Figure 46 illustrates this concept; the left portion represents a normal zoom level where the disconnection is not noticeable. The problem is clearly noticeable om the right portion of the figure, but this is a very high level of zoom that the user would not typically use. Figure 46: Pipe Connection Errors Another tool was utilized in KYPIPE to check for other general errors, such as undefined initial elevations in tanks, an undefined grade in a reservoir, or an extremely high value for pump power. The tool was utilized by selecting Error Check in the Analyze menu. Roughness coefficients were also applied to all pipes in the model to estimate head loss through the pipes due to friction, and the concept behind these roughness values are discussed in Section These values were applied to the model by navigating to the Setups/Defaults tab on the main screen, and then selecting the Pipe Type tab. The chart displayed each pipe material present in the model along with all pipe diameters for each material present in the model. Values were entered for each pipe material and diameter (although roughness coefficients did not vary based on pipe diameter) to represent the roughness factor of a new pipe in the column labeled Reference Roughness. Values were also entered to represent the reduced roughness coefficient of the pipe after 10 years in the Aged Roughness (10 yr) column. These values were also used to estimate roughness of pipes that were greater than 10 years old. To apply these roughness values to pipes in the model, the Select All Pipes option was selected under the Edit menu. In the Edit Pipe Set box in the upper right-hand corner, Roughness 100

109 was selected as the Item to Edit, and Not was chosen as the Operation. The Proceed button was clicked to apply the set roughness values to pipes in the system, and this procedure is shown in Figure 47. Figure 47: Changing Roughness Values of Pipes in KYPIPE The values used for roughness coefficients in the KYPIPE model are shown in Table 10. Table 10: Hazen-Williams Roughness Coefficients Material Reference Roughness Aged Roughness (10 yr) Cast Iron Concrete Ductile Iron PVC PE (Polyethylene) AC (asbestos cement) To create a model that was a close reflection of an actual water distribution system, water demand data also needed to be incorporated into the model. This included allocating the total daily demand to nodes throughout the model and also implementing demand factors to account for varying water use patterns throughout the day. To find the total daily demand of the system, 101

110 the system in question was found in the WRIS Portal (Kentucky Infrastructure Authority, 2010). Under the Planning tab, the total water usage was given in million gallons per year. This value was converted to million gallons per day (MGD), and then gallons per minute (GPM) to match the units used in KYPIPE. In KYPIPE, the Automatic Demand Distribution option was selected under the Analyze menu. In the box labeled Total Demand to Distribute, the total daily demand in GPM for the system was entered. The box labeled Apply this Demand at Junction Nodes was clicked, and this applied the total demand to nodes throughout the system based on pipe diameter. This process is displayed in Figure 48. Figure 48: Demand Allocation in KYPIPE Because water usage in a typical water distribution system has varying water demand patterns throughout the day, it was also necessary to implement demand patterns in the model. In KYPIPE, the Demand Patterns tab was located in the Setups/Defaults tab. The Load option was clicked, and the AWWA.dmt file was selected and the OK button was clicked. This loaded demand patterns established by the American Water Works Association over a 24 hour period. The demand factors were automatically loaded into the row labeled Type

111 Because the junctions in the model were automatically set to the R (residential) type, it was necessary to cut and paste the set of demand factors from the Type 1 row to the Residential row. This ensured the demand factors would be applied to the junctions in the system. This could also be executed by changing all junctions in the system from Dm Type R to Dm Type 1. This process applied demand patterns to the system by changing average hourly demand throughout the day based on time of day and estimated water use. This process is shown in Figure 49. Figure 49: Demand Patterns in KYPIPE Control switches were also added to the model to help simulate the pump schedule usually controlled by the utility. KYPIPE uses control switches to turn pumps on and off based on the level (pressure, head, or HGL) of a certain node in the system. Control switches were applied to certain pumps in the system that had a primary purpose of filling tanks. These control switches caused the pumps to turn off when the tanks reached a high water level (usually close to the maximum tank level) and turn back on when the tank reached a low level (close to the minimum tank level). In creating the model database, the hydraulic grade line (HGL) was used as the measurement in control switches. The pump was typically turned on when the tank reached a few feet above the low level, and it was turned off when the tank was a few feet below the maximum 103

112 level. The control levels were altered based on each specific system. If the tank in question reached its maximum or minimum tank level (causing extreme high and low pressure, respectively), the level at which the tank was turned on or off was moved further from the minimum and maximum tank levels. An example of a control switch setup is shown in Figure 50. This option is reached in KYPIPE by selecting Control Switches in the Other Data tab. Figure 50: Control Switches in KYPIPE 104

113 Appendix C Layout of Database Models 105

114 The following figures display the layout of each distribution system in the model database. The figures display all major system components, including tanks, pumps, reservoirs, water lines, and nodes. The water lines are also displayed using different colors based on pipe diameter. Figure 51: KY 1 System Layout 106

115 Figure 52: KY 2 System Layout Figure 53: KY 3 System Layout 107

116 Figure 54: KY 4 System Layout Figure 55: KY 5 System Layout 108

117 Figure 56: KY 6 System Layout Figure 57: KY 7 System Layout 109

118 Figure 58: KY 8 System Layout Figure 59: KY 9 System Layout 110

119 Figure 60: KY 10 System Layout Figure 61: KY 11 System Layout 111

120 Figure 62: KY 12 System Layout Figure 63: KY 13 System Layout 112

121 Figure 64: KY 14 System Layout Figure 65: KY 15 System Layout 113

122 Appendix D Procedure for Executing Sensor Placement Tool 114

123 D.1 Execution of Tool The following section outlines a detailed, step-by-step procedure for executing the sensor placement tool on a system model in KYPIPE. After the water distribution system model has been created in KYPIPE (procedure outlined in Appendix B), an Extended Period Simulation was set up in KYPIPE. This menu is found in the program by selecting the System Data tab, followed by the EPS (Extended Period Simulation) tab. In the Extended Period Simulation menu, the box next to Use EPS was first checked. Then values were entered for Total Time (hrs), Computational Periods (hrs), Report Period (hrs), Default Power Cost ($/kwhr), Starting Time (hrs 0-24), and Report Time Style. For this study, the total time was set to 24 hours, the computational and report period was set to 1 hour, the default power cost was left as 0, the starting time was set to 0 hours, and the report time style was set to Military Time. This step is shown in Figure 66. Figure 66: EPS Setup in KYPIPE Next, an analysis was run on the distribution system model. This can be executed by clicking the Analyze icon at the bottom right-hand corner of the screen or selecting Analysis in the Analyze menu. The execution process by selecting Analysis in the Analyze drop-down window is shown in Figure

124 Figure 67: Execution of EPS Analysis After the hydraulic analysis was completed, the sensor placement tool was executed. The tool can be initiated by using the shortcut Shift + F7. A window appeared verifying that the sensor placement tool was starting, and the OK icon was clicked. 116

125 Figure 68: Starting Sensor Placement Tool The sensor placement tool window appeared, and the number of sensors to be placed was input to the box in the upper left-hand corner labeled Number of sensors (max 5). In this study, optimal sensor locations were determined for only one and two sensors. The tool is capable of placing up to five sensors in a system. After the number of sensors was determined, the box labeled Set Default Parameters located under the input box for number of sensors in the upper left-hand corner was selected. This menu allowed the user to set values for the total simulation time (hours), WQ computational time (sec), mass injection rate (mg/min), injection start time (hours), injection end time (hours), and detection limit (mg/l). The total simulation time was set to 24 hours, the same time period as the extended period simulation. The WQ computational time was set for 60 seconds because this matched the computational time set in EPANET and used in TEVA-SPOT. It was important to ensure all parameters matched between KYPIPE and TEVA-SPOT for the purposes of this study. However, the user may specify any time period desired for these values. 117

126 The mass injection rate refers to the rate at which the contaminant is being injected in the system. For this study, 15 different injection scenarios were executed, and the values for mass injection rate varied from 250 mg/min to 4000 mg/min. The injection start time and injection end time determine the length of time, and when that time period occurs in the total simulation time. Although time duration of injection varied in this study, the injection start time was always set to 0 hours. The detection limit refers to the concentration of the contaminant that must be present at the sensor node in order for the sensor to detect the contaminant. The sensor will not detect the contaminant until the specified concentration is present. This study used a detection limit of 0.01 mg/l, which was consistent with the value used in TEVA-SPOT. The process of setting parameters for the sensor placement tool is displayed in Figure 69. Figure 69: Setting Parameters in Sensor Placement Tool After all default parameters were set, the box labeled Generate INP File was clicked. After a few seconds, the text in this box turned from Black to Gray, indicating the INP file was 118

127 generated. To begin the sensor placement simulation, the Run icon was clicked, located below the Generate INP File icon. This process is shown in Figure 70. Figure 70: Initiating Sensor Placement Run D.2 Sensor Placement Tool in Progress As the run was executing, the sensor placement tool displayed the progress of the simulation. The box in the lower left-hand corner of the tool window displays useful information about the system, such as the number of nodes, pipes, dead-end nodes, demand nodes, possible injection nodes, and possible sensor nodes. The tool considers possible sensor locations to be all nodes (including tanks, pumps, reservoirs, and junctions) except dead-end nodes. The average travel time to dead-end nodes will generally be much higher, skewing the average times to detection. Possible injection sites are considered to be all non-zero demand nodes, excluding dead-end nodes. Dead-end nodes are considered to be consumption nodes, so any contaminant injected at these nodes will be consumed immediately and the contaminant will not be able to travel further in the system. The reality of this concept may be slightly different, but this assumption was used 119

128 in the sensor placement tool. The bottom left-hand corner of the tool window also displays the default parameters previously entered. Figure 71 displays the sensor placement tool in progress. Figure 71: Sensor Placement Tool in Progress Figure 72 displays the sensor placement tool once it has started the sensor placement portion of the simulation. The box in the lower right-hand corner of the tool, labeled Optimal Sensor Locations displayed the best sensor locations at that point in the simulation. The average travel time of the current optimal solution was also displayed. The optimal sensor locations updated as the simulation continued, and this output (along with the average travel time) changed at least several times before the simulation was complete. 120

129 Figure 72: Sensor Selection Process When the sensor placement tool finished running through all possible sensor locations in the system, it selected the sensor locations with the lowest average travel time as the optimal sensor locations. When the simulation was complete, the window displayed in Figure 73 appeared. To continue, the user clicked the icon labeled OK. Figure 73: Completion of Sensor Placement Simulation 121

130 Once the OK icon was clicked, the KYPIPE map reappeared. In the Optimal Sensor Locations box in the lower right-hand corner of the tool, the selected sensors were displayed, along with the average travel time in hours. The output from the sensor placement tool is displayed in Figure 74 and Figure 75. Figure 74: Completed Sensor Placement Simulation 122

131 Figure 75: Completed Sensor Placement Tool Display D.3 Results Provided by Tool In order to view the selected sensors on the map, the node labels were turned on. This was accomplished by navigating to the Map Setting tab and then the Labels tab. In the Node Labels box, the Name box was checked, along with the Selected Labels Only box. This highlighted the selected nodes in red on the map and created a label with the node name. This process is shown in Figure 76, and the result is displayed in Figure

132 Figure 76: Turning on Node Labels Figure 77: Displaying Selected Nodes on the Map KYPIPE also has several features to observe more detailed results of the sensor placement simulation. In the lower right-hand corner of the sensor placement tool window, the View Report icon was clicked. This file is also accessible through the systemname.kyp file folder, 124

133 located wherever the KYPIPE model was saved. The report is named systemname(report), and is in the form of a text file. The report displays the default parameters used in the simulation, along with the selected sensor nodes and average travel time. This report for the example simulation is shown in Figure 78. Figure 78: Sensor Placement Summary Report Also located in the systemname.kyp file folder, the file named systemname in the WQC file format displays more useful information about the simulation. This text document displays the optimal sensor locations throughout the course of the simulation, along with the average travel time. The first sensors listed were the first combination tried in the simulation. These sensors were replaced when a set of sensor locations was tried that resulted in a lower detection time. The file shows the chronological order of optimal sensors as the simulation progressed. The sensors on the bottom row are the optimal sensor locations chosen. This file is shown in Figure

134 Figure 79: WQC File When the sensor placement tool was executed, an Excel file with the file name systemnametimematrix.csv was generated. The file includes all data used to determine the sensor location with the lowest time to detection. The first column shows all possible sensor nodes, and the first row of nodes represents the injection nodes. The values show the travel times between the injection node and sensor nodes (in minutes). If the cell shows 0 for a travel time, this means that the sensor nodes are too far away from the injection nodes and the contaminant will not reach the sensor within 24 hours. Therefore, the travel time is considered to be 24 hours for calculation purposes. This file can be accessed in the systemname.kyp folder, and a sample of the time matrix file is shown in Figure

135 Figure 80: Time Matrix Excel File 127