ANALYSIS OF THE 1998 WATER-DISTRIBUTION SYSTEM
SERVING THE DOVER TOWNSHIP AREA, NEW JERSEY:

Field Data Collection Activities and
Water Distribution System Modeling

HYDRAULIC MODEL CALIBRATION

Overview

With a computer model of the water-distribution system, we are trying to reproduce the behavior of a "real-world" hydraulic system as closely as feasible in terms of spatial and temporal characteristics. The collection of field data (previously described) provides an opportunity to understand the operation of the real system at a specified number of locations and times. Such efforts are consistent with the findings of the American Water Works Association Engineering Computer Applications Committee which indicate that "true model calibration is achieved by adjusting whatever parameter values need adjusting until a reasonable agreement is achieved between model-predicted behavior and actual field behavior" (AWWA Engineering Computer Applications Committee 1999). Once a model is considered to be calibrated, it can then be used to, among other purposes, estimate hydraulic characteristics of the real-world system at locations where measured data are unavailable or unknown, spatially and temporally.

In the United States, definitive standards to assess the accuracy of model calibration have yet to be agreed upon or established. However, the following calibration criteria have been suggested:

1. An average pressure difference of ±2.2 psi with a maximum difference of ±7.3 psi for a "good" data set, and an average pressure difference of ±4.3 psi with a maximum difference of ±14.2 psi for a "poor" data set (Walski 1983); and
   
   
2. The difference between measured and simulated values should be ±5 psi to ±10 psi (Cesario and Davis 1984).

We have used these criteria as general guidelines and have taken into account the availability and accuracy of the data for the water-distribution system serving the Dover Township area. Therefore, we selected a pressure difference at the test-hydrant locations (difference between measured and simulated pressure) of ±5 psi to ±7.5 psi as the calibration criteria for the model of the Dover Township area water-distribution system.

According to the AWWA Engineering Computer Applications Committee (1999), 10 sources of possible error could cause poor agreement between simulated model values and measured field values. These sources of error, which provide a potential list of factors that can be adjusted during the model-calibration process, are: (1) errors in input data (measured and typographic), (2) unknown pipe roughness values (i.e., Hazen-Williams "C-Factors"), (3) effects of system demands (distributing consumption along a pipe to a single node), (4) errors in data derived from network maps, (5) node elevation errors, (6) errors introduced by time variance of parameter values such as storage tank water levels and pressures, (7) errors introduced by a skeletal representation of the network as opposed to modeling all small-diameter pipes, (8) errors introduced by geometric anomalies or partially closed valves, (9) outdated or unknown pump-characteristic curves, and (10) poorly calibrated measuring equipment including data loggers, tank water-level monitors, and SCADA systems. We will discuss these sources of possible model error as they relate to calibrating the model of the distribution system serving the residents of Dover Township in a subsequent section of this report (see section on "Calibration Procedures").

Requirements for Model Input

The EPANET water-distribution system model (Rossman 1994) was used in conjunction with the collected 1998 field-test data (previously described in the section on "Field-Data Collection Activities") to develop and calibrate a model of the present-day (1998) water-distribution system serving the Dover Township area. Information required to conduct a simulation using EPANET include data describing pipeline characteristics, booster pump characteristics, groundwater-well pumping factors, consumption and diurnal demand patterns, tank geometries and initial water levels, and simulation time parameters. Table 5 describes the set of input data properties needed to model components of the water-distribution system serving the Dover Township area using EPANET. Specific requirements for the creation of a data input file and the necessary data formats are provided in the EPANET Users Manual (Rossman 1994). (The manual is available on the Internet at the following address: http://www.epa.gov/ORD/NRMRL/wswrd/epanet.html. The data for developing the physical network attributes (e.g., pipe lengths, diameters) required for the EPANET input file for simulating conditions of March and August 1998 were obtained from databases supplied by the water utility (Flegal, 1997), and from the field data ATSDR collected (refer to section on "Field-Data Collection Activities"). Sources for the data and model parameter values are listed in Table 6 and are described below.

Junction Data

EPANET identifies junctions (or nodes) as the beginning and ending points associated with each pipe or pipe segment in the model network. Each junction is assigned an alpha-numeric identification label, an elevation, a demand (or consumption) value, and a demand pattern number (Table 5). Because the goal of the investigation is to conduct a population-based assessment, geo-spatial location information for pipe junctions, pipelines, and network facilities is required. Geographic coordinates of the model network (in decimal degrees and New Jersey State Plane coordinates) were determined using GPS equipment to obtain locations of the 50 test and alternate hydrants described above, in conjunction with TIGER/Line™ files (1992) for the Dover Township area (coordinate values of the hydrants are listed in Table 2). These known coordinates were used to geo-reference all model nodes (and links) in the distribution-system network. The model of the 1998 water-distribution system has 14,987 junctions or nodes.

Table 5. Set of input data properties required by EPANET to model the water-distribution system serving the Dover Township area, New Jersey¹

¹From Rossman (1994, 1999).

Table 6. Sources for data and model parameter values used to construct the water-distribution system model, Dover Township area, New Jersey

For the model of the Dover Township area, nodal values of elevation were derived by relating the geo-spatial locations of the pipe junctions to elevation data derived from the USGS 7½-minute DEM quadrangles that cover the Dover Township area using GIS software. For modeling, pipelines were assumed to be located at land surface and not buried below ground level. For junctions assigned to test-hydrant locations, elevations for land surface at the hydrant locations were determined using GPS equipment and verified using DEM data, as previously explained (See section on "Hydrant Selection").

Demand or consumption was assigned to model nodes based on data provided by the water utility. For the Dover Township area, metered consumption data were available for the area serviced by the water utility solely on a quarterly basis. Thus, each meter is read four times per year; however, all meters are not read at the same time or even within a few days of each other. Quarterly consumption data for October 1997 through April 1998, representing the distribution of consumption in the Dover township area, were used for both the March and August 1998 simulations. Metered data specifically representing the distribution of consumption for the August 14-16, 1998, test, or for peak-demand conditions, were not available to investigators for simulating the August 1998 test. To use the model to simulate network configurations that differ from the present-day (1998) network (i.e., historical network configurations), data on total demand and the nodal distribution of consumption should be obtained.

The quarterly consumption data for October 1997 through April 1998 obtained from the water utility for all meter locations in the distribution system were averaged and allocated to model nodes. This was accomplished by using an address-matching technique and GIS software to locate the water meter address provided by the water utility with the closest model node. If more than one meter address was adjacent to a model node, then the consumption for the node was the sum of all metered consumption adjacent to the node. The spatial distribution of average consumption assigned to model nodes is shown in Plate 7. Values range from 0.001 gallons per minute (gpm) to about 6.1 gpm with a mean of about 0.4 gpm. (In EPANET, a positive demand or consumption value indicates outflow from the network; a negative demand value indicates inflow or supply to the network.)

Supply of water to the distribution system can be input to EPANET by assigning a negative demand value. Using this approach, groundwater wells supplying either storage tanks or pumping directly into the distribution system (Table 1) were simulated by assigning the wells to a model node and specifying a negative demand value for the node. The demand for the node was set equal to the rated capacity of the well (Table 1) for individual wells (e.g., well 21), or the combined rated capacity for a group of wells in a well field (e.g., Parkway well field wells). The actual amount of water supplied by the wells on an hourly basis when they were pumping was varied by use of a demand pattern (discussed below) associated with the well node.

The last parameter associated with junction data is the demand pattern. With this parameter, EPANET has the ability to modify the nodal demand data based on the demand pattern. For example, if the water utility serviced residential, commercial, and industrial users, each group of water users might have a different diurnal demand pattern and therefore, nodal demand data would need to be modified depending on the type of use. This is accomplished in EPANET by assigning a demand pattern number to each junction. To enter demand patterns for the Dover Township area, the diurnal demand pattern specific to each type of user, derived from diurnal demand data (e.g., Figure 1), is required. However, investigators did not have this information. Rather, the only demand data available was the data obtained from the water utility's SCADA system for the supply of water to the overall system (Figure 1). Therefore, for the Dover Township area, all model nodes that were assigned a positive demand value (indicating outflow from the system) used the same diurnal demand pattern--the same demand pattern number was assigned to each model junction identified as having a positive consumption value. Each node representing an individual groundwater well or a combination of wells in a well field (thus having a negative demand value) was assigned a unique demand pattern number. (The hourly demand factors associated with the demand pattern numbers are described below in the section on "Pattern Data.")

Tank Geometry and Initial Water-Level Data

Ground-level and elevated storage tanks (Plate 2) are associated with model junctions in EPANET. The parameters used to describe storage tanks in EPANET are listed in Table 5; specific features for each storage tank are listed in Table 7. For this investigation, all storage tanks were modeled as having cylindrical geometries. The initial water level for each tank was determined from data collected by ATSDR staff monitoring the water utility's SCADA system during the March and August 1998 tests. For the ground-level storage tanks, the elevation of the bottom of the tank was determined by using GPS equipment and verified using the USGS 7½-minute DEM data for the Dover Township area. For the elevated storage tanks, the elevation of the bottom of the tank was obtained from data supplied by the water utility (Flegal 1997).

Pipeline Data

Data pertaining to the pipeline characteristics constituting the distribution system network were retrieved from electronic computer-aided-design files supplied by the water utility. Parameters required by EPANET to describe pipes include (Table 5): a pipe identification label, starting and ending node labels, length, diameter, roughness coefficient, and the status of the pipe (open or closed). The model network consists of 16,071 pipe segments or links. Table 8 lists the material type for pipes composing the network as of the end of December 1997, the range of pipe diameters for specific material types of pipe, and estimated values for the Hazen-Williams "C-factors" (or roughness coefficient) assigned to pipes for use in model calibration (see section on "Calibration Parameters" for a discussion of Hazen-Williams "C-factors"). Spatial distribution of the network pipes classified by diameter and by roughness coefficient are shown on Plates 8 and 9, respectively. As shown on Plate 8 and listed in Table 8, the model network of the distribution system is composed of pipes ranging in diameter from 2 in. to 16 in. Additionally, data in Table 8 (column 6) show that most of the network is composed of asbestos cement (60%) and plastic (PVC, 37%) pipes.

Table 7. Storage tank identification, hydraulic data, and dimensions used for EPANET simulations of the water-distribution system serving the Dover Township area, New Jersey

¹Datum is sea level.
²Effective diameter for two tanks simulated as one tank in EPANET.

Table 8. Pipeline characteristics of the water-distribution system, Dover Township area, New Jersey¹

¹Data for water-distribution system network pipelines as of December 1997.
²Data for material type, year first installed, year last installed, and range of pipe diameters from Flegal (1997).
³Values for Hazen-Williams "C" from Rossman (1994, Table 2.2).

Pump Data

Booster pumps are used in water-distribution systems to raise the hydraulic head of water and increase the pressure in certain portions of a system. In EPANET, pumps are modeled as separate links. As described in Table 5, each pump in the model is identified by a numeric identification label (Table 9), a start and end node label, and a head-discharge or pump- characteristic curve. The pump-characteristic curve describes the relationship between the hydraulic head imparted to the fluid (water) as a function of the flow rate of the fluid through the pump (Table 9; Appendix F). In EPANET, the pump-characteristic curve is represented as a function with the form of:

where:

EPANET requires a minimum of 3 points--the shutoff head, h₀, and two additional points on the pump characteristic curve (Table 9; Appendix F) to estimate values for coefficients a and b. Initial pump-characteristic curve data for every booster in the water-distribution system were provided to ATSDR by the water utility (Flegal 1997). These data are listed in Table 9 and shown graphically in Appendix F. The initial pump-characteristic curve data were modified during the calibration process (see section on "Calibration Parameters").

Pump Control Data

To control the on/off cycling of booster pumps, EPANET uses pump control data. These data include (Table 5): a pump identification label, the type of control (time or level), the pump setting (open or closed), and the control setting (time or tank water level). When a "time" control is used, pumps are cycled on and off using the time of day as the controlling criterion. When a "level" control is used, pumps are cycled on and off using storage tank water-level as the criterion. Input data used to simulate the March and August 1998 tests were obtained from field data ATSDR staff collected monitoring the water utility's SCADA system. During the test periods, the water utility used a time-of-day control setting (rather than a tank water-level control setting) to cycle booster pumps on and off (Table 9) .

Table 9. Pump identification, characteristic data, calibrated values, and status during March and August 1998 pressure tests, Dover Township area, New Jersey

¹From Flegal (1997).

Pattern Data

EPANET allows for varying of demand values by using a demand pattern number. Pattern data are entered into EPANET by specifying an alpha-numeric pattern identification label, and then supplying factors by which the nodal demand value is to be multiplied. The 48 hourly demand factors for water consumption used in the simulations are listed and presented graphically in Appendices G (March 1998) and H (August 1998). These factors, when multiplied by the nodal consumption, represent the diurnal demand that occurred during the time of the tests in March and August 1998. The 48 hourly demand factors were derived from the water utility's SCADA system demand data recorded by ATSDR staff during the March and August 1998 tests (Figure 1). The demand factors were obtained by using the demand data recorded by ATSDR staff for March 1998 (Figure 1A) and August 1998 (Figure 1B) averaged over a period of one hour and dividing the values by the average demand of 7.6 MGD for March 1998 or 16.1 MGD for August 1998 (compare Figure 1A with demand factors for March 1998 shown in Appendix G and Figure 1B with demand factors for August 1998 shown in Appendix H).

As described above in the section on "Junction Data," supply of water from groundwater wells to the distribution system and to storage tanks was represented by using a negative base demand equal to the rated capacity of a well or well field (Table 1) at the node corresponding to the well or well-field location. The base demand was modified by multiplying a pumping factor for each hour of the simulation by the base demand (the rated capacity of the well). In this manner, supply was provided to the distribution system equaling the hourly amount metered by the water utility's SCADA system and recorded by ATSDR staff during the tests. Tables listing and graphs showing the pumping factors for wells in operation during the tests are provided in Appendices G (March 1998) and H (August 1998).

Time Parameter Data

EPANET assumes that consumption values, supply rates, and concentrations at source nodes remain constant over a fixed period of time. However, these parameter values can change from one time period to another. To conduct an extended period hydraulic simulation, EPANET requires three time parameters: (1) the duration of the simulation, (2) the hydraulic time-step size, and (3) the pattern time-step size. For the Dover Township simulations, the duration of the simulation was set equal to the duration of the tests-48 hours. The hydraulic and pattern time-step sizes were set equal to 1 hour, which is the default time-step size used by EPANET.

Calibration Procedures

As described in the "Overview" section, model calibration entails adjusting model parameter values until an acceptable match is achieved between measured data and model-simulated values (i.e., pressures at the test hydrants, water levels in the storage tanks, flows from booster pumps, and pumpage from groundwater wells). The Dover Township water-distribution system model was calibrated to the hydraulic and operational data collected during the March 1998 test and then further tested against data collected during the August 1998 test. The model was run as an extended period simulation (EPS) using one-hour hydraulic time steps and demand-pattern factors derived from the control-room data collected during the test. The 10 sources of possible error that could lead to model simulated values not agreeing with measured values discussed previously are listed in Table 10. These sources of error also provide a list of potential model parameters that can be modified during the calibration process. To decide which parameters might require more, less, or no modification, investigators evaluated each parameter as to the qualitative magnitude of error (high, moderate, or low) that could result from uncertainty and variability of the parameter. These evaluations are listed in Table 10. Three of the sources of possible error were evaluated as having a qualitatively high or moderate error magnitude: (1) unknown pipe roughness (Hazen-Williams "C-factor") values, (2) effects of system demands and consumption, and (3) outdated or unknown pump-characteristic curve data. The initial estimates for these three parameters where subjected to possible variation during the calibration process and will be discussed below. The remaining 7 sources of possible error are believed to introduce minor to insignificant errors to model simulations, and therefore, were not modified during the calibration process.

For model calibration (March 1998 test) and testing (August 1998 test), four data comparisons (measured data versus simulated values) were made during the calibration and testing process and these comparisons are summarized in Table 11 and presented in the accompanying appendices. The data comparisons are for (Table 11): (1) pressure at each of the 25 test hydrants (Appendix I [March 1998] and Appendix J [August 1998]), (2) storage tank hydraulic head at each storage tank that was operational during the tests (Appendix K [March 1998] and Appendix L [August 1998]), (3) booster pump flows for each booster pump that was operated during the tests (Appendix M [March 1998] and Appendix N [August 1998]), and (4) groundwater well pumpage for each well that was operated during the tests (Appendix O [March 1998] and Appendix P [August 1998]). In each of these appendices (I-O), graphs and tables that compare measured data with model simulated results are provided.

Table 10. Qualitative evaluation of sources for model error, water-distribution system model, Dover Township area, New Jersey

¹List of error sources from AWWA Engineering Computer Applications Committee (1999)

Table 11. Summary of comparisons between test data and model simulations for the March and August 1998 pressure tests described in report, Dover Township area, New Jersey

Calibration Parameters

Discussions with the water utility indicated that the network pipes were believed to be very clean, and inspections had shown very little debris. In addition, as shown in Table 8, more than one-third of the pipes (in quantity and lengthwise) are made of PVC where the variation in "C-factor" is negligible. Therefore, initial estimates for "C-factor", obtained from published tabular values (Rossman 1994) and listed in Table 8 for every pipe material type, were not varied during the calibration process. A sensitivity analysis, subsequent to model calibration, (see section on "Sensitivity Analysis") confirms that for this distribution system, variation in "C-factor" has little influence on system pressures and flow directions.

ATSDR staff obtained initial estimates of system demand factors by recording demand data (water production by the utility) during the tests from the control-room SCADA output. During the calibration process, individual hourly factors were modified. Factors for the March test differ from factors for the August test because the conditions of the each test represented different demand conditions-winter-time demand for the March test and peak demand for the August test. The rational used for modifying the individual hourly demand factors for each test follows in the discussion below.

Initial estimates for the demand factors were derived from the instantaneous output of the water utility's SCADA system (Figure 1). These data were recorded at random time intervals and then averaged over a period of one hour. As can be seen in Figure 1, these data are quite variable in times of significant system-demand change. Therefore, initial estimates of demand factors, based on an average over a period of one hour, may not have accurately reflected or have been representative of the system's demand. Therefore, investigators felt justified in modifying the initial estimates of the demand factors. It is important to note, however, that although individual hourly factors were modified, the total system-wide demand, 7.6 MGD for March 1998 and 16.1 MGD for August 1998, was not modified during the calibration process.

Calibrated values for the hourly demand factors for the tests are provided in Appendix G (March 1998) and H (August 1998) are in general agreement with the individual system demand patterns obtained from the water utility's SCADA system (compare Figure 1A with the demand factors shown in Appendix G and Figure 1B with the demand factors shown in Appendix H). Use of system-demand factors in conjunction with measured and recorded hydraulic device operations (pump on/off status, groundwater well status, etc.) helped investigators calibrate the model to conditions of March 1998 and further test the calibration to conditions of August 1998.

The last model parameter adjusted during the calibration process was the booster pump-characteristic curves. The water utility provided initial characteristic-curve data to ATSDR (Flegal 1997). These data are listed in Table 9 and shown graphically in Appendix F. However, the source of these data (e.g., manufacturer's data, field testing) could not be determined. Therefore, investigators believed this was a key parameter that could be modified during the calibration process. Modifications to the original data were made to increase or decrease the characteristic curve so that system operations, as observed during the two tests, could be duplicated as closely as possible. These modifications, however, still provide typical and reasonable head-discharge curve relationships for the water-distribution system. The calibrated characteristic-curve data along with the status of the pumps (on/off) are listed in Table 9; they are shown graphically in Appendix F for the March and August 1998 tests. Of the 11 system pumps listed in Table 9, 3 pumps (Holly booster 2, Parkway booster 2, and Route 37) required very little or no modification during the calibration process. Two pumps (South Toms River booster 1 and 2) were not operated during either the March or August 1998 tests.

Next Section     Table of Contents