Professional Documents
Culture Documents
Calibrate Guidelines AWWA PDF
Calibrate Guidelines AWWA PDF
by
Engineering Computer Applications Committee
Hydraulic network simulation models are widely used by planners, water utility
personnel, consultants, and others involved in the analysis, design, operation, and
maintenance of closed-conduit hydraulic systems. Quite possibly the largest application
of hydraulic network models lies with the municipal water supply industry. The results
of network models have been used to assist in long-range master planning, short-term
project design, fire flow studies, daily operations, emergency response, energy
management, rehabilitation, troubleshooting, operator training, and water quality
investigations. Clearly multi-million dollar decisions can be and have been based on the
results provided by hydraulic network models. Consequently the results from the model
must bear close resemblance to the actual performance of the hydraulic system. In other
words, the computer model must be calibrated. The purpose of this paper is to present
guidelines for network model calibration. The degree of accuracy needed for model
calibration will be discussed within the context of the intended use of the model. For
example, a higher degree of calibration is necessary for a model that is used to examine
daily operations than is needed for a model used in long-range planning studies. The
paper will also discuss calibration approaches and testing procedures that can be used to
aid in the calibration effort.
By
Introduction
A computer model of a water distribution system is a valuable tool which can assist
engineers and planners in analyzing the hydraulic performance of water delivery systems.
Computer models of water supply systems are constructed for a variety of uses including:
It is not unusual for the results provided by a computer simulation model to be used in
capital projects involving several million dollars. As a result, it is imperative that the
results provided by the model bear close resemblance to reality. If this is not the case,
then the results provided by the model will be of limited value. Consequently, the model
must be calibrated. Computer model calibration can best be defined as the process of
adjusting data describing the mathematical model of the system until observed
performance, typically pressures and flow rates, are in reasonable agreement with
computer-predicted performance over a wide range of operating conditions.
The purpose of this paper is to provide a little deeper background on the various sources
of error that will produce differences between measured and computed system
performance. Armed with a strong understanding of the errors inherent in a network
model, we will briefly discuss the level of effort that can be involved in performing a
comprehensive network calibration. Finally we propose some calibration guidelines and
attempt to establish some criteria indicating a suitable level of calibration based upon the
intended use of the model. The purpose of this paper is not to establish standards for
model calibration. Rather members of the Engineering Computer Applications
Committee hope that the proposed criteria presented here will foster meaningful
discussion among the United States modeling community regarding the need and validity
of calibration standards.
Sources of Error
Consider the pipe shown in Figure 1. Notice the build-up, called turburculation, on the
interior pipe walls. The actual diameter of this pipe might be closer to 5 ½” or 5 ¼”
instead of 6”. Because the build-up on the pipe walls is totally random and irregular, it is
highly doubtful that the actual diameter, regardless of what it is, will remain the same
throughout the pipe length. Generally speaking it is usually best to use the nominal
diameter in a computer model and adjust pipe roughness values to achieve calibration.
Simulation models require a value for pipe roughness for each individual pipe segment.
This can result in quite a bit of information if one considers that it is not unusual for a
model of a moderately sized system to contain 500-1,000 pipes with larger systems
incorporating 2,000 or more pipelines. Certainly observed pressures can point the
modeler in the correct direction with regard to pipe roughness values. Nonetheless
compensating errors could cause incorrect pipe roughness values to be used.
Consider the simple parallel pipe system shown in Figure 2. Suppose that pressure
measurements have been taken at the node on each end of the pipe segment and the flow
through the system is known. The unknowns are the internal pipe roughness values for
each pipe. Table 1 shows the results of a simple analysis performed on this system.
Column 1 represents an assumed internal roughness for Pipe #1. Column 2 is the
resulting flow due to the measured head loss, the pipe characteristics and the assumed
roughness. Column 3 is the flow in Pipe #2 assuming a total system flow of 1,350 Gpm.
Finally column 4 represents the pipe roughness in Pipe 2 that is due to the head loss, the
pipe characteristics and the pipe flow.
Clearly there are multiple values of pipe roughness in pipes 1 and 2 that produce the same
head loss across the system. So the real question becomes which is the correct set of pipe
roughness values? The only way this question can be answered is to measure the flow in
one of the pipes. With the flow known the correct roughness values can be established.
Frequently in the United States actual pipeline flows throughout the system are not
The problem above illustrates the complexity for a simple two pipe system. The problem
complexity is compounded in real systems. For the system shown in Figure 2, let’s
assume that the flow into the system is not known and that the pressure at the upstream
node is also unknown. Then the number of combinations of flow, pipe roughness and
upstream pressure that would produce a pressure of 47 psi at the downstream node is very
large. Now consider that there are many parallel paths that water may take in real
systems. With the above illustration it is clear how compensating errors can result in
incorrect model parameters even though an exact match between computer predicted and
observed values is obtained.
Grouping or lumping water use at junction nodes instead of placing it at the actual
location where water is withdrawn from the system will produce some differences
between computer predicted and actual field performance. The differences should be
relatively minor assuming, of course, that the customer demand assigned to a particular
junction node is not located far away from the node. A bigger reason for the discrepancy
between computer predicted and field measured results lies with the spatial and temporal
nature of water use.
Water use within a home, and to a lesser extent, a commercial establishment is most
likely stochastic or random. In other words, one cannot predict with absolute certainty
when and how much water will be used at a particular residence. Perhaps the closest one
can get to certainty in residential water use lies with those homes that have automatic
lawn irrigation systems designed to operate at the same time each day. If one were to
meter the instantaneous water use in a specific residence the usage might look something
that shown in Figure 4. Industrial water use has a tendency to be a somewhat less
variable than residential or commercial usage. Any difference between the actual water
use in a distribution system and the estimates of that use supplied to a computer model
will generate errors between computer predicted and field measured performance.
As changes to a system are made then the system maps should be modified to reflect the
changes that have been made. Most water utilities are very good at making these changes
though some are not. Despite the best efforts of the mapping department, sometimes
systems grow at a faster pace than the mapping department can make the changes.
Obviously the accuracy of system maps can have a pronounced effect on the accuracy of
a model calibration effort.
Node Elevations
Frequently in hydraulic network simulation ground elevations will be used to find
pressures at locations throughout the hydraulic network. For most systems, pipes are
buried at a uniform distance below the ground surface usually anywhere between 3-6 ft
deep (1.3-2.6 psi). The pressure difference associated with using the ground level versus
the actual pipe centerline elevation is usually not that great. Moreover, it is easier to
gather elevation data from topographic maps than it is to scour through As-Built
drawings to find the actual elevations.
When pressure measurements are taken for model calibration they are usually taken at
fire hydrants or at some other direct connection to the water system. The elevation of the
pressure gauge at the hydrant can be several feet higher than the ground elevation or
several feel lower than the ground in the case of a meter vault. For calibration purposes
this elevation difference can make a difference. Therefore one should always use the
elevation of the pressure gauge, that is, the location where the pressure measurement is
taken when calibrating a hydraulic network model.
Effect of Time
If the hydraulic simulation is an extended period simulation (EPS), then the calibration
must also consider adjustments that should be made for time-varying conditions.
Generally this will mean adjusting model parameters until there is a reasonable
agreement between computer predicted and observed tank water levels. In addition to
tank water levels, time-varying pressures and pipeline flows can also be compared.
Many times information on tank water levels will be obtained from circular charts such as
the one shown in Figure 5. Notice that there can be fairly substantial changes in tank
water level over a short period of time. This is especially true for small storage tanks
such as standpipes or if the tank capacity is too small for the system it is supplying. If too
large a time step is used in hydraulic calculations then the true time-varying behavior of
Model Detail
When a computer model of an existing system is constructed, quite often a simplified or
skeletal version of the system will be developed instead of one which contains every pipe
in the system. A skeletal version of the system is one that does not include small
diameter pipes or does not include lines that have an insignificant influence the
hydraulics of the overall system. A skeletal model is one which is manageable and yet
accurate – when properly calibrated. Interestingly, Cesario, et al state that the economics
of model construction make a strong case for including all pipes in the mathematical
model [2].
Although skeletalization can have an impact on the calibration process, the effects of
skeletalization on calibration are, for the most part, unknown. If one is having difficulty
in achieving a reasonable calibration with a highly skeletal system, then perhaps it may
be necessary to include more lines in the model. For example, omitting a dense grid of
small diameter mains from a model simply because of their size may be inappropriate if,
as a group, they present a sizeable hydraulic impact on the system.
Geometric Anomalies
Even if high quality information on the physical attributes of the system is available and
good estimates of nodal demand is provided to the simulation model, there can still be
differences between computer predicted and observed performance. Anomalies in the
geometry of the system are usually to blame in these cases. For example, suppose that
we have a condition where two pipes cross one another as illustrated in Figure 6. From
the plan view it may appear that these two lines are connected. The cross-section view
shows otherwise. Obviously the model will not be calibrated if these pipes were assumed
to be connected.
Another issue related to geometric anomalies that could cause differences between
observed and predicted behavior lies with system isolation valves. Most of the time in
computer models gate valves, butterfly valves, etc. are assumed to be fully open. In some
systems this may not be the case. If one has to use unrealistically low roughness values to
obtain a suitable calibration then perhaps there are closed or partially closed isolation
valves in the system.
Not only can inaccurate pump curves cause a discrepancy between computer predicted
and observed system performance, but errors in any boundary element can cause
difficulty with the calibration. These boundary elements include regulating valve
settings, tank levels and pressure zone boundaries. Clearly an incorrect setting for
pressure reducing valve will produce pressures in the computer model that are
inconsistent with field measured values.
Whenever a measurement is taken we must be sure that the equipment or device that was
used to take the measurement is itself calibrated. Pressure measurements that were taken
with a poorly calibrated pressure gauge or pressure transducer are of limited value and
will frequently be discarded. Remember not only should the equipment used to take
measurement be calibrated but the pressure, flow, and tank level monitors for the
SCADA system should also be calibrated.
Even if measuring and monitoring equipment is well calibrated, other problems can arise
to give poor data. Suppose tank water levels is obtained from circular charts such as that
shown in Figure 5. Notice how the chart indicates the tank water level at a given time.
We need to make sure that the timing mechanism in the chart recorder is working. In
other words, does the level on the circular chart at 6:00 a.m. really correspond to 6:00
a.m. or some other time because the timing mechanism is slow or because the chart slips
or because someone installed the chart incorrectly. Issues such as these must be
considered when using analog monitoring equipment. Of course the Y2K issue may
create problems for digital systems that are not compliant.
In the preceding paragraphs we discussed some of the data requirements for hydraulic
network models. Key among this information is internal pipe roughness values for each
individual pipe segment. Providing accurate roughness coefficients for moderate to large
models can be a considerable undertaking. In fact, it is possible that each individual
internal pipe roughness value could be a candidate for adjustment during the calibration.
When calibrating for water quality, the level of effort required to obtain a suitable
calibration can be even more imposing. Most water quality models require information
Obtaining bulk reaction coefficients is not that difficult and procedures are available to
perform the necessary tests. Finding wall reaction coefficients can be much more
challenging. Just as each pipeline can have a unique internal roughness, it is possible that
each pipeline can have a unique wall reaction coefficient. Unlike roughness coefficients,
no well-established test is available for measuring reaction coefficients. Unlike roughness
coefficients, very little data is available for typical reaction coefficients. In fact, there is
some question within the modeling community if “typical” reaction coefficients even
exist.
In the case of water quality models, a match between some observed and predicted water
quality parameter such as chlorine or fluoride concentrations is usually used for
calibration purposes. However in the case of nonconservative species such as chlorine not
only are system-wide concentrations dependent upon the network hydraulics, but they are
a function of any reactions that take place in the system. Clearly one can see that
calibrating for water quality can greatly increase the level of effort required to obtain a
suitable match between observed and computer predicted performance. Interestingly as
Grayman points out, a water quality model can be used to help calibrate a hydraulic
model [5].
Calibration Methods
There have been a number of papers written on the subject of hydraulic network model
calibration and several methods have been developed to assist with the calibration effort.
However the traditional method of trial-and-error (trial-and-effort?) still seems to be the
predominate method of choice. In other words, data describing the model is adjusted –
sometimes randomly – until a suitable match is obtained.
The lack of calibration methods is not surprising once the full complexity of the level of
effort required for a comprehensive calibration is understood. Ormsbee and Reddy
provide a nice summary of the steps necessary to calibrate a hydraulic network computer
model [3]. In their paper they note that direct solution techniques or optimization
approaches offer some potential for simplifying the calibration effort. However for
moderate to large systems calibrating a model can be a considerable undertaking.
Direct parameter calculation methods and optimization techniques are not immune to the
problem of compensating errors. This problem can be addressed somewhat with
optimization methods by placing constraints on the decision variables, e.g. pipe
roughness values, water demand, valve setting, etc. With the direct parameter calculation
methods engineering judgment is usually the filter which captures unrealistic model
parameter values. Nonetheless, the problem of compensating errors may still exist.
It should be clear by now that the level of effort required to obtain an acceptable level of
calibration can be considerable. Water quality calibrations can be even more challenging.
Many modelers in the United States agree that the level of effort required to calibrate a
hydraulic network model will depend upon the intended use of the model [2, 3, 4]. In
other words, it may not be reasonable to expect that a model used for long-range planning
purposes will have the same level of accuracy than a model intended for operations or
water quality investigations. Walski provides some suggestions on a suitable level of
calibration as a function of the use of the model [4].
In the United Kingdom, however, standards or performance criteria that modeler should
strive for have been established. The criteria for flow and pressure are shown below.
Additional criteria also exist including those for extended period simulations [6].
The manner by which hydraulic networks are operated in the United Kingdom goes a
long way to explaining why calibration criteria exist in the UK. For example, in the UK
it is not unusual for water utilities to provide a lower level of fire protection than is
provided in the United States. Perhaps one of the bigger differences lies in metering. In
the UK most individual service connections are not metered. Rather large networks are
typically divided into smaller “demand management areas” each serving about 1,000 –
2,000 connections. The demand areas are independent service areas that are metered at
all inflow and outflow points. As a result, there will be more “in-line” flow metering
Calibration Accuracy
Despite the many variables that contribute to differences between observed and predicted
system performance and even though calibrating a hydraulic network model can be a
time consuming and sometimes frustrating task, the time is right in the United States to
consider adopting some calibration criteria. The results of computer simulations are
frequently used as the basis for capital projects involving many millions of dollars.
Moreover, the 1997 EPA Drinking Water Needs Survey has placed the price tag over the
next 20 years of bringing water distribution systems up to compliance with the Safe
Drinking Water Act at nearly $140 billion [7]. It is likely therefore that in the near future
computer simulation will be even more widely used for planning, design, operation and
water quality investigations.
As stated earlier, the primary purpose of this paper is to suggest calibration guidelines
that can be used in the United States to determine a suitable level of model calibration for
a particular use of the model. The purpose of this paper is not to establish calibration
standards. The Engineering Computer Applications Committee hopes that by providing
these criteria some discussion among the network modeling community in the United
States (and abroad) will ensue regarding the need for such criteria and the validity of the
criteria.
Minimum calibration criteria are presented in Table 2. Note that while a hydraulic
network model may have a number of different uses, four basic use categories are
presented in Table 2. Most of the uses of hydraulic models should fit into one of the
following categories: 1) Planning, 2) Design, 3) Operations and 4) Water Quality. For
example training system operators, developing emergency response plans, or performing
energy efficiency studies would all fall under the Operations category.
Six calibration criteria have been identified. The level of detail criteria relates the degree
of skeletonization that a model may have while the type of time simulation reflects
whether the model should be a steady-state or extended period simulation model. The
number of pressure readings is expressed as a percentage of the number of nodes in the
model while the number of flow readings is presented in terms of the number of pipes in
the model.
Note that the criteria represent minimum criteria. For example, based on the proposed
criteria a model used for long-range planning may have a low level of detail. However
there is nothing to prevent a model used for long range planning from being a model with
a high degree of detail. Note that the level of detail and the type of time simulation will,
in large measure, depend upon the use of the model.
[1] Walski, Thomas M., Sharp. Wayne W. and Shields, F. Douglas Jr. 1988 “Predicting
Internal Roughness in Water Mains,” Miscellaneous Paper EL-88-2, US Army Engineer
Waterways Experiment Station, Vicksburg, Miss.
[2] Cesario, A. L., Kroon, J.R., Grayman, W., and Wright, G. “New Perspectives on
Calibration of Treated Water Distribution System Models,”
[3] Ormsbee, L.E. and Lingireddy, S. “Calibrating Hydraulic Network Models,” Journal
American Water Works Association, Vol. 89, No. 2, pp. 42-50, 1997.
[4] Walski, Thomas M. “Standards for Model Calibration,” 1995 AWWA Computer
Conference, Norfolk, VA.
[5] Grayman, Walter M. “Use of Tracer Studies and Water Quality Models to Calibrate
a Network Hydraulic Mode,” Essential Hydraulics and Hydrology, Haestad Press,
Waterbury, CT., 1998.
[6] Water Authorities Association and WRc, 1989, “Network Analysis – A Code of
Practice,” WRc, Swindon, England.
Table 1
Pipe Roughness Values and Flow for Simple Two-Pipe System
Intended Use Level of Detail Type of Time Number of Accuracy of Number of Accuracy of
Simulation Pressure Pressure Flow Readings Flow Readings
Readings1 Readings
Long-Range Low Steady-State 10% of Nodes ± 5 Psi for 100% 1% of Pipes ± 10%
Planning Or EPS of Readings
Design Moderate to Steady-State 5% - 2% of ± 2 Psi for 90% 3% of Pipes ± 5%
High Or EPS Nodes of Readings
Operations Low to High Steady-State 10% - 2% of ± 2 Psi for 90% 2% of Pipes ± 5%
Or EPS Nodes of Readings
Water Quality High EPS 2% of Nodes ± 3 Psi for 70% 5% of Pipes ± 2%
of Readings
Notes:
1. The number of pressure reading is related to the level of detail as illustrated in the table below.
Figure 1
Pipe 1
L = 2,300 Ft
D = 10”
P = 54 Psi P = 47 Psi
1350 Gpm
Pipe 2
L = 2,800 Ft
D = 8”
Figure 2
Q Q Q Q Q Q Q Q Q
J-23 J-24
Q Q Q Q Q Q Q Q Q
Figure 3
Water Use
Time (Hrs)
Figure 4
6 pm 6 am
2-Hrs
A
Noon
Figure 5
Top View
SideView
Figure 6