Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
7Activity
0 of .
Results for:
No results containing your search query
P. 1
Uncertainty in Water Distribution Network Modeling

Uncertainty in Water Distribution Network Modeling

Ratings: (0)|Views: 308|Likes:
Published by Youssef
Uncertainty in water Distribution Network Modelling
Uncertainty in water Distribution Network Modelling

More info:

Published by: Youssef on Dec 11, 2009
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

06/24/2013

pdf

text

original

 
22
Uncertainty In Water Distribution Network Modeling
Kevin Lansey 
Department of Civil Engineering and Engineering Mechanics,University of Arizona
Introduction
Water distribution network modeling requires input frommany sources. Much of this data cannot be measureddirectly and it is rarely known precisely. Inputimprecision is propagated to uncertain model predictions.This paper discusses a technique for quantifying theuncertainty of model results based on the variability inmodel input.Once the magnitude of the uncertainty is understood, anengineer can use that information in several ways. Basedon the prediction uncertainty in present or futureconditions, safety factors can be included in design of network improvements. To enhance model predictions,additional field data can be collected which will lead tomore precise model parameters. Using the predictionuncertainty, the optimal network conditions andmeasurement locations can be identified. Finally,different network representations can be evaluated bycomparing their predictive uncertainties.The model input needed to compute the pressure headsand pipe flows are the nodal demands, pump and pipecharacteristics, tank dimensions, valve settings, andcoefficients describing energy losses in other appurtenances. Some of this data, such as pipe diameter and length, is accurately known and available fromconstruction plans. Other information for pipes, in particular the pipe roughness, varies over time asencrustation occurs. Nodal demands are difficult to predict since they also change over a short time and because they are modeled as lumped at single locationsrather than distributed along a pipe. As such, both parameters are not measured exactly but estimatedindirectly from other field measurements. This paper focuses on the uncertainties in pipe roughnesscoefficients and nodal demands.
Pipe network hydraulics
Flow in a pipe network satisfies two basis principles,conservation of mass and conservation of energy.Conservation of mass states that, for a steady system, theflow into and out of the system must be the same. Thisrelationship holds for the entire network and for individual nodes. A node is included in a network modelat (1) a demand location and/or (2) a junction where twoor more pipes combine. One mass balance equation iswritten for each node in the network as:where Q
in
and Q
out
are the flows in pipes entering or exiting the node and Q
demand
is the user demand at thatlocation. These demands are uncertain since they areestimated from the local user base that cannot be predicted exactly since they vary nearly continually. Inaddition, the demand is typically represented as a lumpeddemand for users near the node.The second governing equation is a form of conservationof energy that describes the relationship between theenergy loss and pipe flow. The most commonly usedhead loss equation for water networks, the Hazen-Williams equation, will be the only such relationshipconsidered in this paper. In English units, the equationis written as:where D is the pipe diameter, L is the pipe length, Q isthe pipe flow, and C is the Hazen-Williams piperoughness coefficient. h
l
is the head or energy loss in the pipe. H
i
and H
 j
are the energy at nodes at the ends of the pipe measured in dimensions of length. Using thisequation, conservation of energy can be written inseveral ways. Most often, it is written for energy lossaround a loop. The two conservation relationships can be used to develop a set of nonlinear equations that can be solved for the pipe flows, Q, and nodal heads, H,
 
23which due to uncertainty in input are inherentlyuncertain.
Quantifying uncertainty
If the variability in input parameters, C and Q
demand
, can be quantified several techniques exist to determine theresulting uncertainty in model output. Monte Carloanalysis (MC) is a general enumerative approach for computing the statistics of a model’s output.Given the input parameters statistics and probabilitydistributions, one iteration of Monte Carlo analysisconsists of randomly generating a set of input parameters. The numerical model, in this case thehydraulic relationships, is then solved to determine themodel output. In this study, KYPIPE (Wood, 1981) isused to compute the nodal pressure heads.To complete a full analysis consisting of a large number of iterations, a set of random input is generated and their corresponding output is determined. The statistics(e.g. mean, variance, and standard deviation) of themodel output can then be computed. If a sufficientnumber of Monte Carlo iterations are completed, theoutput statistics will converge to their actual values. The probability distribution of the model output can also beexamined.To estimate the predicted head uncertainty, Monte Carloanalysis was applied to several networks and levels of input uncertainty in Araujo (1992). Figure 1 shows oneof the systems considered and Table 1 provides the inputstatistics for pipe roughnesses. Table 2 summarizes theresults estimates of the predicted head variances. It wasassumed that the roughness coefficients followed anormal probability distribution. As the input uncertaintyincreased, the output variability increased at nearly thesame rate.TABLE 1: Pipe roughness statistics for selected pipesPipe13589Low standard deviation0.91.50.41.01.4Med. standard deviation4.37.72.05.27.1High standard deviation9.115.64.110.013.2TABLE 2: Standard deviations of nodal pressure heads for selected nodesLow pipe standarddeviationMedium pipe standarddeviationHigh pipe standard deviation NodeMCFOSMMCFOSMMCFOSM10.040.040.200.210.400.4220.400.412.032.104.004.1330.080.080.420.440.820.8560.030.030.160.170.320.34Trace*0.180.194.745.0818.4419.65*Trace is the sum of the diagonal elements of the covariance matrix, i.e., variancesAs shown by Araujo, the distribution of the pressureheads followed a normal distribution until a high inputuncertainty level was evaluated. This result allows oneto draw conclusions about the confidence level in model
 
24 predictions. For example, the 90 percent confidencelevel that the pressure head at a node is greater than adesired value can be computed. Thus, when designing asystem expansion, an engineer can consider confidencelevels on pressure heads as well as their expected values.Monte Carlo analysis requires a large number of modelevaluations and long computer times to determine thedesired statistics. Alternative approaches that are lesscomputationally intensive have been proposed in thestatistics literature. One popular class of methods isknown as point estimation approaches with the simplest being the first order second moment approximation(FOSM). To estimate the mean and variance of modeloutput, FOSM requires computing the model output at asingle point and determining the derivative of modeloutput to model input (i.e., change of model output dueto a change in model input). In the worst case, thesederivatives can be computed by finite differences withone additional model evaluation for each input parameter. The variance of the model output, e.g., thenodal pressure heads, can be expressed as a function of the variance of the model parameter, e.g., Hazen-Williams roughness coefficient, as:where cov(H) is the covariance matrix of the predicted pressure head and cov(C) is the covariance matrix of theroughness coefficient. The derivative terms are thematrix of representing the change in pressure heads dueto a change in roughness coefficients. The key result inthe cov(H) matrix is the diagonal elements that are thevariances of the individual pressure heads. The meanvalue of the output is assumed to occur when the valueof the input is evaluated.FOSM was used to predict model uncertainties of thenetwork in Figure 1 with results again summarized inTable 2. Although FOSM is an approximationtechnique, the results compare very favorably withMonte Carlo analysis. However, the difference incomputations is dramatic. Thus, FOSM was used for theanalysis in the following sections. FOSM, however, begins to diverge from Monte Carlo analysis at highlevels of uncertainty. This result is due to modelnonlinearity at higher uncertainty levels while theFOSM, by using only the first derivatives, neglects thisnonlinearity.
Collection of field data to reduce uncertainties
If input uncertainty is very high, a model is likely poorlycalibrated and additional field data should be collectedto reduce the input uncertainty. Uncertainty analysis canassist in determining where and under what conditionsadditional data should be taken to improve modelcalibration. Araujo (1992) developed a methodology for comparing the utility of field data collected from a set of  potential demand conditions. The data collectionmethodology is based on lowering the uncertainty levelat one or more nodes through a reduction in one or more pipe roughness coefficient uncertainty.Roughness uncertainty is dependent upon fieldmeasurements. Therefore, a two level gradient wascomputed. The first level estimates the change in piperoughness uncertainty given a demand pattern andassumed measurement locations. Extreme demandconditions should be selected, such as simulated fireloads, that stress the network and cause high energy loss.These types of demand conditions can be induced byopening one or more fire hydrants and supply usefulinformation for improving parameter estimates. Theuncertainty analysis is completed using the FOSMmethod assuming that the predicted demand occursusing:where H
M
is the measured pressure heads and cov(H
M
)is the covariance matrix for the measured heads. Thecovariance of H
M
varies depending upon the device usedfor measurement and the accuracy of determiningelevations. The gradients are evaluated at the bestestimate of the C values and measured heads.The second level of the methodology computes the propagation of the expected new pipe roughnessuncertainties to nodal pressure head predictions. Thesecond step is identical to the uncertainty predictionsdescribed in the previous section. By comparing thereductions in pressure head uncertainties for differentdemand patterns, the conditions that provide the most

Activity (7)

You've already reviewed this. Edit your review.
1 hundred reads
1 thousand reads
d44u liked this
dinam111 liked this
suedas liked this
stuarthenryjames liked this

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->