(2004) Wahl - Uncertainty of Predictions of Embankment Dam Breach Parameters PDF
(2004) Wahl - Uncertainty of Predictions of Embankment Dam Breach Parameters PDF
Breach Parameters
Tony L. Wahl1
Abstract: Risk assessment studies considering the failure of embankment dams often require the prediction of basic geometric and
temporal parameters of a breach, or the estimation of peak breach outflows. Many of the relations most commonly used to make these
predictions were developed from statistical analyses of data collected from historic dam failures. The prediction uncertainties of these
methods are widely recognized to be very large, but have never been specifically quantified. This paper presents an analysis of the
uncertainty of many of these breach parameter and peak flow prediction methods. Application of the methods and the uncertainty analysis
are illustrated through a case study of a risk assessment recently performed by the Bureau of Reclamation for a large embankment dam
in North Dakota.
DOI: 10.1061/共ASCE兲0733-9429共2004兲130:5共389兲
CE Database subject headings: Dam failure; Uncertainty analysis; Peak flow; Erosion; Dams, embankment; Risk management.
vatively designed to underpredict the reported time more often The uncertainty analysis was performed using the database
than they overpredict, but overprediction errors of more than one- presented in Wahl 共1998兲, with data on 108 case studies of actual
half of an order of magnitude did occur several times. embankment dam failures, collected from numerous sources in
The first question that must be addressed in an uncertainty the literature. The majority of the available breach parameter and
analysis of breach parameter predictions is how to express the peak flow prediction equations were applied to this database of
results. The case study datasets used to develop most breach pa- dam failures, and the predicted values were compared to the ob-
rameter prediction equations include data from a wide range of served values. Computation of breach parameters or peak flows
dam sizes, and thus, regressions in log–log space have been com- was straightforward in most cases. A notable exception was the
monly used. Fig. 1 shows the observed and predicted breach peak flow prediction method of Walder and O’Connor 共1997兲,
widths as computed by Wahl 共1998兲 in both arithmetically scaled which requires that the reservoir be classified as a large- or small-
and log–log plots. In the arithmetic plots, it would be difficult to reservoir case. In addition, in the case of the small-reservoir situ-
draw in upper and lower bound lines to define an uncertainty ation, an average vertical erosion rate of the breach must be esti-
band. In the log–log plots, data are scattered approximately mated. The Walder and O’Connor method was applied only to
evenly above and below the lines of perfect prediction, suggesting those dams that could be clearly identified as large-reservoir
that uncertainties would best be expressed as a number of log
共where peak outflow is insensitive to the vertical erosion rate兲 or
cycles on either side of the predicted value. This is the approach
small-reservoir with an associated estimate of the vertical erosion
taken in the analysis that follows.
rate obtained from observed breach heights and failure times. Two
The other notable feature of the plots in Fig. 1 is the presence
other facts should be noted:
of some significant outliers. Possible sources of these outliers
include the variable quality of the case study parameter observa- 1. No prediction equation could be applied to all 108 dam fail-
tions being used to test the predictions and the potential for mis- ure cases, due to the lack of required input data for the spe-
application of some of the prediction equations in the analysis cific equation or the lack of an observed value of the param-
described here due to lack of detailed firsthand knowledge of each eter of interest. Most of the breach width equations could be
case study situation. Such problems should not affect a careful tested against about 70 to 80 cases, the failure time equations
future application of these prediction equations to a specific case, against 30 to 40 cases, and the peak flow prediction equa-
and we do not wish for them to affect the present analysis of the tions against about 30 to 40 cases.
uncertainties of the methods themselves. Admittedly, much of the 2. The testing made use of the same data used to originally
scatter and the appearance of outliers are probably due to the develop many of the equations 共since the 108-dam database
inherent variability of the data caused by the variety of factors was compiled from these and other sources兲, but each equa-
that influence dam breach mechanics, and this variability should tion was also tested against additional cases, the number
be preserved as we analyze the uncertainties of the prediction varying depending on the method. This should provide a fair
equations. To exclude the truly anomalous data 共the statistical indication of the ability of each equation to predict breach
outliers兲 and retain the characteristic variability, an objective out- parameters for future dam failures. 共It is difficult to say ex-
lier exclusion algorithm was applied 共Rousseeuw 1998兲. The se- actly how many additional cases were analyzed for each
lected algorithm has the advantage that its performance is itself method, since the exact number of failures used to develop
insensitive to the presence of the outliers, which overcomes a each method is not indicated clearly in literature for all
common problem encountered when attempting to exclude outli- methods, and some are based on a combination of statistical
ers. analysis of case studies and physically based theory.兲
and Langridge-Monopolis and Froehlich relations. Despite this when the breach reaches its maximum size, before significant
fact, some very fast failures are documented in literature, and this drawdown of the reservoir has occurred. Despite the refinement
possibility is reflected in the prediction intervals determined from of considering large- versus small-reservoir behavior, the Walder
the uncertainty analysis. and O’Connor method was found to have uncertainty similar to
most of the other peak flow prediction methods 共about ⫾0.75 log
cycles兲. However, among the 22 case studies to which the method
Results—Peak Outflow Estimates
could be applied, only four proved to be large-reservoir/fast-
Peak outflow estimates are shown in Table 4, sorted in order of erosion cases. Of these, the method overpredicted the peak out-
increasing peak outflow for the top-of-joint-use scenario. The flow in three cases, and dramatically underpredicted in one case
lowest peak flow predictions come from those equations that are 共Goose Creek Dam, South Carolina, failed 1916 by overtopping兲.
based solely on dam height or depth of water in the reservoir. The Closer examination showed some contradictions in the data re-
highest peak flows are predicted by those equations that incorpo- ported in literature for this case. On balance, it appears that the
rate a significant dependence on reservoir storage. Some of the Walder and O’Connor method may provide reasonable estimates
predicted peak flows and the upper bounds of the prediction limits of the upper limit on peak outflow for large-reservoir/fast-erosion
would be the largest dam-break outflows ever recorded, exceed- cases.
ing the 65,000 m3/s peak outflow from the Teton Dam failure. For this application, results from the Froehlich method were
共Storage in Teton Dam at failure was 356⫻106 m3 ). The length of considered to be the best estimate of peak breach outflow, and the
the reservoir 共about 48 km兲 may help to attenuate some of the results from the Walder and O’Connor method provided an upper
large peak outflows predicted by the storage-sensitive equations, bound estimate.
since there will be an appreciable routing effect in the reservoir
itself that is probably not accounted for in the peak flow predic-
NWS-BREACH Simulations
tion equations.
The equation offered by Froehlich 共1995b兲 clearly had the best Several simulations runs were made using the NWS-BREACH
prediction performance in the uncertainty analysis, and is thus model 共Fread 1988兲. The model requires input data related to
highlighted in Table 4. This equation had the smallest mean pre- reservoir bathymetry, dam geometry, the tailwater channel, em-
diction error and narrowest prediction interval by a significant bankment materials, and initial conditions for the simulated pip-
margin. ing failure.
The results for the Walder and O’Connor method are also The results of the simulations are very sensitive to the eleva-
highlighted. As discussed earlier, this is the only method that tion at which the piping failure is assumed to develop. In all cases
considers the differences between the so-called large-reservoir/ analyzed, the maximum outflow occurred just prior to the crest of
fast-erosion and small-reservoir/slow-erosion cases. This dam the dam collapsing into the pipe; after the collapse of the crest, a
proves to be a large-reservoir/fast-erosion case when analyzed by large volume of material partially blocks the breach and the out-
this method 共regardless of the assumed vertical erosion rate of the flow becomes weir controlled until the material can be removed.
breach—within reasonable limits兲, so the peak outflow will occur Thus, the largest peak outflows and largest breach sizes are ob-
tained if the failure is initiated at the base of the dam, assumed to are large for all methods, and thus it may be worthwhile to incor-
be at an elevation of 423.67 m. This produces the maximum porate uncertainty analysis results into future risk assessment
amount of head on the developing pipe, and allows it to grow to studies when predicting breach parameters using these methods.
the largest possible size before the collapse occurs. Table 5 shows Predictions of breach width generally have an uncertainty of
summary results of the simulations. For each initial reservoir el- about ⫾1/3 order of magnitude, predictions of failure time have
evation, a simulation was run with the pipe initiating at an eleva- uncertainties approaching ⫾1 order of magnitude, and predictions
tion of 423.7 m, and a second simulation was run with the pipe of peak flow have uncertainties of about ⫾0.5 to ⫾1 order of
initiating about midway up the height of the dam. magnitude, except the Froehlich peak flow equation, which has an
There is a wide variation in the results depending on the as- uncertainty of about ⫾1/3 order of magnitude.
sumed initial conditions for the elevation of the seepage failure. The uncertainty analysis made use of a database of informa-
The peak outflows and breach widths tend toward the low end of tion on the failure of 108 dams compiled from numerous sources
the range of predictions made using the regression equations in literature 共Wahl 1998兲. Those wishing to make use of this da-
based on case study data. The predicted failure times are within tabase may obtain it in electronic form 共Lotus 1-2-3, Microsoft
the range of the previous predictions, and significantly longer Excel, and Microsoft Access兲 on the Internet at http://
than the very short 共0.5 to 0.75 h兲 failure times predicted by the www.usbr.gov/pmts/hydraulicsគlab/twahl/
Bureau of Reclamation 共1988兲 equation and the first Von Thun The case study presented here showed that significant engi-
and Gillette equation. neering judgment must be exercised in the interpretation of pre-
dictions of breach parameters. The results from use of the physi-
cally based NWS-BREACH model were reassuring because they
Conclusions fell within the range of values obtained from the regression-based
methods. However, at the same time, they also helped to show
This paper has presented a quantitative analysis of the uncertainty that even physically based methods can be highly sensitive to the
of various regression-based methods for predicting embankment assumptions of the analyst regarding breach morphology and the
dam breach parameters and peak breach outflows. The uncertain- location of initial breach development. The NWS-BREACH
ties of predictions of breach width, failure time, and peak outflow simulations demonstrated the possibility for limiting failure me-
chanics that were not revealed by the regression-based methods.