Professional Documents
Culture Documents
157
[4]
ABSTRACT
Melching, C.S., 1992. An improved first-order reliability approach for assessing uncertainties in hydrologic
modeling. J. Hydrol., 132: 157-177.
First-order reliability analysis involving Taylor series linearization at central values of the uncertain
variables is well known as a means of approximating the output reliability for hydrologic and environmental models. However, linearization at central values has several key flaws - - namely, problems with
non-linear systems and estimation of extreme probabilities
which limit its usefulness. An improved
first-order reliability approach is presented wherein the linearization point varies to match the output level
whose exceedance probability is sought. Thus, the improved approach circumvents some of the problems
of central value linearization while retaining much of its simplicity. The accuracy, relative to Monte Carlo
simulation, of the improved approach and its use are demonstrated for an example of estimating the
probability distribution of peak discharge predicted by two conceptual rainfall-runoff models with
uncertain parameters.
INTRODUCTION
0022-1694/92/$05.00
158
C.S. M EL('HIN(i
comprise the system and its uncertainties. This mean-value first-order reliability
analysis method (MFORM) has worked well for many applications but it has
been found to have several important shortcomings which limit its application.
This paper illustrates an improved (advanced) first-order reliability analysis
method (AFORM), which retains many of the advantages of M F O R M while
circumventing many of MFORM's disadvantages. AFORM is not unknown
in water resources engineering, however, most of its applications have been
presented at specialty conferences (see references in Yen, 1989). Only Sitar
et al.'s (1987) application of AFORM to simple groundwater flow and solute
transport examples and Melching et al.'s (1990) use of AFORM to derive the
probability distribution of peak discharge estimates for specific events
produced by the HEC-1 rainfall-runoff model have reached a wider audience
of water resources professionals. The purpose of this paper is to discuss in
detail the improvements offered by AFORM relative to MFORM and to
show the computational procedure for A F O R M and its accuracy relative to
Monte Carlo simulation and M F O R M for two example rainfall-runoffmodel
applications. Thus, providing water resources professionals with additional
evidence that AFORM is a potentially powerful tool for design and decision
making subject to uncertainties.
GENERAL
RELIABILITY
ANALYSIS
CONCEPTS
(1)
Pr(L <~ R)
RL
1 -
lfR.,
*d
td
(r,
l)drdl
(2)
1 --
fL(l)
fR(r)dr dl
(3)
Generally, for realistic problems the joint PDF of R and L and the
159
R c = Pr(Z >~ O)
(4)
= g(x,.) + ~ ( x i - : q w ) ( t ? g ) ~
E[Z]
~- g(i)
P
VAR(Z)
(6)
P
(7)
i=1 j=l
VAR(Z)
a_2 ~
~ C2a~
s=~
(8)
160
c . s MELCHING
in this paper a foreward difference is used wherein each basic variable value
is increased 1% with the others held constant and the change in Z is calculated.
The system reliability is measured in terms of a reliability index,/3, defined
as
/3 = E[Zl/az
(9)
*(/3)
(10)
UNCERTAINTIES
IN H Y D R O L O G I C
161
MODELING
E[Z]
C*(<-x*)
(ll)
i-[
P
VAR(Z)
a2 ___ Z ~, CffCff E [ ( x i i= l j
X*)(Xj
--
x*)]
where (7* = 8g/Oxi evaluated at x*, and ~ ' = 8g/Oxj evaluated at x*.
(12)
162
c.s. MELCHING
VAR(Z)
= a~ ~-
E (C*a~) 2
(13)
i=1
Hence, for the case of statistically independent variables, the reliability index
for A F O R M is
p
~:t
(14)
(C*
i=|
of course g(x) = 0 at x*. A similar relation may be derived for the case of
correlated basic variables.
A F O R M formulations are fairly simple and straightforward. The key to
this method is the determination of the failure point for the Taylor series
expansion. Several iteration methods have been proposed for determining the
failure point - - e.g. a flow chart of the Rackwitz (1976) method is given in
Fig. 1. The efficiency of these methods is largely a function of whether the
failure surface corresponds to extreme values of the basic variables. It can be
shown that for A F O R M the reliability index, B, (or its absolute value) is
the shortest distance in standardized space between the system mean state
(i.e. where all basic variables are at their mean values) and the failure surface
(Shinozuka, 1983). Hence, optimization algorithms currently employed for
constrained non-linear programming problems can be used for failure point
determination. For example, Cheng (1982) proposed using the well-known
generalized reduced-gradient algorithm to minimize the absolute value of
Z with the constraints chosen such that the proper ~ value is determined.
Also, Shinozuka (1983) describes a Lagrange multiplier approach for direct
minimization of ~.
For linear failure surfaces where all the basic variables are normally distributed, the reliability is exactly given by eqn. (10). For convex failure
surfaces where all the basic variables are normally distributed, Hasofer (1974)
showed that the system reliability is bounded as
*(/~) ~< RL ~< Z~(Y)
(15)
| 63
NO
Evaluate:
-')Iei = (~G/~Xi) at x*it V i
Evaluate:
IC i = (~G/~Xi) at x*il V i
Calculate:
~iNi and aiN 1 V i
I(mqs. lS and 19)
Calculate:
XiNt and OiN t V i
(Eqs. 18 and 19)
Compute
Ci iNt
Compute:
aiNl =
aiNt = " p
{j__El (Cj OjNt)2} I/2
Compute
x*
= I i(t+l)
:
C i iN I
p
{ Z (Cj OjNI)2}i/2
j=1
Icmpute:
~1 (Eq. 14)
Compute :
~t+l using x~(t+1) and
Ig(x~+l)
(Eq. 14)
Reeompute :
x*(t+l ) as above V i using Bt+ 1
NO
YES
Compute:
= (Bt+ I)
Fig. 1. Flow chart of Rackwitz's (1976) iterative algorithm for locating the failure point, x* (independent
basic variables).
assumption has been found acceptable for a number of non-linear performance functions typical of those found in practical engineering problems
(Ang and Tang, 1984, p. 383). Further, this assumption is explored in this
paper using a comparison with Monte Carlo simulation for estimation of the
164
C.S, MELCHING
O ( x*aiN-xiN~/
(16)
(17)
where Fx,(x*) is the CDF ofx~ at x*,fx,(X*) is the PDF ofx~ at x*, a n d f N ( )
is the PDF of the standard normal distribution.
In order to do this, Rackwitz approximated the non-normal distribution
function by a first-order Taylor series expansion. Thus, the mean, ~N, and the
standard deviation, o"m, of the equivalent normal distributions become
~,y =
am =
x* - @-'[Fx,(x*)]aiN
i N{O-' [F~,(x*)]}
L,(x*)
(18)
(19)
Rackwitz and Fiessler (1978) pointed out that this normal transformation
of non-normal variables is exact within the accuracy of the first-order theory
under consideration. Transformation of uniform- and gamma-distributed
loss-rate functions for a reliability analysis of a simple rainfall-runoff-flood
frequency model confirm the accuracy of this normal distribution transformation (Melching et al., 1987).
EXAMPLE
APPLICATION
UNCERTAINTIES
IN H Y D R O L O G I C
MODELING
"
165
T - 2h(0)
(20)
where )t is the correction factor which accounts for bias in model predictions,
h( ) is the model functions which estimate peak discharge, and 0 is the vector
of model parameters.
The union of 2 and 0 is the vector of basic variables, x, for this case.
Example watershed
The Vermilion river watershed at Pontiac, Illinois, USA is a 1500km 2
agricultural watershed located about 120km southwest of Chicago. The
watershed outline is formed by morainal ridges which range in elevation from
230 to 240 m, and with the exception of these ridges the watershed is a nearly
flat plain ranging in elevation from 180 to 200 m. The soil types are primarily
silt loams and silty clay loams which support considerable agricultural
development (about 85% of land in row crops). The gentle slopes and low
permeabilities require the use of an extensive agricultural drainage system
including channelization of the Vermilion river and many of its tributaries. In
the period from 1950 to 1985 the city of Pontiac experienced five significant
flood events. The results reported in this paper are a part of a larger study of
flooding at Pontiac and the assessment of the reliability of flood event forecasting (Melching, 1987).
C.S. MELCHING
166
TABLE 1
Statistics of the basic variables for HEC-1
Parameter
Mean
Standard
deviation
Distribution
Notes
2H
1.726-0.154 In Qp
0.0472
Normal
IL
3.88-0.60 In QB
0.718
Log-normal
0.645
8.94
4.32
Log-normal
Normal
Normal
CL
TC
0.591
38.0
27.7
by the Clark unit hydrograph option using the built-in dimensionless timearea curve as recommended by Ford et al. (1980). The basic variables for
this modeling case are IL (mm), CL (mm h-~), the watershed time of concentration, TC (h), the watershed storage coefficient, R(h), and the model
correction factor 2 . . Table 1 gives the mean value, standard deviation, and
distributional assumptions for the basic variables used in this study. These
assumptions are based on the results of calibrating HEC-1 to 32 flood and
near-flood events occurring on the Vermilion river during the period 1965-1983
as discussed in Melching (1987).
RORB is a quasi-distributed, non-linear conceptual rainfall-runoff model.
In RORB applications the watershed is divided into a number of subcatchments each of which is assigned (ideally) a rainfall depth based on the storm
(or typical storm) isohyetal pattern and a rainfall temporal distribution based
on a representative raingage. This distributed rainfall input is reduced by an
abstraction scheme of initial loss (ILR) - - continuing, constant loss rate
(CLR). Each of the loss parameters have identical values for all subcatchments above a c o m m o n stream gage (if there is only one stream gage all
subcatchments have the same I L R and C L R values), hence, making RORB
quasi-distributed. The rainfall excess for each subcatchment is then routed
through the channel network to the outlet via a cascade of conceptual nonlinear reservoirs.
In this study, the Vermilion watershed was divided into 21 subcatchments.
167
TABLE 2
Statistics of the basic variables for RORB
Parameter
Mean
Standard
deviation
2R
1.737-0.152 In (~pr
0.0629
ILR
3.88-0.60 In QB
0.718
CLR
m
0.564
0.9
C~
0.556
-
65.4
11.6
Distribution
Notes
Normal
NOTES
M o n t e Carlo simulation
168
f . s . MELCHING
(21)
where b = 2 for RORB and is related to the difference between IL and Pi for
HEC-1, the value of IL between these bounds is irrelevant in the calculation
of the rainfall excess. When the iteration scheme moves to such a point,
c~Z/c3IL = 0 which is generally a drastic change from its previous non-zero
value and this leads to problems with iteration convergence. This is more of
a problem with HEC-1, which uses a single hyetograph, than for RORB,
which uses multiple hyetographs (two in this study). Generally, these
169
MFORM
In the case of the initial loss parameter for both models the Taylor series
expansion was performed at the median (i.e. the mean value of the logarithms),
while for all other parameters the expansion was taken at the mean values.
Only the results of M F O R M with the performance function in the form
Z = R - L are reported here. The form Z = (R/L) - 1 performs poorly in
general (Yen et al., 1986), while for this case study the form Z = ln(R/L)
performs poorly, especially when extreme discharge levels are considered
(Melching, 1987).
PRESENTATION OF RESULTS
170
C.S. MELCHING
1.0
MFORM--
0.8
\
Monte
Carlo
0.6-
\ \
u
0.4--
0.2
o.o
I
50
I
i00
Peak Discharge,
150
200
Q (m3/s)
Fig. 2. Comparison of Monte Carlo (MFORM) and AFORM estimates of exceedance probabilities of
peak discharge predictions by HEC-I for the storm of 8 April 1965.
- - the storm event of 13 July 1957 is a convective type event while the others
are frontal (cyclonic) events.
Cornell (1972) noted that the basis for judging the quality of first-order
reliability analysis should not be the exactness of the probabilities estimated
but rather the decision or design parameters derived from the probabilities.
1.0
AFORM
~\
>~ 0 . 8
MFORM -
~\
.o@
0.6
0.4
--
Monte
Carlo
,\
x
\\,
\\\
x~, \ ,
~ o.2
0.0
I
i00
,
200
Peak Discharge,
300
4O0
0 (m3/~],
Fig. 3. Comparison of Monte Carlo (MFORM) and AFORM estimates of exceedance probabilities of
peak discharge predictions by HEC-1 for the storm of 26 April 1959.
UNCERTAINTIES
IN HYDROLOGIC
--
1.0
171
MODELING
\
~-
0.8
'AFORM I
MFORM
0.6
0.4
0.2
0.0 0
i00
x,
~\
"~k \
-~
---
'
Monte
Carlo
'
200
300
400
Fig. 4.
peak
AFORM -0.8 ~
MFOP~M
---
..4
Monte
Carlo
0.6
o.4
0.2
\ ~
0.
i00
200
300
400
peak
172
C.S. MELCHING
1.0
0.8
_ " ~
,~\\~\
'
I
AFORM
MFORM
Monte
Carlo
0.6
0.4
7
M
0.2
I
50
0.0
I
i00
Peak Discharge,
150
200
Q (m3/s)
Fig. 6. Comparison of Monte Carlo (MFORM) and AFORM estimates of exceedance probabilities of
peak discharge predictions by RORB for the storm of 8 April 1965.
1.0
~-_._
0.8
'\
--
"~\
~
'
0.6
__
'_ _
_
Monte
Carlo
\\
,\
0.4
0.2
0.0
i00
200
Peak Discharge,
300
400
Q (m3/s)
Fig. 7. Comparison of Monte Carlo (MFORM) and AFORM estimates of exceedance probabilities of
peak discharge predictions by RORB for the storm of 26 April 1959.
173
1.0
AFORM
0.8
~ ~\
MFORM--
"\
,.Q
0.6
\\
\\,
u
r~
Monte
Carlo
'\
0.4
0.2
0.0
100
300
200
Peak D i s c h a r g e ,
400
Q (m3/s)
Fig. 8. Comparison of Monte Carlo (MFORM) and AFORM estimates of exceedance probabilities of
peak discharge predictions by RORB for the storm of 4 May 1965.
show that MFORM offers very good agreement with Monte Carlo simulation
estimates of the peak discharge CDF for HEC-1 modeling of the 26 April 1959
and 4 May 1965 events and RORB modeling of the 26 April 1959 event. The
agreement is best in the central portion of the CDF and it degrades significantly
at the extremes (as would be expected). The other figures show very poor
1.0
'
'
AFORM
MFORM -0.8
" ~
"~
~
..~
Monte
Carlo
~ ~..
0.6
'\
"-~>
0.4
tB
o oo
i00
200
Peak D i s c h a r g e ,
--"-t
300
-- ~-"'-~--~
400
Q (m3/s)
Fig. 9. Comparison of Monte Carlo (MFORM) and AFORM estimates of exceedance probabilities of
peak discharge predictions by RORB for the storm of 13 July 1957.
174
C.S. MELCHING
TABLE 3
Probabilities of flood level exceedance estimated for HEC-1 using Monte Carlo simulation,
MFORM, and AFORM
Date
Monte Carlo
AFORM
MFORM
21/6/56
16/4/57
24/4/57
13/7/57
22/7/57
9/6/58
12/6/58
14/7/58
26/4/59
23/9/61
10/5/62
5/4/65
8/4/65
23/4/65
4/5/65
12/6/73
0.192
0.108
0.023
0.191
0.164
0.023
0.933
0.019
0.0046
0.160
0.260
0.0026
0.239
0.125
0.034
0.251
0.214
0.036
0.954
0.035
0.000005
0.00026
0.000088
0.000024
0.0094
0.222
0.335
0.0018
0.212
0.077
0.0083
0.222
0.206
0.105
0.962
0.025
0.000003
0.00067
0.00002
0.0023
0.0047
0.216
0.298
UNCERTAINTIESIN HYDROLOGICMODELING
175
TABLE 4
Probabilities of flood level exceedance estimated for R O R B using Monte Carlo simulation,
M F O R M , and A F O R M
Date
Monte Carlo
AFORM
MFORM
21/6/56
16/4/57
24/4/57
13/7/57
22/7/57
9/6/58
12/6/58
14/7/58
26/4/59
23/9/61
10/5/62
5/4/65
8/4/65
23/4/65
4/5/65
12/6/73
0.363
0. | 62
0.057
0.438
0.278
0.228
0.944
0.026
-
0.029
0.425
0.170
0.078
0.504
0.340
0.290
0.960
0.044
0.00033
0.0027
0.0011
0.00072
0.040
0.314
0.552
0.0059
0.378
0.138
0.013
0.467
0.296
0.292
0.951
0.062
0.00091
0.0018
0.00079
0.014
0.024
0.318
0.511
0.026
0.254
0.485
176
C.S. MELCHING
offers a compromise between the accuracy of Monte Carlo simulation (at high
computational cost) and the simplicity of mean value first-order reliability
analysis (with considerable problems for non-linear systems and extreme
failure events). A simple example of the uncertainty in peak discharge
estimates from two rainfall-runoff models demonstrated the accuracy of
AFORM relative to Monte Carlo simulation over a wide range of exceedance
probabilities. This example included several events where non-linearities in
model performance caused serious problems for MFORM and yet AFORM
still performed well for these cases. The example also considered large coefficients of variation for the most sensitive basic variables and non-normal
distributions for the basic variables which can in general lead to problems with
MFORM estimates.
Performance function discontinuities can lead to overestimation of
exceedance probabilities by AFORM. However, this overestimation may not
be significant at the design or decision parameter level. For cases where the
discontinuity effects are small, the agreement between AFORM and Monte
Carlo simulation is quite close.
It is hoped that this paper will stimulate water resources professionals to try
A F O R M for assessing uncertainties for those problems where the computational demands of Monte Carlo simulation are still too great.
REFERENCES
Ang, A.H.-S. and Tang, W.H., 1984. Probability Concepts in Engineering Planning and
Design, Vol. II, Decision, Risk, and Reliability. Wiley & Sons, New York.
Beale, E.M.L., 1960. Confidence regions in nonlinear estimation. J.R. Stat. Soc., Ser. B, 42:
1-25.
Beck, M.B., 1987. Water quality modeling: A review of the analysis of uncertainty. Water
Resour. Res., 23: 1393-1441.
Brown, L.C. and Barnwell, Jr., T.O., 1987. The enhanced stream water quality models
QUAL2E and QUAL2E-UNCAS: documentation and user manual. US Environmental
Protection Agency, Athens, GA, Rep. No. EPA/600/3-87/007.
Cheng, S.-T., 1982. Overtopping risk evaluation for an existing dam. Ph.D. Thesis. Dep. Civ.
Eng., University of Illinois at Urbana-Champaign.
Cornell, C.A., 1972. First-order analysis of model and parameter uncertainty. Proc. Int. Syrup.
Uncertainties in Hydrologic and Water Resources Systems, Vol. 3, 11-14 December 1972,
Tucson, AZ, University of Arizona, Tucson, AZ, pp. 1245-1274.
Ford, D.T., Morris, E.C. and Feldman, A.D., 1980. Corps of Engineers experience with
automatic calibration of precipitation --- runoff model. In: Y. Haimes and J. Kindler
(Editors), Water and Related Land Resources Systems. Pergamon, New York, pp. 467-476.
Garen, D.C. and Burges, S.J., 1981. Approximate error bounds for simulated hydrographs.
ASCE J. Hydraul. Div., 107(HYll): 1519-1534.
Hasofer, A.M., 1974. Reliability index and failure probability. J. Struct. Mech., 3: 25-27.
177
Hasofer, A.M. and Lind, N.C., 1974. Exact and invariant second-moment code format. ASCE
J. Eng. Mech. Div., 100(EM1): 111-121.
Kuczera, G., 1988. On the validity of first-order prediction limits for conceptual hydrologic
models. J. Hydrol., 103: 229-247.
Laurenson, E.M. and Mein, R.G., 1985. RORB-Version 3 Runoff Routing Program User
Manual. Dep. Civ. Eng., Monash University, Clayton, Vic.
LaVenue, M., Andrews, R.W. and Rama Rao, B.S., 1989. Groundwater travel time uncertainty
analysis using sensitivity derivatives. Water Resour. Res., 25: 1551-1566.
Melching, C.S., 1987. A reliability analysis on flood event forecasting with uncertainties. Ph.D.
Thesis. Dep. Civ. Eng., University of Illinois at Urbana-Champaign.
Melching, C.S., Wenzel, Jr., H.G. and Yen, B.C., 1987. Application of system reliability
analysis to flood forecasting. In: V.P. Singh (Editor), Application of Frequency and Risk in
Water Resources. Reidel, Dordrecht, pp. 335-350.
Melching, C.S., Yen, B.C. and Wenzel, Jr., H.G., 1990. A reliability estimation in modeling
watershed runoff with uncertainties. Water Resour. Res., 26: 2275-2286.
Rackwitz, R., 1976. Practical probabilistic approach to design. Comite European du Beton,
Paris. Bull. No. 112.
Rackwitz, R. and Fiessler, B., 1978. Non-normal vectors in structural reliability. Technical
University of Munich, SFB 96 Rep. No. 29, pp. 1-22.
Scavia, D., Powers, W.T., Canale, R.P. and Moody, J.L., 1981. Comparison of first-order error
analysis and Monte Carlo simulation in time-dependent lake eutrophication models. Water
Resour. Res., 17: 1051-1059.
Shinozuka, M., 1983. Basic analysis of structural safety, ASCE J. Struct. Eng., 109: 721-740.
Sitar, N.. Cawfield, J.D. and Der Kiureghian, A., 1987. First-order reliability approach to
stochastic analysis of subsurface flow and contaminant transport. Water Resour. Res., 23:
794-804.
US Army Corps of Engineers, 1985. HEC-1 Flood Hydrograph Package: User's Manual.
Hydrologic Engineering Center, Davis, CA.
Yen, B.C., 1989. Engineering approaches to risk and reliability analysis. In: Y. Haimes and E.Z.
Stakhiv (Editors), Risk Analysis and Management of Natural and Man-Made Hazards.
Am. Soc. Civ. Eng., New York, pp. 22-49.
Yen, B.C., Cheng, S.-T. and Melching, C.S., 1986. First order reliability analysis. In: B.C. Yen
(Editor), Stochastic and Risk Analysis in Hydraulic Engineering. Water Resources Publications, Littleton, CO, pp. 1-36.