Professional Documents
Culture Documents
Abstract: The coincidence of floods on a mainstem and its tributaries may cause significant flood damage downstream of the confluence.
In this study, the coincidence risks of annual maximum (AM) floods on Jinsha River and Min River, China, were analyzed using multivariate
copulas based on both systematic records and available historical information. Archimedean and elliptical copulas were selected to establish
Downloaded from ascelibrary.org by URI LIBRARIES on 12/01/18. Copyright ASCE. For personal use only; all rights reserved.
the joint distributions of flood magnitudes and occurrence dates of the two rivers. The coincidence probabilities of AM floods were inves-
tigated and compared with monthly maximum floods. The results show that the coincidence probabilities of AM flood magnitudes and
occurrence dates are much less than those of monthly maximum floods. Higher coincidence probabilities of AM flood occurrence dates
occur in the period from mid-June to early September with three coincidence peaks, while there is only one coincidence peak in each month
for monthly maximum floods. Ignoring historical flood information would underestimate the coincidence probabilities of flood magnitudes.
The results can provide decision support for cascade reservoir operation on Jinsha River and flood mitigation downstream. DOI: 10.1061/
(ASCE)HE.1943-5584.0001744. © 2018 American Society of Civil Engineers.
Author keywords: Flood coincidence risk; Copula function; Joint distribution; Flood magnitude; Flood occurrence date; Historical flood.
peak by the occurrence date of the two rivers, based on the counted the coincidence number of floods on the Jinsha and Min
assumption that the floods on the two rivers were independent. How- Rivers based on the measured data but did not quantitatively esti-
ever, this implicit assumption is violated because streamflows are mate the coincidence probability of design floods. Chen et al. (2011)
structured by the hydrologic network and have strong spatial depend- used a copula method to analyze the coincidence risks of AM floods
ence. As an alternative, Chen et al. (2012) used a multivariate copula of the Jinsha and Min Rivers within a time interval of 4 days based
function to estimate the coincidence of flood risk by considering both on the systematic AM floods in the period 1951–2007. In this study,
flood magnitude and occurrence time. Although this approach can be the coincidence risk of AM flood magnitudes and occurrence dates
used to simultaneously estimate the coincidence probabilities of of the Jinsha and Min Rivers occurring on a daily basis were esti-
flood magnitude and occurrence time, ignoring historical flood data mated by their copula-based joint distribution based on systematic
may lead to underestimate or overestimate the coincidence flood risk flood records and available historical data. The daily results were
(Reis and Stedinger 2005; Merz and Blöschl 2008a, b; Yan and compared with those of monthly floods.
Moradkhani 2015). Therefore, coincidence flood risk analysis of The streamflow data used in this study were acquired from two
a river mainstem and its tributary needs to be improved.
hydrological stations, Pingshan and Gaochang, which are located
The objective of this article is to present a copula-based method
on the Jinsha and Min Rivers, respectively. The locations of the
for estimating the coincidence risk of floods on a mainstem and its
rivers and hydrological stations are shown in Fig. 1. The drainage
tributary through a multivariate copula. In the method, factors con-
areas of the two stations are 458,592 and 135,378 km2 , respectively
sidered include flood magnitude, occurrence time, and spatial corre-
(Xu and Tong 2012). The AM flood series and the corresponding
lation. For the coincidence flood risk analysis, effective flood data
occurrence dates are then extracted from the daily average stream-
were collected and used, including systematic records and available
flow records at the stations during the period 1950–2010, a 61-year
historical information. A case study on Jinsha River and Min River,
record, as presented in Fig. 2. Detailed data are given in Appendix I.
China, is also presented to demonstrate the method developed in
this study. Discharges, runoff volumes, and AM floods are analyzed and
summarized in Table 1. Of 61 pairs of annual flood peaks of the
Jinsha and Min Rivers, 7 occurred concurrently within a time frame
Study Area and Data of 3 days. However, only one pair occurred on the same day,
i.e., September 1, 1966.
Jinsha River Basin is the upper watershed of the Yangtze River, In addition to the 61-year record, historical flood events were
which flows through eastern Tibet and southwestern China. Jinsha also incorporated into this analysis. The peak discharges and
River flood events are commonly caused by snowmelt or heavy occurrence dates of historical floods (CWRC 1996; Wang 1999;
Fig. 2. Time series of annual maximum (AM) flood peak and corresponding occurrence date for Jinsha and Min Rivers during period 1950–2010:
(a) AM flood peak; and (b) occurrence date of AM flood.
Table 1. Characteristic values of annual daily discharge, runoff, and annual maximum flood on Jinsha and Min Rivers, 1950–2010
Average annual Average annual Maximum AM flood
daily discharge runoff volume Average AM flood
River Station (m3 =s) (108 m3 ) Magnitude (m3 =s) Occurrence date (m3 =s)
Jinsha Pingshan 4,567 1,436 28,600 September 1, 1966 17,871
Min Gaochang 2,711 846.6 31,400 June 29, 1961 17,068
Table 2. Investigated historical floods on Jinsha and Min Rivers Sklar 1983), a multivariate distribution can be written in a copula
Peak discharge and its univariate marginal distributions as
River (m3 =s) Occurrence date
Fðx1 ; x2 ; : : : ; xn Þ ¼ Cθ ðFX1 ðx1 Þ; FX2 ðx2 Þ; : : : ; FXn ðxn ÞÞ
Jinsha 36,900 Mid-September 1924
35,300 July 17, 1860 ¼ Cθ ðu1 ; u2 ; : : : ; un Þ ð1Þ
35,000 July 1892
34,300 August 1905
where Fð·Þ is an n-dimensional distribution function; ui ¼ FXi ðxi Þ
31,100 September 21, 1813
36,900 Late September 1637
is a marginal distribution of random variable X i ; and θ is a copula
parameter to be estimated. If these marginal distributions are con-
Min 51,000 July 18–22, 1917 tinuous, the copula function Cθ ð·Þ can be written (Sraj et al. 2015)
40,800 July 1931
Cθ ðu1 ; u2 ; : : : ; un Þ ¼ FðF−1 −1 −1
X1 ðu1 Þ; FX 2 ðu2 Þ; : : : ; FX n ðun ÞÞ ð2Þ
where FT ðti ; tj Þ = joint distribution function of occurrence dates of Fig. 3 presents the flowchart of the proposed study. The copula
the mainstem and its tributary; and T M and T T = occurrence dates method can estimate the marginal and joint distributions in two sep-
of AM floods on the mainstem and its tributary, respectively. arate processes, giving flexibility in selecting the marginal and joint
probabilistic models. In this study, the P-III and mixed von Mises
Coincidence Risk of Flood Magnitudes distributions were used to quantify the marginal distributions of
When the combined flood discharge exceeds the safety discharge AM flood magnitudes and the occurrence dates on the Jinsha
downstream, coincidence floods can cause significant flood dam- and Min Rivers, respectively, compared with other distributions in-
age downstream. However, if the floods of the mainstem and its cluding Gamma, GEV, lognormal, and LP-III distributions. Then
tributary do not occur on the same day or the interval of occurrence four commonly used Archimedean copulas and two elliptical cop-
dates is large, there is no coincidence flood risk. Therefore, the ulas were used to model the bivariate joint cumulative distribution
coincidence risk of flood magnitude must be analyzed taking into functions (CDFs) of AM flood magnitudes and occurrence dates of
consideration the coincidence of flood occurrence dates. Assuming the two rivers, and the most appropriate copulas were identified
that the flood occurrence date is independent of flood magnitude based on the Rosenblatt transformation with the Cramér–von Mises
(Chen et al. 2012), the coincidence probability of flood magnitudes statistic. Finally, the coincidence risk of the AM flood magnitudes
(P) in this paper is defined as and occurrence dates were estimated from the established joint dis-
tribution, aimed at exploring potential reservoir flood control and
P ¼ P T × P P0 ð12Þ management practices once coincidence floods occurred.
stations are presented in Table 3. Table 3 shows that the values of tematic records, respectively. It can be seen that the exceedance
chi-squared and K-S test statistics did not exceed their critical val- probabilities of floods greater than 20,000 m3 /s increased with the
ues at the 0.05 significance level, 5.99 and 0.174, which implies consideration of historical floods.
that all the distributions performed well in quantifying the marginal A similar analysis of the marginal distribution of monthly maxi-
distributions of AM flood peaks and occurrence dates of the Jinsha mum floods shows that the P-III distribution and Gaussian kernel
and Min Rivers. Furthermore, the values of RMSE were then density performed best in modeling the distributions of monthly
adopted to identify the most appropriate distributions for quantify- flood magnitudes and occurrence dates, respectively. The RMSE
ing the marginal distributions. As presented in Table 3, the P-III values of the probability distributions for monthly flood magni-
distribution performed best in modeling the distributions of AM tudes and occurrence dates are shown in Table 4.
flood peaks at the stations, and the mixed von Mises distribution
performed best in modeling the distributions of AM flood occur-
rence dates, as shown by the minimum RMSE values. Joint Probability Distribution Using Copulas
The fitted marginal distribution curves for AM flood peaks The dependencies of flood magnitudes and occurrence dates be-
and occurrence dates on the Jinsha and Min Rivers are presented tween the Jinsha and Min Rivers were analyzed to estimate the joint
in Figs. 4 and 5, respectively. It can be seen that the P-III and distribution. The estimates of the bivariate Kendall’s tau coefficients
mixed von Mises distributions performed well in simulating the of flood magnitudes and occurrence dates for AM sample series
observed data. were 0.103 and 0.153, respectively. This result indicates that the
Table 3. Statistical test results for marginal distributions of annual maximum flood peaks and occurrence dates
Flood peak Occurrence date
Pingshan Station Gaochang Station Pingshan Station Gaochang Station
2 2 2 2
Distribution χ Dn RMSE χ Dn RMSE χ Dn RMSE χ Dn RMSE
GEV 2.23 0.071 0.092 1.71 0.058 0.102 3.45 0.127 0.241 3.25 0.085 0.162
Gamma 4.72 0.116 0.158 2.57 0.137 0.134 2.23 0.145 0.152 3.68 0.138 0.184
Lognormal 2.64 0.077 0.075 4.14 0.074 0.093 4.12 0.082 0.168 3.54 0.127 0.173
P-III 3.26 0.062 0.041 3.49 0.066 0.065 3.33 0.092 0.186 2.31 0.112 0.195
LP-III 3.13 0.077 0.073 3.15 0.074 0.086 3.25 0.116 0.132 2.89 0.145 0.146
Mixed von Mises — — — — — — 3.58 0.047 0.019 3.15 0.058 0.018
Note: Dn = K-S statistics.
Fig. 4. Pearson Type III (P-III) distributions fitted to annual maximum flood data: (a) Jinsha River; and (b) Min River.
Fig. 5. Mixed von Mises distributions fitted to observed occurrence date of annual maximum flood: (a) Jinsha River; and (b) Min River.
0.030 0.030
Sample data Sample data
0.025 Mixed von Mises distribution 0.025 Mixed von Mises distribution
Probability
Probability
0.020 0.020
0.015 0.015
0.010 0.010
0.005 0.005
0.000 0.000
Jun. Jul. Aug. Sep. Oct. Jun. Jul. Aug. Sep. Oct.
(a) Month (b) Month
Fig. 6. Histograms and probability densities of occurrence date of annual maximum flood: (a) Jinsha River; and (b) Min River.
Table 4. RMSE values of marginal distributions for monthly maximum copulas were estimated by the inversion of Kendall’s tau, and the
flood magnitudes and occurrence dates parameters of the two elliptical copulas were estimated using a
Flood peak Occurrence date maximum likelihood estimation method. Goodness-of-fit tests
were performed to select the most appropriate copula among the
Month Pingshan Gaochang Pingshan Gaochang
six candidate copulas. Table 5 shows the estimated parameters, the
June 0.045 0.056 0.039 0.041 corresponding Cramér–von Mises statistic Sn (Genest et al. 2009),
July 0.052 0.043 0.038 0.033 and the associated p-value of the six copulas. It can be seen that for
August 0.041 0.052 0.027 0.029
AM flood magnitudes and occurrence dates, the GH copula was
September 0.053 0.048 0.035 0.036
October 0.040 0.047 0.031 0.037 the best among the six copulas with the lowest Sn value and the
highest p-value. The RMSE values for the joint distributions ob-
tained through different copulas are also shown in Table 5. Based
on the values of RMSE, it is found that the GH copula was most
correlations of flood magnitudes and occurrence dates between the appropriate for modeling the joint distribution of the AM flood
two stations are positive. magnitudes and occurrence dates of the two rivers as indicated by
The six copulas from the Archimedean and elliptical families the minimum RMSE values, 0.058 and 0.053.
shown in Appendix II were used to construct the bivariate joint The joint CDFs fitted by the GH copula are presented in
CDFs for the AM flood magnitudes and occurrence dates on the Figs. 7(a and b). Fig. 7 shows that the joint CDF FQ ðqJ ; qM Þ had
Jinsha and Min Rivers. The parameters of the four Archimedean a convex surface in the shape of an abrupt rise, while the surface of
Table 5. Estimated values of copula parameter (θ̂), goodness-of-fit test: Cramér–von Mises criterion (Sn ), and associated p-value at 0.05 significance level
and RMSE
Flood magnitude Flood occurrence date
Copula θ̂ Sn p-value RMSE θ̂ Sn p-value RMSE
Ali-Mikhail-Haq 0.412 0.062 0.315 0.104 0.577 0.076 0.493 0.089
Gumbel-Hougaard 1.115 0.031 0.642 0.058 1.181 0.067 0.648 0.053
Frank 0.935 0.046 0.149 0.083 1.404 0.071 0.576 0.064
Clayton 0.230 0.053 0.312 0.076 0.361 0.086 0.350 0.068
Gaussian 0.244 0.042 0.536 0.108 0.201 0.084 0.374 0.083
Student t 0.212 0.039 0.481 0.093 0.223 0.078 0.559 0.063
Note: The lower Sn, higher p-value, and lower RMSE indicate a better copula model.
Fig. 7. Joint distributions of annual maximum flood magnitudes (Q) and occurrence dates (T) of Jinsha River (QJ and T J ) and Min River (QJ and
T J ): (a) flood magnitudes; and (b) flood occurrence dates.
Fig. 8. Comparisons of observed and modeled joint probabilities of observed combinations between Jinsha [annual maximum (AM) flood magnitude
QJ and occurrence date T J ] and Min River (AM flood magnitude QM and occurrence date T M ): (a) QJ and QM ; and (b) T J and T M .
the joint CDF FT ðtJ ; tM Þ was concave with a much shallower slope. The estimates of the bivariate Kendall’s tau coefficients of
This indicates that the nonexceedance joint probabilities of flood flood magnitudes and occurrence dates for monthly maximum
magnitudes of the two rivers increased greatly with the increase sample series are presented in Table 6. It shows that the Kendall’s
in peak discharges, while the nonexceedance joint probabilities of tau coefficients of occurrence dates for monthly maximum floods
the flood occurrence dates slowly increased with the flood occur- were negative in the months June–August, so the AMH and Frank
rence dates. copulas were used for the three months. For the joint distributions
The observed and computed nonexceedance joint probabilities of monthly maximum flood occurrence dates in the other two
are presented in Figs. 8(a and b), in which the x-axis is sorted in months and monthly maximum flood peaks, the four commonly
ascending order of the computed nonexceedance joint probabilities. used Archimedean copulas were used.
Fig. 8 shows that the GH copulas fitted the observed joint distri- Besides the Archimedean copulas, the Gaussian and Student t
bution of flood magnitudes and occurrence dates. The correlation copulas were also used to model the joint distributions of monthly
coefficients of AM flood magnitudes and occurrence dates between maximum flood. The RMSE values of these copulas for monthly
observed and computed joint probabilities are 0.994 and 0.988,
respectively.
Goodness-of-fit tests were also performed to select the best-fit
copula for the bivariate CDFs of the AM flood magnitudes of the Table 6. Values of Kendall’s Tau of monthly maximum flood magnitudes
two rivers with marginal distributions ignoring historical floods. and occurrence dates for Jinsha and Min Rivers
The result shows that the Clayton copula gave the best fit. A lower Monthly maximum
exceedance flood probability resulted in a higher nonexceedance flood June July August September October
joint probability of floods. The nonexceedance joint probability
computed by the Clayton copula has a lower value at a lower flood Flood peak 0.146 0.112 0.105 0.107 0.273
Occurrence date −0.0197 −0.00164 −0.0601 0.0415 0.233
exceedance probability.
maximum flood magnitudes and occurrence dates are shown in southeast monsoon in the Pacific Ocean. Therefore, July and August
Table 7. Based on the RMSE values, it is found that for the joint are the overlapping period of heavy rainfall events in the two river
Downloaded from ascelibrary.org by URI LIBRARIES on 12/01/18. Copyright ASCE. For personal use only; all rights reserved.
distribution of monthly maximum flood magnitudes, the Student t basins. As shown in Fig. 9, two AM flood occurrence dates with
copula was most appropriate for June, and the GH copula was most high estimated peak coincidence probabilities are in July.
appropriate for the months July–October. For the joint distribution The coincidence probability of the AM flood occurrence dates
of monthly maximum flood occurrence dates, the Frank copula was of the two rivers in the period from June 1 to October 31 can be
the best fitted copula for the June–August period, and the GH and obtained by adding together the coincidence probabilities P of AM
Clayton copulas were the best fitted copulas for September and flood occurrence dates on each day, expressed as PT ¼ pit , and
October, respectively. the result was 0.0129. Then the expected coincidence number can
be obtained by multiplying PT by the total number of years from
Estimation of Flood Coincidence Risk 1950 to 2010, yielding 0.79. Based on the measured data from 1950
to 2010, there was only one coincidence occurrence date of AM
flood (September 1, 1966). Therefore, the expected value that was
Coincidence Risk of Flood Occurrence Dates
computed based on the modeled joint probabilities [Eq. (11)] was
With Eq. (11), the coincidence probability of the occurrence date
in accordance with observations.
of AM floods occurring on each day in the period from June 1 to
The estimated daily coincidence probabilities of the monthly
October 31 can be calculated, and the results are shown in Fig. 9.
maximum flood occurrence dates of the two rivers in the months
Higher coincidence probabilities of AM flood occurrence dates
June–October are shown in Fig. 10. It can be seen that the daily
are seen during the period from mid-July to early September than
coincidence probability of the monthly maximum flood occurrence
other periods. Three coincidence peaks occur on July 28, August
14, and August 31, with the associated probabilities of 5.27 × 10−4 , date is greater than that of the AM flood. For monthly maximum
4.09 × 10−4 , and 1.64 × 10−4 , respectively. The results suggest that floods, there is only one coincidence peak in each month, which oc-
the flood control water levels of Xiluodu and Xiangjiaba Reservoirs curs on June 25, July 18, August 15, September 11, and October 7
during this period should not be raised too high above their flood with associated probabilities of 2.26 × 10−2 , 6.55 × 10−3 , 4.64 ×
limit levels of 560 and 360 m, respectively. Instead, the Xiluodu– 10−3 , 9.93 × 10−3 , and 2.36 × 10−2 , respectively. This indicates
Xiangjiaba cascade reservoirs should maintain a flood control stor- that for monthly maximum floods, June and October have great
age to store the flood in the Jinsha River and reduce the flood coincidence probabilities, while August has the smallest coinci-
coincidence risk of the two rivers from mid-July to early September. dence probability.
Major storm events in the Jinsha River Drainage Basin and the Min
River Drainage Basin occur during the months of June to August
(Shen and Yang 2007) and July to August, respectively, which are
influenced by the southwest monsoon in the Indian Ocean and the
Fig. 9. Daily coincidence risk of occurrence date of annual maximum Fig. 10. Daily coincidence risk of occurrence date of monthly max-
floods of Jinsha River and Min River during summer. imum floods of Jinsha River and Min River.
floods in the months of June through October were obtained, and dence risk through flood control operation of Xiluodu–Xiangjiaba
the results are presented in Table 8. It shows that the estimated cascade reservoirs.
coincidence numbers in the months were in accordance with obser- For comparison, the coincidence probabilities of flood magni-
vations. Compared with the AM flood, the coincidence probabilities tudes without consideration of historical floods are also presented
of occurrence dates for monthly maximum floods in June to October in Table 9. It shows that the coincidence probabilities of flood
are higher. The coincidence probability of occurrence dates for magnitudes were underestimated when the historical floods were
monthly maximum floods in October reach a maximum, a corre- ignored.
sponding coincidence number of 8.43. The coincidence probability Presented in Table 10 are the coincidence probabilities of
of monthly maximum flood occurrence dates in August is at its monthly flood peaks of the two rivers in the months June–October.
minimal value, a corresponding coincidence number of 2.17. Compared with AM flood peaks, the coincidence probabilities of
0.1%, 0.2%, 1%, 2%, and 10% of the monthly maximum flood
Coincidence Risk of Flood Magnitudes peaks increase with time during the five months and reach the maxi-
Table 9 shows the coincidence probabilities of flood magnitudes mum in October.
for different design floods on the Jinsha and Min Rivers. It can be
seen that the coincidence probabilities of flood magnitudes for
0.1%, 0.2%, 1%, 2%, and 10% design floods of the two rivers Summary and Conclusions
were 1.79 × 10−6 , 3.60 × 10−6 , 1.88 × 10−5 , 4.33 × 10−5 , and
2.82 × 10−4 , respectively. Floods with higher frequencies resulted Flood coincidence risk represents a key issue in reservoir oper-
in higher coincidence probabilities. The estimated coincidence ation and management. In this study, the coincidence risk of floods
probabilities shown in Table 9 are consistent with the results pre- on the Jinsha and Min Rivers was estimated with multivariate cop-
sented by Chen et al. (2011, 2012): the coincidence risks of flood ulas. Besides the 61-year systematic streamflow records, archived
Table 9. Coincidence probabilities of annual maximum flood peaks of Jinsha and Min Rivers
Min River
Jinsha River
(%) Data 0.1% 0.2% 1% 2% 10%
0.1 SH 1.79 × 10−6 2.46 × 10−6 4.15 ×10−6 5.31 × 10−6 6.81 × 10−6
S 1.65 × 10−8 3.30 × 10−8 1.65 × 10−7 3.29 × 10−7 1.63 × 10−6
0.2 SH 2.46 × 10−6 3.60 × 10−6 6.92 × 10−6 9.24 × 10−6 1.26 × 10−5
S 3.30 × 10−8 6.60 × 10−8 3.29 × 10−7 6.58 × 10−7 3.25 × 10−6
1 SH 4.15 × 10−6 6.92 × 10−6 1.88 × 10−5 2.89 × 10−5 4.97 × 10−5
S 1.65 × 10−7 3.29 × 10−7 1.65 × 10−6 3.28 × 10−6 1.62 × 10−5
2 SH 5.91 × 10−6 1.07 × 10−5 3.85 × 10−5 4.33 × 10−5 1.73 × 10−4
S 3.29 × 10−7 6.58 × 10−7 3.28 × 10−6 6.57 × 10−6 3.25 × 10−5
10 SH 6.81 × 10−6 1.26 × 10−5 4.97 × 10−5 9.27 × 10−5 2.82 × 10−4
S 1.63 × 10−6 3.25 × 10−6 1.62 × 10−5 3.25 × 10−5 1.61 × 10−4
Note: SH = systematic record and historical floods; S = systematic records; and 0.1%, 0.2%, 1%, 2%, and 10% denote the frequencies of design floods.
Table 10. Coincidence probabilities of monthly maximum flood peaks of Jinsha and Min Rivers
Flood frequency
Month 0.1% 0.2% 1% 2% 10%
June 3.94 × 10−7 1.68 × 10−6 4.01 × 10−5 1.54 × 10−4 3.48 × 10−3
July 2.12 × 10−7 8.48 × 10−7 2.10 × 10−5 8.31 × 10−5 1.91 × 10−3
August 1.48 × 10−7 5.91 × 10−7 1.46 × 10−5 5.79 × 10−5 1.33 × 10−3
September 2.93 × 10−7 1.17 × 10−6 2.90 × 10−5 1.16 × 10−4 2.64 × 10−3
October 6.45 × 10−7 2.58 × 10−6 6.37 × 10−5 2.51 × 10−4 5.68 × 10−3
date in ascending order, flood peaks of the Jinsha River occurred reached a maximum in October.
mainly in early September, mid-August, and late July, while The results of this study showed the spatiotemporal character-
flood peaks of the Min River occurred mainly in late July, mid- istics of floods of the Jinsha and Min Rivers, providing coincidence
August, and early July. probabilities for different design floods of the rivers during the
2. The correlations of flood magnitudes and occurrence dates be- period from June 1 to October 31, which can be utilized to guide
tween the two stations were positive. On performing goodness- cascade reservoir flood control operations on the Jinsha River.
of-fit tests for the GH, Clayton, AMH, Frank, Gaussian, and It is suggested that the flood control water levels of Xiluodu
Student t copulas, it was found that the GH copula was the best and Xiangjiaba Reservoirs should not be raised too high above
fitted copula for the two dependence structures and was selected their flood limit levels during the period from mid-July to early
to establish the joint distributions of flood magnitudes and oc- September in order to maintain certain flood control storage and
currence dates of the two rivers, respectively. If the marginal reduce coincidence flood risk.
distribution of flood magnitude was established without con- The accuracy of a joint probability model can greatly affect the
sidering historical floods, the Clayton copula would be most results of this coincidence flood risk analysis with copula-based
appropriate for modeling the joint distributions of the flood joint distributions. To improve model accuracy, the use of nonpara-
magnitudes of the two rivers. metric copulas or two-parameter copulas should be considered in
3. Higher coincidence probabilities of AM flood occurrence dates future studies. In addition, coincidence risk analysis is based on the
of the two rivers occurred in the period from mid-June to early assumption that flood occurrence date and flood magnitude are in-
September than other periods. Three coincidence peaks oc- dependent of each other, which will need to be verified in future
curred on July 28, August 14, and August 31. The coincidence research.
Appendix I. AM Flood Peaks and Occurrence Dates of Pingshan and Gaochang Stations in 1950–2010
Copula Cθ ðu; vÞ θ∈
Ali–Mikhail–Haq uv=½1 − θð1 − uÞð1 − vÞ ½−1; 1Þ
Gumbel–Hougaard expf−½ð− ln uÞθ þ ð− ln vÞθ 1=θ g ½1; ∞Þ
1 ðe−θu − 1Þðe−θv − 1Þ
Frank − ln 1 þ ð−∞; þ∞Þ \ f0g
θ ðe−θ − 1Þ
Clayton −θ
ðu þ v − 1Þ −θ −1=θ ð0; ∞Þ
2
−1 −1 1 s − 2θst þ t2
Gaussian ∫ Φ−∞ðuÞ ∫ Φ−∞ðvÞ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp − dsdt ½−1; 1
2π ð1 − θ2 Þ 2ð1 − θ2 Þ
vþ2
t−1 −1
v ðuÞ ∫ tv ðvÞ
1 s2 − 2θst þ t2 − 2
Student t ∫ −∞ −∞ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 þ 2
dsdt ½−1; 1
2π ð1 − θ2 Þ vð1 − θ Þ
Note: Φ = cumulative distribution function of standard normal distribution; and tv = standard t distribution with v degrees of freedom.
Appendix III. Estimated Parameters of Marginal Distributions for Annual Maximum Flood Peaks and
Occurrence Dates
Estimated parameters
Flood peak Occurrence date
Distribution Probability density function Estimated parameter Pingshan Gaochang Pingshan Gaochang
GEV 1h α i 1 −1 n h α i1 o
α̂ 0.171 0.223 −0.391 −0.214
1 − ðx − δÞ α exp − 1 − ðx − δÞ α
β β β β̂ 4,014.3 3,662.6 25.588 19.428
δ̂ 15,421 13,931 68.594 54.689
Gamma 1 x−δ α̂ 2.304 3.449 129.87 24.709
exp − ðx − δÞα−1 x>0 β̂
ΓðαÞβ α β 4,067.7 3,210 4.027 2.143
δ̂ 9,177 5,996.9 −37.051 −202.02
Lognormal 1 ½lnðx − βÞ − δ2 α̂ 0.523 0.442 0.120 0.04
pffiffiffiffiffiffi exp − x>β β̂ −542.85 −103.82
ðx − βÞα 2π 2α2 7,100.8 4,032.4
δ̂ 9.208 9.372 6.428 5.106
LP-III βα α̂ 9.321 5.623 3.180 14.933
ðln x − δÞα−1 exp½−βðln x − δÞ x > eδ
xΓðαÞ β̂ 0.101 0.140 0.089 0.212
δ̂ 8.832 8.8972 5.414 4.946
P-III βα α̂ 2.687 2.441 1.785 1.543
ðx − δÞα−1 exp½−βðx − δÞ x > δ
ΓðαÞ β̂ 0.0006 0.0004 0.0026 0.0047
δ̂ 9,374.95 7,733.94 26.58 31.96
Mixed von Mises Pn pi p̂i — — 0.143 0.63
exp½ki cosðx − ui Þ 0 ≤ x ≤ 2π
i¼1
2πI 0 ðki Þ 0.470 0.12
0.387 0.25
μ̂i — — 3.08 3.02
3.93 2.38
2.23 1.61
k̂i — — 68.33 3.07
9.53 299.85
3.74 7.90
Note: α̂, β̂, and δ̂ = estimated shape, scale, and location parameters of probability density distribution; i = each component of mixed von Mises distribution;
pi = mixing proportion; and ui and ki = mean direction and concentration parameter for component i of mixed von Mises distribution, respectively.
Massey, F. J. 1951. “The Kolmogorov-Smirnov test for goodness of fit.” Sraj, M., N. Bezak, and M. Brilly. 2015. “Bivariate flood frequency analy-
J. Am. Stat. Assoc. 46 (253): 68–78. https://doi.org/10.1080/01621459 sis using the copula function: A case study of the Litija station on the
.1951.10500769. Sava River.” Hydrol. Process. 29 (2): 225–238. https://doi.org/10.1002
Merz, B., and G. Blöschl. 2008. “Flood frequency hydrology. I: Temporal, /hyp.10145.
spatial, and causal expansion of information.” Water Resour. Res. Stedinger, J. R., and V. W. Griffis. 2008. “Flood frequency analysis in the
44 (8): W08432. https://doi.org/10.1029/2007WR006744. United States: Time to update.” J. Hydrol. Eng. 13 (4): 199–204. https://
Merz, B., and G. Blöschl. 2008. “Flood frequency hydrology. II: Combin- doi.org/10.1061/(ASCE)1084-0699(2008)13:4(199).
ing data evidence.” Water Resour. Res. 44 (8): W08433. https://doi.org Stedinger, J. R., and G. D. Tasker. 1985. “Regional hydrologic analysis. 1.
/10.1029/2007WR006745. Ordinary, weighted, and generalized least squares compared.”
Merz, B., J. Hall, M. Disse, and A. Schumann. 2010. “Fluvial flood risk Water Resour. Res. 21 (9): 1421–1432. https://doi.org/10.1029
management in a changing world.” Nat. Hazards Earth Syst. 10 (3): /WR021i009p01421.
509–527. https://doi.org/10.5194/nhess-10-509-2010. Viglione, A., R. Merz, J. L. Salinas, and G. Blöschl. 2013. “Flood
Michael, Z., and U. Stan. 2013. Statistical decision problems, 45–52. frequency hydrology. III: A Bayesian analysis.” Water Resour. Res.
New York: Springer. 49 (2): 675–692. https://doi.org/10.1029/2011WR010782.
Mirabbasi, R., A. Fakheri-Fard, and Y. Dinpashoh. 2012. “Bivariate drought Wang, C., N. B. Chang, and G. T. Yeh. 2009. “Copula-based flood fre-
frequency analysis using the copula method.” Theor. Appl. Climatol. quency (COFF) analysis at the confluences of river systems.” Hydrol.
108 (1–2): 191–206. https://doi.org/10.1007/s00704-011-0524-7. Process. 23 (10): 1471–1486. https://doi.org/10.1002/hyp.7273.
MWR (Ministry of Water Resources). 2006. Regulation for calculating de- Wang, D., Y. T. Li, J. Y. Deng, and J. J. Fang. 2014. “Preliminary study of
sign flood of water resources and hydropower projects. [In Chinese.] impounding optimization of cascade reservoirs in upper Yangtze River.”
Beijing: Water Resources and Hydropower Press. [In Chinese.] J. Sediment Res. 2: 62–67.
Nelsen, R. B. 2006. An introduction to copulas. 2nd ed. New York: Wang, L. F. 1999. “Overview of the Jinsha River’s historical floods
Springer. features.” [In Chinese.] Sichuan Water Conservancy 20 (3): 46–48.
Peng, Y., K. Chen, H. Yan, and X. Yu. 2017. “Improving flood-risk analysis Xiao, T. G. 2001. “Analysis on flood encounter of Jinsha river and Minjiang
for confluence flooding control downstream using Copula Monte Carlo river.” [In Chinese.] Yangtze River 23 (1): 30–32.
method.” J. Hydrol. Eng. 22 (8): 04017018. https://doi.org/10.1061 Xiao, T. G., and K. Y. Zhao. 2002. “Floods in the Min River.” [In Chinese.]
/(ASCE)HE.1943-5584.0001526. Express Water Resour. Hydropower Inf. 23 (2): 29–30.
Prohaska, S., and A. Ilic. 2010. Coincidence of flood flow of the Danube Xiong, L. H., S. L. Guo, Y. Xiao, and H. F. Yuan. 2005. “Application of
River and its tributaries: Hydrological processes of the Danube River copulas to multivariate hydrological frequency analysis.” [In Chinese.]
Basin, 175–226. Dordrecht, Netherlands: Springer. Eng. J. Wuhan Univ. 38 (6): 16–19.
Prohaska, S., A. Ilic, and B. Majkic. 2008. “Multiple-coincidence of flood Xu, Q., and H. Tong. 2012. “Characteristics of flow and sediment change in
waves on the main river and its tributaries.” In Proc., IOP Conf. Series: Yangtze River in recent 50 years.” [In Chinese.] J. China Hydrol. 32 (5):
Earth and Environmental Science. Bristol, UK: IOP Publishing. 38–47.
Razali, N. M., and Y. B. Wah. 2011. “Power comparisons of Shapiro-Wilk, Yan, B., S. Guo, L. Chen, and P. Liu. 2010. “Flood encountering risk analy-
Kolmogorov-Smirnov, Lilliefors and Anderson-Darling tests.” J. Stat. sis for the Yangtze River and Qingjiang River.” [In Chinese.] Shuili
Model. Anal. 2 (1): 21–33. Xuebao 41 (5): 553–559.
Reis, D. S., and J. R. Stedinger. 2005. “Bayesian MCMC flood frequency Yan, B., S. Guo, and W. Yu. 2013. “Coincidence risk of flood hydrographs
analysis with historical information.” J. Hydrol. 313 (1–2): 97–116. between Yangtze River and Qing River.” [In Chinese.] J. Hydroelectric
https://doi.org/10.1016/j.jhydrol.2005.02.028. Eng. 32 (1): 50–53.
Rosenblatt, M. 1952. “Remarks on a multivariate transformation.” Ann. Yan, H., and F. G. Edwards. 2013. “Effects of land use change on hydrologic
Math. Stat. 23 (3): 470–472. https://doi.org/10.1214/aoms/1177729394. response at a watershed scale, Arkansas.” J. Hydrol. Eng. 18 (12): 1779–
Sackl, B., and H. Bergmann. 1987. “A bivariate flood model and its 1785. https://doi.org/10.1061/(ASCE)HE.1943-5584.0000743.
application.” In Hydrologic frequency modeling, edited by V. P. Singh, Yan, H., and H. Moradkhani. 2015. “A regional Bayesian hierar-
571–582. Dordrecht, Netherlands: Springer. chical model for flood frequency analysis.” Stochastics Environ. Res. Risk
Salvadori, G., and C. D. Michele. 2013. “Multivariate extreme value Assess. 29 (3): 1019–1036. https://doi.org/10.1007/s00477-014-0975-3.
methods.” In Extremes in a changing climate, 115–162. Dordrecht, Yan, H., and H. Moradkhani. 2016. “Toward more robust extreme flood
Netherlands: Springer. prediction by Bayesian hierarchical and multimodeling.” Nat. Hazards
Schweizer, B., and A. Sklar. 1983. Probabilistic metric spaces. New York: 81 (1): 203–225. https://doi.org/10.1007/s11069-015-2070-6.
Elsevier Science. Yan, H., H. Moradkhani, and M. Zarekarizi. 2017. “A probabilistic
Shen, X., and W. Yang. 2007. “Rainstorm characteristics analysis on the drought forecasting framework: A combined dynamical and statistical
lower part of the Jinsha River basin.” [In Chinese.] J. Water Resour. approach.” J. Hydrol. 548: 291–304. https://doi.org/10.1016/j.jhydrol
Res. 28 (1): 39–41. .2017.03.004.
Shiau, J. T. 2003. “Return period of bivariate distributed extreme hydro- Yue, S. 1999. “Applying the bivariate normal distribution to flood fre-
logical events.” Stochastic Environ. Res. Risk Assess. 17 (1): 42–57. quency analysis.” Water Int. 24 (3): 248–254. https://doi.org/10.1080
https://doi.org/10.1007/s00477-003-0125-9. /02508069908692168.