## Abstract

Design flood quantiles are crucial for hydraulic structures design, water resources planning and management, whereas previous multivariate hydrological quantile estimation methods usually do not consider historical flood information. To overcome such limitations, a meta-heuristic inference function for margins (MHIFM) approach, coupling meta-heuristic algorithm with a modified inference function for margins (IFM) method, is developed for modeling the joint distributions of flood peak and volumes with incorporation of historical flood information. Then, the most likely realization (MLR) and equivalent frequency combination (EFC) methods are employed for selecting multivariate design floods on a quantile iso-surface. The Danjiangkou reservoir located in Hanjiang River basin, the first pilot basin of most regulated water resources management policy in China, is selected as a case study. Application results indicate that the MHIFM approach shows good performance for estimating the parameters of marginal and joint distributions; moreover, the MLR method yields safer design flood quantiles than the EFC method in terms of highest routed reservoir water levels. The proposed MHIFM approach associated with the MLR method is safer and more rational for reservoir design, which would provide rich information as the reference for flood risk assessment, reservoir operation and management.

## INTRODUCTION

Flood frequency analysis is performed to quantify the risk of flooding at different spatial locations and to provide guidelines for determining the design capacity of flood control structures. The design capacities of spillways, levees and several other flood control structures, and reservoir operation or management depend on the estimated magnitude of design flood quantiles (Xiao *et al.* 2009). Conventional hydrological frequency analysis mainly focuses on unique hydrological characteristics such as peak discharge or flood volume utilizing univariate distribution, which is unable to consider the interdependence between correlated hydrological variables and fails to provide a complete description of hydrological events (Favre *et al.* 2004). Therefore, a multivariate frequency analysis is needed for characterizing the random hydrological variables' interdependent features, particularly in designing hydraulic structures such as dams (Li *et al.* 2017).

The copula function, enabling the integration of heterogeneous dependence structures and marginal distributions, has been widely applied in constructing flexible joint distributions of interrelated hydrological variables since the introduction of copulas in hydrology and geosciences by De Michele & Salvadori (2003). This method has been applied mainly to the analysis of rainfall storms (e.g. Goel *et al.* 2000; De Michele & Salvadori 2003; Wei & Song 2017), drought frequency analysis (e.g. Shiau *et al.* 2006; Chen *et al.* 2016), flood frequency analysis (e.g. Favre *et al.* 2004; Salvadori & De Michele 2004; Brunner *et al.* 2018), synthetic flood hydrograph (e.g. Gräler *et al.* 2013; Xu *et al.* 2016), regional frequency analysis (e.g. Chebana & Ouarda 2009; Zhang *et al.* 2015a), and several other multivariate domains (e.g. De Michele *et al.* 2005; Soltani *et al.* 2018).

While the copula-based methodology for multivariate frequency analysis is well established and explored, a common problem exists in both univariate and multivariate flood frequency analysis, i.e. the gauged record is rarely long enough to yield a reliable estimation of design flood quantiles, especially for high return periods. Furthermore, the existence of a cyclic variation over periods longer than the duration of the records would introduce further bias (Stedinger & Cohn 1986; Guo & Cunnane 1991). To resolve these problems, hydrologists have considerable interest in augmenting flow records, increasing sample size and estimating uncertainty. For example, the peak over threshold sampling method was used to extract flood series and thus significantly increased sample size compared with the annual maximum sampling method (Bhunya *et al.* 2013). Several pioneering approaches such as regionalization procedures (Reed 2002), Bayesian method (Zhang *et al.* 2015b) and bootstrapping approaches (Yin *et al.* 2018a, 2018b) have been suggested to reduce the design flood estimation uncertainty. The widely acknowledged and reliable technique is to incorporate historical or paleological information into flood frequency analysis, which has been recommended as an official procedure (e.g. NERC 1975; MWR 2006). Several methods, such as historically weighted moments, maximum likelihood, probability weighted moments, and L-moments (Guo & Cunnane 1991; Hosking 1995) have been proposed for the incorporation of historical or paleological information into univariate flood frequency analysis. To the authors' knowledge, only a very few studies address the bivariate characteristics of flood events considering historical information (e.g. Li *et al.* 2013), and incorporating flood information within a multivariate framework is still not available.

Besides the problem of incorporating historical floods, the question of estimating joint flood quantiles still has not yet been properly addressed in the multivariate frequency analysis framework (Chebana & Ouarda 2011; Yin *et al.* 2017). Under the multivariate framework, the choice of an appropriate return period for structure design leads to infinite combinations of correlated hydrological variables. These combinations are generally not equivalent, from a practical point of view, i.e. different joint design floods would lead to differences of flood hydrograph and highest water level which occurs in reservoir routing, which likely cause non-negligible discrepancy in spillway capacity design, reservoir operation and management (Volpi & Fiori 2012; Yin *et al.* 2018a, 2018b). To overcome such limitations and derive unique and rational design flood quantiles, many researchers (e.g. Liu *et al.* 2015; Xu *et al.* 2016; Li *et al.* 2017) adopted the equivalent frequency combination (EFC) method, which assumed that different hydrological variables have the same occurrence likelihood under a given Joint Return Period (JRP). Despite the EFC method having ease of implementation, the statistical and theoretical basis of the equal frequency assumption is questionable (Guo *et al.* 2018). To develop a more rational design flood estimation approach, several pioneering works have been conducted. For example, Salvadori *et al.* (2011) proposed a novel concept of ‘most-likely design realization’ to choose the event with the highest likelihood along the quantile curve. Li *et al.* (2017) developed a conditional expectation combination (CEC) method to derive the quantiles of flood peak and 7-day volume in Geheyan reservoir in China. Yin *et al.* (2017) developed the conditional most likely combination (CMLC) method to depict the interdependence between seasonal floods by using the conditional density function to measure the occurrence likelihood of seasonal flood events. While the CEC and CMLC methods are applicable to estimate the bivariate flood quantiles, they may have less physical realism in the multivariate framework.

Although several studies investigating the interdependence between correlated hydrological variables have been conducted, there is little previous work on deriving multivariate joint design floods with incorporation of historical information, which is extremely crucial in making decisions about flood control and water resources management, particularly for reservoirs with large regulation capacity. This study develops a meta-heuristic inference function for margins (MHIFM) approach to model the joint distributions of flood peak and volumes with incorporation of historical information. Moreover, we derive the general formulae of most likely realization (MLR) and EFC methods for selecting multivariate design flood quantiles under a given JRP.

The remainder of this paper is organized as follows. The ‘Study area and data’ section introduces the systematic record and historical floods in Hanjiang River basin. In the ‘Methodology’ section, we present the methodology used to estimate the parameters of marginal and multivariate distributions and flood quantile selection methods. The ‘Results and discussion’ section addresses a case study in Danjiangkou reservoir. Finally, the ‘Conclusions’ section summarizes the main results and provides necessary explanations.

## STUDY AREA AND DATA

### Danjiangkou reservoir

The Hanjiang River basin, the first pilot basin of most regulated water resources management policy in China, is the largest tributary of the Yangtze River. The Hanjiang River spans 1,570 km with a basin area of 159,000 km^{2}, which is dominated by a sub-tropical monsoon climate. The annual precipitation in Hanjiang River basin varies from 700 to 1,100 mm, associated with large spatial and seasonal variability. The rainfall from June to October accounts for 75% of the annual amount, and the convective rainstorms in the early summer and long-lasting frontal rains in the autumn often lead to intense flash flood hazards. The Danjiangkou reservoir marked in Figure 1 is a key control project located in the middle reach of Hanjiang River basin, which is the source for the middle route of the South–North Water Diversion Project in China. The Danjiangkou reservoir is a multi-purpose reservoir, which has serious conflicts between flood control, water supply and hydropower generation. The inflow of Danjiangkou reservoir is mainly concentrated from late June to early October, and intense runoff during July and September plays a dominant role in water resources utilization and flood risk prevention. The Danjiangkou reservoir has multi-years regulation ability and the flood control storage during flood season is over 11.1 billion m^{3}. Since the Danjiangkou reservoir is the key structural component in the Hanjiang River flood control system and plays a critical role in flood control and water supply in China, it is of great significance to investigate its design floods for guiding reservoir operation and management.

### Systematic record and historical floods

According to the regulation and operation guidelines of the Danjiangkou reservoir, peak discharge, 7- and 15-day flood volumes are considered for deriving the design flood quantiles and the corresponding design flood hydrograph (DFH) (Xu *et al.* 2016). The annual maximum peak discharge, 7-day flood volume, and 15-day flood volume are available with a systematic record of 86 years (1929–2014) from the Bureau of Hanjiang River Water Resources Management. These data series after the dam construction have been restored to natural flow and have been checked for possible long-term trends and abrupt changes, with the test results demonstrating that the dataset is stationary and represented (Xu *et al.* 2016). The dataset has been widely employed in hydraulic structure capacity design, reservoir operation and water resources management in the Danjiangkou reservoir (e.g. Yang *et al.* 2017; Zhou *et al.* 2017; Shen *et al.* 2018). Besides the systematic observations, the historical flood events have been investigated by Changjiang Water Resources Commission (CWRC) in China. The gathered information is from gauging authority records, historical documents, archives, flood marks and stone inscriptions showing the concrete positions of high water stages recorded. As a result, the six largest historical floods (see Table 1) since 1,583 were quantitatively evaluated by the CWRC.

## METHODOLOGY

### Copula functions and JRP

*Q*) and flood volumes to investigate the joint probability of flood occurrence by copulas and to derive the DFHs, in which

*m*represents the number of flood volumes needed for estimating DFHs. The multivariate joint distribution

*H*can be expressed in terms of univariate marginal distributions and the associated dependence function, known as Sklar's theorem (Sklar 1959): where denotes the marginal cumulative density functions (CDFs) of

*X*

_{i}, is the copula function with corresponding parameter vector

**.**

*θ**et al.*2013; De Michele

*et al.*2013; Volpi & Fiori 2014; Salvadori

*et al.*2016). The Danjiangkou reservoir is at risk when either flood peak or flood volume exceeds the design standard, and therefore the multivariate phenomenon for this particular situation is defined based on OR JRP, which can be defined by copulas accordingly (De Michele

*et al.*2005): where

*μ*is the mean inter-arrival time between two consecutive events (in the case of annual maxima

*μ*= 1 year); , are the marginal distributions of peak discharge and flood volumes, respectively.

### Univariate maximum likelihood estimation with historical information

In univariate hydrological frequency analysis, several methods for incorporating historical information have been proposed, among which the maximum likelihood estimation method is widely acknowledged and employed (Hald 1949; Stedinger & Cohn 1986; Chen & Singh 2018). In the annual maximum flood series shown in Figure 2, there is a total of *g* known floods. Of these, *l* is known to be the largest flood during the maximum historical textual period of *n* years, while the *c* large flood occurred during the systematic record (recently gauged data) of *s* year (*c* ≤ *l*, *c**<**s*, *s* ≤ *n* and *g* = *s* + *l* − *c*). A fixed threshold *X*_{0} is assumed to be exceeded by the *l* largest floods and is not exceeded by any of the remaining *n* − *l* floods. It is also noted that the *r* (*r* = *l* − *c*) floods in the pre-gauging period *h* (*h* = *n* − *s*) are known as they are included in the *l* floods which exceed *X*_{0}, and it is assumed that no other floods exceed the threshold during the pre-gauging period.

*f*and

_{X}*F*denote the probability density function (PDF) and the CDF of variable

_{X}*X*, respectively. The likelihood function for the mixed sample of

*s*+

*r*known and

*h*−

*r*unknown data is given as (Condie 1986; Stedinger & Cohn 1986; Guo & Cunnane 1991): where

**is the parameter vector of**

*α**f*and

_{X}*F*.

_{X}*c*flood events exceeding the perception threshold

*X*

_{0}occur during the systematic period

*s*(analogously to the sketch in Figure 2), the

*c*events are removed from the period

*s*and are treated as historical data (Bayliss & Reed 2001). Then, Equation (3) can be expressed as: where

*x*(

_{i}*i*= 1, 2 …

*s*−

*c*) denotes the systematic data less than the threshold

*X*

_{0},

*y*(

_{j}*j*= 1, 2 …

*l*) denotes the

*l*largest floods exceeding the threshold

*X*

_{0}.

*et al.*2013): The maximum likelihood estimates are those values of

**that maximize Equation (5).**

*α*### MHIFM approach

The inference function for margins (IFM) method is the preferred fully parametric method for estimating the parameters of multivariate distributions due to its robustness and ease of implementation (Joe & Xu 1996; Joe 1997; Baharith 2017). The conventional IFM method is only applicable for systematic data series without historical floods, which could not satisfy the requirements of engineering practice in spillway capacity design. Li *et al.* (2013) developed a modified IFM method to estimate the parameters of marginal and bivariate distribution with consideration of historical information. The modified IFM method focused on solving the bivariate problems, failing to incorporate flood information within a multivariate framework, thus limiting the applicability in reservoirs with large capacity. To address such limitations, an MHIFM approach is developed as discussed below.

Let and () denote the systematic data of flood peak (*Q*) and volumes (*W _{i}*), respectively; and () denote the flood peak and volumes of the

*l*largest floods with the same length, respectively;

*Q*

_{0}(or

*W*

_{0i}) is the fixed threshold exceeded by the

*l*largest flood peaks (or volumes) and not exceeded by any of the remaining

*n*−

*l*flood peaks (or volumes). Furthermore, let , denote the univariate marginal PDFs, and (

*q*), denote the univariate marginal CDFs of variables

*Q*and

*W*respectively; denotes the joint PDF.

_{i},*m*equals 1), referring to Equation (5), the likelihood function with historical floods for bivariate joint distribution can be described as (Li

*et al.*2013): where , , denote the parameter vector of

*F*

_{Q}(

*q*) and

*F*

_{W1}(

*w*

_{1}) respectively.

The parameter estimation procedure consists of *m* + 1 separate optimizations. The optimization of multivariate likelihood is a function of the dependence parameter vector. More specifically:

*et al.*1983). It has the capability of jumping out of the local optima for global optimization (Baykasoğlu & Gindy 2001), so it is used to estimate the parameters of marginal and joint distributions in the MHIFM method. The SAA emulates the gradual cooling process (i.e. annealing) of material, during which the temperature decreases stepwise. At each temperature, an initial configuration is generated and employed to evaluate the corresponding objective function

*E*(energy value). Then, a small displacement is imposed on the current configuration to obtain a new objective function

_{a}*E*. If , this new configuration is accepted as the best solution. Otherwise, it is accepted with the probability: where

_{b}*T*is the current temperature and

*K*is the Boltzmann's constant. If this probability is higher than a random value, this configuration is accepted as the best, although it is not better than the previous solution. Once the system has reached equilibrium at the current temperature

*T*, the algorithm continues and reduces temperature gradually until the system reaches its minimum temperature (Kirkpatrick

*et al.*1983; Bagherlou & Ghaffari 2018).

In this study, the optimization problem in SAA is a constrained single-objective optimization minimizing the opposite of the log-likelihood function or *L* separately in Equations (10) and (11), i.e. or −*L*. The main procedures of the SAA for the parameter estimation are summarized as follows:

- (1)
Initialize the temperature

*T*_{0}. - (2)
Set a variation interval for the parameters of distribution and generate initial solution configuration randomly.

- (3)
Calculate the value of objective function

*E*(e.g. −_{a}*L*)._{a} - (4)
Generate a new solution configuration by imposing a small displacement on current configuration and evaluate the value of objective function

*E*(e.g. −_{b}*L*)._{b} - (5)
If , the new solution is accepted. Otherwise, accept the new solution with a probability

*p*(*δE*). - (6)
Repeat steps (4) and (5) at current temperature.

- (7)
Reduce the temperature stepwise and repeat steps (2)–(6) until the stop criterion is satisfied.

### EFC method

After modelling the joint distribution with incorporation of historical information by the developed MHIFM approach, we need to estimate the design flood quantiles, which is necessary for flood control system design, water resources management and reservoir operation. However, for a given JRP , there are innumerable combinations of flood peak and volumes that satisfy Equation (2). Take the bivariate case as an example: the critical level curve as shown in Figure 3 is defined as a bivariate quantile curve under the OR JRP, with countless quantile combinations satisfying the flood prevention standard. Chebana & Ouarda (2011) segmented the *p*-level curve into a central part (i.e. the subset *BC* in Figure 3) and a tail part by employing its geometric properties. All the bivariate events along the quantile curve have the same JRP; however, these events may not be equivalent from a practical point of view (Volpi & Fiori 2012; Yin *et al.* 2017).

To meet the needs in engineering, a unique combination of flood peak and volumes should be determined. Hence, besides Equation (2), one or more constraints characterizing the interdependence between hydrological variables are necessary. The EFC method recommended by the Chinese design flood guidelines (MWR 2006) is widely employed due to its ease of implementation, which assumes that the flood peak and volumes have the same probability of occurrence, i.e. (or ). As shown in Figure 3, the tangent curve on which is depicted as a chosen constraint on the critical level curve and the intersection *A* point demonstrates the unique estimator of the EFC method. Moreover, point *D* depicted in Figure 3 indicates the estimated design flood quantiles by the conventional univariate method, in which only the marginal distribution is used under the assumption that the design frequency of different variables is equal to the joint design frequency.

*u*, and ) can be estimated by the solution of Equation (14), which is expressed as follows: It should be noted that other copulas with more complicated formulas or higher-dimension copulas are sometimes needed and the values of could be estimated by the numerical method such as a trial and error method (Liu

*et al.*2015).

### MLR of flood quantile

As previously stated, all possible flood events under a given JRP differ in terms of their probability of occurrence. Salvadori *et al.* (2011) measured the likelihood of flood events with the joint PDF and proposed the novel concept of ‘most likely design realization’ to choose the point with the highest likelihood along the *p*-level isolines. In this paper, the general formulae of MLR are derived as discussed below.

*et al.*2018a, 2018b) is applied for solving Equation (17) to obtain the MLR. For a given JRP , the Lagrange function is constructed as: where is the Lagrange multiplier.

The nonlinear Equation (20) can be solved by numerical methods, such as the incomplete Jacobian Newton method proposed by Liu & Ni (2008).

## RESULTS AND DISCUSSION

### Parameter estimation for marginal distributions

*α*,

*β*and

*δ*are shape, scale and location parameters of the P-III distribution, respectively.

The series of flood peak discharge (*Q*), 7-day flood volume (*W*_{1}) and 15-day flood volume (*W*_{2}) were considered in the Danjiangkou reservoir following a previous study (Xu *et al.* 2016). The length of the systematic observations is unequivocally given: *s* = 86 years. Since one flood event occurring in 1935 ranked second among the historical and systematic records, this flood is usually taken as an extraordinary flood during the systematic period, and thus *c* = 1 and *r* = *l* − 1. While six historical floods during the pre-gauge period *h* = 347 (i.e. from 1583 to 1929) were investigated, the total largest floods *l* equals 7 and *r* equals 6; the perception thresholds of peak discharge, 7-day flood volume and 15-day flood volume are *Q*_{0} = 3,700 m^{3}/s, *W*_{01} = 110 10^{8} m^{3} and *W*_{02} = 150 10^{8} m^{3}, respectively.

The parameters of the P-III marginal distributions were estimated by the first stage of the MHIFM approach in Equation (11) and are listed in Table 2. To compare the estimated design flood quantiles by the developed MHEFM approach with that by the conventional IFM method, we also estimated the parameters of marginal distributions by the conventional IFM method which only enables the consideration of systematic data.

A Chi-Square goodness-of-fit and the Kolmogorov–Smirnov (K-S) test are used to test the assumption *H*_{0}, that the flood magnitudes follow the P-III distribution. The results in Table 2 show that hypothesis could not be rejected at the 5% significance level because , and the P-III distributions can also pass the K-S test because *D*_{KS} < *D*_{0.05}. The marginal distribution frequency curves of flood peak, 7-day and 15-day flood volumes by employing MHIFM and IFM methods are illustrated in Figure 4, and indicates that these theoretical distributions can fit the observed data reasonably well.

### Joint distribution of flood peak and volumes

Different families of copula functions have been proposed and described by Nelsen (2006), among which the Archimedean family is more desirable for hydrological analyses because it can be easily constructed and applied to whether the correlation among the hydrological variables is positive or negative (Kim *et al.* 2018; Lei *et al.* 2018). The Gumbel–Hougaard (G-H), Frank and Clayton copulas of the Archimedean family were applied to construct the bivariate joint distributions between *Q* and *W*_{1}, and the symmetric and asymmetric copulas were selected as candidates for constructing the trivariate dependence of *Q*–*W*_{1}–*W*_{2}. The expressions of the copulas can be found in Xu *et al.* (2016). The parameters of bivariate and trivariate copulas were estimated by MHIFM and IFM, respectively. The root mean square errors (RMSE) and AIC values were calculated and listed in Table 3 to assess the performance of the parameter estimation approaches. The statistical results in Tables 2 and 3 both indicate that the statistical values under MHIFM method are smaller than those of the IFM method. This implies that incorporating historical information in multivariate frequency analysis is possible and the developed MHIFM approach has good performance for estimating the parameters of marginal and joint distributions.

Table 3 also indicates that the bivariate G-H copula has the smallest RMSE and AIC values for constructing the joint distribution of *Q*–*W*_{1}, and the trivariate asymmetric G-H copula performs best for *Q*–*W*_{1}–*W*_{2}. Therefore, the bivariate and trivariate asymmetric G-H copula is used to model the dependence between the extreme maximum annual flood peak and volumes in this study. Figure 5 demonstrates the bivariate joint CDFs of flood peak and volumes by MHIFM and IFM methods respectively, which show the CDFs constructed by G-H copula is continuous.

### Bivariate design flood quantile estimation

After constructing the joint distribution by MHIFM and IFM methods, the EFC and MLR methods are used to estimate the bivariate design flood quantiles of flood peak and 7-day volume under JRP = 200, = 100, = 50, = 20 and = 10 year, respectively. For the purpose of comparison, the univariate method considering historical floods (Chinese design flood guidelines) was also used to estimate the design flood quantiles by assuming that the univariate return period was to the bivariate return period. Figure 6 demonstrates the bivariate quantile curves and the estimated design flood quantiles under the G-H copula.

To assess the rationality of the estimated design flood quantiles by the EFC and MLR methods, the 95% confidence interval (CI) bounds were derived by the approach proposed by Li *et al.* (2017). The selected flood events and coordinate pairs of the lower and upper bounds of the 95% CI are given in Figure 6, and it is revealed that the estimated design floods by the EFC and MLR methods are both within the 95% CI. This means that the EFC and MLR methods are reasonable for estimating the bivariate flood quantiles in the Danjiangkou reservoir.

The estimated quantiles of univariate and bivariate methods are listed in Table 4. The results in Table 4 indicate that the design flood quantile based on the univariate distributions considering historical floods are less than the EFC and MLR methods. This especially illustrates that the bivariate methods associated with the OR JRP is safer than the univariate distribution method for the engineering practice, and this viewpoint has been confirmed by many authors (e.g. Shiau *et al.* 2006; Zhang *et al.* 2015b; Salvadori *et al.* 2016).

Comparing the results estimated by the MHIFM and IFM approaches in Table 4 and Figure 6, it is shown that for the IFM method the design flood quantiles selected by the EFC and MLR methods are both underestimated compared with the MHIFM method. This is due to the fact that the conventional IFM method fails to consider the historical floods and leads to underestimation for flood frequency analysis. Table 4 also demonstrates that the estimated flood quantiles under the IFM method are less than the univariate estimations when the return period is over 20 years. However, for the 10-year return period, both the EFC and MLR designed events under the IFM method are larger than the univariate design floods. This may imply that, even though the historical floods may have less impact on design floods under the low return period, considering historical floods is essential for hydrologic structure capacity design since the conventional IFM method could not satisfy the flood control standard, particularly for the high return period.

### Multivariate design flood quantile estimation

The Danjiangkou reservoir has large flood control storage during the flood season, and peak discharge, 7-day and 15-day flood volumes are necessary for deriving the design flood quantiles and the corresponding DFH. To meet this need, the trivariate surfaces of *Q*–*W*_{1}–*W*_{2} under the JRP of 200, 100, 50, 20 and 10 years are estimated by MHIFM and IFM methods. Figure 7 demonstrates the trivariate surfaces based on the asymmetric G-H copula function under three JRPs for the sake of saving space. It is demonstrated that numerous schemes of design flood quantiles of flood peak and volumes can satisfy a given flood prevention standard, and the surfaces formed by the IFM method are lower than that by the MHIFM approach. The EFC and MLR methods are employed to select design flood quantiles on the surfaces under both MHIFM and IFM approaches, and the estimated values are listed in Table 5.

It is shown in Table 5 that the MHIFM method has larger design flood quantiles of flood peak, 7-day and 15-day volumes compared with the estimated values by the IFM method. This is true under two quantile selection methods, which means that the parameter estimation approach has far heavier impacts on design flood quantiles than different quantile selection methods. Comparing the results of the EFC and MLR methods, it is found that the MLR method has smaller flood peak and larger volumes compared with the EFC method. Despite the design flood quantiles derived from the EFC and MLR methods, both can meet the flood prevention standards and these different joint design floods may lead to discrepancy of the flood hydrograph and highest water level which occurs in reservoir routing, which supports the necessity for comparing the hydrological loads on the structure of different joint design floods under a given JRP.

### DFH and corresponding reservoir routing

The estimated trivariate design flood quantiles of flood peak, 7-day and 15-day volumes under different methods are applied to derive the DFH, and the resulting highest reservoir water level is selected as an index to evaluate the effects of different hydrological loads on structure. The DFH for a dam is the flood of suitable probability and magnitudes adopted to ensure safety of the dam in accordance with appropriate design standards. The DFH could be derived by amplifying the typical ﬂood hydrographs (TFH), which has been widely used by practitioners (e.g. Yue & Rasmussen 2002; Tan *et al.* 2017). The flood hydrograph with high flood peak or large flood volume is usually selected as a TFH, and thus the flood hydrograph of 1964 is selected as a TFH following a previous study (Yin *et al.* 2017). The peak and volume amplitude method based on the selected TFH is used to derive the DFH (MWR 2006; Xiao *et al.* 2009), and the results are shown in Figure 8. The DFHs are routed through the reservoir with initial water level (flood control limiting water level, 160.0 m) according to the reservoir operation rule in Danjiangkou reservoir, and the corresponding highest reservoir water levels (*Z*_{max}) are calculated and listed in Table 5.

The comparison results listed in Table 5 show that Z_{max} obtained by the trivariate MHIFM method is higher than that by the conventional IFM method. Comparing the results of 200-year JRP, it is found that the differences under two selection methods among the MHIFM and IFM methods are both over 2.0 m. For 20- and 10-year return periods, the Z_{max} of MHIFM is higher, 0.19–0.30 m, than that of the IFM method, which implies that multivariate flood frequency analysis considering historical floods has a significant impact on flood control safety, especially for the high return period. The conventional IFM method, which fails to incorporate the historical floods, may lead to heavier flood hazards for hydraulic structure.

Table 5 also demonstrates the comparative results of Z_{max} estimated by the EFC and MLR methods, and it indicates that these different quantile selection methods have slight impacts on the corresponding Z_{max}. Comparing the Z_{max} by the EFC and MLR methods, it is found that the Z_{max} of the EFC method is lower than that of the MLR method, which is true under the same JRP and parameter estimation approach. For example, for the MHIFM approach, the Z_{max} of the EFC method (170.30 m) is lower than that of the MLR method (170.35 m) under 200-year JRP. This may be due to the fact that the Danjiangkou reservoir has a large amount of flood control storage with annual regulation ability, and therefore the design flood volume is relatively more important than peak discharge for flood prevention safety. As a consequence, the trivariate MLR method with slightly larger 7- and 15-day flood volume is safer for reservoir design compared with the EFC method.

## CONCLUSIONS

Flood quantile estimation is essential and important for deriving the DFH and determining reservoir operation rules. This study developed an MHIFM approach to model the joint distribution of flood peak and volumes with incorporation of historical information. The MLR and EFC method were employed for selecting unique design floods under a given JRP. The DFHs in Danjiangkou reservoir were derived based on trivariate design floods, and through reservoir routing the corresponding highest reservoir water levels (Z_{max}) were obtained. The main conclusions of this study are as follows.

- (1)
The investigated data series consist of systematic record and six historical flood events, and the fitting results indicate that the developed MHIFM approach that incorporates historical information into multivariate distribution has a good performance for estimating the parameters of marginal and joint distributions.

- (2)
The conventional IFM method that fails to consider historical floods produces underestimated design flood quantiles compared with the MHIFM method, and is also inferior to the univariate method with incorporation of historical information, implying that the conventional IFM method is unable to satisfy the flood prevention standard. The reservoir routing results show that Z

_{max}obtained by the trivariate MHIFM approach is higher than that obtained by the conventional IFM approach and the trivariate MLR method yields safer results for reservoir design than the EFC method. - (3)
The developed MHIFM approach associated with the MLR method could provide a more comprehensive and effective way for multivariate quantile estimation to incorporate historical information, particularly when the recorded data series is relatively short, which would provide rich information in engineering practice as the references for flood risk assessment, reservoir operation and management.

## ACKNOWLEDGEMENTS

This study is financially supported by the National Key R&D Program of China (2016YFC0402206) and the National Natural Science Foundation of China (51539009).