## Abstract

Based on the multivariate joint probability distribution of the discharge and water quality indicators, this paper analysed the occurrence probabilities and improvement probabilities of combinations of water quality indicators under different discharge conditions and then presented a method for calculating the optimal discharge to seek a balance between the discharge dispatch and water quality improvement. The method was used to construct the relationship curve between the discharge and joint improvement probability used by a copula function and then calculate the critical point on the curve. The proposed method was applied to the Yi River Basin above Gegou Station with data composed of the discharge and main pollution indicators (NH_{3}-N and COD_{Mn}) from 1982 to 2015. The results showed that the trivariate joint probability distribution can more reasonably reveal the statistical characteristics of different combinations of discharge and water quality indicators. Furthermore, the optimal discharges and the corresponding improvement probabilities that improved NH_{3}-N and COD_{Mn} to different grades were calculated. The calculation method took the interdependence of multiple water quality indicators into account, thereby providing a more reasonable method for using discharge dispatch data to improve the river water quality.

## INTRODUCTION

With the acceleration of urbanization in China, the deterioration of the quality of river water has become increasingly serious. One of the main water pollution control technologies used to improve the water quality is a reasonable discharge dispatch system (Jang *et al.* 2012; Qian *et al.* 2013; Shokri *et al.* 2014). However, the determination of the optimal discharge and the quantitative relationship between the discharge dispatch and the improvement in the water quality need to be further explored, especially with regard to the joint effects of discharge and water quality indicators.

Most of the studies conducted on the relationships between the water quantity and water quality follow two main approaches: mechanism models and non-mechanism models. The former takes into account the relationship of the mutual transformation of pollution factors and the coupling of hydrodynamic modules to simulate and predict concentration changes in pollution factors (Cox 2003; Wang *et al.* 2015; Yu *et al.* 2016). Nevertheless, many interdependent factors affect both the discharge and the water quality; consequently, the calibration of the parameters and the construction of a model are relatively complex. In contrast, non-mechanism models avoid the above-mentioned problems to some extent (Sun *et al.* 2011; Phan *et al.* 2016). Commonly used non-mechanism models include nonlinear algorithms (Liu *et al.* 2015), statistical models (Bonansea *et al.* 2015; Liu *et al.* 2016), Bayesian models and artificial neural network (ANN) models (Chitsazan *et al.* 2015). However, most non-mechanism models require that the variables obey the same marginal distribution or be transformed to the same distribution; this not only leads to the loss of information regarding the original data but also does not conform to the fact that variables such as the discharge and water quality obey different marginal distributions. Thus, most of the previous studies have some limitations on the occurrence probability of water quality combination events (Faruk 2010; Zhu *et al.* 2012). A copula can connect different univariate marginal distributions without changing the dependencies of variables (Zegpi & Fernandez 2010; Dai *et al.* 2014), i.e. it can construct the joint probability distribution of any univariate marginal distribution while retaining the information included in the initial data during the transformation process. Due to these advantages, copulas are widely used in multivariate hydrological analyses and calculations (Rauf & Zeephongsekul 2014). Their application is also becoming increasingly common in the risk analysis of floods (Kao & Chang 2012; Bezak *et al.* 2014; Li & Zheng 2016; Balistrocchi & Bacchi 2017) and the analysis of drought characteristics in watersheds (Hao & AghaKouchak 2013; Yusof *et al.* 2013; Bazrafshan *et al.* 2015; Li *et al.* 2017). In recent years, copulas have also been used to analyse the joint probability of the water quantity and water quality (Zhang *et al.* 2011; Wu *et al.* 2015). In summary, there have been few applications of copulas to the joint analysis of discharge and water quality; those few studies mainly focused on the bivariate joint distribution of discharge and individual indicators. Moreover, investigations of the multivariable joint distribution are relatively rare, and those focused on determining the optimal discharge for improving multiple water quality indicators are even scarcer.

Therefore, this paper presents novel research insomuch that the copula function is not only used to establish the multi-dimensional joint probability distribution of the water quantity and water quality but is also combined with the marginal probability distribution function to deduce the relationship curve between the discharge and the joint probability. Furthermore, the critical point of the improvement in the water quality is analysed from a statistical perspective, thereby extending the application scope of the copula function. The objective of this research is to construct a multivariate joint probability distribution of the discharge and multiple water quality indicators based on copula functions and then to use the joint probability distribution to construct a relationship curve between the discharge and joint improvement probability to analyse the comprehensive improvement of the river water environment under different discharge conditions. Moreover, the optimal discharge is suggested to seek a balance between the discharge dispatch and water quality improvement, providing a more scientific and reasonable basis for river water quality management.

## METHODS

### Multivariate joint probability distribution

*R*can be connected by using a copula function. The Archimedean copula, which is commonly used in hydrology (Schumann 2017), can be generally defined as follows: where is the generation function, is the inverse function of , and is the marginal distribution of a d-dimensional random variable.

Common symmetric Archimedean copulas include the Clayton, Gumbel–Hougaard (G-H), Frank and Ali–Mikhail–Haq (AMH) varieties. Their specific function forms and parameter ranges are shown in Table 1.

. | . | Ranges of . |
---|---|---|

Clayton | ||

G-H | ||

Frank | ||

AMH |

. | . | Ranges of . |
---|---|---|

Clayton | ||

G-H | ||

Frank | ||

AMH |

The establishment of a joint distribution based on a copula is divided into the following two steps:

- 1.
Selection of the marginal distributions of hydrological variables. Using the maximum likelihood method, the parameters of marginal distributions are estimated. Then, the proper marginal distributions are chosen using two goodness-of-fit tests: the Kolmogorov–Smirnov (K-S) test and the root mean squared error (RMSE) test.

- 2.
Determination of the parameter values of the multivariate joint distribution based on copula functions. The parameters of the multivariate joint distributions are estimated by the two-stage estimation method and the inference functions for margins (IFM) (Joe 2005).

### Calculation of the optimal discharge

#### Construction of the relationship curve between the discharge and joint probability

*m*is the number of water quality indicators (

*m*

*=*1, 2, … ,

*n*), and

*j*represents the different management targets (

*j*= 1, 2, … , 5) In addition, the marginal probability corresponding to can be obtained and recorded as . Then, the marginal probability of the discharge sequence

*Q*, where

_{i}*i*is the serial number of the discharge sequence (

*i*

*=*1, 2, …

*n*), is arranged in descending order and recorded as . The new

*m*+ 1 dimensional matrix is constructed by and as shown in Equation (2): Combined with the parameter of the selected copula function, the probability can be obtained. Then, the relationship curve can be drawn and recorded as , which denotes changes in the probability with an increase in the discharge when reaches the

*j*th grade.

### Determination of the critical point on the relationship curve

The critical point on the curve refers to a catastrophe point of the change rate of with an increase in *Q _{i}*. To calculate the critical point on the curve, whether a critical point exists along the curve must be clarified. It is difficult to deduce the relationship between

*Q*and directly because it is relatively complex. Therefore, the relationships between and and between and

_{i}*Q*are deduced individually to clarify the relationship between

_{i}*Q*and .

_{i}- 1.Relationship between and . The Frank copula, which is taken as an example here for the sake of explanation, is expressed as: Following the simplification of the copula: where is the occurrence probability of the
*j*th grade under*Q*conditions, , is the marginal probability of the_{i}*m*th indicator at the*j*th grade, and is the parameter of the copula function. Equation (4) constitutes a linear relationship between and . Similarly, the relationships between and for the other three functions in Table 1 can be deduced to be linear relationships.

- 2.
Relationship between

*Q*and . According to the most commonly used hydrological distribution function in China (Luo & Song 2014), the probability density function (PDF) of_{i}*Q*can generally be selected from the lognormal distribution, gamma distribution or P-III distribution. Furthermore, is a cumulative distribution function (CDF) of_{i}*Q*. Three types of CDFs are shown in Table 2._{i}

. | CDFs . | Ranges of parameters . |
---|---|---|

Lognormal | ||

Gamma | ||

P-III |

. | CDFs . | Ranges of parameters . |
---|---|---|

Lognormal | ||

Gamma | ||

P-III |

Taking the lognormal distribution as an example, the monotonicity of the integrand is analysed first. It can be proven that and are monotonically increasing functions when and that their values are all greater than 0, and thus, the integrand obtained by multiplying these two formulas is also an increasing function. The second derivative can be obtained, following which can be proven to be a convex function. Similarly, the relationship between *Q _{i}* and is also a convex function when

*Q*is subjected to a gamma or P-III distribution.

_{i}*k*is the slope of the curve.

## STUDY AREA

The Yi River is one of the most important rivers in Shandong Province. Gegou Station, which is located at the coordinates 110°28′ east longitude and 35°21′ north latitude in Gegou Village, Linyi City, China, is a controlled hydrological station located in the upper reaches of the Yi River. This study mainly examines the Yi River Basin above Gegou Station. The Yi River extends 110 km above Gegou Station with a watershed area of 5,500 km^{2}, and the terrain is dominantly hilly. In recent years, along with increasing population growth and economic development, non-point source pollution in the basin has polluted the water quality of the upper reaches of the Yi River and has become increasingly serious. In a few months, the water quality has exceeded a grade of V; as a consequence, the aquatic ecosystem in the basin has been severely damaged. Three reservoirs have been built within the basin, namely, the Tianshan Reservoir, Bashan Reservoir and Andi Reservoir. Thus, it is necessary to improve the water quality of the river by employing reasonable discharge dispatch through the use of water storage within these three reservoirs. The drainage map of the Yi River Basin above Gegou Station is shown in Figure 1.

The studied data were collected from Gegou Station, which has monitored the discharge and water levels since 1952 and provides complete water quality data records beginning in 1970. To ensure the validity and representativeness of the data, this paper chose the discharge and water quality data of Gegou Station from 1982 to 2014 (water quality data are recorded every other month) for a total of 153 months. This paper collected a total of six water quality indicators, namely, the chemical oxygen demand (COD), permanganate index (COD_{Mn}), biochemical oxygen demand (BOD), ammonia nitrogen (NH_{3}-N), total phosphorus (TP) and total nitrogen (TN), from Gegou Station. The ratio of each measured concentration to the grade III value of each water quality indicator is shown in Figure 2. The grade III values of the six water quality indicators are obtained from the ‘Environmental Quality Standards for Surface Water’ (GB3838-2002). Figure 2 shows that the excess ratios of the BOD, NH_{3}-N, COD_{Mn} and TN concentrations are significantly greater; however, the length and continuity of the data of the BOD and TN are not as good as those of the data of the NH_{3}-N and COD_{Mn}. Therefore, this paper selected the NH_{3}-N and COD_{Mn} as the research indicators.

## RESULTS AND DISCUSSION

### Multivariate joint distribution of Q, NH_{3}-N and COD_{Mn}

#### Establishment of the marginal distribution and joint probability distribution

Establishment of the marginal distribution. The parameters of the normal, lognormal, gamma and P-III distributions of Q, NH

_{3}-N and COD_{Mn}are estimated by the maximum likelihood method, as shown in Table 3. The results the of K-S and RMSE tests are given in Table 4, which shows that the lognormal distribution is the most appropriate for the Q, NH_{3}-N and COD_{Mn}indicators at Gegou Station on the Yi River.Establishment of the joint probability distribution. Combined with the sample empirical frequency, the K-S and RMSE results are calculated, as shown in Table 5. The results show that the trivariate joint distribution of the set (Q, NH

_{3}-N, COD_{Mn}) is in accordance with the Clayton copula function, and the bivariate joint distributions of the set (Q, NH_{3}-N) and the set (Q, COD_{Mn}) are in accordance with the G-H copula function and AMH copula function, respectively.

. | Q . | NH_{3}-N
. | COD_{Mn}
. |
---|---|---|---|

Normal | , | , | , |

Lognormal | , | , | , |

Gamma | , | , | , |

P-III | , , | , | , |

. | Q . | NH_{3}-N
. | COD_{Mn}
. |
---|---|---|---|

Normal | , | , | , |

Lognormal | , | , | , |

Gamma | , | , | , |

P-III | , , | , | , |

. | Normal . | Lognormal . | Gamma . | P-III . | ||||
---|---|---|---|---|---|---|---|---|

K-S . | RMSE . | K-S . | RMSE . | K-S . | RMSE . | K-S . | RMSE . | |

Q | 0.2828 | 0.1669 | 0.0617 | 0.0287 | 0.1453 | 0.0831 | 0.2426 | 0.1216 |

NH_{3}-N | 0.2900 | 0.1776 | 0.0427 | 0.0177 | 0.1344 | 0.0751 | 0.3751 | 0.1945 |

COD_{Mn} | 0.1565 | 0.0883 | 0.0479 | 0.0247 | 0.0717 | 0.0352 | 0.0771 | 0.0365 |

. | Normal . | Lognormal . | Gamma . | P-III . | ||||
---|---|---|---|---|---|---|---|---|

K-S . | RMSE . | K-S . | RMSE . | K-S . | RMSE . | K-S . | RMSE . | |

Q | 0.2828 | 0.1669 | 0.0617 | 0.0287 | 0.1453 | 0.0831 | 0.2426 | 0.1216 |

NH_{3}-N | 0.2900 | 0.1776 | 0.0427 | 0.0177 | 0.1344 | 0.0751 | 0.3751 | 0.1945 |

COD_{Mn} | 0.1565 | 0.0883 | 0.0479 | 0.0247 | 0.0717 | 0.0352 | 0.0771 | 0.0365 |

*Note:* The numbers in bold denote the minimum RMSE, and the K-S value is 0.10995 at a 5% significance level in this study.

. | Q, NH_{3}-N. | Q, COD_{Mn}. | Q, NH_{3}-N, COD_{Mn}. | ||||||
---|---|---|---|---|---|---|---|---|---|

θ
. | K-S . | RMSE . | θ
. | K-S . | RMSE . | θ
. | K-S . | RMSE . | |

Clayton | – | – | – | – | – | – | 0.8861 | 0.0865 | 0.0369 |

G-H | 0.8919 | 0.0551 | 0.0215 | 0.6239 | 0.0574 | 0.0220 | 0.9219 | 0.1658 | 0.0862 |

Frank | 1.6756 | 0.0832 | 0.0310 | 1.7545 | 0.0689 | 0.0236 | 0.3592 | 0.1537 | 0.7980 |

A-M-H | 0.8085 | 0.0702 | 0.0218 | 1.3839 | 0.0488 | 0.0214 | 0.4901 | 0.1235 | 0.8341 |

. | Q, NH_{3}-N. | Q, COD_{Mn}. | Q, NH_{3}-N, COD_{Mn}. | ||||||
---|---|---|---|---|---|---|---|---|---|

θ
. | K-S . | RMSE . | θ
. | K-S . | RMSE . | θ
. | K-S . | RMSE . | |

Clayton | – | – | – | – | – | – | 0.8861 | 0.0865 | 0.0369 |

G-H | 0.8919 | 0.0551 | 0.0215 | 0.6239 | 0.0574 | 0.0220 | 0.9219 | 0.1658 | 0.0862 |

Frank | 1.6756 | 0.0832 | 0.0310 | 1.7545 | 0.0689 | 0.0236 | 0.3592 | 0.1537 | 0.7980 |

A-M-H | 0.8085 | 0.0702 | 0.0218 | 1.3839 | 0.0488 | 0.0214 | 0.4901 | 0.1235 | 0.8341 |

In addition, sample empirical distributions are compared with the theoretical distributions of the sets (Q, NH_{3}-N), (Q, COD_{Mn}) and (Q, NH_{3}-N, COD_{Mn}), as shown in Figures 3 and 4, for which the correlation coefficients R^{2} are all relatively high. This result illustrates that the selected copulas can appropriately describe the multivariate probability distributions of Q, NH_{3}-N and COD_{Mn}.

#### Analysis of the joint probability distribution of the discharge and water quality

With the established multivariate joint probability distribution of Q, NH_{3}-N and COD_{Mn}, their corresponding contour plots and contour surfaces are shown in Figures 5 and 6, respectively; note that the coordinate axes of Figure 6 use a logarithmic coordinate. Moreover, some typical joint probabilities of Q, NH_{3}-N and COD_{Mn} are given in Table 5.

Figure 5 shows the contour plots of the sets (Q, NH_{3}-N) and (Q, COD_{Mn}). When the value of the bivariate joint probability distribution is between 0.1 and 0.4, the contour plot is more intensive than that between 0.6 and 0.9. This reflects that the degrees of change in the joint probabilities for both NH_{3}-N and COD_{Mn} present a decrease with an increase in the discharge.

Furthermore, as seen from Figure 6, when the value of the trivariate joint probability distribution is between 0.1 and 0.2 or between 0.7 and 0.9, the contour surface has a larger spacing, and Q, NH_{3}-N, COD_{Mn} each change more than that spacing value between 0.3 and 0.6, indicating that a trivariate joint probability from 30 to 60% is the most sensitive to the discharge. Combining Tables 6–8, the degree of increase in the probability does not always increase as the discharge increases.

NH_{3}-N (mg/L)
. | Discharge (m³/s) . | ||||
---|---|---|---|---|---|

10 . | 20 . | 40 . | 60 . | 80 . | |

0.5 | 0.2203 | 0.3302 | 0.4223 | 0.4608 | 0.48093 |

1.0 | 0.3302 | 0.4873 | 0.6152 | 0.6676 | 0.69470 |

1.5 | 0.3826 | 0.5606 | 0.7031 | 0.7607 | 0.79016 |

2.0 | 0.4116 | 0.6006 | 0.7504 | 0.8103 | 0.84070 |

NH_{3}-N (mg/L)
. | Discharge (m³/s) . | ||||
---|---|---|---|---|---|

10 . | 20 . | 40 . | 60 . | 80 . | |

0.5 | 0.2203 | 0.3302 | 0.4223 | 0.4608 | 0.48093 |

1.0 | 0.3302 | 0.4873 | 0.6152 | 0.6676 | 0.69470 |

1.5 | 0.3826 | 0.5606 | 0.7031 | 0.7607 | 0.79016 |

2.0 | 0.4116 | 0.6006 | 0.7504 | 0.8103 | 0.84070 |

COD_{Mn} (mg/L)
. | Discharge (m³/s) . | ||||
---|---|---|---|---|---|

10 . | 20 . | 40 . | 60 . | 80 . | |

4 | 0.1488 | 0.2436 | 0.3355 | 0.3767 | 0.39836 |

6 | 0.2594 | 0.4063 | 0.6947 | 0.0000 | 0.62116 |

10 | 0.3898 | 0.5812 | 0.7372 | 0.7992 | 0.83020 |

15 | 0.4473 | 0.6532 | 0.8147 | 0.8774 | 0.90846 |

COD_{Mn} (mg/L)
. | Discharge (m³/s) . | ||||
---|---|---|---|---|---|

10 . | 20 . | 40 . | 60 . | 80 . | |

4 | 0.1488 | 0.2436 | 0.3355 | 0.3767 | 0.39836 |

6 | 0.2594 | 0.4063 | 0.6947 | 0.0000 | 0.62116 |

10 | 0.3898 | 0.5812 | 0.7372 | 0.7992 | 0.83020 |

15 | 0.4473 | 0.6532 | 0.8147 | 0.8774 | 0.90846 |

NH_{3}-N (mg/L)
. | COD_{Mn} (mg/L)
. | Q (m^{3}/s). | ||||
---|---|---|---|---|---|---|

10 . | 20 . | 40 . | 60 . | 80 . | ||

0.5 | 4 | 0.2233 | 0.2657 | 0.2898 | 0.2978 | 0.3014 |

6 | 0.2764 | 0.3423 | 0.3821 | 0.3955 | 0.4018 | |

10 | 0.3104 | 0.3944 | 0.4471 | 0.4653 | 0.4738 | |

15 | 0.3206 | 0.4106 | 0.4677 | 0.4875 | 0.4968 | |

1.0 | 4 | 0.2609 | 0.3192 | 0.3539 | 0.3655 | 0.3709 |

6 | 0.3345 | 0.4329 | 0.4965 | 0.5186 | 0.5291 | |

10 | 0.3843 | 0.5168 | 0.6077 | 0.6405 | 0.6562 | |

15 | 0.3997 | 0.5440 | 0.6450 | 0.6817 | 0.6994 | |

1.5 | 4 | 0.2735 | 0.3378 | 0.3766 | 0.3897 | 0.3958 |

6 | 0.3549 | 0.4665 | 0.5404 | 0.5665 | 0.5790 | |

10 | 0.4109 | 0.5644 | 0.6732 | 0.7131 | 0.7324 | |

15 | 0.4284 | 0.5966 | 0.7186 | 0.7639 | 0.7858 | |

2.0 | 4 | 0.2793 | 0.3466 | 0.3874 | 0.4012 | 0.4076 |

6 | 0.3645 | 0.4828 | 0.5620 | 0.5902 | 0.6036 | |

10 | 0.4237 | 0.5878 | 0.7061 | 0.7499 | 0.7711 | |

15 | 0.4423 | 0.6227 | 0.7559 | 0.8059 | 0.8302 |

NH_{3}-N (mg/L)
. | COD_{Mn} (mg/L)
. | Q (m^{3}/s). | ||||
---|---|---|---|---|---|---|

10 . | 20 . | 40 . | 60 . | 80 . | ||

0.5 | 4 | 0.2233 | 0.2657 | 0.2898 | 0.2978 | 0.3014 |

6 | 0.2764 | 0.3423 | 0.3821 | 0.3955 | 0.4018 | |

10 | 0.3104 | 0.3944 | 0.4471 | 0.4653 | 0.4738 | |

15 | 0.3206 | 0.4106 | 0.4677 | 0.4875 | 0.4968 | |

1.0 | 4 | 0.2609 | 0.3192 | 0.3539 | 0.3655 | 0.3709 |

6 | 0.3345 | 0.4329 | 0.4965 | 0.5186 | 0.5291 | |

10 | 0.3843 | 0.5168 | 0.6077 | 0.6405 | 0.6562 | |

15 | 0.3997 | 0.5440 | 0.6450 | 0.6817 | 0.6994 | |

1.5 | 4 | 0.2735 | 0.3378 | 0.3766 | 0.3897 | 0.3958 |

6 | 0.3549 | 0.4665 | 0.5404 | 0.5665 | 0.5790 | |

10 | 0.4109 | 0.5644 | 0.6732 | 0.7131 | 0.7324 | |

15 | 0.4284 | 0.5966 | 0.7186 | 0.7639 | 0.7858 | |

2.0 | 4 | 0.2793 | 0.3466 | 0.3874 | 0.4012 | 0.4076 |

6 | 0.3645 | 0.4828 | 0.5620 | 0.5902 | 0.6036 | |

10 | 0.4237 | 0.5878 | 0.7061 | 0.7499 | 0.7711 | |

15 | 0.4423 | 0.6227 | 0.7559 | 0.8059 | 0.8302 |

### Application analysis of the joint improvement of the water quality based on the relationship curve

#### Construction of the relationship curve among Q, NH_{3}-N and COD_{Mn}

Based on the improvement probability of an individual water quality indicator, the curves of recovering NH_{3}-N and COD_{Mn} to grades of III, IV and V are constructed separately, and the fitting function of each curve is computed by MATLAB, as shown in Figure 7. For the joint improvement probability of multivariate water quality indicators, the curve of NH_{3}-N and COD_{Mn} simultaneously recovered to grade III is denoted as (III, III), that of NH_{3}-N recovering to grade III while COD_{Mn} recovers to grade IV is denoted as (III, IV), and that of NH_{3}-N recovering to grade IV while COD_{Mn} recovers to grade III is denoted as (IV, III). These curves are each constructed individually, thereby representing equal and non-equal weights for the recovery of the set (NH_{3}-N, COD_{Mn}), and the fitting function of each curve is computed by MATLAB, as shown in Figure 8.

#### Analysis of the improvement probability and the optimal discharge

According to the bivariate and trivariate joint probability curves, the typical improvement probability values and the corresponding discharges of each individual water quality indicator and of the multiple water quality indicators are given in Tables 9–11.

NH_{3}-N
. | 40.0% . | 45.0% . | 50.0% . | 55.0% . | 60.0% . | 65.0% . | 70.0% . | 75.0% . | 80.0% . |
---|---|---|---|---|---|---|---|---|---|

III | 13.50 | 16.30 | 20.00 | 24.95 | 33.15 | 50.50 | 98.80 | 175.00 | – |

IV | 11.00 | 13.05 | 15.65 | 18.80 | 22.85 | 28.4 | 38.15 | 59.60 | 115.00 |

V | 9.77 | 11.65 | 13.70 | 16.00 | 19.30 | 23.20 | 28.60 | 38.10 | 59.60 |

NH_{3}-N
. | 40.0% . | 45.0% . | 50.0% . | 55.0% . | 60.0% . | 65.0% . | 70.0% . | 75.0% . | 80.0% . |
---|---|---|---|---|---|---|---|---|---|

III | 13.50 | 16.30 | 20.00 | 24.95 | 33.15 | 50.50 | 98.80 | 175.00 | – |

IV | 11.00 | 13.05 | 15.65 | 18.80 | 22.85 | 28.4 | 38.15 | 59.60 | 115.00 |

V | 9.77 | 11.65 | 13.70 | 16.00 | 19.30 | 23.20 | 28.60 | 38.10 | 59.60 |

COD_{Mn}
. | 40.0% . | 45.0% . | 50.0% . | 55.0% . | 60.0% . | 65.0% . | 70.0% . | 75.0% . | 80.0% . |
---|---|---|---|---|---|---|---|---|---|

III | 19.00 | 23.40 | 30.00 | 39.40 | 59.70 | 125.00 | – | – | – |

IV | 10.95 | 13.15 | 14.96 | 18.25 | 12.50 | 26.65 | 33.10 | 46.65 | 76.60 |

V | 9.78 | 11.65 | 13.60 | 16.10 | 18.80 | 22.80 | 26.65 | 35.50 | 48.70 |

COD_{Mn}
. | 40.0% . | 45.0% . | 50.0% . | 55.0% . | 60.0% . | 65.0% . | 70.0% . | 75.0% . | 80.0% . |
---|---|---|---|---|---|---|---|---|---|

III | 19.00 | 23.40 | 30.00 | 39.40 | 59.70 | 125.00 | – | – | – |

IV | 10.95 | 13.15 | 14.96 | 18.25 | 12.50 | 26.65 | 33.10 | 46.65 | 76.60 |

V | 9.78 | 11.65 | 13.60 | 16.10 | 18.80 | 22.80 | 26.65 | 35.50 | 48.70 |

(NH_{3}-N, COD_{Mn})
. | 40.0% . | 45.0% . | 50.0% . | 55.0% . | 60.0% . | 65.0% . | 70.0% . | 75.0% . | 80.0% . |
---|---|---|---|---|---|---|---|---|---|

(III, III) | 15.00 | 21.90 | 38.90 | 138.00 | – | – | – | – | – |

(IV, IV) | 9.80 | 12.00 | 15.10 | 19.20 | 24.40 | 35.80 | 64.50 | – | – |

(V, V) | 8.90 | 10.80 | 13.00 | 16.20 | 20.10 | 25.40 | 28.60 | 52.60 | 108.10 |

(III, IV) | 10.75 | 13.40 | 16.80 | 22.30 | 34.03 | 93.10 | 160.00 | – | – |

(III, V) | 10.10 | 12.30 | 14.90 | 19.20 | 25.30 | 39.80 | 100.60 | 170.00 | – |

(IV, V) | 9.20 | 11.00 | 13.10 | 15.80 | 19.30 | 24.50 | 33.15 | 59.60 | 131.50 |

(IV, III) | 12.70 | 17.25 | 24.75 | 42.90 | 138.20 | – | – | – | – |

(NH_{3}-N, COD_{Mn})
. | 40.0% . | 45.0% . | 50.0% . | 55.0% . | 60.0% . | 65.0% . | 70.0% . | 75.0% . | 80.0% . |
---|---|---|---|---|---|---|---|---|---|

(III, III) | 15.00 | 21.90 | 38.90 | 138.00 | – | – | – | – | – |

(IV, IV) | 9.80 | 12.00 | 15.10 | 19.20 | 24.40 | 35.80 | 64.50 | – | – |

(V, V) | 8.90 | 10.80 | 13.00 | 16.20 | 20.10 | 25.40 | 28.60 | 52.60 | 108.10 |

(III, IV) | 10.75 | 13.40 | 16.80 | 22.30 | 34.03 | 93.10 | 160.00 | – | – |

(III, V) | 10.10 | 12.30 | 14.90 | 19.20 | 25.30 | 39.80 | 100.60 | 170.00 | – |

(IV, V) | 9.20 | 11.00 | 13.10 | 15.80 | 19.30 | 24.50 | 33.15 | 59.60 | 131.50 |

(IV, III) | 12.70 | 17.25 | 24.75 | 42.90 | 138.20 | – | – | – | – |

Taking the function (Equation (9)) with the curvature method and slope method (Equations (5) and (6), respectively), the optimal discharge can be obtained. Accordingly, the optimal discharges under the conditions in which the set (NH_{3}-N, COD_{Mn}) is improved to grades of (III, III), (IV, IV), (V, V), (III, IV), (III, V), (IV, V) and (IV, III) are separately shown in Table 12.

Indicators . | Recovery grade . | Optimal discharge (m³/s) . | Corresponding probability (Curvature) . | |
---|---|---|---|---|

Curvature . | Slope . | |||

NH_{3}-N | III | 25.40 | 22.76 | 0.5413 |

IV | 27.56 | 23.50 | 0.6332 | |

V | 28.30 | 24.15 | 0.6826 | |

COD_{Mn} | III | 27.90 | 22.32 | 0.4825 |

IV | 29.75 | 24.55 | 0.6667 | |

V | 32.30 | 24.76 | 0.7241 | |

(NH_{3}-N, COD_{Mn}) | (III, III) | 20.45 | 17.65 | 0.4393 |

(IV, IV) | 23.80 | 20.03 | 0.5902 | |

(V, V) | 26.74 | 21.75 | 0.6454 | |

(III, IV) | 21.78 | 19.14 | 0.5342 | |

(III, V) | 22.35 | 19.65 | 0.5668 | |

(IV, V) | 24.18 | 21.88 | 0.6352 | |

(IV, III) | 20.81 | 16.60 | 0.4671 |

Indicators . | Recovery grade . | Optimal discharge (m³/s) . | Corresponding probability (Curvature) . | |
---|---|---|---|---|

Curvature . | Slope . | |||

NH_{3}-N | III | 25.40 | 22.76 | 0.5413 |

IV | 27.56 | 23.50 | 0.6332 | |

V | 28.30 | 24.15 | 0.6826 | |

COD_{Mn} | III | 27.90 | 22.32 | 0.4825 |

IV | 29.75 | 24.55 | 0.6667 | |

V | 32.30 | 24.76 | 0.7241 | |

(NH_{3}-N, COD_{Mn}) | (III, III) | 20.45 | 17.65 | 0.4393 |

(IV, IV) | 23.80 | 20.03 | 0.5902 | |

(V, V) | 26.74 | 21.75 | 0.6454 | |

(III, IV) | 21.78 | 19.14 | 0.5342 | |

(III, V) | 22.35 | 19.65 | 0.5668 | |

(IV, V) | 24.18 | 21.88 | 0.6352 | |

(IV, III) | 20.81 | 16.60 | 0.4671 |

Tables 11 and 12 give the joint improvement probabilities of the water quality and its corresponding discharge under various recovery scenarios of the water quality, providing a variety of choices for managers to make water dispatch schemes. Those managers can therefore choose an appropriate discharge according to the storage capacity of a reservoir and the pollution of a river.

### Discussion

Bivariate and trivariate joint probability distributions have been constructed based on copula functions, and the typical joint probability values of Q, NH_{3}-N and COD_{Mn} have been given. Under the condition of NH_{3}-N listed in Table 6, the occurrence probability is 0.6152 when Q is less than 40 m^{3}/s and NH_{3}-N is less than 1.0 mg/L. Under the condition of COD_{Mn} in Table 7, the occurrence probability is 0.6947 when Q is less than 40 m^{3}/s and COD_{Mn} is less than 6.0 mg/L. In contrast, under the conditions of NH_{3}-N and COD_{Mn} in Table 8, the occurrence probability is 0.4965 when Q is less than 40 m^{3}/s, NH_{3}-N is less than 1.0 mg/L and COD_{Mn} is less than 6.0 mg/L. These results show that the trivariate joint probability of the set (Q, NH_{3}-N, COD_{Mn}) is evidently less than the bivariate joint probability of the sets (Q, NH_{3}-N) and (Q, COD_{Mn}). Therefore, if only the bivariate joint probability distributions are considered, it is difficult to reflect the correct statistical characteristics of the discharge and water quality. On the contrary, the trivariate joint probability distribution comprehensively considers Q, NH_{3}-N and COD_{Mn} and can more reasonably illustrate the joint probabilities of different combinations of the discharge and water quality quantitatively.

Based on the results of the improvement probabilities of Q, NH_{3}-N and COD_{Mn} in Tables 9–11, respectively, additional discharge is evidently needed when the improvement probability is increased at a rate of 5%. Moreover, the calculation of an individual indicator can obtain only the improvement probability of that indicator while ignoring the improvement probability of other indicators; consequently, this approach cannot reflect the characteristics of the overall improvement of the main pollutants in a given water environment. For instance, according to the relationship between the discharge and a single indicator, under the condition in which the discharge is 20 m^{3}/s, the improvement probability of only NH_{3}-N recovering to grade III is 50%, which reflects only the improvement probability in which only NH_{3}-N exists in the water environment, and it does not simultaneously reflect the improvement of coexisting pollution indicators and their interactions. However, the joint improvement probability of multiple indicators is analysed through the statistical characteristics of the indicators and the discharge, which can help predict the joint improvement of several important indicators throughout the water environment.

Figures 7 and 8 and Table 12 show that the trivariate improvement probability is smaller than the bivariate probability despite the fact that the optimal discharge for improving multiple water quality indicators is less than that of a single indicator. The main reason is that there are certain synergistic and antagonistic effects among the pollution factors in the water system; as a result, the joint improvement of multiple water quality indicators is more difficult to achieve than the improvement of a single indicator. In general, multiple pollution factors coexist within a river, and thus, the joint improvement of multiple water quality indicators can reflect the relationships between the discharge and multiple indicators more reasonably.

Meanwhile, Table 12 illustrates that the optimal discharge calculated by the curvature method is higher than that calculated by the slope method. Therefore, the value calculated by the curvature method is adopted for the sake of safety.

## CONCLUSIONS

In this study, the trivariate joint probability distribution of Q, NH_{3}-N and COD_{Mn} is established to indicate the occurrence probability of water quality indicators under different discharge conditions. Then, based on the trivariate joint probability distribution, a method for calculating the optimal discharge for the joint improvement of water quality indicators is proposed. The main conclusions are as follows:

Because the trivariate joint probability distribution of the set (Q, NH

_{3}-N, COD_{Mn}) can better reflect the improvement of the water quality with multiple coexisting pollutants, this approach is more effective and comprehensive than the bivariate joint probability distribution of the sets (Q, NH_{3}-N) and (Q, COD_{Mn}) in estimating the combinations of water quality indicators under different discharge conditions.Based on the trivariate joint probability distribution, the existence of a critical point along the curve is proven. Then, the joint improvement probability and the optimal discharges of multiple water quality indicators under different discharge conditions are calculated.

The methods that can be employed to fit the relationship curve of and identify the critical point on the curve are flexible and varied. Therefore, the selection of the proper method needs to be based on the specific situation of the river. Meanwhile, due to the limited availability of data, only a few water quality indicators are selected (i.e. only NH

_{3}-N and COD_{Mn}). Therefore, in the future, the monitoring and accumulation of water quality indicator data should be strengthened to better reveal the relationships between the discharge and multiple water quality indicators.

## ACKNOWLEDGEMENTS

This study was funded by the Non-profit Industry Financial Program of MWR: Key Technology Research and Demonstration Project of Water Ecological Civilization Construction (No. 201401003).

## REFERENCES

*Environmental Quality Standards for Surface Water GB3838-*2002