Vous êtes sur la page 1sur 14

Nat Hazards (2015) 75:13891402

DOI 10.1007/s11069-014-1385-z
ORIGINAL PAPER

Determination of the distribution of flood forecasting


error
Junhong Zhang Lu Chen Vijay P. Singh Hongwen Cao
Dangwei Wang

Received: 4 March 2014 / Accepted: 9 August 2014 / Published online: 11 September 2014
Springer Science+Business Media Dordrecht 2014

Abstract Flood forecasting plays an essential role in enhancing the safety of residents
downstream and preventing or reducing economic losses. One critical issue in flood risk
assessment is the determination of the probability distribution of forecast errors. Several
investigations, which have been carried out to analyze the influence of the uncertainty in
real-time operation or water resources management, assumed that the relative forecast error
was approximately normally distributed. This study investigates whether the flood forecast
error follows the normal distribution. Several distributions were fitted to the flood error
series, and their performances were analyzed using the data from Three Gorges Reservoir
(TGR) and Muma River. Then, the most appropriate distribution was selected. Results
show that the assumption of normal distribution is not justified for the flood forecast error
series of TGR and Muma River. The use of normal distribution for estimating flood risk
may lead to incorrect results.
Keywords Flood forecasting error  Probability distribution function  Three Gorges
Reservoir  Muma River

J. Zhang  H. Cao  D. Wang


State Key Laboratory of Simulation and Regulation of River Basin Water Cycle, China Institute of
Water Resources and Hydropower Research, Beijing 100038, China
e-mail: zjh411891611@163.com
J. Zhang
College of Chemistry and Materials Science, Wuhan 430074, China
L. Chen (&)
College of Hydropower and Information Engineering, Huazhong University of Science and
Technology, Wuhan 430074, China
e-mail: chl8505@126.com
L. Chen  V. P. Singh
Department of Biological and Agricultural Engineering and Zachry Department of Civil Engineering,
Texas A&M University (TAMU), College Station, TX 77843-2117, USA

123

1390

Nat Hazards (2015) 75:13891402

1 Introduction
Flood forecasts, which are important for predicting extreme hydrologic events, are used
to guide management decisions on water resource systems (Alemu et al. 2011; Boucher
et al. 2012). When using stream flow forecasts in real-time reservoir operation models
and water resources management, the uncertainty involved in stream flow forecast
products is an inherent and important characteristic of stream flow forecasting. And those
uncertainties have been identified as a major influencing factor of the value of the
forecast.
One critical issue in reservoir risk assessment is the determination of the probability
distribution of inflow forecast errors or uncertainties. Several investigations, which have
been carried out to analyze the influence of uncertainty on real-time operation or water
resources management, assumed that the relative forecast error at each time is
approximately normally distributed (Christensen 2003; Stedinger et al. 2008). For
example, Li et al. (2010) determined the flood limited water level by considering the
influence of inflow forecast uncertainty following a normal distribution. Zhao et al.
(2011) analyzed the effect of stream flow forecast uncertainty on real-time reservoir
operation by assuming a normal distribution for forecast uncertainty. Yan et al. (2013)
quantified inflow forecast errors and their impact on reservoir flood control operations
by assuming the inflow forecast errors to be normally distributed. The Chinese design
flood guidelines also assumed the distribution of flood forecast errors to be a normal
distribution (MWR 2006).
However, whether the flood error follows a normal distribution is still to be
investigated. In reality, the forecast error or uncertainty does not follow the normal
distribution all the time. For example, Zhao et al. (2013) selected the Three Gorges
Reservoir as a case study and indicated that the real-world forecast uncertainty was
non-Gaussian and heavy tailed. Diao et al. (2007) developed a model based on the
principle of maximum entropy to investigate the distribution of relative flood forecast
errors for some typical reservoirs in China. Zhao et al. (2013) indicate that stream flow
at the coming time t corresponds to multiple prediction made at preceding time, and
forecast uncertainty at the coming time t generally decreases over time, as the previous
forecast is replaced by the new forecast with more hydrologic information. And the
flood forecast error may follow a different distribution, when the flood forecast lead
time increases. Therefore, it may be difficult to justify that the forecast error obeys the
normal distribution all the time. Although forecast uncertainty analysis has been one
research focus in hydrology, there are comparatively less studies on the distribution of
the uncertainty.
The objective of this study was therefore to determine the distribution of the
uncertainty and investigate whether this kind of uncertainty or flood forecast error
obeys the normal distribution. Using different hydrological models or the same model
with different flood forecast lead times, flood relative errors are determined. Several
distributions, which are widely used in hydrology, are selected to fit the flood relative
error series. The performance measurements coupled with hypothesis testing is applied
to justify the use of each distribution. The Three Gorges Reservoir basin and the Muma
River, which is a tributary of Han River, are selected as case studies.

123

Nat Hazards (2015) 75:13891402

1391

2 Methodology
2.1 Flood forecasting models
Rainfall-runoff (RR) models have become a basic tool for flood forecasting and have
become increasingly important for catchment management (Xu et al. 2013). Three conceptual rainfall-runoff models, namely Xinanjiang model, HBV model and Tank model,
were used in this study. Detailed information on these models is given below.
2.1.1 Xinanjiang model
The Xinanjiang model was developed by Zhao et al. (1980) for forecasting flows in the
Xinanjiang reservoir. The model has been widely applied for flood forecasting in large basins
in humid and semi-humid regions of China (Zhao 1992). Since the TGR is located in the south
of China, the Xinanjiang model was used. Its main feature is the concept of runoff formation
on repletion of storage, which denotes that runoff is not produced until the soil moisture
content of the aeration zone reaches field capacity (Zhao 1992). The Xinanjiang model
includes two components, namely runoff generating and runoff routing. It has 17 parameters
that include seven runoff generating component parameters and 10 runoff routing parameters.
These parameters are abstract conceptual representations of non-measurable watershed
characteristics that have to be calibrated by an optimization method. Detailed information
about the meaning and range of parameters can be found in Xu et al. (2013).
2.1.2 Tank model
Tank Model is a synthetic flow model, which was developed and introduced in 1956 by
Sugawara, and has been widely used in many Asian countries, such as Japan and China, to
forecast floods and manage water resources (Ngoc et al. 2013). The Tank model used in
this study has a simple structure with 4 tanks, a surface tank (A), an intermediate tank (B),
a subbase tank (C), and a base tank (D), laid vertically in a series. Each tank has a vertical
outlet at the bottom (except tank D) and one horizontal outlet at the side (except tank A,
which typically has 2 horizontal outlets).
2.1.3 HBV model
The Hydrologiska Byrans Vattenbalans avdelning (HBV) model is a conceptual model of
catchment hydrology, which simulates discharge using rainfall, temperature and estimates
of potential evaporation (Bergstrom 1992;Grillakis et al. 2010). The HBV hydrological
model has a long history, and the model has been applied in more than 30 countries
(Lindstrom 1997). The HBV model has been applied in numerous studies, e.g., hydrological forecasting and computing design floods and climate change studies. (Grillakis
et al. 2010). A description of the model can be found in Bergstrom (1992) and Harlin and
Kung (1992).
2.2 Probability distribution functions for flood forecast error
In order to the select appropriate distributions for fitting flood forecast error series, the
flood forecast error should be calculated first and can be defined as

123

1392

Nat Hazards (2015) 75:13891402

res;t

Qs;t  qt es;t

qt
qt

where res,t is the relative stream flow forecast error, whose stream flow is predicted at time
s for time t. qt denotes the observed flow at time t. Qs,t denotes the flow predicted at time
s for time t (s B t).
There are several probability distributions for modeling hydrological variables. For
example, the Pearson Type III has been recommended by the Chinese Ministry of Water
Resources (MWR 2006) as a uniform procedure for flood frequency analysis in China
(Chen et al. 2010).Griffiths (1989) used the Wakeby distribution to fit the annual maximum
floods. The exponential distribution was selected for fitting drought duration (Zelenhastic
and Salvai 1987; Chen et al. 2013). The following distributions were used for fitting the
flood forecast error data.
1.

Normal distribution
Fx

Zx
0

xl2
1
p e 2r2 dx
r 2p

where x is the forecast error value; l is the mean or expectation of the distribution; and r is
its standard deviation.
2.

Pearson type three distribution (P-III)


Zx a
b x  ua1 bxu
e
dx
Fx
Ca

where a is the shape parameter; b is the scale parameter; and u location parameter, u [ 0.
3.

Generalized extreme distribution (GEV)


x  u 1=n

g
Fx expf1 n
r

where u is the location parameter; r is the scale parameter; and n is the shape parameter.
4.

Generalized Pareto distribution (GP)


Fx 1  1  k

x  u 1=k

r

where k is the shape parameter; r is the scale parameter; and u is the location parameter.
5.

Exponential distribution (EX)


Fx 1  exp

xu

where u, r are the shape and scale parameters, respectively.


6.

Gumbel distribution
Fx expeaxl 

where r is the scale parameter; and u is the location parameter.

123

Nat Hazards (2015) 75:13891402

1393

Fig. 1 Sketch of the TGR intervening basin

7.

Wakeby distribution
a
c
Fx f 1  1  xb  1  1  xd
b
d

where a, b, c, d, and f are parameters.


2.3 Performance measures
The following criteria were used to estimate the goodness of fit of the above-mentioned
distributions.
1.

Bias
Bias

n 

1X
Pthe i  Pemp i =Pemp i
N i1

where Pthe is the theoretical probability, Pthe(i) = F(xi). Pemp is the empirical probabilities
and can be computed by the Gringorten plotting-position formula:

123

1394

Nat Hazards (2015) 75:13891402

Fig. 2 the relative error series with 1-, 2-and 3-day lead time in the TGR

pemp i

i  0:44
n 0:12

10

where Pemp(i) is the cumulative empirical frequency, the probability that a given value is
less than the ith smallest observation in the data set of n observations.
2.

Root mean square error (RMSE)


1
RMSE
n

s


n
X
Pthe i  Pemp i 2


Pemp
i1

The best fitting distribution was identified through the goodness of fit tests.

123

11

Nat Hazards (2015) 75:13891402

1395

Fig. 3 Flood forecast results of Muma River using different hydrological models

3 Study area and data


The Yangtze River, which is the longest river in Asia and the third longest river in the
world, was selected as a case study. It is about 6,300 km long and flows from its source in
Qinghai Province eastward to the East China Sea at Shanghai. This study mainly focused
on the upper Yangtze River, which is 4,529 km long, up to 3/4 of the whole length of
Yangtze River, with a drainage area of 1,006,000 km2, accounting for 55.6 % of the
watershed area. Floods of the Yangtze River (Chang Jiang) in central and eastern China
have occurred periodically and have often caused considerable destruction of property and
loss of life. Among the most major flood events are those of 1870, 1931, 1954, 1998, and
2010 (Chen et al. 2012). For example, in 1998, the entire Yangtze River basin suffered
from tremendous floodingthe largest flood since 1954, which led to the economic loss of
166 billion Chinese Yuan (nearly US$ 40 billion) (Yin and Li 2001).Hence, flood forecasting of the upper Yangtze River is important for flood prevention and disaster relief.
The inflow of TGR consists of three components, the main upstream inflow, the tributary inflow from the Wu River, and the rainfall runoff from the TGR intervening basin, as

123

1396
Table 1 Parameters of the
selected distributions for fitting
the flood error series of TGR

Nat Hazards (2015) 75:13891402

Order

Distributions

Parameters
1-day

2-day

3-day

Normal

l = -0.45
r = 4.52

l = -0.92
r = 6.32

l = -1.81
r = 8.58

P-III

a = 16.00
b = -0.63
l = 24.56

a = 34.60
b = -1.30
l = 26.26

a = 5.54
b = -0.27
l = 18.85

GEV

l = -1.81
r = 4.73
f = 0.39

l = -2.68
r = 6.73
f = 0.43

l = -3.7
r = 9.46
f = 0.55

GP

l = -871
r = 1,849
k = 1.24

l = -1,287
r = 2,812
k = 1.35

l = -1,949
r = 4,693
k = 1.65

Ex

l = -5.55
a = 5.10

l = -8.05
a = 7.13

l = -11.49
a = 9.68

Gumbel

l = -2.58
a = 3.68

l = -3.88
a = 5.14

l = -5.84
a = 6.99

Wakeby

a = 166.9
b = 11.60
c = 2.91
d = 5.28
f = -16.76

a = 193.09
b = 10.05
c = 4.25
d = 1.22
f = -22.59

a = 229
b = 8.75
c = 6.21
d = 0.11
f = -30.9

shown in Fig. 1. The intervening basin has a catchment area of 55,907 km2, about 5.6 % of
the TGR upstream Yangtze River basin. Over the whole intervening basin, there are 40
rainfall gauge stations and 2 hydrological stations (Cuntan and Wulong), which control the
upstream inflow and tributary inflow, respectively (Li et al. 2010). The inflow of the TGR
was the total flow of Cuntan, Wulong, and the intervening basin.
The data set for TGR inflow forecasting includes daily runoff data of the Cuntan and
Wulong hydrological stations and the arithmetic mean of observed rainfall data in the
intervening basin from 2003 to 2006. Since the Xinanjiang model was tested to be the most
appropriate model for predicting the flow of the intervening basin of TGR and has been
applied in the real-time flood forecasting system, this model was used to simulate the
rainfall-runoff process in the intervening basin with different lead time. The data from June
1, 2003, to December 31, 2006, were used for flood forecasting with 1-, 2- and 3-day lead
times. The relative errors of 1-, 2- and 3-day lead times are shown in Fig. 2. It can be seen
that the relative error increases with increasing lead time.
The Han River is the source of water for the middle route of the well-known South-toNorth Water Diversion Project in China. The basin has a subtropical monsoon climate, and
the whole Han basin is divided into three regions: the Danjiangkou reservoir sub-basin
(upper sub-basin), the middle sub-basin, and the lower sub-basin (Chen et al. 2007). The
Muma River, which is a tributary of Han River and is located in the upper Han River basin,
was used as a case study. The daily data during the flood season (June to September) from
the year 1980 to 1990 were employed in this study.
Three conceptual rainfall-runoff models were used in this study to predict the daily flow
of Muma River. The three chosen rainfall-runoff models were Xinanjiang, Tank, and HBV
models. The forecast results of the three models are shown in Fig. 3.

123

Nat Hazards (2015) 75:13891402


Table 2 Results of calculated
bias and RMSE for 1-, 2- and
3-day error series of TGR

1397

Distributions

1 day
Bias

Normal

2-day
RMSE

Bias

3-day
RMSE

Bias

RMSE

0.009

0.049

0.012

0.045

0.021

P-III

-0.003

0.047

-0.004

0.039

-0.005

0.051
0.032

GEV

-0.003

0.052

-0.004

0.045

-0.006

0.038

GP

-0.005

0.081

-0.009

0.075

-0.013

0.072

Ex

0.074

0.124

0.074

0.125

0.079

0.135

Gumbel

0.044

0.079

0.046

0.080

0.053

0.091

Wakeby

0.001

0.014

0.000

0.016

0.000

0.015

Table 3 Results of KS test for TGR


Distributions

1-day

2-day

3-day

p value

Selection

p value

Selection

p value

Selection

Normal

0.000

0.000

0.000

P-III

0.000

0.002

0.022

GEV

0.000

0.001

0.004

GP

0.000

0.000

0.000

Ex

0.000

0.000

0.000

Gumbel

0.000

0.000

0.000

Wakeby

0.008

0.216

0.640

4 Application
4.1 Fitting results for TGR
The Wakeby, exponential, Gumbel, Pearson type three, generalized Pareto, and generalized extreme distribution were used to fit the 1-, 2-, and 3-day relative flood forecast errors
for TGR. The L-moment method was employed to estimate the parameters of these distributions. The estimated parameters are listed in Table 1. The bias and RMSE of the
observed and theoretical flow forecast errors are given in Table 2. The Kolmogorov
Smirnov (KS) test was employed for the goodness of fit test. Their p values were calculated, as given in Table 3. If the p values were small, the KS test was rejected. In other
words, the null hypothesis (H0) was accepted for all the p values greater than a. The fixed
values of a (0.01, 0.05, and 0.1) are generally used to evaluate the null hypothesis (H0) at
various significance levels. A value of 0.05 is typically used for most applications. In this
study, a was equal to 0.05. Symbol H in Table 3 means that this distribution is accepted,
and symbol 9 means that the distribution is rejected. Results indicate that the Wakeby
distribution is the only one that could not be rejected at the 5 % significance level. The
marginal distributions and frequency histograms of the flood forecast errors for TGR fitted
by Wakeby and normal distributions are shown in Fig. 4. In the left figure of Fig. 4, the
line represents the fitted distribution and crosses the empirical frequencies of observations.
Figure 4 indicates that the Wakeby distribution fitted the data well and performs better fit

123

1398

Nat Hazards (2015) 75:13891402

Fig. 4 Marginal distributions and histograms of the relative flood forecast error fitted by GEV and normal
distributions for TGR

123

Nat Hazards (2015) 75:13891402


Table 4 Results of calculated
bias and RMSE for Xinanjiang,
HBV, and Tank models of Muma
River

1399

Distributions

Normal

Xinanjiang

HBV

Bias

RMSE

Bias

Tank
RMSE

Bias

RMSE

-0.058

0.097

-0.057

0.098

-0.058

0.099

P-III

0.007

0.048

0.012

0.047

0.008

0.041

GEV

0.003

0.021

0.004

0.017

0.002

0.013

GP

0.005

0.041

0.010

0.041

0.006

0.032

Ex

0.001

0.040

0.006

0.042

0.001

0.033

Gumbel

-0.027

0.054

-0.026

0.056

-0.028

0.056

Wakeby

-0.073

0.258

-0.104

0.265

0.184

0.315

Table 5 Results of KS test for Muma River


Distributions

Xinanjiang

HBV

Tank

p value

Selection

p value

Selection

p value

Selection

Normal

0.000

0.000

0.000

P-III

0.000

0.000

0.000

GEV

0.287

0.357

0.590

GP

0.064

0.047

0.149

Ex

0.010

0.000

0.030

Gumbel

0.000

0.000

0.000

Wakeby

0.000

0.000

0.000

than the normal distribution. Based on this analysis, the assumption of normal distribution
may not be justified in some cases. Therefore, it is necessary to carry out the hypothesis test
first to verify whether this assumption is true or not when using it.
4.2 Fitting results of Muma River
Three conceptual RR models, namely Xinanjiang, HBV, and Tank models, were used in
this study to predict daily flow of the Muma River. The forecast lead time in Muma River
is 1 day for all the hydrological models. The Wakeby, exponential, Gumbel, Pearson type
three, generalized Pareto, and generalized extreme distribution were used to fit the relative
flood forecast errors of the three models. The L-moment method was employed to estimate
the parameters of these distributions. The bias and RMSE of the observed and theoretical
probability are given in Table 4. The KS test was employed for the goodness of fit test.
Their p values were calculated and given in Table 5. Results indicate that the GEV distribution gives a better fit and can be accepted for all the three cases at the 5 % significance
level. The marginal distributions and frequency histograms of flood forecast errors for
Muma River fitted by Wakeby and normal distributions are shown in Fig. 5. Figure 5
indicates that the selected distribution fits the data well and performs better than the normal
distribution. From this analysis, the assumption of normal distribution is not justified for
the Muma River. Therefore, it is necessary to carry out the hypothesis test first when using
it.

123

1400

Nat Hazards (2015) 75:13891402

Fig. 5 Marginal distributions and histograms of the relative flood forecast error fitted by GEV and normal
distributions for Muma River

123

Nat Hazards (2015) 75:13891402

1401

5 Discussion and conclusion


As far as flood forecasts are concerned, one critical issue in the reservoir risk assessment is
the determination of probability distributions of inflow forecast errors. Several studies,
which have been carried out to analyze the influence of uncertainty on real-time reservoir
operation or water resources management, assumed that the relative forecast error was
approximately normally distributed. However, whether the flood error follows the normal
distribution needs to be investigated. This study uses several distributions to fit the flood
error series with different lead time or calculated by different hydrological models. The
performance of these distributions, especially the use of the normal distribution, is analyzed using the data of TGR and Muma River. The main conclusions are summarized as
follows.
1.

2.

Analysis shows that the assumption of normal distribution is not justified sometimes,
and the use of this assumption for estimating flood risk may lead to incorrect results.
Therefore, it is necessary to carry out the hypothesis test first when using this
assumption.
The flood errors calculated by the same hydrologic model with different lead times
tend to obey the same distribution. For example, the 1-, 2-, and 3-day relative flood
errors for TGR follow the Wakeby distribution.

Acknowledgments The project was financially supported by the National Natural Science Foundation of
China (NSFC Grant 51309104, 51209221), Open Research Fund of State Key Laboratory of Simulation and
Regulation of Water Cycle in River Basin, China Institute of Water Resources and Hydropower Research
(IWHR-SKL-201408), Natural Science Foundation of Hubei Province (No. 2013CFB184) and Wuhan
Planning Project of Science and Technology (2014060101010064).

References
Alemu ET, Palmer RN, Polebitski A, Meaker B (2011) Decision support system for optimizing reservoir
operations using ensemble streamflow predictions. J Water Res PL-ASCE 137(1):7282
Bergstrom S (1992) The HBV modelits structure and applications. SMHI Reports RH No. 4, Norrkoping
Boucher MA, Tremblay D, Delorme L, Perreault L, Anctil F (2012) Hydro-economic assessment of
hydrological forecasting systems. J Hydrol 416:133144
Chen H, Guo SL, Xu CY, Singh VP (2007) Historical temporal trends of hydro-climatic variables and runoff
response to climate variability and their relevance in water resource management in the Hanjiang
basin. J Hydrol 334:171184
Chen L, Guo SL, Yan BW, Liu P, Fang B (2010) A new seasonal design flood method based on bivariate
joint distribution of flood magnitude and date of occurrence. Hydrol Sci J 55(8):12641280
Chen L, Singh VP, Guo S, Hao Z, Li T (2012) Flood coincidence risk analysis using multivariate copula
functions. J Hydrol Eng 17(6):742755
Chen L, Singh VP, Guo S, Ashok KM, Guo J (2013) Drought analysis based on copulas. J Hydrol Eng
18(7):797808
Christensen S (2003) A synthetic ground water modelling study of the accuracy of GLUE uncertainty
intervals. Nord Hydrol 35(1):4559
Diao YF, Wang BD, Liu J (2007) Study on distribution of flood forecasting errors by the method based on
maximum entropy. J Hydraul Eng 38(5):591595
Griffiths GA (1989) A theoretically based Wakeby distribution for annual flood series. Hydrol Sci J
34:231248
Grillakis MG, Tsanis IK, Koutroulis AG (2010) Application of the HBV hydrological model in a flash flood
case in Slovenia. Nat Hazards Earth Syst Sci 10:27132725
Harlin J, Kung CS (1992) Parameter uncertainty and simulation of design floods in Sweden. J Hydrol
137:209230

123

1402

Nat Hazards (2015) 75:13891402

Li X, Guo S, Liu P, Chen G (2010) Dynamic control of flood limited water level for reservoir operation by
considering inflow uncertainty. J Hydrol 391:124132
Lindstrom G (1997) A simple automatic calibration routine for the HBV model. Nord Hydrol 28(3):153168
Ministry of Water Resources (MWR) (2006) Regulation for calculating design flood of water resources and
hydropower projects. Chinese ShuiliShuidian Press, Beijing (in Chinese)
Ngoc T, Hiramatsu K, Harada M (2013) Optimizing parameters for two conceptual hydrological models
using a genetic algorithm: a case study in the DauTieng River Watershed. Vietnam. Jpn Agric Res Q
47(1):8596
Stedinger JR, Vogel RM, Lee SU, Batchelder R (2008) Appraisal of the generalized likelihood uncertainty
estimation (GLUE) method. Water Resour Res 44:W00B06
Xu D, Wang W, Chau K, Cheng C, Chen S (2013) Comparison of three global optimization algorithms for
calibration of the Xinanjiang model parameters. J Hydroinf 15:174191
Yan B, Guo S, Chen L (2013) Estimation of reservoir flood control operation risks with considering inflow
forecasting errors. Stoch Environ Res Risk Assess. doi:10.1007/s00477-013-0756-4
Yin HF, Li CA (2001) Human impact on floods and flood disasters on the Yangtze River. Geomorphology
41:105109
Zelenhastic E, Salvai A (1987) A method of streamflow drought analysis. Water Resour Res 23(1):156168
Zhao RJ (1992) The Xinanjiang model applied in China. J Hydrol 135:371381
Zhao RJ, Zhang YL, Fang LR (1980) The Xinanjiang model. Paper presented at Hydrological Forecasting
Proceeding Oxford Symposium, IASH-AISH Publ. no. 129, Washington, pp. 351356
Zhao TTG, Cai XM, Yang DW (2011) Effect of streamflow forecast uncertainty on real-time reservoir
operation. Adv Water Resour 34(4):495504
Zhao T, Zhao J, Yang D, Wang H (2013) Generalized martingale model of the uncertainty evolution of
streamflow forecasts. Adv Water Resour 57:4151

123

Vous aimerez peut-être aussi