Sie sind auf Seite 1von 15

Journal of Hydrology 318 (2006) 200–214

www.elsevier.com/locate/jhydrol

Stepwise development of a two-parameter monthly


water balance model
Safouane Mouelhi, Claude Michel, Charles Perrin*, Vazken Andréassian
Cemagref, Hydrosystems Research Unit, Parc de Tourvoie, BP 44, 92163 Antony Cedex, France
Received 1 March 2004; revised 18 May 2005; accepted 3 June 2005

Abstract

It has recently been shown that the structures of monthly water balance models can be very parsimonious. In this paper, we
describe how a new two-parameter rainfall-flow model was developed using a stepwise approach. A sample of 410 basins
representing a wide range of climate conditions from semi-arid through temperate to tropical humid was used to develop the
new model, on the basis of previous work on a four-parameter daily rainfall-flow model. Comparisons with several well-known
models showed that the new structure is efficient and provides valuable insight into the rainfall-flow transformation at a monthly
time-step.
q 2005 Elsevier Ltd All rights reserved.

Keywords: Hydrological modelling; Rainfall-runoff models; Water balance models; Lumped parameter models; Continuous simulation;
Monthly time-step

1. Introduction regionalisation and can be further used on ungauged


basins. Since they can be very simple, they are easy to
Over the last century many models have been handle for water resources managers. Of course, their
proposed to represent the transformation of precipi- low level of complexity means that they have to focus
tation into streamflow at the outlet of a basin. When the on the most prominent features of the rainfall-flow
time step is large, these models have usually been called transformation.
water balance models, assuming the response time to be Developing monthly water balance models implies
negligible compared to the time step. a threefold lumping—spatial, temporal and concep-
Monthly water balance models are valuable tools tual—of processes at work. Spatial lumping means
in water resources management, reservoir simulation, that rainfall and evaporation are averaged over the
drought assessment or long-term drought forecasting. basin of interest and are referred to as scalars.
These models are also very useful because, due to Temporal lumping means that only cumulative inputs
their inherent parsimony, they lend themselves to and outputs over a month are considered; we do not
know the actual sequence of rainfall events, but only
* Corresponding author. the resulting monthly aggregation of all inputs
E-mail address: charles.perrin@cemagref.fr (C. Perrin). recorded in a given month. Conceptual lumping
0022-1694/$ - see front matter q 2005 Elsevier Ltd All rights reserved.
doi:10.1016/j.jhydrol.2005.06.014
S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214 201

means that processes are not individualized and some hydraulic conductivity) while the others need
of them are assembled into simple sub-models. calibration.
From a physical point of view, one might argue In this article, we focus on lumped water balance
that, with such a large time step, the processes models involving a small number of free parameters to
underlying the rainfall-flow transformation may be be calibrated from rainfall and discharge data.
distorted out of recognition, leaving little hope of Identifying the relationships between parameters and
success. However, the performances of monthly basin characteristics is considered here as a question
water balance models are evidence of their to be addressed a posteriori (in regionalisation
efficiency. studies) rather than a priori (i.e. based on precon-
The MOSAZ model developed by Jayasuriya et al. ceived ideas of the physical meaning of parameters).
(1991) and presented by Boughton (2005) in his Indeed, we believe that insisting on a priori
review of water balance modelling in Australia, was connections to physical characteristics jeopardises
likely the first attempt to represent the monthly the ability of the data to verify or disprove these
rainfall-flow transformation with only two free conceptual links. The issue of parameter explanation
parameters. A few years later, the model proposed is not addressed in this paper.
by Makhlouf and Michel (1994), developed on French Looking at past experience of rainfall-flow
catchments, proved to be more successful than many modelling at a monthly time step, it seems that
other models in that climate context. Since then, this very few authors have made extensive comparative
model has been widely used in France (Lavabre et al., tests of these water balance models in various
1999, 2002) as well as in Western Africa (Paturel climate conditions (see e.g. Vandewiele and
Ni-Lar-Win, 1998). However, much could be gained
et al., 1995; Niel et al., 2003; Mahe et al., 2004). More
from such assessments to improve model efficiency
recently, Guo et al. (2002) also developed a two-
(Perrin et al., 2001). Therefore, one objective of our
parameter model that can be used in a semi-
research was to try to develop a new or improved
distributed way when dealing with large basins. It
model, even more general in its applicability,
was applied to large basins in China to predict the
making the most of recent experience in this domain
impacts of climate change.
and evaluating the proposed model on a large
Vandewiele et al. (1992) developed a series of
sample of catchments.
three-parameter models that were applied in several It would be tedious to retrace all the stages of
countries (see e.g., Xu and Vandewiele, 1995; Xu model development that have led to the monthly
et al., 1996; Müller-Wohlfeil et al., 2003). Ibrahim model presented in this article. The first step was to
and Cordery (1995) built a four-parameter model for choose a model structure general enough to encom-
Australian basins. Alley (1985) reports on an efficient pass the components of existing monthly models that
four-parameter model—the abcd model—applied to seemed the most efficient. Then, we made systematic
several catchments in the United States. attempts to improve its performance while reducing
In the literature, there are also other monthly its complexity. By testing the various model versions,
models with a much higher number of parameters: see we let hydrological data corroborate or invalidate
among others the CATPRO model (Kuczera et al., hypotheses about the best way to represent the main
1997) with up to 12 parameters used in Australia and features of the rainfall-flow transformation in a
the model developed by Pitman (1978a,b) and later tractable model. Note, however, that leading up to
used by Hughes and Metzler (1998) in Namibia and the rather straightforward approach outlined here,
by Wilk and Hughes (2002) in India, which also there has been a great amount of data mining and
contains 12 parameters. Among the models of empirical research.
intermediate complexity, the model proposed by The next section presents the data used for model
Jothityangkoon et al. (2001) was developed following development and comparative assessment. In the third
a downward approach. In these intermediate models, section, a Parent Model Scheme (here called PMS)
some parameters must be derived from physical basin with five parameters is presented. In the fourth
characteristics (such as slope, soil field capacity and section, we analyse this model to identify the number
202 S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214

of functions and/or parameters that deserve to be - 2.2. Assessment approach


present and/or calibrated. The resulting model struc-
ture including the seemingly most interesting The data record from each basin was divided into
components was further improved to give a new two periods of almost equal length. Every model was
monthly model, called GR2M and detailed in Section assessed in simulation mode on each period using
5. This model is then compared to five well-known parameter values calibrated on the other period,
models from the hydrological literature in Section 6. according to the split-sample technique (Klemeš,
The seventh section is devoted to a discussion of the 1986). This operation was repeated after switching the
water exchange term included in the GR2M model. two periods. At the beginning of each test period, the
The last section gives the main conclusions of this first year of simulated discharge was not used in the
study. computation of evaluation criteria (i.e. the warming-
up period was 1-year long). The overall assessment
was made on cumulative performances from both
tests in simulation mode. When only one basin was
2. Data used for model development and model considered, the criterion of the sum of least squared
assessment methodology errors was chosen to judge the efficiency of the model.
For each basin, a single value of RMSE (root mean
square error, in millimetres) was computed by
2.1. Data combining the sum of squared errors obtained in
simulation mode for each period in turn (Eq. (1)).
The large sample of basins used as a basis for
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
developing and comparing the various models spans u
u 1 X ni X 12
four continents. It has recently been used to explore RMSEi Z t ðQ K Q^ i;j;k Þ2 (1)
the role of complexity in model performances 12ni jZ1 kZ1 i;j;k
(Perrin et al., 2001) and to complete the develop-
ment of the daily lumped GR4J model (Perrin et al., where the subscript i, varying from 1 to N, indicates
2003). Most of the basins are situated in France, the basin, the subscript j, varying from 1 to ni,
others in the United States, Australia, the Ivory indicates the year in the time series of catchment i and
Coast and Brazil (see Fig. 1). Basin surface areas the subscript k corresponds to the month in year j. Q
range from 1 to 50,600 km2. The climate conditions stands for observed runoff, Q^ for calculated runoff.
vary from temperate through semi-arid to tropical The mean monthly rainfall, Pi , and the variance of the
humid, with mean annual potential evapotranspira- monthly rainfalls, VARPi, were computed on the
tion between 630 and 2040 mm, mean annual same dates:
rainfall of 300–2300 mm and mean annual stream-
1 X
ni X12
flow between less than 10 and 2040 mm. Inputs to Pi Z P (2)
the models are essentially areal precipitations 12ni jZ1 kZ1 i;j;k
(potential evapotranspiration rates are interannual
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
mean monthly values). The average record length is u
u 1 X ni X 12
about 15 years in our sample. Fig. 2 shows mean VARPi Z t ðP K Pi Þ2 (3)
interannual values of precipitation, potential evapo- 12ni jZ1 kZ1 i;j;k
transpiration and streamflow for all 410 basins of our
sample. In a few cases, mean streamflow exceeds where P is the observed precipitation. For each basin
mean rainfall, but the corresponding basins have not i, a goodness-of-fit criterion Ci could be computed by:
been excluded from our test sample because of the RMSEi
lack of an objective criterion to do the same critical Ci Z (4)
VARPi
work on the opposite range of behaviours, i.e. when
streamflow might be suspected of being unreason- to yield the overall evaluation criterion C over the
ably low. whole test basin sample:
S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214 203

Fig. 1. Catchment locations in (a) France, (b) the Ivory Coast, (c) Brazil, (d) Australia and (e) the United States.
204 S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214

Fig. 2. (a) Mean streamflow versus mean precipitation and (b) mean potential evapotranspiration versus mean precipitation for the 410 basins.

1 XN assessment penalizes any overparameterization


C Z C (5) phenomenon. Model parameters were calibrated by
N iZ1 i
a local optimisation scheme described by Edijatno
Often, RMSE is compared to the variance of et al. (1999).
observed streamflow, as in the widely used Nash and
Sutcliffe (1970) criterion. In Eq. (4), we used the
variance of observed rainfall to standardise model 3. The parent model scheme (PMS)
error and emphasise the role of the model in
transforming these inputs into streamflow. However, In order to determine which elements are important
since the common approach is to use the Nash and in a monthly rainfall-flow model, we started by
Sutcliffe criterion, we also give its mean value over adopting a parent model structure (PMS) that was
the whole test sample (hereafter called NS) along with later altered in a number of ways. Starting from the

the C-criterion value. Given the size of our test PMS, our approach is close to that described by
sample, any comparison or model ranking is bound to Boughton (2005) as a ‘parameter reduction’ process.

be sensitive to the choice of the overall criterion C. This is a very common approach in hydrological
However, there is little chance of greatly modifying modelling, used among others by Boughton (1984),
our conclusions on the respective value of the Jayasuriya et al. (1991), Chiew and McMahon (1994),
different model components tested here. Perrin et al. (2003), Boughton (2004), whose
To confer even greater robustness on our split- objective is ‘to reduce the number of variable
sample test approach, the parameters were calibrated parameters in a model by either assigning a constant
by using the least square errors computed on the value or else relating the value of a parameter to that
square root transforms of the discharges as the of another, such that one value determines two or
objective function, while the validation criteria were more of the original parameters’ (Boughton, 2005).
computed without prior transformation on the The parent model structure used here is derived
discharge. Many authors advocate the use of the from the experience gained during the development of
square root of the discharges before evaluating the monthly model proposed by Makhlouf and Michel
the square errors to make the most of the information (1994) and the daily GR4J model (Perrin et al., 2003),
imbedded in the data. In addition, the change of whose structures already appeared satisfactory in
criterion between the calibration and validation comparison with the existing ones. The functions
S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214 205

incorporated in the PMS model were selected after a where parameter X1, the maximum capacity of the
trial-and-error process showing their relevance and store, is positive and expressed in millimetres. It
efficiency. The GR4J model has only four free yields excess rainfall P1:
parameters, which severely limits the amount of
P1 Z P C S K S 1 (7)
complexity in modelling the monthly rainfall-flow
transformation. Given the temporal lumping from Due to evapotranspiration (the calculation of actual
daily to monthly time-steps, it should be possible to evapotranspiration depends on the potential value E),
model the rainfall-flow transformation with only two S1 becomes S2:
or three parameters. The PMS model is not overly  
S1 ð1 K jÞ E
complicated and still accommodates many of the most S2 Z  where j Z tanh (8)
features commonly found in the model structures. 1 C j 1 K X1S1 X 1

Since we wanted a general model whose most


important components could then be identified, five Then, the soil moisture storage releases water P2
degrees of freedom were initially introduced to this and takes its new value, S, ready for the next month:
model structure. Fig. 3 outlines this model. S2
The loss function (responsible for the determi- SZ
 X2 1=X2 (9)
nation of effective rainfall) in the PMS model is very 1 C XS21
similar to that of the daily GR4J model. It is based on
a soil moisture accounting store. Due to the rainfall P,
the soil moisture storage, S, becomes S1 obtained by: P2 Z S 2 K S (10)
where X2 is a positive parameter. The sum of P1 and
  P2 is the net rainfall, P3, that enters the routing part of
S C X1 4 P
S1 Z ; where 4 Z tanh (6) the model. A fraction, X3 (0%X3%1), is a first
1 C 4 XS1 X1
discharge component, Q1. The complementary part is

Fig. 3. Diagram of the Parent Model Scheme (PMS) used as a basis for developing a parsimonious rainfall-flow model at the monthly time step.
206 S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214

routed by a quadratic reservoir, with storage R. The values over the 410 basins, one for each parameter.
store capacity, (parameter X4) is a positive number The median value for parameter i was noted Mi. The
expressed in millimetres. The store releases a values found for the test sample and other character-
discharge Q2. The calculations are as follows: istics of parameter distributions are shown in Table 1.
R1 Z R C ð1 K X3 ÞP3 (11) Then, each parameter in turn was set at its median
value, Mi, and five versions involving four parameters
were calibrated and tested against the same data. The
R21
Q2 Z (12) following step involved setting two parameters, Xi and
R1 C X4
Xj, at their median values Mi and Mj, which yielded 10
models with three free parameters, and so on, until we
R Z R1 K Q 2 (13)
reached the version with no free parameters, i.e. the
The new storage R, is ready for the next month’s model with fixed parameters M1,., M5.
computations. The sum of Q1 and Q2 is multiplied by Results of model tests are shown in Table 2. In
the positive parameter X5 to give the actual discharge, 
order to explore the possible range of C-criterion
Q, from the studied basin: values, we also tested the very simple Ol’dekop
Q Z X5 ðQ1 C Q2 Þ (14) (1911) model in the same calibration-simulation
pattern. This model has no free parameters and
On the one hand, X5 can be interpreted as a water computes monthly flow as:
exchange term. If X5 is greater than 1, there is a water  
supply from the outside of the basin; otherwise there P
Qk Z Pk K Ek tanh k (15)
is a loss. To use such a parameter is the easiest way to Ek
model a water exchange with an outside environment where the subscript k indicates the month. This model
other than the atmosphere. On the other hand, X5 can was originally developed for an annual time step and
also be interpreted as a catchment area correction includes no memory device, contrary to the PMS
factor. 
model versions tested here. The C-criterion value for
the Ol’dekop model is 0.628 and the mean Nash–
Sutcliffe criterion NS is K1.490.
4. Stepwise search for an optimal version It is comforting to discover that the version of the
of the monthly water balance model PMS model with no free parameters (version #32 in
Table 1) is far more efficient than the Ol’Dekop model
After selecting the PMS model as an a priori 
since the C-criterion is 0.286 instead of 0.628
reliable starting point to build our monthly model, we
tried to reduce its complexity and identify its most Table 1
useful components. Here, the modelling task is carried Percentiles of parameter distributions obtained for the Parent Model
Scheme (PMS) with, respectively, five and two optimised
out as a fact-finding operation with little allowance parameters (the other parameters of the two-parameter version
made for any prior knowledge of what a monthly were fixed at median values) and for the final GR2M model on the
model should look like. The spatially lumped whole sample of catchments
approach together with a large time step excludes Values of distribution percentiles
most lines of physically based reasoning. Conse-
0.05 0.50 0.95
quently, we adopted an empirical search, to discover
PMS model X1 (mm) 162 330 1670
which structure is most faithful to the actual rainfall-
(5 parameters) X2 (K) 1.38 3.19 39.6
flow transformation. With the proposed structure of X3 (K) 0.00 0.00 0.57
the PMS model, all versions of the model arising from X4 (mm) 0.0 53.0 469
all combinations of free parameters were tested on the X5 (K) 0.02 0.85 1.62
large data sample described above. PMS model X1 (mm) 84 360 1940
(2 parameters) X5 (K) 0.02 0.86 1.73
In order to investigate prospective parameter
GR2M Model X1 (mm) 140 380 2640
values, we first calibrated all parameters. This X5 (K) 0.21 0.92 1.31
calibration produced five distributions of parameter
S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214 207

free parameters, X5 excluded (versions #15, #17 and


#6, respectively), are clearly less efficient (CZ 
0.199, 0.203 and 0.199, respectively and K0.002,
K0.021, K0.000 for NS) than the model with X5 as
the single free parameter (version #31, CZ  0:186,
NS Z 0:497). As to the second best parameter to
calibrate, the solution with calibration of the routing
storage capacity, X4 (version #26, CZ  0:161,
NS Z 0:619) is slightly better than the one where
the soil moisture storage capacity, X1, is calibrated
(version #20, CZ  0:160, NS Z 0:617). The C-
 criterion is further slightly improved when both X1
Fig. 4. Plot of the C-criterion values of the 32 PMS model versions
as a function of the number of free parameters (the line links best and X4 are calibrated in addition to X5 (version #11,
performance for each number of free parameters).  0:155, NS Z 0:635), but this improvement is
CZ
probably too small to justify an additional par-
ameter. Note that two other three-parameter models
(K0.758 instead of K1.490 for NS). If there is one (version #7 with X3, X4 and X5 calibrated and
parameter to calibrate (see versions #27 to #31) it version #14 with X1, X2 and X5 calibrated) give very
should undoubtedly be X5 (version #31, which further  0:158, NS Z 0:629 for #7; 0.160
similar results (CZ

reduces the C-criterion from 0.286 to 0.186 and and 0.624, respectively for #14). However, since X1,
increases NS from K0.758 to 0.497). This is very X4 and X5 are the ones most frequently optimised
surprising since the corresponding process—an within these three best performing structures, one
exchange with the outside of the basin—is generally can conclude that they are the most important ones
omitted in most models. Here, on the contrary, X5 is so in the PMS model. Fig. 4 shows the evolution of
important that the best models with two, three or four model performance when the number of free

Fig. 5. Diagram of the GR2M model.


208 S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214

Fig. 6. Map of the two-parameter experimental probability density of the GR2M model. The scale indicates the probabilities multiplied by 1000.

parameters is increased. It confirms that, as soon as We propose to further refine the outside exchange
the model has two free parameters, very little can be function. It is no longer modelled as if it were the
gained by adding more flexibility. The provisional effect of a change in the effective basin area (which
conclusion is that a two-parameter model is translated into multiplying Q1CQ2 by X5). Instead,
sufficient at a monthly time-step. The parameter the routing reservoir is subject to an outside exchange
distributions characteristics of this two-parameter proportional to its content. This mathematical
PMS model version are shown in Table 1. However, formulation was chosen as the most satisfactory
given the importance of parameter X5, a more among many others. After adding input P3 to the
elaborate way to model the outside exchanges was routing reservoir to obtain level R1, an outside
investigated, as described below. exchange term is computed as follows:

5. The new model: GR2M

The following conclusions were drawn from the


preceding section: (i) the parameter of the percolation
equation, X2, is no longer a free parameter; (ii) there is
no direct discharge by-passing the routing reservoir
(X3Z0) and therefore, Q1Z0; (iii) the capacity of the
routing store, X4, is fixed. Here, the formulation of the
soil moisture part was also kept unchanged. The two
other parameters (X1, the capacity of the soil moisture
reservoir and X5, the parameter of the outside Fig. 7. Distributions of the C and Nash–Sutcliffe criteria obtained by
exchange) were left free. the GR2M model over the 410 catchments.
S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214 209

Table 2
Values of C and NS criteria obtained for all combinations of free parameters within the PMS model over the catchment sample

Version # Parameter values (*Zcalibrated) C criterion NS criterion


(number of
free par-
ameters)
X1 X2 X3 X4 X5
1 (5) * * * * * 0.155 0.641
2 (4) 330 * * * * 0.156 0.635
3 (4) * 3.19 * * * 0.156 0.635
4 (4) * * 0 * * 0.154 0.640
5 (4) * * * 53 * 0.157 0.635
6 (4) * * * * 0.85 0.199 0.000
7 (3) 330 3.19 * * * 0.158 0.629
8 (3) 330 * 0 * * 0.159 0.629
9 (3) 330 * * 53 * 0.177 0.541
10 (3) 330 * * * 0.85 0.229 K0.520
11 (3) * 3.19 0 * * 0.155 0.635
12 (3) * 3.19 * 53 * 0.159 0.627
13 (3) * 3.19 * * 0.85 0.214 K0.259
14 (3) * * 0 53 * 0.160 0.624
15 (3) * * 0 * 0.85 0.199 K0.002
16 (3) * * * 53 0.85 0.203 K0.023
17 (2) * * 0 53 0.85 0.203 K0.021
18 (2) * 3.19 * 53 0.85 0.236 K0.341
19 (2) * 3.19 0 * 0.85 0.214 K0.269
20 (2) * 3.19 0 53 * 0.160 0.617
21 (2) 330 * * 53 0.85 0.247 K0.678
22 (2) 330 * 0 * 0.85 0.229 K0.534
23 (2) 330 * 0 53 * 0.178 0.524
24 (2) 330 3.19 * * 0.85 0.275 K0.633
25 (2) 330 3.19 * 53 * 0.183 0.515
26 (2) 330 3.19 0 * * 0.161 0.619
27 (1) * 3.19 0 53 0.85 0.230 K0.321
28 (1) 330 * 0 53 0.85 0.246 K0.617
29 (1) 330 3.19 * 53 0.85 0.285 K0.812
30 (1) 330 3.19 0 * 0.85 0.275 K0.645
31 (1) 330 3.19 0 53 * 0.186 0.497
32 (0) 330 3.19 0 53 0.85 0.286 K0.758

The negative Nash–Sutcliffe criteria are in bold characters. The fixed parameters take the median value of calibrated parameters in version #1.

F Z ðX5 K 1ÞR1 (16) and the reservoir level is updated to the R value ready
for the computation of the next month:
where X5 (non-dimensional) is a positive parameter. If
X5 is greater than 1, F represents a gain of water for R Z R2 K Q2 (19)
the basin and a loss otherwise. Then, the reservoir The water exchange term must act on the routing
level becomes: reservoir before it empties. If the sequence of
R2 Z X5 :R1 (17) calculations is changed, model performance signifi-
cantly decreases.
Reservoir output is computed by: Since this improved formulation of the water
exchange function substantially modified the model
R22 structure, the choice of fixed parameter values must
Q2 Z (18)
R2 C X4 be challenged and a new analysis similar to the one
210 S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214

described in Section 4 must be made. For the sake basins. To this end, we used the evaluation process
of brevity, only the main results of this analysis are applied in Section 4 to compare GR2M to five models:
presented. The best two-parameter model, called the two-parameter model proposed by Guo et al.
GR2M, is the one where X1 and X5 must be (2002), the MOSAZ model (Jayasuriya et al., 1991), a
calibrated. X4 is set at a fixed value of 60 mm and three-parameter version of the Wilk and Hughes
the X2 parameter at 3. Fig. 5 shows a sketch of the (2002) model involving linear functions of the soil
final GR2M model. The characteristics of the two moisture reservoir, the Thornthwaite and Mather
parameter distributions are set out in Table 1 and T-alpha model as described by Arnell (1992), and
the two-parameter experimental probability density the four-parameter model, abcd, detailed by Alley
function is mapped in Fig. 6. With the new (1985). For the sake of brevity, these models are not
specifications, the C criterion obtained in simu- described here and one should refer to the referenced
lation mode on the 410 catchments is 0.153 and the articles for a detailed description. Table 2 summarizes
NS criterion is 0.642, i.e. a substantial improve- the results (C and NS criteria) obtained for each tested
ment on the best two-parameter version tested in model as well as the Ol’dekop relationship (Eq. (15))
Section 4. One can see in Table 1 that the previous that is used as a baseline model.
version would have required four free parameters
 0:154, NS Z 0:640) to obtain a The GR2M model performs very satisfactorily as
(version #4, CZ
compared to the other models. The good performance
performance similar to (though still lower than) the
of two of the three two-parameter models is likely to
new version. For a more complete view of the
be taken as evidence that the optimal number of
performance range, the distributions of the C
parameters is two, as far as the monthly time step is
criteria (Eq. (4)) and of the Nash–Sutcliffe criteria
obtained on the 410 catchments are shown in concerned.
Fig. 7. The Ol’dekop and abcd models perform quite
poorly. It could be argued that none of them has a
functionality similar to that offered by parameter X5
in GR2M, which may be viewed as a way to force
6. Comparison of GR2M with models from long-term water balance closure. In order to make
the literature the comparison more balanced, an additional free
parameter to scale the rainfall input was introduced
The validity of the new version can be confirmed both in the Ol’dekop and the abcd models. The
by comparing its performance with that of other well- performance of both models is greatly enhanced by
known models applied to the same large sample of this modification. However, the gain is not enough

Table 3
Comparison of the performance (C and NS criteria) of several monthly water balance models in simulation mode

Model Number of free parameters NS (best: 1) C (best: 0)


Ol’dekop (1911) 0 K1.490 0.628
Ol’dekop (1911) with a rainfall 1 K0.170 0.409
multiplying parameter
MOSAZ (Jayasuriya et al., 1991) 2 K0.336 0.220
GR2M (this version) 2 0.642 0.153
Guo et al. (2002) 2 0.278 0.194
Vandewiele et al. (1992) with linear 3 0.325 0.188
relations
Thornthwaite and Mather (Arnell, 1992) 3 0.351 0.216
abcd model (Alley, 1985) 4 K1.717 0.406
abcd model (Alley, 1985) with a rainfall 5 K0.003 0.339
multiplying parameter

Tentative variants of the Ol’dekop and abcd models involve a rainfall multiplying parameter.
S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214 211

Fig. 8. Observed and simulated monthly flows for the Leaf River (a) in the 1950–1968 period (with parameters calibrated on the 1969–1986
period) (Nash–Sutcliffe criterion: 80.9%) and (b) in the 1969–1986 period with parameters calibrated on the 1950–1968 period) (Nash–Sutcliffe
criterion: 82.6%).

for these models to compete with the best ones in However, this issue deserves further discussion
Table 3. A tentative conclusion could be that the (see Section 7).
underground water exchange parameter X5 is not When the data sample is this large, it would not be
just a fudge factor that compensates for bias in basin fair to use only one simulation example to show the
rainfall and/or potential evapotranspiration data. results. But since the well-known Leaf River basin
212 S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214

near Collins, Mississippi (USA) is part of our test


sample, we used this example to illustrate the
simulation results of the GR2M model in Fig. 8.
It is true that the good performance of the GR2M
model on our large test sample is not a proof of its
validity. But it indicates that it is likely to be more
reliable than other competing models. Note that
although the GR2M model appears significantly
better than the other models tested here, the ranking
among the best models might be slightly modified
by using other criteria to measure their performance.
However, the main conclusions are likely to hold
true. For example, the results obtained with the C Fig. 9. Sketch of a basin with input precipitation P and potential
and NS criteria in Tables 2 and 3 are quite evapotranspiration E. Q is the streamflow at the outlet and F (loss or
consistent. gain) is the resultant of all flows crossing the lateral underground
boundary (arrows are just indicative of where the fluxes originate).

7. Is the underground water exchange parameter cylinder on the topographical boundary of the basin
(X5) a fudge factor? forms the lateral boundary of the system. Second, the
system is arbitrarily bounded at the bottom by a layer
Questioning the validity of a parameter such as X5 at sea level such that fluxes through this horizontal
is natural, because some modellers could see it as a boundary are likely to have a zero sum. Since the
fudge correction factor. Indeed this parameter may be lateral underground boundary is an imaginary surface,
interpreted both as (i) a correction parameter that it is not an impermeable casing reproducing itself as
could account for possible biases in precipitation or we move along the river. The resulting algebraic flux
streamflow measurements, or (ii) an underground through this boundary, named F in Fig. 9, cannot be
exchange parameter accounting for the actual water equal to zero.
exchanges with the outside environment. What There is no direct information by which to assess
elements may help to decide whether we can restrict the amplitude and dynamics of F. In our lumped
the interpretation of X5 to the first case? In the modelling approach, hypotheses about F can only be
following, we attempt to answer this question. made indirectly, based on precipitation and outflow
First, the introduction of a parameter to adjust series. F is the resultant of local fluxes that are
rainfall in the abcd model led to an insufficient probably very complex. In some basins, it may be
improvement of model performance (see Table 3). We roughly assimilated to the common underflow that
interpreted this as an indication of the genuine occurs in the valley substratum. In large artesian
necessity to allow for a water exchange between the regions, F could be a gain for the system. In other
basin and its environment other than the atmosphere. basins, F might be small, but the hypothesis of F being
But there is a more compelling reason to introduce equal to zero for every basin at all times is
the underground water exchange parameter X5. The indefensible. Just imagine what occurs when the
real reason is that it is extremely hard to imagine an basin outlet is gradually moved upstream. With F
absence of underground exchange in any basin. Let us remaining equal to zero, there is no transfer through
return to the natural system that the models are the vertical slice of soil between the original system
intended to represent. Fig. 9 shows such a system that and the system obtained by an incremental displace-
is obviously three-dimensional although the term ment of the downstream section. If this phenomenon
basin often refers to the upper surface of the system. occurs everywhere and at every time the most
We must stress that very little is known of the vertical physical image we can get of the system in Fig. 9 is
dimension. In Fig. 9, the underground boundary of the that of a lump of impermeable material. Therefore, F
system is composed of two parts. First, a vertical does exist and cannot be ignored on the grounds that it
S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214 213

is difficult to model. The hypothesis that F is Nascimento (University of Minas Gerais, Belo
negligible must be demonstrated by the fact that the Horizonte, Brazil) for data of Brazilian catchments;
models where it is not included are more efficient than Dr John Schaake and the MOPEX team for US
the others since they are not overburdened by a catchments; the HYDRO database of the French
useless additional component. The main result of our Ministry for the Environment and Météo France for
research is that F, far from being a negligible flux of the hydrological and meteorological data sets of the
water, is likely an important feature of water balance French catchments. This research was partly sup-
modelling. ported by a research grant—provided by the French
Embassy in Tunisia to the first author—which is also
gratefully acknowledged. D.A. Hughes and an
8. Conclusion anonymous reviewer are heartily thanked for their
many suggestions that helped to improve the clarity of
A new model of rainfall-flow transformation at the the original version of this paper.
basin scale is proposed to answer hydrological
questions which are best addressed at the monthly References
time step, e.g. those pertaining to reservoir manage-
ment and long-term drought forecasting. It was not Alley, W.M., 1985. Water balance models in one-month-ahead stream
possible, within the scope of this article, to retrace all flow forecasting. Water Resources Research 21 (4), 597–606.
the development studies undertaken to finalise the Arnell, N., 1992. W., Factors controlling the effects of climate
model. The chosen approach was clearly empirical, change on river flow regimes in a humid temperate environ-
ment. Journal of Hydrology 132, 321–342.
based on a large bulk of data. The position taken by Boughton, W., 1984. A simple model for estimating the water yield
the authors was to trust the data in order to unveil the of ungauged catchments. Civil Eng. Trans., Institution of
structure that best depicts the rainfall-flow transform- Engineers, Australia CE26 (2): 83-88.
ation and to distinguish the components that are Boughton, W., 2004. The Australian Water Balance Model.
important from the ones to which fixed parameter can Environmental modelling and software 19, 943–956.
Boughton, W., 2005. 2005. Catchment water balance modelling in
be assigned. The relative success of GR2M proves the Australia 1960- Agricultural Water Management 71, 91–116.
overriding importance and complexity, at the monthly Chiew, F.H., McMahon, T.A., 1994. Application of the daily
time step, of the loss sub-model to which both free rainfall-runoff model MODHYDROLOG to 28 Australian
parameters are allowed. The vital role of the loss catchments. Journal of Hydrology 153, 383–416.
model is not entirely surprising when the time step is Edijatno, Nascimento, N.O., Yang, X., Makhlouf, Z., Michel, C.,
1999. GR3J: a daily watershed model with three free
as large as a month but it is surprising that the routing parameters. Hydrological Sciences Journal, 44(2), 263-277.
part does not require a single parameter to be Guo, S., Wang, J., Xiong, A., Li, D., 2002. A macro-scale and semi-
calibrated. However, model development is a never- distributed monthly water balance model to predict climate
ending process. It is likely that a more data-faithful change impacts in China. Journal of Hydrology 268, 1–15.
model will be found in the future, possibly with new Hughes, D.A., Metzler, W., 1998. Assessment of three monthly
rainfall-runoff models for estimating the water resource yield of
insights into underground water exchanges. semiarid catchments in Namibia. Hydrological Sciences Journal
43 (2), 283–297.
Ibrahim, A.G., Cordery, I., 1995. Estimation of recharge and runoff
Acknowledgements volumes from ungauged catchments in eastern Australia.
Hydrological Sciences Journal 40 (4), 499–515.
Jayasuriya, L.N., McMahon, T.A., O’Neill, I.C., 1991. Develop-
The authors thank the scientists and institutions ment of a simplified two-parameter rainfall-runoff model with
that provided data sets for model testing: Dr Francis potential for ungauged catchment application. International
H.S. Chiew (University of Melbourne, Australia) for Hydrology and Water Resources Symposium, vol. 2, 498–503.
data sets of the Australian catchments; Dr Eric Servat Jothityangkoon, C., Sivapalan, M., Farmer, D.L., 2001. Process
(Institut de Recherche pour le Développement, controls of water balance variability in a large semi-arid
catchment: downward approach to hydrological model develop-
France) for data sets of catchments in the Ivory ment. Journal of Hydrology 245, 174–198.
Coast; Jane L. Thurman (USDA, United States) for Klemeš, V., 1986. Operational testing of hydrological simulation
data sets in the US (ARS data base); Dr Nilo Oliveira models. Hydrological Sciences Journal 31 (1), 13–24.
214 S. Mouelhi et al. / Journal of Hydrology 318 (2006) 200–214

Kuczera, G., Raper, G.P., Brah, N.S., Jayasuriya, M.D., 1997. The Paturel, J.E., Servat, E., Vassiliadis, A., 1995. Sensitivity of
quest for more powerful validation of conceptual catchment conceptual rainfall-runoff algorithms to errors in input data -
models. Water Resources Research 33 (10), 2325–2335. case of the GR2M model. Journal of Hydrology 168, 11–25.
Lavabre, J., Aranaud, P., Masson, J.M., Folton, N., 1999. Apport de Perrin, C., Michel, C., Andréassian, V., 2001. Does a large number
la modélisation de la pluie en débit pour la connaissance de la of parameters enhance model performance? Comparative
ressource en eau et la prédétermination des crues. La Houille assessment of common catchment model structures on 429
Blanche (3), 67–71. catchments. Journal of Hydrology 242, 275–301.
Lavabre, J., Folton, N., Fouchier, C., 2002. Connaissance régionale Perrin, C., Michel, C., Andréassian, V., 2003. Improvement of a
de la ressource en eau de surface: application au quart sud-est parsimonious model for streamflow simulation. Journal of
français. Hydrology of mediterranean and semi-arid regions. Hydrology 279, 275–289.
IAHS Publ. n8 278, 94–100. Pitman, W.V., 1978a. Flow generation by catchment models of
Mahe, G., Paturel, J.E., Servat, E., Conway, D., Dezetter, A., 2004. differing complexity - A comparison of performance. Journal of
Hydrology 38, 59–70.
The impact of land use change on soil water holding capacity
Pitman, W.V., 1978b. Trends in streamflow due to upstream land
and river flow modelling in the Nakambe River. Burkina-Faso.
use change. Journal of Hydrology 39, 227–237.
Journal of Hydrology 300 (1-4), 33–43.
Vandewiele, G.L., Xu, C.Y., Ni-Lar-Win, 1992. Methodology and
Makhlouf, Z., Michel, C., 1994. A two-parameter monthly water
comparative study of monthly models in Belgium, China and
balance model for French watersheds. Journal of Hydrology
Burma. Journal of Hydrology 134, 315–347.
162, 299–318.
Vandewiele, G.L., Ni-Lar-Win, 1998. Monthly water balance
Müller-Wohlfeil, D.-I., Xu, C.-Y., 2003. Legard Iversen. H., models for 55 bassins in 10 countries. Hydrological Sciences
Estimation of monthly river discharge from Danish catchments. Journal 43 (5), 687–699.
Nordic Hydrology 34 (4), 295–320. Wilk, J., Hughes, D.A., 2002. Calibrating a rainfall-runoff model for
Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through a catchment with limited data. Hydrological Sciences Journal 47
conceptual models. Part I - A discussion of principles. Journal of (1), 3–17.
Hydrology 27 (3), 282–290. Xu, C.Y., Vandewiele, G.L., 1995. Parsimonious monthly rainfall-
Niel, H., Paturel, J.-E., Servat, E., 2003. Study of parameter stability runoff models for humid basins with different input require-
of a lumped hydrologic model in a context of climatic ments. Advances in Water Resources 18, 39–48.
variability. Journal of Hydrology 278, 213–230. Xu, C-Y., Seibert, J., Halldin, S., 1996. Regional water balance
Ol’dekop, E.M., 1911. On evaporation from the surface of modelling in the NOPEX area: development and application of
river basins. Trans. Metorol. Observ. University of Tartu 4, monthly water balance models. Journal of Hydrology 180, 211–
200. 236.

Das könnte Ihnen auch gefallen