Zhang2008 PDF

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Journal of Hydrology (2008) 360, 117– 131

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/jhydrol

Water balance modeling over variable time scales


based on the Budyko framework – Model
development and testing
a,*
Lu Zhang , Nick Potter a, Klaus Hickel a, Yongqiang Zhang a, Quanxi Shao b

a
CSIRO Land and Water, G.P.O. Box 1666, Canberra, ACT 2601, Australia
b
CSIRO Mathematical and Information Sciences, Private Bag No. 5, Wembley, WA 6913, Australia

Received 22 December 2007; received in revised form 27 June 2008; accepted 11 July 2008

KEYWORDS Summary Partitioning of precipitation into evapotranspiration and runoff is controlled


Water balance; by climate and catchment characteristics. The degree of control exerted by these factors
Evapotranspiration; varies with the spatial and temporal scales of processes modeled. The Budyko framework
Ungauged catchments; or the ‘‘limits’’ concept was used to model water balance at four temporal scales (mean
Budyko framework annual, annual, monthly and daily). The method represents a top-down approach to
hydrologic modeling and is expected to achieve parsimony of model parameters. Daily
precipitation, potential evapotranspiration, and streamflow from 265 catchments in Aus-
tralia were used. On a mean annual basis, the index of dryness defined as the ratio of
potential evapotranspiration to precipitation was confirmed to be a dominant factor in
determining the water balance with one model parameter. Analysis of the data, however,
suggested increased model complexity is necessary on finer time scale such as monthly. In
response, the Budyko framework for mean annual water balance was extended to include
additional factors and this resulted in a parsimonious lumped conceptual model on
shorter-time scale. The model was calibrated and tested against measured streamflow
at variable time scales and showed promising results. The strengths of the model are con-
sistent water balance relationships across different time scales, and model parsimony and
robustness. As result, the model has the potential to be used to predict streamflow for
ungauged catchments.
Crown Copyright ª 2008 Published by Elsevier B.V. All rights reserved.

* Corresponding author.
E-mail address: lu.zhang@csiro.au (L. Zhang).

0022-1694/$ - see front matter Crown Copyright ª 2008 Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2008.07.021
118 L. Zhang et al.

Introduction on water balance over variable time scales. Specifically,


the paper presents water balance modeling over timescales
The movement of water through the continuum of soil, veg- ranging from mean annual to daily based on Budyko’s frame-
etation, and atmosphere is an important process. Under- work. This method could be considered as a holistic or sys-
standing the water balance in relation to climate and tems approach to catchment water balance modeling. It
catchment characteristics provides insight into the complex enables the development of water balance models with con-
processes operating over a range of spatial and temporal sistency in process representation and parsimony of model
scales. At the catchment scale, quantifying the effects of parameters, making them suitable for application to unga-
land-use change on water balance and prediction of stream- uged catchments and for predicting land-use change im-
flow in ungauged catchments are the major scientific chal- pacts. The paper begins with a brief description of the
lenges for hydrologists (Zhang et al., 2004; PUB initiative). catchments and data used in this study. The next section
While there have been major advances in our knowledge presents the development of the variable timescale water
of the physical and biophysical processes controlling the balance models and their calibration. Finally, results of
water balance, it remains a difficult task to develop models the model calibration and testing are discussed.
that can be used to make quantitative hydrological predic-
tions at catchment scale. This is partly due to the fact that Catchment description and data
for such a model to be of practical use it must meet the
requirements of parsimony in terms of data inputs and mod- The catchments included in this study have at least five
el parameters. This requires inevitably that the model is complete years of unimpaired daily streamflow data and a
sufficiently simple and that parameters can be estimated catchment area between 50 and 2000 km2. Unimpaired
from known climate and catchment characteristics. streamflow is defined as streamflow that is not subject to
In an attempt to understand relationships between water regulation or diversion. The streamflow data was assembled
balance and climate, Budyko (1958) postulated that long- by Peel et al. (2000). In total, 265 catchments were
term average annual evapotranspiration from a catchment available.
is determined by rainfall and available energy. Based on this Monthly rainfall was estimated from gridded daily rainfall
assumption, Budyko derived a simple water balance model (Peel et al., 2000). The spatial resolution of the gridded dai-
known as Budyko’s curve that showed good agreement with ly rainfall is 5 km by 5 km based on interpolation of over
the long-term water balance data for a number of catch- 6000 rainfall stations in Australia. The interpolation uses
ments in the former USSR. The work of Budyko has led to monthly rainfall data, ordinary Kriging with zero nugget
more theoretical studies trying to understand how climatic and a variable range. Monthly rainfall for each point is con-
and catchment characteristics affecting equilibrium or verted to daily rainfall using the daily rainfall distribution
long-term average water balance (Milly, 1994; Koster and from the station closest to that point. Catchment averaged
Suarez, 1999; Choudhury, 1999; Zhang et al., 2001, 2004; rainfall was estimated from the daily rainfall. Mean daily
Farmer et al., 2003; Raupach et al., 2001; Atkinson et al., potential evapotranspiration was calculated based on the
2002; Potter et al., 2005). The importance of these studies Priestley–Taylor equation (Priestley and Taylor, 1972). De-
is that they showed a hierarchy of control exerted by cli- tails of the calculation can be found in Raupach et al.
mate and catchment characteristics on water balance over (2001).
various temporal scales. At mean annual scale, climatic fac- Potter et al. (2005) and Hickel et al. (2006) showed that
tors such as precipitation and potential evapotranspiration rainfall seasonality is important for the water balance of
are the main control on water balance and inclusion of these Australian catchments. A distinction was made between
two factors may be sufficient for many purposes (Budyko, catchments with different seasonal distributions of rainfall.
1958). These studies also demonstrated clearly the scien- A sine curve was fitted to monthly means of rainfall and po-
tific merits and practical benefits of the method developed tential evapotranspiration and the phase shift between the
by Budyko (1958), which is an example of a top-down ap- rainfall and potential evapotranspiration curves was calcu-
proach (Sivapalan et al., 2003). The main feature of the lated. The amplitude of the sine curve indicates the
top-down approach is its empirical nature, seeking an strength of rainfall seasonality. A phase shift of 6–8 months
understanding of overall catchment behavior and function indicates a winter-dominant rainfall regime and summer-
based on observed data. In the top-down approach, one dominant regime has a phase shift of – 1 to 3 months. If
starts from exploring first-order controls in catchment the amplitude of rainfall divided by mean annual rainfall
water balance, and model complexity is increased only is small (<0.2), a non-seasonal rainfall regime is assumed.
when deficiencies are identified. This means that when a fi- The catchments selected in this study cover large geo-
ner time scale is of interest, one may have to include addi- graphic regions and span a variety of climates including
tional factors such as rainfall seasonality and catchment tropical, dry, and warm temperate. Mean annual precipita-
water storage capacity into the water balance model. It is tion from the catchments varies from 282 to 2886 mm with
equally important that the complexity of the model is sup- different seasonal distributions. The mean annual potential
ported by available data. To meet the challenge of predic- evapotranspiration ranges from 785 to 2095 mm. The vege-
tions for ungauged catchments, knowledge of model tation in the catchments includes native woodlands, open
parameters in relation to climate and catchment character- forests, rainforests, eucalyptus, various species of pine
istics is essential. trees, native and managed grass, and agricultural crops.
This study aims to contribute to the understanding of the The catchments cover soil types from sand through loams
controls exerted by climatic and catchment characteristics to clays with large differences in soil properties such as sat-
Water balance modeling over variable time scales based on the Budyko framework – Model development and testing 119

urated hydraulic conductivity and water holding capacity. proposed based on the ‘‘limits’’ concept (Calder, 1998).
Runoff ratio of the catchments ranges from 0.01 to 0.65, Based on phenomenological considerations, Fu (1981) devel-
representing diverse hydrological conditions. oped the following relationships for estimating mean annual
evapotranspiration:
Water balance modeling over variable time   w 1=w
ET E0 E0
scale ¼1þ  1þ ð5Þ
P P P
Equilibrium water balance model where E0 is potential evapotranspiration, w is a model
parameter with range (1,1). For the purpose of model cal-
The dynamic water balance of a catchment can be written as ibration, we define a ¼ 1–1=w so that a varies between 0
d and 1 and an increase in a corresponds with an increase in
Sc ðtÞ ¼ PðtÞ  ETðSc ; tÞ  Q ðSc ; tÞ ð1Þ evapotranspiration efficiency. The relationship represented
dt
by Eq. (5) is shown in Fig. 1 and details of the solutions are
where Sc(t) is the total stored water in the catchment, and
given in Zhang et al. (2004).
P(t), ET(Sc,t), Q(Sc,t) are precipitation, evapotranspiration,
The method of Fu (1981) is similar to Budyko (1958) in
and total runoff, respectively.When Eq. (1) is integrated
assuming that the equilibrium water balance is controlled
over a time interval T, one obtains
by water availability and atmospheric demand. Zhang et
Sc ðTÞ  Sc ð0Þ al. (2004) argued that Eq. (5) has a physical basis and is a
¼ P  ET  Q ð2Þ
T better model for estimating mean annual evapotranspira-
where P, E, Q are time-averaged water fluxes given by tion compared with other similar empirical equations.
Z Z By combining Eqs. (4) and (5), one obtains the following
1 T 1 T expression for mean annual streamflow:
P¼ PðtÞdt; ET ¼ ETðSc ; tÞdt;
T 0 T 0
Z T  1 1
1a
1 Q ¼ P 1a þ E 01a  E0 ð6Þ
Q¼ Q ðSc ; tÞdt ð3Þ
T 0
Eq. (6) can be used to calculate mean annual stream-
The left-hand side of Eq. (2) accounts for the effect of
flow when mean annual values of rainfall and potential
water storage changes in the time-averaged water balance
evapotranspiration are known. A comparison of observed
and decreases relative to the right hand side as T increases.
and calculated mean annual streamflow and runoff ratios
If T is long enough, e.g. decades, the storage terms can be
for the 265 catchments is shown in Fig. 2. In the calcula-
neglected relative to the fluxes and Eq. (2) becomes
tion, a single value of 0.64 was used for the a parameter
0 ¼ P  ET  Q ð4Þ and it resulted in a mean absolute error (MAE) of 44 mm
which can be thought as the ‘‘equilibrium’’ or ‘‘steady- and the coefficient of efficiency of 0.93. Potter et al.
state’’ water balance. (2005) showed that while Eq. (5) with a single a value pre-
The evaluation of the water balance equation requires dicts overall mean annual evapotranspiration ratio well,
extra information such as catchment physical characteris- calibrated values of the a parameter vary with rainfall
tics, climatic variables, and further relationships. In the seasonality.
case of dynamic water balance models, the extra informa-
tion may include equations for estimating evapotranspira- Interannual variability in water balance
tion and groundwater recharge. Such models are In moving from mean annual to shorter-time scale, i.e.
generally complicated with a large number of parameters decreasing T, one generally has to account for the effect
(Walker and Zhang, 2002). The models for the equilibrium of catchment water storage change on the water balance.
water balance are much simpler, but can provide useful in- Koster and Suarez (1999) assumed that interannual changes
sight into the key processes responsible and they are gen- in catchment water storage are much smaller than the an-
erally robust. nual precipitation, evaporation, and runoff. As a result,
they suggested that the evaporation ratio for any one
catchment in a given year i can be modeled using Bud-
Budyko’s framework yko-like curves:
    
The mean annual water balance can be modeled using the ET E0
method of Budyko (1958) by only considering dominant con- ¼F ;a ð7Þ
P i P i
trols on evapotranspiration. It is a holistic approach that as-
sumes the equilibrium water balance is controlled by water where F() is Eq. (5). Based on the above relationship and by
availability and atmospheric demand. The water availability further assuming negligible interannual variations in poten-
can be approximated by precipitation, the atmospheric de- tial evapotranspiration, Koster and Suarez (1999) showed
mand represents the maximum possible evapotranspiration that the ratio of the standard deviation of annual evapora-
and is often equated with potential evapotranspiration. tion to that of annual precipitation can be approximated
Milly (1993, 1994) and Zhang et al. (2001) further developed as a function of index of dryness (U ¼ E 0 =P):
the Budyko framework by introducing additional controls
such as rainfall seasonality and vegetation characteristics. rET
¼ FðU; aÞ  UF 0 ðU; aÞ ð8Þ
A similar method for modeling evapotranspiration has been rP
120 L. Zhang et al.

1.0 α = 0.80
α = 0.60

α = 0.50

0.8

α =0.33

0.6
E/P
α = 0.23

0.4

0.2 α = 0.09

0.0
0.0 1.0 2.0 3.0 4.0 5.0
Eo/P

Figure 1 Ratio of mean annual evapotranspiration to rainfall (E/P) as a function of the index of dryness (E0/P) for different values
of parameter.

1600 which is called evaporation deviation ratio, U is the index


of dryness. Large values of this ratio indicate that most of
precipitation variability becomes evapotranspiration vari-
1200 ability, whereas small values mean that evapotranspiration
variability is largely insensitive to variability in precipita-
Predicted Q (mm)

tion. From this relationship, Koster and Suarez (1999) con-


cluded that the ratio of the standard deviation of annual
800
streamflow to the standard deviation of annual precipita-
tion can also be expressed as

400 rQ rET
¼1 ð9Þ
rP rP
It should be noted that Eq. (9) was obtained from the
0 relationship: DP ¼ DQ þ DET: However, this is also depen-
0 400 800 1200 1600 dent on the assumption that annual evapotranspiration
Observed Q (mm) anomalies are fixed proportions of annual precipitation
anomalies. This equation can be used to describe interan-
0.8 nual variability of streamflow (Sankarasubramanian and Vo-
gel, 2002). One can also estimate annual streamflow from
Eq. (10) in analogy to Eq. (6) by ignoring interannual water
storage change:
0.6
 1 1
1a
Predicted Q/P

Q i ¼ P 1a
i þ E 0i
1a
 E 0i ð10Þ
0.4 Fig. 3 compares the observed and predicted values of
interannual variability of streamflow using Eq. (9) for all
the catchments with a constant value of 2.63 for a based
0.2 on Zhang et al. (2004). The coefficient of efficiency for
interannual streamflow variability is 0.87 and the mean
absolute error (MAE) is 22 mm. Over 61% of the catchments
0.0 have streamflow variability, defined as standard deviation in
0.0 0.2 0.4 0.6 0.8 annual streamflow, less than half of the average annual
Observed Q/P streamflow (i.e. 100 mm) and a few catchments showed
large variability of more than 400 mm (see Fig. 3). These re-
Figure 2 Comparison of observed and predicted (a) mean sults are consistent with the findings of Sankarasubramanian
annual streamflow and (b) runoff ratio for all the catchments. and Vogel (2002) for catchments in the US and Arora (2002)
Water balance modeling over variable time scales based on the Budyko framework – Model development and testing 121

1.0 600
Streamflow deviation ratio (σQ/σP)

143110
0.8

Observed streamflow [mm]


0.6 400

0.4

200
0.2

0.0
00 .5 11 .5 22 .5 3
Mean index of dryness (E0/P) 0
0 200 400 600
600
Predicted streamflow [mm]
Predicted streamflow variability (mm)

500 600

400 405228

300 Observed streamflow [mm] 400

200

100
200

0
0 100 200 300 400 500 600
Observed streamflow variability (mm)
0
Figure 3 Comparison of observed and predicted streamflow 0 200 400 600
deviation ratio (a) and interannual streamflow variability Predicted streamflow [mm]
calculated as standard deviation in annual streamflow (b).
800

for GCM simulations. The advantage of the method devel- 318900


Observed streamflow [mm]

oped by Koster and Suarez (1999) is that the interannual var- 600
iability in streamflow can be calculated from a simple
relationship using the index of dryness. The scatter in the
results may be mainly due to rainfall seasonality and the
omission of interannual storage variations. It has been sug- 400
gested that inclusion of a soil-moisture index can improve
the model predictions (Sankarasubramanian and Vogel,
2002, 2003). 200
The equilibrium water balance model with annual time
step works reasonably well for three selected catchments
representing different rainfall regimes (see Fig. 4) and this
indicates that the storage effect can be neglected in these 0
catchments. The model was also applied to all the catch- 0 200 400 600 800
ments to test its general applicability and the results are Predicted streamflow [mm]
summarized in Fig. 5. It can be seen that the model works
Figure 4 Comparisons of predicted and observed annual
well for most of the catchments with 90% of them having
streamflow for selected catchments, representing summer-
values of the coefficient of efficiency greater than 0.5 and
dominant rainfall (a), winter-dominant rainfall (b), and non-
less than 3% of the catchments have bias values greater than
seasonal rainfall (c).
10%. The bias in the estimated annual streamflow is within
5% for 85% of the catchments. However, there are some can not be neglected on annual time scale and more com-
catchments where the model performed poorly and this plex models are required to accurately simulate water bal-
may indicate that the impact of catchment water storage ance for these catchments.
122 L. Zhang et al.

15

10

Bias (%)
0

-5

-10

-15
0 20 40 60 80 100
Percentage of catchments where bias value exceeded

1.0

0.8
Coefficient of Efficiency (E)

0.6

0.4

0.2

0.0
0 20 40 60 80 100
Percentage of catchments where E value exceeded

Figure 5 Summary of model calibration for annual streamflow. (a) Percentage of catchments where bias value exceeded, (b)
percentage of catchments with E value greater than or equal to a given E value.

Dynamic water balance models streamflow, especially a slow flow component and it calls
for an increase in model complexity. Here the method of
Changing from annual to shorter-time scales, such as Budyko (1958) for modeling mean annual evapotranspiration
monthly or daily, it is expected that the effect of catchment is generalized as a demand and supply framework. The goal
water storage on the water balance will become even more is to develop a dynamic water balance model for sub-annual
significant. As a result, variations in rainfall, potential time scales using the demand and supply framework, or the
evapotranspiration, and water storage need to be consid- ‘‘limits’’ concept (Calder, 1998) by appropriately defining
ered. To test if the annual water balance equation repre- new state variables and fluxes.
sented by Eq. (10) is capable of accurately simulating The catchment is conceptualized as a system of two sto-
monthly streamflow and to identify any potential deficiency rages: root zone storage and groundwater storage. For the
in the model, Eq. (10) was next applied with monthly cli- catchment water balance, it is assumed that rainfall P(t)
mate data and the results are shown in Fig. 6 for three se- in time step t will be partitioned into direct runoff Qd(t)
lected catchments representing different rainfall regimes. and the sum of the other water balance components:
It is clear that the annual water balance model represented PðtÞ ¼ Q d ðtÞ þ XðtÞ ð11Þ
by Eq. (10) is not able to accurately simulate monthly
streamflow. The predicted streamflow is overly sensitive where X(t) is called catchment rainfall retention and is the
to rainfall with sharp rises and falls in the hydrographs. As amount of rainfall retained by the catchment for evapo-
a result, the model over-predicted streamflow due to a lack transpiration ET(t), change in soil-moisture storage
of storage control and slow release processes. The results S(t)  S(t  1) and recharge R(t).
shown in Fig. 6 suggest that additional processes have to The demand limit for X(t) is the sum of available storage
be incorporated in order to simulate the observed monthly capacity (Smax  S(t  1)) and potential evapotranspiration
Water balance modeling over variable time scales based on the Budyko framework – Model development and testing 123

600 500 0
143110

400
Observed streamflow (mm)

200
400

Precipitation (mm)
Streamflow (mm)
300

400

200
200

600
100

0
0 200 400 600 0 800
Simulated streamflow (mm) 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984
Time

200 300 0
405228

150
Observed streamflow (mm)

200 200

Precipitation (mm)
Streamflow (mm)

100

100 400
50

0
0 50 100 150 200 0 600
Simulated streamflow (mm) 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983
Time

300 400 0
318900

250
Observed streamflow (mm)

300 200
200

Precipitation (mm)
Streamflow (mm)

150
200 400

100

50 100 600

0
0 50 100 150 200 250 300
0 800
Simulated streamflow (mm)
1984 1985 1986 1987 1988 1989 1990 1991 1992 1993
Time

Figure 6 (a) Scatter plots of observed and simulated monthly streamflow for the whole period of records using Eq. (10) for three
selected catchments representing summer-dominant rainfall (143110), winter-dominant rainfall (405228), and uniform rainfall
regime (318900); (b) time series of observed ( ) and simulated (ÆÆÆÆÆ) monthly streamflow of 10 year windows for the same
catchments. Also shown is monthly rainfall distribution.

 
(E0(t)) and is denoted as X0(t), while the supply limit can be X 0 ðtÞ
XðtÞ ¼ PðtÞF ; a1 ð14Þ
considered as rainfall P(t). Following a similar argument to PðtÞ
Budyko (1958), we can postulate that:
XðtÞ=PðtÞ ! 1 as X 0 ðtÞ=PðtÞ ! 1ðvery dry conditionsÞð12Þ where F() is Fu’s curve – Eq. (5), a1 is retention efficiency,
XðtÞ ! X 0 ðtÞ as X 0 ðtÞ=PðtÞ ! 0ðvery wet conditionsÞ ð13Þ i.e., a larger a1 value will result in more rainfall retention
and less direct runoff. The ratio X0(t)/P(t) in Eq. (14) is anal-
Note that these are analogous to the supply and demand ogous to Budyko’s dryness index (Fig. 1). It can be consid-
limits (i.e. dashed lines) in Fig. 1. The catchment rainfall ered as a demand/supply index for the partition of P.
retention X(t) can be calculated as From Eqs. (11) and (14), direct runoff is calculated as
124 L. Zhang et al.

Q d ðtÞ ¼ PðtÞ  XðtÞ ð15Þ where a2 is a model parameter, representing evapotranspi-


ration efficiency. It should be noted that Eq. (22) is identical
At sub-annual time scales, water availability W(t) can be
to Fu’s curve except that rainfall has been replaced with the
defined as
available water W(t), which takes into consideration of the
WðtÞ ¼ XðtÞ þ Sðt  1Þ ð16Þ effect of catchment water storage.
Combining the definition of X(t) with Eq. (16), one It can be noted that Eq. (21) shares a common model
obtains parameter a2 with Eq. (22) and this is because groundwater
recharge is essentially determined by evapotranspiration
WðtÞ ¼ ETðtÞ þ SðtÞ þ RðtÞ ð17Þ efficiency. That is, as evapotranspiration efficiency is great-
While Eq. (16) defines the source of the water availabil- er (e.g. larger values of a2), recharge is diminished. Soil
ity, Eq. (17) determines the partitioning. Next define evapo- water storage can now be calculated as
transpiration opportunity (Sankarasubramanian and Vogel, SðtÞ ¼ YðtÞ  ETðtÞ ð23Þ
2002) as YðtÞ ¼ ETðtÞ þ SðtÞ, we obtain:
Note also that the parameter in Eqs. (21) and (22) must be
WðtÞ ¼ YðtÞ þ RðtÞ ð18Þ identical to ensure that catchment water storage S(t) re-
The demand limit for Y(t) can be considered as the sum mains between 0 and Smax.
of potential evapotranspiration (E0(t)) and soil water stor- Finally, groundwater storage is treated as a linear reser-
age capacity (Smax) and is denoted as Y0(t), while the supply voir, so that the groundwater balance and baseflow can be
limit is the available water W(t). Similar to Budyko (1958), modeled as
we can postulate that:
Q b ðtÞ ¼ dGðt  1Þ ð24Þ
YðtÞ=WðtÞ ! 1as Y 0 ðtÞ=WðtÞ ! 1ðvery dry conditionsÞ
GðtÞ ¼ ð1  dÞGðt  1Þ þ RðtÞ ð25Þ
ð19Þ
YðtÞ ! Y 0 ðtÞ as Y 0 ðtÞ=WðtÞ ! 0ðvery wet conditionsÞ where Qb is baseflow, G is groundwater storage, and d is a
ð20Þ constant. Eq. (24) represents a linear storage–discharge
relationship and constant d is a model parameter. A sche-
The evapotranspiration opportunity Y(t) can be esti- matic diagram of the dynamic water balance model with
mated from the following relationship: limits demand and supply is shown in Fig. 7.
 
E 0 ðtÞ þ Smax
YðtÞ ¼ WðtÞF ; a2 ð21Þ
WðtÞ Model calibration and validation
Thus groundwater recharge R(t) can be calculated from
The dynamic water balance model described above has
Eq. (18). The next step is to calculate evapotranspiration
four parameters a1, a2, d, and Smax. Application of this
ET(t). The demand limit for ET(t) can be considered as po-
model requires estimates of these parameters and the
tential evapotranspiration E0(t) and the supply limit is the
model was calibrated against recorded streamflow at both
available water W(t). Similar to Budyko (1958), evapotrans-
daily and monthly time scales. The split-sample test was
piration ET(t) can be calculated as
  applied where the observed streamflow was divided into
E 0 ðtÞ two roughly equal parts for calibration and validation
ETðtÞ ¼ WðtÞF ; a2 ð22Þ
WðtÞ (Refsgaard and Knudse, 1996).

Qd(t)/P(t) R(t)/W(t)
Partition of water availability
Partition of rainfall

S(t)/W(t)

Y(t)/W(t)
X(t)/P(t) ET(t)/W(t)

X0(t)/P(t) Demand/supply index E0(t)/W(t) (E0(t)+Smax)/W(t) Demand/supply index

Figure 7 A schematic diagram of the dynamic water balance model. Left: Partitioning of rainfall P(t) into catchment rainfall
retention X(t) and direct runoff Qd(t). The demand and supply limits are shown as dashed lines. Right: Partitioning of water
availability W(t) into evapotranspiration ET(t), storage S(t) and recharge R(t). The demand/supply index for evapotranspiration
opportunity Y(t) is (E0(t) + Smax)/W(t), where E0(t) is potential evapotranspiration. The demand/supply index for evapotranspiration
is E0(t)/W(t).
Water balance modeling over variable time scales based on the Budyko framework – Model development and testing 125
Pn
A generalised pattern search method was applied for ðQ sim;i  Q obs;i Þ2
parameter optimization (TheMathWorksInc., 2006). Four ba- F 2 ¼ Pi¼1
n 2
ð27Þ
i¼1 ðQ obs;i  Q obs 
sic objective functions are considered in this study:
Pn Pn
½lnðQ sim;i Þ  lnðQ obs;i Þ2 i¼1 ðQ sim;i  Q sim ÞðQ obs;i  Q obs Þ
F 1 ¼ Pi¼1 ð26Þ F 3 ¼ 1  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn ffi ð28Þ
n 2 2 Pn 2
i¼1 ½lnðQ obs;i Þ  lnðQ obs Þ i¼1 ðQ sim;i  Q sim Þ i¼1 ðQ obs;i  Q obs Þ

500
600
143110 observed simulated
143110

400
Observed streamflow (mm)

400

Streamflow (mm)
300

200
200

100

0
0 200 400 600 0
Simulated streamflow (mm) 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984

200 150
405228 405228 observed simulated

150
Observed streamflow (mm)

100
Streamflow (mm)

100

50
50

0
0 50 100 150 200 0
Simulated streamflow (mm) 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983

300 400
318900 observed simulated
318900
250
Observed streamflow (mm)

300
200
Streamflow (mm)

150
200

100

50 100

0
0 50 100 150 200 250 300
Simulated streamflow (mm) 0
1984 1985 1986 1987 1988 1989 1990 1991 1992 1993

Figure 8 (a) Scatter plots of observed and simulated monthly streamflow for the whole period of records using the dynamic water
balance model for three selected catchments representing summer-dominant rainfall (143110), winter-dominant rainfall (405228),
and uniform rainfall regime (318900); (b) time series of observed and simulated monthly streamflow of 10 year windows for the same
catchments.
126 L. Zhang et al.
Pn !
sidered as a holistic or systems approach to catchment
i¼1 Q sim;i
F 4 ¼ ln Pn ð29Þ water balance modeling and it is expected to achieve parsi-
i¼1 Q obs;i

mony of model parameters. A number of studies concluded
where Qsim,i is the simulated streamflow at time i, Qobs,i is that water balance models with three to five parameters are
the observed streamflow at time i, n is the number of time generally capable of accurately simulating monthly stream-
steps in the calibration period. F1 uses a logarithmic func- flow for humid regions (Xu and Singh, 1998; Vandewiele and
tion and weights the error in low flows more than the errors Ni-Lar-Win, 1998). However, the prospect of these models
in the high flows. F2 was selected to place emphasis on the for arid areas is less encouraging (Xu and Singh, 1998). This
errors in high flows. F3 describes the time shift between the can be chiefly attributed to extremely high temporal and
simulated and observed streamflow. F4 was selected to en- spatial variability in rainfall and occurrence of Hortonian
sure mass balance over the period of the calibration. Final- runoff in arid areas, particularly for summer-dominant rain-
ly, the average of these four objective functions was used as fall regimes (Potter et al., 2005).
the objective function to be optimized. The main differences among various water balance mod-
els pertain to the treatment of evapotranspiration, runoff
generation, and groundwater recharge. The water balance
Results and discussion model developed in this study has several new features.
At monthly time step, most water balance models assume
One of the objectives of this study was to demonstrate how that evapotranspiration is equal to potential evapotranspi-
dynamic water balance can be modeled based on the top- ration when rainfall is greater than potential evapotranspi-
down approach by extending the Budyko framework or the ration (Thornthwaite and Mather, 1955; Wolock and
‘‘limits’’ concept (Calder, 1998). This method can be con- McCabe, 1999). Given that rainfall and potential evapo-

1.0

0.8
Coefficient of Efficiency (E)

0.6

0.4

0.2 Calibration
Validation

0.0
0 20 40 60 80 100
Percentage of catchments where E value exceeded

0.25
winter-dominant
non-seasonal
0.20 summer-dominant
Frequency [-]

0.15

0.10

0.05

0.00
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Coefficient of Efficiency

Figure 9 Percentage of catchments where E values greater than or equal to a given E value (a) and distribution of E values across
three different rainfall regimes (b) at monthly time step.
Water balance modeling over variable time scales based on the Budyko framework – Model development and testing 127

transpiration vary within a month, there exist periods of less than potential evapotranspiration. The critical soil-
deficit and surplus. As a result, this assumption may lead moisture content varies considerably with climate, soil,
to overestimation of evapotranspiration. On the other hand, and vegetation.
when rainfall is less than potential evapotranspiration, Evapotranspiration is treated differently in this study. It
evapotranspiration is generally modeled as a ‘‘ramp’’ func- is modeled using a method similar to that of Budyko (1958)
tion of soil moisture and various relationships have been for mean annual evapotranspiration. The model assumes
developed (Potter, 2006). A common feature of these rela- evapotranspiration will always be less than the potential
tionships is the specification of the critical soil-moisture evapotranspiration, except for the case when a2 ! 1. A lar-
content, defined as the level of soil moisture below which ger a2 value indicates higher evapotranspiration efficiency.
capillary tension restricts actual evapotranspiration to be Sensitivity analyses performed by Zhang et al. (2004)

Figure 10 Comparison of observed and simulated daily flow duration curves in the selected catchments for the calibration period
(a–c) and validation period (d–f).
128 L. Zhang et al.

suggest that a 11% increase in parameter a2 (which is equiv- recharge does not occur until soil-moisture capacity is filled
alent to a 20% increase in parameter w) results in about a (Alley, 1984). This is important for modeling water balance
10% increase in mean annual evapotranspiration for typical in arid regions such as some of the catchments selected in
fitted values of parameter a2. this study. The treatment of recharge in the dynamic water
Another feature of the model developed in this study is balance model is comparable to that of the abcd model
the estimation of runoff and groundwater recharge. It as- (Alley, 1984), but the relationships between evapotranspira-
sumes that direct runoff and recharge can occur even when tion and soil-moisture content are treated differently and
the soil is not saturated and it represents an improvement of the model allows both linear and non-linear relationships
the models based on the threshold concept that runoff or to be modeled.

1.0

0.8
Coefficient of Efficiency (E)

0.6

0.4

0.2 Calibration

Validation

0.0
0 20 40 60 80 100
Percentage of catchments where E value exceeded

1.0
Coefficient of Efficiency (E)

0.8

0.6

0.4

0.2 Calibration

Validation

0.0
0 20 40 60 80 100
Percentage of catchments where E value exceeded

Figure 11 Percentage of catchments where E values greater than or equal to a given E value for (a) ranked daily streamflow and
(b) time series of daily streamflow in the calibration and validation periods.

Table 1 Comparison of model performance in predicting streamflow at annual, monthly, and daily time scales for three
selected catchments
Catchment Coefficient of efficiency Bias (%)
Annual Monthly Daily Annual Monthly Daily
143110 0.82 0.90 0.63 3.39 14.47 21.76
318900 0.80 0.89 0.64 0.51 5.49 14.28
405228 0.82 0.84 0.57 1.06 12.04 24.87
Water balance modeling over variable time scales based on the Budyko framework – Model development and testing 129

An important aspect of this study was to explore the teristics (e.g. flow duration curves). The E values between
trade-off between parameter identifiability and model per- the simulated and observed daily streamflow time series
formance to produce an optimum model for Australian con- are generally lower compared with the E values for the
ditions. Thus, the focus was on maximizing identifiability, so ranked streamflows (Fig. 11b). The results shown in Fig.
that model parameters can be related to catchment charac-
teristics. This is important if the models are going to be used
for predictions of ungauged catchments or of the impact of 0.50
land-use change, for example. Based on these consider- Monthly

ations, the dynamic water balance model developed in this 0.40


Monthly
Daily
study maintains a minimum level of complexity while still
describing the effects of the underlying processes. 0.30

Frequency
The use of Fu’s curve provides a consistent framework
for the modeling of different processes and results in a sim- 0.20

ple model structure, and each of the parameters has a phys-


ical interpretation. For example, Smax is the maximum soil 0.10

water storage in the root zone and it relates to soil and veg-
etation characteristics. The parameter a1 represents catch- 0.00

ment rainfall retention efficiency, i.e., a larger a1 value will

=5
0.

1.

2.

3.

4.
5-

5-

5-

5-

5-
0-

1-

2-

3-

4-
0.

1.

2.

3.

4.
result in less direct runoff and more residual rainfall. The Parameter α 1

parameter a2 is related to evapotranspiration efficiency


and a larger a2 value indicates a higher partitioning of avail- 0.40

Monthly
able water into evapotranspiration. Finally, parameter d is
Daily
commonly referred as the recession constant (Nathan and 0.30 Monthly
McMahon, 1990). Frequency Daily

Results of the model calibration at monthly time step are


0.20
shown in Figs. 8 and 9. It is clear that the model can be well
calibrated against the water balance criteria mentioned be-
fore for a large number of catchments with different rainfall 0.10

regimes. Significant improvement has been made in the sim-


ulated monthly streamflow compared with the results 0.00
shown in Fig. 6 for the three selected catchments (Fig. 8)
5

=5
0.

1.

2.

3.

4.
5-

5-

5-

5-

5-
0-

1-

2-

3-

4-
0.

1.

2.

3.

4.
with an average E value of 0.90. No systematic error has Parameter α 2
been found in the simulated monthly streamflow and the
deficiencies identified in the annual water balance model 0.25
have mostly been overcome in the dynamic water balance Monthly
Daily
model. The majority of catchments have values of the coef- 0.2
Monthly
ficient of efficiency ranging between 0.6 and 0.9 with less Daily

than 4% of the catchments having calibrated E values less 0.15


Frequency

than 0.5 (Fig. 9a). These results indicate that that the dy-
namic water balance model can be accurately calibrated 0.1

for most Australian catchments at monthly time step. The


0.05
validation results shown in Fig. 9 indicate that over 95% of
the catchments have E values greater than 0.43; over 80%
0
of the catchments have E values greater than 0.60; and over
0

0
50

0
15

20

25

30

35

40

45

50
0

47% of the catchments have E values greater than 0.75. Run-


0-

-1

=5
0-

0-

0-

0-

0-

0-

0-

0-
50

10

15

20

25

30

35

40

45

off estimates with these threshold E values can be consid- Parameter S max

ered reasonable, acceptable, and good according to Chiew


0.80
and McMahon (1993). These results are very similar to the
Monthly
modeling results for Australian catchments using a 7 param- Daily
eter rainfall–runoff model (Chiew et al., 2002). There were 0.60 Monthly
Daily
few systematic differences in the coefficient of efficiency
Frequency

for catchments across three rainfall regimes (see Fig. 9b).


0.40
Fig. 10 shows comparisons of the simulated and observed
daily flow duration curves for the selected catchments.
They demonstrate that the dynamic water balance model 0.20

provides a reasonably good fit to the flow duration curve


based on observed daily streamflow. The E values of the 0.00
ranked streamflows are very high for most of the catch- =0 0.
1
0.
2
0.
3
0.
4
0.
5
0.
6
0.
7
0.
8
0.
9
1.
0 =1
0- 1- 2- 3- 4- 5- 6- 7- 8- 9-
0. 0. 0. 0. 0. 0. 0. 0. 0.
ments in the calibration period (Fig. 11a) and the model Parameter d
performed less well in the validation period as expected
(Fig. 11a). The dynamic water balance model is considered Figure 12 Distribution of model parameter values at monthly
capable of accurately simulating daily streamflow charac- and daily time steps.
130 L. Zhang et al.

11b indicate that routing of the simulated daily runoff may important. One can argue that more complicated models
be necessary, especially for large catchments. would be required. Results from this study showed that the
Figs. 5, 9 and 11 together with the results shown in Table interannual variability of water balance in many Australian
1 indicate that the water balance model performed best at catchments can be estimated from the index of dryness using
monthly time scale in terms of E values and bias. At annual the method of Koster and Suarez (1999). Predicted annual
time scale, the E values are consistently high, but lower E streamflow using this method showed good agreement with
values are observed at daily time scale. Based on these re- the observed streamflow for some selected catchments,
sults, it can be argued that monthly time scale is appropri- indicating minimum effect of storage change and rainfall
ate for modeling water balance for the catchments seasonality in these catchments. However, the method per-
concerned. The relatively low E values at daily time scale formed poorly in other catchments. At shorter-time scales,
may be due to lack of flow routing in the model. one has to take into consideration of the effects of soil-mois-
Fig. 12 shows the distribution of the calibrated model ture dynamics on water balance and hence increased model
parameter values at monthly and daily time scales. Of the complexity is required. Following the top-down approach,
four model parameters, a2 showed the highest consistency we extended the mean annual water balance model of Fu
at the two time scales and the parameter reached its limit (1981), a Budyko-like model, to include additional processes
imposed in the calibration in 30% of catchments. The mode and factors. The resulting model is a parsimonious dynamic
of the distribution of the parameter a1 is 1.5–2 at daily time water balance model. The model was calibrated and cross
scale and is increased to 2.5–3 at monthly time scale. These validated against observed streamflow data at both daily
results indicate that less direct runoff is simulated by the and monthly time scales. The results showed that the dy-
model at monthly time scale. This is most likely due to the namic water balance model can be well calibrated and is
decrease in variability of daily rainfall caused by the aggre- capable of accurately simulating monthly streamflow time
gation of the rainfall input to monthly time scale. The series and daily flow duration curves.
parameter Smax exhibited a large range from 50 to 500 mm, Understanding individual hydrological processes and
and its distributions are similar, except for low values. The their relationships with climate and catchment characteris-
distribution of the parameter d shows the largest difference tics is an important step in predicting catchment water bal-
at the two time scales. The fact that the a1, a2 and Smax ance. However, this does not mean that one should try to
parameters are broadly consistent gives confidence in the incorporate every known process into a model. The interac-
identifiability of these parameters. The calibrated soil-mois- tion and co-evolution of these processes may manifest
ture capacity of a catchment Smax should not be dependent themselves in such a way that the overall behavior of the
on the timescale of interest. The retention efficiency a1 catchment can be described by simple relationships. The
and the evapotranspiration efficiency a2 should also not be strength of purpose-derived relationships, such as Budyko
dependent on timescale of calibration as, for example, the (1958) and Fu (1981), is to bypass the need to describe each
average fraction of rainfall that becomes direct runoff of these processes in detail. Rather, they try to encapsulate
(determined by the retention efficiency parameter) should the combined effect of these competing processes within a
scale from daily to monthly timescale. The recession con- parameter space that describes catchments. The develop-
stant d, however, is directly dependent on the timescale ment of the dynamic water balance model followed the
of interest through Eqs. (24) and (25). The larger spread of same path and this proved to be beneficial with respect to
calibrated parameters at monthly timescale indicates that model parsimony and robustness.
the representation of groundwater and baseflow by Eqs.
(24) and (25) is better suited for monthly timescale.

Acknowledgments
Summary and conclusions
The work was supported by the CSIRO Flagship Program
Catchment water balance is controlled by climate variables (Water for a Healthy Country T3.S3.P3: River Murray Up-
and characteristics of the catchment. These factors exhibit lands). We would like to thank John Gallant and Trevor Dow-
considerable spatial and temporal variability and interact ling for providing estimates of MrVBF and Peter Briggs for
with each other. As a result, the degree of control exerted providing potential evapotranspiration data. We would like
by individual factors on water balance varies with the tem- to thank Warrick Dawes, Richard Morton, and three anony-
poral scales considered. We used a top-down approach to ex- mous reviewers for their helpful comments.
plore the effects of these factors on catchment water
balance over variable time scales. At mean annual scale,
the water balance is dominated by climatic factors such as References
average precipitation and potential evapotranspiration.
The conceptual thinking embedded in the Budyko framework Alley, W.M., 1984. On the treatment of evapotranspiration, soil
moisture accounting, and aquifer recharge in monthly water
has led to the development of Budyko-like models and they
balance models. Water Resour. Res. 20, 1137–1149.
proved to be accurate for describing mean annual water bal-
Arora, V.K., 2002. The use of the aridity index to assess climate
ance under a range of climatic and land-use conditions. The change effect on annual runoff. J. Hydrol. 265, 164–177.
key to the success of Budyko-like models is to capture the Atkinson, S.E., Woods, R.A., Sivapalan, M., 2002. Climate and
dominant controls on water balance at the expense of minor landscape controls on water balance model complexity over
processes. For example, at annual time scale, the effects of changing timescales. Water Resour. Res. 38 (12), 1314.
catchment storage and rainfall seasonality may become doi:10.1029/2002WR00148.
Water balance modeling over variable time scales based on the Budyko framework – Model development and testing 131

Budyko, M.I., 1958. The Heat Balance of the Earth’s Surface. US capacity on mean annual water balance for Australian catch-
Department of Commerce, Washington, DC. ments. Water Resour. Res 41, W06007. doi:10.1029/
Calder, I.R., 1998. Water use by forests, limits and controls. Tree 2004WR00369.
Physiol. 18, 625–631. Priestley, C.H.B., Taylor, R.J., 1972. On the assessment of the
Chiew, F.H.S., McMahon, T.A., 1993. Assessing the adequacy of surface heat flux and evaporation using large-scale parameters.
catchment streamflow yield estimates. Aust. J. Soil Res. 31, Mon. Weather Rev. 100, 81–92.
665–680. PUB initiative (<http://www.cig.ensmp.fr/~iahs/>).
Chiew, F.H.S., Peel, M.C., Western, A.W., 2002. Application and Raupach, M.R., Kirby, J.M., Barrett, D.J., Briggs, P.R., Lu, H.,
testing of the simple rainfall–runoff model SIMHYD. In: Singh, Zhang, L., 2001. Balances of water, carbon, nitrogen and
V.P., Frevert, D.K. (Eds.), Mathematical Models of Small phosphorus in Australian landscapes: (2) model formulation
Watershed Hydrology & Applications. Water Resour. Pub., and testing. Tech. Rep. CSIRO Land and Water, Canberra, ACT,
Colorado, pp. 335–367. Australia, 41/01.
Choudhury, B.J., 1999. Evaluation of an empirical equation for Refsgaard, J.C., Knudse, J., 1996. Operational validation and
annual evaporation using field observations and results from a intercomparison of different types of hydrologic models. Water
biophysical model. J. Hydrol. 216, 99–110. Resour. Res. 36, 1495–1510.
Farmer, D., Sivapalan, M., Jothityangkoon, C., 2003. Climate, soil, Sankarasubramanian, A., Vogel, R.V., 2002. Annual hydroclimatol-
and vegetation controls upon the variability of the water ogy of the United States. Water Resour. Res. 38, 1083.
balance in temperate and semiarid landscapes: downward doi:10.1029/2001WR00061.
approach to water balance analysis. Water Resour. Res 39 (2). Sankarasubramanian, A., Vogel, R.V., 2003. Hydroclimatology of
doi:10.1029/2001WR00032. the continental United States. Geophys. Res. Lett. 30 (7), 1363.
Fu, B.P., 1981. On the calculation of the evaporation from land doi:10.1029/2002GL01593.
surface. Sci. Atmos. Sin., 23–31 (in Chinese). Sivapalan, M., Blöschl, G., Zhang, L., Vertessy, R., 2003. Downward
Hickel, K., Zhang, L., 2006. Estimating the impact of rainfall approach to hydrological prediction. Hydrol. Process., 2101–
seasonality on mean annual water balance using a top-down 2111.
approach. J. Hydrol. 331, 409–424. TheMathWorksInc., 2006. Genetic algorithm and direct search
Koster, R.D., Suarez, M.J., 1999. A simple framework for examining toolbox 2 User’s Guide, Natick, MA.
the interannual variability of land surface moisture fluxes. J. Thornthwaite, C.W., Mather, J.R., 1955. The water balance. Publ.
Clim. 12, 1911–1917. Climatol. Lab. Climatorl. Drexel Inst. Technol. 8 (1), 1–104.
Milly, P.C.D., 1993. An analytic solution of the stochastic storage Vandewiele, G.L., Ni-Lar-Win, G.L., 1998. Monthly water
problem applicable to soil water. Water Resour. Res. 29, 3755– balance models for 55 basins in 10 counties. Hydrol. Sci.
3758. J. 43, 687–699.
Milly, P.C.D., 1994. Climate, soil water storage, and the average Walker, G.R., Zhang, L., 2002. Plot-scale models and their appli-
annual water balance. Water Resour. Res. 30, 2143–2156. cation to recharge studies. In: Studies in Catchment Hydrology –
Nathan, R.J., McMahon, T.A., 1990. Evaluation of automated The Basics of Recharge and Discharge. CSIRO Publishing,
techniques for base flow and recession analyses. Water Resour. Australia.
Res. 26 (7), 1465–1473. Wolock, D.M., McCabe, G.J., 1999. Explaining spatial variability in
Peel, M.C., Chiew, F.H.S., Western, A.W., McMahon, T.A., 2000. mean annual runoff in the conterminous United States. Clim.
Extension of unimpaired monthly stream flow data and region- Res. 11, 149–159.
alization of parameter values to estimate stream flow in Xu, C.Y., Singh, V.P., 1998. A review of monthly water balance
ungauged catchments. Report to National Land and Water models for water resources investigations. Water Resour. Man-
Resources Audit, Cent. For Environ. Appl. Hydrol., Univ. of age. 12, 31–50.
Melbourne, Parkville, Vict., Australia. Zhang, L., Hickel, K., Dawes, W.R., Chiew, F.H.S., Western, A.W.,
Potter, N.J., 2006. Statistical-dynamical modeling of catchment Briggs, P.R., 2004. A rational function approach for estimating
water balance: climatic and vegetation controls on hydrological mean annual evapotranspiration. Water Resour. Res. 40,
fluxes. PhD Thesis. The University of Melbourne, Parkville, Vic., W02502. doi:10.1029/2003WR00271.
Australia, 305pp. Zhang, L., Dawes, W.R., Walker, G.R., 2001. Response of mean
Potter, N.J., Zhang, L., Milly, P.C.D., McMahon, T.A., Jakeman, annual evapotranspiration to vegetation changes at catchment
A.J., 2005. The effects of rainfall seasonality and soil moisture scale. Water Resour. Res. 37, 701–708.

You might also like