You are on page 1of 17

Journal

of
Hydrology
ELSEVIER Journal of Hydrology 201 (1997) 272-288

Development and test of the distributed HBV-96


hydrological model
G6ran Lindstr6m*, Barbro Johansson, Magnus Persson, Marie Gardelin,
Sten Bergstr6m
Swedish Meteorological and Hydrological Institute, S-60l 76 NorrkOping, Sweden
Received 11 October 1996; accepted 6 March 1997

Abstract

A comprehensive re-evaluation of the HBV hydrological model has been carried out. The objec-
tives were to improve its potential for making use of spatially distributed data, to make it more
physically sound and to improve the model performance. The new version, HBV-96, uses subbasin
division with a typical resolution of 40 km 2, although any resolution can be used. In addition, each
subbasin is divided into elevation bands, vegetation and snow classes. Automatic weighting of
precipitation and temperature stations was introduced and a new automatic calibration scheme
was developed. The modifications led to significant improvements in model performance. In
seven test basins the average value of the efficiency criterion R z increased from 86 to 89%, with
improvements in both the calibration and verification periods. © 1997 Elsevier Science B.V.

Keywords: HBV-96 Hydrological model; Subbasin; Automatic calibration scheme; Model


performance

1. B a c k g r o u n d

The HBV hydrological model has a long history and the model has found applications in
more than 30 countries. Its first application dates back to the early 1970s (Bergstr6m and
Forsman, 1973). Originally the HBV model was developed at the Swedish Meteorological
and Hydrological Institute (SMHI) for runoff simulation and hydrological forecasting, but
the scope of applications has increased steadily (Bergstr6m, 1995). The model has also
been subject to modifications over time, although the basic modelling philosophy has been
unchanged. This philosophy acknowledges that the model complexity and data demand
must not be in conflict with the operational requirements and can be formulated as follows:

* Correspondingauthor.

0022-1694/97/$17.00 © 1997- Elsevier ScienceB.V. All rights reserved


PII S0022- 1694(97)00041-3
G. LindstrOm et al./Journal of Hydrology 201 (1997) 272-288 273

• the model shall be based on a sound scientific foundation;


• data demands must be met in typical basins;
• the model complexity must be justified by model performance;
• the model must be properly validated;
• the model must be understandable by users.
Today many versions of the HBV model exist, and new codes are constantly being
developed by different groups (see for example Vehvilfiinen, 1986; Killingtveit et al.,
1990; Renner and Braun, 1990). The standard at SMHI has long been a version which
is best characterised as a semi-distributed conceptual model. Experience has shown,
however, that this version has some major drawbacks concerning areal representation, a
fact which limits the use of distributed data. There are also a number of physical
inconsistencies in this commonly used model, such as the lack of an interception routine
and the lack of an elevation correction of evapotranspiration. These inconsistencies
became questionable when the model was to be used for climate impact studies. Finally,
the response function of the model gave rise to some concern when the model became the
standard tool for spillway design flood studies of the hydropower system in Sweden
(Bergstr6m et al., 1992).
In 1993 the Swedish Association of River Regulation Enterprises (VASO) and the
SMHI initiated a major revision of the structure of the HBV model. The objective of
the work was to re-evaluate the existing model and to develop a new model version for
hydrological problems related to hydropower production and design. The model should be
based on the same philosophy of simplicity as the original HBV model, but it should
be more physically reasonable and up-to-date with the current hydrological and
meteorological knowledge. HBV-96 is the final result of this model revision.

2. Methods and data

The core of the project was a re-evaluation of the model routines in the original HBV
model: the routines for input integration, snow accumulation and melt, evapotranspiration,
soil moisture, runoff response and routing. The effects of model resolution in both space
and time were also studied. As far as possible, each modification of the model was tested
by comparison with a reference model simulation by the standard model in a set of test
basins. The model performance was measured over both a calibration period and a
verification period of independent data. The main criterion of model performance was
the commonly used R 2 (Nash and Sutcliffe, 1970):

[Qcom(i) - Qrec(i)] 2
R 2 = l _i=1
[Qrec(i) - Qr~c]2
i=1

where Qcom and Qrec are the computed and recorded hydrographs respectively. All
calibrations were made automatically, by either the method developed by Harlin (1991)
or a further development of that method. It would have been almost impossible to make all
274 G. Lindstrfim et al./Journal of Hydrology 201 (1997) 272-288

the tests by manual calibration. Another advantage with the automatic calibration is
enhanced objectivity, i.e. one is less tempted to calibrate a preferred model version
more carefully than another.
A set of data from ten basins in Sweden was used in various stages of the re-evaluation
process (Fig. 1) and seven of these basins were used in a final validation of the new model.
In most basins the modelled period was 1969-1989, with 10 years for calibration and
10 years for verification.

3. The standard HBV model

The standard HBV model at SMHI was used as a reference in the re-evaluation process.
It has been described in a number of publications (see for example Bergstrrm, 1976 or
Bergstrrm, 1995). The large number of applications of this model, under various physio-
graphic and climatological conditions, has shown that its structure is very robust and
surprisingly general, in spite of its relative simplicity. The number of free parameters,

tsin Area Lakes Forest Calibration Verification


(km2) (%) (%) period period
Iorva 4688 14 3 1979-89 1969-79
altsjOn 1109 7 28 1969-79 1979-89
rSms Vattudal 3851 12 67 1979-89 1969-79
~rrSn 1366 13 24 1969-79 1979-89
~rpshammar 4291 10 74 1979-89 1969-79
usnedal 343 0 28 1969-79 1979-89
ankastr6m 3446 7 72 1979-89 1969-79
~rsebro 3676 9 59 1969-79 1979-89
~rnestorp 174 0 33 1969-79 1979-89
51jes 5975 8 46 1969-79 1979-89

Fig. 1. Geographical location of the basins used in the study and some physiographic data.
G. LindstrOm et al./Journal of Hydrology 201 (1997) 272-288 275

subject to calibration, is moderate, but large enough to give freedom for adjustment under
different conditions.
A recent application of the HBV model to the entire Baltic Sea basin (Bergstr6m et al.,
1996) has shown that the model can be applied to a wide range of scales without
modification of its structure. This can be explained by the simple fact that a large basin
is just a sum of several small ones, but also by the pragmatic model philosophy of the HBV
model and its semi-empirical structure. The best illustration of this is how the soil moisture
accounting is solved.
Soil moisture dynamics is a complex process which requires complex models to be
described in detail. If the problem is limited to modelling of the effects of soil moisture on
runoff generation on the basin scale, the problem can be greatly simplified. Often a bucket
approach is chosen to represent the field capacity and thus the storage capacity of the soil.
It is clear, however, that this approach is too crude and gives a response that is too
categorical. The soil moisture accounting of the HBV model is based on a modification
of the bucket theory in that it assumes a statistical distribution of storage capacities in a
basin. This simple assumption has followed the model ever since its introduction in 1972
and has proved to be very important, as it makes the model independent of scale as long as
this distribution function is stable.
The HBV model has gradually been developed into a semi-distributed model. This
means that a basin may be separated into a number of subbasins and that each one of
these is distributed according to elevation and vegetation. Lakes have a significant impact
on runoff dynamics and the routing in major lakes is, therefore, modelled explicitly. The
presence of major lakes is, therefore, an important factor when choosing substructure
based on subbasins.

4. Model re-evaluation

The re-evaluation of the model was carried out in steps followed by validation and
comparison with the results of the standard model. This process included a vast number of
tests, modifications and rejections of hypotheses. The effects of a number of additional
input variables were also investigated, such as wind speed, cloud cover and humidity. It is
not feasible to give all details of the re-evaluation process in this paper, but the most
important results are summarised below. The reader is referred to the work by Lindstr6m
et al. (1996) for further details.

4.1. Spatial resolution

A finer resolution in space was introduced to provide for better use of spatial data, either
as input to the model or for validation and updating. The new model is based on a subbasin
division according to the digitised standard drainage basins at SMHI. This corresponds to a
spatial resolution of some 40 km 2. As in the standard model, these subbasins are further
divided into zones according to elevation and vegetation. According to the test runs, the
increased resolution in space had little effect on the model performance. Its main potential
lies in updating of the model, for instance by comparison with satellite images.
276 G. Lindstrfm et al./Journal of Hydrology 201 (1997) 272-288

4.2. Resolution in time

The standard HBV model is normally operated on daily time steps, but it is often
believed that this is the cause of some errors in the discrimination of snow and rain
which, in turn, may result in errors in the simulation of runoff. Test runs made with
time steps of 12 h, however, did not lead to any significant improvements compared
with the traditional daily values in the selected basins. One reason for this might be that
the daily data are often of a higher quality than those with a shorter time step. In spite of
this, the option to use higher resolution will remain in the new model version.

4.3. Optimal interpolation of precipitation and temperature

The standard model uses a rather crude weighting routine and lapse rates for computa-
tion of areal precipitation and air temperatures. In HBV-96 a more sophisticated method,
based on optimal interpolation (see for example Daley, 1991) was introduced. This
method is frequently used in meteorological applications, and aims at minimising the
estimation error:

E = Y~(P-Pcom) 2
t

k
Pcom = P ' + y~ wi(Pi, obs-Pi)
i=l

where E is the error in estimated precipitation/temperature, summarised over a number of


time steps t, P is the actual precipitation/temperature for the selected area, Pcom is the
estimated precipitation/temperature for the selected area, P' and PI are first guesses of
the precipitation/temperature for the selected area and for the station i respectively, Pi.obs is
the observed precipitation/temperature for the station i, and wi is the weight for station i.
The derivation of E with respect to wi gives the following system of k equations:
k
"~. [ c o v ( P i , Pj) + c o v ( O i , Oj)lwj --- C o y ( P , Pi)
j=l

i = I .....k

where cov(Pi,Pj) is the covariance between the precipitation/temperature for the stations i
and j, cov(Oi,Oj) is the covariance between the observational errors for the stations i and
j (assumed to be zero if i ~ j), and cov(P,Pi) is the covariance between the precipitation/
temperature for the selected area and station i.
This equation system can be solved if the covariance field is known, and the accuracy of
the method depends on how accurately the covariance field can be described. For the
simulations presented in this paper the description was based purely on data from meteor-
ological stations and general knowledge of the precipitation/temperature pattern. One may
also add the information included in a meteorological model and, for example, take into
consideration the topography and prevailing winds.
Tests comparing optimal interpolation and distance interpolation showed that optimal
G. LindstrOm et al./Journal of Hydrology 201 (1997) 272-288 277

interpolation generally estimated point precipitation more accurately (Johansson, 1994).


The improvements in runoff simulations were achieved mainly for catchments with a
sparse network of meteorological stations and large gradients in precipitation. In these
areas the covariance field added information which is not used in such a consistent manner
when selecting stations weights in earlier HBV model applications.

4.4. Snow

The standard snowmelt routine of the HBV model is a degree-day approach, based on
air temperature, with a water holding capacity of snow which delays runoff. Melt is further
distributed according to the temperature lapse rate and is modelled differently in forests
and open areas. In the re-evaluation process a number of attempts were made to improve
the modelling of snow.
1. A more advanced distribution of the snow accumulation in each elevation zone was
introduced. This gives a more realistic description and simplifies comparisons with
remote sensing data.
2. A wind-dependent snowfall correction factor was tested without any success.
3. The distinct threshold temperature to discriminate snow from rain was replaced by a
gradual change over the temperature interval of typically - I ° C to +l°C around the
threshold temperature.
4. Wind-dependent snowmelt was attempted without success.
5. A season- and weather-dependent degree-day factor was attempted without success.
6. Simulations with the wet-bulb temperature gave improvements in some basins, but not
in general.
It seems like the results from the well-controlled WMO intercomparison of snowmelt
models (WMO, 1986) are confirmed once again. The re-evaluation of the HBV model
shows that the introduction of more advanced routines and additional input data normally
have limited effects on the results of the snow modelling.

4.5. Evapotranspiration

The standard HBV model is run with monthly data of long-term mean potential evapo-
transpiration, usually based on the Penman formula (Penman, 1948). These data are
adjusted for temperature anomalies (Lindstrrm and Bergstrrm, 1992). In the re-evaluation
process a number of alternative calculations, which take the day-to-day variations into
account, were tested, such as a simplification of the Thornthwaite (1948) temperature
index method, the Priestley-Taylor method (Priestley and Taylor, 1972) and various
corrections. None of these gave significantly better results than the other (Gardelin and
Lindstrrm, 1996). However, a correction of evapotranspiration with altitude was intro-
duced based on a study by Evremar (1994), to make the model more physically reasonable.

4.6. Interception and precipitation correction

Interception has previously been neglected or considered part of the snowfall correction
278 G. Lindstrfm et al,/Journal of Hydrology 201 (1997) 272-288

and soil moisture accounting in the HBV model. To some extent this has also been
compensated by the use of uncorrected precipitation data, which means that a simplifica-
tion in the model compensates for systematic errors in input. This is not satisfactory when
internal variables such as snowpack are compared with field data. Therefore, a simple
interception storage was introduced for forested areas (see Lindstr6m et al., 1994).
Interception was neglected for open areas.
It was not possible to show convincingly any improvement of the model performance
by the introduction of the interception routine. On the contrary, the results indicated a
slight deterioration. This could be caused by shortcomings in the computation of
potential evapotranspiration. However, the fact that this model is more realistic justified
this routine as an option in the final HBV-96. The realism of this interception
routine has further been confirmed by comparisons with interception data made by
Tallaksen et al. (1996).

4. Z The response function

The response function controls the dynamics of the generated runoff, and thus its
distribution in time once the water balance is set by the routines for snow and soil
moisture. The response function of the standard model (Fig. 2) is governed by three
recession coefficients (KO, K1 and K4), one percolation capacity (PERC), one threshold
(UZLO) and one parameter in the transformation function (MAXBAS): altogether six
empirical parameters. It has long been felt that this is an overparameterisation which
leads to undue parameter interaction. It is also questionable whether the increase in
response, indicated by a successive increase in recession, should cease above a certain
level in the upper zone, as is the case in the standard model.
To cope with these problems an alternative parameterisation was introduced in the
upper zone of the response function, whereas the lower zone was not changed. The two
recession components and the threshold were replaced by a non-linear drainage equation

r R

UZ
~. . . . . . . . . . . . . . . . . . ~ ' : ~ : ~ : # ~ OO : KO * (UZ-UZLO)
% ~,~

~PERC

LZ

Q4 = K4 * LZ
Fig. 2. The response function of the standard HBV model.
G. LindstrOm et aL/Journal of Hydrology 201 (1997) 272-288 279

expressed as
Q = K . U Z l+~
with parameters K and or. The new response function was generally slightly better than the
old one, and it had the merit of having one parameter less and possibilities to generalise
without affecting the simulations too strongly. Therefore, the new response function was
chosen for HBV-96.

4.8. Model calibration

Although the automatic calibration routine is not a part of the model itself, it was an
essential component in the re-evaluation work. The basic problem is to find a criterion
which in a reasonable way summarises the performance of the model and to optimise this
criterion in an efficient way. A new criterion R V was introduced, which is a compromise
between the traditional efficiency R 2 and the relative volume error RD:
R V = R 2 - wlRDI
In practice the optimisation of only R 2 often results in a remaining volume error. The new
criterion proved to be quite useful as it gave results with almost as high R 2 values and
practically no volume error. The best results were obtained with w close to 0.1.
The automatic calibration method for the HBV model developed by Harlin (1991) used
different criteria for different parameters. With the simplification to one single criterion,
the search method could here be made more efficient. The optimisation is made for one
parameter at a time, while keeping the others constant. The one-dimensional search is
based on the Brent parabolic interpolation (see Press et al., 1992). A modification was
made, however, so that a tolerance in the criterion, instead of in the parameter value, is
specified. The one-dimensional optimisation is repeated in a loop over all the parameters.
A new loop is made as long as the improvement is sufficiently large. After each loop a
search is made in the direction defined by the differences in parameter values between the
two latest loops. A few hundred model runs, much fewer than for the old method, were
normally sufficient for a complete calibration with this new method.

5. Final model structure

The final model structure of HBV-96 is presented schematically in Fig. 3. The figure
only shows the most important characteristics of the model, and some clarifications are
given below.
The classes of land use are the same as in the standard model, i.e. open areas, forests,
lakes and glaciers. Since it has been found difficult to relate model parameters to basin
characteristics, it was not felt justified to introduce more classes. As an example,
Johansson (1993) found a significant relation only between evapotranspiration losses
and forested area. In applications of the HBV-96 model it is possible to use different
values of SFCF, SFDIST, CFMAX, ECORR and the interception storage capacity IC for
different vegetation zones, but the ratios between the values for forested and non-forested
areas are kept constant.
280 G. LindstrOm et al./Journal of Hydrology 201 (1997) 272-288

P, T, EP INPUT
p=T. w~'Pi
t = Z W(TI
ep=T~ wL"EPI
T i PC(z)= PCALT if z< PCALTL
PC(Z)= PCALTUP if z> PCALTL
P=p.(1 + PC(z). (z- z~))
Z T = t - TCALT. (z-z~)
PCALTL EP = ep. (1 - ECALT. (z-z,~))
RF=RFCF.P ifT_> TT + T r l l 2
SF = SFCF- P if T <TT -TFI / 2
.,= ..= EP = ECORR. EP

SNOW
I! II
MELT = CFMAX. (T- TIM) if T > TTM (MELT< SP)
REFR = CFR. CFMAX. ('r- TTM) if T < TIM (REFR< WC)
WC = WG+MELT+RF If WC < WHC • SP
IN = MELT + RF HWC > WHC. SP
SP

0 FC O ~

S(~L ~ EVAPOTRANSPIRAT
~ = (SMr',-C)mA
EP: EP; ( ! ÷ ~ "(T'TM)/Tu
E/VEP= SM4.P
~P =1
CF = CFLUX ! (1-81WFC)
EL= C ~ P L ' EP(-LAG)

RESPONSE
Qo=K,UZ(l+ ~
QI=K~.LZ

..,,h,

time MAXBAS time time

Fig. 3.
G. Lindstr#m et aL/Journal of Hydrology 201 (1997) 272-288 281

5.1. S n o w r o u t i n e

A threshold temperature TT is used to distinguish rainfall from snowfall. If the


p a r a m e t e r 77"/is used the threshold is e x t e n d e d to an interval, and within this interval
precipitation is a s s u m e d to be a mix o f rain and snow, decreasing linearly from 100% s n o w
at the l o w e r end to 0% at the upper end. The s n o w p a c k is a s s u m e d to retain melt water as

INPUT SOIL AND EVAPOTRANSPIRATION


Variables: Variables:
p =
Weighted mean precipitation SM = Soil moisture
t =
Weighted mean temperature R = Runoff from soil
ep =
Weighted mean evaporation EA = Actual evapotranspiration
P~ =
Precipitation at station i El = Interception evaporation
T~ =
Temperature at station i CF = Capillary flow
EP= =
Potential evaporation at EL = Lake evaporation
station i (long-term TM = Subbasin mean temperature
mean values) (long-term daily means)
P = Zone precipitation Parametersand constants:
T = Zone temperature FC = Maximum soil moisture content
EP = Zone potential evapotranapiration LP = Limit for potential evapotranspiration
RF = Zone rainfall BETA = Parameter in soil routine
SF = Zone snowfall CFLUX = Maximum value of CF
Parametersand constants: CEVPL = Lake evaporation correction factor
PC(z) = Elevation correction factor LAG = Time lag for lake evaporation
TCALT = Elevation correction factor ETF = Temperature correction factor
ECALT = Elevation correction factor
z = Zone elevation RESPONSE
Zref = Reference level Variables:
TT = Temperature limit for snow/rain UZ = Storage in upper response box
111 = Temperature interval wi~ 0o = Outflow from upper response box
a mixture of snow and rain LZ = Storage in lower response box
RFCF = Rainfall correction factor Q~ = Outflow from lower response box
SFCF = Snowfall correction factor Q = Outflow from transformation function
ECORR = Evaporation correction factor Parametersand constants:
K = Recession coefficient
SNOW ALFA = Response box parameter
Variable~ PERC = Percolation from upper to
SP = Frozen part of snowpack lower response box
WC = Liquid water in snow K1 = Recession coefficient
MELT = Snowmeit MAXBAS = Transformation function parameter
REFR = Refreezing of liquid water /lake = Internal lake zone
IN = Infiltration to soil
Parametersand constants:
SFDIST = Distribution factor for snowfall
WHC = Water holding capacity
CFMAX = Degree day factor
TrM = Temperature limit for melting
CFR = Refreezing factor

Fig. 3. (continued) Schematic structure of one subbasin in the HBV-96 model, with routines for snow (top), soil
(middle) and response (bottom). Bold letters indicate that the parameter is normally calibrated.
282 G. Lindstrfm et al./Journal of Hydrology 201 (1997) 272-288

l o n g as the a m o u n t d o e s n o t e x c e e d a c e r t a i n f r a c t i o n o f the s n o w . W h e n t e m p e r a t u r e
d e c r e a s e s b e l o w the t h r e s h o l d t e m p e r a t u r e , this w a t e r refreezes. G l a c i e r m e l t will o c c u r
o n l y in g l a c i e r z o n e s a n d f o l l o w s the s a m e t y p e o f f o r m u l a as for s n o w m e l t , b u t w i t h
a n o t h e r d e g r e e - d a y factor. N o g l a c i e r m e l t o c c u r s as l o n g as t h e r e is s n o w in the zone. A
s n o w d i s t r i b u t i o n c a n b e m a d e in e a c h z o n e b y s u b d i v i d i n g it into a n u m b e r o f s u b a r e a s
w i t h d i f f e r e n t s n o w a c c u m u l a t i o n . I n the v a l i d a t i o n runs t h r e e s n o w classes w e r e used.
T h i s a c c o u n t s for r e - d i s t r i b u t i o n o f s n o w , s n o w d r i f t a n d s n o w t h a t is t r a p p e d in c r e e k s a n d
o t h e r i r r e g u l a r i t i e s in r u g g e d terrain.

5.2. E v a p o t r a n s p i r a t i o n a n d soil routine

A s a n a l t e r n a t i v e to u s i n g l o n g - t e r m m e a n v a l u e s o f p o t e n t i a l e v a p o r a t i o n as i n p u t to the
m o d e l , daily v a l u e s c a n b e c a l c u l a t e d as b e i n g p r o p o r t i o n a l to air t e m p e r a t u r e , b u t w i t h
monthly coefficients of proportionality. The potential evaporation can be reduced during
d a y s w i t h p r e c i p i t a t i o n , b y a f a c t o r e x p ( - E P F . P ) w h e r e P is the p r e c i p i t a t i o n a n d E P F is a
model parameter.
F r o m the i n t e r c e p t i o n storage, a n e v a p o r a t i o n e q u a l to the p o t e n t i a l e v a p o r a t i o n will
o c c u r as l o n g as w a t e r is a v a i l a b l e , e v e n if it is s t o r e d as s n o w . If the i n t e r c e p t i o n r o u t i n e is
u s e d it is also p o s s i b l e to r e d u c e the soil e v a p o r a t i o n to a v o i d v a l u e s o f total e v a p o r a t i o n
t h a t are t o o large. T h e i n t e r c e p t i o n r o u t i n e , h o w e v e r , w a s n o t u s e d in the final v a l i d a t i o n

Table 1
Results with the final HBV-96 model compared with the results of the standard HBV-REF model for both
calibration and verification periods. R z is the model efficiency, IRDI is the absolute value of the relative
accumulated difference (volume error) and VS is the mean volume error during the spring flood

Basin HBV-REF HBV-96 HBV-REF HBV-96 HBV-REF HBV-96


g ~% Rz % IRDI (%) IRDh (%) VS (%) VS (%)

Suorva cal. 90.7 90.8 1.2 0.1 1.5 -0.3


Kultsj6n cal. 87.4 89.9 0.4 0.0 0.2 -1.4
Str6ms Vattudal cal. 86.7 87.9 1.6 0.3 1.4 -1.9
Ljusnedal cal. 86.0 89.8 2,8 0.9 -3.8 -2.1
Blankastr6m cal. 84.3 90.5 1.6 0.3 6.3 4.6
Torsebro cal. 93.3 93.7 4.7 0.1 -4.0 -0.1
H61jes cal. 91.5 93.3 1.5 1.2 -2.3 -0.7
Mean cal. 88.6 90.8 2.0 0.4 --0.1 ~0.3
Suorva ver. 77.2 80.4 16.2 11.3 20.2 -8.7
Kultsj6n ver. 81.4 89.5 9.0 6.9 12~4 4.9
Str6ms Vattudal ver. 78.6 80,9 6.2 6.0 -12.2 -11.2
Ljusnedal ver. 84.4 90.5 2.3 2.6 3.4 -5.0
Blankastr6m ver. 86.6 88.8 8.5 2.8 7.0 8.1
Torsebro ver. 88.2 90.8 5.3 3.2 -1.0 -2.6
H61jes ver. 88.7 90.6 2.7 2.1 6.9 5.4
Mean ver. 83.6 87.4 7.2 5.0 -1.3 -1.3
Mean total 86.1 89.1 4.6 2.7 -0.7 -0.8
G. LindstrOm et aL/Journal of Hydrology 201 (1997) 272-288 283

runs. Instead the potential evapotranspiration from forested areas was assumed to be 15%
higher than that from open areas, based on the findings by Johansson (1993). The potential
evapotranspiration was thus a function of the time of the year, the current air temperature,
vegetation, elevation, and, as an option, precipitation.
Evaporation from lakes will occur only when there is no ice. Ice conditions are modelled
with a simple weighting subroutine on air temperature, which results in a lag between air
temperature and lake temperature. It is assumed that the lake is frozen when the weighted
temperature drops below zero.

5.3. Response function and routing

As illustrated in Fig. 3, the response function of the model consists of two reservoirs and
one transformation function. Level pool routing is performed in lakes located at the outlet
of a subbasin. The division into submodels, defined by the outlets of major lakes (not
shown in the figure), is thus of great importance for determining the dynamics of the
generated runoff. The routing between subbasins can be described by the Muskingum
method (see for example Shaw, 1988) or simple time lags.
The novelty with HBV-96 is that the two outlets of the upper reservoir of the earlier

R 2 (%)
I ~ HBV-R1EF
~ .ev.96
100 --

Calibration Validation

6O
11
Fig. 4, Model efficiency measured as g 2 by the reference HBV model and by HBV-96 for the calibration period
(left) and for the independent validation period (right).
284 G. LindstrOm et al./Journal of Hydrology 201 (1997) 272-288

model are now replaced by a non-linear relation between storage and outflow. Each one of
the subbasins has individual response functions.

6. Final model validation

At the end of the re-evaluation process a final version of the model, the HBV-96, was
ready for the final test. Seven basins were chosen for these control runs. The old calibration
routine was used for the old model and the new routine for the new model. The results are
summarised in Table l and Fig. 4, which show the calibration periods and the independent
validation periods separately. The new model systematically gave better results in all
basins. It is further encouraging that these improvements were largest in the independent
period. An average increase in R 2 from 86 to 89% may look small at first sight, but one
should keep in mind that the error corresponds to the unexplained variance, or 1 - R 2. The
unexplained variance was thus reduced from 14 to 11% as an average. Furthermore, the
new method for automatic weighting of input data was only used in four of the seven
basins. The new model, of course, did not always perform better than the old one. The year
with the greatest improvement is shown for two basins in Fig. 5.
The spatial distribution of the modelled snowpack was checked by comparisons with
estimations of snow cover based on N O A A - A V H R R scenes (H~iggstr6m, 1994). Examples

Q (m3/s) 500 - o (m3/s)


5OO
Accdlff (mm) 1000
HBV-REF KULTSJON 400 HBV-96 KULTSJON ., ~r~500
400
l V

300-

200 200 -

100 100-
_. . =t ,1..
0 . . . . . .~,i
0 ' M''I"'A I U I J I J I A ] ' M'"I~'A I M I J I J [ A ]
1989 I 1989 I
100 - Q (m3/s) Q (m3/s)
lO0
-] Accdiff (mm)
HBV-REF LJUSNEDAL 80 q HBV-96 LJUSNEDAL I- +200
+0
80 ~ -'="--

60-

60 9 I -200
40- 4O

20- 20

0 0
M I A I M I d I d I A I
1989

Fig. 5. The year with the largest improvements using the new model (from the validation period). Thick curve is
the computed discharge and thin curve is the recorded discharge.
G. Lindstri~m et al./Journal of Hydrology 201 (1997) 272-288 285

of such comparisons are given in Figs. 6 and 7. The distributed nature of the HBV-96
model version improves the possibility o f using this kind of information for updating of the
model during the snowmelt period.

7. Discussion and conclusions

The development o f the new HBV-96 model version followed a long tradition in
hydrological modelling, with its roots in a debate on modelling strategies initiated by
Nash and Sutcliffe (1970). The basic idea in the early 1970s was to find a model structure
of optimum complexity in relation to the operational requirements and available data. All
increased model complexities should thus be justified by results in the validation
procedure. This strategy resulted in a new generation of conceptual hydrological models
where good simulation results were obtained by surprisingly uncomplicated structures and
with only moderate data demands. A comprehensive overview of some of these models
was given by Singh (1995). The HBV model was one of the models of this generation.
Along with the advent of more user-friendly computers, this model approach proved to be
a very robust and powerful management tool and, nowadays, conceptual hydrological
models of this kind are everyday tools of hydrologists all around the world.
A number of model intercomparisons have confirmed that the relatively simple
conceptual model is an appropriate operational choice in hydrological modelling,
maybe best illustrated by the well-controlled intercomparisons of hydrological models
carried out by the World Meteorological Organisation ( W M O , 1975 and 1986). The
present re-evaluation of the HBV model confirms that we cannot expect any major

clouds

Fig. 6. Snow cover in the Suorva basin 15 July 1993 according to a NOAA-AVHRR analysis (left, after
Hfiggstr6m, 1994) and according to HBV-96 (right). The figure shows mostly snow-covered areas (snow), partly
snow-coveredareas (mixed) and areas with almost no snow (no snow). In the NOAA-imagethe whole basin could
not be analysed due to scattered clouds. The HBV-96 model here has 38 subbasins, compared with 3 subbasins in
the standard HBV model for the same basin.
286 G. Lindstrfm et al./Journal of Hydrology 201 (1997) 272-288

HBV-REF HBV-96
+20 T E M P (*C) +20 -aTEMP (°C)

+o / -^ .....
2° J r V -
.2oY" V - "
30 PREC (mm/d) SNOW (rnm) 500 30 PREC (mm/d) SNOW (ram) ~'00

,L E;,o
0 it~. LI .. ~ i i i . , i JIil~'~.~.=JJll i / 0 o L,.. i .... t ...|o
SNOW COVER 1 0 SNOW COVER -- COMP.
1.0 ~ -- COMP.

300 - Q (m3/s) ACCDIFF (rnm) 800 300 - Q (m3/s) ACCDIFF (mm) 600

0,0j-200~ ~ ~ ~ k

,00
500

200

loo-

o
J
" ~
1991
I M I J I J I
I
600

o
1991
A' I ~ I J 1 J I

Fig. 7. Snow-covered fraction compared with estimations using NOAA-scenes (H~iggstr6m, 1994) according to
the reference model and the new HBV-96 model.

breakthroughs as long as we limit ourselves to standard input data and judge the models by
analysis of fiver runoff only. We also have to realise that the starting point is a model of
high performance, where we have probably used nearly all the available information value
in the crude input data base.
It is interesting to note that it was not easy to single out one factor as being the most
important one in the re-evaluation process, and that the processing of input data and the
new calibration routine were, at least, equally important as physically justified improve-
ments of the process description in the model. The results of the present re-evaluation do
not justify any increased resolution in time or space, unless more detailed data are to be
used as input or for validation. These results are strictly specific to runoff simulation under
these Swedish conditions, and they may have to be reconsidered, even in Sweden, for
smaller catchments of quicker response. The option of higher resolution in space is also a
prerequisite for future integration of spatially distributed field data in the model, and this
type of data is gradually becoming available. Examples are given already in this paper.
Although some of the experiments came out with negative or neutral results, the final
model structure, with its new routines for data processing and calibration of free
parameters, outperformed the old system. These results are particularly encouraging as
G. LindstrOm et aL/Journal of Hydrology 201 (1997) 272-288 287

the improvements are better for the validation periods compared with the periods used for
calibration of the model. This means that we now have a better tool for hydrological
simulation, forecasting, design studies and other applications. We also have a model
which is more physically realistic, as some inconsistencies are replaced by better
parameterisations. This is of great importance for future applications to non-stationary
conditions, such as impact studies o f climate change, and for design studies where the
model is stretched to applications outside the range of calibration of its parameters.
Finally, we now have a more flexible model structure which can be used to take advantage
of more advanced future data from field campaigns or remote sensing. This new structure
also has a greater potential for further development into the field o f water quality
modelling, as effects of land use and other human activities can be described in more
detail.
As the model development process is continuous, the HBV-96 is the latest, but not the
last version of the HBV model. The dramatic spring floods o f 1995, with the highest peak
flows in the 20th century in several rivers in Norway and Sweden, gave rise to some further
concern about model performance under extreme snowmelt conditions. This is only one
example of the continuous feed-back from operational hydrology to the development
process. More is likely to come as HBV-96 is entering operation and results are reported
back.

Acknowledgements

This work has been carried out at the Research Department of the SMHI. It has been
financed by the Swedish Association of River Regulation Enterprises (VASO), the
Swedish Meteorological and Hydrological Institute (SMHI), the Nordic Council of
Ministers (NMR) and the Swedish Natural Science Research Council (NFR). The authors
are also grateful to all colleagues and friends in Sweden and abroad who have shared their
experience in hydrological modelling with us and thus contributed in this process.

References

Bergstr6m, S., 1976. Development and application of a conceptual runoff model for Scandinavian catchments.
SMHI RH07, Norrk6ping.
Bergstr6m, S., 1995. The HBV model. In: V.P. Singh (Ed.), Computer Models of Watershed Hydrology. Water
Resources Publications, Highlands Ranch, CO.
Bergstr6m, S., Carlsson, B., Graham, L.P., 1996. Modelling the water balance of the Baltic basin--preliminary
results. Nordic Hydrological Conference, Akureyri, August 13-15, 1996. NHP-Report No. 40, pp. 449-455.
Bergstr6m, S., Forsman, A., 1973. Development of a conceptual deterministic rainfall-runoff model, Nordic
Hydrol., 4, 147-170.
Bergstr6m, S., Harlin, J., Lindstr6m, G., 1992. Spillway design floods in Sweden: I. New guidelines, Hydrol. Sci.
J., 37, 505-519.
Daley, R., 1991. Atmospheric Data Analysis. Cambridge University Press.
Evremar, ~,., 1994. Avdunstningens h6jdberoende i svenska fj~illomr~denbest/4md ur vattenbalans och med
modellering (The dependence of evaporation on altitude in Swedish mountains, estimated by water balance
and modelling). SMHI Hydrology, No. 52, Norrk6ping.
288 G. Lindstriim et al./Journal of Hydrology 201 (1997) 272-288

Gardelin, M., LindstriSm, G., 1996. Priestley-Taylor evapotranspiration in HBV simulations. XIX Nordic
Hydrological Conference, Akureyri, August 13-15, 1996, NHP-Report No. 40, pp. 648-657.
Harlin, J., 1991. Development of a process oriented calibration scheme for the HBV hydrological model, Nordic
Hydrol., 22, 15-36.
HiiggstriSm, M., 1994. Sn6kartering i svenska fj~illomrhdenmed NOAA-satellitbilder (Snow mapping in Swedish
mountains, by use of NOAA satellite images). SMHI Hydrology, No. 57, Norrk6ping.
Johansson, B., 1993. The relationship between catchment characteristics and the parameters of a conceptual
runoff model--A study in the south of Sweden. Contribution to the Second International Conference on
FRIEND, October, 1993, Braunschweig. IAHS Publication No. 221, pp. 475-482.
Johansson, B., 1994. The use of spatially distributed input data in the HBV model. Nordic Hydrological
Conference, T6rsbavn, August 2-4, 1994, NHP-Report No. 34, pp. 147-156.
Killingtveit, A., Saelthun, N.R., Saether, B., Taksdal, S., yon Hirsch, R., 1990. Programmer till HBV-modellen
(The program code for the HBV model). SINTEF, Norwegian Hydrotechnical Laboratory, Report STF 60
A90033, Trondheim.
Lindstr6m, G., Bergstr6m, S., 1992. Improving the HBV and PULSE-models by use of temperature anomalies.
Vannet i Norden, No. 1, pp. 16-23.
Lindstr6m, G., Gardelin, M., Persson, M., 1994. Conceptual modelling of evapotranspiration for simulations of
climate change effects. SMHI RH No. 10, Norrk6ping.
Lindstr6m, G., Gardelin, M., Johansson, B., Persson, M., Bergstr6m, S., 1996. HBV-96--En areellt fi3rdelad
modell f6r vattenkrafthydrologin (HBV-96--A distributed model for bydropower hydrology). SMHI RH
No. 12, Norrk6ping.
Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models. Part I: a discussion of
principles, J. Hydrol., 10, 282-290.
Penman, H.L., 1948. Natural evapotranspiration from open water, bare soil and grass, Proc. R. Soc. London Ser.
A:, 193, 120-145.
Press, W.H., Teukolsky, S.A., Vetterling, W.T., Flannery. B.P., 1992. Numerical Recipes in FORTRAN. The Art
of Scientific Computing, second edition. Cambridge University Press.
Priestley, C.H.B., Taylor, R.J., 1972. On the assessment of surface heat fluxes and evaporation using large-scale
parameters, Monthly Weather Rev., 100, 81-92.
Renner, C.B., Braun, L., 1990. Die Anwendung des Niederschlag-Abfluss Modells HBV3-ETH (V 3.0) auf
verschiedene Einzugsgebiete in der Schweiz (The application of the HBV3-ETH (V 3.0) rainfall-runoff
model to different basins in Switzerland). Geographisches Institut ETH, Berichte und Skripten Nr 40, Ziirich.
Shaw, E.M., 1988. Hydrology in Practice, second edition. Chapman and Hall, London.
Singh, V.P. (Ed.), 1995. Computer Models of Watershed Hydrology. Water Resources Publications, Highland
Ranch, CO.
Tallaksen, L.M., Schunselaar, S., van Veen, R., 1996. Comparative model estimates of interception loss in a
coniferous forest stand, Nordic Hydrol., 27, 143-160.
Thornthwaite, C.W., 1948. An approach toward a rational classification of climate, Geogr. Rev., 38, 55-94.
Vehvil~iinen, B., 1986. Modelling and forecasting snowmelt floods for operational forecasting in Finland.
Proceedings from the IAHS Symposium: Modelling Snowmelt-Induced Processes, Budapest, IAHS
Publication No. 155, pp. 245-256.
WMO, 1975. Intercomparison of conceptual models used in operational hydrological forecasting. Operational
Hydrology Report No. 7, Geneva.
WMO, 1986. Intercomparison of models of snowmelt runoff. Operational Hydrology Report No. 23, Geneva.

You might also like