A STATISTICAL STUDY FOR AUTOMATIC/RHO_10 A...Owh = waterholding capacity of the anow (% of the...
Transcript of A STATISTICAL STUDY FOR AUTOMATIC/RHO_10 A...Owh = waterholding capacity of the anow (% of the...
-
A STATISTICAL STUDY FOR AUTOMATIC CALIBRATION OF A CONCEPTUAL RDNOFF MODEL
STATISTISK STGDIE FÖR AUTOMATISK KALIBRERING AV EN :BEGREPPSMÄSSIG AVRINNINGSMODELL
Sören Svensson
SMIU Rapporter
HYDROLOGI OCR OCEANOGRAFI Nr RRO 10 ( 1977 )
SVERIGES METEOROLOGISKA OCH HYDROLOGISKA INSTITUT
Norrköping 1977
-
i •
2.
3.
4,
LIST 0F C0NTENTS
Chapter
4. 1
4.2
4-3
4,4
4.5 4.6
5. 1
5.2 5.3
List of contents
SUillI!lary
Introduction
Test catchments
The HBV-model and its free para.meters
Snow routine
Soil moisture routine
Response function
Statistical analysis of the residuals of the HBV-model
Mechanism separation criterion (MSC)
Estimation of the density function
Test of normality and independence
Estimation of the autocorrelation
Significance of the autocorrelation
Analysis of the autocorrelation
A study of response surfaces of criteria of fit
Criteria of fit
Basic assumptions and observations
The response surfaces of criteria of fit and test plottings of Kultsjön
Th f th R2 ·t . e response o e -cri eria to K
2 and C perc
Th f th R2 ·t . e response o e -cri eria to K
1, B and C max route
The response of the R2-criteria to Fe, Sand L /Fe
p of the R2-criteria The response
to csf and S
Th f th R2 ·t . e response o e -cri eria to T and C
0 0
Th f th R2 ·t . e response o e -cri eria to C h and C f w r r
1
2
5 6
8
9 10
13
13
16
17
19
22
23
29
29
31
40
42
42
-
6.
5.4
5.4.1
The response surfaces of eriteria of and test plottings of Stadarforsen
The response of the R2-eriteria to K
2 and C pere
2 The response of the R -eriteria to K
1, C t and B rou e max
The resporise of the R2-eriteria to K and L
0 UZ
Th f th R2 't . e response o e -eri eria to K
0 and K
1 The response of the to Fe, 8 and L /Fe p The response to csf and 8
of the
R2 't . -eri eria
R2 't . -eri eria
5.4.7 The response of the R2-eriteria to T and C
0 0
5.4.8 The response of the R2-eriteria to C h and C f w rr
5.5 Results of the study Conelusions
Appendix A: Density_funetions
List of symbols
List of referenees
Appendix B:
fit
45
45
45
48
48
52
52
52
55 55
-
- 1 -
SUMJ'IIARY
A conceptual runoff model is studied in this work. Especially
the residuals of the model (the differences between the compu-
ted hydrograph and the recorded one) are exami-11.ed. The density
function and the autocorrelation function of the residuals are
estimated and tested.
The model must be calibrated for each new catchment, to which
it is applied. Therefore a criterion of the goodness of fit
between the computed and the recorded hydrograph is required.
Some criteria of fit have been examined concerning their sensi-
tivity to changes in the model parameter setting.
Conclusions of the work are: The residuals of the model are
neither stationary nor independent and normally distributed. A
classification based on the different physical processes, which
govern the discharge, makes the residuals of each class more
stationary and in some sense more normally distributed t~an the
residuals of the material without a:ny classification. Further-
more a criterion of good.ness of fit, based on the classifica-
tion above resembles the subjectively judged fit more than the
simple sum of squares criterion, which has become practice in
applications of runoff models.
-
- 2 -
1 • INTRODUCTION
The purposes of the development of hydrological catchment medels
are mainly (Clarke, 1973):
1 • Forecasting
a) Operational forecast: Estimating streamflow when rain-
falls, losses, streamflows etc.
are given up to the present time.
b) Design forecast: Estimating the flood hydrograph
caused by a hyJ:>othetical heavy
storm (or snowmelt).
2. Extending the discharge redord, where we have got long cli-
matological record but short discharge record.
3. Prediction of the possible effects of proposed pbysical changes to the catchment.
Beföre the model can be taken inta operation, its free parameters
have to be accurately estimated by means of a calibration proce-
dure. This means that the parameters are adjusted until a good
fit is obtained between the computed hydrograph and the observed
one. The procedure can be either a subjective one, relying upon
the hydrologist when deciding which parameters are to be adjusted,
or an automatic one, based on an optimization algorithm.
One major problem concerning the automatic procedure has been the
finding of a matching index, a criterion of fit. This is very im-
portant, because every automatic parameter optimization routine
has to rely upon a nu.merical value of the goodness of fit between
two curves.
The statistical properties of the residuals could be a key toa
better understanding of this problem. The lack of such studies was
pointed out by Clarke (1973).
-
the prol:iabili ty distribution of the rnodel resj_d_,,1als" BowevE.·r~
one fu::cther possibili ty is to do s. nu.mhe:r: of paE3lliE' r-er sstina-
tions based on independent data and from ths :.r6su::i_ ts of ti1e::;e
estimations get an apprehension of the s i se ,:,f" the ,,onfidenct:c
regions of the model parameters.
The a,im of this wo:r:k is to study the statistical pL·operties sf
the residuals in order to find a more app~opriate criterion oi
fi t between the cornputed and the recorded h;ydJ..0ographo This prob-
lem. was approached in two steps.
1. The calibration period was divid.ed ir1to 11 subperiods'1 in
order to obtain stationary subsets of residuals. Some
statistical properties of these residuals were examined.
2. The response su:rfaces of different criteria of fit were
studied when altering the parameter values.
The EJ3V runoff model was used for the study. This rnodel is deve-=
loped s;c the SJYIHI (Swedish Mete,:;rological and Hydrological Insti-
tute) by Bergström (1976). It is in operational use in same
c1;,,t:;hm.ent~ in Sweden and N"orway.
Tl1is stuc'Qr was financed by the SMHI and it was carried out in
co-o:peratinn with the ])epartment of Mathematics at the Lin-
köping University~
-
- 4 -
~-100 -= --!~-- 100 km --:1
(j i 30°
. () I
+
i ig. 2.1. Test catchm t th en 8 for th e HBV model~ e applications of the study of
-
- 5 -
2o TEST CATCBJVJENTS
Two catchments were used in the study: the Kultsjön catchment~
below referred toas Kultsjön, and the Stadarforsen catchment,
below referred toas Stadarforsen (fig. 2.1). Both catchments
are below tirnberline dominated by coniferous forest, and the
soil is mostly moraine ar of a pervious type.
The catchments are representative for large areas in Scandi-
navia.
Table 2.1 Test catchments.
Area Altitude Lakes Catchment River (k:m.2) range (m) %
Stadarforsen Västerdalälven 4 136 835 2,4
Kultsjön 0
Angerm.anälven 1 109 1 040 6
The analysed runoff record of Stadarforsen bega.n 1961-10-01
and ended 1976-03-31. In Kultsjön_ the analysed record bega.n
1961-10-01 and ended 1976-05-18.
-
- 6 -
3. THE HBV-MODEL AND ITS FREE PARAMETERS
The simulation of discharge by the HBV-model. is made in tbree
steps (fig. 3.1).
1. Snow accumulation and ablation.
2. Soil moisture accounting.
3. Generation of _runoff and transformation~of t~e hydrograph.
PRECIPITA TION
SNOW ROUTINE
RAIN, SNOWMELT, _..__~--i.- EVAPORATION
SOIL MOISTURE ZONE
Fig. 3.1 Schematic representation of the HBV-model.
-
- 7 -
Table 3 _ •j Parameters of the HBV-model.
Corrections on input va.riables
P correction factor on rainfall corr
P = precipitation-elevation correction lapse
~ = temperature-elevation correction -1apse T general temperature correction
0
Parameters of the snowroutine
snow fall correction factor
= degree-day melt factor water holding capacity
bottom storage under snowpack
= refreezing coefficient
Parameters of the soil moisture routine
Fe = maximum soil moisture storage
L limit for potential evaporation p 6 erupirical coefficient
Parameters of the response function
storage discharge constant of the upper zone
C perc p
w
B max
C route
tr PI 11 n 11
= Il " Il t1 lower limit for slow dxainage of the upper zone
= percolation capacity into the lower zone
part of the lower zone representing lakes and other wet areas
Il
"
maximum base in the transformation function
parameter relating the base in the transfor-mation function to the generated flow
-
- 8 -
The model para.meters are shown in table 3.1. P , P1
, T1 corr apse apse
and P are set from information outside the calibration procedure w
(maps, experience etc.). The ot:.1ers are calibrated to optimum fit.
3.1 Snow routine
Whenever the air temperature (T) is below a threshold value (T ), all 0
precipitation is regarded as snow and is accu.mulated in the snowpack.
All effects of evaporation and lacking representativenese of the
gauge are put together in one empirical coefficient, the enow fall correc•
tian factor (C8f).
Thus, if T < T then 0
s = C • P, s sf where s = actual snow accumulation (mm),
8
csf = snow fall correction factor, p = precipitation (mm), T = surface air temperature (oc).
Snowmelt is taken care of by the degree-day method:
If T > T, then 0
M = C (T - T ) , 0 0
where M = snowmelt (mm/day),
C = degree-day factor (mm/( 0 c a.ay)). 0
The water retention in the snowpack is described by two parame;.ers.
Owh = waterholding capacity of the anow (% of the snowpack), Sb = bottom storage under the snow (IIIIIl).
Sb has'rarely been used and was therefore omitted in this work.
Refreezing of liquid water in the snowpack:
If T < T and if there is liquid water present in the enowpa.ck, then 0
M = C f • C (T - T ), r r o o
where -M = refreezing rate (mm./a.ay),
C = refreezing coefficient. rfr
-
- 9 -
The area-elevation distribution of the snowroutine in the HBV-
model is described by Bergström (1976). It contains no fre&
parameters and is therefore not very interesting for the cali::J-
ration of the model.
3.2 Soil moisture routine
The behaviour of the soil moisture zone is illustrated in fig.
i.O
0 0
Fig. 3.2 The contribution.s from rainfall or snowmelt, P, to the soil moisture zone, S , and the upper zone, sm s
UZ
Mathematically, this can be described by the following
equations:
s l UZ 6. p
/:, 8sm 6. p
where P
s UZ
s sm Fe
s s
6. UZ
precipitation or snowmelt (mm), = storage in the upper zone (mm),
computed soil moisture storage (mm), = maximum soil moisture storage in the model (mm), = empirical coefficient,
= amount that passes through the soil moisture zone (mm)'
-
- 10 -
~ S = amount that is stored in the soil moisture zone (mm), sm
~· p = precipitation or snowmelt fed into the zone mm by mm.
Potential evaporation is reduced to actual values by the function
in fig. 3. 3.
0
Fig. 3.3 Reduction of potential evaporation, E, to actual, E. p a
Mathematically described by:
E = a
E if S > L p sm - p
s E • sm
p L if S < L, sm p p
where E = potential evaporation, p E = actual evaporation,
a L = limit for potential evaporation.
p
3.3 Response function
Having passed the soil moisture routine the excess water passes
through some reservoirs, where the runoff Q is easily calculated g
in the ma.nner illustrated in fig. 3.4.
-
- 11 -
0. =K·(S-l I 0 0 ~ Vl'
QI= K(Suz
TRANSF. _ __..-;it FUNCTION
Fig. 3.4 The response function of the HBV-model.
C = percolation capacity, perc E = potential evaporation,
p K = storage discharge para.meter of the upper zone,
0
K1 = slow drainage storage discharge para.meter of the
K2
L UZ
p
p w
Qg
Qo
Q,1
Q,2
Qc 81z s
UZ
6 S UZ
upper zone,
= storage discharge para.meter of the lower zone,
= limit for slow drainage of the upper zone,
= precipitation,
= part of the lower zone, representing wet areas,
= total generated runoff,
= runoff generated from the upper zone,
= slow drainage runoff generated from the upper zone,
= runoff generated from the lower zon8
= total computed runoff,
= storage in the lower zone of the medel,
" Il " upper " " " " inflow in the upper zone.
The transforming function multiplies Q with a weight differing g
intime. Q is transformed inte Q according to fig. 3.5. g C
-
- 12 -
WEIGh1
TIME TIME
Fig. 3.5 The effect of the transformation function on the generated hydrograph.
This can be expressed as: t -C • Q,g if (B - C • Qg) .::_ 1 B = max route max route q if (B - C • Q ) < 1 max route g
B = base of the triangu.la.r function q
B = maximum base (days), max
(days),
C t = free parameter (days/(m3/s)), DOU e
Qg = generated runoff from the reservoirs (m3/s),
Time = O: the day on which Q is generated (days). g
-
4. STATISTTCAL Af~ALYSIS OF ~BE RESIDUALS OF THE m:iv-MGD~L
4.1 Mechanism separation criterion (lVISC)"
A study of the hydrographs showed that differe11t pbysical pro-
cesses produced different discha:r·ge patter·ns. Subsequen-tly 8,
study of the residuals showed a similar mixture of different
curve types with different statistical properties.
Chiefly three different processes were assumed.
1. Sno"WIIlelt
2. Rainfall or recession succeding rainfall or sno-wIILelt
(referred toas y-flow)
3. Dry summer- or winter-recession (referred toas low flow).
1'.i:le problem was to find a good cri terion to separa te the days
in order to obtain classes of days with approximately station-
a:.r:y residuals. One method is to divide the calibration period
by visual inspection. The drawbacks of this method are its sub-
jective character and the fact that it does not take the dif-
ferent model mechanisms inta account.
The method was therefore abandoned and the followJ.11g mechaism
separation criterion (MSC) was used.
1. M > 0 snowmelt
> o): y-flow 2. M
3. JVI
where M
s UZ
LIS UZ
< 0
< 0
and
and
(LIS > UZ
( LIS = UZ
0 or S UZ
0 and S UZ
snowmelt (mm),
0) : low flow,
storage in the upper zone of the rnodel (mm),
= inflow in the upper zone of the model (mm).
This MSC has two major drawbacks:
1. It will give different partitions of the calibration
period at different parameter settings. This will
disturb the behavi01.-~1.• of the properties studied.
-
- 14 -
2. It may be difficult to apply to other models.
The main advantage of the chosen MSC is that it is easy to extract
from the model. An example of the use of the MSC is shown in fig. 4.1.
In Kultsjön the transforming function causes a time lag, which hasa
maximum length of one day (B = 2, C t = 0.00103, chapter 3.3). max rou e This small influence is neglected in the mechanism separation.
In Stadarforsen the time lag is much greater (B = 6, C t = O, max rou e chapter 3.3). Here the main part of the generated runoff is delayed by three days. In consequense of this the MSC was displaced three
days.
Fig. 4.1 (See next page.) Residuals (1/s) and MSC of Stadarforsen 65.10.01 - 69.09.30. 1. Snowmelt
2. Rain or recession succeeding rain or
snowmelt.
3. Dry summer- or winter-recession.
The upper curves show the MSC, the
lower curves show the residuals.
-
laj M-J -Jvrvu~t f I-'· 250000 ~ ~ . 150000 _.
50000 t, \~~-~
-50000
-150000
0 N D J F M A M J J A s 0 ~~ I I ._j -3
'{ -\_ _ _J ~2 250000 I .l
150000
_.
50000 (\ \Jl
-1 (/\~ t\ ' -50000 \/ 1-v ,,
-150000
F M A M J J A s ('J N D _J F 1'1 I A -7 M _ j\__ __ 1 Jlfl_
vVVl __ J\ __ -250000
150000
50000
~ /I},\ /1 --, r \ I _,.,~--\-,_ .-- -~- '\._ - -._,j - ... /\ -50000 'v \/\
I
I I I ,_ I
,,J
-150000 _J J I
---- - - -- -- I - - - - -,
A s 0 N D I F M A 1'1 l I :_ l ,_,
-
4.2 Estimation of the density function
Mean and standard deviation of the residu.als were estimated by:
x = 1 i ro (i) - Q (i)l n i=o L'""r c J
s2 ~ ~-1 te [ Cl.,, (i) - Qc (i) - X] 2 where X = mean value of the chosen residu.als (m3/s),
n = number of the chosen residu.als,
~(i)= observed discharge (m3/s),
Q (i)= computed discharge (m3/s), C
S =standarddeviation of the chosen residuals (m3/s).
A_ symmetrical interval, four standard deviations long, was
placed around the mean value. The interval (x - 2s, X+ 2s) was divided inta 40 equ.ally long classes, and the number of residuals
in each class was plotted in a histogram. For comparison a Normal
probability distribution function with X mean and S standard devi-
ation was plotted in the same diagram.
As the interval of four standard deviations is toa short to
contain all the residu.als, the number of exceeding residu.als (NER)
is printed together with each histogram. An exam.ple of a histogram
can be seen in fig. 4.2. (days)
300
200
100
(standard 0
0L...L-'--L.....L.-'--L-L..L-J~
0.L...l.-'-.L...l.-'-'---'-_._~
0......,,_._~_._ ........ __._ ........ o_._ ......... ~ ........ ..._.__._,o deviations)
N - 0 -I I Fig. 4.2 Histogram of the low flow residu.als of Kultsjön 62.10.01 -
- 76.05.1a. Meanvatue = o. 725 (m3 /1,);-i3tanda:rd deviation--=
= 7.07 (m3/s). NER= 145
-
All histograms are presented in Appendix A.
A preliruinary test was made i.:;e, s1;ud:r it ~h,-, ill::oc',-'-- 'c,_1_ ~te
eae:b clase of __cesiduals (x) dlfI8Ii8U_ ,JJ_g11L L~antl~ :r,-,(,Jli 2,-0 :1 ( The simple t-test w-a,s used fo:c this purpose. SoLe appaie:~-;Jy
signific&nt resul ts were fou11d.
However, the t-test is -based on the assumption of i_ndepeudeuc
and N( µ, cr)-distributed variables. The assumption of .norn19,li ty
is not very critical, but the assumption of independence has
to be fullfilled.
As will be shown later, the residuals are not independent. The
test was carried out with regard paid to the interdependence~
according to Hansen (1971). Thus the quantity .2l_ was inter-2 n
preted as the equivalent number of independent x observa-
tions.
Where n is defined by: X
n J R( T) d T X 0
R(T) = autocorrelation function of the chosen residuals.
Using the t-test in this way showed no significant deviation
from zero on the 5 % level.
However, each class of residuals consists of a number of inde~
pendent continuous time periods. Regarding this, one could
construct a more powerful test.
4.3 Test of normality and independence
Ve,ry often in statistical applications the assumptions of mubu-
ally independent and normally distributed variables are ma,le.
To test these assumptions on the 2 mechanisms the X -test was used.
k I: u = . 1 l=
2 (Y.- np.) l l
residuals of different model
-
where u k
n
p. l
Y. l
- 18 -
test va.riable,
number of test classes,
= total number of recorded days,
probability fora normally distributed, stochastic variable
to assume a value within class i,
observed number of days in class i.
If the assumptions hold, then U is approximately x2-distributed with the
number of test classes mL"J.us three degrees of freedom.
Whenever np. < 5, the test class ,j,was merged with a neighbouring l
test class, until the probability to fall within test class limits
was greater than 5/n. This validity limit is ta"i
-
- 19 -
The hypochesis had to be rejected on the 0.7 ja ievel.
Although the material was divided in-i:;o different classesj the
hypothesis H still did not hold. However, the x2-variable (u) 0
was made less by the separation of the different mechanisms.
A visual inspection also shows that the residuals a:r-e more
close to normality in each subclass after the separation.
Therefore H is, in some sense, more valid after the mechanism 0
separation of the material than before.
2 Ji'urthermore, the x -test is very sensitive when used on a mate-rial built upon many observations. The data may be sufficient
in nllJIJ,ber to show their inconsistency with al.mast any hypothe-
sis suggested. (Ha.mon and Hannan, 1963).
The x2-test is unfortunately not able to separate the questions
of distribution and autocorrelation. As will be shown later,
a clearly significant autocorrelation exists in this case.
Moreover, mostly the assumption of normally distributed resi-
duals is not very ciritical. Every result below, obtained by
relying upon Normal distributions may be considered as an
approximation.
4.4 Estimation of the autocorrelation
To estimate the autocorrelation a method described by Jenkins
and Watts (1969) was used.
At first the autocovariance was estimated for each continuous
period j, when one mechanismv.BS working (see chapter 4.1).
where
,.. R. ( T)
J
R. ( T) J
X(t)
N. J
t. J
=
N.-,+t.-1 1 J J
N.-T J
z (x (t+,) t=t.
J
x) ( x ( t) - x) ,
estimated autocovariance of residuals separated
by T days,
'\,(t) - Qc(t) = residual at time t (m3/s), duration of period j (days),
the time at which period j started (days from
the beginning of the discharge record).
-
- 20 -
In order to get an overall estimate of the autocovariance the estimates
from the different periods with the same mechanism working were brought
together as follows,
h N. - '"[ R(,) = Z (-N--
all -r the
periods j
R.(-r)) J
where R(-r) = N
overall estimate of the autocovaxiance,
T ~(N. - ,) = the total number of observations of R(-r) (days). j J
Finally the autocorrelation was estimated by dividing the estimates of
the autocovaxiance by the estimate of the variance.
When there is less than about 400 observations, one should not rely too
much upon the autocorrelation cuxves, because of the extremely skew
probability distribution of the autocovariance.
Kendall and Stuart (1967) actua.lly warned for the use of a correlation
coefficient estimate, based on less than 500 observations. However,
here we have got autocorrelation estimates at surrounding time inter-
vals toa, which gives an indication of the roughness of the estimation
procedu:re. Furthermore the coefficients were estimated from a number of
series differing intime •
.Duxing series of days with rapidly changing MSC the discharge pro-
duced by different physical mechanisms strongly interacts. This may
distuxb the estimation of the autocorrelation function. (See fig. 4.3, where the estimated autocorrelation exceeds 1.) To avoid this phenomenon a
minimum period duration of 5 days was defined. It had to be exceeded without any change in the MSC, if the period was to be used in the
autocorrelation estimation proceduxe. (See fig, 4.4.)
-
- 2·1 ~ X
1· ,:_
O.::JO
0. 70 \
X. ' 0.50 \ ~ " "
\ I.
X \ I J(
\
0.30 \ I
X
:r.
0.10 X \ \~ _ ,L.__J
0 ~ .. 9 \ 0 -o. 10 .; 00 C\JI ~~ X .... I X I
-0.30 I "I
X -o.so X
X X
Fig. 4.3 Estimate of the autocorrelation of the model residu-als at snowmelt in Stadarforsen 61.10.01 - 76.03.31. Variance of the residuals = ·1. 72 • 1 o3 (rn3 / s) 2 • 95 % confidence limits: x > 400 observations to the left of the line: ,
Minimurn period dUJ'.'a tion = 1 day. X
l(
X 0.90
X
0.70 l(
X
o.so X X
)(
X l(
o. 30 )( lx " I )( )( )( )(
l(
0.10 )(
0 o• d 0 -0.10. ~ CD c-J
. )(
-1 )(
I
-0. 30 "I X
" )( X
-0.50
Fig. 4,4 Estimate of the autocorrelacion of the rnodel residu-als at snowmelt in Stadarforsen 61.10~01 3 76.03.31. Variance of the residuals = 2.35 • 107 (m /s) 2• 95 % confidence limits: x I > 400 observations to the left of the line:
Minimurn period duration = 5 days.
-
- 22 -
4.5 Significance of the autocorrelation
It was considered an important task to determine confidence intervals
around the autocorrelation function and especially to tel1 when it
significantly differe:i from zero.
One method to approach this problem isa significance test developed
by Anderson (1941).
The hY}lothesis that the series isa so called circular series built
upon N purely random observations of a normally distributed stochas-
tic variable, is tested. This method was not used because of the
difficulty of merging the different significance limits from separate
time series, differing in length inta an overall significance limit.
Assurning that the residuals are normally distributed and that the
estimated mean value is correct, we will obtain:
N.-,-1
= 1 I:J [R~(v) + R.(v+-r) R.(v-,)] (1 Nj V= -N.+-r+1 J J J
-r+lvl) N.
A
where R . ( -r ) J
N. J
J
estimate of the autocovariance function of period j,
duration of period j (days).
A
This method is described by Hjorth (1976). Assurning the R.(-r) of J
different periods j with the same mechanism working as independent,
we get
Var [iicTJ - ,t; -TJ 2 "Vax [R.(T)] all -r J
periods j
Confidence intervals were computed, assurning the R(-r):s to have
normal probability distributions. Then the 95 % confidence interval was given in the form:
R(-r) ~ 1. 96 • v Var [R(-r) ]' When plotted in the autocorrelation graphs all confidence limits
A
had to be divided by R(o) in order to get the correct scale. The
prodedure led toa slight paradox, the upper confidence limit
might exceed one. This is of course impossible, but the autocor-
J
-
- 23 -
c-o,_r.c:;:ciance g.raph. 'i'he conf'idencc L.1ni t s werG s;('.llU1etricall;:t
_placed. on both s ides of the autocorrela tion est:;_c!!att:.
Analysis of the autocorrelat:'..on
::::h.rring the beginning of the statistfoal study the :.utoco1:T8la-
tion coefficients were believed to contain valuable inforn1a-
tion.
The kind of i ntermi ttent processes descri bed here are not
knmm to be fi tted wi th any standard estirnation methods. .An
aim of this work has been to get unb i ased estima tions. Wi th
increasi,ng number of observations the observed quantity
tums more and more normal , and then an unbiased estimation
is preferable.
The three MSC are:
1 • M > 0 ( snowmelt) ,
2, M 0 and (s or 6S ) > 0 C,·-flow), UZ UZ '
3. M 0 and S = LIS = 0 (low flow). UZ UZ
snowmel t (mm),
s UZ
~ storage in the upper zone of the model
6S UZ
""inflow in the lr lT lt lt Il Il
:uv.ri ng snowrnel t we have got a c l earl y s i gnificant autocorrela-
tion (see fig. 4.5 and 4.6). This shows us that a large resi-dual du.ring one day will gi ve rise to large resi dual s du:ring
the following days. If, for instance, bad representativeness
of the temperature measurements causes false sno-wrne lt one day,
the reservoirs of the model are filled up toan improper level.
This affects the r esiduals du.ring the following days.
-
X
,c
0.90 X
X
0.70 '\_ " X
0.50: X X
X
0.30 X X
,c
X
0.10 X
0 OX
-0.10 . ..: CD X
-0.30
-0.50
X
X
lx
9 ~
X
XI I
X
- 24 -
lC X X X
X X X X
X
0
ci C\I
lC X
lC
o x x•x x x6 T CD N C\I N (I')
0
ID (I')
X X X X X X X X lC X
X 0 X •
0 .,.
X X X X X X X X lC X X X X X
Fig. 4.5 Estimate of the autocorrelation of the residuals of
Stadarforsen (snowmelt) 61.10.01 - 76.03.31.
Variance of the residuals = 2.35 • 103 (m3/s)2
•
95 % confidence limits: x > 400 observations to the left of the line:
X 0.90
X
0.70 X
X
o.so lC X
0,30 X X X X X
X
0,10 X lC
0 X
-0.10 • . 9 . t0 X X ~X X T
X X --: X ,c 400 observations to the left of the line:
X
T
T
-
-0. 30
-0.50
- 25 -
cr12 fact.s tha:t~.
"
lund of noise that lies on top of "d_e o.::.scharge r:ecord
in Kultsjön, due to the methofi_ (3ergströrri ·1976; used
when estimating inflow.
Fic-. 4.7 Estimate of the autocorrelation of the residuals of Stadarforsen ( Y-flow·) 6 L 10. 01 - 76. 03. 31
Vari ance of the res i dual s = 33c (m3/s,) 2
95 % confidence limits: x > 400 observations to the left of the lLne:
-
- 26 -
0.90
0.70
0.50 X
X X
X X 0.30 I X X X
X X X X
X X X
0.10 X (days) X X X
0 0 0 0 0 -0.10. N ~ ~ w
• ci T ~ X - - ~ X X
X
-0.30 X X X X X
X -0.50
X
Fig. 4.8 Estirnate of the autocorrelation of the residuals of Kult-sjön (y-flow) 62.10.01 - 76.05.18.
Variance of the residuals = 258 • (m3/s) 2
95 % confidence lirnits: x > 400 observations to the left of the line:,
I Note that due to the small number of observations the maxi-
mum argument above is just 19 days.
When the MSC shows y-flow (fig. 4.7 and 4.8), we have also got a sig-nificant autocorrelation. However, it is less than in the former case.
A reasonable explanation to this is that the reservoirs may still be-
come filled up toan irnproper level but not up to the same high level
as during snowmelt. This makes the residuals, separated by a shorter
tirne period, independent of each other.
-
"
0.50 "
0,30
o.w
-0.10
-::i. 30
-::i. sa
0.90
0.70
0.50
0.30
0.10
-o. 101· )( -0.30
-o.so
- 27 -
\
" )'. :~ X X
X X X X
:,(
" " " X " X )< " X )C X
)C
0 0 0
----------................ _____ _ X )C X " X )( X " )( X " -1!. • "6" 0 0 0 0
a:i .
st" N (0 . .
ai c-i 0 ..,. N N C\i (I')
Fig. 4.9 Estimate of the autocorrelation of the residuals of Stadarforsen (low flow) 61.1C:.01 - 76,03.31.
Variance of the res i dual s = 19.3 (m3/s) 2 • 95 o/~ confidence limits: x > 400 observati ons everywhere.
)C
>< X X " ( days)
Fig. 4.10 Est imate of the autocorrelati on of the residuals of
Kultsjön (low flow) 62 .10. 01 - 76. 05.18.
Variance of the residuals = 47 .1 • (m3/s) 2
95 % confidence l imits: x > 400 observations everywhere.
-
- 28 -
At low flow there are great differences between Stadarforsen and
Kultsjön (fig. 4.9 and 4.10).
The noise of the hydrograph of Kultsjön leads to the assumption
that any autocorrelation in Kultsjön during low flow is masked by
the noise on top of the recorded bydrograph.
In Stadarforsen this is not the case. A nice smooth recession curve
gives us large autocorrelation estimations.
If the reservoirs of the model are filled up toan improper level
during or before the winter recession, this will cause a very
persistant series of residuals.
-
- 29 -
5. A STUilY OF RESPONSE SURFACES OF CRITERIA OF FIT
5- ·1 Cri teria of fi t
Nash and Sutcliffe (1970) defined the R2-criterion of fit as
the efficiency of the model.
where
l: ( ~ ( t) - ~) 2 - It ( ~ ( t) - Qc ( t) >2 2 t
R =------------------l: (~(t) - ~)2 t
~(t) observed discharge at tirne t (m3/s),
Q (t) computed I! I! ,r It H C
~ ::: arithmetic mean of ~(t).
The R2-criterion of fit is widely spred and it hasa relative
character, which makes it attractive when comparing the fit of
different models, different time periods or different catch-
ments.
These comparisons should not be made between hyclrographs that
differ too much, because of the phenomenon illustrated in fig.
5.1 and 5.2, which could be due to one or both of the follow-
ing explanations.
1. Inevitable errors of roughly constant size cause the
R2-criterion to give high values when applied to hyclro-
graphs with high initial variance and vice versa.
2. rr·he calibration of the model and the model i tself
favour fit during periods with high initial variance,
and so the relative fit is bound to be worse during
periods with low initial variance.
-
0..(1/s) I. 0 0
200
- 30 -
OBSEAVE:D HYDROGflAFti
-4f-- COMPUTEO HYOROGRAPH
R2 =- 0.64
MAY JU 1-1 JUL AUG ' SEP OCT
F:c. 5.1. Lcw P.2-value as a result cf lo,-, initial variance (Stab'c:::r, 19::?;. (From Bergström, 1976,)
Q(L/~) 600
,oo
200
0 BSERVED HYDFWGRAP+-1
COMPUTE O HYDROCiRAPH
a~--..-----~--=~~-~~_..,~-f"-..!t,:,.,,!=.~---.-----1 A:,..,Pc--:R'-'---'---'--'~'---'--....:....;=....:..c..._ ...__....:J:....U;...L"--_._--'-'=--">--'--.-J
-
- 31 -
Let u2 E,ssume that the R2-c:citeri& of ·ilfferent, pnysical p1:0--
cesses do give a more accurate measure of the goodness oi fit
between the computed and the :recorded l--:i.yc,l',.,g-:::. Then the
problem of merging the three cri teris, im;o one arises. One
possibili ty of doing this i8 simply t0 add u;i "'.;l1s three cri teria,
thus achieving the R -criterion. sum
5.2 Basic assumptions and observations
The MSC gives different classifications of the material at
different parameter settings. This sometimes made the R2-
criteria vary unexpectedly (especially the R~-criterion). If
a couple of days are moved to the low flow class from the
other classes, this will probably make the initial variance
(F2 ) of the low flow class greater, but it might possibly not 0
influence the sum ~ (~(t) - Qc(t)) to the same extent. Thus, the R~-cri terion will grow perhaps wi thout a..n.y visible change
in the hydrograph, a phenomenon similar to the one illustrated
in fig. 5.1 and 5.2.
The HBV-model has always been calibrated by visual inspection
of the computed and recorded hydrographs. This means that a
considerable skill in parameter setting has been obtained
during the years. For example, recently on the first try when
calibrating the HBV-model fora new catchment the R2-value of w
the four year ca.l_ibration period was greater than 0.8.
This implies that a fairly good parä.!Jleter setting could be
obtained by a QUalified guess based on experience from other
applications. It is likely that an automatic parameter optimi-
zation routine would accomplish this too.
This assurnption made the work easier, while only the region
around the optimum parameter setting had to be examined. The
subjectively found optimum (tab. 5.1) was used as the actual one, and the parameters were varied around this central point.
-
- 32 -
Table 5, i Criginal settings of the free parameters of the HBV-model.
Parameter Kultsjön Stadarforsen
p 1.330 1. 136 corr
T ( 0 c) 0.5 o.o 0
C f' s__,_ 1. 23 0.90
C (mm/(0 c
0 day)) 3.2 2.0
C ·wh 0.05 0.05
sb (mm) o.o 0.0
C rfr 1 • 0 1 • 0
Fe (mm) 150 150
L (mm) 150 150 p
s 3.0 1. 5
K (1/(s • mm)) 0 12500 0
K1 (1/(s • mm)) 4000 3500
K2 (1/(s • mm)) 300 350
Luz (mm) co 15
C (mm/day) 1.3 1 .o perc 1l (days) 2.0 6.o max
C (day . s/m3) 0.00103 0.00000 route
-
- 33 -
!fote that K in Kultsjön is zero. That version of the :i"TIV-o
model is older than the one used in Stadarforsen. :t is occa-
sionally used when the catchment behaviour is deterrnined to be
sufficiently simple. Because there were no initial values of
K and~ , a variation of these parameters in Kul tsjön was 0 UZ
avoided.
During the original subjective calibration 8-year periods were
used, but due to the limited capacity of tte computer only
4-year periods were used in the study of the response of the
R2-criteria. This limits the value of the comparisons between
the model behaviour of the original parameter setting and the
test settings of this study. The periods studied are in
Stadarforsen 1965-10.01--1969-09-30 and in Kultsjön 1962-10-01
--1966-09-30.
Two, sometimes three, parameters were varied simultaneously,
and the R2-criteria were computed at the different parameter
settings. This resul ted in tables and "three-dimensional 1'
d . f . R2 2 "'- d iagrams o iso- graphs. The R -curves were only u..i.awn aroun
the optimurn point.
If there was a substantial difference between the R2-values at
the optimum and at the original parameter setting,
the cooputed hydrograph corresponding to the
optimum values of the ~arameters was plotted. These hydxo-
graphs were later judged by the model calibraters in order to
estimate the goodness of fit at the new parameter settings.
-
-·· 34 -
K2 100 200 300 400 500 600 700 800 900 1000
C perc
')
12
15
18
21
24
27
33
\ . \
. 727 ~J_JO , 729 . 727 . 725 , 723',----_ 721 ~ .
,727 732 .73~30 .728 ,726 .724
\ . 725 \. 731 . 733 . 733 . 10,~o . 728 • 726 \. 722 \ 730 . 734 . 734 . 733 . 73~"'--_:z30 . 728
, \ \ , 7 3 5 · 7 3 4 . 7 3 3 · 7 ~ 7 30
K2 100 200 300 400 500 600 700 800 900 1000
C perc
9
12
15
18
21
24
27
33
.524 .524 .524 .520 . 519
Fig. 5,3 The response of the R2-criteria to K2 (1/(s • mm)) and
C (10-1 mm/day) at Kultsj ön. (See chapter 5,3; 5.3.1) perc
-
- )) -
H2 3
K2 100 200 300 ~00 500 600 '700 800 900 1000
C perc
9 .145 .233 .254 .248 .232 . 12 .161 .274 . 306 .306 ,293 . 15 .151 . 286 .328 .332 .323 . 18 .141 .292 .344 . 35_1 . 21 .113 .285 • 47 ~.360-:=.=;-. -;:,•
--- --- --- / .316 24 .079 .265 .33 .354 .3Sö .354 ,347 .337 .327 27 .067 .260 350 .341
30 .024 .229 .307 .353 .341 .334
33 .009 .220 .300 .334 . . 358 · . 347 36 .336 . 341 • 341 .339 .335
R2 w
K2 100 200 300 400 500 600 700 800 900 1000
C perc
9 • 777 .783 .783 .782 .700 . 12 .780 .788 .789 .788 . 786 . 15 .781 .792 ,792 790 . 18 .780 .791 .795 .795 . 793 ">
.796 .796 ~
21 • 777 .791 .795 . . ~ . 24 • 772 .']89 .795 .797 .796 .795 .793
-~~787 27 .766 .7 .794 .796 .797 .796 .794 .793 .792 790
30 .759 .781 791 .795 .796 .796 .795 .793 .792 .790
33 .751 • 776 .78 .792 ,794 .795 .795 ,794 .792 .791
36 ~ ,794 ,794 ,793 ,792 .791
Fig. 5.3
-
::-:,f K1_,1_j_ ts 7ör.!.
the j_r_j_ tig_l valucs :)f the ,1,u•;ir_,_ei:,,~::-•s spe tab. 5. 1.
The response of tica ii'~ -cri teria to K0
and C - - - - - - - - - - - = - - _p~r~
')
C::hc va:r·ia";ic:i of e,',ch p_-'- -c.~:-i tericn was c.omputed when al tering K2
and
C , Since t~'lese •ia.1:·ameters kLve thcir grcatest influence on ::..ow per8 ~ -
flov, the R':-criterion was considercd most important. ln fig. 5.3 the . -s ? /
n;-criterion lndicates )
tha-t C and K2
should be ir.creased. The r, perc
i:rrei;ular i,h2.:;ie of ~}'e R~-rc-sponse surfacc is duc to thc variable ') -
cl;,,ssificai;ion ·'Jy the MSC (chapter 5.2).
An increase in C will have ~he effect of increasing the overall perc
flow during low· flow. It will 2.lso cause a shorter duration of the
pea.1.::: flows. l:m increase : . .L i::? ,.,ill accelerate the low flow recession.
A test :;ilr::nting was made at;
K - ''·rJ·n 1;1s • 2 - ,_,, ,..,,
C perc.
w.:--1:'..ch is the o:;itir:rwn point du.ring low flow.
Tne ne,., hydrograph was considerecl to be somewha t ·c,etter than the
old one.
? The response of the R--criteria to K
1, :B and C t
max rou e
The R;-criterion in fig. 5-4 ob-rjously points towards a decrease in K. a...nd C t • A decrease in K
1 ·,1ill ::na.ke the hydrograph morc d.amped. , rou e
and will also cause a consiuerable number oi' days to movc from the
low flow ciass to the Y-flow class. The latter is the reason ?
why we get such odd. informati on :..':::-om the B.3-cri terion here.
C . 1 too, affects the variancc cf t°'J-? hyrirograpL A l ow C t roui;e rou e value will damp the discharge peaks, ~rh:~:::..e a :'ligi value .-,ill make
the hydrograph v~ i n a 111orc rap:, c. way.
-
7
2
·)
5.Li. 'l'he response of tl1e RL-criteria to K 1 (l;(s · :mn)),
r\ (days) and C (clav · s;m3 ) at Kultsjön. max · route C
(See chapter 5,3; 5,3.2)
-
--'---, I -
,-,
'-'route 0 • 00 1 day / ( ,n J / s ) ,
f: na.x 2 days.
Eere the rrwde L demonstrates an unability to move rapi dly between
k"ign a..Dd medium flows by consistently underesti.mating high flows.
This underestimation was ccmsidered serious and ca.used the ealibra-
02:cs to ,1uö.ge the plotti.:.'l.g to be not as good as the original one.
This Ja,ck ·:::i-f a K parameter is likely to be the cause of this 0
d.arrrped b e2haviour of the 11'.odel, since one further storage di se har ge
para.meter w(mld increase t;Je s l ope of the recessi on.
The response of the R2-cri teria to Fe, L /Fe and B - - - - - - - - - - - - - - - - - E - - - - -
The parameters varied i.r. fig. 5.5 affect the evaporation. They also affect the lcvel of the flow peaks succeedi21g dry periods.
·')
TLe c:ptirrur:, of the RL -cirterion seems to be: w
Fe 200 mm,
L /Fe p
O. 6 - ·-L p
2 or 4.
120 l!llll,
s:'wo test plottings are made, one at S = 2 and one at r3 = 4. They do not differ :nuch fron, each other, but they differ from the original
plotting. The decrease in the L /Fe ratio causes an increase in the p
Eivaporation. This l e ads to an 1mder e stimation of the high flow
peaks of sUlillller. Thus, the test plott ings were i nferi or to the
origi.nal one. The medium and. low summer flows are perhaps a bit
better. on the test plotti.r..gs than on the original one. Agai n the
lack of a third storage diseharge parameter (K) i s obvious. 0
:Pig. ").') (See next page). ,·,
The reSp:)nse of the R"-criteria to Fe, L /Fe and 13 at Kultsjön. p
-
1.0
0.6 .,/L---7"-'----,'1,---,.,C--'-~ s ..._ ..._s ,,_, ,,_,s
R~ R. 2 _L., -) __ 1,r:::::.------;.-----,.,.:::----,oi
Fig. 5.5 (See chapter 5.3; 5.3.3)
-
.. 40 -
Tl. ~ th -, 2 . t . ' C , ne response 01 e h -cri eria i,;o · f ano. 13 s_
p and C sf l)oth affect the los ses between precipi tation and runo::'.'f.
The differences ·oetween the c,ptima of these response surfaces and
the R2-values of the original parameter setting was small and no
test plotting was made.
n V
si'
1.0
: • 1
1 • 2
1 • 3
1 • 4
1.0
1 • 1
1 • 2
1.3
1.4
2 P.1
, I • 2. 4, 8.
.699 .690 .671 . 651
~ ~m:-~ri.s .700 . ~ ...... ' •, "- .. L 'I3]__./[ ~?___:>])1 3 =- ·-·-·-----~----,686
.c-01
.396
.334
.680
.603
2.
•)
n"' 2
.678 .675
.608 .610
4.
,466
·16.
.637
.689
,706
,672
.609
-359
. 377
,375
,·,
C si'
'1 • 0
1.1
1 • 2
1.3
1.4
1. 0
1.1
1. 2
1.3
? n3 1 • 2 .. 4, 8, 16.
.317\ .323 .319 . 315,, .306 \,_:/ //
.313 .319 .315 (:309 ,305
~8 .315 \ ··, . 312/ 310 .305 ·-., --- - ____ ,
.303 .301 .3os .304 .302 .. -----"-..
:·299 .20 .305 .301,.,.--:-299
,., c:.. 4. 8. 16.
,1§5-- .774-.. __ !766 .748 ,728 .,,.. ----.... ,,,.-;: .. ----~---------- "'·• .. __ • 783 /0',, Li-·'" 790., ':1:J6 • 758
I • ..._ ""-.. l ·, '-, '-
:Jtl '\.____!]})_ ~ __}23 ) • 768 ·-, . --~--~----~ ------ _ . ..,....,., .-·
• 753 -··:'i-C~---++.:+---:763 . 750
,708 .722 ,730 .724 ,712
5,6 The response of the R~-criteria to r and S at Kultsjön, ....,sf
-
- 41 -R2
1
C 0 1.4 1. 7 2.0 2.3 2.6 2.9 3.2 3.5 3.6
T 0
-1.0 ,670 ,714 ,672 .59e
-0.5 ,612 .709~ .612 ,539 .466 .399
o.o ,468 .610 • 70 , • 716 ,668 ,618 ,550
0.5 • 221 .425 ,532 ,633 .689 ,706 , 717 .697 .665
1.0 ,473 .546 ,582 ,608 .620
1,5 .211 .251 ,307 ,357 ,395
R2 2
" "o 1. 4 1.7 2.0 2.3 2.6 2.9 3.2 3.5 3.8
To
-1.0 .513 ,482 ,55~67 :_. / . -0.5 .449 ,478 -492 .551 • 556 .555 .515 -492 .446
o.o .292 .419 .446 .503 .549 .517
0,5 -570 -318 .450 ,447 -492 .540 ~585
1.0 .433 ,452 ,499 .511 .553
1.5 .344 ,398 ,408 .467 ,500
R2 3
C 1.4 1.7 2.0 2.3 2.6 2.9 3.2 3.5 3,8 0 T
0
-1.0 ,277 ,299
-0,5 .273 .285 .336 .337 ~314
o.o .263 .278 ,337 .338
0.5 .243 .265 .274 .281 .323
1.0 ,286 .292 22
1.5 .274 .282 .285 .288 -295
R2 W
C 0 1.4 1. 7 2.0 2.3 2.6 2.9 3.2 3,5 3,8
T 0
-1.0 , 745 .765 ,746 .704 . -0.5 • 706 ,571
0.0 .610 .697
0.5 .471 . 766
1 .o .656 ,697 ,724 • 741 • 753
1,5 -498 ,540 , 575 .609 .636
Fig. 5.7 The response of the R2-criteria to T and C 0 (mm/( 0 c • day)) at Kultsjön, (See chaiter 5.3; 5.3.5)
-
- 42 -
mh f th R2 't . t T ~ e response o e -cri eria -o 0
and C 0
T affects the starts of the mel t periods, "11hile C affects the 0 0
overall melt ratio (fig. 5.7;.
A test plotting was made at:
T 0.0 oc 0 '
C ') , - I (0 . day)., ~.o mm/ C 0
The test plotting showed out to be inferior to the original plot-
ting. There seems to be nc way to :nake the model fit the recorded
hydrograph on both peak flow and medium flow. One further degree
of freedorn is needeä..
The response of the R2-criteria to C ~ and C f
W11 r r
The paramete:rs Cwh and Crfr influence the behaviour of the rnodel,
when a rnelt period is restarted after a short interruption by toa
low ternperatures. The use of the Crfr parameter started during
this study of the HBV-model, and the value of having such a para-
meter was questioned.
As the response surfaces (fi6', 5 ., f'.) show, C f does not affect the r r R
2-criteria very rnuch. so the C f. parameter seems to be of no use
· r r here.
The optirnum value af C ~ is obviously 0.05. No test plotting was Wll
rnade, since the optimum values of Cwh and_ C f do not deviate from r_r the original ones.
-
C,00
0.10
o. 15
0.20
crfr
Cwh
0,00
0,05
0.10
0.15
0.20
,701
,709
,714
.709
,700
o.oo
.566
.566
.564
.562
-556
,701 ,70~ • 70·: . 701
,704 . 7ce("-=~.l;r-~ 7 ~ J .7~7 -----:tCTt---;-t"tOr--:fi~~~- 677
.665 ,GGo .GG1 .644
.635 .6:;o .62.e .604
R2 2
0.05 0.10 0.20 0,40 1, 00
.566 .566 • 5(;(; . 66
.562 59 ,556 .556 .551
)61 ,557 750 ~-54/43
,547 .543 .54•J ,534
0rfr o.oo 0,05 0,10 0,20 0.40 1.00
Cwh
0.00 .326 ,;,26 .326 .326 .326 •
0.05 .328 .;, ,6 .31 / .~314 .313 ----0.10 .323 ~~-295 0.15 .320 ~286
0.20 .313 .287 .284 .283 .281
0rrr o.oo 0.05 0.10 0.20 0.40 1.00
r ~wh
0.00
0.05
0.10
0.15
0.20
,770
• T(8
,784
,785
• 783
.770 .770 .770 . 770
:-r~. 7 ·'3 5 .. 7P.7 ,785 .790
~ 7~__JSC-. -~ 778
• T,'3 . 77'1_>~. 765
.761 .760 .759 .74e
-
1 r .. \ ... , 50'.)
1. 0 . 6'./ ) .895 , 8;•6 .8?5 .89'.J
1.5 ' . STJ
-
- 45 -
5.4 The response surfaces of criteria 01 fit and test plot-tings of Stadarforsen
In Stadarforsen runoff data of better q_uality tha.n in Kultsjön
were available. For the interpretation of the model parameters
see chapter 3 and for the initial values of the para.meters see
tab, 5. 1 •
5. 4. 1 The response of the R2-criteria to K
2 and C
- - - - - - - - - - - _p~r~
There is no difference between the optimum parameter setting
here (fig. 5,9) and the original one, and so no test plotting
was made.
There has been difficulties in finding the optimum setting of
K2 and C • Eut note that there isa clearly observab~e opti-perc mum in the dry sum.mer and winter recession class here.
5.4.2 The response of the R2-criteria to K1
, C and B route max
The original parameter setting is within the limits of accept-
a.nce in fig. 5.10. The R2-criterion does not vary much when w
B max
is varied in a region around the original parameter set-
ting. To study this apparent independence a test plotting was
made at:
K. 3 500 1/s • mm, I
day/(m3/s), C route 0
B 5 days, max
where there isa small improvernent of the R2-criterion but the w
new plotting had approximately the same fit as the original one,
as judged b;y the calibrators.
-
6
6
5
Fig. 5,10
8 max
i J >c ro ute
~,-,J.~~~~-~---+ TSOO
fa-,,,L,L....,,4.:;ff""'-,.L'----"?'------,~ 2500
,~,
7
6
5
;rhe response
(day • s/m3)
5.4; 5.4.2)
of the Hc::-criteria to K1
(1/(s · mm)), C t rou e
and J (days) at Stadarforsen. (See chapter max
-
.i:>'ig. 5. 11
K0
7~00 10000 12'./JO 1
I. ll.Z
5 ,796 ,850 .870 .0'/5 ,670
10
15
20
25
30
,794
,790
,783
• 774
,851
,850
,846
,838
R2 2
K0
7500 10000 12500 15000 17500 20000 22500 25000
L UZ
5 ,794 ,752 .692 .628 .567
10 ,819 ,809 • 782 • 753 , 721
15 ,8
20
25
30
R2 3
K0
7500 10000 12500 15000 17500 20000 22500 25000
L UZ
5
10
15
20
25
30
~6 ,5 7
5 1\ .543 ~ ,496 .504
,483 .492
.508 ,487
r:,-
550 ,549
.505 .505
,496 .496
,453
.524
,549
. 512 .51.1 ,516 .519
.496 .504 .504 ,504
,491 ,492 ,492
K0
7500 10000 12500 15000 17500 20000 22500 25000
L UZ
5 .sG7 ,895 .902 ,900 .894
10 • 867
15 .865
20 • 861
25 • 855
30
')
(1/ ( s u.m)) 'I'he response of the -::;~ -cri teria to K . 0 , \
and T (mmJ at Stadarforse11. (See chapter 5. 4; 5.4. J_j 77'7
-
- 48 -
5.4.3 Th f th R2 ·t . t K e response o e -cri eria o 0
andL UZ
K affects the top flow recession rate and L the level, at which 0 UZ
this recession rate is activated. A great improvement of the R2-
criteria can be seen here. It is only the R~-criterion that is not
improved when increasing K and L • (Fig. 5.11). 0 UZ
A test plotting was made at:
K 22 500 1/s • IllIIl, 0
L 25 IllIIl. UZ
The plotting corresponds to the information available through the
R2-criteria. The result isa better overall fit, eHpecially at
high flows but perhaps there is same deterioration at low flow.
5.4.4 The response of the R2-criteria to K0
and K1
In fig. 5.12 we have got an example of
practically governs the R2-criterion. w
the optimum of the R2-criterion. w
K 16 500 1/s • mm, 0
K1 3 500 1/s • mm.
2 how the snowmelt (R1
) class
A test plotting was made at
It showed out to be somewhat inferior to the original plotting. The
difference, however, was considered to be of minor importance.
Fig. 5.12 The response of the R2-criteria to K0
(1/(s • mm)) and K1 (1/(s • mm)) at Stadarforsen. (See next page)
-
~ ig. 5. 12
, -..,_,
K 0
8500 10500
Kl
1500 .820 .360
2500 .819 .858
3500 .819 .857
4500 .820 .B5T
5500 .821 .857
2 R1
12;,00
8Sl
879
79
78
78
R2 2
,i,1, 5C" 2-6: ·~ ~ J C 1..- .. ,! ~ _',.)
1· = '.il . C94 I I
\::i:,J \
.69:+ I ,.
I I
·~ . .393 ") . C,/ .. · ,:7:_ .893 .S9:; 89c, • c, c',
.SS8\ .892 . .
K0
8500 10500 12500 14500 16500 18500 20500
1500
2500
. 765 . 744
3500 .835 .836
4500 .824 .827
5500 .810 .816
K 8500 10500 0
Kl
1500 .380 .414
2500 .452 .466
3500 .528 .543 -----
R2 3
12500 14500
.440 .458
.475 .479
,550 .549
4500 ~560
5500 5 , I • 542
.794
16500
.472
.478
.549
-557
.539
.800 . 789
.803 . 794
18500 20500
,550 551
.555 .552
K 8500 10500 12500 14500 16500 18500 20500 0
K 1
1500
2500
3500
l:500
5500
.8tl2
.823
.6iL
. f.6~
.882
901 .90'
.90,1 I
. 90~ \
.903
.916 920 .921
.'?l'T 92':' r'•)":I ,::,'.___)
-917 ,q22 .921, .923 .921
,9lb . 02~~ • 9:__:~ .923 , 92]
.914 ()2,J .922
-
Fig. 5.13
- 50 -
The response of tne R2 -eri teria to Fe (mn ), L /Fe p
and 8 at Stadarforsen. (See enapter 5-4; 5.4.5)
-
5. 14
o.e5
0,90
0,95
1.00
C
csf
0.80
0.85
0.90
0,95
1.00
~
csf
0,80
0,85
0.90
0.95
1.00
.600 ,796 ,799 ,818 ,833
rl 2
0.5 1.0 2.0 4,0 5.0
.284 .11, 1i~~ \ .no .62,1 .262 .707 5 \ ,791 .696
·'"I/ '\; \\ace .235 . 713· .205 .684 d8~6 \a20 .. 729 , 151 .657,1 .s50\ I Je20 . 732:
2 R3
o.s 1.0 2.0 4.0 8.0
.518 .519 .525 ,440 .254
-5~ ~}~1 .473 .310
.s, \ ·rr 1 ·~t .503 ,355 ~o ll:,Pd4 I ·:p .. 522 ,397 • ., • •◄ 1 · ) . "
~I I ' I I ,498 \l,574, .::,5å ,534 ,430
-~"-,e respomw of the ~,/ -cri teria t o C &id s~ at Stadar-
-
- 52 -
5. 4. 5 The ~e~p~n~e _ o: ~h~ ~2 =c~i ~e~i~ ~o _ F~, _ \/~c _ ~ci _1~ _ 'I'he danger of letting the R2-criterion define the parameLer optimum is
1,,r
accentuated in fig. 5.13. A test plotting was made at: \ Fe= 150 mm,
~p a :5° ,m J This is the n.2-o::_JtiI;i.wi-1. The model now underestimates y-flow. From the
w response surface it is clear that the R~-criterion was deteriorated by
this modification of the parameters. This gives an unacceptable error
in the accwnulated flow after the four calibration years.
The R -criterion SUlll
(chapter 5.1), however, has i ts optimUlll at:
Fe 150 mm,
L = 150 mm, p s "' 2,
which is very close to the original setting.
5 4 6 Th f th R2 ·t . t C d • • e response o e -cri eri a o sf an 6
The D -criterion (chapter 5.1) shows optimu.m (comp.are fig. 5.14) at: sum
S 2.0,
csf 0.9.
No test plotting was made, because the original parameter val ues
are close to this optimum~ and further more the test plotting of
chapter 5.4.5 showed no improvement of the nydrograph.
The response of the R2 -cri teria to '=' and C 0 0
According to fig. 5,15 C seemed to be a bit too smal l , and a 0
test plotting was made at:
T 0
C 0
0 °c '
2,3 mm/( 0 c · day).
It shows a small improveruent.
-
5. 15
o.c
0.5
T 0
-1.0
-0.5
C
o.o
0.5
1,0
" T
0
-1.0
0.0
- 1
,724 .645 ,590 ,533 .~HO
,733 .733 .697 .534 .6C2
,657 . 721 ~~~:oo .597 ,639 ~'2:,541
,502 .55n ,6C'i ,656 .G35 ,720
1/ l3
1.4 1. 7 2 .0 2.3 ?.6 2.9
.6;:4 ,590 •. 1,:36 .403 . 357
;tp~? .553 ,491
~i.5:S 640 ,6;7 ,560 ~
.569 . 520 ,)05 .G15 .616 .605
.665 .Gcc
4725 . 72~,
~')-~ • j L.J .7'2t
3.2 3.5
-
- 54 -
Crfr 0.00 0,C5 0.10 O.?O 0,40 C.80
Cwh
o.oo
0.05
0.10
0, 15.
,9C:,
.922
0.20 .912 .874 .8~7 ,826 ,778
R2 2
crfr 0,00 0.05 0.10 0.20 0,40 0.80
cwh
o.oo ,740 ,740 ,740 .740 .740 . 0.05 • 756 -756 .755
0.10 , 773 .746 • 726
0.15 .780 ,733
0.20 • 777 .710 .
crfr o,oo 0.05 0.10 0.20 0.40 o.so Cwh
o.oo
0.05
0.10
0.15 .644
.663 .663 .663 .663 • ~ .652 .650 .656 .657 .659 • .651 1 • .625
, 581
0.20 .643 .622 .612 ,585 .532
crfr o.oo 0.05 0.10 0.20 0.40 o.eo
Cwh
0. 00 • 919 t-9. ........ ~'-"-"---!~
0.05
o. 10
0.15
0.20
,929
,935
.935
,929
:tig. 5.1s ~ 2 ·t . t C l'Le response ur tl1e ..:"!'. -cri eria o and C ~ w.: rtr
at Stadarforsen. (See c:'!1apter 5,4; 5,4,8\_
-
'_'Le s-"ttings cf C::rfr and
,~ettinc then seeEted to be (:'ig. 5- -1 C ·_
C ·rLc C wh
().05,
0.05.
,?J .. _ •:,-,- ...... .:. -· ..L
.A plotting was made a.t th:;_s -::i:: int. ::::'he tes-:: -)lotting manageci
to center the spring floods better t:C-1a.r:,, the c,rigir1"":'._ ::_:ilo:ting
and was t!lerefore considered to be i)ette:,r than the original
one,
5.5 Resu'ts of the study ')
It is obvious that the RL-critcrion is nota good criterion of w
fit for our purposes. The sno'WI!lelt peri od g0Yern2 tie behavi-
our of the R2-cri terion too much, }'or cxample, the ;::-parameter
w of Stadarforsen (and to some extent the one oi:' Kultsjöri too) ., is badly optimi·zec1 by the R"--cri ter-; on.
1,:l
-~
J\Ilother disadvantage of the studied c:ri teria is -:;:1e H...::: -- 3 cri teriot:'.. It ~-s cisturbed by t!lc fact that ch.fferc:--it parame-
ter set tings give r1; ,,. _;rent c l ass i fications of data. J3oth the
initial variance and the sum of the squareo_ residuals may l;e
greatl y changed by this phenoEienon. This -:ou:1 lead to a
ch.ange in R~ • not from a change : : :::. -: , but- "'· __ J.J a rearrange-,, m-2nt of the d2,ta 2.vaila'::Jl e.
In order to develci, aI1 acceptabJc cri tcrion of fi t t r_e s1un o:f
,de H~, :.he R; E,nd the R~-cri teria was also studied. The weak-ness of R; de;cri-Ded above, .:-1owever. c.oes in::::_uence tl:is
/
criterion too.
To overcome tl:is :irawback i t is suggested tL2.t tl,c c2-assifica-
tl.on of d.ata shoul ci not -be chan_ged ,iuring cocparisor_1_ of diffe-
rent parameter settings.
-
Parameter
~
Kc2 (1/(s • mm)) (rnm/day)
perc
{
K1
(1/(s • mm))
C t ( days/m3) rou e 3 (days)
max
t1·0 (1/(s • mm)) (mm) UZ ~~
(1/(s • mm))
(1/(s • mm))
te (mm)
L /Fe p s
tf f 'l.' (oc) ~: (m.m/( 0 c • day))
f~wh trfr
Original setting
350
1. 0
3500
0.0000
6
12500
15
12500
3500
150
1 • 0
1.5
1.5
C).90
0.0
2.0
0.05
1.0
ST.AJ)ARFORSEN
2 t· R1-op lJilum R~-optimum
300
1.0
1500
0.0000
5
25000
30
16500
3500
150
1.0
8.0
8.0
0.85
2.3
0.05
0.05
500
1.5
3500
0.0005
7
25000
25
8500
2500
150
1 • 0
2.0
2.0
0.95
0.5
2.6
0.10
0.05
R~-optimum
350
1 • 0
3500
0.0010
5
7500
5
8500
4500
150
1 .o 1 .o
1.0
0.85
- 0.5
1. 7
0.05
1 .o
R -optimum SUill
350
1.0
3500
0.0000
5
22500
25
14500
4500
150
1. 0
2.0
2.0
0.90
o.o
0. 1 ll
n.05
R2-optimwn
w
350
1 .o
3500
0.0000
5
22500
25
16500
3500
150
1.0
4,0
0,/J5
0. Cl
Table 5. 2 'l'he optimum parameter settings of Stadarforsen as judged by the di fforent cri kria of fi t.
'For comparison the original parameter setting ( the optimu.m of an 8 year period as junged 1,y tr1 e
calibraters) is also printed in this table.
-
c--l.f'\
KULTSJÖN
Original R2 opti- R2 opti- R2 opti- R opti- R2 opti-1 2 3 sum w F,:,earn,,ter setting mwn mwn rnum ffilUil mum " ----··-~··-~-· "'"--~-- --~~ ~~--
I ( ; .· (" 111ni)) 300 rioo 2110 600 450 450 2
>)
' 1 wr,_ (n1m/day) 1. 3 3,G 2 •. 2.7 2.4 2.6
k_ I (1/(s . mm)) 4000 3000 3000 6000 3000 3000
(; (,bJ s/m3) u.00105 o.onos O.ll01 0 0,001 IJ,0005 i r,Ju te ', B (days) ,; 2 2 3 2 2 c'.. \_ msx
.. jFc ( rnm) 150 200 200 100 200 200 ! .
~/Fe 1 .o o.6 0_6 o.6 o.6 o.6 3.0 1 • 0 4.u 2.0 4.0 2.0
,· (B 3 .o 1.0 4.0 2.0 4.0 2.0 \c tsr 1. 23 1.2 1 . 1 1.0 1 • 1 1 • 1 r
(oc) ~:o
0.5 ·- 0.5 0.5 o.o o.o o.o
(rnm/( 0 c . day)) 3.2 2.0 3.5 3.8 2.9 2.6 \ 0
i C wh 0.05 0,05 o.u5
I_I, ()7 0.05 0.05 •;
\C rfr 1. Cl 1.U 0. 0:i 1.-.,.05 (1. 4 1 .o
Tat,12 5, 3 The optimum parameter settings of Kultsjön as judged liy the different cri teria of fi t.
For comparison the original parameter setting (the optimum of an 8 year period as judged
by the calibrators) is also printed in this table.
-
- 58 -
The optimu.m parameter settings, as described by the dif:erent R2-
cri teria and the R -cri terion, .ruay be studied ir ta0. 5. 2 a:1d sum
5.3, where they are compared with results fro□ calibrations '-'Y
visual inspection. Note that the RBV-model is usuaEy c8.librs:tecl
during 8-year periods, but :iue to the computer capaci ty ,:,nly
4-year periods were used in the study of the response surfaces c-f
the R2-criteria. We see that the R -criterion is ;rrcstlv closer sum '
to the original setting than -the R2-criterion (tab. 5.4). w
In Kultsjön, however, any studied parameter optimu.r:1 sugges tec. ::i.y
any of the criteria did not give better hydrographs than the original-
ly pJ_.:)tted one. Possible explanations to this are the bad qual.:.. t:,,-
of both the runoff data and the climate data and the inco□pleteness
of the model type used (the lack of a K -parameter). 0
In Stadarforsen the R -criterion agrees with the visual inspection sum
surprisingly well. If same sort c: fixed classification manages to
make the R~-criterion more reliatle, it is obvicus that the HBV-
rnodel could be automatically ca.:.ibratec in Stadarforsen,
Whether the model can be autornatica:::_ ly calibrated for any catchment,
isa much harder question. This study covers only two catchments and
furtherrnore only four years of each one.
The impression of the author is that the R -criterion or same .-, sum
other linear combination of the R'--cri teria might be the basis oi a
useful crit"rion of fi t. Other q_uantities, such as the SWD c-E" the
residuals and the ratio of over- or underestimations of the hydxo-
graph, may perhaps take part in a criterion oi fit too, since
these two quantities are often used during the subjecti ve caliLra-
tion.
-
K u
L UZ
K 0
i, /Fe -.:,'
C . ., s.r
T 0
(; C
,~ . , .. ,h
C rf~
·:.1J.L..:
X X
X
· . .x.
X:
X
X y
X X
-
6. CONCLUSIONS
The residuals of the H:BV-model are nei ther independent nor s·uationa-
ry distributed dUTing the :Jear. Yet this has been assumed by 1;1any
calibrators of hydrological models. A way to get closer to these
assu.mptions is to base a classification of the calibration data on
the different processes governing the c.ischarge and to consider
each class to be a separate set of residuals. By doing this we do
not get rid of the autocorrelation, but the residuals become more
stationary distributed. In fact it is impossible to get rid of the
autocorrelation of the model residuals, since one single climate
measurement er.ror affects the level of the model storage and thereby
the discharge during a series of days.
Knowing that a classification helps irl making the residuals rnore
stationary, the R~-criterion of fit (chapter 5.1) was computed for l
each class. The sum of these R~-criteria (the R -criterion) showed i sum 2
out to be a better criterion of fit than the formerly used R -w
criterion. Concerning the possibility of automatic calibration of
the .d13V-model at any catchment, data of good Qual ity must be demanded.
There must also be a sufficient number of degrees of freedom in the
model. It might otherwise compensate an unability to fo::i_J_ow the
observed hydrograph by producing a too da.mped hydrograph, overesti-
mating low flows and underestimating high ones.
If these demands are full : j lled, the R -cri teri;'):'.1 gives a better SUID
representation of the goouness of fit tha.rc the R2-criterio~. The w
response surfaces had mostly a regular elliptic shape. Therefore it
seems plausible that an opti.rnization algorit:bm such as Rosebrock 1 s
(1960) method or Powell's (1964) method with slight :nodifications may perform acceptably.
The demand above of many degrees of freedom contains a dil emma. l f a
model is equipped with a sufficient number of degrees of freedom, it
will be able to reconstruct almost any di scharge record from any
climate record. ]ut the headpoint in making a model is not to make
-
it comrlex in ord.er to f:_t tJ: .. e calitratior. pericö c::1ly, (si:1ce
this is ~10 guarantee for fi t dur.i.ng :;,ther perioc.s), but to make
it si:1:1;Je and yet ~it 80th the ,::::al i.bra t: ion r>eriod a.ni tl:'.e inde-
pendent ueriod.s.
-
APP:s'lmIX A
The distributions of the residuals (differences between the
computed and the recorded hydrogTaphs) are shown below. Histo-
gTams showing the residuals both separated and not separated
by the MSC (chapter 4.1) are plotted.
A description of the method used when constructing these
histograms can be föund in chapter 4.2. The prescribed minimum
period duration is explained in chapter 4.4.
(days)
600
Fig. A. 1 Histogram s:,owing the residuals of Stadarforsen
61.10.01 - 76.03.31. Prescribed minimum period dl.lration: 1 day.
I'-'Iean = -589 1/ s.
Standard deviation= 22 503 1/s. Number of residu.als = 5 273. Number of exceeding residuals (NER) 337.
-
A 2
120
80
C\I I
(days)
0 0
0 ...: 0 . C\I
(standard deviations)
Fig. A.2 Histogram showing the snowmelt residuals of Stadarforsen,
61.10.01 - 76.03.31.
60
40
C'I I
Prescribed minimum period duxation: 1 day.
Mean = 1 385 1/s. Standard deviation= 41 431 1/s. Number of residuals = 1 156.
NER= 76.
(days)
0 - ci I 0 0 N (standard deviations)
Fig. A.3 Histogram showing the snowmelt residuals of Stadarforsen,
61.10.01 - 76.03.31. Prescribed minimum period duxation: 5 days.
Mean = -4 645 1/s. Standard deviation= 48 485 1/s.
Number of residuals ~ 794.
NER= 49.
-
300
200
120
80
A 3
( days)
(standard 0
0 o deviations) N
Fig. A.4 Histogram showing the y-flow residuals of Stadar-
forsen 61.10.01 - 76.03.31.
(days)
Prescribed minimum period duration: 1 d.ay
Mea.n = 16 1/s.
Standard deviation= 17 553 1/s. Number of residua.ls = 2 008.
NER= 116.
0 . C
(standard deviations)
Fig. A. 5 Histogram showing the y -flow residua.ls of Stadar-forsen 61.10.01 - 76.03.31. Prescribed minimum period duration: 5 days.
Mean = -590 1/s. Standard deviation= 18 185 1/s.
Number of residua.ls = 1 650.
NER= 94.
-
A 4
(days)
120
BO
40
0 0 0 0 _. 0
. .... c-.i I
Fig. A,6 Histogram showing the low flow residuals of Stadar-
forsen 61.10.01 - 76.03.31.
120
80
(Il I
Prescribed minimum period duration: 1 day,
Mean = -743 1/s. Standard deviation= 5 629 1/s.
Number of residuals = 2 097,
NER= 102.
(days)
..... I
0
c:i C)
.,...; C)
(Il
Fig, A.7 Histogram showing the low flow residuals of Stadar-
m forsen 61.10.01 - 76.03.31, Prescribed minimum period dUJ'..'ation: 5 days. Mean = 1 188 1/s.
Standard deviation= 4 393 1/s.
Nu.mber of residuals = 1 780.
NER = 114.
(standard deviations)
(standard deviations)
-
A 5
( days)
120
80
+o
~:::J......Jl...l......L..1-1.....L...J.....L__LL...L.J-Ll...l...LL..L..LJLL.Ll-Ll..l..LL.l....LJ_J_..t::1....Cn:::J (standard ~ deviations)
C\l I
600
+oo
200
C\I I
0 -I 0 ci 0 N Fig. A.8 Histogram showing the low flow residuals of Stadar-
forsen 61.10.01 - 76.03.31.
(days)
0
.... I
Prescribed minimum period duration: 20 days
Mean = -1 578 1/s. Standard deviation= 3 890 1/s.
Number of residuals = 1 260.
NER= 69.
0 0
ci
(standard deviations)
0
csi
Fig. A.9 Histogram showing the residuals of Kultsjön
62.10.01 - 76.05.18. P1:'23cribed minimum period duration: 1 day.
Mea.n = 374 1/s. Standard deviation= 16 905 1/s.
Number of residuals = 4 977,
NER= 274,
-
A b
( days)
120
80
Fig, A.10 Histogram showing the sno-wrnelt residuals of Kultsjön
62.10.01 - 76.05.18.
60
+o
20
c-., I
Prescribed minimum period duration: 1 day
Mean = -545 1/s. Standard deviation= 30 198 1/s,
Number of residuals = 1 185,
NER =74,
(days)
-I 0 ö 0 0 N (standard deviations)
Fig. A.11 Histogram showing the snowmelt residuals of Kultsjön
62.10.01 - 76,05.1 8. Prescribed mmimum period duration: 5 days.
Mean = -2 400 1/s.
Standard deviation = 33 605 1/s.
Number of residuals = 884,
NER = 50.
..
-
60
40
20
iO
20
C\I I
(days)
.... 1
0
ci 0 . ....
(standard deviations)
0 c-,i
Fig. A.12 Histogram showing the y-flow residuals of Kultsjön
62.10.01 - 76.05.18.
Prescribed minimum period duration: 1 day.
Mean = 411 1/s.
Standard deviation= 14 652 1/s.
Number of residuals = 914,
NER = 47. f ( days) :
(standard deviations)
0
0 L...L....c;;;::L...L--=l.......1.-'--'-'---'-.1...J~..L....IL....L...J.......aL..L..J_L..J......L....I......J......L...I......J..-J....J_J....J.....L..J-l::::1....J~ 0
c-.i I
0 . ..... I
0
0 0
N
Fig. A.13 Histogram showing the y-flow residuals of Kultsjön
62.10.01 - 76.05.18.
Prescribed minimum period duration: 5 days.
Mean = 1 178 1/s,
Standard deviation= 16 067 1/s.
Number of residuals = 607.
NER= 36.
-
A 8
I ( day s)
300 l I
~ 200
100
_._, - ---· I , 1 r--.,._R ... i : ,---_ • . . -H-- .· I 1-titt ! . 7
! I ' lii'·,. lliwnn , i 1· ~i I I , -I !, I ' ...._ __
l..-1.-.I.....I..-L...J..J......L..1-.1......L-L.L....L _ _.__._...1..-1......1..,...;.._.L....L_LI I I 1 :k, N I
0 . -I -,
C
0 0
N
(standard ,::.evid. t:'..ons)
Fig. A.14 r:istogram showing the i.ow flow residuals of Kultsjön
62.10.01 - 76.05.18.
C\I I
(days)
Prescri bed mir1imum period duration: 1 day.
Hean = 725 1/s.
Standard deviation = 7 068 1/s.
}Jumber of reaiduals = 2 876.
0 0
c-.i
(standard deviation)
Fig. A.15 Histogram showing thc l ow flow resici.uals of Kul tsjöE
62.10.01 - 76.05.18.
Prescribed minirnUl!l. perioci. duration: 5 d.ays.
Mean = 530 J./s.
Standard deviation= 6 863 1/s.
Number of residuals = 2 545.
NER= 123.
-
120
80
C'II I
(days)
0 - ci I 0 0 C'II
A 9
(standard deviations)
Fig. A.16 Histogram showing the low flow residuals of Kultsjön
62.10.01 - 76.05.18.
Prescribed minimum period duration: 20 days.
Mean = 370 1/s.
Standard deviation= 6 637 1/s.
Number of residuals = 2 117.
NER.= 100.
-
- 1
AppendL: B
LIST OF SYJVIBOLS
B max
B q
C 0
C perc
C rfr
C route
E a
E p
Fe
F 2 0
K 0
L UZ
M
MSC
n
N
NER
N T
N. J
n X
maximum base in the transformation function
actual base in the transformation function
degree-day melt factor
percolation capacity
refreezing coefficient
parameter in the transformation function
snowfall correction factor
water holding capacity of snow
actual evaporation
potential evaporation
maximum soil moisture capacity in the model
initial variance
stora.ge discharge parameter of the upper zone
slow drainage stora.ge discharge parameter of the upper zone
stora.ge discharge J"''.L'am.eter of the lower zone
limit for potential evaporation
limit for slow drainage of the upper zone
snowrnelt
hlechanism separation criterion
tr,tal number of observations in a class
number of observations in a continous period in a class
number of residuals not contained in the histo-gram. of the estimated density function
total number of observations of R(,)
duration of period j
sum of autocorrelation coefficients
-
B 2
p
p COIT
p lapse
p. l
p w
R sum
R2 w
R2 1
R2 2
R2 3
R( T)
R( T)
R.(T) J
s
s s
s sm s
UZ
T
T 0
precipi tatic,n
rainfall corrsc __ ur, J.c,c t,:.:c.·
area elevation corr
-
T lapse
t
t. J
u
Var(X)
X
X
s
y
]J
a
T
area elevation correction of ternperatlll'e
time
starting time for the j:th period
test variable
variance of the stochastic variable X
a residual regarded as a stochastic variable
mean value of the residuals
parameter of the soil moistlll'e zone
indicates rain or recession succeeding rain or snowmelt
mean value
standard deviation
time step
:B 3
-
B 4
LIST 0F REFERENCES
Andersson, R L (1941)
Bergström, S ( 1976)
Clarke, R T ( 1973)
Hamon & Harman (1963)
Hansen, E (1971)
Hjorth, u ( 1976)
Jenkins & Watts (1961)
Kendall & Stuart (1967)
Nash & Sutcliffe (1970)
Powell, M J Il (1964)
Rosenbrock, H H (1960)
Rudemo & Råde (1970)
Distributioi,. _ 1 "'""'--' ial co:e:c·r.=.,.a ;:;ion c:oefi'icient. Arm. Math. Statis-G, S No. 1
Development and application of a conceptual runoff model for Scandinavian catchments. SMHI Rapporter Hydrologi och 0ceanografi Nr RRC 7
A review of same mathematical models used in bydrology, with observation on their calibration and use. Journal of Hydrology" No. 19
Estimating relations between time series, J. Geophys. Res., 81 (21). 6033-6041
Analyse af hydrologiske tidsserier. Laboratoriet for Hydraulik, Ilanmarks Tekniske H~jskole
Stokastiska processer. Kompendium i matematisk statistik III. Matematiska institutionen, Universitetet i Linköping.
Spectral analysis and its applications
The advanced theory of statistics. Val. 2, 2nd Edn
River flow forecasting through conceptual models. Part I. A discussion of principles • .:: au.mal of Hydrology No. 10
An efficient method of finding the minimum of a function of several variables without calculating derivates. The Computer Journal, Val. 7, p. 155
An automatic method for finding the greatest or least value of a function. The Computer Journal, Val. 3, p. 175
Sannolikhetslära och statistik med tekniska tillämpningar. Ilel 1
-
Nr 1
Nr 2
Nr 3
Nr 4
Nr 5
Nr 6
Nr 7
Nr 8
Nr 9
Nr 10
Nr 11
Nr 1~
Nr 13
Notiser och pre~iminlira rapporter
Serie HYDi.JL0GI
Sundberg-Falkemnark, M Om isbärighet. Stockholm 1963
Forsman, A Snösmältning och avrinning. Stockholm 1963
Karström, U Infrarödteknik i hydrologisk tillämpning: Värmebilder som hjälpmedel i recipientundersökningar. Stockholm 1966
Moberg, A Svenska sJoars isläggnings- och islossningstidpunkter 1911/12 - 1960/61. Del 1. Redovisning av observationsmaterial. Stockholm 1967.
Ehlin, U & Nyberg, L Hydrografiska undersökningar i Nordmalingsfjärden. Stockholm 1968
Milanov, T Avkylningsproblem i recipienter vid utsläpp av kylvatten. Stockholm 1969
Ehlin, U & Zachrisson, G Spridningen i Vänerns nordvästra del av suspenderat materiel från skredet i Norsälven i april 1969. Stockholm 1969
Ehlert, K Målarens hydrologi och inverkan på denna av alternativa vattenavledningar från Mälaren. Stockholm 1970
Ehlin, U & Carlsson, B Hydrologiska observationer i Vänern 1959 - 1968 jämte sammanfattande synpunkter. Stockholm 1970
Ehlin, U & Carlsson, B Hydrologiska observationer i Vänern 17 - 21 mars 1969. Stockholm 1970
Milanov, T Termisk spridning av kylvat\t=nutsläpp från Karlsharrmsverket. Stockholm 1971
Persson, M Ryd~'~ 2..ogi ska undersökningar i Lappträskets represen ta ti va område. Rapport I. Stockholm 1971.
Persson, M Hydroloe:;iska undersökningar i Lappträskets representativa område. Rapport II: Snömiitningar med snörör och snökuddar. Stockholm 1971.
-
Nr 14
Nr 15
Nr 16
Nr 17
Nr 18
Nr 19
Nr 20
Nr 21
Nr 22
Nr 23
Nr 24
Nr 25
Heu.in, L Hydrologi:_;ka w1(:E:Y';.-~Ök:-1ine;SLr i ·✓ cle~lS rc1
-
Nr 2G
Nr 27
Nr 28
Nr 29
Nr 30
Ber ge; 1.,röm, S The application of a simple r2,inf'ull-nrnc,f':~ mcde~- ·"o " 2:1:cri-ment w~ th incor:1_piete data coverage. Stockholm lC:172
W~ndahl, T i bergstrand, E Oceano~rafiska förhållanden 1 svenska kustvatten. Stockholm 1973
Ehlin, U Kylvattenutsläpp 1 SJoar och ha,,. Stockholm 1973
Andersson, U-M & Waldenström, A Mark- och grundvattenstudier i Kassjöåns representativa område. Stockholm 1973
Milanov, T Hydrologiska undersökningar i Kassjöåns representativ~ område. Markvattenstudier i Kassjöåns område. Rapport II. Stockholm 1973
-
Nr RllG 1
Nr RHO 2
Nr RHO 3
Nr RHO 4
Nr Rl!O 5
Nr EU-JO 6
Nr RHO 7
Nr RHO 8
Nr R.c"9:0 9
Nr RHO 10
Weil, J G Verification cf heated wo.ter jet nurnericaJ model. Stock:-1clrr. 197 1~
Svensson, J CalculaLion of poison concentrations from a hypothet::.ce.l accident off tbe Swedish coast. Stockholm 1974
Vasseur, B Temperaturförhållanden i ~venska kustvatten. Stockholm 1975
Svenssor., ~ Beräkning av effekt ::. v vattentransport genoo Sunninge sund till Byfjord.er .. Stockholm 1975
Bergströo, S & Jönsson, S The application of the HBV-rwdel ':, O the F::.leCell researcL bo.sin. Norrköping l976
Wilmot, . ..,, A numerical nocel of the effects of reactor cooling water on fjordcirculation, I och !I (bilaga till I). Norrköping 19'(6
Bergström, S Development and applic2.tion of a conccpt,1al rur.off medel for Scandinavian ca-cchmcr.ts. Norrköpir.e 1 976
Svensson, ,J Se1uinars at SHJ-a 1976-03-29--:J4-U1 on ~ifumerical Y:odels of the Spreading of Cooling Water Norrköping 1976
Simons, .J & :2unkquist, L ~:: Svensson, J Application of a rn; ,':e r i c2.l I:J.odel to Lake V:'~r1.err:: Norrköping 197'7
Svensson, S A Statistical Study for Automatic Calil)ration of a Conceptual Runoff J'llodel Norrköping 1977
-
Tom sida