Professional Documents
Culture Documents
Automatic Calibration of The Tank Model L Talonnage Automatique D Un Mod Le Cisterne
Automatic Calibration of The Tank Model L Talonnage Automatique D Un Mod Le Cisterne
Automatic Calibration of The Tank Model L Talonnage Automatique D Un Mod Le Cisterne
M. SUGAWARA
To cite this article: M. SUGAWARA (1979) Automatic calibration of the tank model / L'étalonnage
automatique d'un modèle à cisterne, Hydrological Sciences Bulletin, 24:3, 375-388, DOI:
10.1080/02626667909491876
Automatic calibration of
the tank model
M . S U G A W A R A 6-13-30 Minami-Karasuyama,
Setagaya-ku, Tokyo 157, Japan
Abstract. The automatic calibration is done not by a hill-top climbing method but by a trial and error
method carried out automatically by a computer program. The feedback procedure is made by compar-
ing some criteria obtained from the observed hydrograph and the calculated hydrograph output from the
working tank model. The two criteria are discharge volume and the shape of the hydrograph. The
feedbacks of these two criteria correspond to dispacement feedback and velocity feedback in automatic
control. The output of the working tank model is composed of components, the outputs from each of the
tanks. Correspondingly, the whole period is divided into subperiods, in each of which each of the
components plays the main part. The volume and shape are calculated in each subperiod and are used for
the adjustment of the respective tanks. The feedback procedure starts from some initial model and
converges very quickly after several (usually less than 15) iterations, and the result obtained is very good.
Résumé. L'étalonnage automatique est effectué non pas en utilisant une méthode 'gravir pente' mais en se
servant d'une méthode de fausse position exécutée automatiquement par un programme à l'ordinateur.
Le processus de réaction se fait en comparant des critères obtenus de l'hydrogramme observé et
l'information de l'hydrogramme calculé du modèle à cisterne fonctionnant. Ces deux critères sont le
volume du débit et la forme de l'hydrogramme. Les réactions de ces deux critères correspondent à la
réaction de déplacement et la réaction de vélocité au réglage automatique. Le résultat du modèle à
cisterne fonctionnant se compose de composants, les résultats de chacun des cisternes. Egalement toute la
durée est divisée en sous-périodes, pendant lesquelles chacun de ses composants joue le rôle principal. On
calcule les critères de volume et de forme en chaque sous-période et les utilise pour l'adjustement des
cisternes respectives. Le processus de réaction commence de quelque modèle initial et converge très vite
après plusieurs (en général moins de 15) itérations, et le résultat obtenu est très bon.
In spite of its simplicity the tank model (Fig. 1) can give good results for rivers in
many regions, and because of its simple structure there are no difficulties with using
computer techniques. The program is easily understood by computer users. The
only difficult problem is the calibration of the model, partly because of its nonlinear
structure and partly because its structure is very difficult for input/output analysis.
The only way is by trial and error. This is not so difficult for those who are
experienced with the model, but it is difficult to teach this method of calibration to
those inexperienced in using the tank model. Therefore to have an automatic
calibration program is the most important point of using a computer for this model.
nr
nr
L
nr
Fundamental principle
The fundamental principle is rather simple. The tank model is composed of two
types of tank (Fig. 2a and b) which can be approximated by a linear model (Fig. 2c)
by moving the side outlets or outlet to the bottom. This linear tank model is a first
nr (b)
(a)
(c)
FIG. 2. (a) and (b), the two types of tank; (c) is the linear model of (a) and (b).
order lag system which can be written f}/[A + (a + /?)], where A is the differential
operator, l/(a + /i) is the time constant, and the ratio of discharge to input is
PI (a + /?). The fundamental principle of parameter modification is as follows.
( 1 ) To change the shape of the hydro graph we must modify a + f>, e. g. to make the
hydrograph steeper we must increase a +/? and vice versa.
(2) To change the total volume of the hydrograph we must modify /?/(a + /?), e.g.
to make the total volume of discharge larger without changing the shape of the
hydrograph we must increase ft and decrease a keeping a + /? unchanged, and vice
versa.
Automatic calibration of the tank model 377
0.2
n£
L 0.05
niô-°5
r= 0.01
ni0.01
0.001
FIG. 3. The initial model.
y2
nrL
y3
nr r
hr L
y5
FIG. 4. Outputs from the four tanks
calculated discharge and not from the observed discharge. For example, in the
definition of the criterion RD(I), the summation is carried out over the falling part of
the calculated hydrograph. The observed hydrograph is disregarded to avoid the
random noise that will appear in it (Fig. 5).
Feedback formulae
According to the fundamental principle given above, adjustment of the parameters
shown in Fig. 6 can be carried out using the criteria RQ(I) and RD(I).
When RQ(I) > \(RQ(I) < 1), we must decrease (increase) the parameter controll-
Automatic calibration of the tank model
A?
nr
FIG. 6. Inlet and outlet parameters.
ing the side outlet, and increase (decrease) the parameter for the bottom outlet. This
adjustment is carried out by dividing the side outlet parameter by ^RQ(I) and
multiplying the bottom parameter by ^JRQÎJ).
When RD(I)>1 (RD(I)<\), we must decrease (increase) both parameters
equally. The adjustment is carried out by dividing both parameters by RD(I).
Feedback by RQ(I) and RD(I) corresponds to the displacement feedback and
velocity feedback in automatic control, respectively. Therefore the feedback of
RQ(l) and RD(l) may be given as follows, where A MO, AMI and AMI are the
modified parameters:
(AMI +AM2) = (A\+A2)j[sjRQ(\) RD(\)}
AM0=A0 ^/RQ(Ï)/RD(\)
A0 = A0 i^RQ{l)/RD(ï) + ^RQi2)/RD(2)] \
AMI =Al/[^/RQ(2) RD(2)]
A2 = (Al +A2)l[jRQ(X) RD(\)]-AM\
Al=AM\
where the equations are written in FORTRAN, i.e. A0, A1 and A2 on the left of the
equations are the modified parameters.
Although adjustments of the parameters of the top tank will have some effects on
the lower tanks, we neglect these effects and adjust the parameters of the second
tank using the criteria RQ(3) and RD(3):
B0 = B0^/RQ(3J/RD(3)
Bl=Bl/[jRQ(3) RD(3)]
In the same way, we can adjust the parameters of the third tank by RQ(4) and
RD{4):
C0 = C0 ^rRQ(4)/RD(4)
Cl=Cl/[y/RQW RD{4)]
The tank model shown in Fig. 6 has no outlet on the bottom of the fourth tank
because at this stage, we consider only the case in which there is no underground
discharge. In this case, the feedback of RD{5) is given by
D\=D\jRD(5)
The feedback of RQ(5), however, cannot be calculated as above. We must control
the water supply from the upper tanks. If RQ(5) > 1 (RQ(5) < 1), we must decrease
(increase) the parameters of the bottom outlets of the upper tanks. The control of
the water supply to the fourth tank is carried out by first adjusting CO of the third
tank, then the change in the third tank caused by the adjustment of CO is compen-
sated by the adjustment of BO, etc. Under such consideration, we obtain the
following formulae:
C0 = C0/RQ(5)
B0 = B0I^/RQ{5)
A0 = A0/ l/RQ(5)
In some cases, some of the RQ(I)s and RD(I)s may show values exceedingly
different from 1. To prevent the feedback of such extreme values, we limit them to
the range Q-, 2), i.e. values of RQ(I) and RD(I) larger than 2 are replaced by 2, and
values smaller than \ are replaced by \.
We expect that by using the iterative feedback described above the tank model
will converge to a good fit. Modifications may be necessary, however, for the
convergence of the feedback system.
RD(r) = sj/RD{f)
At first, RD(I) was replaced by ^jfRD{I) but the modification has now been revised
to the above formula.
(a) (b)
FIG. 7. Lag times between observed and calculated bydrographs.
have large errors, and in the case shown in Fig. 7b RD(l) or RD(2) will have large
errors. To avoid this type of error we modify the definition of the subperiods by
neglecting the day and the next day when the day belongs to subperiod 3,4 or 5 and
the next day belongs to subperiod 1 or 2.
RQ(I) and RD(I) which are near to 1 should not be used for feedback
The main reason why the above feedback system cannot work well must come from
the simultaneous feedback of all RQ(I) and RD(I). Then, the interference between
the criteria and the feedback of RQ(I) and RD(I) which are near to 1 will lead the
feedback in the wrong direction. The criteria near to 1 contain less information but
much noise, therefore such criteria should be neglected in the feedback system. For
this purpose, we set two thresholds, CRQ and CRD, and we feedback only those
criteria RQ(I) and RD(I) which satisfy the following conditions,
\RQ(I)~l\^CRQ
\RD(I)-l\>CRD
Setting CRQ = i and CRD =4, the automatic calibration trial converged for the
River Waga, the test basin for this method. Then, it converged under the condition
CRQ = 10 per cent, CRD = 20 per cent, and also under CRQ = 7 per cent, CRD = 15
per cent but not under the condition CRQ = 5 per cent, CRD= 10 per cent. Here,
converge means that all RQ(I) and RD(I) are restricted within the given threshold
CRQ and CRD. Though it did not converge under the condition CRQ = 5 per cent,
CRD = 10 per cent, the tank model obtained, under the condition CRQ = 7 per cent,
382 M. Sugawara
CRD = 15 per cent, gives an excellent hydrograph, so we believe that this type of
automatic calibration is successful.
Ité(jv)-e*(jv)]2/£i
A' ff
This formula can give good error evaluation to the calculated hydrograph in most
cases but it is bad in the particular case shown in Fig. 8 which occurs very seldom.
Automatic calibration of the tank model 383
FIG. 8. Particular case where the modified mean square error method fails.
Divergence
The iteration of feedback goes on as follows. Within the first five to ten trials, all the
criteria RQ(I) and RD(I) approach the goal quickly. Then the convergence becomes
slow, and after they have reached some value they cannot improve. Then in some
cases divergence or oscillation begins. When divergence begins, the balance between
parameters is broken and the feedback leads to a worse state and cannot return to a
good condition.
We can understand the above phenomena as follows. As the data contain random
noise, RQ(I) and RD(f) will also contain noise. When they are different from 1,
RQ(I) and RD(I) contain a lot of information. But when they approach close to 1,
the proportion of noise becomes larger and feedback by RQ(I) and RD(I) cannot
give a good adjustment. Once the balance between parameters is broken, feedback
accelerates the disintegration.
Flowchart
The details of the feedback system are shown in Fig. 9.
: : : : : : j : : : : : : :
Adjustment for titne lag on observed discharge:
lj(N) = 1 1 - a M N ) * aç(N*1)
Calculation of criteria which are used for adjusting the working tank model.
HQ(I) = i QEtN) / i cCJ) (I = 1.2,J,<*,5)
i'flog iE(h-l) - log jS(N)>
HD(I)
- pTTzrm^r-^rmri n =-^.3»
where 1 means the sum for days belonging to subperiod I, and 1' means the
sum for days N where both days h, ft-1 belong to subperiod I and
log ^E(n-1) - log QE(U) is positive.
RJ(I) - 1 + UP(l)-l)/2
i
Calculation of criteria for evaluation :
M Sag (mean square error of discharge) /(mean discharge)
(mean square error of log ^ ) *
CUE (MSE^ + MSEL^)/2
CRi^D criterion for RQ(l) and RD(l)**
< U ( R y U ) - 1 Î 2 + i(RD(l) - 1 ) 2 ) / {^ + z.l)) 1/<i
CH CR& + Ci^D/4
* the effe t of random fluctuation of time lag is under consideration.
neglecte i H€>( I ) ' 5 and ED (I ) 's are not counted in summation.
FIG. 9. Flow chart of the automatic calibration program of the tank model.
AO = AO / 80(3)
BO = BO
/ BB(3)
B1 = B1
/ HD(3)
AO = AO / VK3(5)
BO = BO / \/HQ(5)
CO = CO / B Q ( 5 )
Stop.
Something is wrong,
Print out the results.
F I G . 9. continued.
Determination of time la
Determination of subperiods:
Same as the previous metnod shown in Fig, 9.
Calculation of criteria which are used for adjusting the working tank model:
Same as the previous method shown m Fig. 9-
I
Preparation for brancning :
AA = <H^(3)-1) - ( H u C O - 1 )
Ad = (fii.CO-1) • Utf(5)-1)
]
I
RD(I) = 1 , (3D(l)-1)/2
I
Calculation of criteria for evaluation, hii,^, HSbLi^, QRb, CK^D and GR :
Same as the previous system shown in Fig. 9.
F I G . 10. Flow chart of the improved automatic calibration program of the tank model.
Division of the duration curve into live sections corresponding to five subperiods of the
working tank model
The most fundamental part of the automatic calibration system is the division into
subperiods by means of the working tank model. On the duration curve of calcu-
Automatic calibration of the tank model 387
A = (A2/A0)/BQ(1)
B = (A1/A0)/Rtj(2)
AO = (A0+A1 + A 2 ) / C ' f ? D ( 2 ) , ( 1 + A + B) )
A1 = B * AG
A2 = A - AO
B = (B1/B0)/RQ(3) AO = A O / H c ( 3 )
BO = ( B 0 * B 1 ) / ( B D ( 3 ) - ( 1 t B ) ) BO = BO/HD(3)
B1 ^ B * BO B1 = B 1 / K D ( 3 )
AO = AO/VWJT
BO = so/ </SMyi
GO = CO / X 0 ( 5 )
Stop.
ng i s w r o n g .
FIG. 10 continued.
lated discharge derived from the working tank model, discharges for the days
belonging to subperiod 1 are usually large and they will gather in the left part, and
conversely discharges for the days belonging to subperiod 5 are usually small and
they will gather in the right part.
After some consideration and hesitation, the duration curve is divided boldly by
the most simple way, dividing it into five subsections successively from the left,
NQ(l) days, NQ(2) days, . . . and NQ{5) days, where NQ(I) is the number of days
belonging to subperiod /.
In the previous system there is an additional condition in defining the subperiods
to avoid the effect of random deviation of the time lag of peak discharge. In this
case, however, this additional condition is not necessary, because in comparing the
duration curves the effect of time lag has no meaning.
Duration curves should be made for every year to prevent mixing of dry years and
wet years.
388 M. Sugawara
Flow chart
After the division of the duration curve into five subsections has been finished, the
definition of RQ(f) and RD(1), the feedback procedure by means of RQ{I) and
RD(I) etc. are easy. The details of the procedure are shown in Fig. 11.
This method has given good results for Japanese rivers and also for many basins
in the Upper Nile region.
Determination of s u b p e n o d s :
Same as the previous method shown in Fig.
except tnat the condition 6) is omitted.
Calculation of criteria which are used for adjustment of the working tank model:
HO(I) = fa L ^ < N 0 ) / fa L Q(riO) (1 = 1,2, -••,5)
I ( 1 QE(NO) - I QE(NO))
HD(1) = -HI it-J. 2 r< _ -> , i.)
Nr «- ÏR "
Where QE(hO) and q(NO) are the calculated and observed discharges and NO is the
number of order, l is the summation over ]'th section, 1 / 1 is summation
the left/right half of the I *th section and I is the ation over years.
In the calculation of HD(2), subsection 1 is joined with subsection 2 to make a
new subsection 2.
|
KD(I) = 1 + (RD(l)-D/2 (1 = 2 , ? , M
FIG. 11. Flow chart of the automatic calibration program of the tank model by means of the criteria
defined by comparison of the duration curves.