You are on page 1of 18

Journal of Hydrology 226 (1999) 3047 www.elsevier.

com/locate/jhydrol

Multilinear discrete lag-cascade model for channel routing


L.A. Camacho, M.J. Lees*
Department of Civil and Environmental Engineering, Imperial College of Science, Technology and Medicine, London SW7 2BU, UK Received 22 February 1999; accepted 17 September 1999

Abstract An extension to the two-parameter multilinear discrete cascade model for channel routing is presented. The linear subcomponent of the extended three-parameter model combines a discrete cascade with a conceptual discrete linear channel element characterised by a time-delay parameter. The inclusion of this explicit advective time-delay makes the model particularly suitable for forecasting applications. The time-variable parameters are related to physical channel characteristics by a method of moments utilising the generalised linear response of a uniform channel of any shape and any friction law. The utility of the extended model is examined through comparison with results of the multilinear model and the solutions of the full St Venant equations for various hydraulic conditions and channel shapes. The results show that without sacricing model simplicity, the additional degree of freedom of the extended model enables better prediction of the ood propagation process described by the full equations. Additionally, it is found that the linear approximation of the proposed mathematical representation gives sufciently accurate predictions over short distances to allow practical utilisation of powerful methods of system parameter estimation. Good results were obtained in uniform channels characterised by open looped-rating curves at the inlet and outlet boundaries. However, the downstream boundary condition considered, based on the Manning equation, is only valid where no hydraulic structures or physical channel changes occur at the downstream location, or where backwater effects due to tides, tributaries or critical control sections do not affect the hydraulic response at the boundary. 1999 Elsevier Science B.V. All rights reserved.
Keywords: Hydrology; Rivers and streams; Unsteady ow; Floods; Flood routing; Flood forecasting

1. Introduction Availability of sufcient and appropriate data describing the hydrogeometric channel characteristics of a long river stretch, or of an entire channel network, is very rare, and in many circumstances accurate eld investigations are prohibited by time schedules and cost. Further, it is unlikely that this situation will change substantially even over the long-term, and especially in developing countries. The application
* Corresponding author. Fax: 44-171-594-6124. E-mail addresses: l.camacho@ic.ac.uk (L.A. Camacho); m.lees@ic.ac.uk (M.J. Lees)

of complete distributed models in data-poor situations is neither justiable nor advantageous. A sensible course of action, therefore, that is feasible under certain hydraulic conditions, is the application of approximate methods in which a lumped system characterisation is sought and extensive use of limited available data is maximised. Weinmann and Laurenson (1979) present a comprehensive review of approximate ood routing methods and demonstrate that all the numerous models available, including storage methods, are essentially variants of either kinematic wave or diffusion analogy models, neither of which account for the inertial terms in the equation of motion. They also

0022-1694/99/$ - see front matter 1999 Elsevier Science B.V. All rights reserved. PII: S0022-169 4(99)00162-6

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

31

highlight the importance of parameter estimation through which, for instance, the simplest models based on linear kinematic assumptions can yield equivalent results to more complex diffusion analogy models. Recent work by Singh et al. (1997) clearly demonstrates this point, showing that the mixing cell method (Wang and Chen, 1996) can be used to convert the diffusion wave equation to a difference form where the diffusion term can be eliminated through selection of an optimal space step size. Singh et al. (1997) also reveal that the optimal step size that results when small step time intervals are considered is equivalent to the characteristic length of the KalininMilyukov (Kalinin, 1971) ood routing method widely used in Eastern Europe. The application of the KalininMilyukov model to a channel reach (Miller and Cunge, 1975) consists of routing the ood wave through a cascade of n-linear conceptual storage elements each with a characteristic length Dx. The resulting model is in practical terms identical to the Nash cascade (Nash, 1960) where an impulse response function, the gamma distribution, is used to represent the unit hydrograph. The method is appealing in that, as for the MuskingumCunge (Cunge, 1969) method, both model parameters (the number of reservoirs in series and the storage coefcient) are related to channel characteristics and hydraulic conditions. Further, it does not produce physically unrealistic reduced outow at the beginning of routed hydrographs, which is a characteristic of the MuskingumCunge method when applied with insufcient spatial resolution (Weinmann and Laurenson, 1979; Ponce and Theurer, 1982; Ponce, 1991; Perumal, 1992a, 1993). It is an accepted fact that the propagation of a ood wave is a non-linear process and therefore application of the linear models previously described may result in crude approximations of actual waves, particularly under conditions dominated by high resistance effects. One approach to incorporating non-linearities into approximate methods has been the development of conceptual non-linear reservoir cascade models (Laurenson, 1964; Whitehead et al., 1979; OConnel et al., 1986; Malone and Cordery, 1989). However, the parameter accounting for the number of reservoirs in the cascade has never been successfully related to channel properties and grid specication, and therefore their predictive ability is limited. Bentura and

Michel (1997) investigated the application of a twoparameter quadratic lag-and-route method. They show that this non-linear model has the ability to reproduce the behaviour of the complete St Venant equations for wave trains in wide rectangular channels. It is apparent from their study, however, that it is difcult to relate the model parameters to channel characteristics and hydraulic conditions. Empirical relationships, which may only be valid within the range of their investigation, were used in their study to estimate the model parameters. In an alternative approach to accounting for nonlinearities, the ood propagation process can be assumed to respond linearly to the input at any point in time, but with the model parameters recalculated in time and space as a function of local ow values. This concept led to the development of the multiple linearisation ow routing model (Keefer and McQuivey, 1974); the non-linear threshold model based on amplitude distribution schemes (Becker and Kundzewicz, 1987); and the multilinear discrete cascade model for channel routing (Perumal, 1994). A similar concept, where the parameters are recalculated in time and space, is also the basis of the variable-parameter MuskingumCunge method (Ponce and Yevjevich, 1978; Perumal, 1992b; Ponce and Chaganti, 1994). In the multilinear discrete cascade model proposed by Perumal (1994), the linear component is based on a discrete representation of the Nash cascade (OConnor, 1976) and the parameters are directly related to the parameters of the KalininMilyukov model. This two-parameter method gives reasonable results when compared with the response of the non-linear St Venant equations for hydraulic conditions characterised by narrow loops in the stagedischarge relationships at the channel inlet. However, Perumals study was limited to uniform prismatic channels with input hydrographs characterised by a single wave, and as with all other approximate methods, it proves to be inaccurate under conditions characterised by open loops in the stagedischarge relationships. Dooge and Harley (1967) and Dooge (1973) present a linearisation of the complete St Venant equations about a reference discharge per unit width applicable to wide rectangular channels with no lateral inow. In a pioneering application of the multilinear approach, Keefer and McQuivey (1974) used this linearised model as their linear sub-model.

32

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

Fig. 1. Multilinear discrete lag-cascade model structure.

However, they concluded that Dooge and Harleys (1967) model is difcult to apply to real data and that the mathematical elegance is lost when computing discrete response functions. Dooge (1973, 1986) compared the channel response of the complete linearised model with the channel response of different linear conceptual models, including the Kalinin Milyukov model, using shape factor diagrams. It was concluded that the solutions for the combination of a pure lag with a cascade of equal linear reservoirs gives the closest correspondence to solutions of the complete linearised equations. It appears reasonable therefore to use this simple three-parameter model as the linear component of the multilinear approach rather than the two-parameter model on which Perumals multilinear discrete cascade is based. This paper presents, and investigates the performance and parameter estimation of a new multilinear model based on the use of a three parameter discrete lag-cascade submodel, which is an extension to Perumals multilinear model. 2. The proposed multilinear discrete lag-cascade method The multilinear modelling approach proposed here is identical to Perumals method (Perumal, 1994) in terms of the time distribution scheme used, but the linear sub-model is dened by a three-parameter model combining a discrete cascade with a discrete linear channel (see Fig. 1). The effect of the conceptual linear channel component is simply to lag the routed hydrograph by an explicit time interval specied by the time delay model parameter. The routing procedure is as follows: 1. Partition the inow hydrograph into P pulses of constant duration equal to the routing time interval Dt. 2. Calculate the linear sub-model parameters nd ; Kd and td for the rst inow pulse I t; t Dt: The parameters depend upon the intensity of the inow

pulse as explained later in Section 3. nd represents the number of reservoirs in the cascade which may not be an integer value; Kd represents the linear storage coefcient for all reservoirs; and td is the time delay parameter of the linear channel. 3. Compute the unit pulse response function of the discrete cascade model associated with pulse I t (see, OConnor, 1976; Perumal, 1994),  nd Dt 1 h1 Dt Kd hm   m nd 2 Kd h Dt Kd m1 m 1

for m 1 where h1 is the discrete unit pulse response ordinate associated with time t Dt; hm is the discrete unit pulse response ordinate at a subsequent time t mDt; m 2; 3; ; M ; Dt is the ood routing time interval; and M is a sufciently large number where hm becomes negligible. 4. Compute the input pulse response associated with the pulse I t; the ordinates of which at a later time t mDt will be given by qt I t hm ; m 1; 2; ; M 5. Apply the discrete linear channel procedure to delay the input pulse response by the time delay td : The linear channel operation applied to the time signal q is simply given by, Qt qt td 3

In discrete time the advective time delay is given as a multiple of sampling intervals or time steps Dt, i.e. Qt qt d where d is dened as the nearest lower integer value of td =Dt: 6. Repeat steps 25 for pulse I t Dt: The complete routed hydrograph is obtained by integrating the constituent pulse responses associated with the P inow hydrograph pulses. Fig. 2 summarises the computational procedure. It is reasonable to expect that the inclusion of the additional parameter, the time delay of the linear channel, will allow a better representation of the non-linear St Venant solutions. Clearly the parsimony of the two-parameter discrete cascade

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

33

Fig. 2. Multilinear discrete lag-cascade model computational procedure.

model is reduced, and therefore inclusion of the additional parameter can only be justied if benets can be obtained in terms of accuracy without sacricing the simplicity of the solution. However, inclusion could also be justied if the additional parameter has clear physical signicance. Since the time delay parameter accounts for a major component of the time delay observed in the propagation of the leading edge of a ood wave downstream, it effectively replaces the large number of cascade reservoirs required to accurately simulate time delay in long channels or in channels characterised by kinematic wave hydraulic conditions, as demonstrated by Nash (1960). The inclusion of this parameter can also be justied by its practical utility in forecasting applications, since it provides explicit information about the time of translation of the ood wave to a downstream location. The method obviously has a linear mode in which the model parameters are calculated for a reference discharge and then kept constant through time. In continuous-time form, and with a single reservoir, this linear mode of the model is analogous to the traditional lag-and-route model (Meyer, 1941; Bentura and Michel, 1997); and, if the lag is set to zero and a series of reservoirs are considered, it is

analogous to the KalininMilyukov model (Kalinin, 1971). As shown later this linear mode of operation proves to be adequate for practical purposes under particular hydraulic conditions. This result is of importance since in this linear mode form the multilinear lag-cascade model can be represented by a general transfer function model, therefore enabling powerful time-series analysis techniques to be used for model identication, parameter estimation, and uncertainty analysis (e.g. Young, 1986). It should be noted that one potential limitation to this method is that formation of irregularities in the peak regions of the simulated hydrographs may occur as a consequence of the selection of inappropriately large time steps, i.e. inappropriate division of the inow hydrograph into pulses (Keefer and McQuivey, 1974; Kundzewicz, 1984). Therefore, in order to avoid this problem, the method could be considered computationally demanding compared to other hydrological ood routing methods since it requires small simulation routing time steps. However, recent advances in desktop computing power is enabling methods that have been previously discarded due to computational intractability, despite their conceptual elegance, to be successfully implemented. We will return to this issue later in Section 5.3.

34

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

3. Model parameter estimation Estimation of the linear sub-model parameters is carried out using a moment or cumulant-matching procedure. This method relates the model parameters to channel properties and hydraulic conditions thereby producing a predictive model of the ood propagation process. 3.1. Cumulants of the multilinear lag-cascade model and the generalised linear channel response The cumulants of the continuous form of the multilinear lag-cascade model are identical to those of the Nash cascade (Nash, 1960), except for the rst cumulant that is increased by the lag term. The rst three cumulants are given by: k1 nK t k2 nK 2 k3 2nK 3 1st-moment about the origin 2nd-moment about centroid 3rd-moment about centroid 4a 4b 4c

k3

3 2 2 1 m 12 F0 1 m 12 F0 m2     y0 2 X 3 S0 X mu0

5c

where F0 is the Froude number; y0 is the ow depth at the steady reference discharge Q0; S0 is the channel bed slope; X is the longitudinal distance at which the hydrograph is computed; and m denotes the ratio of the kinematic wave speed c0 to the average velocity of ow at the reference condition u0,    dQ  c d A A A0   6 m 0 Q0 u0 A0 Note that for wide rectangular channels with a Chezy friction law m 1:5; and with a Manning friction law m 5= 3: 3.2. Cumulant matching It can be readily demonstrated that matching the rst three cumulants of the linear sub-model gives:    3 y0 X 2 1 m 1F0 7a K S0 X 2m mu0 4m 2 1 m 1 2 F 0 9   n y0 2 2 1 m 1F0 S0 X 2 X t 1 mu0
2 3 2 1 m 12 F0 2 1 m 1 F 0

Following Dooge (1973), the model parameters can be related to the parameters of the linearised St Venant model, and therefore to ow conditions and channel characteristics through a cumulant matching technique. The linearisation presented by Dooge and Harley (1967) and Dooge (1973) is based on the differential equations of continuity and momentum governing ow in a wide rectangular channel with Chezy friction and no lateral inow. The linearisation was extended to any shape of channel and any friction law by Dooge et al. (1987a,b) and is referred to as the linear downstream response of a generalised uniform channel. Applying the Laplace transform technique the rst three cumulants of the generalised downstream channel response are (Dooge et al., 1987b): k1 X mu0 1 y0 2 1 m 1 2 F 0 S0 X m   X mu0 2 5a

7b 3 7c

Following Perumal (1994) and OConnor (1976), the discrete sub-model parameters can be related to the continuous-time parameters (Eqs. (7a)(7c)) for time steps Dt where Dt K by, Kd K Dt nd nK Dt K Dt 8a 8b 8c

td t D t
5b

k2

As pointed out by Perumal (1994), when Dt K ; a situation that may arise when simulating kinematic

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

35

20 18 16 Time delay and Travel time [h] 14 12 10 8 6 4 2 0 0 Travel time Time delay

100

200

300 Discharge [m3s 1 ]

400

500

600

Fig. 3. Parameter variation with discharge for a kinematic channel type.

ood waves, Eqs. (7a)(7c) must be used rather than Eqs. (8a)(8c). Also, in order to apply the discrete linear channel procedure (see step 5 in Section 2), the advective time delay d in sampling intervals or time steps Dt; dened as the nearest lower integer value of td =Dt; must be computed to represent Eq. (3). The remainder after division is usually small for small time steps and can be ignored as a rst approximation introducing a negligible error in the computations (see later discussion in Section 5). For larger time steps however, the storage coefcient should be adjusted accordingly to maintain the correct travel time Kd Kd td mod Dt=nd 9

accounted for by the multilinear technique, the approximation will be valid over a wide range. In a manner that is consistent with the methods assumptions and multilinear concepts, the reference discharge upon which the linearisation is based is also recalculated over time. It is assumed that the reference discharge varies according to the intensity of the inow in a form given by Perumal (1994): Q0 Ib aI t Ib 10

where td mod Dt is the signed remainder on dividing td by Dt. Although the accuracy of the model parameterised by Eqs. (8a)(8c) is at most as good as the accuracy of the linearised St Venant model, it is valid within a small range about the reference discharge Q0 and area of ow A0. Further, since non-linearity is

where Ib is the initial steady ow in the reach before the arrival of the ood wave; Ib I t 0; and a is a small coefcient with empirical limits 0 a 0:5: Since the reference discharge is recalculated over time, the linearisation will result in close approximations of the non-linear equations. It is worth noting that the model parameters are computed for the reference discharge Q0, and not for the inow discharge I t at each time (see Fig. 2). Also it may be noted that Eq. (10) describes a unique relationship between the model parameters and the inow discharge, notwithstanding whether the discharge to be routed is on the rising or falling limb of the hydrograph.

36

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

Table 1 Channel congurations used in the study Channel identication Rectangular 1 2 3 4 Width (m) Bed slope Mannings n

25 25 25 25 Width (m)

0.0002 0.0002 0.002 0.0001 Lateral bank slope 45 Top width (m)

0.04 0.02 0.02 0.04 Bed slope Mannings n

Trapezoidal 5

25 Depth (m)

0.0001 Bed slope

0.04 Mannings n

Irregular cross-section 6

0 1 2 5 10 15

10 25 35 50 100 150

0.0001

0.04

Close examination of the relationships (8a)(8c) reveals the convenience of the routing parameters. They are a function of channel characteristics and hydraulic conditions and grid specication. The reservoir storage constant depends upon the intensity of ow but remains constant with the length of the channel considered, as in the two-parameter Kalinin Milyukov model. However, both the lag of the linear channel and the number of reaches in the cascade vary directly with the length, which leads to results that are independent of grid size. The rst moment nd Kd td that represents the wave translation time (travel time) shows a characteristic exponential decay with discharge. The explicit time delay of the leading edge of the hydrograph is also conveniently
Table 2 Channel classications Channel identication 1 2 3 4 5 6 TS0 u0 =y0 4.81 12.07 524.11 1.50 4.36 1.62 p TS0 g=y0 40.78 51.32 740.83 18.11 19.67 20.11 Classication Diffusion analogy Diffusion analogy Kinematic wave Full dynamic Full dynamic Full dynamic

represented, again showing a characteristic non-linear decay with discharge. For example, Fig. 3 shows the relationship of the travel time and time delay with ow for one of the channel types used in later testing (see Tables 1 and 2). Note that the highly non-linear characteristics of low ows are successfully reproduced by the relationships. These observations coincide with observed characteristics of ow propagation and therefore the relationships are considered to be appropriate as a rst approximation. Also note that no matter what reach length is specied, unrealistic reduced ow characteristics are not produced.

4. Application To aid comparison, the test example used by Perumal (1994) is used to evaluate the multilinear lagcascade model and a number of performance criteria are used to compare the solutions of the proposed method with solutions obtained from the full St Venant equations and the original multilinear model. In Perumals test example a synthetic inow hydrograph is routed through uniform rectangular channels of different hydrogeometric congurations as dened

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

37

Inlet rating curves 15 Flow stage [m] Ch.4 Ch. 1 Ch. 2 Ch. 3 0 100 150 200 250 300 350 400 450 500

10

Outlet rating curves 10 Flow stage [m] 8 6 Ch. 2 4 2 0 100 150 200 250 300 350 3 1 Discharge [m s ] Ch. 3 Ch. 4 Ch. 1

400

450

500

Fig. 4. Rating curves at the inlet and outlet boundaries.

in Table 1. Here, additional congurations (No. 4, No. 5 and No. 6) are used to dene channel reaches where diffusion analogy models are clearly not applicable. These channels represent a case where the inertial terms are important, and therefore must be retained in any model attempting to solve the routing problem. Classication criteria (Ponce et al., 1978) for the channels considered are presented in Table 2 where T is wave period or wave duration taken as 40 h considering the hydrograph dened by Eq. (11), and y0, u0 are steady uniform ow depth and mean ow velocity computed for Q0 100 m3 s1 ; respectively. Note that kinematic type models are suitable for channel type 3 whereas for channel types 1 and 2 diffusion analogy type models are required as a minimum level of complexity. The upstream and downstream boundary rating curves for the rst four channels, obtained from the full St Venant equation solutions, are shown in Fig. 4. It can be seen that channel type 3 is characterised by a single relationship between discharge and water stage that is not distinguishable from the steady ow rating curve whereas the rest of the

channels are characterised by open looped-rating curves at the inlet and outlet boundaries. Congurations 5 and 6 are considered in the present paper to illustrate results in uniform channels with nonrectangular cross-sections. The inow hydrograph is expressed as in the original example t I t Ib Ip Ib tp 2 3
1 g 1

exp

1 t = tp g 1

! 11

where Ib is the initial steady ow (100 m 3 s 1) in the reach; Ip is the peak ow (500 m 3 s 1); tp is the time to peak (10 h) and g is the skewness factor (1.15). The hydrograph is routed to distances of 20, 40, 60, 80, 100 and 120 km from the inow section. A single reach is used to compute the downstream hydrograph at the desired location, and the value of the coefcient a in Eq. (10) is maintained at a constant value of 0.2. The full St Venant solutions are obtained using a numerical four-point Preissmann implicit method, with the system of non-linear algebraic equations

38

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

solved by the NewtonRaphson algorithm (Fread, 1985). A weighting coefcient u 0:6 is used for the nite difference scheme. A time step of Dt 15 min and channel sub-reaches of 4 km and are used in the St Venant computations for all channel types except for conguration No. 3 where Dt 10 min is used to full the Courant condition and to minimise errors introduced by the implicit scheme. The upstream boundary condition is dened by the known inow hydrograph and the downstream boundary condition, xed at 120 km downstream, is dened by a looped-rating curve based on the Manning equation for normal ow. The loop is obtained by approximating the friction slope from the momentum equation (Fread, 1993). This boundary condition allows the unsteady wave to pass the downstream boundary with minimal disturbance from the boundary itself, and it is commonly used where no hydraulic structures or physical channel changes occur at the downstream location. Note that it is not valid for locations where backwater effects, such as those caused by downstream tributary inputs or tides, are important or for critical ow sections around hydraulic structures. The peak discharge Qp, time to peak tp, and risetime of the routed hydrograph td for the St Venant solutions are compared with the values obtained from the multilinear lag-cascade model. The model t is given by the coefcient of determination R2 T or NashSutcliffe criterion (Nash and Sutcliffe, 1970), which can be interpreted as the percentage of the variance of the St Venant solution explained by the multilinear lag-cascade model. The statistics are computed from results produced by both models at exactly the same time steps, i 1n; and therefore no interpolation is required.

The linear mode of the multilinear lag-cascade model was tested in each validation experiment by xing the parameters to those calculated for the initial steady discharge Ib. The efciency of the multilinear mode can then be judged by the difference in the coefcient of determination compared to that obtained with the fully nonlinear model. It can be interpreted as the additional variance explained by the multilinear approach, and is given by (Nash and Sutcliffe, 1970):
2 rT 2 R2 T2 RT1 1 R2 T1

12

where the sufxes 1, 2 denote the linear and multilinear mode results, respectively. Additionally the performance of the new model is tested by visually comparing the peak discharge attenuation along the channel (as a percentage of the input hydrograph peak) against proles obtained from the St Venant model. It is worth noting that another convenient evaluation approach is to use a shape factor diagram, as presented by Dooge (1986) and Dooge et al. (1987a). However, this is not possible here since no explicit analytical relationships can be derived to form shape factors for the normalised (multilinear) channel responses. 5. Discussion of results 5.1. Multilinear approach Typical solutions at 20 and 120 km are summarised in Table 3 for channels of type 1, 2 and 3. The

Table 3 Multilinear lag-cascade results at 20 and 120 km for channel types 1, 2 and 3 Channel identication Distance (km) St Venant Qp (m 3 s 1) 1 2 3 20 120 20 120 20 120 430 256 465 347 500 499 tp (h) 12.25 27.50 11.5 21.25 10.83 14.67 td (h) 1.5 8.75 1.25 7.5 1.67 7.17 Multilinear lag-cascade Qp (m 3 s 1) 431 252 468 348 499 498 tp (h) 12.5 29.75 12.0 23.25 10.83 15.33 td (h) 1.5 11.25 1.25 8.25 1.5 6.5 0.9992 0.9916 0.9991 0.9786 0.9983 0.9657 R2 T

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

39

Channel type 1 500 Discharge [m3s1] 400 300 200 100 0 120 km 20 km

10

20

30

40 Channel type 2

50

60

70

80

500 Discharge [m3s1] 400 300 120 km 200 100 0 20 km

10

20

30

40 Time [hr]

50

60

70

80

Fig. 5. Routing results for channel types 1 and 2: the St Venant solution (continuous line); the multilinear lag-cascade model (dotted line); and the multilinear cascade model (dashed line).

60 St. Venant Mult. lag-cascade Mult. cascade (Ch.1)

Percentage of attenuation of inflow peak discharge

50

40 Ch. 1 30

20

Ch. 2

10 Ch. 3 0 0 20 40 60 Distance [km] 80 100 120

Fig. 6. Peak discharge attenuation proles for channel types 1, 2 and 3.

40

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

(a) Travel time 35 Travel time [hr] 30 25 20 15 10 0 20 40 60 (b) Variance 140 120 Variance [hr2] 100 80 60 40 20 0 0 20 40 60 Distance [km] Ch. 2 Ch. 3 80 100 120 Ch. 1 80 100 120 Ch. 1 Ch. 2 Ch. 3

Fig. 7. Variation of the rst two moments of the routed wave with distance: the St Venant solution (dashed line); and the multilinear lag-cascade (circles).

computed multilinear lag-cascade hydrographs for channel types 1 and 2 are compared to the St Venant solutions and Perumals Perumal (1994) multilinear model results in Fig. 5. The results for these prismatic rectangular channels clearly demonstrate the new models ability to accurately describe the ood propagation process. The overall shape and timing of the hydrograph is well reproduced. Although accuracy decreases slightly with distance, at a downstream distance of 20 km the R2 T s are all over 0.99 and at 120 km they are still all over 0.96. Note in particular how for the more dynamic conditions of channel type 1, the multilinear lag-cascade model results are significantly better than the results obtained by the multilinear cascade model. Fig. 6 compares the peak attenuation prole with distance for the multilinear lag-cascade and St Venant models. It is clear that the multilinear lag-cascade model accurately reproduces the observed attenuation characteristics in all cases, and that for channel type 1 the multilinear lag-cascade model represents a significant improvement in accuracy when compared to the

original multilinear model. The combination of a reservoir cascade with a pure lag parameter gives the model the ability to predict the travel time and time delay of the ood wave accurately over distance as revealed by Fig. 7a, which shows the variation of the rst moment of the routed hydrograph (i.e. the travel time or time of translation of the centroid of the wave) with distance for the multilinear lagcascade model and the St Venant solutions. Note that the multilinear lag-cascade model uses a single reach from the input to the downstream location whereas 4 km sub-reaches are used to generate the full St Venant solutions. If more reaches are used in the multilinear lag-cascade modelling a negligible improvement is observed in the overall reproduction of the nal downstream hydrograph. This observation clearly suggests that the results are independent of length subdivision, and is supported by examination of Eqs. (7a)(7c) which show how the parameters vary with distance. It should be noted that 120 km long reaches would not normally be used in actual ood routing applications without subdivisions, and

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

41

500 Inflow St. Venant at 120 km Multilinear mode at 120 km Linear mode at 120 km

450

400

Discharge [m3s1]

350

300

250

200

150

100 0

10

20

30 Time [hr]

40

50

60

70

Fig. 8. Comparison of linear and multilinear mode results for the multilinear lag-cascade model.

therefore the calculated R2 T s can be interpreted as representing a worse case scenario. Fig. 7b demonstrates that the rst two moments of the routed hydrograph are well predicted over distance by the multilinear lag-cascade model over the wide range of hydraulic conditions described by channel types 1, 2 and 3. Another interesting result is that the time delay parameter effectively replaces the large number of cascade reservoirs required to accurately simulate time delay in a long channel. At 40 km the multilinear cascade model requires 2.52, 4.62, and 86.1 reservoirs for channel types 1, 2 and 3, respectively, in comparison with the new multilinear lagcascade that requires just 1.06, 1.72 and 18.67 reservoirs for the same channel types. In channel types 1, 2 and 3 the R2 T s obtained at 120 km with the linear version of the multilinear lag-cascade model are 0.8071, 0.7735 and 0.8195, respectively. The additional variances accounted for by the multilinear mode, computed using Eq. (12), are 0.95, 0.90 and 0.84, respectively. It can therefore be concluded that the multilinear mode is required to accurately predict the non-linear dynamics observed

over long distances in these channels. Fig. 8 illustrates how the non-linear model produces signicantly improved results in the case of channel type 2. In particular note the steepening of the rising limb and the corresponding attening of the receding limb of the multilinear lag-cascade hydrograph. Although the a coefcient in Eq. (10), which is used to compute the variable reference discharge, is xed in these multilinear model tests at a value of 0.2, it is observed that better results are obtained if different values are specied according to the channels hydraulic conditions. In particular for the kinematic ow channel (type 3) a higher value of a 0:4 gives improved results with R2 T s of 0.9998 and 0.9963 at 20 and 120 km, respectively. On the other hand, in channels characterised by highly attenuated proles (types 4, 5 and 6) a very low value of a produces more accurate results. This observation suggests that the linear form of the model, where the parameters are computed for a reference discharge and kept constant throughout time, will give reasonably accurate predictions in such cases. The reader may conclude that the a coefcient is effectively an additional parameter in

42

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

Table 4 Linear mode multilinear lag-cascade results at 20 and 40 km for channel types 4, 5 and 6 Channel identication Distance (km) St Venant Qp (m 3 s 1) 4 5 6 20 40 20 40 20 40 415 348 411 344 383 306 tp (h) 12.0 14.5 12.5 15.25 13.5 17.5 td (h) 1.25 2.0 1.25 2.5 2.0 3.25 Linear mode multilinear lag-cascade Qp (m 3 s 1) 411 343 406 336 394 317 tp (h) 13.25 16.25 13.25 16.75 13.5 17.5 td (h) 2.75 4.75 2.75 4.75 3.25 6.0 0.9831 0.9500 0.9929 0.9784 0.9975 0.9957 R2 T

the model. However, Eq. (10) is only an empirical way of indicating how the reference discharge can be recalculated in time and other forms can be considered (Keefer and McQuivey, 1974; Becker and Kundzewicz, 1987). The a coefcient is equivalent to the input data required in linear models where a reference discharge needs to be specied. It is worth noting, however, the convenience of Eq. (10) for ood forecasting applications since the reference discharge is related to the (known) inow at each time step. The chosen time steps of Dt 15 min for all channels except channel type 3 where Dt 10 min is used, produced solutions that did not produce irregularities in the peak regions of the simulated hydrograph pulses. The use of a ner time resolution of Dt 6 min in all multilinear lag-cascade simulations produced a very small improvement in the reproduction of the full model hydrographs. This result indicates that the truncation error introduced by using an integer advective time delay (as a multiple of Dt) can be ignored as a rst approximation in the solutions using a larger time step of 15 min. The requirement of a small routing time step (usually Dt 1 h) is also a condition required for the distributed model since

the accuracy of the implicit solution may be compromised for larger time steps. Since discharge measurements are commonly available at hourly sampling intervals, linear interpolation is required in both models to dene the input ow at the upstream boundary condition if smaller routing time steps are used. 5.2. Linear mode of the model and channel shape It is observed that the linear mode gives acceptable results for short reaches up to 20 km with R2 T s 0:98 in all channel types as can be seen from the results presented in Tables 4 and 5. Note that the performance of the multilinear lag-cascade model is better than the multilinear cascade model for all channel types. Linear mode results for channel type 4 are presented in Table 4 and Fig. 9a. They demonstrate that the linear form of the multilinear lag-cascade model is able to predict the St Venant solution under highly dynamic conditions over relatively long distances. A R2 T of over 0.95 is observed even at a downstream distance of 40 km. However, note that the linearised model depends on the channel shape since the friction slope depends upon the hydraulics radius.

Table 5 Linear mode multilinear lag-cascade results at 20 and 120 km for channel types 1, 2 and 3 Channel identication Distance (km) Linear mode multilinear lag-cascade Qp (m s ) 1 2 3 20 120 20 120 20 120 421 261 463 352 498 490
3 1

Linear mode multilinear cascade R2 T 0.9791 0.8071 0.9851 0.7735 0.9952 0.7890 R2 T 0.9661 0.7642 0.9853 0.7510 0.9932 0.7866

tp (h) 13.5 33.75 12.75 26.5 11.17 17.67

td (h) 2.25 11.0 1.75 8.25 1.5 6.5

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

43

(a) Channel type 4 450 Discharge [m3s1] Discharge [m3s1] 400 350 300 250 200 150 100 0 20 40 60 40km 20km 450 400 350 300 250 200 150 100 0

(b) Channel type 5 20km

40km

20

40

60

(c) Channel type 6 450 Discharge [m3s1] Discharge [m3s1] 400 350 300 250 200 150 100 0 20 40 Time [hr] 60 100 0 40km 20km 500

(d) Channel type 1 Input 400 300 200 Output 20 km 20 40 Time [hr] 60

Fig. 9. Linear mode multilinear lag-cascade model routing results for dynamic channels: the St Venant solutions (solid line); and the linear mode multilinear lag-cascade model (dotted line).

The linearisation presented by Dooge and Harley (1967) assumes a wide rectangular channel and Chezy friction law. For different channel types the general extension given by Dooge et al. (1987a,b) should be used. Typical results for a trapezoidal cross-section (channel type 5) and an irregular cross section (channel type 6) are also presented in Table 4 and are shown in Fig. 9b and c. These results suggest that use of the linear mode of the model is quite acceptable for practical purposes up to 40 km in channels of similar type to those used here. These results indicate that the method of parameter estimation based on the generalised linear channel response is producing accurate approximations. In particular note the excellent reproduction of the full equation hydrographs for channel type 6. This result for an irregular channel shape and fully dynamic wave is of considerable practical signicance. The integral computational approach of the modelling methodology also allows the prediction of wave trains. For example, Fig. 9d shows that the predicted translation and attenuation of two consecutive

sinusoidal ood waves in a channel of type 1 and at a distance of 20 km is very accurate with a R2 T of over 0.99. 5.3. Benchmark simulations As mentioned previously, the multilinear lagcascade method could be considered computationally demanding since it requires small simulation routing times steps in order to avoid the formation of irregularities in the peak regions of the simulated hydrographs pulses and to minimise the truncation error introduced by using an integer advective time delay (as a multiple of Dt). In order to test the relative efciency of the model a series of benchmark runs were conducted on a Pentium PC (166 MHz processor, 32 MB RAM) using implemented codes for the multilinear lag-cascade and full St Venant equations. To aid comparison both codes were implemented using the same numerical programming language matlab (Mathworks, 1996). The implicit St Venant equations solver uses the NewtonRaphson scheme (Fread et

44 Table 6 Benchmark simulation results Channel identication

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

Routing distance (km)

Simulation time (h)

Run-time (CPU s) Full St Venant solver Dt 0:25 h; Dx 4 km Discrete lag-cascade Dt 0:25 h; single reach Linear mode Multilinear 8 8 13 18 232 249

3 3 1 1 6 6

20 120 20 120 20 120

90 90 90 90 180 180

28 150 35 187 243 880

5 5 7 10 24 33

al., 1985; Fread, 1993; Chow et al., 1988) with an efcient solution technique that takes advantage of the banded (quad-diagonal) structure of the system of simultaneous equations (Fread, 1971). Simulation parameters for the benchmark runs are those of channels No. 1, No. 3 and No. 6 using both the multilinear mode with a 0:2 and the linear mode of the model a 0: Hydrographs were computed at 20 km and 120 km for a real time simulation of 90 h in the former channels and 180 h in the latter using the temporal and spatial discretisation described in Section 4. Table 6 summarises the results from the benchmark simulations. Two main conclusions may be drawn from the table. Firstly, the processing run-time depends on the type of channel used and the hydraulic conditions prevailing in the channel. Kinematic ow in a rectangular channel being the less computationally expensive situation and a fully dynamic wave in an irregular channel being the most computationally demanding case. This is expected for the full equations since the number of iterations required in the NewtonRaphson algorithm generally increase in the dynamic case and interpolations are required to estimate hydraulic variables from the tabular values of channel width and elevation that dene an irregular cross-section. In the case of the multilinear lagcascade model the unit pulse response function is very at for the dynamic conditions and requires a large number of unit pulse response ordinates (recall a large M value in Eq. (2)) for its accurate description. Therefore, the number of operations required in the convolution procedure increases accordingly. In the case of the linear mode of the model however, the

unit pulse response is computed only once in time and so the computational cost is considerably reduced compared to the multilinear mode of the model. Secondly, the multilinear lag-cascade can be considered to be relatively computationally expensive compared to the full equations solver taken into account that in the approximate method only discharge is being computed at a single location whereas depth and discharge are simultaneously computed at all sections in the distributed model. Nevertheless, in the cases analysed savings of computer time of 1000% are obtained in long channels. This difference is important for instance if ow prediction is incorporated within adaptive ood warning schemes (Lees, 1999; Lees et al., 1994) or probabilistic decision support frameworks that undertake Monte Carlo based uncertainty estimation. 6. Conclusions The proposed multilinear discrete lag-cascade ood routing method has the following features: It is based on a discrete linear channel coupled with a discrete cascade of reservoirs. This sub-model is simple, well studied and acceptable among hydrologists. Improved predictions in comparison with Perumals multilinear model (Perumal, 1994) have been obtained for uniform channels with very low slopes and high resistance effects characterised by open looped-rating curves at the inlet and outlet boundaries (i.e. fully dynamic channels).

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047

45

The multilinear approach proves to be effective in simulating the nonlinear characteristics of ood waves. However, there is evidence that the linear mode of operation can provide an adequate description of the ood propagation phenomenon for many practical purposes where long reaches can be sub-divided. This is especially the case for channels that show high ood wave attenuation. A considerable proportion of the ood wave translation process is explained by the explicit time delay of the discrete channel. This description is advantageous in ood forecasting applications because it resumes information about the leadtime available before the ood wave arrives at a downstream location. No deciencies in mass conservation result from the integral convolution approach and multilinear scheme. As with other hydrological ood routing models, it is still not possible to consider backwater effects due to tides and tributaries or the effect of critical control sections at the downstream location. The authors are not aware of further analysis or studies of the assumption of a steady-state rating curve at the downstream end on the accuracy of the solution as suggested by rkowski (1987). Also in line with the Dooge and Napio KalininMilyukov method assumptions, within this method the stage value at a given section cannot be estimated using the corresponding computed discharge at the same section using a steady-ow rating curve (Miller and Cunge, 1975; Weinmann and Laurenson, 1979; Perumal and Ranga Raju, 1998). An approximate method of estimating the model parameters is presented in this paper. The model parameters are related to the parameters of a linearised version of the St Venant equations by the cumulant matching method. The proposed relationships successfully reproduce the non-linear characteristics described by relationships of travel time and time delay with discharge. In particular, the observed high non-linearity under low ow conditions is reproduced. The relationships presented are valid for any shape of channel and any friction law. The model accuracy, with parameters estimated by means of the method described in this paper is

at most as good as the accuracy that the linearised St Venant model can provide. However, the proposed three-parameter model is certainly simpler to compute than the complete linearised model and is implemented in just 60 lines of the numerical programming language matlab (Mathworks, 1996) including data reading, processing and plotting. The multilinear approach uses time-variable parameters that are estimated as a function of a time-variable reference discharge. Effectively, Eq. (10) provides a simple objective method of implementing a multilinear time distribution scheme. In the present study small values for the a coefcient in Eq. (10) 0 a 0:4 were observed to reproduce ood wave non-linearity. A typical a value for a specic type of channel can be chosen from an a priori calibration exercise using observed input and output hydrographs. In any case, for short reaches the linear operation mode with time-invariant parameters produced sufciently accurate results for most practical purposes. Finally, the success of the linear mode is important in that the linear model enables the practical utilisation of powerful methods of system identication in the estimation of model parameters. Indeed the results presented in this paper validate the use of linear transfer function models in previous ood forecasting applications (Lees, 1999; Lees et al., 1994). Recently developed non-linear system identication methods (e.g. Lees, 1999; Young and Beven, 1994) may provide a possible method of parameter estimation for the full nonlinear form of the model. The potential of these nonlinear model parameter estimation techniques for ood routing purposes is currently under investigation.

Acknowledgements We would like to thank Danny Fread and an anonymous reviewer for their thorough reviews that have resulted in substantial improvements to the paper. Luis Camacho is grateful to COLCIENCIAS, the Colombian Institute for Science and Technology Development for nancial support granted during 1997/98.

46

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047 surface water modelling, Proceedings of Baltimore Symposium, 181. IAHS, pp. 115124. Meyer, O.H., 1941. Simplied ood routing. Journal of the Civil Engineering Division, ASCE 11 (5), 306307. Miller, W.A., Cunge, J.A., 1975. Simplied equations of unsteady ow. In: Mahmood, K., Yevjevich, V. (Eds.). Unsteady Flow in Open Channels, Water Resources Publications, Ann Arbor, MI. Nash, J.E., 1960. A unit hydrograph study with particular reference to British catchments. Proceedings of Institution of Civil Engineers 17, 249282. Nash, J.E., Sutcliffe, J.V., 1970. River ow forecasting through conceptual models. Part I a discussion of principles. Journal of Hydrology 10, 282290. OConnel, P.E., Brunsdon, G.P., Reed, D.W., Whitehead, P.G., 1986. Case studies in real-time hydrological forecasting for the U.K. In: Kraijenhoff, D.A., Moll, J.R. (Eds.). River Flow Modelling and Forecasting, Reidel, Lancaster. OConnor, K.M., 1976. A discrete linear cascade model for hydrology. Journal of Hydrology 29, 203242. Perumal, M., 1992. The cause of negative initial outow with the Muskingum method. Hydrological Sciences Journal 37, 391 401. Perumal, M., 1992. Multilinear Muskingum ood routing method. Journal of Hydrology 133, 259272. Perumal, M., 1993. The cause of negative initial outow with the Muskingum methodClosure. Hydrological Sciences Journal 38, 153154. Perumal, M., 1994. Multilinear discrete cascade model for channel routing. Journal of Hydrology 158, 135150. Perumal, M., Ranga Raju, K.G., 1998. Variable-parameter stagehydrograph routing method, Theory. Journal of Hydrologic Engineering, ASCE 3 (2), 109114. Ponce, V.M., 1991. The kinematic wave controversy. Journal of the Hydraulic Division, ASCE 117 (4), 511525. Ponce, V.M., Chaganti, P.V., 1994. Variable-parameter MuskingumCunge method revisited. Journal of Hydrology 162, 433439. Ponce, V.M., Theurer, F.D., 1982. Accuracy criteria in diffusion routing. Journal of the Hydraulic Division, ASCE 108 (6), 747757. Ponce, V.M., Yevjevich, V., 1978. MuskingumCunge method with variable parameters. Journal of the Hydraulic Division, ASCE 104 (12), 16631667. Ponce, V.M., Li, R.M., Simons, D.B., 1978. Applicability of kinematic and diffusion models. Journal of the Hydraulic Division, ASCE 104 (3), 353360. Singh, V.P., Wang, G., Adrian, D.D., 1997. Flood routing based on diffusion wave equation using mixing cell method. Hydrological Processes 11, 1894. Wang, G., Chen, S., 1996. A new model describing convective dispersive phenomena derived by using the mixing-cell concept. Applied Mathematical Modelling 20 (4), 309320. Weinmann, P.E., Laurenson, E.M., 1979. Approximate ood routing methods: A review. Journal of the Hydraulic Division, ASCE 105 (12), 15211536. Whitehead, P.G., Young, P., Hornberger, G., 1979. A systems

References
Becker, A., Kundzewicz, Z.W., 1987. Nonlinear ood routing with multilinear models. Water Resources Research 23, 10431048. Bentura, P.L., Michel, C., 1997. Flood routing in a wide channel with a quadratic lag-and-route method. Hydrological SciencesJournal-des Sciences Hydrologiques 42, 169185. Cunge, J.A., 1969. On the subject of a ood propagation computation method (Muskingum Method). Journal of Hydraulic Research 7, 205230. Chow, V.T., Maidment, D.R., Mays, L.W., 1988. Applied Hydrology, McGraw-Hill, New York. Dooge, J.C.I., 1973. Linear Theory of Hydrologic Systems. Agricultural Research Service, Tech. Bull. No. 1468, USDA, Washington, DC. Dooge, J.C.I., 1986. Theory of ood routing. In: Kraijenhoff, D.A., Moll, J.R. (Eds.). River Flow modelling and Forecasting, Reidel, Lancaster. Dooge, J.C.I., Harley, B.M., 1967. Linear routing in uniform open channels. Proceedings International Hydrology Symposium, Fort Collins, CO 1, 5763. rkowski, J.J., 1987. The effect of the downDooge, J.C.I., Napio stream boundary conditions in the linearised St Venant equations. The Quarterly Journal of Mechanics and Applied Mathematics 40 (2), 245256. rkowski, J.J., Strupczewski, w.G., 1987. The Dooge, J.C.I., Napio linear downstream response of a generalised uniform channel. Acta Geophysica Polonica 35 (3), 277291. rkowski, J.J., Strupczewski, W.G., 1987. PropDooge, J.C.I., Napio erties of the generalised downstream channel response. Acta Geophysica Polonica 35 (4), 405418. Fread, D.L., 1971. Discussion of implicit ood routing in natural channels by M. Amein and C.S. Fang. Journal of the Hydraulic Division, ASCE 97 (7), 11561159. Fread, D.L., 1985. Channel routing. In: Anderson, M.G., Burt, T.P. (Eds.). Hydrological Forecasting, Wiley, Chichester. Fread, D.L., 1993. In: Maidment, D.R. (Ed.). Handbook of Hydrology, 946. McGraw-Hill, New York, pp. 946951. Kalinin, G.P., 1971. Global Hydrology Transactions, Russian, Israel Program for Scientic Translations, Jerusalem, Israel. Keefer, T.N., McQuivey, R.S., 1974. Multiple linearisation ow routing model. Journal of the Hydraulic Division, ASCE 100 (7), 10311046. Kundzewicz, Z.W., 1984. Multilinear ood routing. Acta, Geophysica Polonica 32, 419445. Laurenson, E.M., 1964. A catchment storage model for runoff routing. Journal of Hydrology 2, 141163. Lees, M.J., 1999. Data-based mechanistic modelling and forecasting of hydrological systems. Journal of Hydroinformatics 3, in press. Lees, M.J., Young, P.C., Ferguson, S., Beven, K.J., Burns, J., 1994. An adaptive ood warning scheme for the river Nith at Dumfries. In: White, W.R., Watts, J. (Eds.). Second International Conference on River Flood Hydraulics, Wiley, pp. 6577. Malone, T.A., Cordery, I., 1989. An assessment of network models in ood forecasting. In: Kavvas, M.L. (Ed.). New directions of

L.A. Camacho, M.J. Lees / Journal of Hydrology 226 (1999) 3047 model of stream ow and water quality in the Bedford-Ouse River-1. Stream ow modelling. Water Research 13, 1155 1169. Young, P.C., 1986. Time-series methods and recursive estimation in hydrological systems analysis. In: Kraijenhoff, D.A., Moll, J.R.

47

(Eds.). River Flow Modelling and Forecasting, Reidel, Lancaster. Young, P.C., Beven, K.J., 1994. Data-based mechanistic modelling and the rainfall-ow nonlinearity. Environmetrics 5, 335363.

You might also like