Flood Routing in Long Channels

download Flood Routing in Long Channels

of 13

Transcript of Flood Routing in Long Channels

  • 8/6/2019 Flood Routing in Long Channels

    1/13

    Journal of the Chinese Institute of Engineers, Vol. 28, No. 7, pp. 23-35 (2005) 23

    FLOOD ROUTING IN LONG CHANNELS:

    ALLEVIATION OF INCONSISTENCY AND DISCHARGE DIP IN

    MUSKINGUM-BASED MODELS

    Weihao Chung* and Yi-Lung Kang

    ABSTRACT

    The Muskingum parameters are often expressed as a function of discharge and

    channel properties (Cunge, 1969; Chow et al., 1988) by referring to the coefficients

    of the advection-diffusion (A-D) equation or the equivalent parabolicized Saint Venantequation. As extensively investigated in this paper by Taylor series expansion, the

    Muskingum model and its variants are found tangibly inconsistent to the A-D equation.

    In addition, these Muskingum-based models experience the effect of negative outflows,

    i.e., the well-known dip phenomenon. To avoid these problems, which are both present

    in the conventional routing procedure, this paper introduces an extra term to the tradi-

    tional Muskingum storage function, that is then linked to Gills concept of initial

    storage. Through this technique, not only the dip phenomenon but also the model

    inconsistency can be alleviated, and a fairly satisfactory outflow prediction can thereby

    be achieved. A proper time to employ the extra term is sought by a convolution inte-

    gral which is a result of the Laplace transformation applied to the Muskingum model.

    It also represents the analytical expression of outflow discharge. The convolution

    integral enables us to trace the origin of discharge dip and quantify the shape varia-

    tion of outflow hydrographs. With the convolution integral, compact models for com-

    puting the maximum flow dip, dip, and its occurrence time, tc, are also offered in

    this study. As a concluding example, routings by the traditional Muskingum model,

    Gills procedure, and the newly developed algorithms having the extra term are per-

    formed in a long channel reach of 90 kilometers to test the robustness of each model.

    Key Words: model inconsistency, dip phenomenon, Muskingum model, flood routing,

    storage function.

    *Corresponding author. (Tel : 886-7-7456290; Email :

    [email protected])

    W. H. Chung and Y. L. Kang are with the Department of Civil

    Engineering of the Chinese Military Academy, Fengshan, Taiwan830, R.O.C.

    I. INTRODUCTION

    Due to limited field data and marked topogra-phy changes of river channels, it is usually impos-

    sible and unnecessary to perform high-dimensional

    river flow analyses using complex mathematical

    models. In this situation, flood wave transport and

    dissipation are often simplified as one-dimensional

    problems in a semi-infinite flow domain, provided that

    detailed flow velocity distribution and backwater

    effects induced by downstream perturbations are

    trivial. The Muskingum model is suitable in this

    situation and plays a fairly important role due to itssimplicity and practicability. Generally, one may per-

    form flood routings and risk analyses for a flood event,

    once the Muskingum parameters, Kand X, that ac-

    count for wave traveling time and dispersion,

    respectively, are calibrated using flow data measured

    in the past. Of course strong material scouring or

    deposition along with flood flows must not change

    the parameter values significantly.

    Stability analysis of the Muskingum model or

    its variant Koussiss formulation (Koussis, 1978)can be carried out through the use of von Neumanns

    method (Smith, 1985) which represents the modelsas a series of complex exponentials. By dampening

  • 8/6/2019 Flood Routing in Long Channels

    2/13

    24 Journal of the Chinese Institute of Engineers, Vol. 28, No. 7 (2005)

    out the round-off error, the Muskingum model can

    be shown to be unconditionally stable, while Koussiss

    formulation has the restriction that the Courant num-

    ber must be less than 1 (Chung, 1994). Chow and

    Kulandaiswamy (1971) developed a general lumped

    system model (not concerning the initial storage) by

    Taylors series expansion to simulate various hydro-

    logic phenomena. The input and output of a water-

    shed system were linked by a transfer function which

    could be used to formulate the instantaneous unit

    hydrograph. A five-term model, consisting of inflow

    and outflow discharges and their derivatives, was rec-

    ommended for practical application but the article

    lacked a discussion of discharge dip and model in-

    consistency (see the following paragraph for

    definition). The impact of calibrated parameters on

    varying outflow hydrograph shape was also not

    mentioned. To produce a stable general hydrologicmodel, they outlined the restriction condition but it

    seemed easily violated since the transfer function led

    to three roots that resulted in more unexpected com-

    binations than usual.

    The flaw of the Muskingum model falls in the

    outcome of equation/model inconsistency (Smith,

    1985), that is, a model differs from its original dif-

    ferential equation as it is inverted from a discrete form

    to a differential one. Despite the fact that Muskingum

    model can be regarded as the discrete form of the A-

    D equation or the equivalent parabolicized Saint

    Venant equation, it is shown in section II that the

    former can not be inverted to the latter exactly (i.e.,

    inconsistent with A-D equation). The worst thing

    about model inconsistency is that the inconsistency

    may converge our numerical solutions to unexpected

    ones and thereby mislead our judgment in spite of

    the fact that all parameters are carefully calibrated.

    As shown in this paper, the inconsistency represents

    the very factor leading to the underestimation of peak

    discharge and possibly the time to peak. Satisfac-

    tory routing results in a long channel are seldom ob-

    tained by using the conventional Muskingum model

    or its variants due to the inconsistency problem.

    A n o t h e r p r o b l e m o f t h e c o n v e n t i o n a lMuskingum method is negative discharges or the so-

    called dip phenomenon. The dip is often interpreted

    as a result of choosing either inadequate time inter-

    vals (Hjelmfelt, 1985) or grid sizes (Ponce and

    Theurer, 1982), or the nature of the storage function

    (Nash, 1959) defined by Kand X. Boneh and Golan

    (1979) considered it possible to have a negative in-

    stantaneous unit hydrograph in nature and took it as

    the cause of negative discharges. Gill (1980) viewed

    the same problem as a consequence of introducing

    inadequate initial conditions, but this view was re-

    jected by Singh and McCann (1980). Gill (1980) alsoproposed a time lag period within which outflow

    hydrographs do not change due to the arrival of flood

    at the inlet section of the flood routing reach. This is

    consistent with reality but seems useless in improv-

    ing present-day models. To get a proper fit of out-

    flow hydrographs, Gill (1977) proposed the concept

    of relative storage in which the Muskingum param-

    eters Kand Xare re-calibrated. In the light of Gills

    concept, Aldama (1990) proposed the formula for

    determining K, X, and the initial storage that links a

    relative storage to an absolute one. In his technical

    report he shows the calibration formula with relative

    storage instead of outflow rates. That leads to more

    accurate predictions (Heggen, 1984, and ODonnell,

    1985, for comparisons ) of the time to peak discharge

    and results in smaller root-mean-square errors, with

    a trade off of more marked discharge dips.

    If lengthy complicated mathematical deduction

    is implicated, the Kalman filtering method (Wood andSzollosi-Nagy, 1978; Ngan and Rusell, 1986; Kan,

    1995; and Van Geer et al., 1991) can be employed to

    eliminate the dip phenomenon. The greatest advan-

    tage one may take from the filtering method is to re-

    lax the restriction of fixed Muskingum parameters for

    a given flood event. By substituting optimized pa-

    rameters within each time interval, Kalmans filter-

    ing offers an excellent alternative to route outflow

    hydrographs with almost perfect performance (Kan,

    1995), even if inflow hydrographs with tributary flows

    vary with time in a complex way. This seems to be

    good for the conventional Muskingum method but still

    does not address the real problem causing the dip

    phenomenon. Besides, due to the nature of temporal

    variations, Kalmans filtering parameters can not be

    interchangeably employed through all flood events.

    That is, the parameters must be recalibrated from one

    flood event to another. Thus, before clarifying the

    cause of dip phenomena, it would still be valuable to

    re-study either the Muskingum model or its variants.

    In order to solve the problems mentioned above

    (model inconsistency and discharge dip), a time con-

    stant R is introduced in the conventional Muskingum

    storage function. By substituting the so obtained stor-

    age function into the continuity mass equation it willhopefully yield an extra term to counterbalance the

    factor causing model inconsistency, and consequently

    reduce the dip phenomenon substantially. Following

    this idea, this paper presents two models essentially

    good for flood routings in long channels. The first is

    called the modified Muskingum algorithm (MMA)

    being a variant of the Muskingum model withR. The

    second is called the modified Gills algorithm (MGA)

    containingR and Gills relative storage. Thus, MMA

    includes three basic parameters while MGA has four.

    Robustness of the new algorithms will be tested and

    confirmed through a series of comparisons with theconventional Muskingum model (CMM) and Gills

  • 8/6/2019 Flood Routing in Long Channels

    3/13

    W. Chung and Y. L. Kang: Flood Routing in Long Channels: Alleviation of Inconsistency and Discharge Dip 25

    procedure with Aldamas calibration formula (GILL)

    in a super long channel of 90 kilometers. Test re-

    sults can be used further for confirming the

    postulation, as addressed by Laurenson (1959), that

    a single-valued storage function is not adequate for

    storage routing in a fairly long channel.

    In brief, this paper reveals first the model in-

    consistency embedded in CMM by Taylors series

    expansions, then under various conditions obtains the

    MMAs analytical solutions of outflow rates in a form

    of convolution integral. With these solutions, the ori-

    gin of discharge dip can be found along with the shape

    variation of outflow hydrographs. The connection be-

    tween model inconsistency and discharge dips can

    thereby be explored. By short time approximation,

    the forcing and response functions in the convolu-

    tion integral are simplified to formulate a series of

    compact models for computing tc and dip. Resultsare verified by a sensitivity analysis example. About

    R, its application range is found through the study of

    the response function that may be unbounded under

    certain flow conditions. The best timing for adopt-

    ing R is also proposed through the analysis of the

    modified equation and by the model prediction ofdip .

    Finally, explicit algorithms are developed for param-

    eter estimations of MMA and MGA. Model compari-

    sons are then made to test their applicability and

    robustness in a fairly long channel.

    II. MODEL INCONSISTENCY

    Since the mass balance equation has no need

    to be modified, attaching one or more terms to the

    Muskingum storage function, S, may be the only al-

    ternative for avoiding the inherent inconsistency and

    flow dips. As a rule of thumb, S must be changed in

    such a way that the modified equation remains in the

    same form as the A-D equation (Qt + UQx =DQxx, Q

    = discharge, U= advection coefficient, D = disper-

    sion coefficient) that basically governs diffusion

    waves in open channel flows. It is found, after sev-

    eral trials, that the most compact way is to introduce

    in S an extra term RdO/dtwhere R has a unit of timesquare. This transforms S into the following

    expression:

    S = KO + KX(I O) + RdOdt

    (1)

    where t denotes t ime and I and O represent,

    respectively, the discharges flowing in and out of a

    finite channel reach. In addition to the storage shape

    described by the second and third terms of Eq. (1),

    the newly added term RdOdt

    accounts for the additional

    volume induced by the rate of rise of outflowhydrographs which, aside from nonlinear effects, may

    contribute to a looped rating curve, hence making the

    storage function physically and mathematically more

    complete.

    Substituting Eq. (1) into the mass balance equation:

    dSdt

    = I O (2)

    yields the modified CMM of discharge as below:

    Rd2O

    dt2+ K(1 X)dO

    dt+ O = KXdI

    dt+ I (3)

    which can be reorganized and discretized into the

    following difference form:

    XIn + 1 In

    t+ (1 X)

    On + 1 On

    t+

    On + 1 In + 1

    2K

    + On

    In

    2K+ R

    K(O

    n + 1

    2On

    + On 1

    t2) = 0 (4)

    where n denotes the time level.

    To derive the modified form of Eq. (4), the dis-

    charges associated with different time levels and lo-

    cations in Eq. (4) are first expanded about j and n as

    a Taylors series. For example:

    In + 1 = Qjn + 1 = Qj

    n + Qtjnt+ t

    2

    2Qttj

    n + t3

    6Qtttj

    n +

    On = Qj + 1n = Qj

    n + Qxjnx + x

    2

    2Qxxj

    n + x3

    6Qxxxj

    n +

    wherej represents the spatial location along a channel,

    etc. The discharge On + 1 (= Qj + 1n + 1 ) can be similarly

    expressed as a function ofOjn (= Qj + 1

    n ) by Taylors

    series expansion for a two variable function (Kaplan,

    1981). Substituting all the results into Eq. (4) with K

    = x/Uyields an equation in Xand U. Applying thedifferential operator /t=D2/x2U/x to the re-sultant equation and denoting Oj

    n as Q yields the fol-

    lowing modified equation after some lengthy algebra:

    Qt + 1Qx = 2Qxx + 3Qxxx + 4Qxxxx + ... (5)

    where 1 = x/Kand

    2 = (12

    X)Ux RU3

    x

    3 = (13

    Cr4

    Cr

    2

    12 1

    Pg(1

    Cr2

    ))Ux2

    + ( 1Pg

    12

    +Cr2

    )XUx2 (1 2Pg

    )RU3

    4 = (1

    Pg(1

    2 Cr)X

    Cr

    2Pg2

    1Pg

    (12

    34

    Cr))Ux3

    (1 +

    Cr2

    12 +1

    Pg2

    2Pg )RU

    3

    x

  • 8/6/2019 Flood Routing in Long Channels

    4/13

    26 Journal of the Chinese Institute of Engineers, Vol. 28, No. 7 (2005)

    In the above equation, x = grid size (channel reachlength), t= time interval, Cr = Ut/x = the Cou-rant number and Pg = Ux/D = the grid Peclet number.

    Now, letting the numerical dispersion equal the

    physical one, that is, 2 = D, produces the formulafor computing X:

    X = 12

    1Pg

    RU2

    x2(6)

    To make Eq. (5) consistent with the A-D equation,

    we have 3 = 0 which with Eq. (6) yields the analyti-cal expression ofR as followings:

    R = x2

    U2(

    1 Cr2 12Pg

    2

    6 + 6Cr 12Pg1

    ) (7)

    Eq. (7) can be also re-expressed as an explicit func-tion ofK and X by substituting Eq. (6) into Eq.(7)

    with U= x/K, which gives:

    RK2

    = 18

    (2 6X tK

    )

    + 12

    X2

    4+ X

    2 5

    48(t

    K)2 +

    (3X 1)4

    tK

    112

    (8)

    Aside from the above analytical expressions, the

    following limitations also result for very small x andt:

    R DU3

    x , X 12

    2Pg

    , and 4 D3

    U2

    This gives the limit truncation error of Eq. (5):

    ET = D3

    U2Qxxxx (9)

    Therefore, introducingR in CMM does relax the in-

    consistency embedded in Qxxx terms (see Appendix

    for comparisons), but evokes another unexpected non-zero term 4Qxxxx. Though the limit value of4 maynot be zero, (i.e., the inconsistency problem still

    exists), the discrepancy is alleviated from a third-or-

    der to a fourth-order derivative. Eq.(9) can be further

    inverted to a function ofKand Xthrough the use of

    Eq.(6) and K= x/U:

    ET = x4

    K(1

    2 X R

    K2)3Qxxxx (10)

    Examining Eq. (9) or (10) reveals that model incon-

    sistency becomes worse for a decreasing U (i.e., in-

    creasing K) or an increasing D and x. That is, it iseasy to produce phase errors and underestimate the

    outflow peak discharge of a flood moving slowly but

    undergoing strong dispersion in a long channel. Indeed,

    channel length has dominated impacts on ET due to

    the fourth power ofx in Eq. (10). It is understoodfrom Eq. (5) that ET behaves like a forcing function

    constantly acting on the discharge Q for a period of

    x/U, hence leading to the unavoidable modelinconsistency. This is the fate of CMM and its vari-

    ants with constant coefficients. The inconsistency can

    not be easily removed by simply adjusting time intervals.

    III. ANALYTICAL EXPRESSIONS

    1. The Analytical Solution ofO

    By deducting the base flow rate Qb from O and

    letting I = I Qb and O = O Qb, Eq. (3) can be

    solved for O by Laplace transformation with O = 0and dO/dt = 0 at t= 0. By so doing, we obtain the

    Laplace-transformed O which is divided into three

    branches as below:

    (i) for R = 0: O = 1K(1 X)s + 1

    f;

    (ii) for R < 0: O = 1 ( 1s 1

    s )f;

    (iii) forR > 0:

    O = 2(

    2R)2

    (s + K(1 X)2R

    )2 + (2R

    )2f, if < 0

    same as (ii), if 0

    where

    = e stdt0

    ,

    = K2(1 X)2 4R,

    = ,

    =1

    2R [ K(1 X) + ] ,

    = 12R

    [ K(1 X) ] ,

    and the forcing function of Eq. (3) is given by:

    f = f(t) = KXd Idt

    + I (11)

    Generally,f(t) has a great chance to be negative for a

    short tsince in this case d Idt I/t and tseldom ex-

    ceeds KX.

    The resultant analytical solution of cases (i) (ii)

    and (iii) is then generally expressed by a convolutionintegral of the form:

  • 8/6/2019 Flood Routing in Long Channels

    5/13

    W. Chung and Y. L. Kang: Flood Routing in Long Channels: Alleviation of Inconsistency and Discharge Dip 27

    O = f(t u)g(u)du0

    t

    (12)

    in which the response function g(t) reads (Doetsch,

    1947):

    g(t) = 1K(1 X)

    exp( tK(1 X)

    ) (13)

    for R = 0;

    g(t) = 2 sin(2R

    t)exp( K(1 X)

    2Rt) (14)

    forand R > 0 and < 0;

    g(t) = 1 [exp(t) exp(t)] (15)

    otherwise. Eqs. (14) and (15) are the expression of

    the MMA approach, which becomes MGA if Gillsrelative storage is considered. The storage will lead

    to KandXvalues different from those of MMA; there-

    fore producing different g(t) values and outflow

    hydrographs.

    2. Model Restrictions Stability Analysis

    Due to the quadratic differential form of Eq. (3),

    the required mathematical work for von Neumanns

    stability analysis will become rather discouraging. In

    this case, it will be more straightforward to achieve

    the goal by looking into the response function of Eq.

    (12). Apparently, is always positive ifR < 0. Thisfact makes g(t) and eventually O unbounded as tgoes

    to infinity, meaning routing procedures with R < 0

    must be excluded. It is also noteworthy that, forR0, and are both negative and || < || since K(1

    X) is always positive. Thus, g(t) will remain finite

    positive and will be that given by Eq. (13). The same

    conclusion applies to Eq. (14) considering that the

    dip phenomenon usually occurs at the initial stage of

    a flood event (i.e., a short range of time) which makes

    sin(t) positive. This assures the convergence of cal-

    culated O .

    To keepR positive, the numerator and denomi-nator in the parenthesis of Eq. (7) must have the same

    sign, i.e.,

    1 Cr2 12Pg

    2 > 0 and 6 + 6Cr 12Pg 1 > 0 ,

    or

    1 Cr2 12Pg

    2 < 0 and 6 + 6Cr 12Pg 1 < 0 .

    The above criteria can be further transformed into:

    Pg 2 < (1 Cr

    2)/12 and Pg 2 < (1 + Cr)

    2/4 ,

    or

    Pg 2 > (1 Cr

    2)/12 and Pg 2 > (1 + Cr)

    2/4

    which yields, respectively, the following restrictions

    recognizing that Pg > 0:

    Pg >12

    1 Cr2

    , if Cr < 1 (16)

    or

    Pg >2

    1 + Cr, for all Cr (17)

    A series of numerical tests shows that Eq. (4) is con-

    ditionally stable for R > 0 as has been reasoned by

    Eq. (12) for case (ii). When channel flow is subjected

    to a small perturbation, Eqs. (16) and (17) based on

    the linear theory are useful and become the equiva-

    lent criteria to assure numerical stability.

    3. The Impact ofg(t) on O

    Likef(t) being connected to inflow hydrographs,

    g(t) has a strong impact on the shape of predicted

    outflow hydrographs. By carefully examining Eqs.

    (13) - (15), it is found that the g(t) of each equation

    differ completely from each other in regard to varia-

    tion trends. For example, g(t) of Eq. (14) is an expo-

    nentially decayed sine function. It becomes however,

    a monotonically decreased exponential function for

    Eq. (13), and a combination of Eqs. (13)(14) for Eq.

    (15). The g(t) of Eq. (15) rises from zero at t = 0,

    then increases monotonically with time all the way

    to its maximum value, then falls down monotonically

    again to approach the time abscissa as t goes to

    infinity. Accordingly, the MMA corresponding to Eq.

    (14) will evolve the outflow hydrgraphs with up-

    and-down oscillatory tails, depending totally on the

    relative importance ofK2(1 X)2 and 4R. It is im-possible however, for the CMM associated with Eq.

    (13) to generate such outflow hydrographs. By in-

    corporating with Gills relative storage, MGA has a

    greater chance to meet the criterion for Eq. (15), hence

    it is also unlikely to produce oscillatory tails.Since g(t) affects outflow hydrograph shape, it

    may also change the corresponding phase error of

    time to peak discharge. For example, ifg(t) decays

    exponentially as described by Eq. (13), the negative

    value off(tu) will be amplified but concentrated ina small u through the convolution integral of Eq. (12).

    That is, the dip phenomenon will occur at the initial

    stage of a flood and it quickly decays out as t

    increases, accompanying the underestimation of the

    time to peak discharge. The situation becomes less

    serious for g(t) of Eq. (14) that increases then de-

    creases and takes a longer time to become zero, mak-ing an aggregate of the negative discharges difficult.

  • 8/6/2019 Flood Routing in Long Channels

    6/13

    28 Journal of the Chinese Institute of Engineers, Vol. 28, No. 7 (2005)

    Result in dip is postponed and the time to peak dis-

    charge is over-predicted.

    IV. FLOW DIPS

    1. The Cause

    It has been shown that all g(t)s are positive as

    long as tis adequately small. The function value of

    f(tu) based on Eq. (12) must therefore be negativeat a certain time or within some period if negative

    outflow rates appear. Accordingly, factors causing

    the dip phenomenon are explained as the nature of

    f(t). Since f(t) does not include R, flow dips fully

    depend on KX values and the shape of inflow

    hydrographs . The smal ler KX or the milder

    hydrograph shape is (generates possibly a kinematic

    wave), the more vague the dip phenomenon appears.

    2. The Occurrence Time

    In fact, the dip phenomenon seems to be

    unavoidable, otherwise f(t) must remain positive all

    the way. Taking a Gaussian inflow hydrograph:

    I= Qb + Qp exp{[(ttp)/]2} (18)

    for example, the forcing function can be expressed

    as:

    f(t) = (2KXt tp

    2+ 1)Qpexp[ (

    t tp

    2)2] (19)

    Its function values are all negative as tis smaller than

    the following critical time:

    tc = tp 2

    2KX(20)

    In above, Qp = peak inflow rate, tp = the time to peak

    discharge, = constant. Eventually, the function val-ues off(tu) are also negative for all us less than t,given t< tc. It is obvious then the area enclosed by

    f(tu) but below the time abscissa reaches its maxi-

    mum as t= tc which thus can be roughly identified asthe moment yielding dip . We use the word roughly

    because tc still depends more or less on g(t).

    3. The Magnitude

    To evaluate the magnitude ofdip, attention is

    particularly paid to the extreme case t 0. In thiscase, Eq. (13) and Eqs. (14) and (15) are approached

    by g(t) = [K(1 X)]1 and g(t) = t/R, respectively. Asfor Eq. (19), it can be approximated by f(t) =

    (2KXt tp

    2+ 1)Qp in which represents a sufficiently

    small positive number introduced for replacing the

    exponential function of Eq. (19). With all the results

    substituted into Eq. (12), dip can be computed by:

    dip = Xtc

    2

    (1 X)2Qp (21)

    for R = 0, and

    dip = KXtc

    3

    R2Qp (22)

    forR > 0 since dip usually occurs at the initial stage

    of a flood.

    Although (21) and (22) are simplified models

    based on the approximations within a short time range,

    their validity can be extended for a long time to ex-

    plore the impacts of Muskingum parameters on dip

    without knowing the precise value. For example,

    increasing Kor Xwill increase tc (see Eq. (20)) andeventually the dip; a rapidly rising inflow hydrograph

    (i.e., smaller ) with high peak discharge may induceserious flow dips as well; a smallerR value may cause

    the same result. By comparing Eqs. (9) and (10) with

    (21) and (22), one realizes that model inconsistency

    seems to worsen while the dip decays by just varying

    X, but both become more serious as Kor x increases.It is deduced then thatR or the newly developed stor-

    age model (1) contributes the most to the flood routings

    in a long slightly rough channel carrying rapidly in-

    creased inflow discharges. Generally, conventional

    Muskingum algorithms failed in this flow condition

    due to the prevalent inconsistency and discharge dips.

    V. VERIFICATION

    In what follows, a sensitivity analysis serving

    as a supplemental example, is performed to help

    verify Eqs. (20)-(22). To achieve the verification, it

    requires, aside from the linear theory developed in

    section 3, a complete model to compute tc and dip for

    comparison. We differentiate Eq. (12) with respect

    to tand let the result equal zero. This gives the fol-

    lowing formula by using the Newton-Leibnitz inte-

    gration rule with f(0) = 0:

    J(u; tc)g(u)du0

    tc= 0 (23)

    where J(u; tc) =f(t u)

    t t= tc. Sincef(t) involving the

    shape of inflow hydrographs may become too com-

    plicated to integrate Eq. (23), it is quite difficult to

    express explicitly Eq. (23) as a formula for tc. So we

    are using the numerical integration method to seek

    the equivalent Nvalues:

    J(u; tc)g(u) = 0j = 1

    N (24)

  • 8/6/2019 Flood Routing in Long Channels

    7/13

    W. Chung and Y. L. Kang: Flood Routing in Long Channels: Alleviation of Inconsistency and Discharge Dip 29

    where u = (j 12

    )u and tc = Nu. The worst dis-charge dip is thus calculated according to:

    dip = f(tc u)g(u)du0

    tc(25)

    which can be also estimated by numerical integration

    with an adequately small time interval u.

    1. Determination of Influential Parameters

    The inflow hydrograph is again assumed as a

    Gaussian distribution, as given by Eq. (18). To sat-isfy the initial condition that I(0) = 0, an equivalent

    approximation ofI/Qp = 0.001 (i.e., 0.1% error) is

    imposed on Eq. (18). This allows us to obtain a value for a given tp by the formula 0.001 = exp[(tp/)2], hence inflow hydrograph shape on tp and Qponly. Accordingly, the factors affecting tc and dip/

    Qp (see Eqs. (24)-(25)) are just tp, Kand X, since tcand dip/Qp are independent ofQp and R is a function

    ofKandXonly. xs impact can not be ignored sinceit involves the action time off(t). To meet the

    purpose, we choose a fairly long channel reach, say

    90 km, to route outflow hydrographs. Model robust-ness can thereby be tested.

    2. Results

    Figures 1~3 are the calculated results based on

    Eqs.(24)(25) forR = 0 and R 0. The adopted peakdischarge Qp and the time to peak tp of Eq. (18) are

    assumed as 12.5 m3 /s and 6.69 hr, respectively. Any

    nonzero R value is theoretically determined from Eq.

    (8). It can be found from Figs. 1~3 that the critical

    time tc increases monotonically at a decaying rate with

    the increasing K, Xand tp, whether R is included or

    not. The increment oftc becomes even more obvious

    when R is included. This coincides with the afore-

    mentioned deduction that g(t) of Eq. (14) does putoff the time to peak discharge and increase tc as a

    result. It is also found from the figures that tc seems

    to be bounded by an upper-limit, tp, forR = 0, imply-

    ing that the worst discharge dip induced by CMM

    appears before the peak time of an inflow hydrograph.

    As shown in Figs. 1~3, dip exhibits a more com-

    plicated variation trend compared with that oftc. In-

    stead of monotonically increasing, dip increases then

    decreases with the increasing tp. Dip also increases

    with the increasing Kand Xwhether R is included or

    not. Fig. 1 reveals that the incorporation ofR will

    weaken the dip for Kgreater than about 10 hours. Thegreater the K is, the stronger the impact ofR on

    K(hr)

    tc

    (R = 0)

    tc

    dip (R = 0)

    dip

    tc

    (R = 0)

    tc

    dip (R = 0)

    dip

    K(hr)

    10

    8

    6

    4

    2

    0

    -2

    -4

    -6

    -8

    tc

    (hr)anddi

    p

    (m3/s)

    10

    8

    6

    4

    2

    0

    -2

    -4

    -6

    -8

    tc

    (hr)anddip

    (m3/s)

    5 10 15 20 250

    5 10 15 20 250

    X

    X

    tc

    (R = 0)

    tc

    dip (R = 0)

    dip

    tc

    (R = 0)

    tc

    dip (R = 0)

    dip

    7

    6

    5

    4

    3

    2

    1

    0

    -1

    -2

    -3

    tc

    (hr)andd

    ip

    (m3/s)

    10

    8

    6

    4

    2

    0

    -2

    -4

    -6

    -8

    tc

    (hr)anddip

    (m3/s)

    0.60.50.40.30.20.10

    0.60.50.40.30.20.10

    Fig. 1 Influence of Kon the variation of the minimum negative

    discharge (dip ) and the corresponding critical time ( tc)

    computed with or without the incorporation ofR. (Qp =

    12.5 m3/s, tp = 6.69 hr, top: X= 0.2, bottom: X= 0.4)

    Fig. 2 Influence of Xon the variation of the minimum negative

    discharge (dip ) and the corresponding critical time (tc)

    computed with or without the incorporation ofR. (Qp =

    12.5 m3/s, tp = 6.69 hr, top: K= 5 hr, bottom: K= 15 hr)

  • 8/6/2019 Flood Routing in Long Channels

    8/13

    30 Journal of the Chinese Institute of Engineers, Vol. 28, No. 7 (2005)

    reducing the dip becomes. Increasing Xfrom 0.2 to

    0.4 also strengthens the effects ofKbut worsens the

    dip value at the same time. This indicates that for a

    long channel or a slowly moving flood flow, one

    should try to incorporate R as wave dissipation gets

    weaker. In brief, R contributes the most to the

    routings of outflow hydrographs with nearly un-dis-

    sipated flood peaks in a long channel.

    Figure 2 makes clear the influence ofXon flow

    dips. As mentioned previously, the increase ofXdoesincrease the absolute dip value of Eq. (25) and the

    increment grows rapidly with the increasing X. In

    addition, R has no impact on the reduction of dis-

    charge dips, as Kis less than about 5 hours. So, there

    is no need to introduceR for routing high-speed flood

    flows in short channels. At this point, CMM would

    be sufficient.

    Figure 3 shows that it is more necessary for a

    flood event with quick increasing inflow rates (i.e.,

    small tp) than a slow one (i.e., big tp) to adopt R for

    reducing dip values. Based on the point where two

    dip curves intersect, it can be deduced that the effec-tive range ofR on dip increases with the increasing

    K. Nevertheless, R becomes less important as tp ex-

    ceeds a certain value, indicating there is no necessity

    to include R for mild inflow hydrographs. All the

    results are consistent with the conclusions of models

    proposed in previous sections.

    VI. OPTIMIZATION PROCEDURES FOR NEW

    MODELS WITH CONSTANT R

    1. Modified Muskingum Algorithm (MMA)

    In this section, MMA is developed with the cali-

    bration ofK, Xand R by the least square optimiza-

    tion rather than the linear theory (i.e., by K= x/U,Eq. (6), and Eq. (8)), recognizing that most hydro-

    logical engineering problems are nonlinear. To

    achieve this and also for symbol saving, Eq. (4) is

    reorganized into a more compact form of second-or-der accuracy as follows:

    On + 1 = (4+ 6 1)On 2On 1

    + (1 2+ 2 2)In

    + (1 2 2 2)In + 1 (26)

    where

    = 1K(1 X)t

    = 1KXt

    = 1R

    with the denominator = 2K(1 X)t+ 2R + t2.The optimal , , and values for Eq. (26) will

    be determined by minimizing a target function de-

    fined below using the least square optimization:

    Tg = (On + 1 Qn + 1)2

    n = 1

    (27)

    where On + 1 and Qn + 1 denote, respectively, the pre-

    dicted and measured outflow discharges at the time

    level n + 1. Differentiation of the target function withrespect to , , and and by letting the results equalzero leads to three simultaneous algebraic equations

    in the matrix form of:

    A2 BA CAAB B2 CBAC BC C2

    =

    F1F2F3

    (28)

    whereA = 4On 2In 2In + 1,B = 2In 2In + 1, C= 6On

    2On 1 2In 2In + 1,D = OnInIn + 1, F1 =A(Qn + 1

    +D), F2 =B(Qn + 1

    +D), F3 = C(Qn + 1

    +D). Symbol denotes the summation from n = 1 toN. Ndenotes the

    tp (hr)

    tp (hr)

    tc (R = 0)

    tc

    dip (R = 0)

    dip

    tc (R = 0)

    tc

    dip (R = 0)

    dip

    8

    6

    4

    2

    0

    -2

    -4

    -6

    tc

    (hr)anddip(m

    3/s)

    10

    8

    6

    4

    2

    0

    -2

    -4

    -6

    tc

    (hr)anddip(m3/s)

    1210864

    20

    12

    108

    6420

    Fig. 3 Influence of tp on the variation of the minimum negative

    discharge (dip ) and the corresponding critical time ( tc)

    computed with or without the incorporation ofR. (Qp =12.5 m3/s, X= 0.3, top: K= 5 hr, bottom: K= 10 hr)

  • 8/6/2019 Flood Routing in Long Channels

    9/13

    W. Chung and Y. L. Kang: Flood Routing in Long Channels: Alleviation of Inconsistency and Discharge Dip 31

    total number of time steps. K, X, and R are readily

    obtained once , , and are given. Flood routingthen proceeds by Eq. (26).

    2. Modified Gills Algorithm (MGA)

    In what follows, MGA will be developed by in-

    corporating the CMM with R and Gills relative

    storage. The storage S in Eq. (2) can be identified as

    either the absolute or the relative storage of a river

    reach, without affecting mass conservation. To make

    it consistent, let the S in Eq. (1) denote the absolute

    storage but in Eq. (2) it is regarded as relative storage.

    Since absolute storage is rarely obtained, Gill

    (1977) proposed the modified expression S = Sr where Sr represents the relative storage and theinitial storage. With the initial storage, the total num-

    ber of involved calibrated parameters is increasedwithout yielding the unwanted roots of the general

    sys tem mode l recommended by Chow and

    Kulandaiswamy (1971). Following Gills modifica-

    tion and the least square technique proposed by

    Aldama (1990) the sum of square errors between pre-

    dicted and measured relative storages becomes:

    Tg = (+ 1In + 2O

    n + 3On

    Srn)2

    j = 1

    N

    (29)

    where O= dO/dt, 1 = KX, 2 = K(1 X), 3 =R andsr

    n = the measured relative storage at time level n. Dif-

    ferentiation of Eq. (29) with respect to and lettingthe result equal zero yield the explicit expression for

    . Substituting this expression into the minimized

    Tg for j, j = 1 ~ 3, gives the following matrix equa-tion in tensor notation:

    Mijj = fi (30)

    where:

    M11 =InIn 1NImIn

    M12 =InOn 1NImOn

    M13 =InO

    n 1

    NImO

    n

    M22 =OnOn 1NOmOn

    M23 =O nOn

    1NO

    mOn

    M33 =OnO

    n 1

    NOmO

    n

    and M21 = M12, M31 = M13, M32 = M23. The forcingfunctionsfi read:

    f1 =InSrn 1

    NImSr

    n

    f2 =OnSrn 1NOmSr

    n

    f3 =On

    Srn

    1NOm

    Srn

    In the above, Sr at the time level n is computed ac-

    cording to the discrete form of Eq. (2) with Sr= 0 for

    n = 1. Once variables , j,j = 1 ~ 3, are computed byEq. (30) using Cramers rule, K, X, and R can be de-

    termined and Eq. (26) is again employed to compute

    the outflow rate at each time level. One of the most

    attractive things about MMA or MGA is that they are

    formulated without breaking the beauty/simplicity of

    Muskingum-type models and all parameters are cali-

    brated out explicitly.

    VII. MODEL COMPARISONS

    This section compares the traditional (i.e., CMM

    & GILL) and modified (MMA & MGA) routing mod-

    els via numerical experiments. The equation govern-

    ing the experiment is the parabolicized Saint Venant

    equation that neglects inertia terms but implicitly in-

    cludes pressure forces. For wide rectangular

    channels, the parabolicized equation has a form of

    advection-diffusion properties and reads:

    Q

    t+ 5

    3

    Q

    Bh

    Q

    x=

    Q

    2BSf

    2Q

    x2(31)

    Its solution will be sought by solving simultaneously

    with the continuity equation below:

    ht

    = 1B

    Q

    x(32)

    where Q = discharge, B = channel width, h = water

    depth, and Sf= the energy gradient based on Mannings

    formula. The observed outflow hydrograph required

    for parameter estimations is borrowed from the nu-

    merical solution of Eqs. (31)-(32) at x = 0.8L so as to

    reduce the influence of the downstream boundary

    condition and satisfy the basic assumption that routingsare performed for a semi-infinite river reach. The

    numerical solution serves as a benchmark solution for

    model comparisons aside from parameter estimations.

    Solving Eq. (31) requires the prescription of up- and

    down-stream boundary conditions of which the former

    is introduced as a Gaussian distributed inflow

    hydrograph and the latter is stated as a fixed value of

    normal water depth corresponding to a given base flow

    rate.

    Figure 4 shows the Gaussian inflow hydrograph

    and the four predicted outflow hydrographs routed

    by CMM, GILL, MMA, and MGA. Concerning thetest of model forecasting ability and robustness, the

  • 8/6/2019 Flood Routing in Long Channels

    10/13

    32 Journal of the Chinese Institute of Engineers, Vol. 28, No. 7 (2005)

    numerical experiment is performed in a 10 m wide,

    90 km long rectangular channel with a bottom slope

    of 0.0004 and a Mannings roughness coefficient of

    0.029. The base flow rate is assumed to be 17.5 m3/s

    with which normal velocity and water depth are

    computed. Since the channel length for this particu-

    lar flow case is quite long, the numerically solved

    outflow peak discharge will be much lower than that

    of the inflow, in spite of the fact that the channel bed

    is only a little rougher. This condition satisfies fairly

    well the best criteria for employing R.

    As one can see clearly from Fig. 4, CMM com-

    pletely fails due to its significant model inconsistency,

    regardless of the nonobvious discharge dip. The large

    under-estimated outflow peak discharge and time to

    peak (phase error) reveal that CMM is not adequate

    for flood routing in long channels. When Gills con-

    cept of relative storage is adopted, the outflowhydrograph labeled GILL is predicted with better

    precision, though the dip value becomes greater. The

    predicted outflow peak discharge is now quite close

    to the benchmark solution in spite of the evident phase

    error. For the case with R but without using Gills

    relative storage (i.e., MMA), the dip in the routed

    outflow hydrograph is much reduced as expected.

    Also, the predicted outflow peak discharge is just

    slightly higher than the benchmark solution. Though

    there exists the problem of anti-phase-error (i.e., the

    time to peak is over-estimated), what really discour-

    aged us was the oscillatory tail attached to the pre-dicted outflow hydrograph since its amplitude was

    too large to be ignored. The tail, though damping

    out rapidly, resulted in a large root-mean-square er-

    ror defined as Erms =1

    N(Oj Qj)

    2j = 1

    N

    , hence re-

    ducing the applicability of MMA. The reason why

    MMA evolves such an apparent oscillatory tail is be-

    cause it strongly meets the criteria for adopting Eq.(14) the value is as low as 61.4 hr2 being thehighest of the four models in Table 1. In contrast,

    GILL is good for the prediction of recession curves

    while MMA is for the rising limb. Consequently,

    MGA has been adopted for predicting the outflow

    hydrograph. It is found that the prediction is so ac-

    curate that the hydrograph almost has coincided with

    the target one. Indeed, the phase error has become

    trivial, thus the outflow peak discharge is deemed

    adequate, and the dip value shows lower than that of

    the GILL curve. By realizing the least square opti-

    mization minimizes the overall instead of the localerror of a predicted outflow hydrograph, MGA may

    not remedy its dip to the minimum. Aside from the

    smallest root-mean-square error, the most pleasant

    result about MGA is that it does not evolve the oscil-

    latory tail. This is because its absolute value justexceeds 11 hr2 and itsR value is 2.1 times lower than

    that of MMA; this fact serves to fasten the exponen-

    tially decaying rate of the sinusoidal wave and mak-

    ing the oscillatory tail invisible. Indeed, the newly

    introducedR, after linked to the relative storage, re-

    duces the dip of GILL, alleviates model inconsistency,

    and shrinks the range of phase errors to the minimum

    among the four models. As a result, MGA offers thebest predictability for the flood routing in long

    channels. The same conclusion can be drawn from

    Table 1 by comparing the corresponding parameter

    values.

    Table 1 lists the root-mean-square errors of all

    models and the parameter values used for routing the

    outflow hydrographs in Fig. 4. In the last two

    columns, tc and dip values are obtained from Eqs. (20)

    ~(22). The associative values are 0.08 and 0.02 forEqs. (21) and (22), respectively. It can be seen eas-

    ily that the K and X values of CMM differ greatly

    from those of other models. The main reason is surelydue to the super long channel reach and the adopted

    t(hr)

    Inflor

    Outflow

    CMM

    GILL

    MMA

    MGA

    220

    170

    120

    70

    20

    -30

    -80

    Q(

    m3/s)

    10 20 30 40 50 600

    Fig. 4 Comparisons of the routed outflow hydrographs by CMM,

    GILL, MMA, and MGA for a 90 km long channel with n =

    0.029 and S0 = 0.0004

    Table 1 The calibrated parameters and root-mean-

    square errors of all models in Fig. 4

    K X R E rms tc dip

    (hr) (hr2) (m3/s) (hr) (m3/s)

    CMM 127 0.085 26.6 6.38 -9.01GILL 7.43 0.392 22.0 5.57 -47.6

    MMA 3.87 0.472 16.4 21.2 4.91 -7.86

    MGA 7.23 0.379 7.79 9.71 5.51 -34.9

  • 8/6/2019 Flood Routing in Long Channels

    11/13

    W. Chung and Y. L. Kang: Flood Routing in Long Channels: Alleviation of Inconsistency and Discharge Dip 33

    storage function. Nevertheless, the least square op-

    timization is the key procedure leading to such a large

    Kvalue for CMM, since in this case K is no longer

    determined by x/U, but is deduced from the statisti-cal point of view. Owing to the excessively large K

    value (about an order of magnitude greater than the

    others) of CMM,R becomes quite effective and plays

    a vital role in the routing procedures of MMA and

    MGA. Furthermore, the KXvalue of CMM, given in

    Table 1, is 3.7 times higher than those of other models,

    implying the largest tc outcome as a result. The

    smallest Xvalue in Table 1 also implies that CMM

    shows the most significant inconsistency with the

    correct A-D expression as indicated by Eqs. (9) and

    (10). Since the predicted tc and dip values are so close

    to those in Fig. 4 (as can be clearly seen), models

    (20)(21)(22) are successfully developed regardless of

    their simplicity. Finally, it will not help too much toadopt R or relative storage alone since GILL and

    MMA produce roughly the same root-mean-square

    errors. Only by using both can MGA obtain the small-

    est error. For a shorter channel of 30 km, Fig. 5(a)-

    (c) shows about the same results: MGA is superior to

    all other variants of CMM. At this point, the postu-

    lation by Laurenson (1959) may be further addressed:

    flood routings in fairly long channels can be improved

    greatly by modifying the conventional storage func-

    tion to include RdO/dt. As for the threshold length

    used to classify a channel as long or short, it is

    suggested to be 30 km in regard to the present analy-

    sis and most of the routing length successfully

    adopted for CMM in the past. The threshold value is

    chosen without rigorous theoretical background but

    based on our best educated judgment.

    VIII. CONCLUSIONS

    In order to solve the inconsistency problem, this

    paper introduces a time constantR to the conventional

    Muskingum storage function. The newly added term

    improves the inaccuracy from Qxxx to Qxxxx, and hence,

    reduces the phase error and refines outflow peak

    discharge. The inaccuracy is found to be proportionalto x4 or to D3/U2 by the linear theory of Taylor se-ries expansion. As a rule of thumb,R must be dropped

    out from the Muskingum storage function if it is cali-

    brated as negative.

    Based on a small time approximation, the oc-

    currence time ofdip (i.e., tc) and its magnitude can

    be easily determined according to Eqs. (20)-(22).

    Accordingly, dip occurs before an inflow hydrograph

    reaches its peak. The higher the peak value is, the

    larger the dip becomes. Other parameters such as K,

    X, and tp also have positive effects on discharge dips.

    Increasing R or decreasing the growth rate of aninflow hydrograph, however, decreases the dip

    A careful study of the convolution integral re-

    veals that flow dips result from the negativef(t) and

    predicted outflow hydrograph shapes relate closely

    to g(t) which governs as well the phase error. Also,

    parameters affecting the normalized dip (dip/Qp) are

    only tp, K, and X, of which Xvaries in an opposite

    way with model inconsistency as Kdoes in a positive

    way. Thus, the best timing for usingR is to route the

    outflow hydrographs in a long but a slightly rough

    channel that carries rapidly rising inflow dischargeswith high peak values.

    Fig. 5 Performance of traditional and modified routing models

    for a 30 km long channel

    t(hr)

    (a) n = 0.029, S0 = 0.004

    Inflor

    Outflow

    CMM

    GILLMMA

    MGA

    220

    170

    120

    70

    20

    -30

    Q(

    m3/

    s)

    5 10 15 20 25 300

    t(hr)

    (b) n = 0.08, S0 = 0.0004

    Inflor

    Outflow

    CMMGILL

    MMA

    MGA

    220

    170

    120

    70

    20

    -30

    Q(

    m3/s)

    5 10 15 20 25 300

    t(hr)

    (c) n = 0.08, S0 = 0.0001

    Inflor

    OutflowCMM

    GILL

    MMA

    MGA

    220

    170

    120

    70

    20

    -30

    Q(

    m3/s)

    5 10 15 20 25 300

  • 8/6/2019 Flood Routing in Long Channels

    12/13

    34 Journal of the Chinese Institute of Engineers, Vol. 28, No. 7 (2005)

    Adopting Gills concept of relative storage will

    increase the dip value by trading off a more precise

    portrayal of outflow peak discharge. CMM and GILL

    tend to underestimate the time to peak discharge but

    MMA overestimates it, primarily due to the nature of

    the response function g(t). As

  • 8/6/2019 Flood Routing in Long Channels

    13/13

    W. Chung and Y. L. Kang: Flood Routing in Long Channels: Alleviation of Inconsistency and Discharge Dip 35

    Koussis, A. D., 1983 Unified Theory for Flood and

    Pollution Routing, Journal of Hydraulic

    Engineering, ASCE, Vol. 109, No. 12, pp. 1652-

    1663.

    Laurenson, E. M., 1959, Storage Analysis and Flood

    Routing in Long River Reaches,Journal of Geo-

    physical Research, Vol. 64, pp. 2423-2431.

    Nash, J. E., 1959, A Note on the Muskingum Flood-

    Routing Method,Journal of Geophysical Research,

    Vol. 64, pp. 1053-1056.

    Ngan, P., and Rusell, S. O., 1986, Example of Flow

    Forecasting with Kalman Filter, Journal of Hy-

    draulic Engineering, ASCE, Vol. 112, No. 9, pp.

    818-832.

    ODonnell, T., 1985, A Direct Three-Parameter

    Muskingum Procedure Incorporating Lateral

    Inflow,Hydrological Science Journal, Vol. 30,

    No. 4, pp. 479-496.Ponce, V. M., and Theurer, F. D., 1982, Accuracy

    Criteria in Diffusion Routing,Journal of Hydraulic

    Division, ASCE, Vol. 108, HY6, pp. 747-757.

    Singh, V. P., and McCann, R. C., 1980, Some Note

    on Muskingum Method on Flood Routing,Jour-

    nal of Hydrology, Vol. 48, pp. 343-361.

    Smith, G. D., 1985,Numerical Solution of Partial Dif-

    ferential Equations: Finite Difference Methods,

    3rd edition, Oxford University Press, New York,

    USA.

    Van Geer, F. C., Te Stroet, C. B. M., and Zhou, Y.,

    1991, Using Kalman Filtering to Improve and

    Quantify the Uncertainty of Numerical Ground-

    water Simulations 1: the Role of System Noise

    and Its Calibration, Water Resources Research,

    Vol. 27, No. 8, pp. 1987-1994.

    Weinmann, P. E., and Laurenson, E. M., 1979, Ap-

    proximate Flood Routing Methods: A Review,

    Journal of Hydraulic Division , Vol. 105, HY12,

    pp. 1521-1536.

    Wood, E. F., and Szollosi-Nagy, A., 1978, An Adaptive

    Algorithm for Analyzing Storm-Term Structural

    and Parameter Changes in Hydraulic Prediction

    Models, Water Resources Research, Vol. 14, No.

    4, pp. 577-581.

    Manuscript Received: May 11, 2004

    Revision Received: Apr. 10, 2005

    and Accepted: May 24, 2005

    APPENDIX

    Of all the approximate methods, as addressed

    by Weinmann and Laurenson (1979), the one

    typified by the Koussis model (Koussis, 1978;

    Koussis, 1983; Koussis and Snenz, 1983) is the most

    general. Koussis model assumes basically the lin-

    ear variation of discharge within any time interval.

    The Laplace transformation yields this model of the

    form:

    Qj + 1n + 1 = Qj

    n + 1(1 + 1

    Cr) + Qj

    n( +1

    Cr)

    + Qj + 1n () (A1)

    where

    =2 + Pg PgCr2 + Pg + PgCr

    (A2)

    To derive the modified equation of Eq. (A1), the

    same procedure as described in section II is used toyield:

    Qt + UQx = Qxx [(1 + 1

    )U2t

    2 Ux

    2]

    numerican dispersion

    + ET (A3)

    where the truncation errorET reads:

    ET

    = Qxxx[(2 + )U3t2

    6(1 )+ Ux

    2

    6 U

    2xt2(1 )

    +DUt

    1 ]

    Qxxxx[DUxt2(1 )

    (1 + )DU2t2

    2(1 )+ D

    2t2

    ]

    Qxxxxx[UD 2t2

    2(1 )] Qxxxxxx(

    D3t2

    6)

    + O(x3, x2t, xt2, t3) (A4)

    Equating the numerical dispersion term in the bracket

    of Eq. (A3) to the physical dispersion D yields the

    expression for as given by Eq. (A2). Further analy-sis of the extreme case that x 0 and t 0 leads

    the truncation error given by Eq. (A4) to the follow-ing nonzero result

    ET = D2

    UQxxx (A5)

    Since has a limit value of 1 and t/(1 ) a limitvalue ofD/U2. It is apparent then that the Koussiss

    model of differential form is not identical to the ex-

    pected A-D equation, that is, model inconsistency

    occurs.