Mwakalila Feyen Hydro Logical Processes

  • Upload
    oscar

  • View
    221

  • Download
    0

Embed Size (px)

Citation preview

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    1/15

    HYDROLOGICAL PROCESSESHydrol. Process. 15, 22812295 (2001)DOI: 10.1002/hyp.257

    Application of a data-based mechanistic modelling (DBM)approach for predicting runoff generation in semi-arid

    regions

    S. Mwakalila,1* P. Campling,1 J. Feyen,1 G. Wyseure2 and K. Beven3

    1 Institute for Land and Water Management (ILWM), K.U.Leuven, Belgium2 Faculty of Agricultural and Applied Biological Sciences, K.U.Leuven, Belgium

    3 Institute of Environmental and Natural Sciences, Lancaster University, UK

    Abstract:

    This paper addresses the application of a data-based mechanistic (DBM) modelling approach using transfer functionmodels (TFMs) with non-linear rainfall filtering to predict runoff generation from a semi-arid catchment (795 km 2)in Tanzania. With DBM modelling, time series of rainfall and streamflow were allowed to suggest an appropriatemodel structure compatible with the data available. The model structures were evaluated by looking at how wellthe model fitted the data, and how well the parameters of the model were estimated. The results indicated that aparallel model structure is appropriate with a proportion of the runoff being routed through a fast flow pathwayand the remainder through a slow flow pathway. Finally, the study employed a Generalized Likelihood UncertaintyEstimation (GLUE) methodology to evaluate the parameter sensitivity and predictive uncertainty based on the feasibleparameter ranges chosen from the initial analysis of recession curves and calibration of the TFM. Results showedthat parameters that control the slow flow pathway are relatively more sensitive than those that control the fast flowpathway of the hydrograph. Within the GLUE framework, it was found that multiple acceptable parameter sets give

    a range of predictions. This was found to be an advantage, since it allows the possibility of assessing the uncertaintyin predictions as conditioned on the calibration data and then using that uncertainty as part of the decision-makingprocess arising from any rainfall-runoff modelling project. Copyright 2001 John Wiley & Sons, Ltd.

    KEY WORDS data-based mechanistic modelling approach; transfer function models; Generalized Likelihood UncertaintyEstimation; parameter sensitivity and predictive uncertainty

    INTRODUCTION

    Rainfall-runoff models are important tools for water resource planning, development and management.

    However, limitations of both model structures and the data available on parameter values, initial conditions

    and boundary conditions, generally make it difficult to apply any hydrological model without some form of

    calibration or conditioning of model parameter sets.

    The data-based mechanistic (DBM) modelling addressed in this study is based on the concepts of lettingthe data suggest an appropriate model structure compatible with the inputoutput data available, but then

    evaluating the resulting model to see if there is a mechanistic interpretation that might lead to insights

    not otherwise gained from modelling based on theoretical reasoning. These concepts have been applied

    successfully in humid climates, e.g. see Young and Beven (1994), Young (1998), and Beven (2001). Similarly,

    in the application of an IHACRES model the data are also allowed to suggest the form of the transfer function

    used for flow routing rather than specifying a fixed structure beforehand (see, Jakeman et al., 1990, 1993;

    Jakeman and Hornberger, 1993; Schreider and Jakeman, 1996; Post and Jakeman, 1996; Sefton and Howarth).

    * Correspondence to: Dr. S. Mwakalila, Institute for Land and Water Management Katholieke Universiteit Leuven, Vital Decosterstraat 102,B-3000 Leuven, Belgium. E-mail: [email protected]

    Received 28 June 2000

    Copyright 2001 John Wiley & Sons, Ltd. Accepted 16 January 2001

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    2/15

    2282 S. MWAKALILA ET AL.

    The results generally show that the parallel transfer function is a suitable structure for rainfall-runoff simulation

    at the catchment scale, with one fast flow pathway and one slower flow pathway. The fast flow pathway

    provides the major part of the predicted storm hydrograph, and the slower pathway the major part of the

    recession discharge between storm periods.

    This study focuses on extending the application of these concepts to semi-arid catchments and also to assess

    the parameter sensitivity and discharge prediction limits of the identified model structure.

    Compared with humid climate regions, in semi-arid regions that cover a large part of the world there are

    a lot of problems in predicting runoff generation. In general, the two interrelated underpinning problems

    in rainfall-runoff modelling in such regions are due to scarcity of data and data uncertainties. The success

    of any catchment rainfall-runoff model depends critically on the data available to set it up and drive it.

    Nearly all rainfall-runoff models require data on some combination of rainfall, discharge, evapotranspiration

    and catchment morphology. The basic sources of error and uncertainty in each of these data types and

    methods to account for these uncertainties are discussed in detail in Melching (1995). He noted that the

    primary uncertainties in using point-rainfall measurements to describe the true rainfall input for a catchmentare: depth measurement error in the gauge due to equipment malfunction; representativeness of ground-

    level precipitation at the gauging point; gauge location; the areal-mean rainfall; gauge-network areal-mean

    rainfall versus true areal-mean rainfall; rainfall spatial variability; rainfall temporal variability; lack of

    synchronization between time clocks for rain and stream gauges; and lack of synchronization between time

    clocks for the various rain gauges in the catchment. The availability of discharge data is important for

    the model calibration process. Discharge data are, however, generally available at only a small number of

    sites in semi-arid regions. It is also an integrated measure, in that the measured hydrograph will reflect

    all the complexity of flow processes occurring in the catchment. However, the discharge data available

    are generally point data and may not be error free. If any model is calibrated using data that are in

    error, the effective parameter values will be affected and the predictions for other periods that depend on

    the calibrated parameter values will be affected. This will be an additional source of uncertainty in the

    modelling.

    The main objective of this paper therefore, is to address the application of a DBM modelling approach

    to predict runoff generation from a semi-arid catchment. Furthermore, the paper employs a Generalised

    Likelihood Uncertainty Estimation (GLUE) methodology using Monte Carlo simulation (MCS) as key

    component to assess the model parameter sensitivity and predictive uncertainty. This type of technique has

    been used on several occasions (e.g. Beven and Binley, 1992; Beven, 1993; Romanowicz et al., 1994; Freer

    et al., 1997; Franks et al., 1998; Cameron et al., 1999), but not in the context of semi-arid regions.

    METHODOLOGY

    The methodology in this study has three distinct stages. At the first stage a DBM approach was used to suggest

    an appropriate transfer function model (TFM) structure compatible with the available set of daily rainfall-

    runoff data. At this stage it is the idea that the data should be allowed to indicate which model structure is

    most appropriate rather than specifying a structure beforehand. Therefore model structures were evaluated by

    looking at how well the model fitted the data, and how well the parameters of the model were estimated.

    The second stage involves the separation of the slow flow component from the total discharge in order to get

    a hydrologically more realistic model structure. At the third stage the GLUE methodology approach using

    MCS as a key component was employed for performance evaluation of the identified model structure for the

    purpose of parameter sensitivity analysis and uncertain estimation.

    Data used

    The data used in this study were rainfall and streamflow records available on a daily basis for 197080

    period for the Great Ruaha catchment at Salimwani gauging station (IKA 8A) in Tanzania. The study area

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    3/15

    PREDICTING RUNOFF GENERATION 2283

    covers an area of about 795 km2. The gauging station is located at 34 07 E longitude and 8 54 S latitude.

    It is a sub-basin of the Great Ruaha River Basin, which is draining an area of about 68 000 km2. The annual

    mean temperature varies from about 18 C at the higher altitudes to about 28 C at the lower and drier part

    of the basin. The rainfall regime in the basin is typical of the unimodal type, with a single rainy season from

    November through May and no rainfall during the rest of the year. As the main process causing rainfall is

    convection, the rainfall is highly localized and with the incoming wind direction and the topography being

    the main determining factors of spatial patterns. The mean annual rainfall ranges from 800 mm in the lower

    part of the catchment to 1000 mm in the highlands. Owing to the generally high potential evaporation rates

    (14001800 mm year1) compared with the yearly rainfall, most of the catchment has an annual rainfall

    deficit. The runoff patterns in the basin correspond closely to the unimodal rainfall patterns prevailing in

    the whole catchment. The streams start rising in NovemberDecember, experience a maximum flow in

    MarchApril and have their recession period from May to October. As rainfed cultivation is very unreliable

    in the area due to the low and unreliable rainfall, farmers in the area depend very much on irrigation. The

    major part of irrigation takes place downstream of the gauging station. In all types of irrigation, rice is thedominant crop, which is grown during the rainy season (DANIDA, 1995).

    Initial evaluation of general linear TFM

    In this stage of the methodology, the general linear TFM was evaluated to identify an appropriate

    model structure. Within the DBM approach a non-linear rainfall filter function [Equation (1)] was used to

    produce an effective rainfall, which was then related to discharge using a generalized linear transfer function

    [Equation (2)].

    Ut / RtQtn 1

    where Ut is the effective rainfall, Rt is the rainfall input at time t, Qt is the discharge; parameter n controls

    the non-linearity of fast runoff generation (if n D 0, the rainfall filter is linear as for subsurface recharge).Here, the current discharge was used as an index of the contribution of antecedent moisture status of the

    catchment on runoff generation. Therefore, the interpretation of RtQtn can be referred to as a contributing

    area function.

    The general form of the discrete time linear transfer function that forms the basis of the TFM package can

    be presented as follows:

    Qt Db0 C b1z

    1 C C bMzM

    1 a1z1 a2z2 C C aNzNUt 2

    where Qt is a simulated discharge, Ut is the effective rainfall, z is a backward difference operator, a and b

    are coefficients treated as parameters, such that parameter a is related to the mean residence time of fast flow

    pathway and parameter b is a gain parameter that scale the differences in total volumes of input and output;

    M and N represent the total number ofb and a parameters respectively, and is the time delay. The derivationof Equation (2) can be found in Beven (2001). Parameter estimation is carried out using a simplified refined

    instrumental variable (SRIV) technique which has been shown to be robust to data errors. For more detail

    see Young (1984), Young and Beven (1994) and Beven (2001).

    The higher the model order (i.e. the greater the number of a and b coefficients), the larger the standard

    errors of the estimated parameter values will tend to be. Very large standard errors on the coefficients are an

    indication that a model is over-parameterized. One way of checking for over-parameterization is by looking

    at how well the coefficients of the model are estimated. The aim in this study, is to find a model structure

    that gives a good fit but is parsimonious in having a small number of coefficients.

    Therefore, using a daily rainfall-runoff data set an appropriate TFM structure was identified, by finding the

    best values of (N, M, ) and the corresponding coefficients.

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    4/15

    2284 S. MWAKALILA ET AL.

    Performance evaluation of TFM structures

    Underlying the model identification step is the idea that the data should be allowed to indicate which model

    structure is most appropriate, rather than specifying a structure beforehand. There may be many models giving

    an acceptable fit to the data. In this study, therefore, model structures were evaluated in terms of the following

    criteria: NashSutcliffe efficiency R2, which indicates how well the model fits the data; Young Information

    Criterion (YIC), which combines a measure of goodness of fit with a measure of how well the parameters

    are estimated.

    R2, introduced as the efficiency measure for rainfall-runoff modelling by Nash and Sutcliffe (1970), is

    statistically known as the coefficient of determination indicating the proportion of the variance in the data

    explained by the model, and can be defined as:

    R2 D 1 e

    2

    o2

    3

    where o2 is the variance of the observed output data calculated over all time steps used in fitting the model

    and e2 is the variance of the differences between observed and simulated outputs at each time step. As the

    model fit improves, the value of R2 will approach unity. If the model is no better than fitting the mean of the

    observed outputs e2 D o

    2, the value of R2 will be zero or less.

    The YIC is defined as:

    YIC D ln

    e

    2

    o2

    C ln

    1

    k

    k

    e2 Pii

    i

    4

    where i, i D 1 . . . k , are the model coefficients, and Pii is the ith diagonal element of a scaled parameter

    covariance matrix. From its definition, the first term of YIC is simply a relative logarithmic measure of

    how well the model explains the data: the smaller the model residuals the more negative the term becomes.

    The second logarithmic term, on the other hand, tends to become large (less negative) when the model is

    over-parameterized and the parameter estimates are poorly defined. Consequently, as whole, the criterionattempts to identify a model that explains the data well but with the minimum of statistically well-defined

    (low-variance) parameters.

    Separating a slow flow component from total discharge

    Since it was learnt that the general linear TFM is underestimating the long-term recession flows, therefore,

    the slow flow component was separated off from total discharge in order to get a hydrologically more realistic

    model structure. Therefore, to allow for non-linearity the linear storageoutflow relationship was generalized

    by an exponential reservoir model as presented by Beven et al. (1995):

    Qb D Qoes/m 5

    where Qb is the outflow from the saturated zone considered as baseflow, Qo is the discharge when recessionof streamflow commences, s is the local storage deficit and m is a model parameter considered as the scaling

    depth of the effective catchment soil profile.

    Solution of Equation (5) for a pure recession in which inputs are assumed to be zero shows that discharge

    has an inverse or first-order hyperbolic relationship to time as:

    1

    QbD

    1

    QoC

    t

    m6

    A plot of 1/Qt against time t should plot as straight line with a slope of 1/m, where Qt is the recession flow

    at time t. Therefore, Equation (5) was used for separating the slow component based on the following major

    steps.

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    5/15

    PREDICTING RUNOFF GENERATION 2285

    (1) The catchment average storage deficit before each time step was updated by subtracting the unsaturated

    zone recharge and adding the baseflow (slow flow component) calculated for the previous time step, thus:

    StC1 D St C Qo eSt/m cRt 7

    where Rt is the observed rainfall at time t and parameter c represents the proportion of rainfall that

    percolates as recharge to the groundwater.

    (2) If an initial discharge QtD0 is known and assumed to be only the result of drainage from the saturated

    zone, Equation (5) can be inverted to give a value for S at time t D 0 as:

    S D m

    ln

    QtD0

    Qo

    8

    Equations (7) and (8), therefore, form the basis for separating the baseflow from the total discharge during

    storm events. Then, given the time series of discharge Qt and rainfall Rt at time t, the time series of directrunoff (fast flow component) Qd was calculated as:

    Qd D Qt Qo eSt/m 9

    Performance evaluation of identified model structure

    The performance evaluation of the identified model structure was based on the framework of the GLUE

    methodology for the purpose of parameter sensitivity analysis and predictive uncertainty estimation. The main

    objective here was to examine those parameters to which the model simulation results were most sensitive

    and to assess the probability of simulated discharge being within a certain prediction interval. The prediction

    interval was defined by 5 and 95% limits (i.e. a 90% probability that the predicted discharge value lies within

    the interval).

    The main procedure was based on the following major steps:

    (1) determining a feasible parameter range from initial analysis of recession curves and calibration of TFM;

    (2) using a MCS method to choose parameter values from uniform distributions spanning specified ranges of

    each parameter;

    (3) using a likelihood value to divide acceptable simulations from unacceptable simulations;

    (4) normalizing likelihood values for the parameter sets that yield acceptable simulations, such that the sum

    of the normalized likelihood values equals unity;

    (5) using the same generated parameter sets and associated likelihood value and model outputs to assess

    parameter sensitivity and predictive uncertainty.

    The modelling efficiency R2 in Equation (3) was used as a likelihood measure for determining acceptable

    simulations. With the MCS technique, 5000 runs of the model were made with different randomly chosenparameter sets. The feasible range of parameters to be sampled was based on an initial evaluation of the

    TFM and recession flow analysis. Those simulations with R2 > 0 5 were retained, those that did not meet

    this performance criterion were deleted. Finally, 4000 simulations were retained as behavioural. In this study,

    therefore, R2 D 0 5 was considered as a behavioural threshold indicating the proportion of the variance in

    the data explained by the model. This value was considered as a minimum value to accept the simulations

    (Beven, 2001).

    In the GLUE procedures a prior likelihood estimate can be updated with a new (posterior) likelihood

    measure calculated from the prediction of a new set of observations. The idea here is to check if the model,

    which performs well during the conditioning or calibration period, will continue to perform well in other

    periods. If this is not the case then its combined likelihood measure will be reduced. The Bayes equation

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    6/15

    2286 S. MWAKALILA ET AL.

    was used here to combine likelihood measures, the type of combination using the Bayes equation may be

    expressed in the form:

    Lp ijY DLo iL ijY

    C10

    where Loi is the prior likelihood of the ith parameter set, LijY is the likelihood calculated for the

    current evaluation given the set of observations Y, LpijY is the posterior likelihood and C is a scaling

    constant to ensure that the cumulative posterior likelihood is unity. The special characteristic of the Bayes

    equation is its multiplicative operation, if any evaluation results in a zero likelihood, the posterior likelihood

    will be zero regardless of how well the model has performed previously. This may be considered as an

    important way of rejecting non-behavioural models: it may cause a re-evaluation of the data for that period

    or variable; it may lead to the rejection of all models.

    The GLUE methodology used here is one technique for conditioning of model parameter sets in the face

    of rejecting the idea that there is an optimum parameter set in favour of the concept of the equifinalityof different models and parameter sets. Parameter interactions and non-linearity in the model responses are

    handled implicitly. Errors in input data and the observation data are also handled implicitly. Thus the likelihood

    measure reflects the ability of a particular model to predict a particular series of observations (which may

    not be error free) given a particular set of inputs (which may not be error free). There is thus an implicit

    assumption that, in prediction, error structures will be similar in some broad sense to those in the evaluation

    period.

    RESULTS AND DISCUSSIONS

    Initially, models were calibrated for five periods of 2 years in length within the decade (197080). The second

    sub-period, 197273, was found to give the best calibrations. Therefore, the time series of daily rainfall and

    discharge (for the 197273 observation period) were allowed to suggest an appropriate model structure withinthe general transfer function modelling framework. The results in Table I indicate that a first-order (1, 1, 0)

    transfer function is an appropriate model structure with parameter n in Equation (1) ranging between 05 and

    Table I. DBM model results of total observed rainfall and discharge in 197273 period

    Model ordera R2 YIC a parameters b parameters

    1, 1, 0 079 873 a1: 09592(00017) b0: 00202(00008)

    1, 1, 1 078 843 a1: 09542(00022) b0: 00223(00010)

    1, 1, 2 077 841 a1: 09561(00021) b0: 00213(00009)

    1, 2, 2 078 602 a1: 09610(00020) b0: 00473(00043)b1: 00281(00044)

    1, 2, 1 079 473 a1: 09554(00026) b0: 00350(00042)b1: 00133(00044)

    1, 2, 0 078 383 a1: 09335(00024) b0: 00096(00042)b1: 00129(00044)

    2, 2, 0 079 261 a1: 06958(02429) b0: 00080(00043)a2: 02471(02316) b1: 00199(00061)

    2, 1, 1 079 170 a1: 08573(01549) b0: 00239(00035)a2: 00937(01482)

    a Model order D N, M, (N is the number of a parameters, M is the number of b parameters, is adelay).

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    7/15

    PREDICTING RUNOFF GENERATION 2287

    10.00

    0.00

    10.00

    0 100 200 300 400 500 600 700 800

    Time (days)

    Error

    0.00

    4.00

    2.00

    6.00

    10.00

    12.00

    0 200 400 600 800

    Discharge(mm/da

    y)

    Qo

    Qs

    Time (days)

    0

    20

    40

    60

    80

    0 200 400 600 800

    Rainfall(mm)

    Time (days)

    Figure 1. Rainfall, observed discharge Qo and simulated discharge Qs and error

    10 for a non-linear rainfall filter function. However, this model structure is underestimating the long-term

    recession flows, as depicted in Figure 1.

    As can be noted from Table I, it seems a second-order transfer function model (2, 2, 0 and 2, 1, 1) does

    better in terms of R2 and YIC, presumably because errors are dominated by peaks during the wet season

    and are relatively small in the dry period recession (see Figure 1). This indicates that, in order to get a

    hydrologically more realistic model, the slow flow component should be separated off first before evaluating

    the TFM structures.

    The DBM model results after separating the slow flow component are presented in Figure 2 and Table II,

    and the synthesized total and baseflow hydrograph is given in Figure 3. A graphical comparison between the

    observed and simulated fast flow time series shows a good fit. This justifies that the slow flow component

    should be separated off first before calibrating the TFM.

    In connection with Figure 3, this implies that, from the general linear TFM [Equation (2)], the effec-

    tive rainfall can be linearly related to the fast flow component of the hydrograph as presented in the

    Equation (11):

    Qst D aQst1 C bRtQdtn 11

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    8/15

    2288 S. MWAKALILA ET AL.

    Table II. DBM results after separating the slow flow Qb component

    Model ordera R2 YIC a parameters b parameters

    1, 1, 0 092 1096 a1: 09580(00009) b0: 00257(00005)

    1, 1, 1 091 1074 a1: 09559(00010) b0: 00268(00006)

    1, 1, 2 090 1046 a1: 09550(00011) b0: 00271(00006)

    1, 2, 2 090 565 a1: 09565(00012) b0: 00378(00034)b1: 00115(00035)

    1, 2, 1 091 523 a1: 09568(00011) b0: 00342(00032)b1: 00080(00033)

    1, 2, 0 091 377 a1: 09570(00010) b0: 00227(00031)b1: 00036(00032)

    a Model order D N, M, (N is the number of a parameters, M is the number of b parameters, is adelay).

    0.00

    2.00

    4.00

    6.00

    8.00

    10.00

    0 200 400 600 800

    Time (days)

    D

    ischarge(mm/day)

    Qd (mm)

    Qd(mm)

    Figure 2. Observed Qd and simulated Q0d fast flow component

    0.00

    4.002.00

    6.00

    8.00

    10.00

    12.00

    0 200 400 600 800

    Time (days)

    Dischar

    ge(mm/day)

    Qt(mm)

    Qb(mm)

    Figure 3. Total discharge Qt and slow flow component (baseflow) Qb

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    9/15

    PREDICTING RUNOFF GENERATION 2289

    where Qst is a simulated fast flow component, Rt is the observed rainfall, Qdt is the estimated (from

    Equation (9)) fast flow component, t is time, and a, b, and n are optimized model parameters. The slow flow

    component can be represented by Equation (12) as:

    Qbt D Qo eSt/m 12

    where Qbt is a simulated slow flow component, St is the catchment storage deficit estimated from

    Equations (7) and (8), and Qo, m and c are optimized model parameters.

    The identified final model structure can be presented as shown in Figure 4, which indicates that a parallel

    model structure is appropriate with a proportion of the runoff being routed through a fast flow pathway and

    the remainder through a slow flow pathway.

    As an example, the synthesized model performance of behavioural simulations for feasible parameter

    ranges using the framework of the GLUE approach is shown in Table III. Table III, shows that there may

    be many representations of a catchment that may be equally valid in terms of their ability to produceacceptable simulations of the available data. Multiple acceptable parameter sets give a range of predictions.

    This may actually be an advantage, since it allows the possibility of assessing the uncertainty in predictions

    Effective rainfall

    Fast Flow Pathway(Surface Flow)

    Qs(t)= aQs(t1)+ bRt(Qd(t))n

    Recharge

    DischargeQtRainfallRt Nonlinear rainfallfilter function

    Slow Flow Pathwya(Base Flow)

    Qb(t) = Qoe-st/m

    Figure 4. Final block diagram of the identified model structure

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    10/15

    2290 S. MWAKALILA ET AL.

    Table III. A sample of identified model performance after using a GLUE approach

    Set of parameters Model performance (likelihood measure)

    c m Qo n a b 1971 1972 1973 1974 1975 1976

    040 5804 462 096 0742 0043 091 092 096 088 084 084033 4989 440 084 0770 0044 091 092 096 087 085 084033 5014 446 085 0777 0040 090 093 096 087 084 083040 6233 715 084 0768 0041 091 092 095 086 084 083026 5130 461 077 0781 0046 090 091 095 086 084 083028 5705 567 093 0686 0049 085 093 095 088 082 082033 5447 741 073 0739 0049 089 093 094 085 083 083034 4917 626 069 0855 0034 094 089 094 083 084 083042 4527 677 082 0796 0031 092 091 093 083 082 082048 4970 405 099 0575 0047 090 089 093 084 082 081

    041 5870 523 063 0807 0041 093 090 093 082 083 083009 6038 451 055 0840 0049 090 088 094 082 084 082036 5491 488 091 0807 0027 086 091 092 083 080 081054 5393 521 088 0767 0037 094 083 094 081 085 083020 6358 416 075 0767 0044 083 092 092 084 080 081051 5044 747 086 0846 0029 095 083 094 080 084 083042 5349 606 063 0877 0031 096 085 093 079 084 082040 6189 477 082 0815 0028 088 090 092 082 080 081014 5375 797 065 0869 0032 088 090 093 081 081 081035 5686 429 075 0809 0032 088 092 091 082 079 081029 4920 629 090 0653 0046 082 092 092 085 079 079026 5712 718 060 0826 0037 088 092 091 081 079 081020 4870 793 064 0780 0044 084 092 090 081 078 080018 4891 627 088 0782 0033 080 091 092 083 079 079009 6032 489 056 0801 0048 084 092 090 081 079 080

    047 4616 786 071 0708 0041 092 088 090 079 079 080053 6064 775 085 0848 0026 094 081 092 078 083 082033 5032 663 075 0803 0030 086 091 090 080 078 079042 6306 710 057 0894 0028 096 084 091 077 082 081045 5688 697 073 0733 0037 089 088 089 079 077 079020 6055 510 092 0675 0043 077 091 090 084 077 077049 4746 549 061 0846 0030 095 083 091 076 080 081044 4745 619 086 0526 0049 087 089 089 079 076 078026 4566 460 072 0752 0039 083 092 089 080 076 078039 4716 718 100 0883 0014 086 088 090 079 077 078011 5633 767 055 0880 0030 086 091 089 078 077 080050 4684 675 095 0815 0019 090 085 090 078 078 079044 5338 724 060 0886 0023 094 085 090 076 078 080042 4628 671 077 0670 0039 087 090 088 078 076 078

    as conditioned on the calibration data and then using that uncertainty as part of the decision-making process

    arising from any rainfall-runoff modelling project.

    The dotty plots in Figure 5 represent a projection of a sample of points on the goodness of fit response

    surface onto individual parameter dimensions. Each dot represents one run of the model from an MCS out

    of 4000 simulations with different randomly chosen parameter values. It will be seen that for each parameter

    there are good simulations across a wide range of parameters. The good models are those that plot near the

    top. This indicates that for each parameter set there are good simulations across a wide range of parameters.

    There are generally also poor simulations across the whole range of each parameter sampled. Whether a

    model gives good or poor results is not a function, therefore, of individual parameters, but of the whole set

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    11/15

    PREDICTING RUNOFF GENERATION 2291

    1

    0.8

    0.6

    0.4

    0.2

    0 0

    0.2

    0.4

    0.6

    0.8

    1

    0 0.5 1 0.5 0.7 0.9

    c-Parameter value n-Parameter value

    Likelihoodmeasure

    Likelihoodmeasure

    * Simulations

    0

    0.2

    0.4

    0.6

    0.8

    1

    0.5 0.6 0.7 0.8 0.9

    a-Parameter value

    Likelihoodmeasure

    * Simulations

    0

    0.2

    0.4

    0.6

    0.8

    1

    0.01 0.02 0.03 0.04 0.05

    b-Parameter value

    Likelihoodmeasure

    * Simulations

    * Simulations

    1

    0.8

    0.6

    0.4

    0.2

    0

    45 65 85

    m-Parameter value

    Likelihoodmeasure

    * Simulations

    1

    0.8

    0.6

    0.4

    0.2

    0

    4 5 6 7 8

    Qo-Parameter value

    Likelihoodmeasure

    * Simulations

    Figure 5. Parameter distributions of behavioural simulations after being conditioned with the 197071 period data of G. Ruaha catchmentat IKA 8A station

    of parameter values and the interactions between parameters. As a projection of the response surface, the

    dotty plots cannot show the full structure of the complex parameter interactions that shape that surface. In

    one sense, however, that does not matter too much, since it primarily allows us to identify where the good

    parameter sets are as a set.

    An impression of the parameter sensitivity can be gained from Figure 6, which depicts the cumulative

    likelihood weighted distributions for each parameter after evaluating each behavioural parameter set on data

    from the 197075 simulation period. Lacking any prior information about the covariation of the individual

    parameters, each was sampled independently from uniform distributions across the feasible range identified

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    12/15

    2292 S. MWAKALILA ET AL.

    Figure 6. Rescaled likelihood weighted distributions for parameters after being conditioned with the 1970 71 period data at IKA 8A station

    from initial evaluation of the TFM and recession flow analysis. A straight diagonal line on each plot would

    represent the prior distributions. Parameter c and parameter m show the strongest departures from the prior

    distributions. The other four parameters (Qo, n, a and b) are much less well conditioned by the observations.

    The possible reasons for parameter sensitivity can be related to the physical interpretation of parameters

    controlling the slow flow and fast flow pathways of the hydrograph: parameter n controls the non-linearity of

    fast runoff generation, parameter a control the residence time of the fast flow pathway; parameter b controls

    the differences in total volumes of input and output; parameter m controls the effective catchment soil profile;

    parameter c controls the proportion of rainfall that percolates as recharge to the groundwater; parameter Qo

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    13/15

    PREDICTING RUNOFF GENERATION 2293

    represents the discharge when the recession of streamflow commences. Therefore, owing to the shallow wet

    soil in the catchment, evapotranspiration would not be expected to be strongly limited (indirectly controlled

    by the n parameter), neither would the residence time for surface flow (controlled by the a parameter). Thus

    any changes in the c and m parameters would result in variations in recharge, which has a major influence

    on streamflow.

    Since dotty plots show that for each parameter there are good simulations across a wide range of parameters,

    all of these good parameter sets will, however, give different predictions. In this study, therefore, the resulting

    uncertainty in the predictions was estimated by weighting the predictions of all the acceptable models by their

    associated degree of belief (as quantified by the likelihood measure values).

    Figures 7 and 8, therefore, present prediction limits for stream discharge from the selected catchment after

    conditioning in the 197071 simulation period and updating with the likelihoods on individual years for the

    entire 1970 75 period data. Figure 9 depicts the prediction limits after testing the behavioural parameter sets

    with new rainfall and discharge data for the 19761978 observation period. For most of the simulations the

    observations are bracketed by the prediction limits, but there are periods when the observations fall outsidethe prediction limits, especially during the wet period. This indicates some deficiencies in the input data or

    model structure used. As noted from Table I, it seems that a second-order transfer function model (2, 2, 0

    and 2, 1, 1) does slightly better in terms of R2. However the YIC has the most negative value for a first-order

    0.00

    2.00

    4.00

    6.00

    8.00

    10.00

    13-Dec-69 01-Jul-70 17-Jan-71 05-Aug-71 21-Feb-72 08-Sep-72

    Date

    Discharge(mm/day) Observed discharge

    95% Prediction limit5% Prediction limit

    Figure 7. Prediction limits after being conditioned and updated with the 197072 data

    0.00

    2.00

    4.00

    6.008.00

    10.00

    12.00

    27-Nov-72 15-Jun-73 01-Jan-74 20-Jul-74 05-Feb-

    75

    24-Aug-

    75Date

    Discharge(mm/day) Observed discharge

    95% Prediction limit5% Prediction limit

    Figure 8. Prediction limits after conditioning and updating with the 197375 data

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    14/15

    2294 S. MWAKALILA ET AL.

    0.00

    2.00

    4.00

    6.00

    8.00

    10.00

    12.00

    02-Dec-75 19-Jun-76 05-Jan-77 24-Jul-77 09-Feb-78 28-Aug-78Date

    Discharge(mm/day)

    Observed discharge95% Prediction limit5% Prediction limit

    Figure 9. Prediction limits after testing with the new 197678 discharge data

    transfer function (1, 1, 0) with high R2 (good fit). Therefore, the second-order transfer functions were not

    tried owing to the fact that they have larger standard errors of the estimated parameter values, an indication

    that a model is over-parameterized. It should be noted here that, the aim in this study was to find a model

    structure that gives a good fit but is parsimonious in having a small number of coefficients.

    The authors suggest that much of the errors may be due to an inadequate representation of the spatial

    patterns of rainfall within the catchment. Although the models used in this study are not fully distributed

    process-based models, these results are perhaps representative of the type of accuracy that might be achieved

    in predicting the discharge while representing the state of the catchment as a single set of process-related

    parameters. Further, it proves to be surprisingly difficult in rainfall-runoff modelling to bracket all the discharge

    observations for at least 90% confidence interval.

    CONCLUSIONS

    With a DBM modelling approach, the simplest TFM structure that is consistent with the observations is made

    up of storage elements for fast flow (surface flow) and slow flow (baseflow) components. The advantage of

    the DBM approach is that the data are allowed to suggest the form of the transfer function used rather than

    specifying a fixed structure beforehand.

    On the basis of the GLUE approach used in this study and the findings, it can be concluded that for

    any rainfall-runoff model there is no single best parameter set that represents the catchment for a range

    of rainfall-runoff responses, but rather a range of different sets of model parameter values may representthe rainfall-runoff process equally well. Therefore, there may be many representations of a catchment that

    may be equally valid in terms of their ability to produce acceptable simulations of the available data. The

    parameter sensitivity analysis with the GLUE approach is essentially a non-parametric method, in that

    it makes no prior assumptions about the variation or covariation of different parameter values, but only

    evaluates sets of parameter values in terms of their performance. This study has found that recession flow

    parameters that control the slow flow component are relatively more sensitive than those of the fast flow

    component.

    Within the GLUE framework, multiple acceptable parameter sets give a range of predictions. This may

    actually be an advantage, since it allows the possibility of assessing the uncertainty in predictions as

    conditioned on the calibration data and then using that uncertainty as part of the decision-making process

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)

  • 8/6/2019 Mwakalila Feyen Hydro Logical Processes

    15/15

    PREDICTING RUNOFF GENERATION 2295

    arising from the modelling project. However, the present study has found that during the wet period the

    rainfall-runoff modelling does not bracket all the discharge observations for at least 90% confidence interval.

    This is due to errors in the input data used being dominated by peaks during the wet season and being

    relatively small in the dry period recession. Most of these errors may be due to an inadequate representation

    of the spatial patterns of rainfall within the catchment. Although the models used in this study are not fully

    distributed process-based models, these results are perhaps representative of the type of accuracy that might

    be achieved in predicting the discharge while representing the state of the catchment as a single set of

    process-related parameters.

    ACKNOWLEDGEMENTS

    This work was performed during the doctoral study programme of the first author at the Institute for Land

    and Water Management (ILWM), Katholieke Universiteit Leuven (K.U.Leuven), Belgium. The study was

    funded by the Belgian Administration for Development Co-operation (BADC/ABOS) and the Universityof Dar es Salaam (UDSM), Tanzania. The first author is pleased to acknowledge this financial support.

    He is also grateful for the accessibility to the streamflow data and rainfall data provided by the Water

    Section of the Department of Civil Engineering at UDSM and the Directorate of Meteorology respec-

    tively.

    REFERENCE

    Beven KJ, 1993. Prophecy, reality and uncertainty in distributed hydrological modelling. Advances In Water Resources 16: 41 51.Beven KJ. 2001. Rainfall-Runoff Modelling: The Primer. Wiley: Chichester.Beven KJ, Binley AM. 1992. The future of distributed models: model calibration and uncertainty prediction. Hydrological Processes 6:

    279298.Beven KJ, Lamb R, Quinn P, Romanowicz R, Freer J. 1995. TOPMODEL. In Computer Models of Watershed Hydrology , Singh VP (ed.).

    Water Resources Publications: 627 668.Cameron D, Beven KJ, Tawn J, Blazkova S, Naden P. 1999. Flood frequency estimation by continuous simulation for a gauged uplandcatchment (with uncertainty). Journal of Hydrology 219: 169187.

    DANIDA. 1995. Joint study of integrated water and land management in the Great Ruaha Basin. A Final report. DANIDA/World Bank,1995.

    Franks SW, Gineste Ph, Beven KJ Mert Ph. 1998. On constraining the predictions of a distributed model: the incorporation of fuzzy estimatesof saturated areas into the calibration process. Water Resources Research 34: 787797.

    Freer J, McDonnel J, Beven KJ, Brammer D, Burns D, Hooper RP, Kendal C. 1997. Topographic controls on subsurface stormflow at thehillslope scale for two hydrologically district small catchments. Hydrological Processes 11(9): 13471352.

    Jakeman AJ, Littlewood IG, Whithead PG. 1990. Computation of the instantaneous unit hydrograph and identifiable component flows withapplication to two small upland catchments. Journal of Hydrology 117: 275300.

    Jakeman AJ, Chen TH, Post DA, Hornberger GM, Littlewood IG, Whithead PG. 1993. Assessing uncertainties in hydrological response toclimate at large scale. In Macroscale Modelling of the Hydrosphere , Vol. 214. Wilkinson WB (ed.). IAHS Publication: 3747.

    Jakeman AJ, Hornberger GW. 1993. How much complexity is warranted in a rainfall-runoff model? Water Resources Research 29:26372649.

    Melching CS. 1995. Reliability estimation. In Computer Models for Watershed Hydrology , Singh VP (ed.). Water Resources Publications:69117.

    Nash JE, Sutcliffe J. 1970. River flow forecasting through conceptual models, part I A discussion of principles. Journal of Hydrology 10:282290.

    Post D, Jakeman AJ. 1996. Relationships between catchment attributes and hydrological response characteristics in small Australian mountainash catchments. Hydrological Processes 10: 877892.

    Romanowicz R, Beven KJ, Tawn J. 1994. Evaluation of predictive uncertainty in non-linear hydrological models using a Bayesian approach.In Statistics for the Environment II. Water Related Issues , Barnett V, Turkman KF (eds). Wiley: 297317.

    Schreider SYU, Jakeman AJ. 1996. Modelling rainfall-runoff from large catchment to basin scale: the Goulburn valley, Victoria. HydrologicalProcesses 10: 863876.

    Sefton CEM, Howarth SM. 1998. Relationships between dynamic response characteristics and physical descriptors of catchments in Englandand Wales. Journal of Hydrology 211: 116.

    Young PC. 1984. Recursive Estimation and Time Series Analysis . Springer-Verlag: Berlin.Young PC. 1998. Data-based mechanistic modelling of environmental, ecological, economic and engineering systems. Environmental

    Modelling Software 13: 105122.Young PC, Beven KJ. 1994. Data-based mechanistics modelling and the rainfall-flow non-linearity. Environmetrics 5: 335363.

    Copyright 2001 John Wiley & Sons, Ltd. Hydrol. Process. 15, 22812295 (2001)