128: Rainfall-runoff modeling: Transfer Function

Models
PETER C YOUNG
Centre for Research on Environmental Systems & Statistics (CRES), Lancaster University,
Lancaster, UK (Also, Centre for Resource & Environmental Studies, Institute of Advanced Studies,
Australian National University, Canberra ACT 2020, Australia)

This article discusses continuous and discrete-time transfer function (TF) models within the context of the Data-
Based Mechanistic (DBM) modeling of hydrological systems. Although, at a superficial level, TF models simply
provide a convenient and concise way of presenting differential or difference equations in an input-output form,
they are much more than this. In particular, they allow for the consideration of linear dynamic systems in simple
algebraic terms and for their analysis within a dynamic systems and control context. This is useful in various
ways, particularly where a higher order TF model is identified and estimated directly from hydrological data.
Often, this data-based model can then be decomposed into serial, parallel, and feedback connections of first-
order systems that can be interpreted in hydrologically meaningful terms, thereby facilitating the model’s use in
hydrological systems analysis. The article also shows that TF models can be considered in time-variable and
state-dependent parameter (SDP) form, so allowing for them to describe nonstationary and nonlinear systems.
And, since it is straightforward to consider all TF models in stochastic terms, they provide a powerful vehicle
for uncertainty, forecasting, and risk analysis. The practical utility and power of DBM TF models is illustrated
by two real hydrological examples.

INTRODUCTION (RF) processes. This is followed by a tutorial section that

From a conceptual standpoint, most mathematical models
of hydrological systems are formulated on the basis of
natural laws, such as dynamic conservation equations,
often expressed in terms of continuous-time (CT) linear
or nonlinear differential equations. The objective of Data-
Based Mechanistic (DBM) modeling (e.g. Young and Lees,
1993; Young, 1998 and the prior references therein) is to
infer the nature and structure of such models directly from
hydrological data, using powerful methods of statistical
identification and estimation, and to then interpret the
model equations in physically meaningful terms. One
important generic model class that facilitates such DBM
modeling studies is the Transfer Function (TF) family of
models, which is the subject of this article. The article
starts by reviewing briefly some of the background to the
use of TF models in hydrology, concentrating on prior
publications concerned with the modeling of rainfall-flow
discusses the formulation of linear, constant parameter TF
models in the derivative operator s = d/dt, which are
simply the CT, TF form of ordinary differential equations.
It then proceeds to develop and discuss the discrete-time
(DT) equivalents of these CT models, namely, TFs in the
backward shift operator z−1 (see later text), before outlining
methods for the statistical identification and estimation of
both CT and DT models from noisy time series data. In this
manner, the models take on a natural stochastic form that is
appropriate to their use in applications such as hydrological
forecasting, uncertainty, and risk analysis. Two practical
examples help demonstrate the practical utility of the TF
approach in hydrology, one concerned with RF modeling
and the other with the dynamics of solute transport and
dispersion in rivers. Finally, the article also shows that TF
models can be extended to incorporate time-variable and
“state-dependent” parameters, so allowing them to describe
nonstationary and nonlinear stochastic systems.

Encyclopedia of Hydrological Sciences. Edited by M G Anderson.

 2005 John Wiley & Sons, Ltd.
2 RAINFALL-RUNOFF MODELING
TF MODELS IN HYDROLOGY
The most common application of TF models in hydrology
is connected with the modeling of RF processes, where
the objective is to characterize the dynamic relationship
between rainfall and total measured flow in the river. Note
that the term “rainfall-flow” is used here, with an emphasis
on the total flow in the river rather than the ubiquitous term
Rainfall-Runoff (RR). As we shall see later, this is because
the most recent work on the application of data-based TF
models in this context has shown how they are able to
quantify other components of the flow; most importantly,
those arising from groundwater processes that dominate the
base flow in the river.
The characterization of the nonlinear dynamic relation-
ship between rainfall and river flow is one of the most
interesting and challenging modeling problems in hydrol-
ogy. It has received considerable attention over the past
40 years, with mathematical and computer-based models
ranging from simple black-box representations to complex,
physically based catchment models. Other articles in this
Encyclopedia will deal more comprehensively with this
topic and there are many books that consider it, either in
whole, or in part: for example, Anderson and Burt (1985:
see particularly, the chapter by Wood and O’Connell); Shaw
(1994), Singh (1995) and Beven (2001). But, in attempting
to consider the historical context of this topic, we should
recall Beven’s comments in the latter book:
“It is now virtually impossible for any one person to be aware
of all the models that are reported in the literature, let alone
know something of the historical framework of the different
initiatives.”
On this basis, it is clear that any treatment of such extensive
and quite often controversial literature will tend to be
somewhat subjective.
Wheater et al. (1993) have categorized RF models into
four, broad types: conceptual models, physics-based mod-
els, metric models and Hybrid Metric-Conceptual (HMC)
models. Conceptual and physics-based models tend to be
driven by deterministic, reductionist thinking. As a result,
they are often very large and suffer from problems of over-
parameterization that make rigorous statistical identification
and estimation from the available data difficult or even
impossible. At the other extreme, metric models, such as
the neural network (e.g. Hsu et al., 1993) and neuro-fuzzy
types (e.g. Jang et al., 1997), are the epitome of “black-box”
modeling, revealing very little of their internal structure that
has any physical meaning.
The HMC models are an attempt to combine the ability
of metric models to efficiently characterize the observa-
tional data in statistical terms (the “principle of parsimony”
(Box and Jenkins, 1970); or the “Occam’s Razor” of antiq-
uity), with the advantages of simple conceptual models that
have a prescribed physical interpretation within the cur-
rent scientific paradigm. Transfer functions have become a
very useful tool in HCM modeling because their outward
“black-box” nature obscures an underlying ability to explain
the more physically meaningful mechanisms that under-
lie this input-output description; mechanisms that can be
revealed using the “system theoretic” procedures of TF
decomposition. In one important sense, TF models pro-
vide a rather natural model form for hydrologists because
the impulse response of a continuous-time TF is equiva-
lent to the Instantaneous Unit Hydrograph (IUH); while the
pulse response of the discrete-time TF is equivalent to the
discrete-time Unit Hydrograph (UH) (or TUH): see later
Section “The unit hydrograph and finite impulse response
TF models”.
System theoretic modeling ideas have had quite a strong
influence in hydrology over many years: to name but a
few contributions, for example, see Amorocho and Hart
(1964); Salas et al. (1980); Wood (1980); O’Connell and
Clarke (1981); Sorooshian (1983); Gupta (1984); Bras and
Rodriguez-Iturbe (1985) and Singh (1988). However, such
models do not often attempt to represent the nonlinear
dynamics inherent in the transformation of rainfall to runoff
and, therefore, may not always perform well over long peri-
ods of time when the catchment storage and soil wetness is
changing over a wide range. Sometimes, piecewise linearity
has been assumed to handle some aspects of the nonlinear
behavior (e.g. Natale and Todini, 1976a,b); while, in other
cases, the model parameters have been allowed to vary with
time (Young, 1974; Whitehead and Young, 1975; White-
head et al., 1976; Young and Wallis, 1985; Young, 1986).
The latter Time-variable Parameter (TVP) approaches to
TF estimation have led on to a more overt treatment of
nonlinearity, sometimes by the use of nonlinear black-box
models (e.g. the early example of Hsu et al., 1993), and
sometimes by the extension of TF models to allow for non-
linearity, as considered in this article. As we shall see, this
latter approach has the advantage of allowing for the easier
inclusion of conceptual ideas and the retention of phys-
ical meaning within the estimated model structure. This
desire to associate a physical meaning with TF models
has a rather natural appeal to hydrologists and was proba-
bly first adumbrated by Young (1974) and Whitehead and
Young (1975). The idea of a Physically Realizable Trans-
fer Function (PRTF: see Han, 1991) tackled this problem
by ensuring that the estimated TF model maintained mass
conservation and did not have any nonphysical, oscillatory
modes of dynamic behavior. It was also the basis for the
work of Jakeman et al. (1990) and provided the philosophi-
cal and methodological underpinning of the DBM approach
to modeling RF processes (Young, 1993; Young and Beven,
1994) that is discussed further in the following sections of
this article.
Finally, it must be emphasized that TF models are not
restricted to RF processes: since they provide a generic
tool for modeling linear and nonlinear, stochastic dynamic
 RAINFALL-RUNOFF MODELING: TRANSFER FUNCTION MODELS
systems, they can be applied to data-based modeling
problems in many areas of science, engineering, and the
social sciences (see the wide variety of examples in Young,
1998). In the hydrological context, they have already been
used for flow routing down rivers (see Section “Continuous-
time TF models”) so that, together with the rainfall-flow
TF models, they allow for characterization of a whole river
catchment. As we shall see in a later example, TF models
are also fundamental to the Aggregated Dead Zone (ADZ)
modeling of pollution transport and dispersion in rivers and
soils. Given this flexibility, it seems likely that they will be
used even more widely in the future as a basic tool for
modeling in hydrological and environmental sciences.
CONTINUOUS-TIME TF MODELS
In order to introduce TF models, let us consider first a
conceptual catchment storage equation in the form of a
continuous-time, linear storage (tank or reservoir) model:
see, for example, the review papers by O’Donnell, Dooge
and Young in Kraijenhoff and Moll (1986); or more
comprehensive treatments, such as the recent books by
Beven (2001) and Dooge and O’Kane (2003). Here, the
rate of change of storage in the channel is defined in terms
of water volume entering the linear storage element (e.g.
river reach) in unit time, minus the volume leaving in the
same time interval, namely,
dS(t)
= GQi (t − τ ) − Qo (t) (1)
dt
where Qi (t − τ ) represents the input flow rate delayed by
a pure time or “transport” delay of τ time units to allow
for pure advection; and G is a gain parameter inserted
to represent gain (or loss) in the system. Making the
reasonable and fairly common assumption that the outflow
is proportional to the storage at any time, that is,
Qo (t) = T S(t)
and substituting into (1), we obtain,
dQo (t)
T = GQi (t − τ ) − Qo (t) (2)
dt
This equation is a first-order, linear differential equation
model whose response, from an initial steady flow condi-
imp
tion, to a unit impulsive change Qi of the input flow at
time t = t0 , is given by
Qo (t) = Qe + Qi e−(t−t0 )/T
imp
where Qe is the initial steady flow level. This has a typical
hydrograph recession shape, with a decay Time Constant,
3
T , that defines the Residence Time of the model. As we
shall see later, combinations of two or more such first-
order models exhibit a typical UH form (e.g. Dooge, 1959;
Beven, 2001).
By introducing the derivative operator s, that is, s =
d/dt, and collecting like-terms together, it is easy to see
that equation (2) can be written as,
(1 + T s)Qo (t) = GQi (t − τ )
so that, dividing throughout by 1 + T s, we obtain the
following continuous-time TF form of equation (2),
G
Qo (t) = Qi (t − τ ) (3)
1+Ts
where,
G
H (s) =
1+Ts
represents the TF in terms of the derivative operator s. Note
that the same letter s (or sometimes p) is used to represent
the related Laplace transform operator. Considered in these
Laplace transform terms, it is possible to utilize Laplace
transform methods to handle initial conditions on the
variables in the model and analytically compute its response
(here Qo (t)) to variations in input variable (here Qi (t)).
However, this is not essential in the present context,
although the interested reader should find the article a good
primer for the study of Laplace transform methods (e.g.
Schwarzenbach and Gill, 1979).
Physically Interpretable Parameters
The TF model 3 is characterized by three parameters: G,
T , and τ . However, there are five, physically interpretable
model parameters associated with the model that are worth
discussing. The Steady State Gain (SSG), denoted by G,
is obtained by setting the s operator in the TF to zero
(i.e. d/dt = 0 in a steady state). It shows the relationship
between the equilibrium output and input values when a
steady input is applied. For this reason, if the input and
output have similar units, G is ideal for indicating the
physical losses or gains occurring in the system. In the case
of a flow-routing model, for example, it indicates whether
water has been added (G > 1) or lost (G < 1) between the
upstream and downstream boundaries; and the percentage
of water lost or gained can be defined by Loss Efficiency
LE = 100(1 − G), which will be negative if G > 1.0. As
pointed out above, the Residence Time or Time Constant T
is the time required for the storage element output to decay
to e−1 or 0.3679 of its maximum value in response to an
impulsive input. Finally, the pure Advective Time Delay τ
indicates the time it takes for a flow increase upstream to
be first detected downstream: and Tt = T + τ defines the
Travel Time of the system. These five parameters typify the
4 RAINFALL-RUNOFF MODELING

equilibrium and dynamic characteristics of the TF model Rainfall-flow

and provide a physical interpretation of the TF model in subsystem
terms of its mass transfer and dispersive characteristics. G1 Flow routing
subsystem
1 + T1s
TF Manipulation and Block Diagrams u (t ) G3 G3 y(t)
1 + T3s 1 + T3s
The first-order TF model (3) is often written in the form, G2
g0 g0 1 + T2s
Qo (t) = Qi (t − τ ) i.e. H (s) = (4)
s + f1 s + f1
where, Figure 1 Block diagram of a hypothetical catchment
G 1 model consisting of a parallel pathway, rainfall-flow
g0 = ; f1 =
T T subsystem in series with a two reach, flow-routing
subsystem
because this is the form in which the model is normally
estimated (see later). Typically, a Channel or Flow-routing
model for a river catchment will contain a number of again of the form (3) with τ = 0, but this time connected
elemental models, such as 3 or 4, connected in a manner in series as a “Nash Cascade”. The upstream input to
that relates to the structure of the catchment. For instance, a this complete system is an effective rainfall measure,
serial connection of n such elements constitutes the lag-and- represented by u(t), that is generated from the gauged
route model of a single river channel (Meijer, 1941; Dooge, rainfall r(t) (not shown) by a nonlinear relationship N (s)
1986) and, with τ = 0 and all elements identical, it becomes involving a measure of the catchment wetness (see later
the well known “Nash Cascade” model (Nash, 1959). More example). The downstream output of the complete system
complex river systems can be represented by a main channel is denoted by y(t).
of this type, with tributaries modeled in a similar manner. One advantage of the TF formulation of the model shown
Also, a typical TF model between effective rainfall and in Figure 1 is that it allows for the computation, using Block
flow often contains a parallel connection of two or more Diagram Algebra, of a single, multiorder TF that represents
such storage elements (see Young, 1992; 2001b,c, 2002a, the total system. Here, TFs connected in parallel are
2003). Related serial and parallel arrangements characterize additive; while those connected in series are multiplicative.
the mass conservation equation of the ADZ model for the Consequently, in this case, the two submodels can be
transport and dispersion of a solute in a river channel (e.g. represented by the following composite TFs:
Wallis et al., 1989 and the prior references therein). And
ADZ-type models can lead to more complex interactive Hrf (s) = H1 (s) + H2 (s); Hff (s) = H3 (s).H3 (s)
water quality models, including chemical and biological
interaction, such as dissolved oxygen-biological oxygen So that, with the above definitions of H1 (s), H2 (s) and
demand (DO-BOD) models (e.g. Beck and Young, 1975; H3 (s),
Whitehead and Young, 1975; Young, 1999b). Practical
examples of both RF and ADZ modeling are described G1 G2
later in Section “STATISTICAL IDENTIFICATION AND Hrf (s) = +
1 + T1 s 1 + T2 s
ESTIMATION OF TF MODELS”.
(G1 + G2 ) + (G1 T2 + G2 T1 )s
The TF formulation allows for the visual representation =
of a total system model in the form of a Systems Block (1 + T1 s)(1 + T2 s)
Diagram. Figure 1 is a typical example of such diagram
that represents a catchment model consisting of a effective and
(G3 )2
RF submodel involving two different first-order TFs of the Hff (s) =
form (3) with τ = 0, (1 + T3 s)(1 + T3 s)

G1 G2 Now, since Hrf (s) and Hff (s) are connected in series, the
H1 (s) = ; H2 (s) =
1 + T1 s 1 + T2 s TF of the total system H (s) is obtained as the multiplication
of these two composite TFs: that is,
that are connected in parallel. The flow output of this
submodel forms the input to a flow-routing submodel (G1 + G2 ) + (G1 T2 + G2 T1 )s
composed of two identical, first order TFs, H (s) = Hrf (s).Hff (s) =
(1 + T1 s)(1 + T2 s)
G3 (G3 )2
H3 (s) = ×
1 + T3 s (1 + T3 s)(1 + T3 s)
 RAINFALL-RUNOFF MODELING: TRANSFER FUNCTION MODELS
Multiplying out these expressions, we see that the complete
H (s) is a 4th order TF that can be manipulated to the form:
g0 + g1 s
H (s) = (5)
s + f1 s + f2 s 2 + f3 s + f4
4 3
where we will leave the definition of the parameters fi , i =
1, 2, . . . , 4 and gj , j = 0, 1 in 5, as an exercise for the
reader (hint: carry the above analysis with the TF represen-
tation 4 rather than 3: then it will be clear why the former
representation is better for analysis, although it lacks the
direct physical interpretation of the latter, which provides
a better form for the block diagram representation).
The General, Multi-Order TF Model
It is clear from the above example that, in general, serial,
parallel (or even feedback) connections of elemental first
order TF models, such as 3 or 4, lead to a multi-order
TF model. The block diagram algebra for a feedback
connection is a little more complicated but, since it is not
particularly relevant in the current context, the interested
reader should consult a standard text on the subject (e.g.
Schwarzenbach and Gill, 1979) to find the details. In the
present context, the general multi-order TF form is as
follows:
G(s)
x(t) = u(t − τ ) y(t) = x(t) + ξ(t) (6)
F (s)
where F (s) and G(s) are polynomials in s of the follow-
ing form:
F (s) = s p + f1 s p−1 + f2 s p−2 + · · · + fp
G(s) = g0 s q + g1 s q−1 + g2 s q−2 + · · · + gq
in which p and q can take on any positive integer val-
ues. Here u(t) and x(t) denote the deterministic input and
output signals of the system at its upstream and down-
stream boundaries, respectively; τ is a pure advective time
(transport) delay affecting the input signal u(t); and y(t)
is the observed output, which is assumed to be contami-
nated by a noise or stochastic disturbance signal ξ(t). This
noise is assumed to be independent of the input signal
and it represents the aggregate effect, at the downstream
boundary, of all the stochastic inputs to the system, includ-
ing distributed unmeasured inputs, measurement errors, and
modeling error. Multiplying throughout equation (6) by
F (s) and converting the resultant equation to alternative
ordinary differential equation form, we obtain:
d p y(t) d p−1 y(t)
+ f1 + · · · + fp y(t)
dt p dt p−1
d q u(t − τ )
= g0 + · · · + gq u(t − τ ) + η(t) (7)
dt q
5
where η(t) = F (s)ξ(t) is a modified noise signal generated
by the manipulation of the equation. The structure of this
model, in either form (6) or (7), is defined by the triad
[p q τ ].
If the noise ξ(t) can be assumed to have specific
statistical characteristics, then this can help in obtaining low
or minimum variance estimates of the TF model parameters.
But, as we see later when discussing the identification
and estimation of TF models, such assumptions are not
an essential requirement of statistical estimation and good,
low variance parameter estimates can be obtained without
invoking them. Also, depending on the objectives of
the modeling study, it may be necessary, in a complete
system consisting of many sub-elements such as (3) or
(4), to consider noise inputs within the system, associated
with collections of subelements that have distinct physical
meaning: for example, stochastic lateral inflows.
Discrete-time, Sampled Data TF Models
To date, the most popular form of TF modeling has been
carried using the DT equivalents of the models (4) and (6).
In the case of equation (4), this discrete-time TF model
takes the form:
b0
Qo,k = Qi,k−δ (8)
1 + a1 z−1
Here, Qo,k is the downstream flow measured at the kth
sampling instant, that is at time kt, where t is the
sampling interval in time units. Qi,k−δ is the input flow
at time (k − δ)t time units previously, where δ is the
advective time delay, normally defined as the nearest inte-
ger value of τ/t (thus incurring a possible approximation
error); and z−1 is the backward shift operator, that is,
z−r Qo,k = Qo,k−r . Of course, this model can be written
in its “difference equation” form, namely:
Qo,k = −a1 Qo,k−1 + b0 Qi,k−δ (9)
which is obtained by simple cross multiplication and
application of the z−1 operator. This reveals that the flow
Qo,k at the kth sampling instant is a proportion −a1 (note
that, in the present context, a1 will be a negative number
less than unity, so that this is a positive proportion) of its
value Qo,k−1 at the previous (k − 1)th sampling instant, plus
a proportion b0 of the delayed upstream flow input Qi,k−δ
measured δ sampling instants previously.
The values of the parameters a1 and b0 in equations (8)
and (9) can be related to the parameters of the model
(4) in various ways depending upon how the input flow
Qi (t) is assumed to change over the sampling interval
between the measurement of Qi,k−1 and Qi,k (since it is
not measured over this interval). The simplest and most
common assumption is that it remains constant over this
6 RAINFALL-RUNOFF MODELING
interval (the so-called zero-order hold, ZOH, assumption),
in which case the relationships are as follows:
g0
a1 = − exp(−f1 t) b1 = {1 − exp(−f1 t)}
f1
Note that, because these relationships are functions of the
sampling interval t, for every unique CT model such as
(4), there are infinitely many DT equivalents (8), depending
on the choice of t, all with different parameter values
defined in equation (10). Following from the definition of
this first-order DT model at the chosen t, the general
multiorder equivalent of the general CT model (6) is defined
as follows:
B(z−1 )
xk = uk−δ yk = xk + ξk
A(z−1 )
where,
A(z−1 ) = 1 + a1 z−1 + a2 z−2 + · · · + an z−n
B(z−1 ) = b0 + b1 z−1 + b2 z−2 + · · · + bm z−m
Normally n = p but m may be equal or greater than q. The
structure of this DT model is defined by the triad [n m δ].
Note finally that the TFs in both (11) and (3) are composed
of ratios of polynomials (in z−1 and s respectively), so they
are sometimes referred to as Rational Transfer Functions.
The Unit Hydrograph and Finite Impulse Response
TF Models
The reader can verify that the division of the numer-
ator polynomial B(z−1 ) by the denominator polynomial
A(z−1 ) in the TF model (2) normally results in a infinite
dimensional polynomial G(z−1 ) = g0 + g1 z−1 + g2 z−2 +
· · · + g∞ z−∞ , so that the equation can also be written in
the alternative form:
∞

the CT case; ai , i = 1, 2, . . . , n; bj , j = 0, 1, . . . , m; and
yk = gi uk−i + ξk ; gj , j = 0, 1, . . . , δ − 1 = 0 (12)
i=δ
This will be recognized as the discrete-time form of the
convolution integral equation associated with the solution
of differential equations (here with a pure time delay of δ
sampling intervals or δt time units) and, once again, we
see that the TF model is simply a discrete-time equivalent
of a continuous-time differential equation model.
Equation (12) has important connotations in hydrology
because, if the noise ξk = 0 for all k, then its unit impulse
response (that is the response of the equation to a unit
impulse input: uk = 1.0 for k = 1; uk = 0 for all k > 0),
is equivalent to the hydrological UH. Sometimes, indeed,
the equation (12) is referred to as a TF model, although
its infinite dimensional nature is an obvious restriction.
Despite this disadvantage, some hydrologists have used the
equation directly in RF modeling by considering it in a
Finite Impulse Response (FIR) form, in which the upper
limit of the summation is set to some value r < ∞ where
(10)
it is considered that the ordinates of the impulse response
have become small enough to be ignored.
Unfortunately, it can be shown that the number of FIR
model parameters (or “weights”), gi , i = 1, 2, . . . , r, is
nearly always much larger than the number of parameters
in the rational TF form (11). As a result, the statistical esti-
mates of these parameters will normally have unacceptably
high variance and have to be constrained in some man-
ner. For instance, Natale and Todini (1976) use quadratic
programming to compute the constrained least squares esti-
mates of the FIR parameters in order to ensure that they are
(11) all positive and, if necessary for conservation purposes, they
all sum to unity (i.e. the model has unity steady state gain).
Even with this approach, however, it is difficult to recom-
mend the FIR model because the rational TF model (11),
normally with far fewer parameters, is not only entirely
equivalent to the infinite dimensional IR model, without
any approximation, it is also easier to estimate from RF
data using the methodology discussed in the next Section
“STATISTICAL IDENTIFICATION AND ESTIMATION
OF TF MODELS”.
STATISTICAL IDENTIFICATION AND
ESTIMATION OF TF MODELS
In statistical terms, the modeling of a system using a TF
model is a problem of time-series analysis (e.g. Box and
Jenkins, 1970; Young, 1984) that involves identification
and estimation of the model from time-series data. Here,
Identification means the definition of the most appropriate
model structure: in particular, the values of the elements
in the triads [p q τ ] and [n m δ]. Estimation is then
the estimation of the parameters that characterize this
structure (fi , i = 1, 2, . . . , p; gj , j = 0, 1, . . . , q; and τ in
δ in the DT case). Much has been written about the
identification and estimation of the DT model (11) but
much less on the CT model 6. This is almost certainly
for two reasons: first, it is easier to analyze the DT
model in statistical terms; second, in these days of digital
instrumentation and computers, the DT model provides a
more convenient model form. For these reasons, attention
will be concentrated initially on the DT model (11) and we
will return to the more physically transparent CT modeling
later.
Discrete-time TF Model Identification and
Estimation
Much has been written on the identification and estimation
of discrete-time TF models and the reader is advised to

consult one or more of the many texts on the subject (e.g.
Ljung and Söderstrom, 1983; Young, 1984; Norton, 1986).
Here, we will not attempt to review this literature but simply
refer briefly to a number of topics that are particularly
pertinent to the use of TF models in hydrology.
The Box– Jenkins Model
The best known work on the identification and estimation
of the DT model (11) is by Box and Jenkins (1970:
see Chapter 10), to which the reader is directed for
a comprehensive description of how the model can be
estimated using the Maximum Likelihood (ML) approach,
under the assumption that the noise ξk has rational spectral
density. This assumption leads to the formulation of the
following Box –Jenkins model:
B(z−1 ) D(z−1 )
yk = uk−δ + ξk ξk = ek ek = N (0, σ 2 )
A(z−1 ) C(z−1 )
or,
B(z−1 ) D(z−1 )
yk = −1
uk−δ + −1
ek ek = N (0, σ 2 )
A(z ) C(z )
where,
C(z−1 ) = 1 + c1 z−1 + c2 z−2 + · · · + cs z−s
D(z−1 ) = 1 + d1 z−1 + d2 z−2 + · · · + dr z−r
and ek is defined as normally distributed “white noise”,
defined in statistical terms as N (0, σ 2 ): that is a zero mean,
serially uncorrelated sequence of normally distributed ran-
dom variables with variance (an additional unknown param-
eter) σ 2 . We see, therefore, that the “colored” noise ξk is
assumed to be generated from the white noise ek by passing
it through a “filter” that has a similar TF form to the sys-
tem model, but with both polynomials in z−1 having leading
unity elements (monic polynomials). This particular type of
filter is called a AutoRegressive Moving Average (ARMA)
model (see Box and Jenkins, 1970). With the noise process
defined in this manner, it is clearly necessary to identify the
order of the noise model polynomials C(z−1 ) and D(z−1 )
and estimate the associated parameters ci , i = 1, 2, . . . , s
and di , i = 1, 2, . . . , r and σ 2 .
The ARMAX Model
In the systems and control literature, the most widely
used approach to DT model estimation is probably the
Prediction Error Minimization (PEM) method (e.g. Ljung
and Söderstrom, 1983), which is the preferred method in the
Matlab Identification Toolbox. PEM estimation is most
often applied to the alternative AutoRegressive Moving
Average eXogenous (ARMAX) TF model form: that is,
A(z−1 )yk = B(z−1 )uk−δ + C(z−1 )ek ; ek = N (0, σ 2 )
RAINFALL-RUNOFF MODELING: TRANSFER FUNCTION MODELS 7
Despite the popularity of this ARMAX model in the
systems literature, there are good theoretical and practical
reasons for preferring the model (13), as discussed below.
Refined Instrumental Variable (RIV) Estimation
One approach to TF model estimation that is simpler and
more robust, because it does not rely on the assumption
that the noise has rational spectral density, is the Refined
Instrumental Variable (RIV) method, which has proven
particularly useful and robust in hydrological applications.
The RIV approach (see Young and Jakeman, 1979; Young,
1984) is based on the TF models (11) and (13) and
was initially developed with environmental applications
in mind, where robustness to the kinds on nonstandard
conditions met in the environment, including hydrology,
is very important. The RIV method has three primary
advantages that are pertinent to the discussion here.
1. In the case of the full Box–Jenkins model (13) and
following from an important theorem by Pierce (1972),
the estimates of the “system” parameters in A(z−1 )
(13)
and B(z−1 ) are asymptotically independent of those
for the “noise” parameters in C(z−1 ) and D(z−1 ). This
has important consequences on parameter estimation
(Jakeman and Young, 1981, 1983) that justify the
RIV estimation of the TF model (13) rather than the
ARMAX model (14). They also expose deficiencies in
the ARMAX model estimates if the data are generated
by the model (13).
2. Like all instrumental variable methods (see e.g. Young,
1984), the RIV approach does not necessarily require
the concurrent estimation of a noise model and, indeed,
remains statistically efficient (i.e. the parameter esti-
mates are optimal in the sense that amongst all esti-
mates, they have the minimum variance) if the noise
ξk is itself white (i.e. C(z−1 ) = D(z−1 ) = 1.0). Used in
this manner, the RIV algorithm is, for obvious reasons,
called the Simplified RIV (SRIV) algorithm (Young,
1985).
3. As pointed out by Wellstead (1978) and considered fur-
ther by the present author and his colleagues (Young
et al., 1980; Young, 1989), model structure identifica-
tion (definition of the triad [n m δ]) is greatly facilitated
by the special properties of the Instrumental Product
Matrix (IPM) that is a fundamental component of IV
estimation and has optimal statistical properties in the
case of RIV/SRIV estimation. This has been exploited
in the development of the Young Information Criterion
(YIC) order identification criterion mentioned later.
4. The RIV/SRIV algorithms for DT models utilize a
special form of adaptive prefiltering. This is further
exploited in the RIV method for continuous-time
systems (the RIVC algorithm: Young and Jakeman,
(14) 1980; Young, 2002b) to bypass the need for direct
8 RAINFALL-RUNOFF MODELING
differentiation of the input and output signals. This
algorithm is discussed further in the following section.
Continuous-time TF Model Identification and
Estimation
Following on the point 4. in the previous subsection, the
statistical estimation of the CT model (6) is straightforward
if completely continuous-time data are available. However,
the hydrologist is normally confronted by discrete-time,
sampled data and the problem of modeling continuous-time
models such as this, based on discrete-time, sampled data
at sampling interval t, is not so obvious. This problem
can be approached in two main ways.
• The Indirect Approach: Here, the identification and
estimation steps are first applied to the DT model (11)
or (13). This estimated model is then converted to the
CT model (6), again using some assumption about the
nature of the input signal u(t) over the sampling interval
t. In the first-order, ZOH case, this conversion is given
by the relationships in equation (10) but, in more general
terms, the multiorder conversion must be carried out in
a computer, using a conversion routine such as the d2cm
algorithm in the Matlab Control Toolbox.
• The Direct Approach: Here, the RIVC algorithm,
together with an appropriate order identification criterion
(see the following text), is used to identify the most appro-
priate, identifiable CT model structure defined by the triad
[p q τ ]; and then estimate the TF parameters fi , i =
1, 2, . . . n, gj , j = 0, 1, . . . m and τ that characterize this
structure. Of course, some approximation will be incurred in
this estimation procedure because the inter-sample behavior
of input signal u(t) is not known and it must be interpo-
lated over this interval in some manner (see the preceding
text).
Model Structure Identification
Finally, as mentioned previously, an important aspect of TF
modeling, in both continuous and discrete-time, is Model
Structure Identification (the definition of the model struc-
ture triads). This can be approached in various ways. For
example, Box and Jenkins (1970) discuss the topic at length
and Akaike (1974) suggested an alternative approach that
has since spawned several other, related methods. Within
the context of the instrumental variable methods discussed
previously, however, model structure identification is based
conveniently on two statistical criteria. First, the Coeffi-
cient of Determination, RT2 , a normalized measure based
on the variance of the error between the sampled output
data yk and the simulated (CT or DT) model output at the
same sampling instants (this is normally equivalent to the
well known Nash-Sutcliffe Efficiency measure (Nash and
Sutcliffe, 1970); and 100 × RT2 is the percentage of the
variance of the output data explained by the model). Note
that RT2 should not be confused with its well known rel-
ative R 2 , as used in classical regression and time-series
analysis, which is defined in terms of the one-step-ahead
prediction errors. In general, RT2 is a more discerning mea-
sure of model adequacy in a dynamic systems context than
R 2 since it quantifies the ability of the model to explain
the whole of the output data from the input data, without
any reference to the output data. In contrast, R 2 only mea-
sures the ability of the model to predict one-step-ahead on
the basis of the latest input and output data. The second
statistic is the YIC, which is a heuristic measure of model
identifiability and is based on the properties of the IPM
matrix (see the previous discussion). These statistics are
discussed in more detail in Young (1989) and Appendix 3
of Young (2001b).
TWO PRACTICAL EXAMPLES
TF models can be applied in many areas of hydrology
and provide a good basis for DBM modeling. Two typical
examples are outlined in this section. The first is a TF
model of the relationship between effective rainfall and
flow in an Australian river catchment. This DBM model
could provide a direct basis for flow forecasting at the flow
measurement station or it could be extended to incorporate
flow-routing TF elements down the river in order to allow
for forecasting over the whole catchment area (that may
involve other RF modules). The second example is an
ADZ model for the transport of a conservative dye tracer
in a Florida wetland area. This DBM model describes the
physical aspects of solute advection and dispersion in the
wetland and so could provide the physical basis for more
complex, nonconservative water quality modeling involving
biological and chemical mechanisms.
Rainfall-flow Modeling
Figure 2 shows a portion of effective rainfall u(t) and flow
y(t) data from the River Canning, an ephemeral River in
Western Australia. A longer set of these data has been
analyzed comprehensively in Young et al. (1997), using DT
modeling. The effective rainfall series plotted in part (b) of
Figure 2 is generated from the effective rainfall nonlinearity
identified in this earlier study, with the effective rainfall
generated by the equation
u(t) = r(t)y(t)γ (15)
where r(t) is the measured rainfall and γ is a power law
parameter. Here, the flow y(t) is acting as a surrogate
measure of the catchment storage, rather than attempting to
model this storage separately in some conceptual manner
(see e.g. Jakeman et al., 1990; Wheater et al., 1993; Young
2001b, 2003). In the following, we obtain direct and
indirectly estimated, linear CT models between the effective

4 occurs in the very short term (within one sampling interval;
Flow (cumecs)
3
2 time of about 2.53 days, probably the result of surface and
1
0 flow, that passes through a much longer 15.9 day residence
(a) 1977.2 1977.3 1977.4 1977.5
80
Effective rainfall (mm)
60
40
20
0 a similar [2 3 0] structure TF with the following parameter
1977.2 1977.3 1977.4 1977.5
(b) Time (h)
Figure 2 Daily effective rainfall (b) and flow (a) data for
the River Canning in Western Australia. The dotted line
appears to me to have come out as a grey line (a) is the
simulated output of the DBM model (see text)
rainfall u(t), as defined in the above manner, and the
flow y(t). Of course, the DT model is useful in its
own right and provides an excellent basis for applications
such as flow forecasting (see e.g. Lees et al., 1994; Lees,
2000a,b; Young, 2001c, 2002a). However, the transparency
of the CT model in revealing, more obviously, the physical
interpretation of the model makes it more useful as a vehicle
in DBM modeling and the use of the model in wider
catchment studies.
Direct CT Identification and Estimation
Let us consider first identification and estimation based
on the measured daily data shown in Figure 2. The RIVC
algorithm, in combination with the RT2 and YIC statistics,
identifies a [2 3 0], second order model with the parameter
estimates:
fˆ1 = 0.457 (0.032) fˆ2 = 0.0248 (0.0045)
ĝ0 = 0.0138 (0.001) ĝ1 = 0.0505 (0.002)
ĝ2 = 0.0046 (0.0008)
where, here and elsewhere in the paper, the figures in
parentheses are the estimated standard errors obtained from
the RIVC algorithm. This model output (shown as the
dotted line in (a) of Figure 2) explains the data very
well with RT2 = 0.980 (98% of the output flow variance
explained by the simulated CT model output) and a residual
variance of 0.00723 (standard deviation 0.085 cumecs,
where the maximum flow is 3.86 cumecs). The model
can also be decomposed by standard TF decomposition
(see earlier text) into a parallel pathway form. The three
pathways are: (i) an instantaneous effect, accounting for
7.4% of the flow, which is a measure of rapid flow that
RAINFALL-RUNOFF MODELING: TRANSFER FUNCTION MODELS 9
here one day); (ii) a quick pathway, accounting for 54.1%
of the flow, that represents a linear store with a residence
shallow, subsurface processes in the catchment; and (iii) a
slow or base-flow pathway, accounting for 38.5% of the
1977.6 1977.7 1977.8
time, linear store, probably representing the effects of
deeper groundwater processes and the displacement of old
water.
Indirect CT Identification and Estimation
The first step of indirect CT modeling involves DT identifi-
cation using the RIV algorithm (see earlier). This identifies
1977.6 1977.7 1977.8
estimates:
â1 = −1.6034 (0.008) â2 = 0.6244 (0.007)
b̂0 = 0.0140 (0.001) b̂1 = 0.0151 (0.002)
b̂2 = 0.0252 (0.0013)
However, since this is a discrete-time model, it has to
be converted to continuous-time form. The d2cm algo-
rithm in Matlab (see Section “CONTINUOUS-TIME TF
MODELS”.) accomplishes this conversion, using a ZOH
approximation (i.e. the input effective rainfall is assumed
to be constant over the sampling interval). The resulting
CT model parameter estimates are:
fˆ1 = 0.4711 fˆ2 = 0.0264
ĝ0 = 0.0140 ĝ1 = 0.0514 ĝ2 = 0.0049
In this case, the model is quite similar to the directly
estimated CT model in the previous subsection (i). But note
that there are no standard errors quoted here: this is because
these are not available after the conversion and must be
computed separately in some manner, usually by the less
convenient and lengthy Monte Carlo simulation analysis.
This lack of directly available estimation statistics is
not the only disadvantage of the indirect approach. Unlike
the CT model parameters, which are independent of the
sampling interval t, the estimated DT model parameters
are a function of t, so that different numerical estimates
of the parameters are obtained for different values of
t. Moreover, the quality of these estimates is also a
function of t: in particular, if the sampling interval is too
small, then the estimates are poorly defined statistically; the
resulting CT model parameter estimates will be impaired;
and the whole estimation process is less robust (see Young,
2004a, for discussion on the reasons for this). Conveniently,
it is under these same, rapid sampling (small t) conditions
that the direct CT approach works best. So, in practical
terms, indirect CT model estimation works very well
when the data are not sampled too rapidly but direct CT
estimation is preferable at rapid sampling intervals.
10 RAINFALL-RUNOFF MODELING
Aggregated Dead Zone (ADZ) Modeling of Tracer
Data
Tracer experiments can provide important information on
how solutes are transported and dispersed in hydrological
systems. This example is concerned with TF modeling of
the input-output data shown in Figure 3 (TF modeling is
also applied to a different set of tracer data in the arti-
cle Chapter 134, Downward Approach to Hydrological
Model Development, Volume 3). These tracer data were
obtained from a bromide tracer experiment carried out in a
Florida wetland area receiving treated domestic wastewater
for further nutrient removal. The experiment was part of
a study carried out by Chris Martinez and Dr. William R.
Wise of the Environmental Engineering Sciences Depart-
ment, University of Florida for the City of Orlando. The
bromide tracer was injected 765 meters upstream of a weir,
at which samples were taken with a sampling interval t of
2 h. As we shall see, TF models are not only able to explain
these tracer data well, but following the requirements of
DBM modeling, they can also be interpreted in ADZ model
terms that have physical meaning (see Wallis et al., 1989).
Here, we will consider the discrete-time TF model and
utilize the SRIV algorithm (see earlier text) to identify
the model order and estimate the parameters. However,
continuous-time TF estimation using the continuous-time
SRIV algorithm yields similar (albeit not identical) results
and the discrete-time analysis is used here simply for illus-
tration.
Although the impulsive (“gulp”) input does not perturb
the system continually (in the systems literature, it is said
Wetland tracer experiment data (Martinez, 2001)
4.5
4 associated with oscillatory response and this is difficult to
Tracer concentration (mg L−1)
3.5
3 based on mass conservation (Wallis et al., 1989) and so,
2.5
2 when considered in TF terms.
1.5
1 are constrained to be real. This was carried out by means of
0.5
0 most parametrically efficient model was obtained, both [3 2
100 200 300 400 500 600 700 800
Time (h)
Figure 3 Wetland tracer experiment data: the input uk
is an impulsive or ‘‘gulp’’ application of bromide tracer
(dashed line) and the output yk is the concentration of
bromide measured every 2 h at a downstream weir (full
line)
to be “not persistently exciting”), the SRIV algorithm has
no difficulty identifying and estimating a low order model.
The best identified TF, based on the RT2 and YIC criteria, is
either 3rd or 4th order, but subsequent analysis, described
later, suggests that the latter is superior from a physical
standpoint. The estimated [4 2 22] (4th order denominator,
2nd order numerator and a 22 sampling interval pure time
delay) takes the form:
B̂(z−1 )
yk = uk−22 + ξk (16)
Â(z−1 )
where,
Â(z−1 ) = 1 − 3.67z−1 + 5.06z−2 − 3.11z−3 + 0.72z−4
B̂(z−1 ) = 0.00103 − 0.00101z−1
Here the “hat” denotes the estimated value; yk is the
observed tracer concentration at the weir and uk is the
impulsive input of tracer (186.33 mg l−1 ), both measured at
the kth sampling instant. Note that, although the standard
errors on the parameters are available they are not cited
specifically. However, the uncertainties on the physically
meaningful, derived ADZ model parameters are discussed
later. Note also that the large advective time delay of 22
sampling intervals (44 h) is the time taken for the solute to
first reach the weir. The noise ξk is small and so the model
explains the data very well with RT2 = 0.997 (i.e. 99.7% of
the output variance is explained by the model).
Unfortunately, despite its ability to describe the data
very well, the model (16) is not immediately acceptable
from a DBM standpoint, primarily because the eigenvalues
√
are {0.988, 0.964, 0.860 ± 0.132j }, where j = (−1) is
the complex number. Here, the pair of complex roots are
justify in ADZ modeling terms. In particular, the elemental,
single reach ADZ model is a first-order differential equation
other than in exceptional circumstances, multiple reach
ADZ models must be characterized by real eigenvalues
In the present circumstances, the most obvious approach
is to re-estimate the model in a form where the eigenvalues
constrained, nonlinear least squares optimization using the
leastsq optimization procedure in Matlab. To ensure that the
900
22] and [4 2 22] models were considered in this analysis but
the latter yielded much the best constrained model, which
has the following form:
B̂(z−1 )
yk = uk−22 + ξk (17)
Â(z−1 )
 RAINFALL-RUNOFF MODELING: TRANSFER FUNCTION MODELS 11

where, obtained by partial fraction expansion of the TF in equation

(17). This decomposition is shown in the block diagram of
Â(z−1 ) = (1 − 0.980z−1 )(1 − 0.855z−1 )3 Figure 5 which, it will be noted, has exactly the same form
as Figure 1, except for the addition of the pure advective
B̂(z−1 ) = 0.00127 − 0.00121z−1
time delay, τ , demonstrating the widely applicable, generic
nature of TF models such as these.
This model is well defined statistically and it explains
The block diagram in Figure 5 shows that the solute is
99.7% of the experimental data (RT2 = 0.997), the same as
partitioned initially into a parallel pathway form with a
the unconstrained model (1). Figure 4 compares the model
“quick-flow” pathway, consisting of a single ADZ reach
output (full line) with the measured output yk (circular
with a residence time of 12.8 h, and a “slow-flow” pathway
points).
with a much longer residence time of 99 h. The two
Unlike the TF model (16), the model (17) not only has pathways then add together and the solute passes further
four real eigenvalues, as required, but three of these are downstream, through two more ADZ reaches, each with
repeated, so defining three identical ADZ reaches. These a residence time of 12.8 h. The total travel time for this
eigenvalues define ADZ residence times (time constants) complete system is 181.5 h (the sum of the 44 h advective
of 99 h and 12.8 h (×3), giving a total estimated residence time delay and the cumulative overall time constant of
time for the wetland cell of 137.4 h (99 + 3 × 12.8). 137.4 h). This means that the dispersive fraction (see Wallis
One particular, physically meaningful, decomposition and et al., 1989; Young and Wallis, 1994; Young, 1999a) is
interpretation of the model defined in this manner is again 0.76 (i.e. 137.4 ÷ 181.5): in other words, 76% of the water
appears to be effective in dispersing the solute. This is a
4.5 very high proportion, reflecting the nature of the system
in this case, with a much higher potential for dispersion
4 of tracer than in normal, faster moving streams, where
the dispersive fraction is normally in the range 0.3 to 0.4.
Tracer concentration (mg L−1)

3.5
The inferred responses of the two pathways are plotted in
3 Figure 4: the dotted line shows the estimated concentration
changes in the quick-flow pathway (effectively three 12.8 h
2.5
residence time ADZs in series), which accounts mainly for
2 the initial response measured at the weir; while the dashed
line shows the estimated changes in the slow-flow pathway
1.5
(effectively a 99 h residence time ADZ in series with two
1 12.8 h ADZs), and these are mainly responsible for the
raised tail of the measured response.
0.5
It is possible to compute estimates of other physical
0 attributes associated with the model. First, the steady state
0 100 200 300 400 500 600 700 800 900 gains associated with the two parallel pathways define the
Time (h)
partitioning of the flow, with 33% of flow associated with
the quick pathway and 67% with the slow pathway. And
Figure 4 Comparison of the DBM model output (full line)
and tracer experiment data (circular points). Also shown since the flow rate is known in this example, the Active
are the inferred slow-flow component (dashed line) and Mixing Volumes (AMVs: Young and Lees, 1993), based
quick-flow component (dotted line) on the estimated partitioned flow, are 361 m3 in the quick

Quick flow
pathway

G1 x(t)
1 + T1s
Advective y(t)
u(t) u(t − t) G1 G1
time + 1 + T1s 1 + T1s
+
delay
G2
1 + T2s

Slow flow
pathway

Figure 5 Block diagram of a transfer function decomposition for the ADZ model that can be interpreted in physical terms
as a parallel-serial decomposition of ADZ reaches. T1 = 12.8 h; T2 = 99 h; τ = 22 h
12 RAINFALL-RUNOFF MODELING
Slow path residence time
1 1
0.8
Normalized frequency
0.6
0.4
0.2
0 0
95 100 105
(a) Residence time (h)
Figure 6 MCS analysis results: normalized histograms of the slow-flow (a) and quick-flow (b) pathway residence times
pathway and 5656 m3 in the slow pathway. As a result,
the total estimated AMV is 5656 + 3 × 361 = 6739 m3 ,
which seems reasonable when compared with the 9749 m3
for the total volume of the wetland, estimated by physical
measurement. This suggests that about 70% of the wetland
is important in dispersing the tracer (and, therefore, the
waste water) and compares reasonably with the dispersive
fraction derived percentage of 76%.
Of course, all of the results above are statistical estimates
and so they are inherently uncertain. The advantage of the
DBM/TF approach is that we can quantify and consider the
consequences of this uncertainty (see Young, 1999a). For
instance, based on the covariance matrix of the parame-
ters produced by the SRIV estimation analysis, empirical
probability distributions, in the form of histograms, can
be computed for the “derived” physical parameters, such
as the residence times, partition percentages, AMVs, total
AMV, and steady state gain, using Monte Carlo Simulation
(MCS) analysis. Figure 6 is a typical example of such anal-
ysis: it shows the normalized empirical distributions for the
two residence times obtained by MCS using 10 000 random
realizations (the procedure used here is discussed in Young,
1999a).
Of course, it should be noted that the parallel decom-
position of the estimated TF used here is not unique:
there are other decompositions that are just as valid and
give precisely the same input–output response. For exam-
ple, two other examples are: (i) a serial connection of the
three ADZs with residence times of 12.8 h, in parallel with
another serial connection, this time of an ADZ with res-
idence time of 99 h in series with two 12.8 h residence
Quick path residence time
0.8
Normalized frequency
0.6
0.4
0.2
12.4 12.6 12.8 13
(b) Residence time (h)
time ADZs the same as those in the first parallel pathway;
(ii) various decompositions including feedback processes.
However, the latter seem less supportable in physical terms
and are rejected according to the DBM ethos.
Finally, how can decompositions of ADZ reaches, such
as those shown in Figure 5 be interpreted in terms of the
wetland system? The most plausible mechanism is that the
pure time delay allows for the flow-induced pure advection
of the solute (chemical engineers would call this “plug-
flow”). The quick-flow parallel pathway then represents
the “main stream-flow” that is relatively unhindered by
the vegetation; while the slow-flow pathway represents
the solute that is captured by the heavy vegetation and
so dispersed more widely and slowly before rejoining the
main flow and eventually reaching the weir. It is this
latter pathway, which we have shown above accounts for
some 67% of the flow, together with the large associated
dispersive fraction of 76%, that is most useful in terms
of nutrient removal, since it allows more time for the
biological activity to take place.
NONSTATIONARY AND NONLINEAR TF
MODELS
Although the RF model discussed in Section “DISCRETE-
TIME TF MODEL IDENTIFICATION AND ESTIMA-
TION” is characterized by an input nonlinearity, the exam-
ple has emphasized the constant parameter, linear TF part
of the model (which defines the shape of the catchment
UH). However, the nonlinearity is very important because
 RAINFALL-RUNOFF MODELING: TRANSFER FUNCTION MODELS
it defines the way in which the model responds under dif-
ferent catchment wetness conditions. So, in general, it is
necessary to consider extensions to the constant parameter,
linear TF model that allow for the consideration of more
complex hydrological processes such as this. Fortunately, it
is possible to extend TF models to handle “nonstationary”
situations described by TVP models or “nonlinear” situ-
ations characterized by State-Dependent Parameter (SDP)
models. The latter model class encompasses a wide vari-
ety of nonlinear, stochastic, dynamic phenomena, including
chaotic systems.
Although such TVP and SDP models are beyond the
scope of this article, they are introduced briefly in the fol-
lowing subsections. Both types of models can be considered
in either continuous or discrete-time. However, since statis-
tical estimation is more straightforward in the DT case, it
often provides the more practical approach. Consequently,
the brief descriptions that follow are limited to this DT
situation.
Nonstationary Time-Variable Parameter (TVP)
Models
In the DT case, the TVP form of the TF model (11) can be
written as follows:
B(k, z−1 )
xk = uk−δ yk = xk + ξk (18)
A(k, z−1 )
where,
A(k, z−1 ) = 1 + a1,k z−1 + a2,k z−2 + · · · + an,k z−n
B(k, z−1 ) = b0,k + b1,k z−1 + b2,k z−2 + · · · + bm,k z−m
Here, all the parameters are assumed to be functions of the
time index k, that is it is assumed that they may vary over
time in an unknown manner that needs to be estimated from
the data.
Some of the latest research on TVP estimation is reported
in Young (1999b, 2000, 2004b) where the reader will
find a complete description of the recursive estimation
algorithms that allow for the estimation of the time-variable
parameters. A simple example would be a model such as
(8) in which the “gain” parameter b0 = b0,k is assumed
to vary over time in order to reflect any changes in the
dynamic behavior of the channel that may occur over a
wide range of flow conditions. An approach similar to
this is suggested by Cluckie (1993), who refers to the
gain as a “model scaling factor” and points out that it
is analogous to the “variable proportional loss method”
of defining effective rainfall. An example of adaptive
gain adjustment that has been implemented in practice is
the adaptive flood warning system for the River Nith at
Dumfries, in south west Scotland (Lees et al., 1993, 1994).
Here, a recursive estimation algorithm has been used for
13
over a decade to continually update a gain parameter,
based on telemetered data from the catchment, in order to
provide an adaptive capability. In this case, however, the
adaptive gain parameter is only required to “fine-tune” an
already computed effective rainfall input generated from
the measured rainfall by a “state-dependent parameter”
nonlinearity, as discussed in the next section.
Nonlinear State-Dependent Parameter (SDP)
Models
Again, in the discrete-time case, the SDP form of the TF
model (11) can be written as follows:
B(wk , z−1 )
xk = uk−δ yk = xk + ξk (19)
A(vk , z−1 )
where,
A(vk , z−1 ) = 1 + a1 (v1,k )z−1 + a2 (v2,k )z−2
+ · · · + an (vn,k )z−n
B(wk , z−1 ) = b0 (w0,k ) + b1 (w1,k )z−1 + b2 (w2,k )z−2
+ · · · + bm (wm,k )z−m
Here, the possibility that the parameters may be functions
of other “state” variables is investigated. In other words, it
is assumed that any parameter in the model may vary over
time as a function of one or more other variables (e.g. the
input uk or output yk and their past values), so introducing
nonlinear behavior into the model. In the SDP model
(19), these variables are denoted by vi,k , i = 1, 2, . . . , n
and wj,k , j = 0, 1, . . . , m and they are, respectively, the
elements of the vectors vk and wk .
The first applications of SDP estimation in hydrology
are described in Young (1993); Lees et al. (1993, 1994)
and Young and Beven (1994); while the latest research on
this topic is reported in Young (2000, 2001a) and Young
(2002). A simple example of SDP estimation is provided by
the estimation of the effective rainfall nonlinearity (15) of
the RF model discussed in Section “DISCRETE-TIME TF
MODEL IDENTIFICATION AND ESTIMATION”. Here,
SDP estimation identifies that, in order to explain the RF
data well, the numerator parameters b0 , b1 and b2 in a
linear TF model of the RF data need to vary significantly
as functions of the flow: that is, they are identified as
SDPs b0 (yk ), b1 (yk ) and b2 (yk ). However, the inclusion of
a power law nonlinearity such as equation (15), operating
on the input rainfall, removes the need for these variations,
leaving the model as a serial connection of the nonlinearity
(15) and the linear, constant parameter TF, as described
in Section “Discrete-time TF model identification and
estimation”. Other related examples have been discussed
recently in Young (2001b,c, 2002a, 2003).
14 RAINFALL-RUNOFF MODELING
CONCLUSIONS
At one level, TF models are simply elegant and convenient
representations of continuous-time, stochastic differential
equation models, or their discrete-time equivalents. This
article demonstrates, however, that they are much more than
this: they also provide a powerful, generic method of mod-
eling hydrological systems. This method provides a useful
block diagram interpretation of the input–output dynamics
of the system under study that can be decomposed straight-
forwardly into physically meaningful sub-systems that are
in an ideal form for DBM modeling studies. The utility of
the method is further enhanced by the powerful methods of
statistical identification and parameter estimation that have
been developed for such TF models over many years. These
are not only robust in practical application, but they also
provide valuable information about the stochastic nature
of the system and allow for the quantification of the uncer-
tainty associated with the model parameter estimates and its
output predictions. Moreover, these estimation methods are
normally available in a recursive form that allows for their
use in on-line, real-time applications, where they provide a
natural basis for adaptive modeling and data assimilation.
Recent developments in recursive time-variable and state-
dependent parameter estimation have enhanced the utility of
TF models still further, so that they are now able to charac-
terize both nonstationary and nonlinear stochastic systems.
Finally, it should be emphasized that DBM models, in
general, and data-based TF models, in particular, naturally
require real time-series data for their identification and
estimation. Also, their primary use is for improving our
understanding of the system dynamics and/or for imple-
menting model-based prediction, as in flood warning and
forecasting applications, where the main changes that are
likely to occur in future relate to the input variables (mea-
sured rainfall and upstream flows). They are not particularly
appropriate, at least in their basic form, for assessing the
impacts of changes to the nature of the system (“what-if”
studies) unless the derived, physically meaningful param-
eters associated with the DBM model (as, for example, in
the case of the Florida wetland example) can be associ-
ated with physical environmental characteristics, such as
the topography or other physical features of a catchment
(“regionalization studies”: see e.g. Kokkonen et al., 2003
and the Chapter 134, Downward Approach to Hydrolog-
ical Model Development, Volume 3 in this Encyclopedia).
In this regard, it is likely that the derived parameters will be
“bulk” parameters that relate to the behavior of the system
at the catchment scale, rather than the kind of “process”
parameters used in hydrological simulation models.
It is clear, however, that when suitable, measured time-
series data are available, DBM/TF models should always
be used as an adjunct to simulation model synthesis
because they define well the “dominant mode” behavior
of the system that is identifiable from the historical data
(see Chapter 134, Downward Approach to Hydrological
Model Development, Volume 3 in this Encyclopedia). In
this way, they can alert the model builder to those aspects of
the simulation model that are not well defined in relation to
the available data and so are of a more speculative nature.
Acknowledgments
The author is grateful to Professor Tony Jakeman of the
Centre for Resource and Environmental Studies, Australian
National University for providing the Canning River data;
and to Chris Martinez and Dr. William R. Wise of the Envi-
ronmental Engineering Sciences Department, University of
Florida, for providing the tracer experiment data.
FURTHER READING
O’Donnell T (1986) Deterministic catchment modelling. In River
Flow Modelling and Forecasting, Kraijenhoff D.A and Moll J.R
(Eds.), D. Reidel: Dordrecht, pp. 11 – 37.
Young P.C, McKenna P and Bruun J (2001) Identification of
nonlinear stochastic systems by state dependent parameter
estimation. International Journal of Control, 74, 1837 – 1857.
Young P.C, Parkinson S.D and Lees M (1996) Simplicity out of
complexity in environmental systems: Occam’s Razor revisited.
Journal of Applied Statistics, 23, 165 – 210.
Young P.C and Pedregal D (1999) Recursive and en-bloc
approaches to signal extraction. Journal of Applied Statistics,
26, 103 – 128.
Young P.C and Tomlin C.M (2000) Data-based mechanistic
modelling and adaptive flow forecasting. In Flood Forecasting:
What Does Current Research Offer the Practitioner? BHS
Occasional paper No 12, Lees M.J and Walsh P (Eds.),
Centre for Ecology and Hydrology on behalf of the British
Hydrological Society: pp. 26 – 40.
REFERENCES
Akaike H. (1974) A new look at statistical model identification.
IEEE Transactions on Automatic Control, ACl9, 716 – 722.
Anderson M.G. and Burt T.P. (Eds.) (1985) Hydrological
Forecasting, John Wiley: Chichester.
Amorocho J. and Hart W.E. (1964) A critique of current
methods of hydrologic systems investigation. Transactions of
the American Geophysical Union, 45, 307 – 321.
Beck M.B. and Young P.C. (1975) A dynamic model for DO-
BOD relationships in a non-tidal stream. Water Research, 9,
769 – 776.
Beven K.J. (2001) Rainfall-Runoff Modelling: the Primer, John
Wiley and Sons: Chichester.
Box G.E.P. and Jenkins G.M. (1970) Time-series Analysis,
Forecasting and Control, Holden-Day: San Francisco.
Bras R.L. and Rodriguez-Iturbe I. (1985) Random Functions and
Hydrology, Addison-Wesley: Reading.
 RAINFALL-RUNOFF MODELING: TRANSFER FUNCTION MODELS
Cluckie I.D. (1993) Real-time flood forecasting using weather
radar. In Concise Encyclopedia of Environmental Systems,
Young P.C. (Ed.), Pergamon Press: Oxford, pp. 291 – 298.
Dooge J.C.I. (1959) A general theory of the unit hydrograph.
Journal of Geophysical Research, 64, 241 – 256.
Dooge J.C.I. (1986) Theory of flood routing. In River Flow
Modelling and Forecasting, Kraijenhoff D.A. and Moll J.R.
(Eds.), D. Reidel: Dordrecht, pp. 39 – 65.
Dooge J.C.I. and O’Kane J.P. (2003) Deterministic Methods in
Systems Hydrology, Balkema: Lisse.
Gupta V.K. (1984) The Identification of Conceptual Watershed
Models, Ph.D. Dissertation, Case Western Reserve University,
Cleveland.
Han D. (1991) Weather Radar Information Processing and Real-
Time Flood Forecasting, Ph. D. thesis, Department of Civil
Engineering, University of Salford.
Hsu K., Gupta H.V. and Sorooshian S. (1993) Artificial
neural network modeling of the rainfall-runoff process. Water
Resources Research, 29, 1185 – 1194.
Jang J.S.R., Sun C.T. and Mizutani E. (1997) Neuro-Fuzzy and
Soft Computing, Prentice Hall: NJ.
Jakeman A.J. and Young P.C. (1981) On the decoupling of system
and noise model parameter estimation in time-series analysis.
International Journal of Control, 34, 423 – 431.
Jakeman A.J. and Young P.C. (1983) Advanced methods of
recursive time-series analysis. International Journal of Control,
37, 1291 – 1310.
Jakeman A.J., Littlewood I.G. and Whitehead P.G. (1990)
Computation of the instantaneous unit hydrograph and
identifiable component flows with application to two small
upland catchments. Journal of Hydrology, 117, 275 – 300.
Kokkonen T.S., Jakeman A.J., Young P.C. and Koivusalo
H.J. (2003) Predicting daily flows in ungauged catchments
model regionalization from catchment descriptors at the
Coweeta Hydrologic Laboratory, North Carolina. Hydrological
Processes, 17, 2219 – 2238.
Kraijenhoff D.A. and Moll J.R. (1986) River Flow Modelling and
Forecasting, D. Reidel: Dordrecht.
Lees M.J. (2000a) Advances in transfer function based flood
forecasting. In Flood Forecasting: What Does Current Research
Offer the Practitioner? Lees M.J. and Walsh P. (Eds.), BHS
Occasional paper No 12, Centre for Ecology and Hydrology on
behalf of the British Hydrological Society: pp. 41 – 55.
Lees M.J. (2000b) Data-based mechanistic modelling and fore-
casting of hydrological systems. Journal of Hydroinformatics,
2, 15 – 34.
Lees M., Young P.C. and Ferguson S. (1993) Adaptive flood
warning. In Concise Encyclopedia of Environmental Systems,
Young P.C. (Ed.), Pergamon Press: Oxford, pp. 234 – 236.
Lees M., Young P.C., Beven K.J., Ferguson S. and Burns J.
(1994) An adaptive flood warning system for the River Nith at
Dumfries. In River Flood Hydraulics, White W.R. and Watts J.
(Eds.), Institute of Hydrology: Wallingford, pp. 65 – 75.
Ljung L. and Söderstrom T. (1983) Theory and Practice of
Recursive Estimation, MIT Press: Cambridge.
Meijer H. (1941) Simplified flood routing. Civil Engineering, 11,
306 – 307.
Nash J.E. (1959) Systematic determination of unit hydrograph
parameters. Journal of Geophysical Research, 64, 111 – 115.
15
Nash J.E. and Sutcliffe J.V. (1970) River flow forecasting
through conceptual models: discussion of principles. Journal
of Hydrology, 10, 282 – 290.
Natale L. and Todini E. (1976a) A stable estimator for
linear models: 1. Theoretical development and Monte-Carlo
experiments. Water Resources Research, 12(4), 667 – 671.
Natale L. and Todini E. (1976b) A stable estimator for linear
models: 2. real world hydrologic applications. Water Resources
Research, 12(4), 672 – 675.
Natale L. and Todini E. (1976) A stable estimator for linear
models: 2. real world hydrologic applications. Water Resources
Research, 12, 672 – 676.
Norton J.P. (1986) An Introduction to Identification, Academic
Press: London.
O’Connell P.E. and Clarke R.T. (1981) Adaptive hydrological
forecasting – a review. Hydrological Sciences Bulletin, 26,
179 – 205.
Pierce D.A. (1972) Least squares estimation in dynamic
disturbance time-series models. Biometrika, 59, 73 – 78.
Salas J.D., Delleur J.W., Yevjevich V. and Lane W.L. (1980)
Applied Modeling of Hydrologic Time Series, Water Resources
Publications: Littleton.
Schwarzenbach J. and Gill K.F. (1979) Systems Modelling and
Control, Edward Arnold: London.
Shaw E.M. (1994) Hydrology in Practice, Third Edition, Chapman
& Hall: London.
Singh V.P. (1988) Hydrologic Systems: Rainfall-Runoff Modeling,
Vol. 1, Prentice-Hall: Englewood Cliffs.
Singh V.P. (Ed.) (1995) Computer Models of Watershed
Hydrology, Colorado: Water Resources Publications.
Sorooshian S. (1983) Surface water hydrology: on-line estimation.
Review of Geophysics and Space Physics, 21, 706 – 721.
Wallis S.G., Young P.C. and Beven K.J. (1989) Experimental
investigation of the aggregated dead zone model for
longitudinal solute transport in stream channels. Proceedings
of the Institution of Civil Engineers, Part 2, 87, 1 – 22.
Wellstead P.E. (1978) An instrumental product moment test for
model order estimation. Automatica, 14, 89 – 91.
Wheater H.S., Jakeman A.J. and Beven K.J. (1993) Progress and
directions in rainfall-run-off modelling. In Modelling Change in
Environmental Systems, Jakeman A.J., Beck M.B. and McAleer
M.J. (Eds.), Wiley: Chichester, pp. 101 – 132.
Whitehead P.G. and Young P.C. (1975) A dynamic-stochastic
model for water quality in part of the Bedford-Ouse River
system. In Computer Simulation of Water Resources Sys-
tems, Vansteenkiste G.C. (Ed.), North-Holland: Amsterdam, pp.
417 – 438.
Whitehead P.G., Young P.C. and Hornberger G.H. (1976) A
systems model of stream flow and water quality in the Bedford
Ouse River – Part l: stream flow modeling Water Research, 13,
1155 – 1169.
Wood E.F. (Ed.) (1980) Workshop on Real Time Forecast-
ing/Control of Water Resource Systems, Pergamon Press: New
York.
Young P.C. (1974) Recursive approaches to time-Series analysis.
Bulletin of the Maths and its Applications, 10, 209 – 224.
Young P.C. (1984) Recursive Estimation and Time-series Analysis,
Springer-Verlag: Berlin.
16 RAINFALL-RUNOFF MODELING
Young P.C. (1985) The instrumental variable method: a practical
approach to identification and system parameter estimation.
In Identification and System Parameter Estimation, Barker
H.A. and Young P.C. (Eds.), Pergamon Press: Oxford, pp.
1 – 16.
Young P.C. (1986) Time-series methods and recursive estimation
in hydrological systems analysis. In River Flow Modelling and
Forecasting, Kraijenhoff D.A. and Moll J.R. (Eds.), D. Reidel:
Dordrecht, Chap. 6, pp. 129 – 180.
Young P.C. (1989) Recursive estimation, forecasting and adaptive
control. In Control and Dynamic Systems: Advances in Theory
and Applications, Vol. 30, Leondes C.T. (Ed.), Academic Press:
San Diego, Chap. 6, pp. 119 – 166.
Young P.C. (1992) Parallel processes in hydrology and water
quality: a unified time series approach. Journal of the Institute
of Water and Environmental Management, 6, 598 – 612.
Young P.C. (1993) Time variable and state dependent modelling
of nonstationary and nonlinear time series. In Developments in
Time Series Analysis, Subba Rao T. (Ed.), Chapman & Hall:
London, pp. 374 – 413.
Young P.C. (1998) Data-based mechanistic modelling of
environmental, ecological, economic and engineering systems.
Environmental Modelling and Software, 13, 105 – 122.
Young P.C. (1999a) Data-based mechanistic modelling, gen-
eralised sensitivity and dominant mode analysis. Computer
Physics Communications, 117, 113 – 129.
Young P.C. (1999b) Nonstationary time series analysis and
forecasting. Progress in Environmental Science, 1, 3 – 48.
Young P.C. (2000) Stochastic, dynamic modelling and signal
processing: time variable and state dependent parameter
estimation. In Nonstationary and Nonlinear Signal Processing,
Fitzgerald W.J., Walden A., Smith R. and Young P.C. (Eds.),
Cambridge University Press: Cambridge, pp. 74 – 114.
Young P.C. (2001a) The identification and estimation of nonlinear
stochastic systems. In Nonlinear Dynamics and Statistics, Mees
A.I. (Ed.), Birkhauser: Boston, pp. 127 – 166.
Young P.C. (2001b) Data-based mechanistic modelling and
validation of rainfall-flow processes. In Model Validation:
Perspectives in Hydrological Science, Anderson M.G. and Bates
P.D. (Eds.), John Wiley: Chichester, pp. 117 – 161.
Young P.C. (2001c) Advances in Real-time Forecasting. Centre
for Research on Environmental Systems and Statistics, Report
No. TR/176, Lancaster University, p. 38.
Young P.C. (2002a) Advances in real-time flood forecasting.
Philosophical Transactions of the Royal Society: Mathematical,
Physical and Engineering Sciences, 360, 1433 – 1450.
Young P.C. (2002b) Comments on the estimation of continuous-
time transfer functions. International Journal of Control, 75,
693 – 697.
Young P.C. (2003) Top-down and data-based mechanistic
modelling of rainfall-flow dynamics at the catchment scale.
Hydrological Processes, 17, 2195 – 2217.
Young P.C. (2004a) Identification and estimation of continuous-
time hydrological models from discrete-time data, Proceedings,
British Hydrological Society BHS2004 Conference, Imperial
College, London.
Young P.C. (2004b) Identification of time varying systems,
Encyclopedia of Life Support Systems (EOLSS), UNESCO-
EOLSS: (www.eolss.net).
Young P.C. and Beven K.J. (1994) Data-based mechanistic
modelling and the rainfall-flow nonlinearity. Environmetrics,
5, 335 – 363.
Young P.C. and Jakeman A.J. (1979) Refined instrumental
variable methods of recursive time-series analysis: Parts I,
single-input, single-output systems. International Journal of
Control, 29, 1 – 30.
Young P.C. and Jakeman A.J. (1980) Refined instrumental
variable methods of recursive time-series analysis: Part III,
extensions. International Journal of Control, 31, 741 – 764.
Young P.C. and Lees M. (1993) The Active Mixing Volume
(AMV): a new concept in modelling environmental systems.
In Statistics for the Environment, Barnett V. and Turkman K.F.
(Eds.), John Wiley: Chichester, Chap. 1, pp. 3 – 44.
Young P.C. and Wallis S.G. (1985) Recursive estimation: a
unified approach to identification, estimation and forecasting of
hydrological systems. Applied Mathematics and Computations,
17, 299 – 334.
Young P.C. and Wallis S.G. (1994) Solute transport and dispersion
in channels. In Channel Networks, Beven K.J. and Kirby M.J.
(Eds.), John Wiley: Chichester, Chap. 6, pp. 129 – 173.
Young P.C., Jakeman A.J. and McMurtrie R. (1980) An
instrumental variable method for model order identification.
Automatica, 16, 281 – 294.
Young P.C., Jakeman A.J. and Post D.A. (1997) Recent advances
in data-based modelling and analysis of hydrological systems.
Water Science and Technology, 36, 99 – 116.