See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/29626455

The PDM rainfall-runoff model

Article  in  Hydrology and Earth System Sciences · January 2007

DOI: 10.5194/hess-11-483-2007 · Source: OAI

CITATIONS READS

462 1,224

1 author:

Robert Moore
UK Centre for Ecology & Hydrology
87 PUBLICATIONS   3,692 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Robert Moore on 10 June 2019.

The user has requested enhancement of the downloaded file.

Hydrol. Earth Syst. Sci., 11(1), 483–499, 2007
The PDM rainfall-runoff model
www.hydrol-earth-syst-sci.net/11/483/2007
© Author(s) 2007. This work is licensed
under a Creative Commons License.

The PDM rainfall-runoff model

R.J. Moore

Centre for Ecology and Hydrology, Wallingford, Oxon, OX10 8BB, UK

Email: rm@ceh.ac.uk

Abstract
The Probability Distributed Model, or PDM, has evolved as a toolkit of model functions that together constitute a lumped rainfall–runoff
model capable of representing a variety of catchment-scale hydrological behaviours. Runoff production is represented as a saturation excess
runoff process controlled by the absorption capacity (of the canopy, surface and soil) whose variability within the catchment is characterised
by a probability density function of chosen form. Soil drainage to groundwater is controlled by the water content in excess of a tension
threshold, optionally inhibited by the water content of the receiving groundwater store. Alternatively, a proportional split of runoff to fast
(surface storage) and slow (groundwater) pathways can be invoked with no explicit soil drainage function. Recursive solutions to the Horton-
Izzard equation are provided for routing flows through these pathways, conveniently considered to yield the surface runoff and baseflow
components of the total flow. An alternative routing function employs a transfer function that is discretely-coincident to a cascade of two
linear reservoirs in series. For real-time flow forecasting applications, the PDM is complemented by updating methods based on error prediction
and state-correction approaches. The PDM has been widely applied throughout the world, both for operational and design purposes. This
experience has allowed the PDM to evolve to its current form as a practical toolkit for rainfall-runoff modelling and forecasting.

Keywords: rainfall-runoff model, PDM, flooding, updating, forecasting

Introduction from reservoirs or constant abstractions, forms the model

The Probability Distributed Model or PDM is a fairly general output.
conceptual rainfall–runoff model which transforms rainfall The components of the PDM model are described in detail
and potential evaporation data to flow at the catchment outlet as the main subject of this paper. Formulations for the
(Moore, 1985, 1999; Moore and Bell, 2002; Moore et al., probability-distributed soil moisture store are first derived
2005; Institute of Hydrology, 1992, 1996; CEH Wallingford, based on mass balance principles with the addition of
2005). Figure 1 illustrates the general form of the model. rainfall, losses to evaporation, drainage to groundwater
Runoff production at a point in the catchment is controlled (recharge) and the production of direct runoff. The direct
by the absorption capacity of the soil (treated together with runoff and recharge are routed via surface and subsurface
canopy and surface detention) to take up water. This can be storages, representing fast and slow pathways to the basin
conceptualised as a simple store with a given storage outlet. Representations of these surface and subsurface
capacity. By considering that different points in a catchment storages — by a choice of Horton-Izzard equation solution
have differing storage capacities and that the spatial variation or by a transfer function discretely coincident with a cascade
of capacity can be described by a probability distribution, it of two linear reservoirs — are set down next. This is
is possible to formulate a simple runoff production model followed by an overview of the model parameters and their
which integrates the point runoffs to yield the catchment calibration, discussion of example applications and a
surface runoff into surface storage. Groundwater recharge summary of the typical model form that is invoked within
from the soil moisture store passes into subsurface storage. the PDM toolkit. Having completed the description of the
The outflow from surface and subsurface storages, together PDM as a simulation model, methods of updating provided
with any fixed flow representing, say, compensation releases for real-time forecasting applications are outlined. Methods

483
R.J. Moore

Surface
P storage
Direct
runoff
S2

qs Surface
E runoff

q
S1
Recharge
Groundwater
Probability- qb Baseflow
storage
distributed soil
moisture storage
S3

Fig. 1. The PDM rainfall-runoff model

of empirical state-correction for the PDM and more generic as follows. First imagine that stores of all possible different
error-prediction techniques are described, and their relative depths are arranged in order of depth and with their open
merits discussed. The paper ends with a historical tops arranged at the same height: this results in a wedge-
perspective on the evolution of the PDM to its current state shaped diagram as depicted in Fig. 2(b). If the basin is
of development together with some concluding remarks. initially dry so that all stores are empty and rain falls at a
net rate P for a unit duration, then stores will fill to a depth
P unless they are of lesser depth than P when they will fill
Probability-distributed soil moisture and spill. During the interval the shallowest stores will start
store generating direct runoff and at the end of the interval stores
Consider that runoff production at any point within a river of depth P will just begin to produce runoff, so that the
basin may be conceptualised as a single storage, or tank, of upper triangular area in Fig. 2(c) denotes the depth produced
capacity c´, representing the absorption capacity of the soil from stores of different depth over the unit interval. Since,
column at that point. The storage takes up water from in general, there are more stores of one depth than another
rainfall, P, and loses water by evaporation, E, until either the actual runoff produced over the basin must be obtained
the storage fills and spills, generating direct runoff, q, or by weighting the depth produced by a store of a given depth
empties and ceases to lose water by evaporation. Figure 2(a) by its frequency of occurrence, as expressed by f(c). Now,
depicts such a storage, whose behaviour may be expressed at the end of the interval stores of depth less than P are
mathematically by generating runoff: let this critical capacity below which all
stores are full at some time t be denoted by
C * { C * (t ) ( C P in the present example). The
*

­ P  E  cc  S 0  P ! cc  E proportion of the basin containing stores of capacity less

qc ® (1)
¯0 P d cc  E than or equal to C* is

probc d C *  ³ o f c dc.
C*
where So is the initial depth of water in storage, and where F (C * ) (2)
P, E and qc represent the depth of rainfall, evaporation and
the resulting direct runoff over the interval being considered. The function F(.) is the distribution function of store capacity
Now consider that runoff production at every point within and is related to the density function, f(c), through the
a river basin may be similarly described, each point differing relation f(c) = dF(c)/dc. This proportion is also the
from another only with regard to the storage capacity. The proportion of the basin generating runoff, so that the
storage capacity at any point, c, may then be considered as contributing area at time t for a basin of area A is
a random variate with probability density function, f(c), so
that the proportion of the river basin with depths in the range Ac (t ) F C * (t )  A. (3)
(c, c + dc) will be f(c)dc.
The water balance for a river basin assumed to have The instantaneous direct runoff rate per unit area from the
storage capacities distributed in this way may be constructed basin is the product of the net rainfall rate, p(t), and the
484
 The PDM rainfall-runoff model

V t + 't  ³ tt +'t qW dW C* (t + 't)

³ C* (t) F c dc . (6)
P E
During dry periods potential evaporation will deplete the
cc qc
water content of each storage. It will be assumed during
such depletion periods that water moves between storages
of different depths so as to equalise the depth of stored water
at different points within the basin. Thus at any time all
(a) Point representation of runoff production by a single store
stores will have a water content, C*, irrespective of their
capacity, unless this is less than C* when they will be full:
storage
element
the water level profile across stores of different depths will
probability density, f(c)
therefore always be of the simple form shown in Fig. 2(c).
C* The assumption which allows redistribution of water
between storages of different size during depletion periods
ci
is particularly important for real-time applications of the
store model where the possibility of updating the store contents
capacity, c is envisaged. Moore (1985) shows how this assumption,
when not invoked, leads to a more complex water accounting
procedure which is less amenable to real-time empirical state
adjustment schemes. Particularly important is that a unique
(b) Basin representation by storage elements of different depth and relationship exists between the water in storage over the
their associated probability density function basin as a whole, S(t), and the critical capacity, C*(t), and in
turn to the instantaneous rate of basin runoff production,
q(t). Specifically, and referring to Fig. 2(c), it is clear that
Direct P the total water in storage over the basin is
runoff
f
S t  cf(c)dc + C * t  ³ C * t f c dc
C * t 
C*
³0  

(7)

store ³0
C * t 
1  F c dc .
capacity, c
For a given value of storage, S(t), this can be used to obtain
(c) Direct runoff production from a population of stores
C*(t) which allows the volume of direct runoff, V(t+Dt), to
be calculated using Eqn. (6) together with Eqn. (5).
The dependence of evaporation loss on soil moisture
Fig. 2. Definition diagrams for the probability-distributed
content is introduced by assuming the following simple
interacting storage capacity component
function between the ratio of actual to potential evaporation,
Eic / Ei , and soil moisture deficit, Smax – S(t):
proportion of the basin generating runoff, F(C (t)); that is *

­ S  S t  ½
be
Eic
q(t ) S t  F C * t  . (4) 1  ® max ¾ ; (8)
Ei ¯ S max ¿
During the i’th wet interval, (t, t+Dt), suppose rainfall and
either a linear (be=1 so Eic ( S (t ) / S max ) Ei ) or quadratic
potential evaporation occur at constant rates Pi and Ei, so
form (be=2) is usually assumed. Here, Smax is the total
that net rainfall pi = Pi – Ei. Then the critical capacity, C*(t),
available storage, and is given by
will increase over the interval according to

f f (9)
C * W  C * t  + S i W  t  t d W d t + 't , (5) S max ³o cf(c)dc ³o 1  F c  dc c,

the contributing area will expand according to Eqn. (3), and where c is the mean storage capacity over the basin.
the volume of basin direct runoff per unit area produced Further loss as recharge to groundwater may be introduced
over this interval will be by assuming that the rate of drainage over the interval, di,

485
R.J. Moore

depends linearly on basin soil moisture content at the start Si Pi  Eic  d i . (14)
of the interval i.e.
During a period when no runoff generation occurs then, for
di kg
1
S t   S t  bg
(10) this general case, soil moisture storage accounting simply
involves the calculation
where kg is the drainage time constant and bg the exponent
of the recharge function (usually set to 1) and St is the S W  S t   S i W  t  t d W d t + 't, 0 d S W  d S max .
threshold storage below which there is no drainage, water (15)
being held under soil tension. An alternative formulation is
available which allows recharge to depend on both soil and When runoff generation does occur then the volume of
groundwater storage for use in catchments where soil/ runoff produced, V(t+Dt), is obtained using Eqn. (6), and
groundwater interactions are important. Consider recharge then continuity gives the replenished storage as
max
into a groundwater store of maximum capacity S g . Then
­S t   S i 't  V t  't  S t  't  d Smax .
a groundwater deficit ratio may be defined as S t  't  ®
¯S max otherwise (16)
S g  S g t 
max
g t  max (11) If basin storage is fully replenished within the interval
Sg
(t,t+Dt) then V(t+Dt) should be computed from continuity
where Sg(t) denotes the groundwater storage at time t. This as
ratio can be used to define a groundwater demand factor
between 0 and 1: V t + 't  S i 't  S max  S (t )  . (17)

§ g t  ·
E
­ The above completes the procedure for soil moisture
f t 
° ¨ ¸ g t   D
® © D ¹ (12) accounting and determining the value of runoff production
°̄ 1 otherwise according to a probability-distributed storage capacity
which achieves a maximum for values of the deficit ratio model. Figure 3 provides a graphical representation of this
g(t) in excess of a. It is then reasonable to suppose that the procedure for a wet interval (t,t+Dt) during which soil
recharge depth over the interval, Di, will increase with soil moisture storage is added to by an amount DS(t+Dt) = pi Dt
storage, S(t), and with the groundwater demand factor, f(t), – V(t+Dt), and a volume of direct runoff, V(t+Dt), is
according to generated.

Di D sat + S max  D sat  f t  S t  . (13)

S max Siǻt
1.0
Here the recharge depth at saturation Dsat=qsat Dt, with qsat
Distribution
the outflow from the groundwater storage when Sg(t) equals function
S gmax (i.e. the maximum rate of recharge). Note that the F(c)
S(t) ǻS(t+ǻt)
drainage rate over the interval is di=Di/Dt. There are thus
only three parameters: a , b and qsat (with S gmax thereby
implied from its storage function). It is seen that, for a 0.5
saturated soil store, recharge is diminished when the
groundwater demand factor is less than a, when the soil
ceases to be freely draining. This formulation derives from
a reparameterised form of percolation model used in the V(t+ǻt) cmax
National Weather Service rainfall–runoff model (Burnash
et al., 1973; Gupta and Sarooshian, 1983). 0
A third recharge formulation is available which assumes 50 100 150
that there is no soil drainage, di. Direct runoff is split between *
C (t)
*
C (t+ǻt)
a fraction a which goes to make up surface runoff and a Storage capacity, c mm
fraction (1–a) going to groundwater storage. Fig. 3. The storage capacity distribution function used to calculate
basin moisture storage, critical capacity, and direct runoff
With both losses to evaporation and recharge, the net
according to the probability-distributed interacting storage capacity
rainfall, pi , may be defined in general as model.

486
 The PDM rainfall-runoff model

A specific application of the procedure can be developed

for a given choice of probability density function. Analytical ^
V t  't  S i 't  S max 1  C * (t ) / cmax 
b1
 1  C * t  't  / cmax 
b1
`.
solutions of the integrals in the probability-distributed (20d)
storage capacity model component (specifically Eqns. (6)
and (7)) are presented in Appendix A for a range of possible The relationship between rainfall and runoff implied by
distribution types. After a number of trials on alternative the above expressions, for given conditions of soil moisture,
distributions, a Pareto distribution of storage capacity is now is presented in Fig. 5. A related, if not similar, procedure
most widely used in practice and will be used here to forms the basis of the Xinanjiang model developed by Ren
illustrate application of the method. The distribution function Jun Zhao and co-workers in China (Zhao and Zhuang, 1963;
and probability density function for this distribution are Zhao et al., 1980; Zhao and Liu, 1995). More recently it
has been popularised and extended in the form of the Arno
b rainfall–runoff model in Italy (Todini, 1996) and, for large-
§ c ·
F c  1  ¨¨1  ¸¸ 0 d c d c max (18) scale applications, the VIC land surface model in the USA
© c max ¹ (Wood et al., 1992). Indeed, Moore (1985) traces the origins
b 1 of such probability-distributed principles in hydrology back
dF c  b § c ·
f c  ¨¨1  ¸¸ 0 d c d c max (19) to the pioneering contribution of Bagrov in 1950, working
dc cmax © c max ¹ in what was then the USSR.
Extended forms of these relations are available in the PDM
where parameter cmax is the maximum storage capacity in toolkit to represent the case where storage capacity is limited
the basin, and parameter b controls the degree of spatial by both the upper bound cmax and a lower bound cmin such
variability of storage capacity over the basin. These that cmin d c d cmax in the Pareto distribution (see Appendix
functions are illustrated in Fig. 4: note that the rectangular A and Moore and Bell, 2002). Note that when C*(t) drops to
distribution is obtained as a special case when b=1, and the value of cmin as a result of evaporation, the basin water
b=0 implies a constant storage capacity over the entire basin. storage S(t) also equals this value. Equation (15) is used to
The following relations apply for Pareto distributed storage update S(t) during ‘dry periods’ allowing it to fall below
capacities: c min and ultimately to zero. Only once net rainfall has
replenished water storage above cmin does runoff generation
S max cmax / b  1 , (20a) occur and the calculation of C*(t) and V(t) is resumed. Thus
the initial catchment response to rainfall can be moderated
S t  ^
S max 1  1  C * t  / c max 
b 1
`, (20b) or delayed if the minimum storage capacity is taken to be
greater than zero.
^
C * t  c max 1  1  S t  / S max 
1 /  b 1
`, (20c)

0.02 1.0

b=2
Probability
density Distribution
function function
f(c) F(c)
b=0
0.01 b=2 0.5

b=0.5
b=1

b=0.5 cmax

0 0
0 100 200 0 100 200
Storage capacity, c mm Storage capacity, c mm

(a) Probability density function (b) Distribution function

Fig. 4. The Pareto distribution of storage capacity.

487
R.J. Moore

100

200
cmax = 140
80
Direct b = 0.4 Initial
Runoff 60 Storage
V mm (Smax = 100) S mm
40

20

100
0

0 100 200
Net rainfall, P-E mm
Smax

Fig. 5. Rainfall-runoff relationship for the probability-distributed interacting storage capacity model, using the Pareto
distribution of storage capacity.

where S { S (t ) is the volume of water held in the storage

per unit area, k is the storage rate coefficient and m the store
Surface and subsurface storages exponent. Note that a=mk1/m and b=(m–1)/m.
The probability-distributed store model partitions rainfall Recursive solutions of the Horton-Izzard equation are
into direct runoff, groundwater recharge and soil moisture provided in the PDM for a choice of non-linear storage form:
storage. Direct runoff is routed through surface storage: a linear, quadratic, cubic, exponential and general non-linear.
‘fast response system’ representing channel and other fast A cubic form is usually considered most appropriate to
translation flow paths. Groundwater recharge from soil water represent the groundwater storage. In this case where
drainage is routed through subsurface storage: a ‘slow q k S 3 , an approximate solution utilising a method due
response system’ representing groundwater and other slow to Smith (1977) yields the following recursive equation for
flow paths. Both routing systems can be defined in the PDM storage, given a constant input u over the interval (t, t+Dt):
by a variety of non-linear storage reservoirs or by a cascade
of two linear reservoirs (expressed as an equivalent second
order transfer function model constrained to preserve S t  't  S t  
1
3kS 2 t 
^exp 3kS 2 t 't   1`u  kS 3 t  .
continuity). (24)
The nonlinear storage model is specified by the Horton-
Izzard equation (Dooge, 1973) Discharge may then be obtained simply using the nonlinear
dq relation
= au  q  q b , q>0 ,  f<b<1, (21)
dt
qt  't  k S 3 t + 't  . (25)
where q { q(t ) is the rate of outflow and u { u (t ) is the
rate of inflow to the store per unit area, and a and b are Solutions for the other nonlinear forms are presented in
parameters. This derives from the continuity equation for Moore (1983) and Appendix A of Moore and Bell (2002).
the store When used to represent groundwater storage, the input u
dS will be the drainage rate, di, from the soil moisture store
= u  q, (22) and the output q(t) will be the ‘baseflow’ component of flow
dt
qb(t). Explicit allowance for groundwater abstractions is
and the nonlinear storage form of the momentum equation incorporated in an extended version of the PDM which can
also make use of well level data (Moore and Bell, 2002).
q k Sm, k>0 , m>0 (23) The most commonly used representation of the surface
storage component is a cascade of two linear reservoirs,

488
 The PDM rainfall-runoff model

with time constants k1 and k2, expressed as the discretely Here Dt is the time interval between times t–1 and t and it is
coincident transfer function model (O’Connor, 1982): assumed that the input ut is constant over this interval. In
this case the input is the volume of direct runoff, V(t),
qt  G 1 q t 1  G 2 q t - 2 + Z 0 u t + Z 1 u t 1 (26) generated from the probability-distributed soil moisture store
and the output qt will be the surface flow component of the
with total basin runoff, qs(t). The total basin flow is given by
qs(t) + qb(t), plus a constant flow, qc, representing any returns
G 1 G 1 + G 2 , G 2 G 1 G 2 , G 1 exp 't k1 , G *2 exp 't k 2  or abstractions.
* * * * *

k 1 G 1  1  k 2 (G 2  1) Model parameters, calibration and

* *

Z0 k1 z k 2
k 2  k1
example application
The parameter and structure options in the model are
k 2 G 2  1G 1  k 1 G 1  1G 2
* * * *
Z1 k1 z k 2 summarised in Table 1. Note that a rainfall factor, fc, is
k 2  k1
1  1+ 't k1 G *1
incorporated in the model to allow conversion of a rainfall
Z0 k1 k 2
observation to rainfall, P, thereby compensating for effects
such as lack of raingauge representativeness. The time
Z1 G *
1  1+ 't k1 G *1 k1 k 2 . (27)
constants kg and kb are equivalent to k-1 in the general non-

Table 1. PDM model parameters

Parameter name Unit Description

fc none rainfall factor

td hour time delay
Probability-distributed store
cmin mm minimum store capacity
cmax mm maximum store capacity
b none exponent of Pareto distribution controlling spatial
variability of store capacity
Evaporation function
be none exponent in actual evaporation function
Recharge function
1: Standard
kg hour mm b g  1 groundwater recharge time constant
bg none exponent of recharge function
St mm soil tension storage capacity
2: Demand-based
a none groundwater deficit ratio threshold
b none exponent in groundwater demand factor function
qsat mm h-1 maximum rate of recharge
3: Splitting
a none runoff factor controlling the split of rainfall to surface
and groundwater storage routing when no soil recharge
is allowed
Surface routing
k1, k2 hour time constants of cascade of two linear reservoirs
Groundwater storage routing
kb hour mmm-1 baseflow time constant
m none exponent of baseflow non-linear storage
Constant flow addition
qc m3 s-1 constant flow representing returns/abstractions

489
R.J. Moore

linear storage function q=kSm. can be used to look at the response over seasons or years as
The calibration of the PDM model is carried out within a well as zoomed-in to flood hydrographs of special interest.
generic Model Calibration Shell environment. This The objective function used to assess model performance
Calibration Shell provides for both automatic optimisation can be censored to exclude flows below a minimum value
and informal visually-interactive parameter estimation. The so that it is not unduly influenced by long hydrograph
former uses a simplex direct search procedure (Nelder and recessions.
Mead, 1965) modified following suggestions made by Gill The PDM model has been widely applied to a variety of
et al. (1981). Informal estimation is supported by an catchments in different countries. These include England,
interactive visualisation tool which allows the user to see Wales, Scotland, Belgium, Hong Kong, India, South Java,
the changing hydrograph response as a chosen parameter Thailand and China. An example of its use as a simulation
value is varied. This can be an invaluable aid to model is shown in Fig. 6. The flows are for Beverley Brook
understanding the model response and the nature of the gauged at Wimbledon Common in London within the
dependence between parameters. Error response function Thames basin, draining an area of 44 km 2. HYRAD
plots for a selected pair of parameters can also be used to recalibrated radar rainfall data (Moore, 1999), which
investigate parameter interdependence. combines radar and raingauge information, is used to form
The seasonal response of a model, dominated by aspects the catchment average rainfall employed as input to the
of the model structure and a certain subset of the model model in this case.
parameters controlling the water balance, can be investigated
through a preliminary calibration at a daily time-step. A 15-
Summary of the PDM rainfall-runoff
minute time-step can then be used to establish the model
parameters dominating the short-term dynamics of runoff model
response and translation. Storm events spanning several Whilst the design of the PDM model provides for a range
years can be included in the optimisation at a 15-minute of different structural forms, the most commonly used
time-step. This is achieved by prescribing the event periods configuration comprises the following components:
to be included in the optimisation and switching to a daily
time-step between events, for purposes of continuous water (i) A probability-distributed soil moisture storage
accounting. Other time-steps to the 15-minute and daily component to effect separation between direct runoff
interval can be used. If long (several years) continuous 15- and subsurface runoff. This is based on a Pareto
minute datasets are available, then modelling can be done distribution of soil moisture storage capacity over the
at this interval. The Calibration Shell visualisation facilities catchment.

Fig. 6. PDM model simulation of Beverley Brook at Wimbledon Common using HYRAD recalibrated radar rainfall data as input. Observed
flow: bold line; simulated flow: dashed line; “baseflow”: small dots. The negative ordinate shows soil moisture deficit as a dashed line and
rainfall on a proportional scale.

490
 The PDM rainfall-runoff model

(ii) A surface storage component which transforms direct operating in updating mode. These two updating methods
runoff to surface runoff. This employs a two-linear are available for use with the PDM and are detailed below.
reservoir cascade formulated as a transfer function with
dependence on two past outputs and the current and
STATE CORRECTION
previous input. The coincidence of this four parameter
discrete time model with the two parameter continuous The term ‘state’ is used to describe a variable of a model
time model that preserves continuity allows the transfer which mediates between inputs to the model and the model
function model to be (a) re-parameterised as a two output (Szollosi-Nagy, 1976). In the case of the PDM
parameter model, and (b) to be used at different discrete rainfall-runoff model, the main input is rainfall and basin
time intervals. flow is the model output. Typical state variables are the water
contents of the surface and groundwater stores, S2 and S3,
(iii) A groundwater storage component which receives and of the probability-distributed soil storage, S1 (Fig. 1).
drainage water from the distributed soil moisture storage The flow rates out of the conceptual stores can also be
as input and contributes the groundwater component regarded as state variables: examples are qs, the flow out of
of total runoff as output. A cubic non-linear storage the surface storage, and qb, the flow out of the groundwater
routing function is adopted to effect this transformation. storage.
When an error, H Q  q Q  (qs  qb ) , occurs between the
Methods are available to update the PDM model with model prediction, q, and the observed value of basin runoff,
reference to observed flows for real-time flow forecasting Q, it would seem sensible to ‘attribute the blame’ to mis-
applications. An empirical state-correction scheme provides specification of the state variables and attempt to ‘correct’
a range of options for correcting internal model water the state values to achieve concordance between observed
contents or flow rates to yield more accurate updated model and model predicted flow. Mis-specification may, for
forecasts. As an alternative, an ARMA error-prediction example, have arisen through errors in rainfall measurement
scheme is available which exploits the persistence in model which, as a result of the model water accounting procedure,
errors to obtain improved forecasts. These methods are are manifested through the values of the store water contents,
described in outline next. or equivalently the flow rates out of the stores. A formal
approach to state correction is provided by the Kalman filter
algorithm (Jazwinski, 1970; Gelb, 1974; Moore and Weiss,
Updating the PDM rainfall-runoff 1980a,b). This provides an optimal adjustment scheme for
model incorporating observations, through a set of linear
operations, for linear dynamic systems subject to random
INTRODUCTION variations which may not necessarily be Gaussian in form.
If observed flows are not used, except for initialisation, a For non-linear dynamic models, such as the PDM, an
model is said to be operating in simulation mode, acting as extended form of Kalman filter based on a linearisation
a function which transforms rainfall and potential approximation is required which is no longer optimal in the
evaporation to river flow. A model which has been calibrated adjustment it provides. The implication of this is that simpler,
in simulation mode may be extended to use observed flows intuitive adjustment schemes can be devised which
by addition of further structure and associated parameters. potentially provide better adjustments than the more
These might take the form of rules for adjusting model states complex and formal extensions of the Kalman filter which
(state-correction) or predicting future errors (error- accommodate non-linear dynamics through approximations.
prediction). The former are heavily dependent on the Such schemes which make physically sensible adjustments
structure of the simulation mode model, whilst the latter are called ‘empirical state adjustment schemes’. A simple
are essentially independent. Parameter-adjustment is not example is the apportioning of the error, H , between the
considered in the PDM. The view is taken that this approach surface and groundwater stores of the PDM in proportion
confuses the issue of correct model identification, which is to their contribution to the total flow. Mathematically this
properly carried out through a controlled calibration may be expressed as
procedure. Parameter variability is better addressed by
improving the structural form of the model than by tracking qb* qb  D gbH (28a)
its variation in real-time.
A model incorporating observed flows either through qs* qs  1  D g sH (28b)
state-correction or error-prediction will be said to be

491
R.J. Moore

where in accuracy at longer lead times. This behaviour appears to

be associated with a combination of some or all of the
D qb qs  qb  (29) following: large gain factors, time lags between the
correction of a state value and the appearance of an effect
and the superscript * indicates the value after adjustment. on the modelled flow, and rapid increases in the model error
The ‘gain’ coefficients, gb and gs, when equal to unity, yield (often due to timing errors on the rising limb). The latter is
the result that q*b  q*s equals the observed flow, Q, thus also a problem for error prediction schemes. Optimal values
achieving exact correction of the model flow to equal the for the gain factors tend to be greater than unity (over-
observed value. Values of the coefficients other than unity relaxation) whilst time lags can occur because correction of
allow for different adjustments to be made, and gb and gs soil moisture may not affect runoff until the next wet period.
can be regarded as model parameters whose values are
established through optimisation to achieve the ‘best’ fit
between state-adjusted forecasts and observed flows. A ERROR PREDICTION
generalisation of the above is to define a to be State-correction techniques have been developed based on
qb adjustment of the water content of conceptual storage
D (30) elements in the belief that the main cause of the discrepancy
E1qs  E 2 qb
between observed and modelled runoff will arise from errors
and to choose the incidental parameters b1 and b2 to weight in estimating basin average rainfall, which in turn
the apportionment towards or away from one of the flow accumulate as errors in water storage content. Rather than
components; in practice b1 and b2 are assigned values of 10 attribute the cause directly and devise empirical adjustment
and 1.1 to apportion more of the error adjustment to the procedures the structure of the errors may be analysed and
surface store. Note that the adjustment is carried out at every predictors of future errors developed based on this structure
time step and the time subscripts have been omitted for which can then be used to obtain improved flow forecasts.
notational simplicity. The scheme with a defined by Eqn. A feature of errors from a conceptual rainfall-runoff model
(29) is referred to as the proportional adjustment scheme is that there is a tendency for errors to persist so that
and that defined by Eqn. (30) is the super-proportional sequences of positive errors (underestimation) or negative
adjustment scheme. Replacing a and (1–a) in Eqns. (28a,b) errors (overestimation) are common. This dependence
by unity yields the simplest non-proportional adjustment structure in the error sequence may be exploited by
scheme. developing error predictors which incorporate this structure
An adjustment to the probability-distributed soil moisture, and allow future errors to be predicted. Error-prediction is
S { S1 , may also be made, either of the proportional form now a well established technique for forecast updating in
real time (Box and Jenkins, 1970; Moore, l982). Error-
S* S D gg H (28c) prediction is available as an alternative to empirical state
correction in the PDM. Predictions of the error are added to
or the direct form of gain with a equal to unity. the deterministic model prediction to obtain the updated
It should be noted that all the above forms of adjustment model forecast of flows. In contrast to the state correction
utilise the same basic form of adjustment employed by the scheme, which adjusts values internally within the model,
Kalman filter in which an updated state estimate is formed the error prediction scheme is wholly external to the
from the sum of the current state value and the model error deterministic model operation. The importance of this is that
multiplied by a gain coefficient. However, instead of error prediction may be used in combination with any model,
defining the gain statistically, as the ratio of the uncertainty be it of TF, conceptual or ‘physics-based’ form, and for
in the observation to that of the current state value, it is first representing rainfall-runoff or channel flow processes.
related to a physical apportionment rule multiplied by a gain Consider that qt " is the forecast of the observed flow,
factor. This gain factor acts as a relaxation coefficient which Qt  " , at some time t+ " , made using the PDM rainfall-runoff
is estimated through an off-line optimisation using past flood model. Since qt " will have essentially been obtained by
event data. transformation of rainfall into flow through the PDM model
State-correction is essentially a form of negative feedback conceptualisation of the catchment, it will not have used
and, although usually very effective, this feedback can previous observed values of flow, except for the purposes
sometimes give rise to an over- or under-shooting behaviour of model initialisation. It will consequently be referred to
characterised by high accuracy at short lead times but with as a simulation-mode forecast to distinguish it from a real-
degraded accuracy at moderate lead times before a recovery time, updated forecast which incorporates information from

492
 The PDM rainfall-runoff model

observed flows. Kt  " t I1 Kt "1 t  I 2 Kt  " 2 t  I K p t  " p t

The error, Kt " , associated with this simulation-mode
forecast is defined through the relation
 T1at " 1 t T 2 at " 2 t   T a q t " q t , " 1,2,  (37)
where
Qt  " qt  "  Kt  " . (31) ­0 "i ! 0
at  " i t ® (38)
¯at  " i otherwise
If the simulation-mode error Kt " may be predicted using an
error predictor which exploits the dependence structure of
and at "i is the one-step ahead prediction error
these errors, then an improved forecast may be obtained.
Let K t " t denote a prediction of the simulation-mode error,
at  "  i { at  " i t  "  i 1 Kt  " i  Kt  " i t  " i 1
Kt  " , made " steps ahead from a forecast origin at time t
Qt  " i  qt  "  i t  " i 1 , (39)
using an error predictor. (The suffix notation t  " t should
be read as a forecast at time t  " given information up to
and
time t.) Then a real-time forecast, qt  " t , made " time units
ahead from a forecast origin at time t may be expressed as Kt  "  i t Kt  "  i Qt  "  i  qt  "  i for "  i d 0 . (40)
follows:
q q K . (32) The prediction Eqn. (37) is used recursively to produce the
t " t t " t " t
error predictions Kt 1 t , Kt  2 t , ..., Kt  " t , from the available
The real-time forecast error is values of at, at-1, ... and Kt , Kt 1 , ... .
Using this error predictor methodology, the PDM model
at  " t Qt  "  qt  " t (33) simulation-mode forecasts, qt+l, may be updated using the
error prediction Kt  " t , obtained from Eqn. (37) (and the
which, depending on the performance of the error predictor, related Eqns. (38–40), to calculate the required real-time
should be smaller than the simulation-mode forecast error forecast, qt  " t , according to Eqn. (32). Note that this real-
time forecast incorporates information from the most recent
Kt  " Qt  "  qt  " . (34) observations of flow through the error predictor, and
specifically through calculation of the one-step ahead
Turning now to an appropriate form of error predictor it is forecast errors, at "i , according to Eqn. (39).
clear that a structure which incorporates dependence on past Alternative error predictor schemes may be devised by
simulation-mode errors is required. Thus the autoregressive working with other definitions of the basic errors: for
(AR) model example by using proportional errors. One such scheme can
be formulated by starting with the logarithmic model
Kt I1 Kt 1  I 2 Kt  2    I z Kt  z  at (35)
log Qt  " log qt  "  Kt  " (41)
is an obvious candidate, where at is the residual error
(uncorrelated), and { Ii } are parameters. However, a more so that the simulation-mode error is now defined as
parsimonious form of model is of the autoregressive-moving
average (ARMA) form Kt  " logQt  " q t  "  . (42)

An error predictor for Kt 1 may be formulated in the normal

Kt I1 Kt 1  I2 Kt  2    I p K t  p
way using Eqns. (37) and (38) with the one-step ahead
 T1at 1  T 2 at  2    T q at  q  at (36)
prediction error given by
which incorporates dependence of past residual errors, at-1,
at-2, ... . at  " i Kt  " i  Kt  "  i t  " i 1 . (40)
In general, the number of parameters p+q associated with
the ARMA model will be less than the number z associated Instead of Eqn. (32) the real-time forecast, qt  " t , takes the
with the AR model, in order to achieve as good a level of form
approximation to the true simulation-mode error structure.
The ARMA model may be used to give the following error qt  " t  
qt  " exp Kt  " t . (41)
predictor:
The PDM uses automatic optimisation to estimate the ARMA

493
R.J. Moore

error-predictor parameters for a chosen model structure. These early developments employed convolution to route
Often a third order autoregressive, with dependence on three direct runoff (and drainage to recharge) to the basin outlet,
past model errors, provides an appropriate choice for UK using the probability-distributed principle applied to the
conditions and a 15 minute model/data time interval. time-of-travel to the catchment outlet. Specifically, the
probability density function (pdf) of translation time was
taken as equivalent to the kernel function or instantaneous
DISCUSSION OF UPDATING THE PDM unit hydrograph. Invoking a pdf of inverse Gaussian form
Whilst error-prediction provides a general technique which provided a physical link with the St. Venant equations of
is easy to apply, its performance in providing improved open-channel flow and the convection-diffusion equation.
forecasts will depend on the degree of persistence in the It was not until 1986 that the model schematic of Fig. 1 first
model errors. Unfortunately, in the vicinity of the rising limb appeared (see Fig. 3 of Moore, 1986). The parallel routing
and peak of the flood hydrograph this persistence is least formulation made explicit by this schematic was inspired
and errors show a tendency to oscillate rapidly and most by the simplified catchment model schematic of Dooge (Fig.
widely; dependence is at its strongest for errors on the falling 1-8 in Dooge, 1973; Fig. 2.3 in Dooge and O’Kane, 2003).
limb, where improved forecast performance matters least. Here, soil moisture feedback controls the partitioning of the
In addition, timing errors in the model forecast may lead to direct storm and groundwater response of total runoff,
erroneous error predictions being made, a problem which deliberately avoiding the less identifiable separation of
is also shared by the technique of state-correction. The overland flow and interflow. Convolution was replaced by
general applicability and popularity of error-prediction as non-linear storage routing as being simpler and easier to
an updating tool commends its use as an ‘off-the-shelf’ update for real-time application. The exposition of the
technique, but empirical state adjustment schemes should Horton-Izzard equation in Dooge (1973), and reference to
also be considered as viable alternatives to the use of error- the source and related texts, provided the foundation for
prediction. State-correction is normally the preferred choice the recursive solutions to this equation used for storage
with the PDM model. routing in the PDM (Moore, 1983); Moore and Bell (2002)
provides further details and reference sources. This was
supplemented by the transfer function representation of a
Historical perspective and concluding cascade of two linear reservoirs efficiently parameterised
via one or two storage coefficients (O’Connor, 1982) as a
remarks further routing option.
The PDM rainfall-runoff model had its origins in the search This early development of a generalised rainfall-model
for model formulations that were well suited to automatic toolkit was undertaken deliberating without recourse to a
parameter estimation: specifically, models that avoided ‘brand name’ association, recognising that many models
threshold behaviours and were parsimonious of parameters. share rather similar conceptual elements. However, with the
Initially, this focused on invoking a probability distribution need to develop the River Flow Forecasting System (RFFS)
of storage capacity to replace a simple single store for operational use, the requirement for a software product
representation of runoff production (Moore and Clarke, arose and the PDM brand name introduced (Moore and
1981). The threshold behaviour of the single store results Jones, 1991; Institute of Hydrology, 1992).
in a discontinuity in the gradient of the model’s objective As the PDM toolkit evolved to represent a wider range of
function and difficulties of parameter optimisation. These hydrological behaviours it became clear that the original
problems were circumvented by the probability-distributed aim of a model well suited to automatic parameter estimation
representation and, more importantly, led to a more realistic had been compromised. It became evident that a rainfall-
representation of runoff production across a catchment. A runoff model that performed well was typically not easy to
simplification of the theory (Moore, 1985) — allowing optimise automatically; the parallel configuration of routing
storage elements to interact with each other so as to equalise stores and more complex recharge functions particularly
the depth of stored water across a catchment — led to the introduced parameter interdependence. Thus a more realistic
formulations now employed in the PDM. Over time, representation of hydrological behaviour is likely to have
solutions were obtained for a range of distributions including problems of parameter estimation. An acceptance of this
exponential, Pareto (reflected power), rectangular, position has led to increasing reliance being placed on
triangular, power and lognormal (see Appendix A). The manual calibration using interactive visual support tools,
theory also allowed for representations of evaporation and using automatic optimisation only for late-stage refinement.
drainage to recharge as a function of catchment soil moisture. It is not clear that recent advances in calibration of rainfall-

494
 The PDM rainfall-runoff model

runoff models (Duan et al., 2003) will change this position Institute of Hydrology, 1992. PDM: A generalized rainfall-runoff
model for real-time use, Developers’ Training Course. National
radically, although the ability to search the parameter space
Rivers Authority River Flow Forecasting System, Version 1.0,
more comprehensively with increased computer power may March 1992, 26pp.
help. Institute of Hydrology, 1996. A guide to the PDM. Version 1.0.
Updating of the PDM for real-time forecasting Wallingford, UK. 45pp.
Jazwinski, A.H., 1970. Stochastic processes and filtering theory.
applications benefited from experience gained in the 1970s Academic Press, New York, USA. 376 pp.
and 1980s (Moore and Weiss, 1982a,b; Moore, 1982, 1986). Moore, R.J., 1982. Transfer functions, noise predictors and the
This experience concerned the recursive state-parameter forecasting of flood events in real-time. In: Statistical analysis
of rainfall and runoff, V.P. Singh (Ed.), Water Resources Publ.,
estimation techniques based on the Kalman filter (Jazwinski, Littleton, Colorado, USA. 229–250.
1970) and the transfer function noise models of ARMA form Moore, R.J., 1983. Flood forecasting techniques. WMO/UNDP
popularised by Box and Jenkins (1970). Formal Kalman Regional Training Seminar on Flood Forecasting, Bangkok,
Thailand. 37pp.
filter techniques for state-correction were abandoned in
Moore, R.J., 1985. The probability-distributed principle and runoff
favour of the simpler empirical state-correction methods production at point and basin scales. Hydrol. Sci. J., 30, 273–
described here. The ARMA noise models were used as the 297.
basis of the error-predictor technique available as an Moore, R.J., 1986. Advances in real-time flood forecasting
practice. Symposium on Flood Warning Systems, Winter meeting
alternative method of updating the PDM forecast. The of the River Engineering Section, Inst. Water Engineers and
parameters of these updating methods are well suited to Scientists, 23pp.
automatic optimisation, once the PDM simulation model Moore, R.J., 1999. Real-time flood forecasting systems:
Perspectives and prospects. In: Floods and landslides:
parameters have been first estimated. Integrated Risk Assessment, R. Casale and C. Margottini (Eds.),
In spite of its widespread application, the PDM has not Springer, Berlin, Germany. Chapter 11, 147–189.
been formally published in extensive form in the open Moore, R.J. and Bell, V.A., 2002. Incorporation of groundwater
losses and well level data in rainfall-runoff models illustrated
literature. The detail has been confined to the manuals and using the PDM. Hydro. Earth System Sci., 6, 25–38.
guides provided with the software product (CEH Moore, R.J., Bell, V.A. and Jones, D.A., 2005. Forecasting for
Wallingford, 2005). Only rather general reviews of the PDM flood warning. C. R. Geoscience, 337, 203-217.
have been given in published papers on flood forecasting Moore, R.J. and Clarke, R.T., 1981. A distribution function
approach to rainfall-runoff modelling. Water Resour. Res., 17,
techniques (Moore, 1999; Moore et al., 2005). This paper 1367–1382.
has rectified this omission. Moore, R.J. and Jones, D.A., 1991. A river flow forecasting system
for region-wide application. Invited paper, MAFF Conference
of River and Coastal Engineers 1991, 8–10 July 1991,
References Loughborough University, 12pp.
Moore, R.J. and Weiss, G., 1980a. Recursive parameter estimation
Box, G.E.P. and Jenkins, G.M., 1970. Time series analysis of a non-linear flow forecasting model using the extended
forecasting and control, Holden Day, San Francisco, 553pp. Kalman filter. In: Real-time hydrological forecasting and
Burnash, R.J.C., Ferral, R.L. and McGuire, R.A., 1973. A control, P.E. O’Connell, (Ed.), Proc. 1st Int. Workshop, July
generalized streamflow simulation system: conceptual modelling 1977, Institute of Hydrology, Wallingford, UK. 264pp.
for digital computers. Report of the Joint Federal State River Moore, R.J. and Weiss, G., 1980b. Real-time parameter estimation
Forecast Centre, U.S. National Weather Service and California of a nonlinear catchment model using extended Kalman filters.
Department of Water Resources, Sacramento. In: Real-time forecasting/control of water resource systems, E.G.
CEH Wallingford, 2005. PDM Rainfall-Runoff Model. Version 2.2, Wood and A. Szollosi-Nagy (Eds.), Pergamon Press, Oxford.
Centre for Ecology & Hydrology, Wallingford. (Includes Guide, 83–92.
Practical User Guides, User Manual and Training Exercises). Nelder, J.A. and Mead, R., 1965. A simplex method for function
Dooge, J.C.I., 1973. Linear theory of hydrologic systems. Tech. minimisation. Computer J., 7, 308–313.
Bull. 1468, Agric. Res. Service, US Dept. Agric., Washington, O’Connor, K.M., 1982. Derivation of discretely coincident forms
327 pp. of continuous linear time-invariant models using the transfer
Dooge, J.C.I and O’Kane, J.P., 2003. Deterministic methods in function approach. J. Hydrol., 59, 1–48.
systems hydrology. Balkema, The Netherlands. 309pp. Smith, J.M., 1977. Mathematical Modelling and Digital Simulation
Duan, Q., Gupta, H.V., Sorooshian, S., Rousseau, A.N., Turcotte, for Engineers and Scientists, Wiley, New York, USA. 332pp.
R. (Eds.) 2003. Calibration of watershed models. Water Science Szollosi-Nagy, A., 1976. Introductory remarks on the state space
and Application 6, American Geophysical Union, Washington, modelling of water resource systems. Int. Inst. for Applied
USA. 345pp. Systems Analysis, RM-76-73, 81pp.
Gelb, A. ed., 1974. Applied optimal estimation. MIT Press, Todini, E., 1996. The ARNO rainfall-runoff model. J. Hydrol.,
Cambridge, USA. 374 pp. 175, 339–382.
Gill, P.E., Murray, W. and Wright, M.H., 1981. Practical Wood, E.F., Lettenmaier, D.P. and Zartarian, V.G., 1992. A land-
optimisation. Academic Press, London, 401pp. surface hydrology parameterisation with subgrid variability for
Gupta, V.K. and Sarooshian, S., 1983. Uniqueness and general circulation models. J. Geophys. Res, 97(D3), 2717–
observability of conceptual rainfall-runoff model parameters: 2728.
The Percolation process examined. Water Resour. Res., 19, 269–
276.

495
R.J. Moore

Zhao, R.J. and Liu, X.R., 1995. The Xinanjiang model. In: Zhao, R.J., Zhuang, Y., Fang, L.R., Lin, X.R. and Zhang, Q.S.,
Computer models of watershed hydrology, V.P. Singh (Ed.), 1980. The Xinanjiang model. In: Hydrological Forecasting
Water Resources Publications, Colorado, USA. 215–232. (Proc. Oxford Symp., April 1980), IAHS Publ. no. 129, 351–
Zhao, R.J. and Zhuang, Y., 1963. Regionalisation of the rainfall- 356.
runoff relations (in Chinese). Proc. East China College of
Hydraulic Engineering.

Appendix A:
Probability-distributed storage models
Analytical solutions of the integrals occurring in the probability-distributed storage model component are
presented below for a number of different distribution types.

PARETO DISTRIBUTION
b
§ c c ·
F c  1  ¨¨ max ¸¸ c min d c d cmax
© cmax  cmin ¹
b 1
b § cmax  c ·
f c  ¨¨ ¸¸ cmin d c d cmax
cmax  cmin © cmax  cmin ¹
f bcmin  cmax
c S max ³ 1  F c dc
o b 1

S t  ³o
C* ( t )
1  F c dc ^
cmin  (c  cmin ) 1  ( cmax  C * (t )) ( cmax  cmin ) 
b1
`
^
C * t  cmin  (cmax  cmin ) 1  ( S max  S (t )) (c  cmin ) 
1/ b 1
`
V t  't  S i 't  ( S (t  't )  S (t ))

RECTANGULAR DISTRIBUTION
c
F c 
cmax  cmin
1
f c 
cmax  cmin
f cmax ( cmax / 2)  cmin 
c S max ³ 1  F c dc
o cmax  cmin
C * t  ª C * t  º
S t  ³ 1  F (c) dc C * t  «1  »
o
¬ 2( cmax  cmin ) ¼

496
 The PDM rainfall-runoff model

ª ª 2 S t  º º
1/ 2

C t  ( cmax
*
 cmin ) «1  «1  » »
«¬ ¬ cmax  cmin ¼ »¼
C * t  't 
V t  't  ³ F c dc
C * t 

(i) C * t  't  d cmax V t  't 

C
*2
t  't   C * t 
2(cmax  cmin )

(ii) C * t  't  ! cmax V t  't  S i 't  Smax  S t 

EXPONENTIAL DISTRIBUTION

F c  1  exp c / c 

f c  c 1 exp c / c 
f
S max ³ 1  F c dc
o
c

S t 
C * t 
³ 1  F c dc
o
 
c 1  exp  C * t  / c 
C * t  c log 1  S t  / c 

C *  t  't 
V t  't  ³ F c dc
C * t 

 
S i 't  c exp  C * t  't / c   exp  C * t  / c  

TRIANGULAR DISTRIBUTION

­ ª cc º
2

° 2« min
» cmin d c d c
° ¬ cmax  cmin ¼
°
F c  ®
° ª cmax  c º
2

°1  2 « » c d c d cmax
°¯ ¬ cmax  cmin ¼

­ 4cmax  c 
° cmin d c d c
° cmax  cmin 
2

°
°° 4c  c 
f c  ®
max
c d c d cmax
° cmax  cmin 
2

°
°
°
°¯0 otherwise
497
R.J. Moore

1
S max cmin  cmax  cmin  c
2
C * t 
S t  ³ 1  F c dc
2cmax  C * t 
o 3

(i) c d C * t  d cmax S t  c
3cmax  cmin 
2

cmin d C * t  d c S t  C * t  

2 C * t   cmin 
3

(ii)
3cmax  cmin 
2

C*(t) may be obtained by solving the expression for S(t) above, for example by Newton-Raphson as
shown later, or as the solution of a cubic equation in C*(t).
C * t  't 
V t  't  ³ F c  dc
C * t 

(i) C * t  't  d cmin V t  't  0

2C * t  't   cmin 
3

cmin d C t  't  d c V t  't   C * t   S t 

* 3
(ii)
3cmax  cmin 
2

(iii) c d C * t  't  d cmax

V t  't  C * t  't   c 

2 cmax  C * t  't  3

 C * t   S t  
3cmax  cmin 
2

(iv) C * t   S i 't d cmax V t  't  S i 't  S max  S t  .

LOGNORMAL DISTRIBUTION

ª log c  ] º 1 ª  log c  ]  º
F c  ) « » erfc « »
¬ V ¼ 2 ¬ V 2 ¼
where F(.) is the standardised normal distribution function

1 ª u2 º
) x 
x

2S ½ ³o «¬ 2 »¼ du,
exp

and parameters z and s are the mean and standard deviation of the logarithms of storage capacity.

1 ­ log c  ] 2 ½
f c  exp® ¾.
cV 2S  2V 2
½
¯ ¿
498
 The PDM rainfall-runoff model

f ª V2º
S max ³o 1  F c  dc c exp «] 
¬
»
2 ¼

C * t 
S t  ³ 1  F c dc
o

1ª § V2 · § V 2 ·º
C * t   «C * t erfcuo   exp 2 ¨¨ ]  ¸¸ erfc ¨¨ uo  ¸»
¸
2 ¬« © 2 ¹ © 2 ¹¼»
§ log C * t   ] ·
where uo ¨¨ ¸¸ .
© V 2 ¹

C*(t) may be obtained by solving the above expression for C*(t), for example by Newton-Raphson
(as shown later).

C * t  't 
V t  't  ³ F c dc
C * (t )

1ª * § V2 · § V 2 ·º *
«
2 «¬
C t  't  erfcu1   exp¨
¨ ]  ¸
2 ¸¹
erfc ¨ u1 
¨ ¸  
¸ » C t   S t 
© © 2 ¹ »¼

§ log C * t  't   ] ·
where u1  ¨¨ ¸¸ .
© V 2 ¹

NEWTON-RAPHSON SOLUTION OF S(t) FOR C*(t)

The Newton-Raphson solution of

C * t 
S t  ³ 1  F c dc
o

for C*(t) is based on the iterative scheme

T i 1 T i  f T i  / f ´T i 

where T i is the estimate of C*(t) at the i’th iteration, and

Ti
f T i  S t   ³ 1  F c dc
o

f ´T i  wf T i  / wT i 1  F T i  .

Thus the following iterative solution results

^ Ti
T i 1 T i  S t   ³ 1  F c  dc /  F T i   1
o
`
where q0 is set to the best initial estimate of C*(t); qn is accepted as the best estimate of C*(t)
when the step (T n  T n 1 ) becomes sufficiently small.

499

View publication stats