Hydrology for the Water Management of Large River Basins (Proceedings of the

Vienna Symposium, August 1991). IAHS Publ. no. 201,1991.

HYDROLOGICAL FORECASTING AND

UPDATING PROCEDURES

P. SERBAN
Institute of Meteorology and Hydrology, Romania
A.J. ASKEW
World Meteorological Organization, Geneva, Switzerland

ABSTRACT Some of the inputs of greatest potential value for the

management of rivers are forecasts of future flows. The heart of
any flow forecasting system is a hydrological model. WMO has
conducted a series of projects for the intercomparison of such
models. The last of these was implemented from 1985 to 1989 and
involved a total of 14 models from 11 countries. Central to the
project was a study of the updating procedures used. This paper
considers the role of flood forecasting in the management of large
rivers, describes the updating procedures used in forecasting sys-
tems and comments on their relative efficiency.

INTRODUCTION

The 1970s and 1980s have seen a great increase in the world's population and in
the level of expectation of both individuals and nations. This has led to greatly
increased demands on the world's limited resources of fresh water; demands not
only for consumption, energy production and waste disposal, but also for the
maintenance of a balanced and healthy aquatic environment. One response has
been the refinement of a range of techniques for use in the development of water
resources and the management of river basins (e.g. Yeh, 1985).

A NEED FOR DATA

Techniques for water-resource systems analysis can operate on a deterministic

basis, assuming full knowledge of the past and future of all relevant processes,
both physical and socio-economic. The results obtained can be very valuable, but
no-one can doubt that the deterministic assumption is a major over-simplification.
Efforts to introduce even a minimum of realistic uncertainty, however, lead
immediately to greatly increased complexity and calls for far more data and
computing power. There is still much to be done in refining the tools of water-
resource systems analysis, particularly those that can operate with limited data.
Increasing effort is being channelled into the collection and interpretation
of relevant socio-economic data. There is, however, a worrying lack of work

357
P. Serban &A. J. Askew 358

directed to the collection and processing of the geophysical data required. There
is in fact ample evidence of a world-wide downward trend in such work, invol-
ving both developed and developing countries (WMO, 1990a).
Past geophysical, more specifically hydrological, data are essential in
assessing not only the magnitude of a body of fresh water and its extent in space,
but also its variability with time. The frequency of past floods and droughts can
be assessed with a fair degree of accuracy, but estimates of the future frequency
of such events is clouded with considerable uncertainty. Yet it is for the future
that water-resource systems are to be designed. Between the known past and the
uncertain future, lies the present when existing systems must be operated as
efficiently as possible. It is here that hydrological forecasting has a vital role to
play.

THE ROLE OF HYDROLOGICAL FORECASTING

Hydraulic structures are usually operated in accordance with a set of rules which
take account of the current state of the system, the demands being made upon it,
and some estimates of future rainfall, streamflow, evaporation and the like.
Sometimes these estimates are not explicitly required as input because the rules
have been devised on the assumption that, for the current period in time, the
various hydrological variables will take on their average values for the relevant
month or season of the year. However, whether they are explicit or implicit,
these estimates of future conditions are forecasts: Often very poor forecasts. It is
well to recognize this fact and avoid the pit-fall of viewing them as characteristics
of the natural system. The average daily discharge for November is a character-
istic of the natural system; the assumption at 00 hrs on 1 November that the
actual discharge for the coming 24 hours will equal this value is human judge-
ment. Given some on-line information about the state of the natural system, in
other words, some relevant hydrological data, it should be possible to improve
the level of judgement and hence improve the operation of the structure concer-
ned. At the very least, the wise operator will take note of the current level of
flow and its rate of change and act accordingly. At the least, all important water-
resource systems should incorporate the real-time collection of hydrological data
and their use in developing forecasts of future conditions relevant to the operation
of the systems. The costs of collecting the data and preparing the forecasts are a
fraction of the cost of constructing and operating the system itself and will
usually be far outweighed by their potential benefit in improved safety and
operational efficiency. There is therefore a strong case to be made for the use of
hydrological forecasts in the operation of water-resource systems. However, if
they are to be used to their full potential, it is important that both the systems
themselves and their operating rules be designed to take account of the availabili-
ty of hydrological forecasts and of the anticipated precision of the forecasts.
Whether explicitly or implicitly, the design of a water-resource system makes
some assumption as to die manner in which it will be operated. Whether explicit-
ly or implicitly, the design of operating rules makes some assumption as to the
nature and precision of forecasts of future hydrological conditions. It is preferable
for both sets of assumptions to be made explicitly and to be well founded.
 359 Hydrological forecasting and updating procedures

WMO PROJECTS

The importance of hydrological forecasting, for both water-resource management

and disaster mitigation, has long been recognized by the World Meteorological
Organization (WMO). Over the years it has undertaken a series of projects
designed to encourage the development of improved forecasting techniques and to
provide information to those faced with a choice of techniques. Central to any
hydrological forecasting system is a model which takes as input the relevant data
on the past and present state of the natural system and produces an estimate of its
future state. During the past twenty years WMO has undertaken three projects for
the intercomparison of flood forecasting models (WMO, 1975, 1986 and 1990b).
The WMO projects each included a numerical intercomparison of the
performance of the various models. In the first two cases, these intercomparisons
were based on the ability of the models to simulate periods of flow extending
over a number of years, given the relevant input (rainfall, temperature, etc.), but
without any opportunity to update the forecasts at any time. There were important
technical reasons why updating could not be allowed. Nevertheless, all concerned
were aware of the lack of reality that this engendered. If, as said earlier, the wise
operator of a hydraulic structure would never in practice ignore the current water
level or discharge rate, it is even less likely that a forecast office would issue
forecasts for years at a time without ever correcting them for current conditions.
The third WMO project (WMO, 1987 and Askew, 1989) was therefore
undertaken explicitly to take advantage of developments in micro-computers
which allowed an intercomparison of models which permitted updating of
forecasts. The project was implemented from 1985 to 1989 under the title "real-
time intercomparison of conceptual models" (WMO, 1990b). A total of 14 mo-
dels from 11 countries were compared using four standard data sets. As was
expected, the inclusion of updating had a major impact, not only on the organiza-
tion of the project, but also on the results obtained. There is no doubt that, when
used in operational practice, all hydrological models are associated with some
form of updating procedure and any description of a model is incomplete without
a description of the updating procedure used with it.
The various types of hydrological models have been described at length
elsewhere (e.g. Becker & Serban, 1990), but updating procedures are not so
widely discussed. This is the topic of the remainder of this paper.

FORECAST UPDATING

A major problem which occurs when applying a hydrological model in real-time

is the fact that, in general, the simulated hydrograph is different from the
measured hydrograph up to the time of the forecast. This difference is due to:
(a) errors induced by the model input data;
(b) imperfections in the model, or in some of its component parts (computation
of snowmelt or effective rainfall, flood routing, etc.);
(c) model calibration with only a limited amount of data;
(d) changes in time in certain characteristics of the basin;
P. Serban &A. J. Askew 360

(e) errors in the determination of the discharge hydrograph at the gauging

station.
Errors between simulated and measured hydrographs are, generally, of three
types (Fig. 1):
(a) volume or amplitude errors, generally due to the modelling of infiltration, or
to errors in model input data;
(b) errors in the timing of the simulated and measured hydrograph or phase
errors mainly induced by the routing component of the model;
(c) hydrograph shape errors mainly introduced by the transfer component of the
model: for instance, a unit hydrograph.

3 - amplitude errors b - phase errors C - shape errors

Tin» TITO Timt

Fig. 1. Definition of types of error between measured (-) and simulated (...)
hydrographs (a) amplitude errors; (b) phase errors; (c) shape
errors.

Various combinations of these three types of error can occur in operational

practice.
Hydrological forecasting models consist of a "simulation model" and a
procedure for forecast updating (Fig. 2). The "simulation models" make use of
input data, either measured or estimated, relating to precipitation (P) and other
meteorological factors (M), such as air temperature, wind speed, humidity, etc.,
in order to compute discharges (QS). These models do not take into consideration
the discharges (QM) recorded at the river station up to the time that the forecast
is derived.
For application in real-time, updating procedures have been developed
which take account of the errors between the computed and measured discharges
in order to correct the values derived by the "simulation" models over the entire
forecasting interval.
Continuous updating procedures can be applied for forecasts at each time
step. It is important to note that these procedures differ from the periodic
recalibration of models - referred to as "periodic updating" by some authors.
Periodic recalibration of model parameters is necessary when the characteristics
of a basin or of a river bed change in time, in particular as a result of man's
influence. It is quite independent of the continuous updating that is the subject of
this paper.
 361 Hydrological forecasting and updating procedures

C/M

'ÔIMULATION UPDATING QF_

Fond M.
MOOFL p/focsùwe

J JW Time

Fig. 2. Block diagram of a forecasting procedure (a) and the principle

effect of updating.
Legend:
P rainfall j time of preparation of forecast
M other meteorological factors QF forecasted hydrograph
QS simulated hydrograph e simulation error
QM measured hydrograph e forecasting error

TYPES OF UPDATING PROCEDURE

Forecast updating procedures differ depending on the variables modified (Table 1

and Fig. 3). These are:
(a) procedures which update the model input variables;
(b) procedures which update the variables of state in the model of the system;
(c) procedures which update the model output variables.

TYPE OF UPDATING PROCEDURE

UPDATING OF MODEL INPUT UPDATING OF STATE VARIABLES UPDATING OF MODEL OUTPUT

VARIABLES OF THE MODELLED SYSTEM VARIABLES

MANUAL - MANUAL- MANUAL-

AUTOMATED AUTOMATED AUTOMATED
INTERACTIVE INTERACTIVE INTERACTIVE

.TRIAL - ERROR KALMAN FILTER AUTOREGRESSIVE

OR MODELS
PROCEDURE EXTENDED KALMAN FIUER

Fig. 3. Types of updating procedure.

P. Serban & A. J. Askew 362

Table 1. Characteristics of forecast updating procedures (information based

on WMO (1987 & 1990b)).

Updating Variables to be directly updated

procedure
MODEL
Automated Manual- Inputs State variables Outputs
interactive

1. UBC - Canada X Discharges

X Water equivalent of
snow cover
2. CEQUEAU X Rainfall and Discharges
Canada snowmelt
3. CRM X Baseflow indicator Discharges
Czechoslovakia
4. GAPI X Discharges
Hungary andflood
volume
5. SMAR X Discharges
Ireland
6. CLS X Unit hydrograph
Italy
7. VIDRA -
Romania
X Flood volume
and shape of
the hydro-
graph
8. HBV Air temperature
Sweden and precipitation
9. SRM Snow covered area Discharge
Switzerland/USA
10. TANK Rainfall and
Japan snowmelt
11. SSARR Air temperature
USA and precipitation
12. HFS Amount of water in
USA the model reservoirs
13. NAMS11 Discharges of
Denmark tributaries
(lateral
inflow)
14. NAMKAL Amount of water in
Denmark the model reservoirs

Certain models make use of updating procedures for both input and output
variables, whereas others use the updating of both the state and output variables.
Among the models taking part in the WMO intercomparison project three updated
only the model inputs, three only the state variables and four only the outputs,
whereas one of them updated bodi the inputs and outputs and three of them
updated the state variables and the model outputs as well.
It is noted that a few models also update the model parameters. In general,
this procedure is not to be recommended because in most models the parameters
 363 Hydrological forecasting and updating procedures

are not independent and the modification of one parameter would require the
modification of other parameters. The parameters are usually determined by
calibrating the model and any adjustment of them amounts to a re-calibration. It
is useful to study the interaction between parameters when calibrating the model,
but such a study is not part of the model's operational application.
Updating procedures can be classified as "automated" or "manual-interacti-
ve", depending on the way they are applied. Automated procedures are complete-
ly reproducible and objective. In most instances, they are preferred to manual
procedures (Table 1). Manual-interactive procedures are based on the practical
experience of the forecaster and therefore include a certain amount of subjectiv-
ity. The application of manual procedures, such as was done in the WMO
project, can involve the use of interactive programmes which are based on a real
"man-machine" dialogue.
Some models use both automated and manual procedures. In these cases
the automated adjustment schemes are used only as instruments for assisting the
user in adjusting the model to the observed conditions. After applying the
automated procedures, the forecaster reviews the state of the system and compa-
res it with the data measured over the basin. If any unanticipated errors or
unusual circumstances have arisen, the user changes certain variables on a
subjective basis and runs the automated procedure once again.

DESCRIPTION OF UPDATING PROCEDURES

Updating procedures for model input variables

The primary input variables that may be updated are precipitation and air
temperature. In some cases snowmelt is updated as an intermediate parameter,
being snowmelt model output and rainfall-runoff model input.
Most procedures which update input variables are interactive and of the
"trial-error" type because with most models it is difficult to solve the reverse
problem, that is, to determine the model input when the model output and
parameters are given, while taking into account the errors which influence
computation quantities and the non-linearity of the hydrological system itself. A
block-diagram for a "trial-error" updating procedure is given in Fig. 4.
The most important stages in these types of procedure, to be carried out at
each forecasting moment j , are:
(a) computation of the error "e" between the measured and simulated hydro-
graphs;
(b) comparison of the error with a pre-defined acceptable level of error;
(c) selection of the input variables to be adjusted, plus the adjustment incre-
ment for each variable and the maximum number of increments of change
allowed in any computation period;
(d) rerun of the model using the adjusted input variables.
P. Serban & A. J. Askew 364

Time of forecastj

Model run
I
1--0

ML 3
Error e;r
Computation
Vs/j
J=Jtf

imputed tiydrqgraph
Com,
'Inin the admittea
error //'mitt

yss A/O

Àin température and snow covered

area (or ônowmett) are adjusted 6y
i A M; and t/ie precipitation
Prec/pitat/on
(if any) 6y i à Pi
cd/'asfed iylAP/

i-if-/

'<«W NO
~@
yes
T/ie made/ /j rerun aver
the interval j-k -f-J

Fig. 4. Block diagram of updating procedure of the "trial-and-error" type

for input variables.
ef error between computed
QSV during the ith iteration and the corresponding measured Qmt hydrograph
TEj accepted error threshold (typically 5%)
i iteration counter
nm max. number of iterations
K modelled system memory

Updating procedures for the state variables of the modelled system

State variables that may be adjusted include: areal extent of snow cover, snow
depth, snow water equivalent, the amount of water contained in model reservoirs
simulating water storage at the surface and in the unsaturated and saturated zones
and in water courses.
One justification for state variable updating is that errors in model inputs
are accumulated and appear as errors in the water content of the stores of the
conceptual rainfall-runoff model which, if not corrected, will in turn lead to
erroneous output variables. The water equivalent of the snow cover is generally
 365 Hydrological forecasting and updating procedures

updated using snow course data. The amount of water stored in the conceptual
reservoirs of the model is often updated by means of a Kalman filter. This can be
applied to linear models or, with the aid of an extended version of the Kalman
filter, to nonlinear models.
The Kalman filter may be integrated with transfer models of the ARMA
type (Wood & Szôllôsi-Nagy, 1980), or with conceptual hydrological models,
such as: the HFS model (Georgakakos et al, 1988) and the NAMKAL model
(Refsgaard et al., 1983). The application of the linear Kalman filter to the study
of a physical system (Fig. 5) requires (Gelb, 1974):
(a) a description of the system dynamics as a system of linear equations of the
form:

Xj + , = #Xj + rUj + Wj (i)

(b) a definition of the measurement equation relating the measurements generally

carried out on the system output with the variables of state:

Zj = HXj + Vj (2)

where:
X is the vector of the variables of state, describing the system evolution;
U a control vector containing the input variables;
# the transition matrix;
T the input adjustment matrix;
W the modelling error vector;
Z the measurement vector;
H the measurement selection matrix;
V the measurement error vector.
The matrixes # , V and H defining the characteristics of the modelled
system can be time-constant or variable in time.
The V and W errors are considered independent and normally distributed.

V~N(0,R); W~N(0,Q); E [V*., ^ 1 = 0

Randcm Initial Fiandom

unmeasured conditions measurement
disturbance noise
Wj •Xltol-Xo 'Vj
System Measurement

input Previous state Output ,

Xj*1=<}>Xj^ Tuj +Wj Xj Xj Zj - H x j +Vj Zj
Uj

Variables of state Xj

XM ^ Unit
delay
Xj.1

Fig. 5. Sketch representing a physical system.

P. Serban &A. J. Askew 366

As the X state quantities at any time step are only estimates of the "true"
X values, the covariance matrix for the estimation errors is defined as:

Pj^E^Xj-Xj^MVXjij-.n (3)

Once the initial values for X,, and P c are pre-established, the equations of
the linear Kalman filter for the forecasting and the updating stages are the
following:

Forecast at moment i

- the vector of state:

x J + 1|J = #x j l j +ru j (4)

- the matrix of estimation errors:

P j + 1 | J = # P j l J # T + Q1 (5)

Forecast updating using the measurement at moment i + 1

- the correction matrix Kj + [

r
Kj + 1 = P j + 1|jH [ H P j + 1 | j H r + P j + 1]-" (6)

- the vector of state

A A A

Xj + i | j + i = Xj + 1 | j + Kj + 1 [ Z j + 1 - HXj + 11 j] (7)

- the matrix of estimation errors

Pj + 1 | j + 1 =PJ + 1 | j - K j + 1HPJ + 1 | j (8)

The algorithm is repeated by substituting into eqs. (4) and (5) the estima-
ted values of the state vector Xj + , | j + , and of the matrix of errors Pj + , | s + x
obtained through equations (7) and (8).
The term HXj + 1 | i of relation (7) represents the estimation Zj + t | j of the
observed value, Zj + l, using the equation:

Zj + 1 | J = HXJ + I | J (9)

The use of a Kalman filter helps by considering the most significant error
sources (input and output variables, non-optimal parameter values) and because of
this the discharges can be computed within given confidence limits.
 367 Hydrological forecasting and updating procedures

Updating procedures for the model output variables

Output variables that may be updated include: discharge, flood volume, hydro-
graph shape and lateral inflow.
The discharge updating procedures most widely used in practice (e.g.
UBC, CEQUEAU, SMAR and NAMS models) are those based on autoregressive
models, AR, that were fitted to the errors "e" between the computed and
measured hydrographs:

ej = a ^ . , + 8363.2 + ... + a ^ + bj (10)

where a„ % ... a,, are coefficients of the autoregressive model and the bj are
residual (uncorrelated) errors.
The order and coefficients of the autoregressive model are determined by
means of the "e" error series over a relatively short period of time before the
time of forecast. The forecast error («j + k | j) at time j + k using the AR model is
given by:

(ej + klj) — a I ^ +k-l |j + «2^ +k-2|j + a3ej + k-3 |j + •••• (11)

Updating procedures applied to state variables and based on the Kalman

filter require the linearization of the model and its re-writing in state space form.
The programming is difficult and, in addition, the memory occupied by the
program is large and it takes quite a long time to run. This procedure is optimum
from a mathematical standpoint and yields satisfactory results with short lead
times. The efficiency of the procedures is still open to discussion, however, for
longer lead times. There is a particular difficulty with respect to flood crests,
because the hydrological conditions in the region of the crest differ considerably
from the conditions that apply at the time of forecasting.
Updating procedures based on Kalman filters are advantageous in as much
as they make it possible to determine confidence limits for forecast discharges,
which are very useful in assessing the likely accuracy of the forecasts.
Procedures which update model output variables and are based on autore-
gressive models are easy to program and apply and do not take much computati-
on time. Their efficiency depends on the degree of error persistence between the
measured and computed hydrographs. Unfortunately, it is in the vicinity of flood
crests that error persistence is least and errors show a tendency to oscillate both
rapidly and widely. Likewise, an error in the time of occurrence of forecast
discharge can give rise to larger errors where an autoregressive model is applied.
Updating procedures based on autoregressive models or Kalman filters
yield very good results for hydrological forecasting as regards amplitude errors.
The efficiency of these procedures is still open to discussion, however, witii
respect to phase and shape errors (Fig. 1).
In order to eliminate the deficiencies of the AR models and of the Kalman
filter in the case of phase and shape errors, procedures have been developed for
identifying error types: amplitude, phase and shape. It is claimed that these
procedures can update a hydrograph as successfully as an experienced forecasting
hydrologist (Sittner & Krouse, 1979; Serban & Plesa, 1986; Rungo et al., 1989).
P. Serban&A. J. Askew 368

CONCLUDING REMARKS

The freshwater resources of the world are limited and under pressure, but there
need be no limitation on mankind's ingenuity in finding ways of maximizing the
use of these resources. Essential to any such efforts are hydrological data:
historical data on which to design new water-resource systems and real-time data
for use in operating the systems to maximum efficiency.
Operational efficiency can be greatly aided by the accurate forecasting of
future hydrological conditions, essential to which is the application of a hydrolo-
gical model in conjunction with an effective updating procedure. The choice and
correct application of an updating procedure can be as important as the choice of
model. The WMO project completed in 1989 went a long way in describing and
evaluating the various updating procedures in current use, but there is more that
can be done in refining these techniques in the years ahead.

ACKNOWLEDGEMENTS The authors express their gratitude to the Secretary-

General of WMO for his permission to publish this paper.

REFERENCES
Askew, A.J. (1989) Real-time intercomparison of hydrological models. In: Symposium on Surface
Water Modelling (Proc. Baltimore Symp., May 1989), 125-132. IAHS Publ. 181.
Becker, A. & Serban, P. (1990) Hydrological Models for Water-resource System Design and Operati-
on. Operational Hydrology Report No. 35, WMO-No. 740, WMO, Geneva, Switserland.
Gelb, A. (1974) Applied Optimal Estimation. MIT Press, Cambridge, Mass., USA.
Georgakakos, K.P., Rajaram, H. & Li, S.G. (1988) On Improved Operational Hydrologie Forecas-
ting of Stream/low. EHR. Report No. 325.
Refsgaard, J.C., Rugbjerg, M.& Markussen, L.M. (1983) Application of the Kalman Filter to Real-
Time Operation and to Uncertainty Analyses in Hydrological Modelling. In: Symposium on
Scientific Procedures Applied to the Planning, Design and Management of Water Resources
Systems (Pros. Hamburg Symp., August 1983), 273-282. IAHS Publ. 147.
Rungo, M., Refsgaard, J.C. & Havno, K. (1989) Improvement of the Updating Routine in the MIKE
11 Modelling System for Real-Time Flood Forecasting. Proceedings of the HYDROCOMP'89
Conference, Dubrovnik, Yugoslavia.
Serban, P. & Plesa, V. (1986) A Conversational System Man-machine for Flood Forecasting in
Basins with Hydraulic Structures. Proceedings of XIII Conference of the Danube Countries
on Hydrological Forecasts. Belgrade. Yugoslavia
Sittner, W.Y. & Krouse, K.M. (1979) Improvement of Hydrologie Simulation by Utilizing Observed
Discharge as on Indirect Input. NOAA Technical Memorandum NWS HYDRO-38, USA.
Wood, E.F. & Szôllôsi-Nagy, A, eds (1980) Real-Time Forecasting/Control of Water Resource Sys-
tems. Pergamon Press, Oxford, UK.
WMO (1975) Intercomparison of Conceptual Models used in Operational Hydrological Forecasting.
Operational Hydrology Report No. 7, WMO-Publ. No. 429, WMO, Geneva, Switzerland.
WMO (1986) Intercomparison of Models of Snowmelt Runoff. Operational Hydrology Report No. 23,
WMO-No. 646, WMO, Geneva, Switzerland.
WMO (1987) Real-Time Intercomparison of Hydrological Models. Technical Report to CHy No. 23,
WMO/TD-No. 255, WMO, Geneva, Switzerland.
WMO (1990a) Water Resource Assessment: Progress in Implementation of the Mar del Plata Action
Plan and a Strategy for the 1990s. WMO/Unesco report - in press, WMO, Geneva, Switzer-
land.
 369 Hydrological forecasting and updating procedures

WMO (1990b) Simulated Real-time Intercomparison of Hydrological Models. Operational Hydrology

Report - in press, WMO, Geneva, Switzerland.
Yeh, W.W-G. (1985) Reservoir management and operations models: a state-of-the-art review. Wat.
Resour. Res. 21 (12), 1797-1818.