Sunteți pe pagina 1din 47

Guidelines for

rainfall-runoff modelling

Towards best practice model application


Vaze, J., Jordan, P., Beecham, R., Frost, A., Summerell, G.

Copyright Notice
2012 eWater Ltd
Legal Information
This work is copyright. You are permitted to copy and reproduce the information, in an unaltered form, for
non-commercial use, provided you acknowledge the source as per the citation guide below. You must not
use the information for any other purpose or in any other manner unless you have obtained the prior written
consent of eWater Ltd.
While every precaution has been taken in the preparation of this document, the publisher and the authors
assume no responsibility for errors or omissions, or for damages resulting from the use of information
contained in this document. In no event shall the publisher and the author be liable for any loss of profit or
any other commercial damage caused or alleged to have been caused directly or indirectly by this
document.
Citing this document
Vaze, J., Jordan, P., Beecham, R., Frost, A., Summerell, G. (eWater Cooperative Research Centre 2011)
Guidelines for rainfall-runoff modelling: Towards best practice model application.
Publication date: March 2012 (Version 1.0)
ISBN 978-1-921543-51-7
Acknowledgments
eWater CRC acknowledges and thanks all partners to the CRC and individuals who have contributed to the
research and development of this publication.
We acknowledge the inputs from the hydrology group in DERM, Queensland, and Mark Alcorn from SA
Department for Water. We thank Matthew Bethune, Peter Wallbrink, Dugald Black, Jin Teng, Jean-Michel
Perraud, Melanie Ryan, Bill Wang, David Waters, Richard Silberstein, Geoff Podger, David Post, Cuan
Petheram, Francis Chiew and Andrew Davidson for useful discussions.
eWater CRC gratefully acknowledges the Australian Governments financial contribution to this project
through its agencies, the Department of Innovation, Industry, Science and Research, the Department of
Sustainability, Environment, Water, Population and Communities and the National Water Commission

For more information:


Innovation Centre, Building 22
University Drive South
Bruce, ACT, 2617, Australia
T: +61 2 6201 5834 (outside Australia)
Support: 1300 5 WATER (1300 592 937)
E: contact@ewater.com.au
www.ewater.com.au

Guidelines for rainfall-runoff modelling

Contents
1

Introduction ................................................................. 5

1.1

Background ........................................................................................................................... 5

1.2

Definition of Best Practice ..................................................................................................... 5

1.3

Scope .................................................................................................................................... 6

2
2.1

Overview of procedure for rainfall-runoff modelling .... 8


Problem definition ................................................................................................................. 8
Problem statement and setting objectives ............................................................................ 8
Understanding the problem domain ...................................................................................... 8
Metrics and criteria and decision variables ........................................................................... 9
Performance across multiple catchments and subcatchments ............................................. 9

2.2

Option modelling ................................................................................................................... 9


Methodology development .................................................................................................... 9
Collate and review data ...................................................................................................... 10
Setting up and building a model ......................................................................................... 10
Calibration and Validation ................................................................................................... 10
Sensitivity/uncertainty analysis ........................................................................................... 12
Documentation and Provenance ........................................................................................ 12
Model acceptance and accreditation .................................................................................. 13
Use of accepted/accredited model...................................................................................... 13

Model choice ............................................................. 14

3.1

Model selection ................................................................................................................... 14

3.2

Available models ................................................................................................................. 15


Empirical methods .............................................................................................................. 15
Large scale energy-water balance equations ..................................................................... 16
Conceptual Rainfall-Runoff Models .................................................................................... 16
Landscape daily hydrological models ................................................................................. 17
Fully distributed physically based hydrological models which explicitly model hillslope and
catchment processes .......................................................................................................... 17

4
4.1

Collate and Review Data........................................... 20


Catchment details ............................................................................................................... 21
Location of gauges (streamflow, rainfall and evaporation) ................................................. 21
Topography and Catchment Areas ..................................................................................... 21
Soil types ............................................................................................................................ 21
Vegetation ........................................................................................................................... 21

Guidelines for rainfall-runoff modelling

Water Management Infrastructure ...................................................................................... 22


4.2

Flow data ............................................................................................................................ 22

4.3

Rainfall ................................................................................................................................ 23

4.4

Evapotranspiration .............................................................................................................. 24

Statistical Metrics for Testing Performance .............. 25

Calibration and validation .......................................... 27

6.1

Calibration ........................................................................................................................... 27

6.2

Validation ............................................................................................................................ 27

6.3

Calibration and Validation of Models to Single Gauge Sites, Multiple Gauge Sites and
Regionalisation of Model Parameter Sets ......................................................................... 29

6.4

Automated, Manual and Hybrid Calibration Strategies ....................................................... 30


Manual Calibration .............................................................................................................. 30
Automated Calibration ........................................................................................................ 31
Hybrid Calibration Strategies .............................................................................................. 32
Selection of Objective Functions for Automated and Hybrid Calibration ............................ 33

6.5

Calibration of Regression Models ....................................................................................... 37

Uncertainty and Sensitivity Analysis ......................... 38

7.1

Sensitivity Analysis ............................................................................................................. 39

7.2

Application of Multiple Parameter Sets ............................................................................... 39

7.3

More Advanced Quantitative Uncertainty Analysis ............................................................. 40

7.4

Consideration of Uncertainty in Practical Applications of Rainfall Runoff Models .............. 40

Concluding remarks .................................................. 42

References ................................................................ 43

Guidelines for rainfall-runoff modelling

Introduction

1.1

Background
Reliable estimates of stream flow generated from catchments are required as part of the
information sets that help policy makers make informed decisions on water planning and
management. The characteristics of the streamflow time series that influence water resources
system modelling and planning can include the sequencing of flows on daily and longer time
steps, spatial and temporal variability of flows, seasonal distribution and characteristics of high
and low flows.
The best available estimate of streamflow would be expected to come from water level
observations made at a gauging station, converted to flow estimates using a well defined and
stable rating curve. However, these observations are only available for limited number of
gauging locations and for limited time span. Estimates for ungauged locations and for a much
longer time period are needed for contemporary water management, and ways to make
estimates for future possible conditions are also required.
A range of methods are available to estimate streamflow from catchments, using observed
data wherever possible, or estimating by empirical and statistical techniques, and more
commonly using rainfall-runoff models. The modelling approach used to estimate streamflow
by different water agencies and consultants varies across Australia and depends on the
purpose of the modelling, time and money available, and the tools and skills available within
the organisation. With increasing levels of inter-agency collaboration in water planning and
management, development of a best practice approach in rainfall runoff modelling is desirable
to provide a consistent process, and improve interpretation and acceptability of the modelling
results.
The purpose of this document is to provide guidance on the best practice for implementing fit
for purpose rainfall-runoff models, covering topics such as setting modelling objectives,
identifying data sources, quality assuring data and understanding its limitations, model
selection, calibration approaches, and performance criteria for assessing fitness for purpose

1.2

Definition of Best Practice


Best Practice Modelling can be defined as a series of quality assurance principles and actions
to ensure that model development, implementation and application are the best achievable,
commensurate with the intended purpose (Black et al., 2011).
What is in practice best achievable, commensurate with the intended purpose may be
subject to data availability, time, budget and other resourcing constraints. Hence, what is
meant by the term Best Practice Modelling can vary. Not only does it depend on the
circumstances of the project, particularly providing results that are fit for the intended purpose,
but it also depends to a great degree on interpretation in peer review. This, in turn, will be
influenced by the general state of knowledge and technology in the modelling field, which can

Guidelines for rainfall-runoff modelling

be expected to progressively develop over time (such as new remote sensing data sources
coming on line, and new computing hardware and software), as well as data, time, budget and
resourcing constraints. Best Practice Modelling provides for a strategic approach to
modelling which enables the trade-offs that may be imposed by these constraints to be better
managed, and assists in identifying priorities for addressing model and data limitations.

1.3

Scope
The eWater CRC has prepared generic Best Practice Modelling guidelines (Black et al., 2011).
They aim to provide for an integrated approach that enables interactions and feedbacks
between all domains relevant to water management (e.g. hydrological, ecological,
engineering, social, economic and environmental) to be considered.
The procedure in that guidance is intended to be flexible enough to accommodate variations in
the meaning of the term Best Practice Modelling and also allow for continuous improvement
as the state of knowledge and technology in the modelling field develops and improves.
The eWater CRC will also provide guidelines to support the BPM guidelines in specific areas of
hydrological modelling that relate to the products that they are developing. This guideline is
intended to address rainfall-runoff model application with the objectives being to provide water
managers, consultants and research scientists with information on rainfall-runoff models and
how to choose one that is fit for purpose, the data available to develop them, and the
calibration and validation methodologies.
There are a number of different purposes that a rainfall runoff model may be applied within an
overall water resources or catchment modelling framework, such as eWater Source. Most of
these purposes relate to providing information to support decision making for some water
management policy. In particular, this can involve:

Understanding the catchment yield, and how this varies in time and space, particularly
in response to climate variability: seasonally, inter-annually, and inter-decadally.

Estimating the relative contributions of individual catchments to water availability over a


much larger region, e.g. valley or basin scale.

Estimating how this catchment yield and water availability might change over time in
response to changes in the catchment, such as increasing development of farm dams,
or changes in land-use and land management.

In some instances with a high quality network of long term stream gauges, most of this type of
information can be estimated from the observations. However, the more common case is
where there is some combination of short term stations, variable quality data, and gaps in
spatial coverage. In these cases, spatial and temporal gaps in the information can be
estimated using rainfall runoff models to:

Infill gaps caused by missing or poor quality data in an observed data series for a
gauged catchment.

Estimate flows for a gauged catchment for periods before the observed flow record
started or after when the observed flow record ends.

Estimate flows for an ungauged catchment.

Estimating flows from ungauged subcatchments within an overall gauged catchment.

Guidelines for rainfall-runoff modelling

Forecast flows for some immediate future period (typically for a period of between a few
days and a few months), conditioned on current (or recent) observations of the
catchment state.

Assess the impacts of human influences within a gauged catchment (for example
landuse or vegetation cover change) and simulating the flows that would have occurred
for the historical climate sequence with modified catchment conditions. This may
include assessment of catchment conditions that may be non-stationary in either the
observed record or for the simulation.

Assess the potential impact of climate variability and/or climate change on flows from a
gauged catchment.

In some cases, several of the above purposes may be satisfied by rainfall runoff modelling for
the same catchment. There are similarities in the approach that is taken to rainfall runoff
modelling, even though the purpose may differ.

Guidelines for rainfall-runoff modelling

Overview of procedure for


rainfall-runoff modelling
The generic guidelines (Black et al., 2011) outline a procedure for applying a hydrological
model. This can be summarised as occurring in 4 phases:
1

Project management,

Problem definition,

Option modelling,

Compare Options and select the best.

This guidance deals only with problem definition and option modelling because the first and
last phases are discussed sufficiently for the purpose of rainfall-runoff modelling in the generic
guidelines. A further reason is that rainfall-runoff modelling is usually only a part of a larger
hydrological modelling project and these phases would be most appropriately considered in
the context of that larger project. Specific aspects of project management and option
comparison that are directly applicable to the development of a rainfall-runoff model, such as
accreditation, are dealt with at appropriate points in this guidance.

2.1

Problem definition
Problem statement and setting objectives
The problem to be addressed must be clearly articulated to minimise the risk that the wrong
tool will be used for the job. The problem statement will give direction on what objectives will be
considered in developing the rainfall-runoff model. As many water management decisions will
often have more than one goal it will be important to ensure these are all identified.
Sometimes it can be useful to express objectives in a hierarchy that shows primary objectives,
secondary objectives and so on. In this regard, consideration should also be given to
possible additional future objectives and goals that could be met based on the current project
or on future projects that build upon the model established in the current project. The decision
on which option offers the best solution should be based upon whether, or how well, each
option meets the agreed objectives (also see section 2.2.1 and 2.2.2 in the generic
guidelines).

Understanding the problem domain


The choice of the rainfall-runoff model will vary based on the purpose the modelling is being
done for, e.g., to understand seasonal low flow characteristics for an in-stream environmental
need; or to assess over-bank flow frequency; or to estimate overall catchment yield on an
average annual basis. The model selected, data required, and calibration approach adopted

Guidelines for rainfall-runoff modelling

should reflect this requirement. Where the same model may be used for two or more different
purposes, there may also be a need to focus the calibration on a number of different flow
regimes. If rough flow estimates are required over large areas and the runoff generation
methodology should be consistent then the data and modelling process will differ again.

Metrics and criteria and decision variables


Model calibration is a process of systematically adjusting model parameter values to get a set
of parameters which provides the best estimate of the observed streamflow (in the case of
rainfall-runoff models). The process of determining which particular set of parameter values
are best for the intended purpose depends on what comparison metrics are used. Metrics
should be used to quantify the acceptability of the developed model. In all cases graphical
assessment and statistical results of some sort will be assessed to identify the ability of the
calibrated model to reproduce the flows calibrated against.
Different metrics will be more effective in determining model appropriateness to meet different
objectives. What these are should be considered when the problem is being defined.
Understanding appropriate metrics allows model acceptance criteria to be identified.

Performance across multiple catchments and subcatchments


In some situations, the purpose of rainfall runoff modelling is to produce an estimate of the
runoff at a single location where there is a streamflow gauge. If this is the case, the calibration
and validation process may be performed for the single gauged catchment. This approach is
justifiable in situations where gauged data is available for most of the period that flow results
are required for and the purpose of the rainfall runoff model is to infill missing data during the
period of record. It may also be justifiable where there is a requirement to extend the period of
record at the single gauge.
A much more common situation is that flow time series estimates are required at several
locations and that gauged streamflow data is also available at several locations. The locations
where flow estimates are required may or may not overlap with the locations where the flow
data is also available. An objective of any project that involves the application of rainfall runoff
models to multiple catchments or subcatchments should be to demonstrate consistency in the
rainfall runoff model response between those catchments and to explain systematic
differences in the hydrological response between catchments and subcatchments in an
appropriate and logical manner.

2.2

Option modelling
This section describes the process of developing a rainfall-runoff model, further details on key
components are provided in later sections.

Methodology development
The models and methodology employed should be appropriate for the purpose that the model
will be used for. The choices made will be directed by the problem definition developed and
any other information at hand to the modeller. Detail on the models available and guidance on
selecting models and methodology that is fit for purpose is provided in Section 3.

Guidelines for rainfall-runoff modelling

Collate and review data


The amount and quality of data available to develop a model should be determined at the
outset of the project. This can influence the selection of models, the performance criteria, and
the approach to calibrate models. A bare minimum data set sufficient to make an approximate
estimate of mean annual catchment yield would include catchment area along with spatial and
temporal characteristics of rainfall and potential evapotranspiration (PET). A comprehensive
data set would include long-term streamflow measurements and rainfall and PET data
collected at one or more locations within the catchment along with land use coverage,
vegetation cover and impervious area information, including changes over time.
The quality of the data should be reviewed prior to using to detect errors, non-stationarity if
any, and understand uncertainties that may influence estimates. Some methods are
discussed in section 4.

Setting up and building a model


The catchment characteristics are considered along with the knowledge on data available and
any other information available to the modeller. The rainfall-runoff model chosen is
conceptualised and an initial parameter set is identified.
When the model is first set up consideration should be given to all constraints which are
limiting and their effects on the modelling. Section 5 provides more details associated with this
step.

Calibration and Validation


Model calibration is a process of systematically adjusting model parameter values to get a set
of parameters which provides the best estimate of the observed streamflow (in the case of
rainfall-runoff models).
The term validation, as applied to models, typically means confirmation to some degree that
the calibration of the model is acceptable for the intended purpose (Refsgaard and Henriksen,
2004). In the context of rainfall runoff modelling, validation is a process of using the calibrated
model parameters to simulate runoff over an independent period outside the calibration period
(if enough data is available) to determine the suitability of the calibrated model for predicting
runoff over any period outside the calibration period. If there is not enough data available, the
validation may be performed by testing shorter periods within the full record.
It is normal in research studies to split the observed data sets into calibration and validation
period prior to the study, to demonstrate the performance of the model under both sets of
conditions. Use of this approach can cause problems in practical applications if a model
demonstrates acceptable performance for the calibration data set but produces unsatisfactory
results for the validation data set. An alternative approach in this situation is to calibrate the
rainfall runoff model to all available data but to demonstrate that the performance of the model
is satisfactory over different sub-sets of the period that observed data is available.
Further discussion of model calibration and validation is provided in Section 6.
It is a very common situation in a project that involves rainfall runoff modelling for flow time
series to be required for several catchments or subcatchments within the model domain and
for data to be available from two or more stream flow gauges to facilitate calibration and

10

Guidelines for rainfall-runoff modelling

validation. At locations where gauged flows are available and flow estimates are required, two
options are available to the modeller:

The rainfall runoff models may be calibrated independently for each gauged catchment.
In this case, independent parameter sets will be derived for the rainfall runoff models of
each catchment; or

A joint calibration may be performed, with rainfall runoff models calibrated with
consistent parameters to fit to the gauge records from two or more gauges. In this case,
a single set of rainfall runoff model parameters will be produced for all of the catchments
that represent a compromise to fit the flows at all of the gauges within that group.

Consideration should be given at the outset of modelling to the approach that will be used for
dealing with flows from multiple catchments and subcatchments and from multiple gauges.
The strategy for dealing with this issue should be documented at this point and revised, if
necessary, during the process of calibrating and validating the models.
Calibration of a rainfall runoff model normally involves running the model may times, trialling
different values of parameters, with the aim of improving the fit of the model to the calibration
data. Calibration can be facilitated:

Manually, with the modeller setting the parameter values, running the model to inspect
the results and then repeating this process many times;

Using automated optimisation, with an optimiser algorithm running the model hundreds
or thousands of times with different parameter values; or

Using a hybrid approach of automated optimisation phases, interspersed with manually


implemented trials of parameter sets.

Defining the calibration and validation approach before commencing a modelling project can
maximise the efficiency of the calibration process, whilst avoiding the temptation to overfit
the model to noise in the observational data. A calibration strategy should therefore outline
the:

gauge locations where model calibration and validation will be performed;

viable or allowable ranges for each model parameter value;

known constraints, dependencies or relationships between parameter values


(for
example, the total of the three partial area parameters in AWBM, A1, A2 and A3 must
sum to 1);

period for calibration at each gauge location;

period for validation at each gauge location;

expected level of uncertainty in observations introduced by measurement uncertainty;

metrics to be used to test calibration and validation performance;

whether manual or automated calibration strategy will be adopted, or how a hybrid


strategy of progressive manual and automated calibration will be implemented.

If an automated or hybrid optimisation strategy is to be used, further details should be


defined at the outset of the calibration process on:

algorithms to be used for optimisation of parameter values;

objective function(s) that will be used to test the calibration performance;

Guidelines for rainfall-runoff modelling

11

weightings that may be applied in computation of objective functions, to encourage


fitting to different parts of the flow regime (typically the relative weightings to high,
medium and low flows); and

the set of model parameters that will be optimised during calibration and constraints on
the allowable range of values for each parameter.

Ideally, calibration strategy should be documented prior to the commencement of the


calibration process. It may be appropriate for the calibration strategy to be reviewed during the
calibration.

Sensitivity/uncertainty analysis
Relevant sources of uncertainty in typical order of importance include:
1

Model input data including parameters, constants and driving data sets,

Model assumptions and simplifications of what the model is representing,

The science underlying the model,

Stochastic uncertainty (this is addressed under variability below),

Code uncertainty such as numerical approximations and undetected software bugs.

The potential impacts of the above sources of uncertainty on the decisions that will be made
using the model should be considered early in the modelling process and then re-examined
once the model has been calibrated, validated and applied for scenario runs. Uncertainty
becomes more important for estimation of events in the tails of the probability distribution,
floods and droughts, than it is for consideration of events that are closer to the centre of the
probability distribution (such as estimation of the mean annual runoff from a catchment).

Documentation and Provenance


Documentation is an important requirement for model acceptance. Its role is:
1

To keep a record of what was done so that it can be reviewed and reproduced,

To provide source or background material for further work and research,

To effectively communicate the results from models, and

To effectively communicate the assumptions made during the modelling process and
the decisions made by the modeller during implementation of the model.

Good documentation supports a study and it will also enable someone coming along later to
see what decisions were made, what was done to underpin the decisions and why, particularly
if aspects of the project are revisited at some later time.
Provenance, as it might relate to hydrological modelling studies simply means the ability to
trace the source/lineage of the data, model and modelling results. Reasons for providing
provenance in rainfall runoff modelling include:

12

Accountability and a full audit trail for all modelled results.

Repeatability; ability to re-create a results data set using current data or better
understanding.

Guidelines for rainfall-runoff modelling

Reproducibility; ability to re-create a results data set exactly using all original data,
workflow ordering, assumptions and parameters.

Versioning of both entire workflow and systems implementation. Versioning of the


subcomponents and data sets will be the responsibility of those who govern them but
must be captured by the system.

The degree of provenance required depends on the application and/or how the modelling
system is intended to be used by the individual or organisation in future. Current best practise
provenance is to save all input data and model/parameters version and workflow history such
that the outputs can be reproduced in future if required. In the future the ability to register and
resolve the type and identity of objects within the modelling process should reduce the
requirement to capture and archive these objects, especially as modellers take greater
advantage of services based point of truth data streams and modelling systems, and rely less
on ad hoc locally managed resources.

Model acceptance and accreditation


The aim of model acceptance is to gain agreement that the model is fit for purpose.
Information available from the model accreditation process (Reporting, QA documentation,
Peer review) provides model development details and review results which will enhance
model acceptance.
Peer review plays an important part, especially with stakeholders that are external to the
organisation undertaking the model development. It is important for establishing the
credibility, reliability and robustness of results and the methodology used to obtain the results.
It is undertaken by people with specialist understanding in fields relevant to the project.

Use of accepted/accredited model


Once a calibrated model is evaluated against good quality data and has undergone thorough
review process (model acceptance and accreditation), it can be used for modelling to support
water management planning and policy decisions (provided that the model was accredited for
similar purpose).

Guidelines for rainfall-runoff modelling

13

Model choice

3.1

Model selection
Model selection is made based on an understanding of the objectives and the system being
modelled (http://www.toolkit.net.au/Tools/Category-Model_development; CRCCH 2005a, b).
The WMO (2008, 2009) report include the following factors and criteria as being relevant when
selecting a model:
1

The general modelling objective; e.g. hydrological forecasting, assessing human


influences on the natural hydrological regime or climate change impact assessment.

The type of system to be modelled; e.g. small catchment, river reach, reservoir or large
river basin.

The hydrological element(s) to be modelled; e.g. floods, daily average discharges,


monthly average discharges, water quality, amongst others.

The climatic and physiographic characteristics of the system to be modelled.

Data availability with regard to type, length and quality of data versus data requirements
for model calibration and operation.

Model simplicity, as far as hydrological complexity and ease of application are


concerned.

The possible transposition of model parameter values from smaller sub catchments of
the overall catchment or from neighbouring catchments.

The ability of the model to be updated conveniently on the basis of current


hydrometeorological conditions.

Other things that should be considered are:

14

The level of modelling expertise available.

Whether the model is going to be used on its own, or if it is going to be used in


conjunction with other models.

Freedom of choice may be limited by a desire to minimise problems of different models


for much the same purpose in the same project area, or to avoid problems of different
models in adjoining project areas, particularly where the models are linked in some way
in the future or results compared in some way.

Whether uncertainty will be explicitly modelled. If uncertainty is to be explicitly included,


what types of uncertainty are to be modelled (e.g. climatic uncertainty, uncertainty in
climate change projections, uncertainty in rainfall runoff model parameter values); what
approaches will be used to generate the replicates to represent uncertainty and how
many replicates will be required to adequately quantify uncertainty.

Whether simulation or optimisation, or a combination of both, is adopted.

Guidelines for rainfall-runoff modelling

Whether the model is to be used for hindcasting or forecasting when being applied in
predictive mode.

In essence the governing principle in choosing a model should be that it should not have more
parameters requiring calibration or a greater level of detail than the available data can support,
to minimise problems of spurious results and false calibrations.

3.2

Available models
Rainfall runoff models can be represented by a range of approaches, in order of increasing
complexity as:

simple empirical methods (e.g., curve number and regression equations);

large scale energy-water balance equations (e.g., Budyko curve);

conceptual rainfall-runoff models (e.g. SIMHYD, Sacramento, AWBM)

landscape daily hydrological models (e.g., VIC, WaterDyn);

fully distributed physically based hydrological models which explicitly model hillslope
and catchment processes (e.g., SHE, TOPOG).

These categories have been used for ease of description, and there is overlap between these
model types. Although these approaches vary in terms of the complexity with which they
represent the rainfall-runoff transformation processes, all of them conceptualise the real
processes using some sets of mathematical equations (and hence are all conceptual models
of the physical environment). Similarly, conceptual rainfall-runoff models run in distributed
mode can be classed as being landscape daily hydrological models. This section provides a
discussion of the characteristics of each of these model types, along with a broad assessment
of the strengths and weaknesses of each approach for rainfall runoff modelling (Table 3-1).

Empirical methods
Empirical methods to rainfall runoff modelling typically involve the fitting and application of
simple equation(s) that relate drivers of runoff response to flow at the catchment outlet.
Empirical equations are most often derived using regression relationships.
Common predictor variables may include rainfall for the catchment, flow observed at another
gauge in the vicinity, evapotranspiration, groundwater levels, vegetation cover and the
impervious area within the catchment. Where rainfall is used as a predictor variable,
regression relationships derived almost always include a non-linear relationship between
rainfall and runoff.
All catchments incorporate storage elements, including interception by vegetation, storage
within the soil column, groundwater storage and storage within stream channels. Catchment
storage typically results in runoff from the catchment being an integrated function of the
climatic conditions for the catchment over some period prior to the period for which runoff is to
be calculated by the model. Therefore, empirical models that produce acceptably accurate
simulations of runoff are either applied at sufficiently long time steps that changes in internal
water storage within the catchment can be ignored (e.g. annual time step) or applied to
represent an integration of the climatic conditions that occurred over some time period prior.
As a practical example, for most catchments a regression model that only includes daily
rainfall on the current day is likely to produce a very poor estimation of daily runoff but a model

Guidelines for rainfall-runoff modelling

15

for predicting daily runoff that used individual values of daily rainfall for several days prior may
produce acceptable runoff estimates.
Empirical regression relationships are often developed using spreadsheets. They can also be
fitted using more sophisticated statistical analysis packages, which may more easily facilitate
the investigation of predictor variables. For general information on the development of
regression relationships, the modeller is referred to NIST/SEMATECH e-Handbook of
Statistical Methods (NIST and SEMATECH, 2010) or to a University Level statistics text book.
Empirical regression equations are best suited to situations where there are two flow gauges
on the same stream with partially overlapping periods of record, which are therefore subject to
similar climatic drivers, and the regression equation is used to extend the simulated flow to the
combined period of record from both sites. They can also produce adequate simulations for
neighbouring gauged catchments with overlapping periods of record in situations where the
two catchments are subject to similar rainfall timeseries and are relatively similar
hydrologically.

Large scale energy-water balance equations


The large scale energy-water balance methods are based on the hypothesis of available
energy and water governing large scale water balance (precipitation, evaporation and runoff).
These are usually developed using large scale observed data sets, eg. the Budyko curve
(Budyko, 1958) was developed using mainly European data, and numerous other forms have
been proposed to improve estimates in local regions and to account for different land cover
types (Arora, 2002). One of the more popular forms of the Budyko method is the Fu (1981)
rational function equation (Zhang et al., 2004) where a single parameter, , in the equation can
be calibrated against local data to tune the method for the local conditions. The inputs to these
equations are rainfall and potential evapotranspiration (PET) and the output is runoff at mean
annual time step.

Conceptual Rainfall-Runoff Models


Conceptual rainfall runoff models represent the conversion of rainfall to runoff,
evapotranspiration, movement of water to and from groundwater systems and change in the
volume of water within the catchment using a series of mathematical relationships. Conceptual
rainfall runoff models almost always represent storage of water within the catchment using
several conceptual stores (or buckets) that can notionally represent water held within the soil
moisture, vegetation, groundwater or within stream channels within the catchment. Fluxes of
water between these stores and in and out of the model are controlled by mathematical
equations.
Most applications of conceptual rainfall runoff models treat the model in a spatially lumped
manner, assuming that the time series of climatic conditions (notably rainfall and potential
evapotranspiration) and the model parameter values are consistent across the catchment.
There have been implementations in more recent times of conceptual rainfall runoff models in
semi-spatially distributed and spatially distributed frameworks. In distributed application, the
catchment is defined by grid cells or subcatchments within the catchment that are assigned
the same rainfall runoff parameter values but different time series of climatic inputs so that
different grid cells or subcatchments within the catchment produce different contributions to
the overall runoff. This is effectively a series of lumped rainfall runoff models, with lumped sets
of model parameters that are applied with spatially distributed rainfall.

16

Guidelines for rainfall-runoff modelling

Conceptual rainfall-runoff models have been widely used in Australia for water resources
planning and operational management because they are relatively easily calibrated and they
provide good estimates of flows in gauged and ungauged catchments, provided good climate
data is available.
In Australia there are six widely used conceptual rainfall-runoff models; AWBM (Boughton
2004), IHACRES (Croke et al. 2006), Sacramento (Burnash et al. 1973), SIMHYD (Chiew et
al. 2002), SMARG (Vaze et al., 2004) and GR4J (Perrin et al. 2003). The input data into the
models are daily rainfall and PET, and the models simulate daily runoff. The models are typical
of lumped conceptual rainfall-runoff models, with interconnected storages and algorithms that
mimic the hydrological processes used to describe movement of water into and out of
storages. They vary in terms of the complexity of the catchment processes that they try to
simulate and in terms of the number of calibration parameters which vary from four to
eighteen.

Landscape daily hydrological models


These models are based on the concept of landscape processes and they model the typical
landscape processes using simplified physical equations (VIC model, Liang et al., 1994;
2CSALT, Stenson et al., 2011; AWRA-L, Van Dijk, 2010). A catchment is usually
conceptualised as a combination of landscapes which are delineated using some combination
of outputs from digital elevation model analysis, underlying geology, soil types and land use.
Often these models have been designed to reproduce other variables in addition to streamflow
(e.g. distributed evapotranspiration, soil moisture, recharge, salinity), and as a result have a
greater complexity to methods that target streamflow alone.

Fully distributed physically based hydrological models which


explicitly model hillslope and catchment processes
The physically based models are based on our understanding of the physics of the
hydrological processes which control the catchment response and use physically based
equations to describe these processes. A discretisation of spatial and temporal coordinates is
made at a very fine scale for the entire catchment and the physical equations are solved for
each discretised grid to obtain a solution.

Guidelines for rainfall-runoff modelling

17

Table 31 Assessment of Strengths and Weaknesses of Different Rainfall Runoff


Model Structures
Criteria

18

Model Type
Empirical

Large Scale
Energy-Water
Balance

Conceptual

Landscape
Daily

Fully
Distributed
Physically
Based

Typical Run Time


Step

Can be
daily if daily
flow from
another
gauge is
used as a
predictor
variable.
Otherwise
typically
only
applied at
annual (or
longer)
time scale

Typically only
applied for
mean annual
runoff,
although
pattern of flows
from a nearby
gauge may be
used to
disaggregate
annual totals to
monthly or
daily time steps

Daily,
although
shorter run
time steps
are possible
if sufficient
climatic
data is
available at
this shorter
time step

Daily, although
shorter run
time steps are
possible if
sufficient
climatic data is
available at
this shorter
time step

Minutes to
hours to
maintain
numerical
stability,
although often
forced with daily
data and
assumed
patterns used to
disaggregate to
shorter time
steps

Typical Number of
Parameters

1 to 5

2 to 4

4 to 20

10 to 100

10 to 1000's

Risk of over-fitting
or
over-parameterising
the model.

Low

Very Low

Moderate

High

Very High

Need for high


resolution spatial
data layers

None to
Moderate

Low to
Moderate

Low

High

Very High

Strength of
Apparent
Connection
between Model
Parameters and
Measurable
Physical Catchment
Characteristics

None

None

Weak for
most
parameters
(although
impervious
area or
interception
may be
exceptions)

Moderately
weak

Claimed to be
strong by
proponents but
can be difficult
to validate this
claim

Run time on
typically available
computer platforms
for 100 years of
daily data

<1 second

<1 second

<1 to 60
seconds

10 seconds to
100 minutes

1 minute to
several hours

Guidelines for rainfall-runoff modelling

Criteria

Model Type

Ability to implement
multiple runs for
automated
calibration

Not
typically
required optimum
parameters
can be
obtained
by least
squares
fitting that
does not
require
multiple
runs

Not typically
required

Very Good.
Run times
are typically
sufficiently
low to
facilitate this
and tools
are
available
(Rainfall
Runoff
Library and
Source) to
facilitate this

Good. Run
times likely to
be sufficiently
low to facilitate
this in most
circumstances,
however tools
for calibrating
such models
using
automated
routines are
not as widely
available

Poor. Run times


are generally
too long to
consider
automated
calibration

Typical
Performance in
Regionalisation

Moderate
at annual
time steps.
Usually
very poor
at shorter
time steps
(e.g. Daily)

Good at annual
time steps.
Usually very
poor at shorter
time steps (e.g.
Daily)

Moderate at
daily time
steps

Proponents
claim to be
superior for
regionalisation
to conceptual
rainfall runoff
models

Proponents
claim this to be
a strength of
distributed
models but in
reality the large
number of
parameters
required may
compromise the
application of
distributed
models to
regionalisation

Representation of
non-stationarity in
catchment
conditions

Not
possible

Often applied
to explicitly
represent
non-stationarity
in vegetation
cover for mean
annual runoff
signal

Usually
difficult, due
to lack of
physical
meaning for
many model
parameters

Possible

Possible

Typical
performance of
model when
applying to a very
different climatic
period to that used
for calibration

Poor

Moderate when
used to
estimate
impact on
mean annual
flow but flows
disaggregated
to shorter time
steps are likely
to be poorly
estimated

Variable can be good


in some
catchments
but poor in
others

Variable - can
be good in
some
catchments
but poor in
others

Variable - can
be good in
some
catchments but
poor in others

Typical level of
expertise with this
type of model within
Australian water
industry

Strong

Moderate

Strong

Weak

Very weak

Likelihood that
previously
calibrated models
are available for
catchment to be
modelled.

Moderate
to Low

Moderate

Very High

Low

Very low

Guidelines for rainfall-runoff modelling

19

Collate and Review Data


Climatic data is the most important driver of any rainfall runoff modelling process. The
calibration and validation of models also involves comparison to observed streamflow data.
Major causes of difficulty in calibrating rainfall-runoff models are errors and uncertainties in the
input data (see Kavetski et al, 2003). A discussion of these problems can be found in the
collection of papers in Duan et al 2003. Checks should therefore be performed on the input
data and the comparison data set for calibration and validation to be used in rainfall runoff
modelling before any attempt is made to apply or calibrate the models. The intent here is to
investigate the integrity of the data, whether observations are in the first instance plausible
(e.g., is the volume in a hydrograph less than the product of the rainfall and catchment area).
Investigations into data to be used for rainfall runoff modelling should include checks of:

Stationarity of the data time series , i.e. has there been any systematic or step change in
the statistical properties over the time of data collection, and if so why;

Spatial coherence of data, i.e., is the data consistent with regional spatial and temporal
patterns and trends;

Accuracy of the spatial location of the gauging site;

Consistency in the approach used to date and time stamp the data, particularly for data
provided by different agencies;

Procedures use for spatially interpolation of point observations to gridded data


estimates or estimated series across catchment areas

e.g., time series plots at different levels of temporal aggregation, ranked plots, residual mass
curves, double mass curves, etc. This will pick up patterns as well as identify anomalies which
may be potential data QA issues.
Other checks and analysis, including regional consistency of runoff depths, rain-runoff ratios,
rating confidence limits, period of record, whether rainfall and PET is observed or interpolated,
base-flow characteristics, checks for stationarity and variability over time, etc would also be
useful. It is important that prior knowledge is considered.
One major factor which will apply across all types of time series data used is that the time base
must be kept consistent so that the data applies to the same time period. An example is
where flow data time steps should be matched to the rainfall data time step. In Australia, daily
rainfall data is commonly recorded as the depth of rainfall that occurred in the 24 hours
preceding 9 am on the date of the recorded data. In contrast, daily streamflow totals are often
quoted for the 24 hour period commencing on the nominated date, resulting in the recorded
flow data being offset by 1 day forward of the rainfall data. Where possible the flow data should
be extracted at a time step to match the rainfall dataset. HYDSTRA databases allow this where
the records are at short time intervals. In other cases shifting the recorded time series by one
day for either the rainfall or flow time series may be required to produce consistent time series
for modelling.
The remainder of Section 4 outlines the data types, sources, availability, accuracy,
manipulations (such as gap filling) and any other issues.

20

Guidelines for rainfall-runoff modelling

4.1

Catchment details
Location of gauges (streamflow, rainfall and evaporation)
The streamflow recorded at the catchment outlet is a combined response to the spatial
distribution of rainfall and evaporation across the catchment. There are uncertainties
associated with the streamflow measurements due to rating curve errors as well as due to
extrapolation outside the limits of the rating curve. There is spatial variability in rainfall (and to
smaller extent evaporation) across a catchment which is not captured when undertaking
lumped catchment modelling using a single rain gauge. There might be problems with the
location of the rain gauge in terms of capturing a representative rainfall for all the rainfall
events especially for catchments with high rainfall gradients.

Topography and Catchment Areas


The catchment area for a catchment represents the contributing area to the catchment outlet
where the streamflow is measured. The catchment boundaries (and the corresponding
catchment area) can either be derived from topographic map layers or using the catchment
digital elevation model (DEM) and a standard package such as ARCGIS. It is usually easier to
determine catchment area for the catchments located in steeper terrain compared to those
located in very flat areas (especially when using DEM).
Slope and dominant aspect may provide useful explanatory variables for estimating routing
parameters or for regionalisation of rainfall runoff parameters between catchments.

Soil types
A catchments rainfall-runoff response is related to the soil types in the catchment. The surface
soil characteristics determine the infiltration rates and so the contributions from different flow
components (surface runoff, throughflow and base flow). Soils information can be obtained
from any soils field work that has been undertaken in the catchment or from large scale soil
properties maps (e.g. Australian Soils Atlas, Northcote et al., 1960). In most practical
applications of conceptual rainfall-runoff models in Australia, soil information is seldom directly
used as input in the calibration process because the inherent spatial variability in soil
properties within a catchment is typically sufficiently large that it has been difficult to
demonstrate statistically robust relationships between conceptual model parameters and soil
properties.

Vegetation
Land cover/vegetation cover in a catchment can often be correlated with the amount of
interception storage/loss and actual evapotranspiration in a catchment. The land cover across
the catchment can be derived from large scale vegetation mapping based on satellite imagery
or remotely sensed data. Vegetation cover data has not typically been used explicitly in
directly determining rainfall runoff model parameters, although there have been some recent
studies which have demonstrated the importance of catchment land cover in rainfall-runoff
model calibration and for predictions in ungauged basins (Zhang and Chiew, 2008; Vaze et al.,
2011c).

Guidelines for rainfall-runoff modelling

21

Water Management Infrastructure


Water management infrastructure within a catchment can allow humans to make very
substantial modifications to flows within a catchment. Water management infrastructure may
include large dams, farm dams and off stream storages, extractions, man-made canals or
diversion pipelines and discharges from sewage treatment plants. Locations of these
infrastructures should be identified where they exist within the catchment so that their potential
impact on streamflows may be assessed. Recorded flows at the catchment outlet may require
adjustment to allow for the influence of water management infrastructure located upstream of
each of the flow gauging locations.

4.2

Flow data
Reliable measurements of streamflow data are critical for successfully calibrating a
rainfall-runoff model to a catchment. The streamflow data for all the gauged locations can be
obtained from the respective state government water management agencies or from the
Bureau of Meteorology (in Australia). Considerations in checking streamflow data include:

22

the agency collecting the data and the quality assurance procedures (if any)
implemented by that organisation during data collection;

reliability of the rating of levels to flows for the gauge;

the accuracy, extent and currency of cross sections surveyed at the gauge site.
(Surveyed cross sections may only extent to the top of the stream bank and gauging for
flows extending onto the floodplain may use a cross section that is inaccurate);

vegetation and substrate material for the channel bed, channel banks and floodplain
and the influence of assumptions made about these on gauged flows;

the ratio of the highest flow outputs to the highest flow that the gauge has been rated for;

how hydraulically stable (variable over time) the rating site is;

examination of potential backwater effects for the site from influences that are
downstream of the site, such as stream confluences, bridge crossings, culverts, dams
or weirs;

hysteresis effects leading to different flow rates for the same recorded level on rising
and falling limbs of hydrographs;

how well maintained the gauging site and instrumentation used for measuring water
levels has been;

any changes to the gauging instrumentation over time;

the length of time since the last rating at high flows;

length of record at the site;

availability of quality codes with the flow data record;

proportion of missing data;

trends in when data is missing from the record (i.e. Is there any bias toward high or low
flow periods, particular seasons, or are the gaps just random?) and how this might
influence any infilling procedures; and

Guidelines for rainfall-runoff modelling

if there are a number of gauges closely located that basically represent the same
catchment the data sets may be able to be combined to give a longer record for the site.

Assessment of the above factors will inform whether the data is useful in calibration of the
model, independent validation of the model or whether the data should be ignored.

4.3

Rainfall
Rainfall is the main driver of runoff and so reliable measurements of rainfall are critical for
successfully calibrating a rainfall-runoff model to a catchment. There are several sources for
obtaining climatic data for a particular catchment:

Site observations from Bureau of Meteorology climate database.

Site observations taken from monitoring sites collected by other organisations that may
exist outside of the Bureau of Meteorology database. Many jurisdictional databases
contain rainfall records.

Gridded data products, such as the Bureau of Meteorologys Australian Water


Availability Project (AWAP) or Queensland Centre for Climate Applications SILO data
set.

It is important to be aware of how this data has been collected and what data quality control
methods have already been applied to the data prior provision of the data set as this may
influence the modelling results. This is particularly relevant to gridded products, such as SILO
and AWAP (SILO, Jeffrey et al., 2001; AWAP, Jones et al., 2009), as these data sources
generally use different algorithms to convert time-series observations at data points to gridded
data products.
In a small catchment, considerably better results may be obtained from using rainfall station
data from the BOM (http://www.bom.gov.au/climate/) or locally collected data than a gridded
data set that smoothes observations from a smaller number of more sparsely located sites. In
some cases it may be appropriate to adjust the station data, normally by a percentage, if the
mean catchment rainfall can be defined using other sources e.g. isohyetal detail.
In large catchments there is spatial variability in rainfall across a catchment which is not
captured when undertaking rainfall-runoff modelling using rainfall time series from the rain
gauges. If using a single rain gauge, there might be problems with the location of the rain
gauge in terms of capturing a representative rainfall for all the rainfall events. If using a
spatial rainfall product (SILO or AWAP in Australia), there can be uncertainties introduced
because of the method used for interpolating rainfall between rain gauges and changes in the
rain gauge network over time. Interpolation methods currently used are more suited to areas
where rainfall varies less over space and in time. They do not account well for orographic
effects, and rainfall networks in Australia historically have not captured the spatial and
temporal variations in tropical and monsoonal areas well.
Vaze et al., 2010b discusses testing carried out considering the effects of using different
rainfall data sets on the calibration and simulation of conceptual rainfall-runoff models. They
conclude that considerable improvements can be made in the modelled daily runoff and mean
annual runoff with better spatial representation of rainfall. Where a single lumped
catchment-average daily rainfall series is used, care should be taken to obtain a rainfall series
that best represents the spatial rainfall distribution across the catchment. However where only
estimates of runoff at the catchment outlet are required, there is little advantage in driving a

Guidelines for rainfall-runoff modelling

23

rainfall-runoff model with different rainfall inputs from different parts of the catchment
compared to using a single lumped rainfall series for the catchment.

4.4

Evapotranspiration
The measured pan evaporation data can be obtained for all the locations with the evaporation
gauges installed (in Australia from the Bureau of Meteorology (BoM) basic records). In
Australia there are also some spatial climate products which use point evaporation
measurements recorded by the BoM and use an interpolation schemes to produce spatial
evaporation surfaces (SILO, Jeffrey et al., 2001; AWAP, Jones et al., 2009).
The network of pan evaporation recording stations in Australia is sparse in comparison to
stream flow and rainfall networks, although there is some compensation in that typically
potential evapotranspiration exhibits substantially higher spatial correlation than rainfall or
stream flow. This limits the ability to accurately represent the true spatial and temporal
variability in evaporation in models however the spatial variability in evaporation is much
smaller compared to the variability in rainfall.
The BoM network records pan evaporation. Modelling requires potential evapotranspiration
(PET). There are a number of methods to convert pan evaporation to PET including Penman
Monteith, Mortons and accepted pan factors. These use climatic variables in the conversion
calculation including solar radiation, temperature, vapour pressure, and wind speed which are
recorded at some pan recording stations but not all. This further limits the network available to
draw data from.
When all the required data is available the conversion calculations will use the records but
often some variable is missing and estimates of that variable are made and used. Where
there is no data for the climatic variables, calculated pan to PET conversion factors from a
nearby station can be used to derive PET from pan evaporation.
Commonly the spatial products have interpolated layers for a range of climatic factors and the
spatial PET layer is calculated from data in these layers rather than interpolating PET
calculated at recording stations.

24

Guidelines for rainfall-runoff modelling

Statistical Metrics for Testing


Performance
There are many performance measures used to consider the acceptability of a rainfall-runoff
model. In all cases visual assessment and statistical results of some sort will be assessed to
identify the ability of the model to reproduce the flows it is calibrated or validated against. All
may contribute to best practice and which measures are more appropriate will be directed by
the modelling objective. A number of commonly used visual assessment techniques are
outlined in Table 51. Statistical performance measures and their relevance in various study
types are listed below in Table 52.

Table 51 Plots for assessing model performance


Plot

Assessment and Purpose

Daily and monthly plots (linear


and log)

Used to check the general size, shape and timing of hydrographs.


Linear plots will better show medium and high flows and log plots
low flows. Baseflow and recession characteristics can be
reviewed. If recessions are frequently too flat then this can
indicate that the interflow and baseflow are not represented
correctly.

Scatter Plot

Scatter plots show the ability of the model to match flows on


actual time steps. They show the flow ranges where the model is
more accurate. Linear and log plots will show the variability
across the various flow ranges. Often a line of best fit is shown to
indicate the bias of the model in estimating flows.

Ranked Plots

Commonly referred to as frequency of excedence or flow duration


graphs, they show the percentage of time a flow is exceeded over
the modelled period. They show whether the modelled output
can replicate the observed flow regime. Flow duration curves are
effective diagnostics to ensure that both the variability and the
seasonal pattern are captured.

Cumulative mass or cumulative


residual mass curves

Scatter plots and flow duration curves do not examine the time
sequence of events. A model could appear to be replicating the
flow regime however the replication of regimes during wet and dry
periods may not be adequate. A cumulative residual mass curve
is a cumulative plot of residuals (flow value - mean of all values).
A residual, and therefore slope of the curve, will be positive during
wet periods as flows are higher than average and during dry
periods the slope will be negative. If the curves diverge there
may be a data issue. If they diverge consistently in all wet or all
dry periods it is likely that model parameterisation for wet periods
or dry periods may not be appropriate.

Plotting average daily or monthly


flows (average of all Days,
average of all Januaries)

A simple diagnostic to ensuring that the model can replicate


seasonality characteristics.

Guidelines for rainfall-runoff modelling

25

Table 52 Statistical performance measures (metrics) and their relevance in


various study types (Y Yes, N No)
Metric

Purpose
Runoff
Yield

Climate
change

Landus
e
change

Low
flow

Water
quality

Peak flow
/ floods

Difference in total runoff

Difference in total runoff over different


seasons of the year*

Difference in total runoff contained


within high, medium and low parts of the
flow duration curve

Y (high
flows)

Difference in proportion of time that


cease to flow occurs

Difference in the slope of logarithm of


flow versus time for baseflow recession
periods

Mean square error between observed


and modelled runoff

Coefficient of determination (often


referred to as r)

Nash Sutcliffe coefficient of efficiency


on daily flows

Nash Sutcliffe coefficient of efficiency


on monthly accumulated flows

Nash Sutcliffe coefficient of efficiency


calculated using logarithm transformed
flows

* Definition of seasons to be used will vary depending upon the climatic zone that the catchment is in. For
tropical areas, two seasons (a wet season from December-April and dry season from May-November) may
be appropriate. In Southern Australia, it may be appropriate to consider the four conventional calendar
seasons (Dec-Feb, Mar-May, Jun-Aug and Sep-Nov).

** Definitions of high, medium and low flow ranges will depend upon the purpose of the study and the
catchment. Typical ranges might be High flows: days in observed data in the 0 to 20% probability of
exceedance range; Medium flows: days in observed data in the 20 to 80% probability of exceedance range;
Low flows: days in observed data with greater than 80% probability of exceedance and above the cease to
flow level at the gauge. Adjustment of the low and medium flow ranges may be required particularly in
response to the probability of cease to flow conditions at the gauge.

26

Guidelines for rainfall-runoff modelling

Calibration and validation

6.1

Calibration
Model calibration is a process of optimising or systematically adjusting model parameter
values to get a set of parameters which provides the best estimate of the observed streamflow.
Virtually all rainfall runoff models must be calibrated to produce reliable estimates of
streamflow because there has been little evidence identified of strong links between physical
characteristics of catchments and the parameters of rainfall runoff models (Beven, 1989).
Models should always be calibrated to observed data to demonstrate that the model can
produce observed flow time series with an acceptable level of accuracy. The acceptable level
of accuracy will depend upon the statistics of the flow data to be reproduced, which is
determined by the purpose that the model will be applied for.
A model may be available that has been previously calibrated for a catchment as part of
another study. In this situation, the calibration performance of the model should be re-tested
before it is applied because the purpose for developing the model may be different between
the earlier and later applications, which may influence the calibration objectives.
When calibrating a model it should always be kept in mind that there are always going to be
tradeoffs, for example between getting wet, dry, and average conditions correct, and those
tradeoffs will be driven by the purposes the model will be used for.

6.2

Validation
Model validation is a process of using the calibrated model parameters to simulate runoff over
an independent period outside the calibration period (if enough data is available) to determine
the suitability of the calibrated model for predicting runoff over any period outside the
calibration period. If there is not enough data available, the validation may be performed by
testing shorter periods within the full record.
Model validation is one of the most important steps in rainfall-runoff modelling as the
performance of the calibrated model in the validation period provides us confidence in the
modelling results when the calibrated model is used for simulating streamflow outside the
measured streamflow period or when the model is used for predicting streamflow under future
climate change scenarios.
Validation has often been achieved using a split sample process, whereby a period of
observed data (say the first two-thirds of the available record) are used for calibration and the
remaining one-third are used for validation. The model that was calibrated using the calibration
data set is run for the validation period without changing the model parameters and the
goodness of fit statistics are computed for the validation period. The split sample approach
assumes that both the catchment and the climatic conditions that it is subject to are stationary
in nature across the entire period that recorded data is available for. Evidence of stationarity
(or non-stationarity) in catchment conditions that would affect the hydrological response during
the period of recorded data should be checked using independent data sources (such as aerial
photography, satellite imagery, landuse, topographic or other spatial information).

Guidelines for rainfall-runoff modelling

27

A more sophisticated calibration approach can involve multiple calibration and validation
periods. As in the simple split sample approach, the model is calibrated to a calibration period
and then performance is tested over the validation period without changing the model
parameter values. This approach is then repeated multiple times, with each replicate using
different start and end dates for the respective calibration and validation periods. This allows a
range of model performance statistics for calibration and validation periods to be reported.
There will be some instances with this calibration and validation approach where the calibrated
parameters perform well against the calibration data set, but performs poorly against the
validation data set. In research type investigations, where the modeller may be comparing
different rainfall-runoff models, calibration methods, or objective functions, the validation
results can be used directly to help decide the best model or method or objective function.
However, in practical applications, a modeller may have to decide either not to change the
calibrated parameters and report the poor results, or to recalibrate the model because the
performance is unacceptable.
The modeller may choose the latter option, and may then recalibrate and compare against the
validation data set several times until the calibrated parameters perform acceptably against
both data sets. However, as the validation data set has now been used to change the
calibrated parameters, it is no longer an independent data set and has in effect indirectly
become part of the calibration data set.
This risk of having much poorer performance in validation than calibration may be mitigated by
ensuring as far as possible both data sets have similar flow distributions, An arbitrary
approach to splitting the data, e.g., at the midpoint, may result in half of the data being in a
much wetter period. A model calibrated to these conditions would not be expected to perform
well under the drier conditions in the validation data set. More alternate approaches should be
considered on how to split the data set, perhaps into non-contiguous periods, to ensure overall
flow distributions are similar in each period.
Data is a valuable resource, and should be used to greatest effect. In most Australian
conditions, long data sets are needed to adequately represent climatic variability. An
alternative approach to having split samples is to use the complete data set to calibrate the
model, then to report its performance for different sub-periods, e.g., first half and last half, or
decadally, or driest X year period and wettest X year period. The objective would be to have a
comparatively persistent performance across all these periods. This does not necessarily give
you an independent assessment of performance, but does report on performance under
different conditions.
Transposition of model parameter values from gauged to ungauged catchments may be
tested using a spatial variant on split sample validation. Under this approach, component
models from a gauged catchment with the calibrated parameter values for that catchment can
be applied to another gauged catchment to test the uncertainty and bias introduced from
transposition. Uncertainty ranges can be established by testing contributions flow series
produced by model outputs with parameter sets adopted from several different gauged
catchments. Examples of the performance of these transposition approaches are discussed in
Viney et al. (2009) and Chiew (2010).
Generally the same metrics used to assess the performance of the model during calibration
are also used to assess model performance during validation. The model performance during
validation is almost always poorer than during calibration because model parameters are
deliberately not specifically fitted to the data for the validation period.

28

Guidelines for rainfall-runoff modelling

6.3

Calibration and Validation of Models to Single Gauge Sites,


Multiple Gauge Sites and Regionalisation of Model
Parameter Sets
It is a very common situation in a project that involves rainfall runoff modelling for flow time
series to be required for several catchments or subcatchments within the model domain and
for data to be available from two or more stream flow gauges to facilitate calibration and
validation. At locations where gauged flows are available and flow estimates are required, two
options are available to the modeller:

The rainfall runoff models may be calibrated independently for each gauged catchment.
In this case, independent parameter sets will be derived for the rainfall runoff models of
each catchment; or

A joint calibration may be performed, with rainfall runoff models calibrated with
consistent parameters to fit to the gauge records from two or more gauges. In this case,
a single set of rainfall runoff model parameters will be produced for all of the catchments
that represents a compromise to fit the flows at all of the gauges within that group.

The advantage of the joint calibration approach is that, assuming some degree of
homogeneity in the rainfall runoff response of the selected gauged catchments, the parameter
sets produced should be more robust when applied to other catchments with similar response
that were not used for the calibration.
If an automated calibration process is used for joint calibration of multiple catchments, the
objective function used for automated calibration to the gauged catchments will be a weighted
average of the objective function values produced at the individual gauges. Options for
selecting the weighting values are:

All gauged catchments contribute equally to the overall objective function;

Weights are assigned according to the length of available record (e.g. number of days
with data) at each site;

Weights are assigned according to the inverse of the correlation coefficient in gauged
flows between one gauge and one or more of the other gauges in the set (i.e. gauges
with strongly correlated recorded flows are assigned lower weighting factors than
gauges that have weaker correlations with other gauges);

Some combination of the above factors.

There are three main methods of developing flow data sets in residual ungauged catchments
between upstream and downstream gauges:
1

Calibrate a model to the difference in flow between the gauged upstream flows routed to
downstream (adjusted for known transmission losses) and downstream gauges.

Adjust a flow data set from a nearby catchment using either recorded or generated data,

Apply parameter values from other calibrated models and use catchment appropriate
climate data.

There are two main methods of developing flow data sets in ungauged catchments:
1

Develop a regression equation between flows for the ungauged catchment and gauged
catchments and apply this equation to transpose the flow, or

Guidelines for rainfall-runoff modelling

29

Apply parameter values from other calibrated models and use catchment appropriate
climate data.

Generally in the second case parameters for a neighbouring or nearby catchment are used but
climate data and catchment characterises of the catchment of interest are applied in the
model. Many studies have shown that selecting a donor catchment based on spatial proximity
gives similar or better results than selecting a donor catchment based on catchment attributes
(Merz and Bloschl 2004, Oudin et al 2008; Parajka et al. 2005; Zhang and Chiew 2009).

6.4

Automated, Manual and Hybrid Calibration Strategies


Calibration of hydrological models can be conducted using manual or automated methods, or
a combination of the two approaches (see Boyle et al, 2000 and Brdossy, 2007 for
frameworks for combining manual and automated methods of model calibration). Calibration
involves the adjustment of model parameter values to improve the fit of model output data to
observations to a level that is acceptable.
In case of manual calibration, definition of goodness of fit is usually produced as a
combination of statistical indices and visual inspection of the observed and simulated
hydrographs. Whereas in case of automated calibration, definition of goodness of fit is
usually produced using an objective function. The objective function translates the observed
and modelled outputs into a single number, so that the results of successive calibration
iterations can be compared. Automated calibration routines use a defined algorithm that runs
the model multiple times, adjusting model parameter values according to a strategy that is
intended to improve the value of the objective function.
The sections that follow give information on the calibration methods available and their
relevance in various study types (shown in Table 3-1) which dealt with model choice
appropriate for intended purposes.

Manual Calibration
Manual calibration involves the modeller selecting a set of parameters for their model, running
the model once and then examining the output statistics from the model (from the list
discussed in Section 5). The modeller would then revise the values of one or more parameters
and repeat the above process. This may continue many times until the model achieves the
desired level of performance.
The match between simulated and observed streamflow can be visually assessed as either a
time series, or as flow duration curves or residual mass curves. The visual assessment can
identify general deficiencies in the matching of the hydrologic regime, e.g., high flow events
under or over estimated, baseflows under or over-estimated or the seasonal response of the
model not captured appropriately. Software that stores the results of conceptual storages and
fluxes for graphing, and interpretation of these results in the context of model structure is also
useful to identify which parameter values need adjusting and in which direction in order to
improve results.
Guidelines are available from the developers of the Sacramento model that describe how to
estimate key parameter values directly from analysis of recorded hydrographs (Burnash,
1995). A range of realistic parameter values has also been recommended to guide initial
estimates.

30

Guidelines for rainfall-runoff modelling

Strengths:
1

Encourages a deeper understanding of model structure and its applicability to


catchment hydrology, rather than treating as a black box.

Allows for hydrologist to consider performance against a broad range of performance


metrics, and make appropriate adjustments.

Takes into account understanding of the data and the catchment.

Allows a logical checking at each change.

Produces a greater appreciation of strengths and limitations of calibrated result.

Weaknesses:
1

Repeatability is limited. Different people may get different parameters and output flow
time series.

More effort and time required to complete a calibration.

Difficult to manually calibrate models with more than about 10 parameters.

Only amenable to calibration for one gauge at a time and difficult to simultaneously
calibrate flows from multiple gauged catchments that may have similar hydrological
response. As a result, in catchments with multiple gauges very different parameter sets
for independently calibrated catchments may be produced that dont represent the
underlying similarity in hydrological response that would be expected from those
catchments.

Automated Calibration
In computer optimisation routines, an objective function, which itself is selected by the
modeller is used to formulate the calibration problem and the computer undertakes the
calculations.
Strengths:
1

Can operate with one simple objective function or use of multiple objective functions.

Can undertake calibrations for multiple catchments in a short period of time.

Relatively easy.

Repeatable. Different people will get same parameter values.

Weaknesses:
1

Is confined by the optimisation routine chosen and how the objectives are set.

Is dependent on the computer routine being set up accurately to reflect the choices in 1.

Lack ability to check the relationships between the calibrated parameter values
produced as the calibration proceeds, which may cause investigation of sets of
parameters that are infeasible (unless appropriate checks are build within the
calibration algorithms).

Software is required to automate the optimisation process.

Guidelines for rainfall-runoff modelling

31

Parameter values commonly become trapped against the minimum and maximum
constraints of the allowable parameter ranges set by the user. If the user does not check
for this, the parameter set chosen may be sub-optimal as the best parameter set may
have parameter values that lie outside the constraints set by the user at the time the
optimisation is initiated.

There are two global optimisation methods included in Source: Shuffled Complex Evolution
(SCE) and Uniform random sampling. The analysis undertaken as part of the testing with data
from 200+ catchments in southeast Australia showed that there is an advantage in following a
global optimiser with a local optimiser to fine tune the calibrated parameter values. The
Rosenbrock method is included in the framework as a local optimiser. The testing results
suggest that the combination of SCE followed by the Rosenbrock should be used (Vaze et al.,
2011a,d).

Hybrid Calibration Strategies


Best practice for model calibration would normally involve a hybrid of manual and automated
calibration approaches. A typical hybrid strategy would involve:
Implementation of an automated calibration routine using a global optimiser (such as SCE),
with most (if not all) of the possible parameters allowed to vary within the widest ranges that
are physically plausible for that model and catchment.
Visual inspection of the time series and other statistical measures (see Section 5) from the run
with the parameter set that produced the optimum value of the objective function. This
inspection should aim to identify aspects of the model that are performing well and those that
are performing poorly, to provide insight on the particular parameters that are influencing the
(poor or good) quality of the fit.
Implementation of one or many trials of manual calibration, adjusting the value of one or two
parameters at a time that are likely to be particularly influential on aspects of the modelled
results that the modeller observes are undesirable. The outcome of these manual calibration
runs would provide insight about fixed values for particular parameters, or tighter ranges for
constraining the value of particular parameters in subsequent automated trials. An example
might be that the global optimiser implemented to optimise the NSE finds the baseflow
recession parameter (KBase) for an AWBM model as 0.95/day, but examination of the flow
duration curve and time series reveals that adjusting the parameter to 0.99/day produces a
much better fit to baseflow recessions without compromising the fit to high flows. The modeller
could then choose to either fix KBase at 0.99/day for subsequent trials or to constrain the
KBase parameter within a tighter range (say 0.985 to 1.0/day) for subsequent trials.
Implementation of further trials of automated calibration using a global optimiser but this time
with tighter constraints on most or all of the parameters. Appropriate ranges for parameters
could be assessed either from the previous (manual) calibration step, or from the ranges of
each parameter from the 20, 50 or 100 runs producing the best value of the objective function
from the first automated calibration.
Visual inspection of the time series and statistical measures produced by runs that produced
the optimum value of the objective function and also from runs that produced near to the
optimum value.
Steps 2 through 5 may be repeated several times by the modeller, allowing different sub-sets
of parameters to be fixed and allowing for different allowable ranges for each parameter.

32

Guidelines for rainfall-runoff modelling

Implementation of a local automated optimiser, such as a Rosenbrook search, starting from


the parameter set that is most desirable to the modeller identified in Step 5.
Over the last decade or so, there has been a debate in some parts of the hydrological
modelling community with calibration (particularly automated calibration) to a single objective
function value. This has led to the development of calibration approaches that consider
multiple objectives (Gupta et al., 1998; Madsen et al., 2000). Viney et al. (2009) have
investigated the use of hybrid objective functions (Nash Sutcliffe Coefficient of Efficiency with
various bias constraints) in optimisation of rainfall runoff models for the purposes of estimation
of catchment runoff. Kuczera et al, (2006) and Kavetski et al. (2006a, 2006b) have suggested
an alternative Bayesian approach to model calibration that attempts to deal with
multi-objective optimisation by considering possible calibration solutions that overcome many
of the limitations of single objective optimisation. While these approaches have been
demonstrated in the research domain, practical tools for use by general practitioners for
multi-objective calibration are generally not easily accessible.

Selection of Objective Functions for Automated and Hybrid


Calibration
Automated calibration requires the use of an objective function to direct the optimisation
process. The Source calibration tool implements single objective function optimisation. The
calibration tool reduces the comparison between the observed and modelled data during the
calibration period to a single number to be optimised. Other tools (such as Insight) perform
multiple objective optimisation, which searches for many options of possible solutions that are
optimum for at least one of the multiple objectives that are chosen. Multiple objective
optimisation has not been implemented within the Source calibration tool, although it may be
introduced in later versions. Multiple objective optimisation has the advantage that it explicitly
presents the modeller with many possible parameter sets and it then allows the modeller to
make the choice after the optimisation has been run between parameter sets.
Objective functions are normally chosen as one of the statistics that are discussed in Section
5, or a combination of two statistics from this table. The objective functions that are currently
implemented within the Source calibration tool are:
1

Match to Nash Sutcliffe Coefficient of Efficiency of Daily Flows

Minimise Bias between Observed and Modelled

Match to Nash Sutcliffe Coefficient of Efficiency of Daily Flows but Penalise Biased
Solutions (NSE-Bias)

Match to Nash Sutcliffe Coefficient of Efficiency of Monthly Flows

Match to Nash Sutcliffe Coefficient of Efficiency of Monthly Flows but Penalise Biased
Solutions

Combined Match to Nash Sutcliffe Coefficient of Efficiency and Match to Flow Duration
Curve (Daily) (NSE-FDC)

Combined Match to Nash Sutcliffe Coefficient of Efficiency and Match to Logarithm of


Flow Duration Curve (Daily) (NSE-logFDC)

The choice of any particular objective function will depend on the intended application. Each of
the pre-defined objective functions are formulated to put emphasis (reproduce as closely as
possible) on different flow characteristics.

Guidelines for rainfall-runoff modelling

33

The Nash-Sutcliffe Efficiency (NSE) is commonly used in hydrological modelling to describe


the agreement between the modelled streamflow and observed streamflow time series. A NSE
value of 1.0 indicates that all the modelled daily streamflows are the same as the observed
daily streamflows, and an NSE value of less than zero indicates that that the model simulations
are poorer than simply using the mean observed daily streamflow as the streamflow estimate
for every single day. The daily NSE tends to produce solutions that match high and moderate
flows very well but often will produce poor fits to low flows. NSE will also tend to favour
solutions that provide a good match to the timing and shape of runoff events.
The minimise bias (number 2 in the list) objective function will produce a match on the overall
volume of flow generated but often will produce a poor fit to the timing of flows. It should
normally only be used to refine calibrations determined from other methods with the local
(Rosenbrook) optimiser.
The NSE-Bias objective function (number 3 in the list) formulation makes sure that the models
are calibrated predominantly to optimise NSE while ensuring a low bias in the total streamflow.
It avoids solutions that produce biased estimates of overall runoff, which can produce marginal
improvements in low flow performance over the NSE objective function. However, NSE-Bias
will still be strongly influenced by moderate and high flows and by the timing of runoff events,
which can still often result in poor fits to low flows.
The NSE of monthly flows and NSE-Bias of monthly flows (numbers 4 and 5 on the list above)
can be useful for initial calibration because they tend to find solutions that will match the overall
movement of water through the conceptual stores in the rainfall runoff model, without being
influenced by the timing of individual runoff events.
The NSE-FDC and NSE-logFDC (numbers 6 and 7 on the list above) are hybrid objective
functions that compromise between the fit to the timing of high and moderate flows from the
NSE component and the fit to the shape of the whole flow duration curve (FDC). The
NSE-logFDC (number 7) will produce a closer fit to low flows than the NSE-FDC (number 6).
They require the modeller to select a weighting before the implementation of the optimisation
to the NSE and FDC / logFDC components. Higher weightings (between 0.5 and 1) shift the
balance toward the NSE component (and consequently toward matching timing and high
flows), whilst lower weightings (between 0 and 0.5) shift the balance toward the FDC or
logFDC component (and consequently toward ignoring timing but matching the prevalence of
baseflows). The modeller should therefore consider the influence that the weighting of the two
components in the combined objective function has on the outcomes of their automated
calibration. The weighting can be adjusted during successive runs of automated calibration:
e.g. commencing with a value of 0.5, then modifying to a value of say 0.2 if there is a desire for
more emphasis to be placed on matching the flow duration curve, particularly for low flows.
Automated calibration using four of the objective functions were tested with a global optimiser
(SCE) followed by a local optimiser (Rosenbrook) for a large number of catchments in Eastern
Australia. Testing of the results of the calibration with each of the four objective functions
indicate that the NSE-Bias objective function provides the best estimates of daily flows, timing
and volume ratios in majority of the catchments when compared to the other three objective
functions. If daily sequencing is not important for the intended application, any of the other
three objective functions can be used. If the main interest is to reproduce the low flows whilst
compromising the fit to moderate and high flows, the NSE- logFDC objective function should
be used (see Figure 1 on the next page).
The above discussion underlines the importance at the end, and often during, the calibration
process of testing the performance of parameter values produced by automated calibration
against multiple metrics that are appropriate for the problem that the calibrated rainfall runoff

34

Guidelines for rainfall-runoff modelling

model is being applied to. Following rainfall runoff calibration, the model should perform
adequately against a number of different metrics that represent the characteristics of interest
from Table 51 and Table 52. Automated calibration presents a risk of selecting model
parameters that reproduce aspects of the flow regime that are well detected by the selected
objective function used in the optimisation at the cost of ignoring other important aspects of the
flow regime.

Further Guidance on Calibration and Validation of Conceptual Rainfall Runoff Models


This section contains a number of miscellaneous guidance about calibration and validation of
conceptual rainfall runoff models.
There are a number of potential problems that can occur during calibration, particularly during
implementation of automated calibration. The following lists a number of potential problems
and issues that should be considered:

Daily flow duration curve (mm)

100
Observed
NSE low flow fdc
NSE
NSE fdc
NSE Bias

10

0.1

0.01
1

10

100
Days

1000

10000

10

100
Days

1000

10000

Daily flow duration curve (mm)

80

60

40

20

Figure 1: Calibration results for different objective functions

Values of one or more parameters can become trapped near the minimum or
maximum values of the range that the modeller has allowed. The modeller should
always check the value of each parameter against the range of values that the particular
parameter was allowed to vary within. A parameter value that is close to one of the limits
(say within 1% of the minimum or maximum value) may indicate that the allowable

Guidelines for rainfall-runoff modelling

35

range for the parameter provided as an input to the automated calibration process was
too narrowly defined and that widening the allowable range for the parameter may allow
the automated optimisation to find a significantly better overall result. Of course, there
may be a physical constraint that means that the parameter range and optimum value is
appropriate, e.g. the depth of soil moisture stores cannot be negative and the optimum
value for a particular parameter may then fall close to 0 mm.

36

The model may produce a good fit to the objective function but poor performance on
other statistics. As discussed in Section 5, the performance of the model should always
be checked against a number of different statistical measures.

The model may produce a good fit to calibration data but poor fit to validation data. This
may indicate some inherent non-stationarity in the hydrological response of the
catchment.

The model may produce an inappropriate split between surface and baseflow
contributions. In this case, objective functions that are weighted toward matching the
baseflow contribution or constraining the ranges of parameters that particularly control
the baseflow response should be considered.

If an automated calibration routine is implemented, the model may fail to converge to an


acceptable solution. In this case, the input data (particularly rainfall and streamflow)
should be checked. If this appears reasonable, the calibration should be re-run for a
much larger number of iterations and possibly with wider ranges on the parameters to
let it find an optimal solution.

Local optimisers (Rosenbrock) may find only a local optimum solution and miss a
solution that would otherwise produce a much better fit. This is why global optimisers
should always be implemented as the first step of any automated calibration routine and
the local optimiser only used as the final step in the optimisation process to sharpen
the fit.

The structure of most rainfall runoff models is such that there is a strong degree of
interaction between model parameters. In other words, changing the value of one
parameter of the model can result in changing the value of one or more of the other
parameters of the model to produce a fit to the observed data that is equivalent or
almost equivalent.

Performance should be tested against a range of metrics for different flow regimes
(high, medium, low), as well as for wet and dry periods. A good all-round performance
indicates that the calibration is representing the catchment under most conditions.

If the purpose of the modelling is to derive time series of constituent loads for water
quality modelling, it is usually the case that runoff and constituent loads are dominated
by episodic events. Therefore, in looking at the performance of the model, higher
weighting should be placed on statistics that emphasise runoff events than on low flows.
The baseflow index (the ratio of mean annual baseflow to mean annual flow) produced
by the modelled outputs should be similar to the baseflow index from gauged
streamflow data. As a guide, the baseflow index should be within 10% in the modelled
and observed data. Significant differences in baseflow index may result in inaccurate
predictions of loads, particularly when an Event Mean Concentration (EMC)/Dry
Weather Concentration (DWC) generation model is adopted. The baseflow index
produced by the model can normally be corrected relatively easily by manually
calibrating parameters in the model that control recharge to and leakage from baseflow

Guidelines for rainfall-runoff modelling

stores, or constraining the range of allowable values of these parameters in an


automated calibration approach.

6.5

In case of nested catchments, some gauges in the catchment will be better rated and
provide more accurate flow measurements than others, particularly in larger flow
events. In deriving a combined objective function to fit parameters across the whole
catchment, weightings should be adjusted in the combined objective function toward
gauges with more reliable data in preference to gauges with less reliable data.

A check should be performed that the mean annual runoff from the model at gauging
sites is close to the observed mean annual runoff. As a guide, the mean annual runoff
should match to within 5%.

Gauges should be identified that appear to be outliers, where a good calibration


performance is difficult to obtain. This may identify problems that are not immediately
evident in the data, such as additional inflow or extractions not included in the model,
losses/gains to/from groundwater or rating errors at high flows.

Comparison of the data over the model calibration and validation periods should be
used to identify other issues such as potential changes in instrumentation, rating or
instrument malfunction that may not have been detected during earlier quality
assurance checks on the recorded data.

Land use change or events such as bushfire can be evident in the observed data but
may not have been included in the model, which will impact on the performance of the
calibration. Measures to address this should be considered, with options of either
adjustment to the structure of the model to explicitly represent these effects, or
calibrating the model on a different period of data.

It is generally more difficult for a model calibrated over a wet period to predict runoff over
a dry period compared to a model calibrated over a dry period to predict runoff over a
wet period (Vaze et al., 2010). If possible it is suggested that the calibration flow data
should include both wet and dry periods (rainfall periods about 20% higher and lower
than the long term mean annual rainfall for the catchment).

Calibration of Regression Models


Empirical regression models are most commonly applied to provide an estimate of flow at one
gauge location, based upon recorded flows at another nearby flow gauge. They may also
incorporate other drivers of runoff response, such as rainfall on the previous day and possibly
total rainfall for a defined number of preceding days.
Regression models are commonly fitted by minimising the least square of the residuals
between flows from the recorded and modelled flow time series. Least squares fitting can be
easily conducted within a spreadsheet, using the common tools provided to undertake those
analyses.
Regression models almost always assume that the catchment is stationary in terms of its
runoff generation characteristics. Care should also be taken to confirm that the period used to
fit the model is climatically representative of the overall simulation period. If the simulation
period is substantially wetter or drier than the model calibration period it is very likely that a
regression model will produce biased results, due to the inherently non-linear rainfall to runoff
conversion relationship that is observed in most catchments.

Guidelines for rainfall-runoff modelling

37

Uncertainty and Sensitivity Analysis


Relevant sources of uncertainty in typical order of importance include:
1

Model input data including parameters, constants and driving data sets,

Model assumptions and simplifications of what the model is representing,

The science underlying the model,

Stochastic uncertainty (this is addressed under variability below),

Code uncertainty such as numerical approximations and undetected software bugs.

The potential impacts of the above sources of uncertainty on the decisions that will be made
using the model should be considered early in the modelling process and then re-examined
once the model has been calibrated, validated and applied for scenario runs. Uncertainty
becomes more important for estimation of events in the tails of the probability distribution,
floods and droughts, than it is for consideration of events that are closer to the centre of the
probability distribution (such as estimation of the mean annual runoff from a catchment).
The major uncertainty is most likely associated with the input data sets. There are
uncertainties associated with the measured rainfall, PET and streamflow (associated with the
rating curve). There is spatial variability in rainfall across a catchment which is not captured
when undertaking lumped catchment modelling. If using a single rain gauge, there might be
problems with the location of the rain gauge in terms of capturing a representative rainfall for
all the rainfall events. If using a spatial rainfall product (SILO, Jeffrey et al., 2001 or BAWAP,
Jones et al., 2009), there can be uncertainties introduced because of the method used for
interpolating rainfall between rain gauges and changes in the rain gauge network over time. It
is up to the modeller to check the integrity of the input data and be aware of its limitations. If
data problems cannot be fixed the implications for model calibration (particularly where
optimisation is involved) and interpreting model results should be recognised.
Rainfall runoff models are inherently a simplification of the actual hydrological processes that
are occurring within a catchment. The assumptions and simplification reflected in the structure
of the hydrological model introduce uncertainty into the predictions produced by the model,
which is sometimes referred to as structural uncertainty. Even if the real catchment did behave
in a manner that was identical to the structure represented by the rainfall runoff model, there
would be uncertainty introduced by the unknown values of the parameters of the rainfall runoff
model.
The risks presented by uncertainty introduced by numerical approximations and undetected
software bugs can be mitigated by rigorous testing of the computer software. There is an entire
discipline of the information technology industry devoted to development and application of
tests for software programs. The testing regime for a rainfall runoff model should include:

38

Numerical testing of the results of the software code against an independently coded
solution that has been constructed in a spreadsheet or in a previous implementation
(e.g. in FORTRAN);

Testing of the user interface;

Guidelines for rainfall-runoff modelling

7.1

Testing the results that are produced by the model for extreme values of each of the
model parameters, i.e. parameter values at the minimum and maximum values of their
feasible limits;

Testing that the model internally preserves a mass balance for each time step, i.e. the
change in water stored within the conceptual stores in the model is equal to rainfall less
evaporation less seepage losses less total runoff;

Testing that one version of the model replicates the results of a previous version,
particularly where there is ongoing development of the model and/or the framework that
the model is implemented within.

Sensitivity Analysis
A relatively simple means of producing a quantitative estimate of uncertainty from a rainfall
runoff model is sensitivity analysis. In sensitivity analysis, one parameter of the model is
typically varied at a time and the model re-run to test the change in the output produced by the
change in the single input parameter. A typical approach might be to modify the value of each
of the model parameters by +10% to check the percentage change in the mean annual runoff,
followed by modifying each of the model parameters by -10% to check the mean annual runoff.
This analysis would provide the modeller with an appreciation of how sensitive the mean
annual runoff generated for the catchment is to a simple change in each of the input parameter
values. Consistent with the discussion in Section 5, several metrics that reflect the purpose of
the model should be checked during the sensitivity analysis.
Some deficiencies of sensitivity analysis are that:

7.2

It only tests for the influence of parameter uncertainty and ignores uncertainty
introduced by the model structure;

The common method of testing for the same proportional change in each parameter
(e.g. the +/- 10% change in each parameter value in the example above) does not
usually reflect the fact that some parameters are inherently more difficult to estimate via
model calibration and therefore more uncertain than other parameters. It can therefore
produce misleading inferences because small changes in the value of one parameter
may produce large changes in the output but because of this sensitivity the value of this
parameter may be defined to within a relatively narrow range;

There is typically strong interaction between parameters of rainfall runoff models, so


that the change in the value of one parameter would be compensated for by the change
in the value of one or more of the other parameters of the model. Sensitivity analysis on
single parameters ignores parameter interaction.

Application of Multiple Parameter Sets


The process of calibrating rainfall runoff models using an automated optimiser normally
involves a search of many thousands of possible parameter sets. While the focus of
optimisation is often on the single parameter set that best calibrates the model, the
optimisation process will normally return a large number (typically hundreds) of parameter sets
that produce a performance from the model that is almost as good as the optimum parameter
set. The results of the rainfall runoff model produced by these almost as good parameter sets
can be used to quantify the uncertainty in the model predictions produced by parameter

Guidelines for rainfall-runoff modelling

39

uncertainty. A typical approach, if 2000 model runs had been completed during the automated
calibration process, would be to select the 100 model runs that produced the best value of the
selected objective function to characterise the uncertainty in the time series produced by the
rainfall runoff model.
This approach avoids the second and third deficiencies of sensitivity analysis because:

Parameters that are well defined via the calibration process will be represented by a
relatively tight distribution of values amongst the parameter sets that almost produce
the optimum calibration;

Interaction between parameter values is retained by selecting each parameter set.

However, a limitation of this method is that uncertainty introduced by model structure is


ignored, which was also the case for sensitivity analysis.

7.3

More Advanced Quantitative Uncertainty Analysis


A further advancement on the approach of testing many possible parameter sets that almost
calibrate the model would be to calibrate several different rainfall runoff model structures to the
same gauged data time series, which would return multiple possible parameter sets for each
of the model structures, and then to choose combinations of model structure and parameter
set that almost produce the optimum value of the objective function. An example might involve
automated calibration of AWBM, SMARG, SimHyd, Sacramento, GR4J and IHACRES models
to a particular catchment and then selecting the top 300 sets of parameters (along with the
model structure) that produced the optimum value of the objective function from amongst all
the model runs with all six selected model structures.
The Bayesian Total Error Analysis methodology (BATEA) provides the opportunity to directly
address all sources of uncertainty (input, model and response error) in the calibration of
conceptual rainfall-runoff (CRR) models. BATEA has demonstrated the potential to overcome
the parameter biases introduced by poor conceptualisations of these sources of errors and
enhance regionalisation capabilities of hydrological models (Thyer et al., 2007). Application of
BATEA is currently within the domain of applied research.

7.4

Consideration of Uncertainty in Practical Applications of


Rainfall Runoff Models
A hierarchy of techniques exist, as described above, for quantitative estimation of uncertainty
for practical applications of rainfall runoff models.
The potential impact of uncertainty in rainfall runoff modelling on the decisions that are to be
made using that model should be assessed before quantitative estimation of uncertainty is
undertaken. In many situations, decisions will be made informed from the best estimate
rainfall runoff model results and will implicitly ignore uncertainty.
Sensitivity analysis, despite its limitations, at least can be informative about the possible range
of solutions that might be produced due to uncertainty in the parameter values. Consideration
of uncertainty in this reasonably simplistic manner can often provide an indicative range of the
possible range of solutions that may be produced.
Application of multiple parameter sets is facilitated using existing tools that are available to
hydrological modellers (including PEST and the Source calibration tool) and it overcomes two

40

Guidelines for rainfall-runoff modelling

of the major limitations of simple sensitivity analysis. In situations where quantifying


uncertainty, at least the uncertainty introduced by model parameter uncertainty, is important
then this is a viable approach that should be applied by the modeller.
More advanced uncertainty analysis techniques, including BATEA, are likely to advance as
further research is conducted but the tools to implement these techniques are not currently
available to most practitioners. If uncertainty in the model results is a primary driver of the
decisions to be made on the basis of the rainfall runoff model results then these more
advanced techniques of uncertainty quantification should be adopted.

Guidelines for rainfall-runoff modelling

41

Concluding remarks
There are a number of methods of undertaking rainfall-runoff modelling and all have strengths
and weaknesses given the data available and the purpose that the model output will be used
for. Simple mean annual flow assessments may utilise simple equations based on the size of
the project and outcomes required.
Generally where the model results will be used for planning processes on larger scale
catchments more detailed models are appropriate. As indicated earlier, conceptual lumped
parameter models are generally the models of choice in Australia.
More complex models than these tend to be data intensive but should be considered when the
results (generated flows or some other parameter such a water quality constituent) may be
heavily biased if a simpler method is used to model rainfall-runoff.
The conceptual rainfall-runoff models can generally be calibrated to reproduce the daily
observed streamflow well and the transfer of parameter values from a gauged catchment
nearby can model runoff reasonably well in ungauged areas.
Current areas of research that are likely to further develop best practice in rainfall runoff
modelling are:

42

More advanced methods for regionalisation of rainfall runoff model parameters from
gauged to ungauged catchments;

Development of more sophisticated approaches for use of ensembles of results from


several different rainfall runoff models to provide an optimum estimate of the flow time
series;

Development of more sophisticated and robust approaches for using parameter sets
derived from multiple gauged (donor) catchments in ungauged catchments;

Use of remotely sensed data sets, such as actual evapotranspiration, soil moisture and
vegetation cover to inform the structure, parameter values and calibration of rainfall
runoff models.

Guidelines for rainfall-runoff modelling

References
Arora, V. K., 2002. The use of the aridity index to assess climate change effect on annual
runoff. Journal of Hydrology, 265, 164 177.
Brdossy, A. (2007) Calibration of hydrological model parameters for ungauged catchments,
Hydrology and Earth System Sciences, 11 (2): 703-710
Black, D., Wallbrink, P., Jordan, P., Waters, D., Carroll, C., Blackmore, J. (2011). Guidelines
for water management modelling: Towards best-practice model application.
eWater Cooperative Research Centre. ISBN 978-1-921543-46-3.
Beven, K. (1989). Changing ideas in hydrology the case of physically-based models. Journal
of Hydrology, 105, pp 157-172.
Boughton, W. C. (2004). The Australian water balance model. Environmental Modelling and
Software, 19, 943956.
Boyle, D.P., Gupta, H.V. and Sorooshian, S. (2000) Toward improved calibration of hydrologic
models: Combining the strengths of manual and automatic methods, Water
Resources Research, 36 (12): 3663-3674.
Budyko, M. I. (1958). The Heat Balance of the Earth's Surface, translated from Russian by N.
A. Stepanova, 259 pp. U.S. Dep. of Commerce., Washington.
Burnash, R. J. C., Ferral, R. L. and McGuire, R. A. (1973). A Generalised Streamflow
Simulation System Conceptual Modelling for Digital Computers. Joint Federal
and State River Forecast Center, Sacramento, Technical Report, 204 pp.
Burnash, R.J.C. (1995). The NWS River Forecast System - Catchment Modeling, pp311-366
in V.P.Singh (ed) Computer Models of Watershed Hydrology, Water Resources
Publications (Colarado, US).
Hydrology, 311-366.Chiew, F. H. S., Peel, M. C. and Western, A. W. (2002). Application and
testing of the simple rainfall-runoff model SIMHYD. In: Singh, V. P. and Frevert,
D. K., (eds.), Mathematical Models of Small Watershed Hydrology and
Applications. Water Resources Publications, Littleton, USA, pp. 335367.
Chiew, F.H.S. (2010). Lumped conceptual rainfall-runoff models and simple water balance
methods: overview and applications in ungauged and data limited regions.
Geography Compass, In Press.
CRCCH (2005a), General approaches to modelling and practical issues of model choice.
Cooperative Research Centre on Catchment Hydrology, (Canberra).
http://www.toolkit.net.au/tools/ModelChoice/MC-2.pdf
CRCCH (2005b), Water Quality Models: Sediments and Nutrients, Cooperative Research
Centre on Catchment Hydrology, (Canberra).
http://www.toolkit.net.au/tools/ModelChoice/MC-1.pdf
Croke, B. F. W., Andrews, F., Jakeman, A. J., Cuddy, S. M. and Luddy, A. (2006). IHACRES
Classic Plus: A redesign of the IHACRES rainfall-runoff model. Environmental
Modelling and Software, 21, 426427.

Guidelines for rainfall-runoff modelling

43

Duan, Q. Y., Gupta, V. K. and Sorooshian, S. (1993). Shuffled complex evolution approach for
effective and efficient global minimization. J. Optimiz. Theory Appl. 76(3),
501521.
Duan, Q, H.V. Gupta, S. Sorooshian, A.N. Rousseau and R Turcotte (ed.) (2003). Calibration
of Watershed Models., Water Science and Application No 6, American
Geophysical Union, Washington D.C , USA.
Fu, B.P., 1981. On the calculation of the evaporation from land surface (in Chinese). Scientia
Atmospherica Sinica, 5, 23 31.
Gupta, H. V., S. Sorooshian, and P. O. Yapo, (1998) Toward improved calibration of
hydrologic models: multiple and noncommensurable measures of information,
Water. Resourc.Res., 34:751-763.
Jeffrey, S. J., Carter, J. O., Moodie, K. B. and Beswick, A. R. (2001). Using spatial interpolation
to construct a comprehensive archive of Australian climate. Environmental
Modelling and Software, 16, 309330.
Jones, D. A., Wang, W. and Fawcett, R. (2009). High-quality spatial climate data-sets for
Australia, Australian Meteorological and Oceanographic Journal 58:4, pp.
233-248.
Kavetski, D and Kuczera, G. and Franks, S.W. (2006a) Bayesian analysis of input uncertainty
in hydrological modeling: 1. Theory, Water Resources Research, 42 (3):
W03407.
Kavetski, D and Kuczera, G. and Franks, S.W. (2006b) Bayesian analysis of input uncertainty
in hydrological modeling: 2. Application, Water Resources Research, 42 (3):
W03408.
Kuczera, G, Kavetski D, Franks S and Thyer M (2006) Towards a Bayesian total error analysis
of conceptual rainfall-runoff models: Characterising model error using
storm-dependent parameters, J. Hydrology 331(1-2): 161-177.
Kavetski,D., S.W. Franks and G. Kuzera (2003) Confronting Input Uncertainty in
Environmental Modelling. In Calibration of Watershed Models, Water Science
and Application No 6, American Geophysical Union, Washington D.C , USA.
pp49-68
Liang, X., Lettenmaier, D.P., Wood, E.F., and Burges, S.J., 1994. A Simple hydrologically
Based Model of Land Surface Water and Energy Fluxes for GSMs, J. Geophys.
Res., 99(D7), 14,415-14,428.
Madsen, H. (2000) Automatic calibration of a conceptual rainfall-runoff model using multiple
objectives, J. Hydrology, 235 (3-4): 276-288.
Merz, R., and Bloschl, G. (2004). Regionalisation of catchment model parameters, Journal of
Hydrology, 287, 95-123.
Nash, J.E. and Sutcliffe, J. V. (1970). River flow forecasting through conceptual models, I, A
discussion of principles. J. Hydrol, 10, 282290.
NIST and SEMATECH (2010). NIST/SEMATECH e-Handbook of Statistical Methods,
http://www.itl.nist.gov/div898/handbook/, United States Department of
Commerce, Last updated 23 June 2010, Accessed 7 July 2011.

44

Guidelines for rainfall-runoff modelling

Northcote, K. H. with Beckmann, G. G., Bettenay, E., Churchward, H. M., Van Dijk, D. C.,
Dimmock, G. M., Hubble, G. D., Isbell, R. F., McArthur, W. M., Murtha, G. G.,
Nicolls, K. D., Paton, T. R., Thompson, C. H., Webb, A. A. and Wright, M. J.
(1960-1968). Atlas of Australian Soils, Sheets 1 to 10. With explanatory data
(CSIRO Aust. and Melbourne University Press: Melbourne).
Oudin, L., Andreassian, V., Perrin, C., Michel, C. and Le Moine, N. (2008). Spatial proximity,
physical similarity, regression and ungauged catchments: A comparison of
regionalisation approaches based on 913 French catchments. Water Resources
Research, 44, W03413, doi:10.1029/2007WR006240.
Parajka, J., Merz, R. and Bloschl, G. (2005). A comparison of regionalisation methods for
catchment model parameters, Hydrology and Earth System Sciences, 9,
157-171.
Perrin, C., Michel, C. and Andreassian, V. (2003). Improvement of a parsimonious model for
streamflow simulations. Journal of Hydrology, 279, 275289.
Refsgaard and Henriksen (2004) Modelling guidelines--terminology and guiding principles,
Advances in Water Resources 27, 71-82.
Rosenbrock, H. H. (1960). An automatic method for finding the greatest or least value of a
function, Computer Journal, 3, 175184.
Stenson, M. P, Littleboy, M. and Gilfedder, M. (2011). Estimation of water and salt generation
from unregulated upland catchments. Environmental Modelling & Software, 26,
1268-1278.
Thyer, M.A., Renard, B., Kavetski, D., Kuczera, G. and Srikanthan, S. (2007) Bayesian Total
Error Analysis For Hydrological Models: Preliminary Evaluation Using Multi-Site
Catchment Rainfall Data, In Oxley, L. and Kulasiri, D. (eds) Proc. MODSIM 2007
International Congress on Modelling and Simulation. Modelling and Simulation
Society of Australia and New Zealand, December 2007, pp. 74-80. ISBN :
978-0-9758400-4-7.
Van Dijk, A. I. J. M. (2010), AWRA Technical Report 3. Landscape Model (version 0.5)
Technical Description, WIRADA / CSIRO Water for a Healthy Country Flagship,
Canberra
(http://www.clw.csiro.au/publications/waterforahealthycountry/2010/wfhc-aus-wa
ter-resources-assessment-system.pdf).
Vaze, J., Barnett, P., Beale, G.T.H., Dawes, W., Evans, R., Tuteja, N.K., Murphy, B., Geeves,
G. and Miller, M. (2004). Modelling the effects of landuse change on water and
salt delivery from a catchment affected by dryland salinity in south-east Australia,
Hydrological Processes, Vol. 18, pp. 1613 - 1637.
Vaze, J., Post, D. A., Chiew, F. H. S., Perraud, J.-M., Viney, N., Teng, J., (2010) Climate
nonstationarity - Validity of calibrated rainfall-runoff models for use in climate
change studies, J. Hydrol. (2010), doi:10.1016/j.jhydrol.2010.09.018.
Vaze J, Chiew FHS, Perraud JM, Viney NR, Post DA, Teng J, Wang B, Lerat J, Goswami M.
(2011a). Rainfall-runoff modelling across southeast Australia: datasets, models
and results. Australian Journal of Water Resources, Vol 14, No 2, pp. 101-116.

Guidelines for rainfall-runoff modelling

45

Vaze, J., Post, D. A., Chiew, F. H. S., Perraud, J.-M., Teng, J., Viney, N. (2011b). Conceptual
RainfallRunoff Model Performance with Different Spatial Rainfall Inputs, Journal
of Hydrometeorology, Vol. 12, No. 5. pp 1100-1112, doi:
10.1175/2011JHM1340.1.
Vaze, J., Perraud, J-M., Teng, J., Chiew, F.H.S., Wang, B. (2011c). Estimating regional model
parameters using spatial land cover information implications for predictions in
ungauged basins. 19th MODSIM Congress, Perth, Australia, 12-16 December
2011. http://mssanz.org.au/modsim11.
Vaze, J., Perraud, JM., Teng, J., Chiew, F. H. S., Wang, B., Yang, Z. (2011d). Catchment
Water Yield Estimation Tools (CWYET). 34th IAHR World Congress, 27th June to
1st July, Brisbane, Australia.
Viney N, Vaze J, Chiew F, Perraud, J, Post D and Teng J. (2009). Comparison of multi-model
and multi-donor ensembles for regionalisation of runoff generation using five
lumped rainfall-runoff models in: 18th World IMACS Congress and MODSIM09
International Congress on Modelling and Simulation, July 2009, Cairns:
Modelling and Simulation Society of Australian and New Zealand and
International Association for Mathematics and Computers in Simulation:
3428-3434.
WMO (2008) Guide to Hydrological Practices. Volume I: Hydrology From Measurement to
Hydrological Information. WMO No. 168. 6th ed. World Meteorological
Organisation, Geneva, Switzerland. ISBN 978-92-63-10168-6.
WMO (2009) Guide to Hydrological Practices. Volume II: Management of Water Resources
and Application of Hydrological Practices. WMO No. 168. 6th ed. World
Meteorological Organisation, Geneva, Switzerland. ISBN 978-92-63-10168-6.
Zhang, L., Hickel, K., Dawes, W.R., Chiew, F.H.S., Western, A.W., Briggs, P.R., 2004. A
rational function approach for estimating mean annual evapotranspiration. Water
Resources Research, 40(2), W02502.
Zhang, Y. Q. and Chiew, F. H. S. (2008). Can remote sensing data improve short-term
rainfall-runoff simulation?, paper presented at Water Down Under 2008, April
2008, Adelaide: Engineers Australia.
Zhang, Y. and Chiew, F. H. S. (2009). Relative merits of different methods for runoff
predictions in ungauged catchments. Water Resources Research, 45, W07412,
doi:10.1029/2008WR007504.

46

Guidelines for rainfall-runoff modelling

eWater Cooperative Research Centre


eWater Limited ABN 47 115 422 903
Innovation Centre, Building 22
University Drive South
Bruce, ACT, 2617, Australia
T: +61 2 6201 5168
contact@ewater.com.au
www.ewater.com.au
2012 eWater Ltd

S-ar putea să vă placă și