FLOOD EVALUATION, HAZARD

DETERMINATION AND RISK

MANAGEMENT

INTERNATIONAL COMMISSION ON LARGE DAMS COMMISSION INTERNATIONALE DES GRANDS

BARRAGES
61, avenue Kléber, 75116 Paris
Téléphone : (33-1) 47 04 17 80 - Fax : (33-1) 53 75 18 22 http://www.icold-cigb.org./
COMMITTEE FLOOD EVALUATION & DAM SAFETY

COMITE ESTIMATION DES CRUES ET SECURITE DES

BARRAGES

Chairman/Président
Canada M. TREMBLAY

Vice Chairman/Vice-Président
Switzerland/Suisse B. JOOS

Active Members/Membres actifs

Australia / Australie P. HILL
Canada (Co-opted) Z. MICOVIC
N. GAUTAM
China / Chine Y. HUANG
Czech Republic / République tchèque R. KUCERA
Dominican Republic / République dominicaine J.M. ARMENTEROS RUIS
France E. PAQUET
Germany / Allemagne U. MULLER
Iran M. NOURY
Ireland / Irlande B. O’MAHONY
Italy / Italie A. BONAFE
Japan / Japon M. KASHIWAYANAGI
G. ORUKAWA
Korea / République de Corée K.S. JUN
Mexico / Mexique H. MARENGO
Netherlands / Pays-Bas J.H. VAN DUIVENDIJK
South Africa / Afrique du Sud J. SMITHERS
Spain / Espagne L. BERGA
Sweden / Suède A. SODESTROM
United Kingdom / Royaume-Uni A. WARREN
United States / États-Unis K. NG
TABLE OF CONTENTS

FOREWORD
TABLES
FIGURES
1. INTRODUCTION
2. FLOOD VOLUME
2.1. INTRODUCTION
2.2. GENERALITIES
2.3. CONSIDERATIONS ABOUT FLOOD PROCESS
2.4. ANALYSIS OF FLOOD HYDROGRAPH
2.5. FLOOD RECESSION ISSUE
2.5.1 Chiumbe River – Angola
2.5.2 Mistassibi River – Canada
2.6. FLOOD VOLUME – STATISTICAL ANALYSIS
2.7. HYDROGRAPH RECONSTITUTION – RAINFALL EVENT(S)
2.8. HYDROGRAPH RECONSTITUTION – COMPLEX EVENTS
2.9. RELATION BETWEEN FLOOD PEAK AND VOLUME
2.10. DETERMINISTIC APPROACH
2.11. STOCHASTIC MODELING
2.12. FLOOD VOLUME – EXTREME VALUES
2.13. IMPACT OF CLIMATE CHANGE ON FLOOD VOLUME
2.14. RECOMMENDATIONS
2.15. REFERENCES
3. STOCHASTIC APPROACH TO FLOOD HAZARD DETERMINATION
3.1. INTRODUCTION
3.2. BASIC PRINCIPLES OF STOCHASTIC APPROACH TO
FLOOD HAZARD
3.3. MAIN ASPECTS OF STOCHASTIC FLOOD HAZARD
MODELLING FOR DAM SAFETY
3.3.1 Stochastic simulation of reservoir inflows
3.3.1.1 Storm seasonality
3.3.1.2 Precipitation magnitude‐frequency relationship
3.3.1.3 Temporal and spatial distribution of storms
3.3.1.4 Air temperature and freezing level temporal patterns
3.3.1.5 Watershed model antecedent conditions sampling
3.3.1.6 Initial reservoir level
3.3.2 Simulation of reservoir operation – flood routing
3.3.3 Stochastic simulation of the availability of discharge facilities
3.3.4 Simulation procedure
3.4 ESTIMATING UNCERTAINTY IN STOCHASTIC FLOOD
MODELLING
3.4.1 Sensitivity analyses
3.4.2 Uncertainty analyses
3.4.3 Characterization of uncertainties for selected model
components
3.4.3.1 The 48‐hour basin‐average precipitation‐frequency
relationship for watershed
3.4.3.2 Watershed response to fast runoff
3.4.3.3 The 1,000 mb air temperature and freezing level
3.4.3.4 Storm seasonality where precipitation magnitudes are
unrestricted
3.5 STOCHASTICALLY DERIVED FLOOD FREQUENCY
RESULTS WITH UNCERTAINTY BOUNDS
3.6 CONCLUDING REMARKS
3.7 REFERENCES
4 RESERVOIR INFLOW PREDICTION FOR PROACTIVE FLOOD RISK
MANAGEMENT
4.1 FUNDAMENTALS AND BASIC PRINCIPLES OF HYDROLOGICAL
AND HYDRO-METEOROLOGICAL INFLOW PREDICTION
4.1.1 Forecasting/prediction methods
4.1.2 Decision making based on potentially uncertain forecasts
4.2 PREDICTION-BASED FLOOD RISK MANAGEMENT ON DIFFERENT
TEMPORAL SCALES
4.2.1 Long-term scale
4.2.1.1 Inflow forecasting/prediction strategies
4.2.1.2 Excursus: Multi-criteria optimization
4.2.1.3 Considering uncertainty on the long-term scale
4.2.2 Seasonal scale
4.2.2.1 Inflow forecasting/prediction strategies
4.2.2.2 Considering uncertainty on the seasonal scale
4.2.3 Short-termed, single-event scale
4.2.3.1 Prediction strategies and aspects of operational short-
term inflow forecasting
4.2.3.2 Excursus: Assessing the benefit of a forecast
depending on lead time
4.2.3.3 Considering uncertainty on the short-term scale
4.3 DISCUSSION AND RECOMMENDATIONS
4.4 CASES STUDY
4.5 REFERENCES

APPENDIX A RESERVOIR INFLOW PREDICTION FOR

PROACTIVE FLOOD RISK MANAGEMENT -
CASE STUDIES
FOREWORD

Design criteria for the construction or rehabilitation of dams and

appurtenant structures generally starts with a section defining the design flood.
Methods for determining flood design have evolved over the years, but these
must continue to progress given the degree of uncertainty in the assessment of
extreme floods and the fact that factors external to floods must be considered in
the risk assessment.

Five bulletins have been published by the Flood Evaluation and Dam
Safety Committee since its formation:

- Project Flood Selection - Current Methods (No 82, 1992);

- Dams and Floods - Guidelines and Case Studies (No. 125, 2003);
- Role of Dams in Flood Prevention - Summary (No. 131, 2006);
- Integrated Flood Management (No. 156, 2010);
- Flood Assessment and Dam Safety (No. 170, 2018).

This bulletin follows the lead of the previous bulletin. It discusses issues
of concern to committee members and we hope it will allow users to better
understand some of the challenges ahead, approaches to solve problem
encountered in this field of expertise and future trends.

It consists of three main chapters following the introductory chapter.

Chapter Two discusses the main aspects related to the volume of floods. In
general, a flood is often associated with the consequences due to its the peak
flow. However the volume of floods is also an important aspect to consider.

Chapter Three is a follow-up to the previous bulletin, addressing in more

detail the stochastic approaches to flood risk assessment. These analyses
include factors that are independent of flood but may have an impact on the safety
of the dam(s) in the system under study.

Finally, the last chapter deals with the forecast aspects related to the
proactive management of floods. Case studies to illustrate short-, medium- or
long-term management challenges are presented in Appendix A of this
document.

Il It should be mentioned that all members of the committee participated

in one way or another in the design of the chapters of this bulletin (work done
mostly on a voluntary basis). I thank them personally.
I must highlight the contribution of the main authors of each chapter. By
sharing their expertise and enthusiasm to work in this area, their work has
become the core of this bulletin. Those are :

- Chapter 2 Bernard Joos Switzerland

Luis Berga Spain

- Chapter 3 Zoran Micovic Canada

- Chapter 4 Uwe Muller Germany

Case Study Masayuki Kashiwayanagi Japan

Michel Tremblay
Technical Committee Chairman
Flood Evaluation and Dam Safety Committee

December 2019
TABLES

Table 3.1 Hydrometeorological Inputs to SEFM for BC Hydro’s Campbell

River System
Table 3.2 Qualitative sensitivity of the La Joie Dam maximum reservoir
level to various hydrometeorological inputs and watershed
model parameters
Table 3.3 Five model components selected for inclusion in the uncertainty
analysis for derivation of flood frequency curves and uncertainty
bounds for the La Joie Dam watershed
Table 3.4 Five model components selected for inclusion in the uncertainty
analysis for derivation of flood frequency curves and uncertainty
bounds for the La Joie Dam watershed
Table 3.5 Restrictions on 48‐hour basin‐average precipitation expressed
as percentage of PMP

Table 4.1 Temporal scales and their connection to typical reservoir

management problems, associated with inflow estimation and
flood risk management
Table 4.2 Suitability of different methods for inflow prediction and
forecasting according to the temporal scale of interest

Table A1.1 (a) Supply Rates and required reliabilities for the three supply
levels and (b) the diversion rates from reservoir RB for the three
diversion levels.
Table A1.2 Summary of reliability assessment results for the three supply
levels R(SL1) to R(SL3) and the fitness functions FF1 to FF3 for
selected solutions for the scenarios (A1B, B1, B2) in 2021-2050
and under recent conditions (status quo). Reliabilities, which do
not meet target reliabilities according to Table A1.2(a) are
marked in red color.
Table A3.1 Criteria for proactive drawdown operation for flood mitigation
Table A3.2 Verification of the criteria of drawdown operation
FIGURES

Figure 2.1 Simulated floods with water exchanges aquifer-river

Figure 2.2 Flood hydrographs from rainfalls of various extents, same
volume
Figure 2.3 Flood hydrographs from rainfalls of various patterns, same
volume
Figure 2.4 Flood hydrographs from rainfalls of same pattern, various
volumes
Figure 2.5 Chiumbe River – Flood Hydrographs and Recession Patterns
Figure 2.6 Mistassini River – Typical Spring Flood Hydrographs
Figure 2.7 Mistassibi River (1963-2013) –
Spring Flood Maximum daily volume – Statistical analysis
Figure 2.8 Mistassibi River – Spring Flood Volume – Statistical analysis
Figure 2.9 Mistassibi River – Spring Flood –Volume vs Number of days
Figure 2.10 Classical (left) and proposed (right) approach to
considering a flood
Figure 2.11 Equivalence of the flood volumes
Figure 2.12 Observed and FIM-simulated Mistassini 1976 flood
Figure 2.13 Mistassibi River – Spring Flood Hydrograph –
Without consideration of recession characteristics
Figure 2.14 Mistassibi River – Spring Flood Hydrograph – With
consideration of recession characteristics
Figure 2.15 Mistassibi River – Comparison of the spring flood volume for
different durations
Figure 2.16 Flood Volumes - Envelope curve of extreme flood volumes –
Kv = 10.5
Figure 2.17 Relationship between specific volume (Vs) and catchment area

Figure 3.1 Probability plot and frequency histogram of storm seasonality

for the Campbell R. watershed in Canada
Figure 3.2 Computed 72‐hour precipitation‐frequency curve and 90%
uncertainty bounds for the 1193 km 2 Strathcona Dam
watershed
Figure 3.3 SCHADEX concept of centered rainfall event
(with central and adjacent rainfalls)
Figure 3.4 Spatial storm template (October 1984 storm)
Figure 3.5 Temporal storm template (October 2003 storm)
Figure 3.6 Range of 12‐hour persisting 1,000 mb dewpoint temperatures
utilized by dewpoint temperature probability model (December
example)
Figure 3.7 Air temperature lapse‐rates for day of maximum 24‐hour
precipitation for storms in Northwestern Washington and
Central California
Figure 3.8 Indexed 1,000 mb temperature and freezing level temporal
patterns for the Oct. 2003 storm
Figure 3.9 Stochastic flood simulation flow chart for the 1193 km 2
Strathcona Dam watershed in British Columbia, Canada
Figure 3.10 Scatterplot showing moderate to high sensitivity of peak
reservoir inflow to freezing level and storm temporal distribution
for the La Joie Dam watershed in Canada
Figure 3.11 Flowchart for Monte Carlo frameworks for stochastic flood
analysis and uncertainty analysis
Figure 3.12 Comparison of predicted and measured 48‐hours basin‐
average precipitation for Bridge River watershed based on
multiple‐regression prediction equation
Figure 3.13 Computed 48‐hour precipitation‐frequency relationship and
90% uncertainty bounds for the Bridge River basin
Figure 3.14 Likelihood functions for fast runoff timing constant for Bridge
River watersheds
Figure 3.15 La Joie reservoir 72‐hour inflow magnitude‐frequency
calibration (SEFM and observed inflow)
Figure 3.16 Example of simulated flood‐frequency relationships for peak
hourly reservoir inflow at La Joie Dam for eleven plausible
model configurations
Figure 3.17 La Joie Dam – Frequency curve for peak hourly reservoir inflow
simulated by stochastic flood model
Figure 3.18 La Joie Dam – Frequency curve for peak hourly reservoir level
simulated by stochastic flood model
Figure 4.1 Conceptual illustration of the dependency of predictive
uncertainty on increasing lead time and decreasing spatial
scale.
Figure 4.2 Illustration of the Pareto principle. Figure modified after Müller,
2014
Figure 4.3 Statistical Moments of monthly flow data - Modified after Müller
(2014).
Figure 4.4 Convergence of the standard deviation of synthetic time series
(y axis) depending on the number of generated series (given on
the x axis). Modified after Grundmann (2010).
Figure 4.5 Quantile-based flow climatology for river gauge Görlitz at
Lausitzer Neiße River, Germany
Figure 4.6 An ESP illustrated by J. C. Schaake in 1978, from “Extended
streamflow prediction techniques: description and applications
during 1977”, taken from Wood et al., 2016
Figure 4.7 AUC skill score for two different QPF products (COSMO-DE
and COSMO-DE-EPS) depending on different lead times
[1, 3, 6, 9, 12, 15, 18, 21 hrs] and different spatial scales
(from 43,200 km² down to 10.55 km²).
Taken from Schütze et al. (2016)
Figure 4.8 Operation of Malter reservoir during the 2002 European Floods
under different pre-release scenarios
Figure 4.9 Simple assessment of the forecast benefit (expressed as peak
reduction) using a release strategy based on available look-
ahead time (lead time)
Figure 4.10 Benefit (peak flow reduction) as a function of lead time (data
according to Figure 4.9)
Figure 4.11 Comparison of a raw ensemble (upper panel) and a processed
ensemble as “true” estimate of the predictive uncertainty
(below). From Klein et al. (2016)

Figure A1.1 Comparison of monthly and annual statistical properties of

exemplary inflow data (inflow to Lehnmühle Reservoir) for 120
simulations from the KNN model (circles), historical flows (solid
gray circles) and “shortened” (i.e., resampled) time series
(diamond). Taken from Müller and Schütze (2017).
Figure A1.2 (a) Map of the reservoir system in the federal state Saxony in
Germany and (b) illustration of the reservoir system in the
simulation model OASIS.
Figure A1.3 Modified Level-diagram showing the Pareto fronts for all
optimization epochs: recent conditions (blue); 2021-2050
conditions A1B (red), A2 (pink) and B1 (gray).
Figure A2.1 Investigation area and precipitation ground stations
Figure A2.2 Use of indices without forecasts
Figure A2.3 Use of indices with seasonal circulation forecasts from NOAA
Figure A2.4 SPI calculated with observed / NOAA forecast data, examples
from different sites
Figure A2.5 Storage volume with/without SPI-based hedging rule
(blue=with, red=without)
Figure A3.1 Dams in Kumano river basin
Figure A3.2 Ikehara dam and its reservoir
Figure A3.3 Flood mitigation scheme
Figure A3.4 Modification of reservoir operation rule
Figure A3.5 Criteria for typhoons as shown in Table A3.1
Figure A3.6 GPV locations for the forecast of precipitation in the basin
Figure A3.7 Simulation of the Ikehara reservoir operation
1. INTRODUCTION

EXTREME FLOODS AND THEIR CONSEQUENCES

As part of the hydrological studies required to assess and/or ensure the safety of
dams, dikes and hydraulic structures, we spend a lot of time evaluating the
characteristics of floods and more particularly peak inflows in the system.

It should be remembered that the primary purpose of dam safety studies is rather
to evaluate the potential consequences when the dam and other components of
the system are confronted with extreme events and that does not only depend on
the peak discharge during major floods.

Knowledge of the system as well as the parameters that may impact the security
of its components is essential. For example, the routing capacity of the system
(run-of-river, daily, seasonal, annual, or multi-year reservoir) plays a role in the
system's ability to respond to major floods. A small reservoir (small in comparison
to its inflows) will be more sensitive to the peak discharge during a major flood
than the flood volume; the time to react to such event could be relatively short. A
large reservoir will normally be more sensitive to the flood volume, depending on
its routing capacity. This will allow more time to react to the flood, however the
consequences related to a failure could be much more important.

Other factors must also be considered, factors that will be more or less
homogeneous and that will depend, in part, on characteristics not controlled by
the dam owners. We will mention here the initial conditions on the watershed, the
temporal distribution and intensity of the precipitation on the basin, the snow
cover and the temperature of the air (if applicable), the vegetation and the rate of
urbanization, ...

As previously mentioned, knowledge of the system under study is therefore a

critical factor in assessing the consequences of extreme events and determining
the best way to respond to these events. To do this, the system must be well
documented while considering that the situation will continue to evolve over time
and the main trends must be identified.

If the knowledge of the system is essential, we often have to face limitations

resources, time and knowledge available to perform the analysis.

Not having the same resources or knowledge of a system does not mean that we
must accept more risk, but rather it could mean that we have to be more
conservative in the design or the maintenance of the system, which also have an
impact.
If we have resources and time required to perform more detailed analyses, we
still have to ask ourselves a few questions :

- Is our knowledge of the different variables having an impact on dam

safety sufficient to establish a realistic distribution? What are the
correlations between a parameter and the other parameters?

- If we are evaluating various scenarios for a parameter (let’s consider

parameters related to climate changes), are all scenarios
equiprobable?

We know that we will not be able to consider all the combinations causing a failure
of the system, however are we realistic in our risk estimates for the scenario
identified ?

Aristotle stated that « the true and the approximately true are apprehended
by the same faculty; it may also be noted that men have a sufficient natural
instinct for what is true, and usually do arrive at the truth. Hence the man who
makes a good guess at truth is likely to make a good guess at probabilities. »

However, to ensure the safety of a dam, a good instinct is not enough and
understanding of all the components and risks to the system is essential.

THE BULLETIN

This bulletin is divided in three main chapters.

Chapter 2 focuses on an aspect of floods that is often more important for large
reservoirs, i.e. flood volume. For small reservoirs with low regulation capabilities,
peak flow is the most important feature, as the reservoir can not have a significant
impact on outflow. On the other hand, for larger reservoirs, an adequate
assessment of the flood volume and its relation to the peak flow is essential. This
chapter reviews some aspects related to flood volumes and provides a specific
point of view on some approaches or perceptions, for example:

- The effect of the temporal distribution of the precipitation over the

peak discharge and the flood volume is not always as important as
is commonly believed over the peak discharge and the flood volume;

- The recession phase of a flood on a watershed depends on the

physical characteristics of the basin and the pattern does not depend
on the size of the flood;

- It is possible to build a hydrograph caused by two distinct events by

knowing the characteristics of the flood on a watershed;
- It is often common practice to consider that the peak discharge of a
major flood and the flood volume have a similar recurrence period.
This may be true for floods caused by a single event, but it seems to
be less so for floods caused by a combination of events.

- The large floods observed around the world can serve for
comparison purpose to validate or predict the characteristics of large
floods on other watersheds.

Chapter Three is a follow-up to our previous technical bulletin (No 170), reviewing
the current trends in the evaluation of extreme flood. While the previous bulletin
reviewed possible approaches with a particular focus on stochastic approaches,
this bulletin focuses exclusively on stochastic approaches, describing in more
detail the different phases of such studies.

Stochastic approaches do not focus only on uncertainties related to floods and

their characteristics (peak, volume, hydrograph), but rather on the consequences
following extreme floods, i.e. in this context at the level reached in the reservoirs
and the consequential risks for the various components of the system.

Validation of the results is also discussed in the chapter. Although it is difficult to

validate the results obtained for extreme floods (given the low probability of these
events), it is easier to validate the proposed approach for floods of more common
periods of recurrence. Aspects related to sensitivity analyses and uncertainties
are also addressed. The uncertainties associated with stochastic studies will be
addressed specifically in the next bulletin.

The final chapter addresses the challenge of proactive flood forecasting to

minimize potential consequences and damage. The chapter is more focused on
the operation of existing systems, so it is more for operators who are looking to
improve the management of their systems.

This chapter introduces first the basics and fundamentals of proactive flood
management. Then the different horizons of analysis are considered, either short,
medium or long term. Different approaches are then discussed, i.e. by moving
from analyzes based on the observed history (long term), to the forecast of a
single event (short term). In addition, there is nothing preventing stochastic
analyses to be done depending on the study horizon in order to evaluate the
potential risks.

As the preceding chapters show, a flood with the same peak flow can have very
different consequences for a given system. Good forecasting of inputs and proper
management of the water system can limit the consequences of an event.

Appendix A presents case studies provided by the committee’s members.

TO CONCLUDE

We all know that there is room for improvement in the field of dam safety. In the
field of hydrology and flood assessment, we must face phenomena continuously
changing. We must rely on the data of the past, our present and imperfect
knowledge of the phenomena involved and the trends we anticipate for the future
to assess the risks we must face. Some of these elements are more obvious,
such as:

Climate changes :

- Non-stationarity of the various aspects related to flood assessment;

- The difficulty of establishing the impacts of climate change for

extreme events or for combinations of events;

- The probabilities related to the different scenarios studied.

The uncertainties related to the analyzes carried out:

- Limited information mainly for extreme hydrological conditions and

for combinations of various conditions (including hydrological
conditions, mechanical failures, human factors and others).

- The links or correlation between these parameters;

- The different ways to take into account uncertainties, including:

o Security factors;

o Assumptions and approximations;

o and others

Some of these elements will be addressed in the next bulletin, where they may
be addressed by or in partnership with other committees. This applies more
specifically to all aspects of climate change that require expertise beyond the
scope of this committee.

In conclusion, we could say that we must continue to progress in the field of dam
safety. With respect to this committee, we need to continue to develop our
knowledge in flood assessment for dam safety and related events that may
endanger dams and hydraulic structures. Each step forward shows us that our
knowledge and tools are limited, but we must look to the future, since we still
have so much to learn.
As Albert Einstein said : « Learn from yesterday, live for today, hope for
tomorrow. The important thing is not to stop questioning ».
2. FLOOD VOLUME

2.1. INTRODUCTION

Uncertainties in the determination of floods sometimes substantially affect the

design of hydraulic structures. A common source of errors is due to the
imprecision of hydrological measurements. Another source, of different nature,
may also lead to serious biases, not being safe from inconsistencies. It is closely
linked to the treatment and interpretation of the recorded data.

Flood determination usually includes the definition of the river flow hydrograph
during and after a storm. The attention usually focuses on the peak of the flood
and not so much on its volume; the recession behaviour of the river flow is
accepted as is, in general without further consideration. The river base flow at the
beginning of the flood draws even less attention.

These elements however play an important role in the intensity of the flood and
the overall shape of its hydrograph.

2.2. GENERALITIES

Flood determination through stochastic approaches is one of the main trends to

evaluate floods occurring on a specific drainage area. Stochastic approaches for
flood determination are reviewed in chapter 3 of the present bulletin. However,
because of the overall resources required for such analyses, these stochastic
approaches are preferably performed on major and well-known systems. They
would be less attractive to implement on small systems or for new dam projects.

Presently, most of the guidelines of different countries are based on the results
of statistical flood analyses or the deterministic evaluation of the PMF. Most of
the papers related to the statistical analyses of floods focus, for small drainage
areas, on the peak discharge or the determination of the hydrograph for a single
rainfall event for small drainage areas. As mentioned by Guillaud (2002):

“The entire discussion so far has focused on flood peak, which is the easiest
characteristic on the flood to obtain. Another important characteristic of a flood,
perhaps more important than the peak is the volume. However, it is usually
difficult to determine the volume of the flood caused by a meteorological event
without ambiguity, because of the interference with other meteorological events.
An alternative is to define the runoff volume over certain durations (5 days, 10
days, 30 days, …) and to reconstruct a hydrograph for each frequency.

Flood peak is essential to know only for the design of hydraulic structures for
projects with no or little storage capacity, for example for run-of-river projects.
When a project comprises a reservoir, it is important to know the volume of the
flood in order to take the storage effect into account in the evaluation of the size
of the structures.”
Similarly, the Canadian Dam Association Guidelines mentioned that “Statistical
analysis is required for estimating the flood peaks and volumes associated with
a range of annual exceedance probabilities (AEPs). In addition to the peaks, the
volumes and the associated hydrographs for the floods of interest are usually
required for reservoir routing or dam-breach and downstream channel routing.
This analysis is done on a seasonal basis and is of greater significance for
storage reservoirs that have large fluctuations in water levels and are designed
to capture spring runoff. For run-of-the-river facilities, only the peak annual flood
is usually required.”

It is unlikely that a direct relation between the peak and the volume of specific
floods will exist for a particular drainage area, but the two characteristics are
correlated. If statistical analyses of peak flood can be performed relatively easily
on natural discharge of a drainage area, the challenge increases when the
volume of the flood corresponding to a specific frequency (return period) must be
determined. In this case, it is not only the maximum annual (or seasonal)
discharge that must be considered, but also the volume of the flood, which will
depend on the duration of the event and will differ for each event.

For small drainage areas (say, less than 50 km2), it is very likely that the peak
discharge and the volume correspond to similar frequency, since the flood is
normally caused by a single rainfall event. For large drainage areas, the situation
is more complex, since peak discharge and volume for a specific frequency can
depend on a combination of events (such as snowmelt and rainfall event(s), in
the northern countries). Another difficulty consists in the spatial and temporal
correlation of such events over a large drainage area.

The present section will present an overview on the role of the flood volume for
dam safety, including:
- Consideration about flood process;
- Flood recession issues;
- Statistical analyses (peak / flood volume);
- Hydrograph reconstitution (rainfall / complex events);
- Extreme values (flood volume);
- Case studies;
- Recommendations.

2.3. CONSIDERATIONS ABOUT FLOOD PROCESS

When attempting to estimate the river behaviour during an important storm,

hydrologists are in general used to focus on the magnitude of the meteorological
event. The peak precipitation is considered to play an important role in this
respect. However, at least as important as the precipitation intensity is the
precipitation volume. The exact timing and location of the storm cell(s) is usually
not well known, and formulating hypotheses about their behaviour is no
guarantee to gain insight in the phenomenon. Actually, as long as the overall
volume and centre of gravity of a storm are correctly determined, the exact
location of the storm’s cells, their succession and magnitude do not play a
determinant influence on the peak discharge of locations further downstream.
Indeed the influence of local and limited meteorological events plays a minor role
in the evolution of the catchment runoff. This is especially true if the event location
is not situated in the immediate vicinity of the considered river point. Methods not
requiring the formulation of particular hypotheses may be not very versatile in
their approach and treatment of the phenomenon, but they would present the
advantage of limiting the possible biases related to these assumptions. In this
sense, sophisticated and detailed methods are not necessary the most
appropriate to treat this type of problem.

The inertia of flood processes is quite large – the movement of water masses
takes time, especially in the underlaying aquifer. Water exchanges between river
and aquifers all along its course play an important role to smooth out the river
discharge, either in absorbing water from the river, when this one has a
particularly high level, or in feeding it when its flow tends to decrease. The direct
correlation between the rainfall and the river regime is disconnected by the transit
through the aquifers.

During a flood, a given length of river contains a multiple of the volume of water
usually present on the same stretch, even during the high flow season. During a
large flood, this multiple can reach large proportions.

2.4. ANALYSIS OF FLOOD HYDROGRAPH

A simplified qualitative, one-dimensional model allows visualizing the behaviour

of a flood when aquifer and river exchange water. At each time step, water
moves:
- from upstream to downstream in the river (from left to right on the graph
below);
- from aquifer to river or conversely, depending on the relative water levels;
- between neighbour aquifers, depending on the elevation of their water
tables.
A precipitation of limited extent (in time and space) is postulated at the beginning
of the considered area. Since the vast majority of the flood is due to the water
collected by the watershed, it is assumed that the totality of the precipitation is
absorbed by the aquifer, to be transferred later to the river (there is no overland
flow). To keep the calculations simple, no losses are considered.

Figure 2.1 illustrates the progression of the wave in time and space.
10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.0 10.0 26.8 26.8 26.8 26.8 26.8 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.0 10.3 27.0 40.2 40.2 40.2 39.9 23.3 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.0 10.6 25.7 40.8 51.0 51.0 50.4 35.3 20.2 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.0 10.9 24.3 38.8 51.7 59.6 58.7 45.3 30.8 17.9 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.0 11.0 23.1 36.5 49.4 60.4 65.4 53.4 40.0 27.1 16.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.0 11.2 22.0 34.4 46.6 58.0 66.2 60.0 47.6 35.4 23.9 14.6 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.2 21.0 32.4 43.9 54.9 63.8 61.8 54.0 42.5 31.4 21.3 13.6 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.3 20.1 30.6 41.3 51.8 60.6 60.7 56.6 48.6 38.0 28.0 19.2 12.7 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.3 19.2 28.9 39.0 48.9 57.3 58.4 56.4 51.6 43.7 34.1 25.1 17.4 12.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.4 18.5 27.4 36.8 46.2 54.2 55.9 55.0 52.2 47.1 39.4 30.6 22.6 16.0 11.6 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.4 17.8 26.0 34.8 43.6 51.3 53.4 53.2 51.5 48.2 42.9 35.6 27.6 20.5 14.8 11.3 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.3 17.2 24.7 32.9 41.2 48.5 51.0 51.3 50.3 48.1 44.5 39.1 32.2 25.0 18.7 13.9 11.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.3 16.6 23.6 31.2 39.0 46.0 48.7 49.4 49.0 47.5 44.9 41.0 35.7 29.2 22.7 17.2 13.1 10.7 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.3 16.1 22.5 29.6 37.0 43.6 46.5 47.6 47.6 46.6 44.7 41.8 37.8 32.6 26.6 20.8 16.0 12.5 10.6 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 11.2 15.6 21.5 28.1 35.1 41.3 44.4 45.8 46.1 45.6 44.2 41.9 38.8 34.8 29.8 24.3 19.1 14.9 12.0 10.4 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.2 11.2 15.1 20.6 26.8 33.3 39.3 42.4 44.1 44.7 44.5 43.5 41.8 39.3 36.1 32.1 27.3 22.3 17.7 14.1 11.6 10.3 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.2 11.1 14.7 19.7 25.5 31.7 37.3 40.6 42.4 43.3 43.5 42.8 41.4 39.4 36.8 33.5 29.6 25.1 20.5 16.5 13.3 11.3 10.3 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.2 11.1 14.4 19.0 24.3 30.1 35.5 38.8 40.8 41.9 42.4 42.0 41.0 39.4 37.2 34.4 31.1 27.3 23.1 19.0 15.4 12.7 11.0 10.2 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.2 11.0 14.0 18.3 23.3 28.7 33.8 37.1 39.2 40.6 41.3 41.2 40.5 39.2 37.3 35.0 32.2 28.9 25.3 21.4 17.7 14.6 12.2 10.8 10.2 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.2 11.0 13.7 17.6 22.3 27.4 32.2 35.5 37.7 39.3 40.1 40.4 40.0 38.9 37.4 35.3 32.9 30.1 26.9 23.4 19.9 16.6 13.8 11.8 10.6 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.9 13.4 17.0 21.4 26.1 30.7 34.0 36.3 38.0 39.0 39.5 39.3 38.6 37.3 35.6 33.4 30.9 28.1 25.1 21.8 18.5 15.6 13.2 11.5 10.5 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.9 13.1 16.5 20.5 25.0 29.3 32.6 34.9 36.7 37.9 38.6 38.7 38.2 37.2 35.7 33.8 31.5 29.0 26.3 23.4 20.3 17.4 14.7 12.6 11.2 10.4 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.8 12.9 16.0 19.7 23.9 28.1 31.2 33.6 35.5 36.9 37.7 38.0 37.7 37.0 35.7 34.0 32.0 29.8 27.3 24.6 21.9 19.0 16.3 14.0 12.2 11.0 10.3 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.8 12.7 15.5 19.0 22.9 26.8 29.9 32.3 34.3 35.8 36.8 37.3 37.2 36.7 35.7 34.2 32.4 30.4 28.1 25.7 23.1 20.5 17.9 15.5 13.4 11.8 10.8 10.3 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.7 12.5 15.1 18.3 22.0 25.7 28.7 31.1 33.1 34.7 35.9 36.5 36.7 36.3 35.5 34.3 32.7 30.9 28.8 26.5 24.1 21.7 19.2 16.9 14.7 12.9 11.5 10.6 10.2 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.7 12.3 14.7 17.7 21.2 24.7 27.6 30.0 32.0 33.7 34.9 35.7 36.1 35.9 35.3 34.3 33.0 31.3 29.4 27.3 25.0 22.7 20.4 18.1 16.0 14.0 12.4 11.3 10.5 10.2 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.7 12.1 14.3 17.1 20.4 23.7 26.5 28.9 31.0 32.7 34.0 35.0 35.4 35.5 35.1 34.3 33.1 31.6 29.9 27.9 25.8 23.7 21.5 19.3 17.1 15.2 13.4 12.0 11.0 10.4 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.6 11.9 14.0 16.6 19.6 22.8 25.5 27.9 29.9 31.7 33.1 34.2 34.8 35.0 34.8 34.2 33.2 31.9 30.3 28.5 26.5 24.5 22.4 20.3 18.2 16.3 14.5 12.9 11.7 10.9 10.3 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.6 11.8 13.7 16.1 19.0 21.9 24.5 26.9 28.9 30.7 32.2 33.4 34.1 34.5 34.4 34.0 33.2 32.0 30.6 28.9 27.1 25.2 23.2 21.2 19.2 17.3 15.5 13.9 12.5 11.4 10.7 10.3 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.5 11.7 13.4 15.7 18.3 21.1 23.6 25.9 28.0 29.8 31.4 32.6 33.5 34.0 34.1 33.8 33.1 32.1 30.9 29.4 27.7 25.8 23.9 22.0 20.1 18.2 16.4 14.8 13.3 12.1 11.2 10.6 10.2 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.5 11.5 13.1 15.2 17.7 20.4 22.8 25.0 27.1 28.9 30.5 31.8 32.8 33.4 33.6 33.5 33.0 32.2 31.1 29.7 28.1 26.4 24.6 22.7 20.9 19.1 17.3 15.7 14.2 12.9 11.8 11.0 10.5 10.2 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.5 11.4 12.9 14.9 17.2 19.7 22.0 24.2 26.2 28.0 29.7 31.0 32.1 32.8 33.2 33.2 32.8 32.2 31.2 30.0 28.5 26.9 25.2 23.4 21.6 19.9 18.1 16.5 15.0 13.6 12.5 11.5 10.8 10.4 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.4 11.3 12.7 14.5 16.7 19.0 21.3 23.4 25.4 27.2 28.8 30.3 31.4 32.2 32.7 32.8 32.6 32.1 31.3 30.2 28.9 27.4 25.8 24.1 22.3 20.6 18.9 17.3 15.8 14.4 13.2 12.1 11.3 10.7 10.3 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.4 11.2 12.5 14.2 16.2 18.4 20.5 22.6 24.5 26.4 28.0 29.5 30.7 31.6 32.2 32.4 32.4 32.0 31.3 30.4 29.2 27.8 26.3 24.7 23.0 21.3 19.6 18.0 16.5 15.1 13.8 12.7 11.8 11.1 10.6 10.3 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.4 11.1 12.3 13.9 15.8 17.8 19.9 21.9 23.8 25.6 27.3 28.8 30.0 31.0 31.7 32.0 32.1 31.8 31.3 30.5 29.4 28.2 26.7 25.2 23.6 21.9 20.3 18.7 17.2 15.8 14.5 13.4 12.4 11.5 10.9 10.5 10.2 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.4 11.0 12.1 13.6 15.3 17.3 19.3 21.2 23.0 24.8 26.5 28.0 29.3 30.4 31.1 31.6 31.8 31.7 31.2 30.5 29.6 28.5 27.1 25.7 24.2 22.6 21.0 19.4 17.9 16.5 15.2 14.0 12.9 12.0 11.3 10.8 10.4 10.2 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.3 11.0 12.0 13.3 15.0 16.8 18.7 20.5 22.3 24.1 25.8 27.3 28.6 29.7 30.6 31.1 31.4 31.4 31.1 30.6 29.7 28.7 27.5 26.1 24.7 23.1 21.6 20.1 18.6 17.1 15.8 14.6 13.5 12.6 11.8 11.1 10.6 10.3 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.3 10.9 11.8 13.1 14.6 16.4 18.1 19.9 21.7 23.4 25.1 26.6 28.0 29.1 30.0 30.7 31.1 31.2 31.0 30.5 29.8 28.9 27.8 26.5 25.2 23.7 22.2 20.7 19.2 17.8 16.4 15.2 14.1 13.1 12.2 11.5 11.0 10.5 10.3 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.3 10.8 11.7 12.8 14.3 15.9 17.6 19.3 21.0 22.7 24.4 25.9 27.3 28.5 29.5 30.2 30.7 30.9 30.8 30.5 29.9 29.1 28.1 26.9 25.6 24.2 22.7 21.3 19.8 18.4 17.1 15.8 14.7 13.6 12.7 11.9 11.3 10.8 10.5 10.2 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.3 10.8 11.5 12.6 14.0 15.5 17.1 18.8 20.4 22.1 23.7 25.2 26.6 27.9 28.9 29.7 30.2 30.5 30.6 30.3 29.9 29.2 28.3 27.2 26.0 24.7 23.3 21.8 20.4 19.0 17.6 16.4 15.2 14.1 13.2 12.4 11.7 11.1 10.7 10.4 10.2 10.1 10.0 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.3 10.7 11.4 12.4 13.7 15.1 16.7 18.3 19.9 21.5 23.1 24.6 26.0 27.2 28.3 29.2 29.8 30.2 30.3 30.2 29.8 29.2 28.4 27.5 26.3 25.1 23.7 22.4 21.0 19.6 18.2 17.0 15.8 14.7 13.7 12.8 12.1 11.5 11.0 10.6 10.3 10.2 10.1 10.0 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.2 10.6 11.3 12.3 13.4 14.8 16.3 17.8 19.3 20.9 22.5 24.0 25.4 26.6 27.7 28.7 29.4 29.8 30.0 30.0 29.7 29.3 28.6 27.7 26.6 25.5 24.2 22.9 21.5 20.1 18.8 17.5 16.3 15.2 14.2 13.3 12.5 11.8 11.3 10.8 10.5 10.3 10.1 10.1 10.0 10.0 10.0 10.0 10.0
10.0 10.1 10.2 10.6 11.2 12.1 13.2 14.5 15.9 17.3 18.8 20.3 21.9 23.3 24.8 26.0 27.2 28.1 28.9 29.4 29.7 29.8 29.6 29.2 28.6 27.8 26.9 25.8 24.6 23.3 22.0 20.7 19.3 18.1 16.9 15.7 14.7 13.7 12.9 12.2 11.6 11.1 10.7 10.4 10.2 10.1 10.0 10.0 10.0 10.0 10.0
10.0 10.0 10.2 10.6 11.1 11.9 13.0 14.2 15.5 16.9 18.3 19.8 21.3 22.8 24.2 25.5 26.6 27.6 28.4 29.0 29.4 29.6 29.5 29.2 28.7 28.0 27.1 26.1 25.0 23.8 22.5 21.2 19.9 18.6 17.4 16.2 15.2 14.2 13.3 12.6 11.9 11.4 10.9 10.6 10.4 10.2 10.1 10.0 10.0 10.0 10.0
10.0 10.0 10.2 10.5 11.0 11.8 12.8 13.9 15.1 16.4 17.8 19.3 20.7 22.2 23.6 24.9 26.1 27.1 28.0 28.6 29.1 29.3 29.3 29.1 28.7 28.1 27.3 26.4 25.3 24.1 22.9 21.7 20.4 19.1 17.9 16.8 15.7 14.7 13.8 13.0 12.3 11.7 11.2 10.8 10.5 10.3 10.2 10.1 10.0 10.0 10.0
10.0 10.0 10.2 10.5 11.0 11.7 12.6 13.6 14.8 16.1 17.4 18.8 20.2 21.6 23.0 24.3 25.5 26.6 27.5 28.2 28.7 29.0 29.1 29.0 28.7 28.1 27.5 26.6 25.6 24.5 23.3 22.1 20.9 19.6 18.4 17.3 16.2 15.1 14.2 13.4 12.7 12.0 11.5 11.0 10.7 10.4 10.3 10.1 10.1 10.0 10.0
10.0 10.0 10.2 10.4 10.9 11.5 12.4 13.4 14.5 15.7 17.0 18.3 19.7 21.1 22.5 23.8 25.0 26.0 27.0 27.8 28.3 28.7 28.9 28.9 28.6 28.2 27.6 26.8 25.9 24.8 23.7 22.6 21.3 20.1 18.9 17.8 16.7 15.6 14.7 13.8 13.0 12.4 11.8 11.3 10.9 10.6 10.4 10.2 10.1 10.0 10.0
10.0 10.0 10.2 10.4 10.8 11.4 12.2 13.1 14.2 15.3 16.6 17.9 19.2 20.6 21.9 23.2 24.4 25.5 26.5 27.3 28.0 28.4 28.6 28.7 28.5 28.2 27.6 26.9 26.1 25.1 24.1 23.0 21.8 20.6 19.4 18.3 17.1 16.1 15.1 14.2 13.4 12.7 12.1 11.6 11.1 10.8 10.5 10.3 10.2 10.1 10.0
10.0 10.0 10.1 10.4 10.8 11.3 12.0 12.9 13.9 15.0 16.2 17.5 18.8 20.1 21.4 22.7 23.9 25.0 26.0 26.9 27.6 28.1 28.4 28.5 28.4 28.1 27.7 27.1 26.3 25.4 24.4 23.3 22.2 21.0 19.9 18.7 17.6 16.6 15.6 14.6 13.8 13.1 12.4 11.9 11.4 11.0 10.7 10.4 10.3 10.1 10.1
10.0 10.0 10.1 10.3 10.7 11.2 11.9 12.7 13.7 14.7 15.8 17.0 18.3 19.6 20.9 22.2 23.4 24.5 25.5 26.4 27.2 27.7 28.1 28.3 28.3 28.1 27.7 27.2 26.5 25.7 24.7 23.7 22.6 21.5 20.3 19.2 18.1 17.0 16.0 15.1 14.2 13.4 12.7 12.2 11.6 11.2 10.9 10.6 10.4 10.2 10.1

Figure 2.1 Simulated floods with water exchanges aquifer-river

The water volume of the flood in of course conserved as long as the flood wave
has not hit the downstream boundary of the model.
To evaluate the behavior of the system, three storms have been postulated, with
identical precipitation volume but different precipitation patterns. They extend on
respectively seven, five and three stretches of river and the precipitations occur
at the same period. Their centre of gravity along the river is similar:

- Storm 1 120 120 120 120 120 120 120

- Storm 2 168 168 168 168 168

- Storm 3 280 280 280

For this case, Figure 2.2 represents the three hydrographs of the resulting floods
in the river at a section close to the downstream extremity of the first storm.
Despite the qualitative nature of the model, it is possible to see that the peak
timing and peak magnitude are fairly similar, despite the important dispersion of
the precipitation in space (ratio of extension 1 to 2.3) and in magnitude (2.3 to 1).
The magnitude of the peak remains in a range narrower than +/– 5%. The
volumes have been conserved.
60
55
50 Flood 1
45 Flood 2
40 Flood 3
Discharge

35
30
25
20
15
10
0 4 8 12 16 20 24 28 32 36 40 44 48
Time

Figure 2.2 Flood hydrographs from rainfalls of various extents, same volume

Another scenario can be considered, postulating the same spatial extent for the
storm, but a different pattern of precipitation. For this scenario, three storms of
equal volume are also considered, but with various distributions of their intensity
along the river.

- Storm 1 100 200 300 400

- Storm 2 250 250 250 250
- Storm 3 400 300 200 100

Figure 2.3 illustrates that the influence of the precipitation pattern is not so
determinant for the timing or the magnitude of the peak flow. Despite ratios of 1
to 4 (resp. 4 to 1) for the extremity’s precipitations, the three maximum flows lay
in a range narrower than +/– 5%.
70

60
Flood 1
50 Flood 2
Flood 3
Discharge

40

30

20

10
0 4 8 12 16 20 24 28 32 36 40 44 48
Time
Figure 2.3 Flood hydrographs from rainfalls of various patterns, same volume

A last scenario evaluates the influence of the storm volume on the river response.
The spatial extent of the storm is again identical in all three cases, and the
precipitation rate is uniform. The volume varies of +/– 20% around a reference
value.

- Storm 1 120 120 120 120

- Storm 2 100 100 100 100
- Storm 3 80 80 80 80

Figure 2.4 clearly indicates a different type of river response, at short distance
downstream of the storm area. The variation range of the precipitation volume is
mirrored in the various river peak discharge: the range of the peak flow reaches
abourt +/– 12% of the central value.
35

30 Flood 1
Flood 2
25 Flood 3
Discharge

20

15

10
0 4 8 12 16 20 24 28 32 36 40 44 48
Time
Figure 2.4 Flood hydrographs from rainfalls of same pattern, various volumes

The main learning from these simulations is that the detailed precipitation pattern
is not overwhelmingly significant for the river response. The pattern may show a
fairly broad variation range without fundamentally impacting on the timing and
magnitude of the resulting flood, as long as the total volume of rainfall is identical.
More sensible however is this river response on the precipitation volume.

2.5. FLOOD RECESSION ISSUE

The knowledge of the subterranean world is so fragmented and limited that an

exact (or at least a reasonably representative) modelling of all sub-catchments is
virtually impossible. Similarly, an insight in the detailed precipitations pattern is
not everywhere possible. A very valuable indication however is given by the
results of the combined effect of all sub-catchments and their tributaries on the
river flow at the outlet of the watershed. One observes that this overall reaction,
all things considered and all meteorological situations experienced, is fairly
regular and repetitive.

For a given river flow, the recession pattern of the flood is nearly always the same,
and this independently from the magnitude of the preceding peak flow. This can
be interpreted by the fact that, the river flow being essentially driven by the
underground water (except perhaps during the peak flow period, when surface
flow prevails), a specific river discharge corresponds to a given (yet unknown)
combination of aquifers water storage.
2.5.1. Chiumbe River - Angola

Figure 2.5 pictures the behaviour of the Kwanza river in Chiumbe (Angola); its
watershed area is about 115 000 km2. Clearly visible are the gradual – yet
irregular – increase of the river discharge during the rainy season, the yearly peak
reached during a storm occurring after the river has significantly grown up, and
the recession patterns following the various peaks. The zoomed area on the right
of the figure shows the great similarity of all recessions. Several are perturbed by
late precipitations of the rainy season. Without however these rainy events, their
recession behaviour is quite regular, as for the curve highlighted by green circles,
from the peak time until the occurrence of a last strong rainfall event, and even
beyond this moment.

Figure 2.5 Chiumbe River – Flood Hydrographs and Recession Patterns

Actually, the recession behaviour of the entire watershed is only minimally

influenced by the magnitude of the peak flow during the rainfall season. As long
as it is not interrupted or perturbed by a late precipitation, the recession is almost
identically repeatable. The highlighted line shows that, for the discharge value
1500 m3/s, the slope of the recession curve just before the last precipitation event
and the resumed recession just at the end of the storm are virtually identical.

In a sense, this line (offsetting the intermediate precipitation perturbation) can be

seen as a signature of the watershed. This signature materializes the overall
behaviour and (unknown) inter-relationships of all aquifers and tributaries of a
river when the aquifers can aliment the river with their water into the river without
being perturbed by new rainfalls. It integrates all the internal (and unknown)
processes of the watershed and appears to be largely independent from the
history of the flood. Due to the location of this river, snow did not play any role in
its hydrograph.
The knowledge of this typical recession pattern provides a fundamental hint to
the underground behaviour of a watershed. Of course, the exact description of all
spatio-temporal characteristics and behaviours of this inner world is probably
bound to remain inaccessible. But its signature is easy to determine; it indicates
in an all-integrated way how the entire water catchment would typically behave
as soon as its aquifers would be filled with water. This reaction would directly lead
to know the evolution of the discharge that can be expected at a point of interest
along the river.

2.5.2. Mistassibi River - Canada

In Northern areas, the spring flood (mainly caused by snowmelt) is very often the
largest annual flood, at least for its volume and very often also for its peak
discharge. This type of flood can extend over several weeks, depending of the
depth of the snowpack, the air temperature, the size of the drainage area and the
rainfall events during this period.

Figure 2.6 illustrates typical large flood hydrographs on the Mistassibi River
(Canada), with a drainage area of about 9 300 km². On the Mistassibi River, the
spring flood occurs during a period of six to eight weeks; it can even reach twelve
weeks from the beginning of the snow melt to the end of the recession. As shown
on Figure 2.6, the main flood was observed on 1976. Depending of the sequence
of temperature observed, the timing of the peak discharge can vary by six weeks
from one year to the other.

Even if the local conditions are quite different from the conditions observed on
the Chiumbe River, the flood hydrographs present similar patterns, i.e. the
increase of the flow until the peak discharge is reached varies significantly
between the floods; however, the slope of the recession pattern is quite similar
(except for periods when rainfall is observed).
Figure 2.6: Mistassini River – Typical Spring Flood Hydrographs

An attempt of catching the behaviour of the 1976 flood with a simple numerical
model based on the recession curve is presented in section 2.6.

2.6. FLOOD VOLUME – STATISTICAL ANALYSIS

Peak discharge and flood volume for specified durations can be estimated
through statistical analysis. This is particularly true when most of the flood volume
is generated by snowmelt (so-called “spring flood”). As mentioned previously,
statistical flood analyses can be performed for different durations (e.g. 5, 10, 30,
45, 60, 90 days) to allow for reconstituting one or several hydrographs for a
specific frequency of the spring flood.

Statistical analyses can also be performed on the total duration of the spring
flood; however, the criteria to determine the beginning and the end of the flood
cannot be determined with certainty if the flood depends on more than one single
event. If the beginning of the spring flood corresponds to an increase of the
temperature to start the snowmelt process, the end of the spring flood is often
difficult to establish, since rainfall events and the characteristics of the drainage
area can have an impact on the overall process. Another alternative consists in
evaluating the flood volume observed between two specific dates (e.g. between
March 15th and June 15th) and perform the statistical analysis based on this
assumption.
A statistical analysis of the flood volume was performed on the discharge
observed on the Mistassibi River (Quebec, Canada). This river is unregulated,
which gives the possibility to observe the characteristics of the flood.
Figure 2.7 presents an example of a statistical analysis of the spring flood
maximum daily volume for the Mistassibi River. This figure presents the
observations available on this river (51 spring floods) and the lognormal relation
proposed to estimate the maximum daily volume.

Figure 2.7: Mistassibi River (1963-2013) –

Spring Flood Maximum daily volume – Statistical analysis

The same exercise can be repeated for different durations; Figure 2.8 illustrates
the estimated volume of the spring flood for different periods of recurrence and
durations 1. For this case, the durations vary from one day (daily maximum
volume) to 90 days.

1 For this exercise, a lognormal distribution was considered, but any other
statistical distribution could have been used; the conclusions should remain
unchanged.
Figure 2.8: Mistassibi River – Spring Flood Volume – Statistical analysis

Figure 2.9 presents the same information, but considering the duration in the X
axis. This figure illustrates the variation of the flood volume over time.
Based on this information, it should be possible to prepare a hydrograph
respecting the characteristics of the flood for a specific period of recurrence.
Figure 2.9 Mistassibi River – Spring Flood –Volume vs Number of days

2.7. HYDROGRAPH RECONSTITUTION – RAINFALL EVENT(S)

A method is presented here that considers the flood volume as a central factor of
the flood estimation. The method seeks the river response that fits both the net
precipitation volume and the recession pattern. The initial base flow plays a non-
negligible role in the disposition of the flood.

The method reveals to be robust; examples below illustrate the respective

influence of the key factors (precipitation volume, recession pattern, base flow)
on the shape of the hydrograph. Two of the most conspicuous advantages of the
method are its simplicity of use and the high consistency of the estimated river
flow, in particular with the magnitude of the storm at the origin of the flood and
the observed recurrent river behaviour.

The flood pattern is usually considered as an added discharge over the river base
flow. Various methods help estimate the shape and magnitude of the peak.
Conceptually, floods are thus commonly seen as superimposed on top of the
naturally receding base flow, as illustrated on Figure 2.10 (left). Very often, the
base flow is even considered to remain constant over the duration of the flood
and beyond (or even to grow during the flood). The total duration of the added
flood and its volume are not always clearly determined. In addition, the junction
between the end of the flood and the base flow is somewhat unclear.
Another approach considers that the amount of water brought by the storm
perturbs the natural recession of the river flow by increasing its discharge, as
seen on Figure 2.10. After peaking, the river flow recedes and reaches again the
level it had at the beginning of the flood. It is legitimate to consider that, as long
as no new precipitation occurs, the recession occurring beyond this point
corresponds to the usual natural base flow recession, as if no flood had occurred
(Figure 2.10 right). In this sense, the flood can be seen as a temporary
intercalation of additional water during the general recession process of the river
flow. Flood volume and duration can be exactly calculated.

"Superimposed" flood "Intercalary" flood

Crue "superposée" Crue "intercalaire"

Figure 2.10 Classical (left) and proposed (right) approach to

considering a flood

The demonstration of the quantitative equivalence of both models focuses on the

two shaded areas of the figure above (and beyond for the superimposed flood).
It is assumed that the recession pattern is an invariable response of the
watershed and is independent from the history of the flood. On Figure 2.11, the
superimposed flood is composed of the partial areas A, B and C; the intercalary
flood of the areas A, B and D. The areas A and B being part of both floods, the
equivalence of C and D must thus be proven.

The volume of the undisturbed base flow (as if there had not been any flood)
consists of areas D and E, from the beginning of the flood until the base flow
recedes to zero. The base flow "shifted" by the flood duration is made of the areas
C and E. As the recession pattern is invariable and assumed independent from
time, as both recessions start from the same discharge (by definition, see
horizontal red line), the two receding base flows are strictly equal in shape and
define the same water volume.
On the figure, C + E is therefore equal to D + E; the volume materialised by C is
hence equal to the volume attributed to D. Finally, A + B + C = A + B + D; both
hydrographs correspond to the same volume. This volume represents the net
runoff of the flood period, i.e. the total precipitation minus all the losses.
C + E = D + E
A

C = D
B

C
D
E

Figure 2.11 Equivalence of the flood volumes

An example of the application of this approach is visible on Figure 2.12. The 1976
flood of the Mistassini River has been simulated by four successive floods, all
with an identical recession pattern. For each flood, the river flow and the flood
volume have been adjusted in order to fit the observation. The blue area
represents the simulated volume of the flood series. The various peaks can be
exactly reproduced, in magnitude and timing; the close fit of the recession curves
confirms that they belong to the same family. The periods of growing river flow
show some small discrepancies due to not considered short rainfall events; in
addition, a very last small rainfall early July has not been considered.

To within half a percent, the integration of the blue area indicates a volume of
4’000 x 106 m3. This corresponds to the official estimates.
Figure 2.12 Observed and FIM-simulated Mistassini 1976 flood

2.8. HYDROGRAPH RECONSTITUTION – COMPLEX EVENTS

The reconstitution of a flood hydrograph for a specific frequency taking into

account the results of the statistical analysis of the peak discharge and the flood
volume (for specific durations) can lead to some discrepancies.

An intuitive approach consists in modifying an observed hydrograph considering

the results of the statistical analysis to reproduce the peak discharge and the
volume of the flood for different duration. Figure 2.13 illustrates this approach
based on the largest spring flood observed on the Mistassibi River (1976). These
hydrographs respect the peak discharge and the volume estimated by statistical
analysis for different durations, however the recession pattern does not respect
the pattern observed on historical floods, which is a characteristic of the drainage
area.
Figure 2.13 Mistassibi River – Spring Flood Hydrograph –
Without consideration of recession characteristics

As shown previously in this chapter, the flood recession follows a specific pattern
which is independent of the flood duration. Under this assumption, a large flood
should take more time to return to the normal conditions than a smaller flood
event; this will have an impact on the reconstitution of the flood hydrograph.
Figure 2.14 illustrates the reconstitution of the same flood hydrographs, but this
time considering the recession pattern of the drainage area. The flood peak and
volume are respected and the results appear more realistic.

It should be noted that such hydrograph(s) is only representative of one possible

pattern for specific period of recurrence. Such analyses should be performed
with different patterns (for the same period of recurrence) to evaluate the
consequences of floods. Intuitively, for systems with relatively large reservoirs,
the later the flood peak discharge occurs, the more critical will be the
consequences, since the reservoir will be at higher level at the occurrence of the
peak discharge.
Figure 2.14 Mistassibi River – Spring Flood Hydrograph –
With consideration of recession characteristics

2.9. RELATION BETWEEN FLOOD PEAK AND VOLUME

For drainage areas where major floods are caused by a single event, the flood
peak discharge and the flood volume could have similar periods of recurrence.
However, for larger drainage areas with more complex flood conditions, the
situation is different. The longer the flood duration will be, lower should be the
correlation between the peak discharge and the volume.

As example, an analysis was performed on the Mistassibi River data. As shown

on Figure 2.15, the coefficient of determination (R2) between the maximum daily
volume of the flood (daily peak discharge) and the five-day maximum volume is
about 98.2%. It decreases significantly for longer comparison periods.
Figure 2.15 - Mistassibi River –
Comparison of the spring flood volume for different durations

It is not possible to generalize conclusions based only on this specific case;

however, this shows that precautions must be taken in the reconstitution of the
flood hydrograph, since the relation between the peak daily volume and the
overall flood volume is not straight forward. It also illustrates the risk of performing
an analysis with only one hydrograph, since numerous combinations are
possible.

2.10. DETERMINISTIC APPROACH

The flood evaluation for different return periods can be performed using a
deterministic approach, particularly for locations where observed discharges data
are not available or only for a short duration. Results of the statistical and
deterministic analyses of the flood volume for rainfall event(s) should lead to
results in a similar range; this is particularly the case for small drainage areas, for
which a single rainfall event is usually considered.
A deterministic approach is also mainly used to evaluate the Probable Maximum
Flood, considering various possible scenarios maximizing the consequences on
the system 2.
The following elements must be considered for applying similar approaches:
a. Main rainfall event
The main rainfall event causing the flood normally corresponds to the expected
flood return period; higher is the probability to see such event, higher the number
of combinations that can generate a similar flood discharge or volume.
b. Antecedents events
Antecedent events are particularly important to establish the conditions prevailing
before the occurrence of the main rainfall event. This will have an impact on the
river discharge before the event and, even more important, on the soil moisture.
The more saturated the soil will be, the faster the response time of the system
will be (increasing the peak but also the volume for a specific duration). A similar
situation can be observed if a major rainfall event occurs when the soil is frozen.

Similarly, the base flow does not simply represent a reference river discharge on
which the peak flow is added but is an integral in the building up of the flood
structure. The net peak (additional discharge over the peak flow) generated by
an incoming flood volume is not a constant independent from the inflow pattern.

To evaluate the PMF, a large rainfall should be considered shortly before the
PMP to saturate the soil and ensure a maximum runoff.
c. Snow
The snow cover and the snowmelt period will have a direct impact on the flood
volume and the peak discharge. Both factors are important, since a rapid
snowmelt of a large snow cover will most likely generate large floods. When the
snow cover is an important part of the flood volume, the spring flood is very often
the largest one of the year, triggered by the combination of snowmelt and rainfall.
Deep snow covers increase the likelihood of large floods.

Before melting, the snow cover must be primed by warm temperatures bringing
it close to the melting point. A realistic temperature sequence must be developed
based on the observed conditions in the drainage area.
d. Subsequent events
Events following the main event can have a significant impact on the flood volume
and its duration. The impact of the subsequent events is particularly significant
until the reservoir returns to its maximum operation level (MOL). The longer the

2 The PMF scenario with the maximum peak discharge is not automatically the
worst scenario (maximum water level) for a dam with regulation capability. It
happens that a spring PMF (with lower peak discharge, but larger volume)
reaches a higher water level in a reservoir than a summer PMF (rainfall event).
duration to return to the MOL, the more vulnerable is the system in case of a new
large flood.
The subsequent events often occur in period(s) when large discharges have been
observed. This is particularly true for the evaluation of the PMF. For floods with
lower return periods however, it is not obvious to determine a subsequent
sequence of events, since there is in general no direct relation between the main
rainfall event and the subsequent events.
e. Initial conditions
Since oftentimes the objectives of such studies consist in determining the
maximum water level in a reservoir corresponding to a specific period of
recurrence, the initial conditions of the system are an important factor. The
expected volume of storage available before the flood will depend of the period
of the year. Normally the MOL is considered for a rainfall flood event.

However, for a spring flood (with a large percentage of the flood volume
generated by the snow cover), the reservoir level and the mode of operation
during the first part of the flood (until it reaches the MOL) will depend on the
expected conditions at this time of the year. The volume available for flood routing
will be larger; it is therefore unlikely that the spillway will be operated at full
capacity at the beginning of the flood. It may even not reach this discharge at all,
because of the uncertainty related to the final flood volume.
f. Comments
The evaluation of the flood volume for rainfall event(s) on large drainage areas
or for spring flood is complex, since it involves different events or conditions as
discussed above. If it is possible to identify the most likely scenario(s) to generate
a PMF, the number of scenarios to determine the 1:100, 1:1000 or 1:10,000 year
flood is almost infinite, since the combination of events leading to lower return
periods depends on too many combinations of parameters.

Comparison of the results from flood statistical analyses of spring flood volume
and deterministic analyses of the PMF volume can lead to inconsistencies. For
example, the extrapolation of the flood volume for a 1:10,000 year return period
can be higher than the volume of the PMF. Some explanation can be proposed:
- The statistical analyses overestimate the flood volume. The number of
recorded floods (usually a few tens) considered in a flood extrapolation
to the range of 1:1,000 years or more does not always guarantee a high
quality estimation, since the trends are not always well defined;
- Some of the parameters used in the deterministic analyses were
underestimated (for instance subsequent events). Since a period of
analysis of several weeks can follow the PMP, it is difficult to consider
realistic rainfall events during this period;
- Most probably a combination of both factors.
2.11. STOCHASTIC MODELING

Stochastic modelling can be seen as an answer to the evaluation of the flood

volume, whose results can be compared with those of a stochastic analysis of the
flood peak and flood volume or with the deterministic evaluation of the PMF.
Stochastic modelling appears to be particularly interesting for floods caused by a
set of events (such as the combination of snowmelt and rainfall events) and for
systems of reservoirs in cascade. Such modelling will have to reproduce the
succession of precipitation and air temperature along the year (to determine if the
precipitation will be rain or snow), the flood routing through the drainage area
including soil moisture variations, infiltration, losses, snowmelt and the reservoir
management (i.e. the operation of the control structures), to determine the
maximum water level to be observed at the site(s). Thousands of years will have
to be stochastically generated and simulated to estimate the probability related
to very large floods.
One of the main challenges in this process consists in representing adequately
the statistical distribution of each parameter and the temporal correlations
between them, such as:
- The relation and duration between precipitation events;
- The distribution of the precipitation on large drainage areas;
- The relation between air temperature and precipitation (rainfall or
snowfall);
- The variation of the air temperature and the snowmelt process (if
applicable).

The situation becomes even more complex for large drainage areas, since spatial
correlations and sometimes orographic effects at different locations must also be
considered. At the same time, the probability of deficiencies in the system and
“human” actions can play a significant role in the spatial and temporal evolution
of the flood.

The accuracy of the physical model(s) to represent large flood events must also
be considered, because of the limitations on the information available to calibrate
the model for such large floods. Usually a large number of assumptions (explicit
or implicit) are made at the basis of such a model; this of course generates larger
uncertainties.

If it is well known that the extrapolation of large floods may be subject to a

significant degree of uncertainty 3, because of the limitations of the sample
available to fit the statistical distribution. Similar concerns should also be
considered about the parameters considered in stochastic modelling and the final
results obtained.

3 CDA Guidelines “Flood statistics are subject to a wide margin of uncertainty,

which should be taken into account in decision-making.”
The use of stochastic modelling to evaluate flood characteristics and the
corresponding flood level is discussed in more details in bulletin 170 and in the
present bulletin (chapter 3).

2.12. FLOOD VOLUME – EXTREME VALUES

Data of maximum floods observed in several countries around the world date
back to 1984, when the International Association of Hydrological Sciences (IAHS)
published the “World Catalogue of Maximum Observed Floods”. Also, ICOLD
Committee on “Dams and floods “published, in 2003, the Bulletin 125 on Dams
and Floods, which contributed with more significant data related to maximum
floods, mainly for dams and reservoirs. Recently, in 2014, a new and more
extensive review of maximum floods has been carried out on the data of flows
and volumes of maximum floods.
For the analysis of the peak discharge, the envelope curves method with the
Francou-Rodier (F-R) equation can be used. The F-R equation is the relationship
between the peak flow and the catchment area:
𝑲𝑲
𝑸𝑸 𝑨𝑨 𝟏𝟏−𝟏𝟏𝟏𝟏
=� �
𝑸𝑸𝟎𝟎 𝑨𝑨𝟎𝟎
where:
- Q= Peak flow (m3/s)
- A= Catchment area (km2)
- Q0 = 106 m3/s
- A0 = 108 km2
- K= Francou-Rodier coefficient

For each peak discharge the coefficient K is calculated by:

𝒍𝒍𝒍𝒍𝒍𝒍𝒍𝒍 − 𝒍𝒍𝒍𝒍𝒍𝒍𝑸𝑸𝟎𝟎
𝑲𝑲 = 𝟏𝟏𝟏𝟏 · �𝟏𝟏 − �
𝒍𝒍𝒍𝒍𝒍𝒍𝒍𝒍 − 𝒍𝒍𝒍𝒍𝒍𝒍𝑨𝑨𝟎𝟎

The database on flood volumes comes from the surveys carried out by ICOLD; it
consists of 187 records on volume of maximum floods in dams and reservoirs
from 15 most significant countries in this field.
The methodology used is similar to that used for the analysis of the peak flows,
assessing the relationship between the flood volumes and the catchment area,
through the equation:
𝑲𝑲𝑽𝑽
𝑽𝑽 𝑨𝑨 𝟐𝟐− 𝟏𝟏𝟏𝟏
=� �
𝑽𝑽𝟎𝟎 𝑨𝑨𝟎𝟎
where:
- V= Flood volume (hm3)
- V0 = 50 × 106 hm3
- A0 = 108 km2
- Kv = Coefficient of flood volume

Therefore, for each flood the coefficient Kv is calculated by:

𝐥𝐥𝐥𝐥𝐥𝐥 𝑽𝑽 − 𝐥𝐥𝐥𝐥𝐥𝐥 𝑽𝑽𝟎𝟎
𝑲𝑲𝑽𝑽 = 𝟏𝟏𝟏𝟏 · �𝟐𝟐 − �
𝐥𝐥𝐥𝐥𝐥𝐥 𝑨𝑨 − 𝐥𝐥𝐥𝐥𝐥𝐥 𝑨𝑨𝟎𝟎

Figure 2.16 shows the relationship between the flood volume and the catchment
area for the floods analysed with the available data. It defines an envelope curve
of the extreme flood volumes with a value of Kv = 10.5

Figure 2.16 – Flood Volumes -

Envelope curve of extreme flood volumes – Kv = 10.5

The highest value was in Brazil (Tocantis reservoir) with a Kv= 10.5, in a 1980
flood.

Figure 2.17 shows the relationship between specific volume and the catchment
area. The specific volume, a measure of the volume generated per unit of the
catchment area, is expressed by :
𝑉𝑉
𝑉𝑉𝑠𝑠 =
𝐴𝐴
where:
- Vs = Specific volume (mm)

Figure 2.17 – Relationship between specific volume (Vs)

and catchment area

It should be noted that there is no systematized database on flood volumes

around the world. The results presented are a preliminary analysis, which should
lead to prepare a more detailed database of flood volume. Such database can be
used to perform initial evaluation of flood volumes on drainage areas presenting
similar conditions, mainly for validation purposes. It should be further expanded
in the future.

2.13. IMPACT OF CLIMATE CHANGE ON FLOOD VOLUME

It is widely recognized that climate change will increase the variability of extreme
events. The increase in air temperature will have an impact on the maximum
rainfall that can be observed in several regions of the world; this will in turn have
a direct impact on the floods peak discharge and the flood volume.
In the northern areas and for spring floods, the impact on the volume of the major
floods will be generally less important on the spring flood than the impact on the
peak flows, since for a watershed, projected reductions in the snowpack volume
may partially offset the expected increases in rainfall (Ouranos 2015). In this
case, the volume of the flood could be similar but it may occur over a shorter
period, since the snowmelt season will possibly be shorter (which could lead to
higher peak discharge). However, this conclusion cannot be generalized because
the regional conditions can change significantly over the world. Some recent
meteorological events will probably have some impacts on our understanding of
their characteristics and of their consequences 4.

2.14. RECOMMENDATIONS

- Representing adequately the volume of the design flood and considering

this volume in the design of the dam and its hydraulic structures is
essential to adequately consider the storage effect of the reservoir and
to optimize the size of the structures. Whatever the approach selected to
evaluate the flood hydrograph, an estimate should in any case check the
volume of the resulting flood and validate it comparatively with the
precipitation/snowfall volume.
- The recession of the flood is not depending on the peak discharge, but
on the characteristics of the drainage area and the river; it is important to
respect the recession pattern in the hydrograph reconstitution.
- For a same peak discharge and a same flood volume, the shape of the
hydrograph can have an impact on the maximum level in a reservoir
(depending of the operation rule of the reservoir). It is important to
perform a sensitivity analysis on the shape of the hydrograph. Intuitively,
a hydrograph with a late peak discharge could have more impact than a
hydrograph with an early peak discharge, since the reservoir could then
be at a higher level.
- For precipitations concerning an entire water catchment, an estimate
based solely on the precipitation volume and the watershed signature
may give worthwhile indications to cross-check the results of traditional
estimation methods.
- There is no guarantee that a direct relation can be found between the
peak discharge and the flood volume for floods deriving from a
combination of events (such as a flood from snowmelt). However, without
further information, a conservative assumption will be to consider that the
1:N-year peak discharge is corresponding to the 1:N-year flood volume
for any duration.
- Reconstitution of a flood hydrograph must respect the physical
characteristics of the drainage area; the recession period of the
hydrograph follows a pattern independent from the flood magnitude.

4 For example, Hurricane Harvey released about 1300 mm of rain in the Houston area
(USA) in 2017. The hurricane remained stationary for a few days, moved away from
the area and came back a few days later.
2.15. REFERENCES

Alberta Transportation - Transportation and Civil Engineering Division - Civil

Projects Branch, “Guidelines on Extreme Flood Analysis”, November 2004.

Bacchi, B., Brath, A., Kottegoda, N.T., “Analysis of the Relationships Between
Flood Peaks and Flood Volumes Based on Crossing Properties of River Flow
Processes”, Water Resources Research, Vol. 28, No. 10, pp 2773-2782, October
1992

Carter, R.W., Godfrey. R.G., “Storage and Flood Routing”, Manual of Hydrology:
Part 3. Flood-Flow Techniques, GEOLOGICAL SURVEY WATER-SUPPLY
PAPER 1543-B, Methods and practices of the Geological Survey, 1960.

Gaál, L., Szolgay, J., Kohnová, S., Hlavčová, K., Parajka, J., Viglione, A., Merz
R., and Blöschl, G., “Dependence Between Flood Peaks and Volumes: A Case
Study on Climate and Hydrological Controls”, Hydrological Sciences Journal, 60
(6) 2015.

Guillaud, C. “A review of the reliability of extreme flood estimates”. In Proceedings

of Canadian Dam Safety Conference, Victoria B.C, 2002

Joos B., “Flood Integration Method (FIM)”. ICOLD proceedings, Stavanger, 2015

International Committee on Large Dams, “Dams and Floods – Guidelines and

Case Histories”. Bulletin 125, ICOLD, Paris, 2003
International Committee on Large Dams, “Flood Evaluation and Dam Safety”.
Bulletin 170, ICOLD, Paris, 2016

Louie, P.Y.T. and Hogg, W.D. “Extreme Value Estimates of Snowmelt”, Canadian
Hydrology Symposium, pp 64-76, 1980

Berga, L. Personal Communication, 2018.

Micovic, Z. “An Overview of Three Hydrologic Flood Hazard Estimation Methods
Used by BC Hydro”, ICOLD 2013 International Symposium, Seattle, USA.

Molini, A., Katul, G.G., and Porporato, A., “Maximum Discharge from Snowmelt
in a Changing Climate“, Geophysical Research Letters, VOL. 38, L05402, 2011

Newton, D.W. “Realistic assessment of maximum flood potentials” Journal of

Hydraulics Div., Amer. Soc. of Civ. Engrs., v.109 no.6 pp. 905-918, 1983.
Ouranos, “Probable Maximum Floods and Dam Safety in the 21st Century
Climate”. Report submitted to Climate Change Impacts and Adaptation Division,
Natural Resources Canada, 39 p, 2015.

Pramanik, N., Panda, R.K., Sen, D., “Development of Design Flood Hydrographs
Using Probability Density Functions”, Hydrological Processes, Hydrol. Process.
24. 415-428 (2010).
SNC-Lavalin Inc., “Gestion du réservoir Gouin – Étude complémentaire”, Juin
2001
Wang, Cheng, "A joint probability approach for the confluence flood frequency
analysis", Retrospective Theses and Dissertations. Iowa State University, Paper
14865 The Gumbel mixed model for flood frequency analysis, S Yuea, T.B.M.J
Ouardaa, B Bobéea, P Legendre1, b, 1, , P Bruneau1, b, 1, , Journal of Hydrology
Volume 226, Issues 1–2, 20 December 1999, Pages 88–100, 2007

Sheng Yue; Taha B. M. J. Ouarda; Bernard Bobée; Pierre Legendre; and Pierre
Bruneau , “Approach for Describing Statistical Properties of Flood Hydrograph”
http://ascelibrary.org/doi/abs/10.1061/(ASCE)1084-
0699(2002)7:2(147)#sthash.sIFP4f7S.dpuf
3. STOCHASTIC APPROACH TO FLOOD HAZARD
DETERMINATION

3.1. INTRODUCTION

The traditional concept of the Inflow Design Flood (IDF) has been and is still being
used to size the dam and its designated flood discharge facilities (i.e. spillway,
low level outlets) so that the dam could safely pass either a flood of pre‐
determined probability of exceedance or the Probable Maximum Flood (PMF).
The IDF standard is directly linked to the dam hazard classification so that low
hazard dams are designed using smaller IDF than high hazard dams. For high
(or extreme) hazard dams, two general world trends have developed (ICOLD,
2003):

1. USA, UK, Canada, Australia and countries under their economic and
technological influence use the PMF methodology. The PMF is defined as the
most severe “reasonably possible” combination of rainfall, snow accumulation,
air temperatures, and initial watershed conditions. The PMF is a deterministic
concept and its probability of occurrence cannot be determined. Theoretically, it
represents the upper physical flood limit for a given watershed at a given season.
In reality, PMF estimates are typically lower than the theoretical upper limit by
some variable amount that depends on the available data, the chosen
methodology and the analyst’s approach to deriving the estimate (Micovic et al.,
2015).
2. Most European countries use probabilistic methods to derive an inflow
flood characteristic (typically peak flow of certain duration) with return periods
ranging from 1,000 to 10,000 years.
For lower‐hazard dams, the IDF selections criteria vary but typically include either
a percentage of the PMF or return periods shorter than 1,000‐years (ICOLD,
2003).

In recent years, an increasing number of dam owners have started to apply

various forms of risk informed decision making process in their dam safety
assessments regarding flood hazard. For example, the IDF selection guidelines
published by US Federal Emergency Management Agency (FEMA, 2013)
suggests that besides the traditional prescriptive approach to IDF selection, a
risk‐ informed hydrologic hazard analysis should be carried out at the discretion
and judgment of dam safety regulators and owners ‘‘for dams for which there are
significant trade-offs between the potential consequences of failure and the cost
of designing to the recommended prescriptive standard’’. The guidelines suggest
that an integral part of risk‐informed hydrologic hazard analysis is the
development of hydrologic loads that can consist of peak flows, hydrographs, or
reservoir levels and their annual exceedance probabilities (AEP). Another
example is the latest guidelines by Australian National Committee on Large Dams
on selection of acceptable flood capacity for dams (ANCOLD, 2017). While the
guidelines retained the PMF concept as part of a simplified risk procedure for
extreme hazard dams, there is a clear emphasis on risk assessment even for the
cases when the PMF needs to be derived, it is recommended that its
“reasonableness” be considered and assessed using the procedure outlined in
Nathan et al. (2011) so that the degree of conservatism implicit in the PMF is
justified and properly aligned with dam safety decisions regarding potential dam
upgrade costs.
Note that risk informed approach to flood hazard for dam safety implies knowing
the probability of dam overtopping due to flood. This practically means that the
full probability distribution of reservoir levels needs to be derived so that the
exceedance probability of the reservoir level corresponding to the dam crest
could be quantified. The IDF standard is necessary for sizing the surcharge
storage, height of a dam and outlet works, and could be useful in assessing safety
of dams with fairly steady reservoir level (where a full pool assumption is not
unreasonable) and without active discharge control systems such as gated
spillways or low level outlets. However, in regard to individual dams or dam
systems with fluctuating reservoir elevation and active discharge control systems,
the IDF concept is inadequate for use in flood hazard risk assessment analyses.

For those types of dams and dam systems, the IDF, by characterizing inflow to
the reservoir, does not provide the necessary information (i.e. magnitude and
probability) on the flood hazard in terms of hydraulic forces acting on the dam
itself (peak reservoir level). The commonly used solution for this problem is to
route the IDF through the reservoir and determine the resulting peak reservoir
level, and thereby obtain at least some information (magnitude but not probability)
on the flood hazard acting on the dam. In addition, the IDF concept typically
assumes that, during an extreme flood, everything operates according to the plan,
i.e. accurate reservoir level measurements, spillway gates open as required,
necessary personnel available on site, communication lines fully functioning. In
other words, the IDF concept does not address possibility of “operational flood”
in which a dam could fail due to a combination of a flood that is much smaller
than the IDF and one or more operational faults.

The number of possible combinations of unfavourable events causing such a

failure is very large and increases with the complexity of the dam or system of
dams. Consequently, the probability of dam failure due to an unusual combination
of relatively usual unfavourable events, which individually are not safety critical,
is larger than the probability of dam failure solely due to an extremely rare flood.
Baecher et al. (2013) stated that, for a complex system such as flow control at a
dam, the number of possible combinations of unfavorable events is
correspondingly as large as the probabilities of any one combination occurring is
small. As a result, the chance of at least one pernicious combination occurring
can be large. There are many examples of “operational flood” failures; two North
American examples are illustrated below:
- Canyon Lake Dam on Rapid Creek in South Dakota failed on June 9th,
1972, resulting in 238 fatalities. The reason for the dam failure was not the lack
of flood passing capacity but the inability to use the spillway which was clogged
by debris.
- Taum Sauk Dam in Missouri overtopped and failed on December 14th,
2005. The reason for the overtopping was not high inflow but the error in reservoir
level measurement (the pressure transducers that monitored reservoir levels
became unattached from their supports causing erroneous water level readings,
i.e. reported reservoir levels that were lower than actual levels). In addition, the
emergency backup reservoir level sensors were installed too high, thereby
enabling overtopping to occur before the sensors could register high reservoir
level.

Clearly, risk informed decision making for dam safety requires more than the IDF
concept. In order to have any scientifically‐based idea of the probability of dam
overtopping due to floods, it is necessary to focus on estimating probabilities of
peak reservoir level. The process can be described as follows:

- The reservoir inflow of a certain probability of exceedance is only the

starting value that gets modified by a complex interplay of starting reservoir level,
reservoir operating rules and decisions, and reliability of discharge facilities,
personnel and measuring equipment on demand.

- At the end of the process, the reservoir outflow and associated peak
reservoir level have different exceedance probability than the reservoir inflow that
started the process.

- The exceedance probability of the peak reservoir level is what determines

the probability of dam failure due to flood hazard; the probability of the reservoir
inflow is no longer the major driving parameter, but only one of the inputs needed
to calculate the probability of the peak reservoir level.

Note that the peak reservoir level, unlike the reservoir inflow, is not a natural and
random phenomenon and its probability distribution cannot be computed
analytically (e.g. by using statistical frequency analysis methods). The probability
of the peak reservoir level is the combination of probabilities of all factors that
influence it, including reservoir inflows, initial reservoir level, reservoir operating
rules, system components failure, human error, measurement error, as well as
unforeseen circumstances. Thus, the approach to estimating the full probability
distribution of the peak reservoir level consists of some kind of stochastic
simulation that includes as many of these factors and scenarios as possible. It is
a complex multi‐disciplinary analysis which is currently beyond technical
capabilities of some dam owners. However, without it, the proper risk‐informed
dam safety management is not possible. The main goal of stochastic simulation
approach to flood hazard is to carry out probabilistic analysis of various flood
characteristics (inflow, outflow, peak reservoir level) resulting from floods on a
dam system and derive the continuous probability distributions which could then
be used to evaluate exceedance probabilities of various reservoir levels including
the level corresponding to the dam crest (dam overtopping level) as well as the
level resulting from the PMF. That way, different design criteria could be
considered and evaluated at various flood frequency levels, thereby departing
from widely used strict “pass/fail” deterministic design criteria.

3.2. BASIC PRINCIPLES OF STOCHASTIC APPROACH TO FLOOD

HAZARD

In deterministic approaches, a particular flood characteristic (e.g. inflow, outflow,

routed reservoir level) is the result of a fixed combination of meteorological,
hydrological and reservoir routing‐related inputs. For instance, the peak reservoir
level resulting from the PMF is derived using a fixed combination of the following
inputs:

- Rainfall magnitude and its spatial and temporal distributions over the
watershed (typically provided in form of Probable Maximum Precipitation)

- Initial snowpack accumulation within the watershed

- Air temperature sequence during the PMF event

- Initial soil moisture content of the watershed

- Initial reservoir level

- Availability and operating sequence of discharge

On the other hand, stochastic approaches treat those inputs as variables instead
of fixed values considering the fact that any flood characteristic could be caused
by an infinite number of different combinations of inputs. The variation of the
flood‐producing input parameters is achieved by stochastic sampling either from
empirical distributions or from theoretical probability distributions fitted to
observed data. Note that most stochastic flood analyses assume process
stationarity, i.e. probability distributions of input parameters are not changing over
time. Consequently, they do not capture potential changes in
hydrometeorological parameters and their inter‐relationships that could result
from long‐term climate change.

Another thing that all stochastic approaches to flood hazard for dam safety have
in common is the use of a deterministic watershed model (i.e. rainfall‐runoff
model) to convert rainfall and snow/glacier melt into watershed runoff, which
ultimately becomes the reservoir inflow. In terms of watershed model simulation,
the stochastic flood hazard methods could employ:

- Event‐based simulation where a watershed model is used to convert

rainfall storm event of certain probability into a flood hydrograph (typically 3‐7
days duration). Initial watershed conditions such as soil moisture content and
snowpack accumulation have to be assumed and described stochastically.
- Continuous simulation where watershed model is used to convert
historical or synthetic rainfall time series into a continuous reservoir inflow record
from which flood events of interest can be directly extracted. In this case, initial
watershed conditions are continuously accounted for by the watershed model,
which is an obvious advantage over event‐based simulation approaches.

Boughton and Droop (2003) presented a review of continuous simulation for

design flood estimation. Despite their theoretical advantages, it should be noted
that continuous simulation models face the challenge of model complexity
needed to accurately represent the full range of watershed runoff, from droughts
and low flows to very large floods typically needed for dam safety applications.
Nathan (2017) provided an excellent discussion on some particular issues that
should be carefully considered prior to using continuous simulation models to
derive flood frequency curves over a probability range of relevance to dam safety
(i.e. return periods of 1,000‐years and beyond). An example of continuous
simulation used for design flood estimation is the GRADE method (Hegnauer et
al., 2014) developed in the Netherlands and used to derive design discharges for
the rivers Rhine and Meuse with drainage areas of 165,000 and 21,000 km2,
respectively. Stochastic weather generator based on the nearest‐neighbour
resampling is used to produce rainfall and temperature series that preserve
statistical properties of the original, historically observed data. Synthetic rainfall
and temperature data are generated at multiple locations simultaneously in order
to preserve their spatial distribution over the watershed, without making
assumptions about the underlying joint distributions. The continuous record of
50,000 years of daily weather data is simulated using a simple nonparametric
resampling technique where daily rainfall amounts are resampled from the
historical record (56‐year for the Rhine basin; 73‐year for the Meuse basin) with
replacement.

Note that this approach does not generate daily rainfall amounts greater than
those observed in the historical record; however, the technique of resampling with
replacement creates different temporal patterns resulting in multi‐day rainfall
amounts higher than those observed in the historical record. A watershed model
is used to calculate runoff from this synthetic rainfall and temperature series, and
the runoff is then routed using a hydrodynamic model to account for complexities
associated with retention and flooding along particular river stretches. This
procedure yields the continuous record of 50,000 years of daily discharges at or
near the points where rivers Rhine and Meuse enter the Netherlands. Finally, the
flood values for the various return periods are obtained by ranking the annual
maximum discharges in the generated 50,000‐year sequence in the ascending
order, where the rank in this ordered set determines the return period. The main
source of uncertainty in the GRADE method is the relatively short length of the
historical precipitation and temperature series used in the stochastic weather
generator. Using less than 100 years of historical data to generate 50,000 years
of synthetic data affects the ability to accurately capture year‐to‐year variability
over the long periods of time. For instance, resampling from a relatively wet
baseline series will result in relatively wet long‐duration synthetic series, which in
turn will increase uncertainty associated with derived flood discharge values,
especially for higher return periods. This large uncertainty is reflected in the
GRADE flood frequency results for the Meuse River, where the best estimate of
10,000‐yr flood of 4,400 m3/s is given with the 95% uncertainty range of 3,250 to
5,550 m3/s.
Finally, there are stochastic approaches that fit somewhere in between event‐
based simulation and continuous simulation – they could be called semi‐
continuous or hybrid approaches. They utilize a watershed model that has been
calibrated to satisfactorily represent hydrological behaviour of the watershed over
a long continuous period for which historical record of climate input data is
available. This creates a continuous database of watershed initial conditions
which can be stochastically sampled at any time/season of the year and
combined with a rainfall event of a certain duration and probability, sampled from
rainfall magnitude‐frequency curve. The end result is thousands of flood
hydrographs ranging in magnitudes from common to extreme. The advantage
over event‐based simulation approach is that statistical distributions of initial (pre‐
storm) watershed conditions are likely more realistic since they do not have to be
arbitrarily assumed. The advantage over continuous simulation approach is that
there is no need to carry out the difficult task of generating thousands of years of
continuous synthetic rainfall and temperature sequences of questionable
accuracy. Examples of semi‐continuous or hybrid stochastic flood models are
SEFM (Schaefer and Barker, 2002) and SCHADEX (Paquet et al., 2013).

3.3. MAIN ASPECTS OF STOCHASTIC FLOOD HAZARD MODELLING

FOR DAM SAFETY

There are three distinct aspects of stochastic flood simulation for a hydroelectric
system consisting of a single or multiple dams and reservoirs.

1. Simulation of natural runoff from the local watershed and inflow into the
reservoir.

2. Simulation of reservoir operating rules (if any), i.e. flood routing for a
single reservoir or a system of multiple dams and reservoirs.
3. Simulation of on‐demand availability of various system components such
as failure of different discharge facilities, telemetry errors, human operator errors,
or some combination of those.

Ideally, all three aspects are combined within the stochastic simulation
framework, and multi‐ thousand years of extreme storm and flood annual maxima
are generated by computer simulation. The simulation for each year contains a
set of climatic and storm parameters that are sampled through Monte Carlo
procedures based on the historical record and collectively preserved
dependencies among different hydrometeorological inputs. Execution of a
rainfall‐ runoff model combined with reservoir routing of the inflow floods through
the system and stochastically modelled failure/availability of various system
components provides the computation of a corresponding multi‐thousand year
series of annual flood maxima. Simulated flood characteristics such as peak
inflow, maximum reservoir release, inflow volume, and maximum reservoir level
are the parameters of interest.
However, it is extremely difficult if not impossible to accurately cover all three
aspects of stochastic flood simulation due to the enormous complexity of a
dam/reservoir system and all possible interactions among its components. That
is why in practical applications not all flood producing factors are modelled to the
same extent ‐ some are treated as stochastic variables, and some are fixed or
not modelled at all. For instance, the third aspect of flood hazard simulation
mentioned above (stochastic simulation of on‐demand availability of various
system components) is rarely carried out due to complexities and difficulties in
describing probability distributions of variables such as spillway gate failure or
human error. A recent study by Micovic et al. (2016) attempted to cover all three
aspects of stochastic flood hazard simulation on a system of three dams and
reservoirs with seasonally fluctuating reservoir levels and active discharge control
systems. The aim of the study was to examine how the inclusion of spillway gate
failures likelihood functions in the stochastic flood modelling framework affects
the probability of dam overtopping. The results indicated that dams are much
more likely to be overtopped due to an unusual combination of relatively common
individual events than due to a single extreme flood event. The all three aspects
of stochastic simulation framework are discussed in the following sections.

3.3.1. Stochastic simulation of reservoir inflows

An example of hydrometeorological inputs to stochastic flood model and the

dependencies that exist in the stochastic simulation of a particular input are listed
in Table 3.1. Note that natural dependencies are prevalent throughout the
collection of hydrometeorological variables. The natural
dependencies/correlations are preserved in the sampling procedures with a
particular emphasis on seasonal dependencies. For instance, the sampling of
freezing levels is conditioned on both month of occurrence and 24‐hour
precipitation magnitude. The watershed conditions for soil moisture, snowpack,
and initial reservoir level are all inter‐related and inherently correlated with the
magnitude and sequencing of daily, weekly and monthly precipitation. These
inter‐relationships are established through calibration to historical streamflow
records in long‐term continuous watershed modelling and the state variables are
stored for each day of the calibration period (typically 25‐years or more). The
inter‐dependencies are preserved for each flood simulation through a resampling
procedure.
Table 3.1. Hydrometeorological Inputs to SEFM
for BC Hydro’s Campbell River System

Model input Dependencies Probability Comments

model
Storm Independent Normal End‐of‐month
seasonality distribution storm
occurrences
72‐hour Independent 4‐parameter Developed from
precipitation Kappa regional
magnitude distribution precipitation
analyses and
isopercental
spatial storm
analyses
Temporal/spatial Resampling 15 prototype
distribution of Independent from equally‐ storms, 72‐hour
storms likely historical to 144‐hour long
storms time‐series
Air temperature Temporal Resampling Pattern indexed
and freezing patterns are from historical to 1,000 mb
level temporal matched one‐to‐ storms temperature and
patterns one to prototype freezing level for
storms day of max. 24hr
precipitation
Air temperature Storm Physically‐ For day of
at 1,000 mb magnitude based stochastic maximum
model 24-hour
precipitation in
storm
Air temperature Independent Normal For day of
lapse‐rate distribution maximum
24-hour
precipitation in
storm
Freezing level 1,000‐mb Physically‐ For day of
temperature, based stochastic maximum
temperature model 24-hour
lapse‐rate and precipitation in
storm storm
magnitude
Watershed Seasonality of Resampling of Sampled from
model storm historical antecedent
antecedent conditions Oct condition files.
conditions 1983 – present Sampled year is
(snowpack, soil independent,
moisture) sampled month
corresponds to
month sampled
from seasonality
of storm
occurrence.
Initial reservoir Seasonality of Resampling of Sampled from
level storm and historical recorded
watershed conditions with reservoir level
model the same data. Sampled
antecedent reservoir year has similar
conditions operating rules antecedent
as the current precipitation as
ones (1998 – year sampled for
now) watershed
model
antecedent
conditions.

3.3.1.1. Storm seasonality

The seasonality of storm occurrence is defined by the monthly distribution of the

historical occurrences of storms with widespread areal coverage that have
occurred over studied area. This information is used to select the date of
occurrence of the storm for a given stochastic simulation. The basic concept is
that the seasonality characteristics of extraordinary storms used in stochastic
flood simulations should be the same as the seasonality of all significant storms
in the historical record. The term “significant” is somewhat subjective, but it
usually refers to storm events where precipitation maxima for a given storm
duration exceeds a 10‐ year return period at three or more precipitation gauges
within the studied area. This criterion assures that only storms with both unusual
precipitation amounts and broad areal coverage would be considered in the
analysis.

Figure 3.1 shows an example of this procedure applied to the 1,463 km2 Campbell
River watershed on Vancouver Island, BC, Canada and using the 72‐hr storm
duration. The procedure resulted in identification of 69 significant storms within
the 1896‐2009 period. A probability‐ plot was developed using numeric storm
dates (9.0 is September 1st, 9.5 is September 15th, 10.0 is October 1st, etc.) and
it was determined that the seasonality data could be well described by a Normal
distribution. A frequency histogram was then constructed based on the fitted
Normal distribution to depict the twice‐monthly distribution of the dates of
significant storms for input into a stochastic simulation framework.

Figure 3.1 Probability plot and frequency histogram of storm seasonality for the
Campbell R. watershed in Canada

Figure 3.1 shows that significant historical storms have occurred in the period
from early October through about mid‐March with a mean date of December 21st.
The probability of occurrence of a storm for any given mid‐month or end‐of‐month
can be determined from the incremental bi‐ monthly frequencies depicted in the
Figure 3.1 histogram (e.g. zero probability for September mid‐month, and
probability of 0.0228 for September end‐month).

3.3.1.2. Precipitation magnitude-frequency relationship

Generally speaking, floods could result from storms of various durations. This is
especially true for very large watersheds (e.g. > 10,000 km2), where significant
floods could originate from intense short‐duration storms covering only a part of
the watershed, or from a wide‐spread general synoptic storm of longer duration
that cover the entire watershed area.

The direction of the storm and its speed of movement over the watershed is also
an important factor. Consequently, proper stochastic flood modelling process
should be sampling rainfall storms of different duration from their respective
frequency distributions and weight modelled floods according to the observed
frequency of different duration rainfall storms used to produce said floods.

This approach implies the existence of separate rainfall‐frequency relationship

for all considered storm durations. However, the stochastic modelling is often
simplified by choosing so called “critical storm duration” for a particular watershed
and deriving a rainfall‐frequency relationship only for that duration. This
simplification is particularly effective in watersheds with drainage area sizes
under approximately 5,000 km2 where it was relatively easy to determine the
typical storm duration from precipitation gauges in the area, and where exclusion
of other storm durations in stochastic flood simulation would have relatively minor
effect on overall accuracy.
The vast majority of precipitation stations record on a daily basis which results in
logical choices of the 1‐day, 2‐day, 3‐day or 4‐day duration for the precipitation‐
frequency analysis. For instance, BC Hydro experience shows that for
watersheds in Pacific Coastal zone of British Columbia the 72‐hr duration (3‐day)
is the most representative of the typical storm duration, whereas for some
watersheds in Interior British Columbia it is the 48‐hr duration. Similarly,
Électricité de France generally uses the 3‐day duration in their SCHADEX
stochastic model for most French watersheds, with some small and flash‐flood
prone watersheds being treated with shorter storm durations.
The magnitude of precipitation relevant to dam safety analyses is typically several
orders of magnitude more extreme than what has been observed in the historic
record. As such, the estimation of this range of rainfall presents special difficulties
and requires the extrapolation of relatively short historical data records. This
extrapolation is challenging, especially considering that rainfall input is generally
the most significant contributor to resulting stochastically derived flood
hydrographs. There are different ways to do this extrapolation and obtain
precipitation magnitude‐frequency relationship over the entire probability domain
of interest, including return periods of 10,000 years and beyond. A couple of
examples are described here:

1. In the SEFM model (Schaefer and Barker, 2002), the precipitation

magnitude‐frequency analysis is done by utilizing the regional L‐moment analysis
(Hosking and Wallis, 1997). By employing the regional precipitation‐frequency
analysis we can compensate for the short length of available hydrometeorological
record by considering a larger study area. This approach takes advantage of the
situation that the size of the study area is much larger than the typical areal
coverage of storms for the duration of interest, and there will be many storms in
the regional dataset with return periods higher than indicated by the chronological
length of the historical record.

Applying this approach to the Campbell River watershed above Strathcona Dam
on Vancouver Island, BC, Canada included assembling storm data from all
locations that were climatologically similar to the Campbell River region.
Precipitation annual maxima series data were assembled for the critical duration
(72‐hour in this case) from all stations on Vancouver Island and stations between
latitude 47° and 52° N from the Pacific Coast eastward to the crest of the Coastal
Mountains (Canada) and Cascade Mountains (USA). This totaled 143 stations
and 6,609 station‐years of record for stations with 25‐years or more of record.
The precipitation‐frequency relationship (Figure 3.2) was developed through
regional L‐ moment analyses of point precipitation and spatial analyses of
historical storms to develop point‐to‐area relationships and determine basin‐
average precipitation for the watershed using the 4‐parameter Kappa distribution.
The uncertainty bounds were developed through Latin‐hypercube sampling
method (McKay et al., 1979; Wyss and Jorgenson, 1998) where regional L‐
moment ratios and Kappa distribution parameters were varied to assemble 150
parameter sets and perform Monte Carlo simulation using different probability
distributions for individual parameters.
Figure 3.2 Computed 72‐hour precipitation‐frequency curve and 90%
uncertainty bounds for the 1193 km2 Strathcona Dam watershed

2. In the SCHADEX model (Paquet et al., 2013), the precipitation

magnitude‐frequency analysis was catchment‐specific (a regional approach is
used if local data are lacking) utilizing weather patterns classification and Multi‐
exponential distribution linked to specific weather patterns (Garavaglia et al.,
2010). The underlying hypothesis is that a rainfall sampling based on days having
similar atmospheric circulation patterns will provide more homogeneous sub‐
samples which will in turn reduce uncertainty associated with extrapolation of
short sample size. The rainfall stochastic generator of SCHADEX is based on the
concept of a 3‐day event so‐called “centered rainfall event”. Therefore,
SCHADEX only develops precipitation magnitude‐frequency for the central daily
rainfall (dark blue in Figure 3.3 below), and precipitation quantiles of the day
before and the day after (adjacent rainfalls, light blue in Figure 3.3) are estimated
using the probabilities of the ratios “Pa‐/Pc” and “Pa+/Pc”, computed from rainfall
events identified in the historical record. Note that Pa‐, Pc and Pa+ represent
daily rainfall amounts for the day before the central day, the central day and the
day after the central day, respectively.
Figure 3.3 SCHADEX concept of centered rainfall event
(with central and adjacent rainfalls)

3.3.1.3. Temporal and spatial distribution of storms

The process of stochastic storm generation requires both spatial and temporal
storm templates that are scalable. The spatial and temporal storm templates are
linearly scaled by the ratio of the desired basin‐average precipitation of certain
duration to the basin‐average precipitation of the same duration observed in a
selected storm template (i.e. prototype storm). These storm templates should be
prepared from as many historically observed storms as possible in order to
capture diversity among storms in terms of spatial and temporal distribution of
precipitation. Typically, 10 to 20 storm templates should be enough to capture
storm diversity over a given watershed, for watersheds sizes up to about
5,000 km2. Larger watersheds should be divided into zones of a size suitable for
describing the spatial and temporal variability of storm types that may affect the
watershed on a given day, with separate storm analyses carried out for each zone
within a watershed.

These kinds of decisions are typically site‐specific and depend of detailed

meteorological analysis of the given area, including use of measurable
parameters such as geopotential height contour maps for different pressure
heights, precipitable water, convective available potential energy and the scale
of the precipitation footprint (synoptic, mesoscale, or local convective storms)
determined by the percentage of gauges in the area exceeding a specified daily
precipitation threshold. For instance, there could be large areas with fairly simple
climatology, whereas some relatively small watersheds could have complex
climatology with multiple storm types with different spatial, temporal and seasonal
characteristics, resulting in a mixed population of storms and floods.
In the presented example of the 1,463 km2 Campbell River watershed located in
Pacific Coastal zone of British Columbia, as mentioned earlier in Section 3.1.2, it
was fairly straightforward to determine the typical storm duration from
precipitation gauges in the area, meaning that exclusion of other storm
durations/types in stochastic flood simulation would have relatively minor effect
on overall accuracy.
Spatial storm templates for a given storm are developed analyzing rainfall data
and using GIS analyses to compute basin‐average precipitation of certain
duration (e.g. 24, 48, 72‐hr). The analyzed rainfall data could be hourly or daily
as well as from point measurements or from radar, depending on what type of
rainfall data is required as an input into a rainfall‐runoff model used in stochastic
flood simulation. An example of spatial storm template (a 72‐hour precipitation
for the October 1984 storm over the Strathcona Dam basin) is shown in
Figure 3.4.

Temporal storm templates could be developed as a 3‐day template with the

centered main rainfall and adjacent rainfalls as assumed in SCHADEX
methodology and described in the previous section (Figure 3.3). A more elaborate
way to develop temporal storm templates is utilized by the SEFM method where
hourly rainfall data from many point‐measurements station within a given
watershed are used to obtain the basin‐average rainfall of specified duration
(e.g. max.72‐hr). This is followed by the examination of the 10‐day period of
precipitation encompassing the max. 72‐hour precipitation using daily synoptic
weather maps, radiosonde data and air temperature temporal patterns.

This procedure leads to the identification of the time span during which there was
a continuous influx of atmospheric moisture from the same air mass where
precipitation was produced under similar synoptic conditions. The identified time
span provides the starting and ending times for the precipitation segment that is
independent of surrounding precipitation and scalable for stochastic storm
generation. An example of this type of temporal storm template is shown in
Figure 3.5 which depicts the observed 10‐day period of basin‐average
precipitation for the storm of October 14‐23, 2003 for the Strathcona Dam basin,
with the portion of the hyetograph (in blue) that was identified as the independent
scalable segment of the storm and therefore adopted for use as a prototype storm
for stochastic storm generation.
Figure 3.4 Spatial storm template (October 1984 storm)

Figure 3.5 Temporal storm template (October 2003 storm)

3.3.1.4. Air temperature and freezing level temporal

patterns

The usual approach is to first stochastically simulate 1,000 mb air temperature

(i.e. temperature at or near the sea level), followed by stochastic simulation of air
temperature lapse‐rates that are required for computing both freezing levels and
air temperature values within the full elevation range of a given watershed.

Within the stochastic flood simulation framework, 1,000 mb air temperatures

during extreme storms could be simulated using a variety of approaches. One
example is a physically‐based probability model for 1,000 mb dewpoint
temperatures derived from monthly maximum dewpoint data (Hansen et al.,
1994). This probability model utilizes monthly upper limit dewpoint data and the
magnitude of the maximum 24‐hour precipitation within the storm relative to 24‐
hour PMP. The 1,000 mb dewpoint temperatures are drawn from a symmetrical
Beta Distribution bounded by lower and upper bounds as shown in Figure 3.6 for
December in the Vancouver Island region, BC, Canada. A separate relationship,
similar to Figure 3.6, is used for each month because 1,000 mb dewpoint
climatology changes with season. Higher maximum 1,000 mb dewpoints are
possible in the fall months of October and November than in the colder winter
months of January and February, which implies that freezing levels tend to be
somewhat lower for storms in the colder winter months.

Figure 3.6 Range of 12‐hour persisting 1,000 mb dewpoint temperatures

utilized by dewpoint temperature probability model (December example)
The next step is to stochastically generate air temperature lapse‐rates. For
example, analyses of upper air sounding data from Northwestern Washington
and Central California stations reveal that air temperature lapse‐rates on the day
of maximum 24‐hour precipitation for noteworthy storms are well described by
the Normal Distribution (Figure 3.7). The mean value was found to be 5.1°C/1000
m, which is near the saturated pseudo‐adiabatic lapse‐rate. Similar results were
found if examining the data from Washington or California separately, so the data
from the two regions were combined to provide a larger sample for computing the
distribution parameters. Stochastically generated 1,000 mb air temperatures and
temperature lapse‐rate described above are used in computing resulting freezing
levels for each storm.

Figure 3.7 Air temperature lapse‐rates for day of maximum 24‐hour

precipitation for storms in Northwestern Washington and Central California

Once air temperatures and freezing levels have been stochastically generated,
we could start deriving their temporal patterns that are used in computing
snowmelt runoff. Temporal patterns for air temperature and freezing level are
matched one‐to‐one with each prototype storm. These temporal patterns are
developed so that they could be rescaled by stochastically drawn values of air
temperature and freezing level. Scalable 1,000 mb air temperature and freezing
level temporal patterns are constructed by subtracting index values (the highest
6‐hour averages observed during the day of maximum 24‐hour precipitation) from
observed data.

An example of indexed 1,000 mb temperature and freezing level temporal

patterns constructed for the Oct. 2003 storm over the Strathcona Dam basin is
shown in Figure 3.8. During stochastic flood simulation, index values for 1,000
mb temperature and freezing level are stochastically simulated for each prototype
storm and then used to rescale the indexed temporal patterns (e.g. Figure 3.8)
by adding the simulated index values. This procedure yields simulated 1,000 mb
temperature and freezing level hourly time‐series similar to those historically
observed and therefore physically plausible.

Figure 3.8 Indexed 1,000 mb temperature and freezing level temporal patterns
for the Oct. 2003 storm

3.3.1.5. Watershed model antecedent conditions sampling

As mentioned earlier, a continuous watershed model (rainfall‐runoff model) is

calibrated to realistically simulate hydrological behaviour of the watershed over a
continuous period of many years, typically dependent on the availability of climate
input data. The model state variables (snowpack accumulation, soil moisture
conditions, glacier mass balance, base flow from seasonally‐varying aquifer
supplies, etc.) are computed continuously by the model over the long simulation
period. This creates a continuous database of watershed initial conditions which
could be stochastically sampled at any time/season of the year and assumed to
occur at the onset of a stochastically generated storm event. Using this approach,
we ensure that the full range of synthetic storm events is simulated with the full
range of watershed’s historically observed internal states (e.g. extreme rainfall on
a dry watershed, rain‐on‐snow events, average rainfall on a saturated watershed,
etc.).

3.3.1.6. Initial reservoir level

A resampling approach is typically used to determine the reservoir elevation at

the beginning of the stochastic flood simulation. Reservoir operating rules,
especially in complex systems of two or more reservoirs usually change over time
due to various reasons. Reservoir operating rules represent a compromise
among the needs of various stakeholders. As a result, hydropower operation,
flood protection, environmental constraints and recreational demands limit the
reservoir operating range. It is therefore important to resample reservoir level
data only from the period in historic record that reflects current system/reservoir
operating rules. This may lead to a fairly short record length for reservoir level
data, especially if the changes in reservoir operating rules were implemented
relatively recently. In these cases, it may be prudent to use a reservoir simulation
model and combine historical inflow data with the new reservoir operating rules,
which would yield a longer record of synthetic reservoir levels reflecting the
current operation.

3.3.2. Simulation of reservoir operation – flood routing

After thousands (or millions) of inflow hydrographs have been stochastically

derived and coupled with initial reservoir levels, they have to be routed through a
reservoir to obtain peak reservoir level and outflow hydrographs that enter any
downstream reservoir(s). Flood routing simulations are carried out with the aim
to realistically capture the way a reservoir (or a system of reservoirs) may be
operated during flood events ranging in annual exceedance probabilities from 1/2
to beyond 1/10,000. In reality this is a rather complex decision making process
involving various factors such as the inflow forecast, flood magnitude,
downstream environmental conditions existing at the time of the inflow, corporate
pressures, and other site‐ specific and season‐specific conditions.

Generally speaking, for very large floods the routing procedure becomes fairly
simple, because the main goal is to save the dam from overtopping (or if
overtopping is tolerable, to prevent the reservoir reaching an elevation that
endangers the stability of the dam) and utilize as much discharge capacity as
possible (i.e. fully open spillway gates, low level outlet gates, etc.). In such cases
the safety of the dam takes precedence over every‐day operational constraints
such as, for example, downstream discharge limits pertaining to residential
flooding, environmental and recreational requirements. Things are more
complicated when routing smaller floods (e.g. return periods between 10 and 50
years or so) because there could be ways to safely pass the flood and still satisfy
other constraints, especially if the inflow forecast is reasonably reliable.

For example, a dam operator may decide to start spilling moderate amounts of
water (within downstream discharge limits) for several days before forecasted
inflow arrives so that there is enough storage in the reservoir to absorb the
incoming flood while keeping spill below downstream discharge limits. Note that
this strategy depends on the accuracy of the inflow forecast and flood magnitude
– when flood magnitude exceeds certain level, the flood routing options diminish
and the only strategy becomes fully utilizing available discharge facilities and
avoid dam overtopping.
The mentioned pre‐spilling strategy and a dam operator decision making process
could be modelled using the following two‐step approach:
1. The first step in flood routing simulations is to route the stochastically‐
generated inflow hydrographs through the reservoir (or system of reservoirs)
using the standard flood routing procedures
2. In the second step, the results from the standard flood routing performed
in the first step are treated as a "forecast", and the inflows are then routed using
modifications reflecting the pre‐spill or any other dam operator’s strategy that
deviates from the standard routing procedure. Although the second step is
performed with the benefit of complete foresight, the routing modification and the
dam operator’s decisions could be developed recognizing that the real inflow
forecast is uncertain.

The two‐step approach above, when incorporated in stochastic flood simulation,

provides an additional “operational” component to the whole process and results
in more realistic modelling of small to moderate floods, where there could be
some flexibility in the routing procedure.

3.3.3. Stochastic simulation of the availability of discharge

facilities

Generally, the reservoir routing of an extreme flood in deterministic dam safety

analyses (i.e. the PMF) is performed by assuming a conservatively high initial
reservoir level typically combined with the assumption that all spillway gates open
as required to pass flood discharge. Some analyses assume the “n‐i” rule where
“n” is the number of spillway gates and “i” is the number of spillway gates
assumed unavailable. However, depending on values for “n” and “i”, this
approach may be too conservative for some spillway configurations.

During the passage of a flood, one or more of spillway gates may be unavailable
for various reasons including debris jams, human error and mechanical or
electrical malfunctions. Due to numerous interconnected components of a
spillway gate system and consequently the infinite number of reasons for failure,
it is impossible to directly compute the probability of gate failure on demand. This
probability has to be estimated through some kind of analysis which should
include as many failure modes as possible, with consideration of site‐specific
knowledge of the state of various gate system components as well as frequency
and thoroughness of gate testing.

One example of such analysis was carried out during gate reliability assessments
of five US Army Corps of Engineers Huntington District dams (Lewin et al., 2003).
Gate and equipment testing for these five dams of 3‐4 times per year was
considered relatively infrequent. In addition, many components, such as relays
or limit switches were tested only when the gate test used those particular
components. During 18 gate tests, there were three instances when a gate failed
to operate correctly. A fault tree analysis of gate failure modes indicated that for
a gate opening in ideal conditions, similar to those under which gate tests were
carried out, a probability of failure on demand was assessed to be of the order of
1 in 10. The probability of multiple failures of gates during an extreme flood was
estimated to be at least 1 in 100 per demand due to a common cause failure
regardless of the number of gates in an installation.
The availability of powerhouse discharge during flood is also uncertain. Flood‐
inducing rainfall could be very severe and could conceivably cause erosion or
landslides that might result in transmission line failures, making generation
impossible. Severe rainfall could also cause other powerhouse‐disabling damage
such as powerhouse flooding or penstock damage. Perhaps the most realistic
way of simulation would be to somehow tie the availability of the powerhouse to
the storm magnitude (e.g. if the simulated 72‐hour basin‐average precipitation
has the annual exceedance probability of 1/1,000 and greater, the powerhouse
discharge is disabled).

The importance of spillway gate reliability could be very high and typically
increases as the size of reservoir surcharge storage decreases. For example,
Micovic et al. (2016) showed that for a dam with very small reservoir surcharge
storage, simulating possibility of random spillway gate failures during the flood
increased the resulting annual probability of dam overtopping by five orders of
magnitude over the case in which all spillway gates are assumed operable.

3.3.4. Simulation procedure

An example schematic of stochastic flood simulation involving the steps

described in Section 3 of this chapter is shown in Figure 3.9. Precipitation to
runoff conversion in this example was done by the UBC Watershed Model (Quick,
1995; Micovic and Quick, 2009).
Figure 3.9 Stochastic flood simulation flow chart for the 1193 km2 Strathcona
Dam watershed in British Columbia, Canada

3.4. ESTIMATING UNCERTAINTY IN STOCHASTIC FLOOD MODELLING

Stochastic flood modeling is technically sound in its principle since it attempts to

derive probabilities of extreme floods from the physically plausible modelling
framework (considering the available historical data and state‐of‐art
hydrometeorological modelling tools). However it should be stated that it is
impossible to estimate with certainty the upper tail of statistical distribution for a
flood parameter such as peak inflow or peak reservoir level. In his review of BC
Hydro’s Mica Dam stochastic flood modelling Vit Klemes (2000) stated that:

“... the factual necessary information (data + knowledge of the processes

involved) to which these technologies could be applied simply does not exist and
it cannot be manufactured just because it is needed for risk analysis. In this sense
the uncertainties are irreducible and the best we can hope for is to arrive at their
rough quantitative estimates and, based on them, to outline plausible credibility
windows or bands ...”

In general, a calibrated watershed model and hydrometeorological inputs

developed from historical data represent one plausible description of the “true
state of nature” for the behavior of a watershed. It must be recognized that
uncertainties exist in the estimation of hydrometeorological inputs and watershed
model parameters due to both aleatoric and epistemic reasons. Consequently,
there are many alternative combinations of probabilistic and deterministic models
and model parameters that could plausibly describe the “true state of nature”. For
instance, the extrapolation of plausible weather patterns is based on historical
precedent. Therefore, the potential for future climate change is one of the
uncertainties in this approach that needs to be handled via adoption of an
appropriate level of conservatism in selection of design parameters, combined
with a thorough uncertainty analysis. The goal of the uncertainty analysis is to
derive a mean‐frequency curve and uncertainty bounds for the various flood
characteristics in a manner that reasonably captures the current understanding
of the hydrologic behavior of the watershed as well as the effect of uncertainties
in estimating the flood‐frequency characteristics.

Considering physical complexity and number of parameters involved in the

process of flood simulation and routing through a system of dams and reservoirs,
it is practically impossible to accurately quantify associated uncertainties. It is
therefore prudent to use a parsimonious approach in the selection of
hydrometeorological inputs and watershed model parameters to be included in
the uncertainty analysis. This generally requires using a two‐step approach:

- Step 1 involves sensitivity analyses (and some engineering judgement) to

determine which hydrometeorological inputs and model parameters have the
greatest effect on the magnitude of the flood outputs of interest for a particular
dam or system of dams.

- In step 2, the uncertainty analysis is performed on the inputs/parameters

identified in Step 1.

3.4.1. Sensitivity analyses

Global Sensitivity Analysis (Saltelli et al., 2001) is recommended for use with
stochastic flood modeling applications, as with any other applications utilizing
Monte Carlo sampling approaches. This kind of sensitivity analysis is capable of
examining sensitivity with regard to the full range of parameter distribution. As
such, Global Sensitivity Analysis can measure the effect of interactions between
parameters and handle non‐linear behavior. In contrast, Local Sensitivity
Analysis (e.g. “one‐at‐a‐time” sampling) examines sensitivity only with regard to
point estimates of parameter values, which results in the sensitivity measure
being affected by the choice of parameter values.

Figure 3.10 depicts examples of the type of scatterplots produced from the Monte
Carlo simulations that were used to assess the sensitivity of the peak hourly
reservoir inflow to hydrometeorological inputs such as freezing level (i.e. the
altitude above which precipitation is not liquid) and temporal distribution of storm
precipitation.
Figure 3.10 Scatterplot showing moderate to high sensitivity of peak reservoir
inflow to freezing level and storm temporal distribution for the La Joie Dam
watershed in Canada

Similar scatterplots are produced for all flood characteristics (inflow, outflow,
reservoir level) and all relevant hydrometeorological inputs and rainfall‐runoff
model parameters; those exhibiting both high sensitivity and high level of
uncertainty in parameter estimation are typically selected to be included in the
uncertainty analysis. An example of this procedure as applied to the 1,000 km2
La Joie Dam watershed in British Columbia, Canada is presented in Table 3.2,
which shows a qualitative listing of the sensitivity of the peak reservoir level to the
various hydrometeorological inputs and watershed model
components/parameters. Table 3.2 also contains a qualitative assessment of the
relative magnitude of uncertainties for those inputs/parameters.
Table 3.2 Qualitative sensitivity of the La Joie Dam maximum reservoir level to
various hydrometeorological inputs and watershed model parameters

Model component Sensitivity to Magnitude of Comments

model component uncertainty
Storm seasonality Moderate Moderate Large sample set of
storms Greater
uncertainty for early‐
season storms
48‐hour basin‐ High Moderate to high Large sample of
average precipitation‐ annual maxima
frequency Uncertainty highest
relationship for for extreme events
watershed
Temporal and spatial High Low to moderate Diverse sample of
distribution of storms spatial/temporal
patterns for 22
prototype storms
Antecedent Moderate Low Adequate sample of
precipitation historical antecedent
precipitation
Antecedent soil Moderate Low Adequate sample of
moisture conditions historical antecedent
soil moisture
conditions
Baseflow Low Low Adequate sample of
historical baseflow
conditions
1,000 mb air Moderate Low to moderate Utilized in computing
temperature freezing level
Freezing level Moderate to high Moderate Runoff volume is
sensitive to freezing
level in winter months
Rainfall‐runoff Moderate Low to moderate Long record of
modeling historical flows for
watershed model
calibration
Snowmelt runoff Moderate Low to moderate Long record of
modeling historical flows for
watershed model
calibration
Watershed response Moderate Moderate to high Long record of
to fast runoff historical flows for
watershed model
calibration
Uncertainty in
response timing
Computation of peak Low to moderate Low to moderate Stage‐discharge
reservoir level via curves are typically
spillway stage‐ based on field testing
discharge curve and hydraulic model
results

The relative rankings in Table 3.2 were reviewed and candidates for inclusion in
the uncertainty analysis were identified as those inputs/parameters where there
was both moderate to high sensitivity and where there was a higher level of
uncertainty in estimation of the input/parameter. This assessment resulted in
identification of five components of the stochastic modelling framework to be
included in the uncertainty analysis for the La Joie Dam watershed as shown in
Table 3.3.
Table 3.3 Five model components selected for inclusion in the uncertainty
analysis for derivation of flood frequency curves and uncertainty bounds for the
La Joie Dam watershed

Model component Anticipated contribution to Comments

total uncertainty
48‐hour basin‐average High Uncertainty highest for
precipitation‐frequency extreme events
relationship for watershed
Watershed response to fast Moderate to high Affects magnitude and timing
runoff (watershed model of reservoir inflow flood peak
timing parameter)
1,000 mb air temperature Moderate Components for computing
air temperature and freezing
level hourly time‐series during
storms. They affect snowmelt
runoff particularly at high
elevations.
Freezing level Moderate
Storm seasonality Moderate Uncertainty for months where
precipitation magnitudes are
unrestricted

3.4.2. Uncertainty analyses

The first task in the process of uncertainty analysis is to identify sources of

uncertainty and how they can be characterized in the analysis. The process of
characterizing the various sources of uncertainty is best done in a team
environment where interaction of perspectives increases the understanding of the
flood behavior and facilitates identifying the most significant sources for a
particular watershed. It is important to distinguish between two main categories
of uncertainty:

- Aleatoric uncertainty is irreducible and associated with natural variability

of all flood producing factors including both atmospheric processes and
watershed hydrological response. For example, flood magnitude at a given
watershed at a specific time of year will vary not only due to atmospheric inputs
(rainfall and snowmelt) but also due to the chance occurrence of prior climatic
conditions and soil properties that led to soil moisture saturation level being what
it was at the time of flood. Values of flood producing factors are subject to chance
and the primary purpose (and greatest value) of stochastic flood modeling is in
addressing the aleatoric uncertainties associated with the hydrometeorological
inputs by treating those inputs as stochastic variables instead of fixed values.

- Epistemic uncertainty is associated with our lack of knowledge about a

particular variable or process and it may be reduced by a combination of research
and additional data acquisition. Some typical sources of epistemic uncertainty in
the flood modelling include: rainfall‐runoff model parameter uncertainty due to
incomplete understanding of the underlying physics of watershed hydrological
response; measurement errors in representation of watershed physical features;
selection of inappropriate theoretical probability distribution for describing
meteorological inputs (e.g. rainfall); uncertainties in reservoir storage‐elevation
curve or spillway discharge rating curve used in flood routing.

The aim of the uncertainty analysis employing the Monte Carlo procedures is to
derive a sample set of flood‐frequency relationships for flood characteristics of
interest by considering a sample set of "plausible model configurations". In this
context, the term "plausible model configurations" represents alternative
combinations of hydrometeorological inputs and alternative watershed model
parameters that could reasonably describe the "true state of nature" and are
selected from the global sensitivity analysis. All of the alternative combination of
hydrometeorological inputs and model parameters are "plausible" within the limits
of sampling variability of historical data, state‐of‐knowledge of the
hydrologic/hydraulic processes and flood modeling experience/judgment of the
analysts.

Figure 3.11 Flowchart for Monte Carlo frameworks for stochastic flood analysis
and uncertainty analysis
The flowchart in Figure 3.11 describes the process of conducting an uncertainty
analysis using a Monte Carlo framework and is based on the concept presented
by Nathan and Weinmann (2004). The inner loop is used to derive a flood‐
frequency relationship for flood characteristics for a given set of models/sub‐
models and model parameters and explicitly incorporates aleatoric uncertainty.
The outer loop represents alternative combinations of models/sub‐ models and
model parameters (alternative configurations) and represents epistemic
uncertainty in development of alternative plausible flood‐frequency relationships
for flood characteristics.
A number of alternative combinations of models/sub‐models and model
parameters could be assembled using sampling methodology (e.g. Latin‐
hypercube) to create a sample set of "plausible model configurations"
representing reality. This sets the number of repetitions for the outer loop in
Figure 3.11. Typically 10‐20 alternative model configurations are adequate to
reasonably determine the mean flood‐frequency relationship for flood
characteristics and to characterize the magnitude of the uncertainty bounds. One
practical option is to assemble 11 alternative model configurations because the
mean frequency curve and uncertainty bounds are often computed in a non‐
parametric manner by simple ranking of flood outputs for specific AEPs from the
group of alternative model configurations. Using the Cunnane’s (1978) non‐
parametric plotting‐position formula would result in 95th and 5th percentile flood
outputs (e.g. peak reservoir level) being the highest and lowest reservoir levels
generated from the 11 model configurations, respectively. Similarly, the median
value would be the 6th largest value and the mean value would be computed from
the 11 reservoir levels generated for a specific AEP.

A variation of this approach could be to combine all estimates into one, therefore
building a kind of predictive distribution. This would be particularly useful in
regulatory contexts and jurisdictions in which no mention is made of confidence
intervals (e.g. France), and where dam safety assessments are based on a single
value corresponding to a specific AEP (currently 10‐3 or 10‐4).

3.4.3. Characterization of uncertainties for selected model

components

Characterizations of uncertainties for each of the five selected model components

included in the uncertainty analysis of the La Joie Dam stochastic flood modelling
example (Table 3.3) are described in the following sections.

3.4.3.1. The 48-hour basin-average precipitation-frequency

relationship for watershed

The magnitude of precipitation relevant to dam safety analyses is typically several

orders of magnitude more extreme than what has been observed in the historic
record. As such, the estimation of this range of rainfall presents special difficulties
and requires the extrapolation of relatively short historical data records. This
extrapolation is challenging, especially considering that rainfall input is generally
the most significant contributor to resulting stochastically derived flood
hydrographs.
Due to hydrometeorological homogeneity, the 1,000 km2 La Joie Dam watershed
was coupled together with the adjacent 2,710 km2 Terzaghi Dam watershed,
forming the 3,710 km2 Bridge River basin for which the basin‐average
precipitation‐frequency relationship was developed. Precipitation annual maxima
series data were assembled for the critical storm duration (48‐hours in the case
of Bridge River basin) from all stations within and near the watershed. This totaled
178 stations and 7,589 station‐years of record for stations with 15 or more years
of record.

The precipitation‐frequency relationships for the Bridge watershed was

developed through regional L‐moment analyses of point precipitation and spatial
analyses of historical storms to develop point‐to‐area relationships and determine
basin‐average precipitation for the watershed using the 4‐parameter Kappa
distribution, which provided the best fit to the observed data sample. Two stations
(Downton Lake and Bralorne Upper) from within the Bridge basin were chosen
as explanatory stations for the 48‐hours basin‐average precipitation‐frequency
relationship. A multiple‐regression relationship was developed between the 48‐
hours precipitation maxima observed at explanatory meteorological stations and
maximum 48‐hours basin‐average precipitation for the Bridge River watershed
observed in the 22 historical storms (Figure 3.12). Monte Carlo methods were
used to generate 48‐hours precipitation‐frequency relationships for the
watershed accounting for sampling variability and uncertainties in estimation of
the various parameters employed in the computation (Figure 3.13). This
approach addressed the following sources of uncertainty associated with
development of the watershed precipitation‐frequency relationship:
- Estimate of mean for point precipitation used in regression,
- Regional L‐Cv,
- Regional L‐Skewness,
- Regional probability distribution,
- Point to area regression
Figure 3.12. Comparison of predicted and measured 48‐hours basin‐average
precipitation for Bridge River watershed based
on multiple‐regression prediction equation

Figure 3.13. Computed 48‐hour precipitation‐frequency relationship and 90%

uncertainty bounds for the Bridge River basin
3.4.3.2. Watershed response to fast runoff

Empirical likelihood functions were developed for the Fast Runoff Timing
Constant (FRTK) parameter within the UBC Watershed Model to characterize
uncertainties in estimation of parameter values for modelled Bridge River
watersheds. The FRTK parameter controls the timing of the watershed response
to fast runoff generation and affects the magnitude and timing of the flood
hydrograph peak. The best‐estimate values for FRTK were determined through
calibration of the watershed model to long‐term streamflow time‐series and to
historical floods. The shapes of the likelihood functions were based on experience
gained from modelling and calibration of the UBC Watershed Model at basins
throughout the world that were hydrologically similar to the Bridge River
watershed. Figure 3.14 shows the developed likelihood functions for both Bridge
River sub‐basins (La Joie and Terzaghi). The summary statistics for these
likelihood functions (in days) were:
- 0.5 for the mode/best estimate;
- 0.736 for the mean; and
- 0.357 for the standard deviation.

From these summary statistics, eleven values were selected for each watershed
using Latin‐ hypercube sampling methods.

Figure 3.14 Likelihood functions for fast runoff timing constant

for Bridge River watersheds
3.4.3.3. The 1,000 mb air temperature and freezing level

Uncertainties in the freezing level for a given stochastic storm simulation were
modelled through adjustment of the indexing value of the 1,000 mb air
temperature. This approach results in adjustment of the indexing value for the
freezing level and for setting the maximum freezing level for a given month. The
value of the 1,000 mb air temperature adjustment was found to be 1.3°C through
calibration to the historical flood‐frequency relationship at La Joie watershed,
which is the highest‐elevation watershed in the Bridge System with the highest
snowmelt runoff contribution.

This calibration should always be carried out as part of the stochastic flood
simulation process to ensure that the flood‐frequency characteristics predicted by
a stochastic model (SEFM in this case) were consistent with the behavior of
historically observed flood volumes. Observed reservoir inflow data for La Joie
was available for the 1961‐2010 period (50 years). Annual maxima 3‐day inflow
volumes were computed for the period of record and exceedance probabilities
determined using a plotting position approach. The 3‐day annual maxima were
scaled by 1.03 (standard adjustment for 3‐day to 72‐hour conversion) for
comparison with the 72 hour runoff volumes produced by SEFM. The 72‐hour
duration was chosen to capture the peak runoff from a wide range of storm
durations including some of the storm templates that exceeded the 48‐hour
critical duration discussed in Section 3.4.3.1.

During the calibration, SEFM model parameters had to be adjusted to achieve

good match with the historically observed reservoir inflows. In the La Joie case,
only the 1,000 mb temperature (generated by a physically‐based stochastic
model described in Section 3.3.1.5) needed to be adjusted upward by 1.3°C; a
minor adjustment and well within the uncertainty of parameter estimation. After
model calibration was completed, the 72‐hour runoff volumes from the SEFM
simulations closely matched the GEV frequency curve based on the observed
data to the 500‐year return period (Figure 3.15). Differences between simulated
and recorded flood‐frequency values for floods larger than the 25‐year return
period are the result of sampling variability in the stochastic model.
Figure 3.15 La Joie reservoir 72‐hour inflow magnitude‐frequency calibration
(SEFM and observed inflow)

Uncertainty in the 1,000 mb temperature adjustment was characterized °as being

equally‐likely over a range of +/‐ 2.0°C centered on the calibration value of 1.3°C.
Table 3.4 lists the eleven pairings of 1,000 mb air temperature adjustment and
maximum freezing level for the cool season (November to April).
Table 3.4 Uncertainty characteristics for sample set of eleven freezing level
parameters for Bridge River watersheds

1,000 mb Freezing Maximum freezing level (m)

Sample temperature level
set adjustment adjustment Nov Dec Jan Feb Mar Apr
(°C) (m)
1 ‐0.7 0 2,700 2,300 2,100 2,100 2,200 2,600
2 ‐0.3 0 2,800 2,400 2,200 2,200 2,300 2,800
3 0.1 0 3,000 2,500 2,300 2,300 2,400 2,900
4 0.5 0 3,100 2,600 2,500 2,400 2,500 3,000
5 0.9 0 3,200 2,800 2,600 2,500 2,600 3,200
6 1.3 0 3,400 2,900 2,700 2,700 2,800 3,300
7 1.7 0 3,500 3,000 2,800 2,800 2,900 3,500
8 2.1 0 3,700 3,200 3,000 2,900 3,000 3,600
9 2.5 0 3,800 3,300 3,100 3,100 3,200 3,800
10 2.9 0 4,000 3,500 3,300 3,200 3,300 3,900
11 3.3 0 4,200 3,600 3,400 3,400 3,500 4,100
3.4.3.4. Storm seasonality where precipitation magnitudes
are unrestricted

Precipitation magnitudes in stochastic flood simulations typically do not have an

upper limit. However, restrictions could be placed within the stochastic flood
modelling framework on the months where precipitation magnitudes may exceed
the Probable Maximum Precipitation (PMP) estimate. In general, precipitation
magnitudes are allowed to exceed PMP estimates only in those months
occupying the central body of the storm seasonality. This period varies depending
on regional climatology and stretches from October to mid‐March for the
Campbell River watershed (Figure 3.1), but is somewhat shorter (October
through February) for the Bridge River region. In those months that are external
to the central body of the seasonality data, restrictions are placed on precipitation
magnitudes and the upper limit to precipitation is set at the PMP estimate (Micovic
et al., 2015) for a particular month.
There is uncertainty in the seasonality of long‐duration storms with regard to the
months where precipitation magnitudes should be unrestricted. This uncertainty
was modeled by considering precipitation magnitudes to be unrestricted for the
central body of the seasonality distribution and then extending outwards to
consider additional months as unrestricted. Eleven sample sets for storm
seasonality were created (Table 3.5) where several monthly groupings were
applicable to several sample sets. In particular, the monthly groupings for sample
sets 4, 5, 6 and 7 and the percent PMP restrictions generally correspond to the
Bridge River region’s climatology.
Table 3.5 Restrictions on 48‐hour basin‐average precipitation expressed as
percentage of PMP

Monthly values of maximum 48‐hour precipitation expressed as percentage of PMP

Sample Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr
set
1 U U U U U U U U U U
2, 10, 11 75% U U U U U U U U U
3, 8, 9 52% 75% U U U U U U U 75%
4, 5, 6, 7 52% 53% 77% U U U U U 77% 54%
Note: ”U” corresponds to precipitation magnitudes being unrestricted

3.5. STOCHASTICALLY DERIVED FLOOD FREQUENCY RESULTS WITH

UNCERTAINTY BOUNDS

Monte Carlo computer simulations were used to develop magnitude‐frequency

relationships for maxima of reservoir inflow, outflow, and reservoir elevation for
individual dams or for a series of dams within a hydroelectric system. For the
examples presented in this chapter (Campbell River and Bridge River
hydroelectric systems), these relationships were based on 10,000 computer
simulations for each of the eleven "plausible model configurations". This
approach is based on the total probability computation procedure developed by
Nathan and Weinmann (2001) which greatly reduces the number of simulations
that would otherwise have been required to develop the flood‐frequency
relationships. In particular, it should be noted that each of the eleven "plausible
model configurations" was comprised of a random combination of the parameter
values for the eleven sample sets, per standard Latin‐hypercube methodology.

Flood outputs of interest from the stochastic flood model were presented as
probability‐plots developed using a non‐parametric plotting position. This
approach avoids the problems often encountered in selecting and fitting a
probability distribution, particularly for flood outputs such as reservoir levels that
have been greatly affected by anthropogenic factors such as imposed reservoir
operating procedures.
Each of the eleven plausible model configurations produces one flood‐frequency
relationship for a flood characteristic of interest. For example, Figure 3.16 depicts
the eleven flood‐frequency relationships for the peak reservoir inflow for the
1,000 km2 La Joie Dam watershed in British Columbia computed using the
stochastic flood model. The mean flood‐frequency curves and uncertainty bounds
for the peak reservoir inflow for and peak reservoir elevation for the same dam
are shown in Figures 3.17 and 3.18, respectively. Note that Figure 3.18 is
particularly important in terms of dam safety considerations and decisions
because it provides information on the probability of dam overtopping.

Figure 3.16 Example of simulated flood‐frequency relationships for peak hourly

reservoir inflow at La Joie Dam for eleven plausible model configurations
Figure 3.17 La Joie Dam – Frequency curve for peak hourly reservoir inflow
simulated by stochastic flood model

Figure 3.18 La Joie Dam – Frequency curve for peak hourly reservoir level
simulated by stochastic flood model

3.6. CONCLUDING REMARKS

Stochastic flood simulation for dam safety is a multi‐disciplinary analysis aiming

to realistically capture interactions among components of an extremely complex
system, considering both natural and anthropogenic inputs. It requires technical
expertise in several fields such as meteorology, hydrological modelling and
sophisticated statistical analyses, as well as significant amount of
hydrometeorological data. Consequently, it is relatively expensive and currently
carried out only by a limited number of dam owners in several countries. Also, the
considerable data requirements prevent applying this approach in many
developing countries at the present time – however that is likely to change in
future with continuous increase in computing power and advancements in both
remote sensing and data acquisition technologies.
As mentioned earlier, the stochastic simulation approach to flood hazard yields
the full continuous probability distributions of reservoir elevation which could then
be used to evaluate annual exceedance probabilities of various reservoir levels
including the level corresponding to the dam crest (dam overtopping level) as well
as the level resulting from the PMF. The approach is recommended for risk
assessment studies associated with high hazard/high consequence dams (or a
national system of flood protection dykes like in the Netherlands) where potential
costs associated with various risk reduction measures are several orders of
magnitude higher than the cost of a stochastic flood simulation study.

Conversely, the highest standard‐based criterion for dam safety is the PMF which
is typically based on the Probable Maximum Precipitation (PMP). In theory, both
the PMP and the PMF represent notional upper limits of rainfall and flood
magnitude, respectively, and there is zero chance of them being exceeded. In
reality, they change over time as new data and methodologies become available,
as well as depending on approaches and level of subjective judgement different
analysts use in deriving the estimates.

Micovic et al. (2015) analyzed the main meteorological components of a PMP

and demonstrated that the uncertainty associated with final PMP estimates (and
corresponding PMF results) can be very high, implying that the PMP and the PMF
are not true physical limits but simply convenient engineering concepts. Thus, a
dam owner using the PMF standard for dam safety may wonder what happens if
the current PMF changes or how much larger (or smaller) it could be. Or, what
are the risk levels associated with different PMFs as well as the costs of potential
risk reduction measures.
These dilemmas could be illustrated by the example of the Cherry Creek Dam in
USA, described in Downton et al. (2005). In 1995, the 24‐hr PMP for the
watershed was estimated by the U.S. National Weather Service to be 536 mm.
The U.S. Army Corps of Engineers (USACE) then used this PMP to derive the
PMF and concluded that the dam could safely control only 75% of the PMF.
Consequently, USACE informed the public that the dam was unsafe and
proposed several alternative improvements to the dam, reservoir, and basin,
costing up to $250 million. In 2000, the PMP study was carried out by another
analyst who came up with the lower 24‐hr PMP value of 401 mm (i.e. about 75%
of the 1995 value). Note that this lower PMP would result in lower PMF (assuming
all other PMF inputs remain unchanged) and indicate that the dam could safely
handle more than 75% (and possibly 100%) of the new PMF, likely bringing the
costs of dam safety improvement down significantly from the original $250 million
estimate.

The Cherry Creek Dam is a good example where additional information obtained
through a stochastic flood study could be applied to enhance dam safety critical
decisions and justify cost and effort of undertaking the study. The full probability
distribution (with uncertainty) of all flood characteristics would help in assessing
the conservatism (i.e. AEP) of different PMF estimates as well as in making risk
informed decisions on potential dam safety upgrades.
In addition, the availability of the full probability distribution of various flood
characteristics allows us to estimate joint probabilities and identify potential failure
modes that cannot be analyzed using strict “pass/fail” deterministic design criteria
(e.g. dam failure due to a modest flood combined with a failure of one or more
spillway gates).

As mentioned in the introductory section of this chapter, dams often fail during a
non‐extreme flood, due to a combination of two or more relatively usual
unfavourable events, which individually are not safety critical. Therefore, the
stochastic flood simulation approach for dam safety introduced in this chapter,
despite being technically complex and carried out by only a small number of
specialists, represents a valid and necessary alternative to standard‐based
deterministic methods, especially when risk estimates and associated dam safety
decisions/operating policies are required over the full range of hydrologic loading
conditions.

Finally, with intrinsic levels of uncertainty that will never be resolved, “satisfying
yourself that you have results that can be relied on to make dam safety critical
decisions” becomes a matter of acceptance of some residual level of risk.
Applying the most rigorous scientific analyses to improve the understanding of
these uncertainties is the best that any owner or regulator can do to satisfy
themselves they have done as much as is practicable in arriving at their
decisions.

3.7. REFERENCES

ANCOLD (2017), Guidelines on selection of acceptable flood capacity for dams.

Baecher, G.B., Hartford, D.N.D., Patev, R.C., Rytters, K., Zielinski, P.A. (2013),
Understanding spillway systems reliability, in Proceedings of ICOLD 81st Annual
Symposium, Seattle, WA, USA, pp. 1241‐1250.

Boughton, W., Droop, O. (2003), Continuous simulation for design flood

estimation ‐ a review, Environmental Modelling and Software, 18 (4), 309‐318.

Cunnane, C. (1978), Unbiased plotting positions – a review, Journal of Hydrology

37, 205‐222.

Downton, M.W., Morss, R.E., Wilhelmi, O.V., Gruntfest, E., Higgins, M. (2005),
Interactions between scientific uncertainty and flood management decisions: two
case studies in Colorado. Environmental Hazards 6, 134–146.
FEMA (2013), Selecting and accommodating inflow design floods for dams,
FEMA report P‐94.

Garavaglia, F., Gailhard, J., Paquet, E., Lang, M., Garçon, R., Bernardara, P.
(2010), Introducing a rainfall compound distribution model based on weather
patterns sub‐sampling. Hydrology and Earth System Sciences, 14 (6), 951–964.

Hansen, E.M., Fenn D.D., Corrigan, P., Vogel, J.L., Schreiner, L.C., Stodt, R.W.
(1994), Probable Maximum Precipitation for Pacific Northwest States – Columbia
River (including portions of Canada), Snake River and Pacific Coastal Drainages,
Hydrometeorological Report No. 57, US National Weather Service, Silver Spring,
MD, USA.

Hegnauer, M., Beersma, J.J., van den Boogaard, H.F.P., Buishand, T.A.,
Passchier, R.H. (2014), Generator of Rainfall and Discharge Extremes (GRADE)
for the Rhine and Meuse basins – Final report of GRADE 2.0, Report for
Rijkswaterstaat, Ministry of Infrastructure and the Environment, The Netherlands,
Reference No. 1209424‐004‐ZWS‐0018.

Hosking, J.R.M., Wallis, J.R. (1997), Regional frequency analysis ‐ an approach

based on L‐ moments, Cambridge University Press, Cambridge, UK.

International Committee on Large Dams (2003), Dams and floods, guidelines and
case histories. Bulletin 125, ICOLD, Paris.

Klemes, V. (2000), Review of stochastic modelling of extreme floods for Mica

Dam, personal communication, October 2000.
Lewin, J., Ballard, G., Bowles, D.S. (2003), Spillway gate reliability in the context
of overall dam failure risk, in Proceedings of the 23rd Annual USSD Conference,
Charleston, SC, USA.

McKay, M.D., Conover, W.J., Beckman, R.J. (1979), A comparison of three

methods for selecting values of input variables in the analysis of output from a
computer code, Technometrics 21, 239‐245.

Micovic, Z., Quick, M.C. (2009), Investigation of the model complexity required in
runoff simulation at different time scales, Hydrological Sciences Journal, 54(5),
872‐885.

Micovic, Z., Schaefer, M.G., Taylor, G.H. (2015), Uncertainty analysis for
Probable Maximum Precipitation estimates, Journal of Hydrology 521, 360‐373.

Micovic, Z., Hartford, D.N.D., Schaefer, M.G., Barker, B.L. (2016), A non‐
traditional approach to the analysis of flood hazard for dams, Stochastic
Environmental Research and Risk Assessment 30, 559‐581.

Nathan, R.J. (2017), Current trends in the evaluation of extreme floods, In Flood
evaluation and dam safety, Bulletin 170, ICOLD, Paris.
Nathan, R.J., Weinmann, P.E. (2001), Estimation of Large to Extreme Floods,
Book VI, In Australian Rainfall and Runoff – A Guide to Flood Estimation, The
Institution of Engineers, Canberra, Australia.
Nathan, R.J., Weinmann, P.E. (2004), An improved framework for the
characterisation of extreme floods and for the assessment of dam safety, In
Hydrology: Science & Practice for the 21st Century, Vol. 1, pp. 186‐193,
Proceedings of British Hydrological Society, London, UK.

Nathan, R.J., Hill, P.I., Weinmann, P.E. (2011), Achieving consistency in

derivation of the Probable Maximum Flood, in Proceedings of the ANCOLD 2011
Conference on Dams, Melbourne, Australia.

Paquet, E., Garavaglia, F., Gailhard, J., Garçon, R. (2013), The SCHADEX
method: a semi‐ continuous rainfall‐runoff simulation for extreme flood
estimation, Journal of Hydrology 495, 23–37.

Quick, M.C. (1995), The UBC Watershed Model, in Computer models of

watershed hydrology, edited by V.J. Singh, pp. 233‐280, Water Resources
Publications. Highlands Ranch, CO, USA.

Saltelli A, Chan K., Scott E.M. (2001), Sensitivity Analysis, John Wiley and Sons,
1st edition, New York, NY, USA.

Schaefer, M.G., Barker, B.L. (2002), Stochastic Event Flood Model (SEFM), In
Mathematical models of small watershed hydrology and applications, edited by
V.J. Singh and D.K. Frevert, pp. 707‐748, Water Resources Publications,
Highlands Ranch, CO, USA.

Wyss, G.D., Jorgenson, K.H. (1998), A user’s guide to LHS: Sandia’s Latin
Hypercube Sampling software, Sandia National Laboratories, Report
4. RESERVOIR INFLOW PREDICTION FOR PROACTIVE
FLOOD RISK MANAGEMENT

4.1. FUNDAMENTALS AND BASIC PRINCIPLES OF HYDROLOGICAL

AND HYDRO-METEOROLOGICAL INFLOW PREDICTION

Inflow prediction is the art of estimating the further development of reservoir inflow
(and—additionally—preferably a measure of uncertainty/reliability) with the aim
of fostering specific demands and informed decisions connected to reservoir
operation at a wide range of temporal scales. The fundamental problem tied to
the task of forecasting is, that an unknown future state has to be inferred from a
mostly limited and erratic prior knowledge of the considered system. Regarding
the aforesaid, forecasting strategies, methods, etc. have to be somewhat oriented
toward the specific situation, problem, or management context.

Following this philosophy, the herein touched aspects should be discussed

alongside the different temporal scales at which specific management problems
are dealt with and management decisions are effective (Table 4.1).
Table 4.1 Temporal scales and their connection to typical reservoir
management problems, associated with inflow estimation
and flood risk management

Temporal scale Long-term Seasonal Single-event

Dam
Dam release
Storage planning release/spillway
management
management
Management Design and
Withdrawal (Withdrawal
problem optimization of rule
management management)*
curves
Climate change Water quality (Water quality
adaption management management)*
*Set in brackets since measures are only effective to a certain extent on the respective
temporal scale

4.1.1. Forecasting/prediction methods

There is extensive literature reviewing the methods and their applicability for
reservoir inflow forecasting, e.g., Easey et al. (2006) and WMO (2011), to name
only two sources. When dealing with the matter, an individual assessment of the
current literature is strongly recommended. Typically, in the literature, methods
which are applicable for reservoir inflow forecasting are classified according to
Table 4.2. Furthermore, the table lines out the suitability of these methods
regarding the temporal scale, and therefore the set of reservoir management
problems, given in Table 4.1.
Table 4.2 Suitability of different methods for inflow prediction and forecasting
according to the temporal scale of interest

Method/suitability
Statistical Dynamic
Temporal scale Downscaling
methods modeling
Long-term + + —
Seasonal + + +
Single-event + — +

Statistical methods incorporate mathematical models that only address the

input-output transfer-behavior of the considered system and explicitly do not
regard the underlying physical processes. Typical representatives for this group
of models are regression models, fuzzy-logic models and artificial neural network
models. The clear advantages of statistical models are their usually
straightforward and robust conceptualization, parameterization and operation.
Typical shortcomings are the high demands regarding the amount and quality of
required data and the limited spatial transferability alongside with partly limited
temporal extrapolation behavior (considering event magnitudes which are not
covered by underlying empirical data).
Dynamic modeling methods rely on an abstract, mathematically expressed
conceptualization of the physical processes which are inherent to the considered
system. This representation can either be directly oriented towards the occurring
processes (usually by incorporating the governing physical conservation laws) or
rely on a more abstract, yet representative portrayal of reality, e.g., via variable
storage-tank models, which are very common in engineering hydrology (Ponce,
1994, to name only one). Dynamic modeling has the highest potential for an
accurate representation of reality. On the other hand, dynamic models are
virtually arbitrarily demanding in terms of conceptualization, parameterization and
operation.

Downscaling methods hold an intermediate position in the way that they can
rely on either statistical or dynamic modeling. The aim of downscaling is to project
the output of long-term (e.g., climate models), seasonal, or short-termed
circulation models (“weather models”, “atmospheric models”, or “numerical
weather prediction [NWP] models”) to spatial scales which are applicable for
inflow prediction, e.g., for the underlying rainfall-runoff/water balance modeling.
Circulation models (CMs) can be applied at global, regional or local scale;
commonly, nested model chains are operated, where a global CM (GCM) yields
the boundary condition for a regionally nested regional CM (RCM) and the RCM
in turn drives a local CM (LCM) with typical horizontal resolutions of few to some
kilometers.
It should be mentioned and underlined that the decision for or against a specific
modeling approach should essentially be as simple as possible and as complex
as necessary.

4.1.2. Decision making based on potentially uncertain forecasts

Uncertainty is inherent to the Earth system and to our knowledge of this system.
The knowledge of uncertainty can help in fostering and informing decisions. In
this light, uncertainty becomes useful and could be considered as a characteristic
of reliability, something which is definitely desired in operational engineering
tasks and also, e.g., for reservoir-related dimensioning purposes. Regarding the
herein discussed problem of inflow forecasting and ignoring the fact at this point
that there are different sources of uncertainty, it is of very importance to
understand that the uncertainty of the inflow forecasting problem is steadily
dependent on lead time and the spatial scale or scope (Figure 4.1); uncertainty
usually will rise with increasing lead times and a decreasing spatial scale.
The aim of uncertainty analysis is to gather information on the chance of
occurrence of a specific event (e.g., a specific flow rate or flow volume) in the
form of a probability distribution, also called the predictive uncertainty. It is
important to note that this probability distribution is conditional on the
aforementioned errors and uncertainties, as well as their statistical
interdependence and therefore represents the prior knowledge of the future for
the point in time where the forecast is created.

Hydrological forecasts are affected by uncertainty, emerging from various

sources (Grundmann, 2010; Klein et al., 2016). Basically, uncertainty is
distinguished in aleatoric and epistemic uncertainty. Aleatoric uncertainty is the
statistic uncertainty, emerging from the limited knowledge of the Earth system
and its abstraction in the context of models. Epistemic or systematic uncertainty
emerges from an insufficient knowledge of the considered system and its state
and causes input parameter uncertainty, structural/model uncertainty, model
parameter uncertainty (e.g., time steps, convergence criteria, etc.), and process
parameter uncertainty (i.e., the physically-based or conceptual “hydrological”
parameters).
Figure 4.1 Conceptual illustration of the dependency of predictive uncertainty
on increasing lead time and decreasing spatial scale

This said, it is clear that when decisions should be informed and when looking at
the usually meso-scale catchment areas of multi-purpose reservoirs, a significant
amount of predictive uncertainty will be present, should be assessed and, in turn,
could be made available for informed decision making. However, uncertainty
quantification is not straightforward when it is performed according to the state of
the art in sciences and technology and can be demanding in terms of
methodological and/or computational demand. With reference to the
aforementioned principle of Occam’s razor (“As simple as possible and as
complex as necessary”), uncertainty estimation should therefore be only applied
when these premises are on the table and clear to all involved decision makers.
In contrast, a limited, methodologically insufficient uncertainty estimation should
only be pursued with caution since it will likely lead to wrong or erratic decisions
(Todini, 2016).

Most crucial for taking informed decision is the step of risk assessment. For
instance, when there is information on target storage threshold exceedance for a
specific lead time in terms of a probability 5, it has to be clear what the “risk
tolerance” should be or, in other words, when action is taken (e.g., evacuate water
to provide volume for retaining predicted flood events). It is indicated that risk
assessment is placed on a preferably solid ground of information, for instance by
simulating costs and benefits emerging from specific management decisions with

5 E.g., resulting from an uncertainty analysis.

regard to multi-criteria operational requirements (e.g., flood protection and water
supply). Last not least, it should be mentioned that emotion always plays a role
in framing a decision (e.g., Kahneman & Tversky, 1979; Druckman & Mcdermott,
2008). That is why altered/optimized/improved management plans should always
be developed and discussed with the operational personnel on the ground and
potentially affected stakeholders.

4.2. PREDICTION-BASED FLOOD RISK MANAGEMENT ON DIFFERENT

TEMPORAL SCALES

As already presented (Table 4.1), inflow forecasting strategies for different flood
risk management purposes are oriented to a specific temporal scale (long-term,
seasonal, short-term/single event). Typically, different management decisions
are made on the basis of specific model-based analyses. For instance, for long-
term dimensioning tasks a (statistical) time-series model may be appropriate,
whereas for operational flood-control purposes, a real-time rainfall-runoff
simulation approach, coupled to a weather forecast model may be preferred.
Briefly said, there is no overarching approach capable of fostering the wide range
of reservoir-related management and planning decisions (Table 4.2).

However, the recently emerging paradigm of seamless prediction aims at

bridging the gaps between event-related forecasting to seasonal prognoses and
even long-term considerations on the climate scale (e.g., Meissner et al., 2014).
The advent of seamless prediction is directly connected to (a) the nesting of
regional and local circulation models (RCMs/LCMs) within global circulation
models (GCMs) and (b) the trend towards using (dynamic) GCMs as climate
models, which was foremost driven by increasingly available computational
power. An analog co-evolution is to be expected on the “hydrology side” where
highly detailed models are more and more driven on regional up to continental
spatial scales for analyses ranging from short-term water balance considerations
up to long-term climate impact studies (e.g., see Samaniego et al., 2017).

4.2.1. Long-term scale

Long-term inflow prediction/estimation is important in two different contexts:

(1) for the initial dimensioning of the storage of a planned reservoir and (2) for
verifying, adapting and potentially improving management strategies for changing
(i.e., non-stationary) boundary conditions, namely as a consequence of climate
change.
4.2.1.1. Inflow forecasting/prediction strategies on the long-
term scale

Long-term inflow prediction/estimation can be achieved by statistical/stochastic

modeling on the one hand and downscaling from weather/climate model output
on the other hand (see Table 4.2).

Historically, the state of the art in long-term inflow prediction was empirically
based on observed (or estimated, e.g., via continuous rainfall-runoff modeling)
inflow time series of limited length. The simplest (but still to a certain extent
skillful!) strategy would be drawing a climatology out of the empirical observations
and use it as a predictor for inflow estimation. To strengthen the statistical basis
of the underlying data and to finally infer information on the reliability of results,
component-based time series models then became the method of choice.
Such models typically assume that a time series is composed by superimposed
information of statistical moments (e.g., expected value, variance, skewness and
kurtosis), periodicity and autocorrelation, trend (i.e., non-stationarity), and a
stochastic (i.e., random) component (“noise” or residuals). Such an approach can
be formalistically expressed as:

𝑥𝑥𝑖𝑖 = 𝑑𝑑𝑖𝑖 + 𝑝𝑝𝑖𝑖 + 𝑡𝑡𝑖𝑖 + 𝑒𝑒𝑖𝑖

with xi :… time series elements,

di :… deterministic component,
pi :… periodicity,
ti :… trend component, and
ei :… stochastic/probabilistic component;
a multiplicative composition is also applicable.
The general approach of time series modeling is to estimate the components of
the (additive or multiplicative) time series model by applying suitable methods to
the empirical observation data (e.g., estimation of statistical moments, trend
analyses, spectral analyses, like Fourier transformations, etc.) in form of a so-
called decomposition 6. This way, a (stochastic) time series model is
parameterized and can then be used to reproduce the characteristics of the
observed time series by running it sufficiently often. The set of output time series
can then be delivered to the actual management/dimensioning/planning
evaluations (e.g., a sequent-peak procedure for supply storage dimensioning;
Potter, 1977 or more complex, so-called generalized reservoir-system operation
models [GRSOMs]; Müller, 2014) which provides an empirical distribution of
statistically equivalent results which in turn can deliver a reliability/probability
information (e.g., supply security) via an appropriate frequency analysis.

6 There is a plethora of decomposition techniques. Despite still being often used, the
classical decomposition based on the mentioned additive or multiplicative time series
concept has a number of flaws and several much better methods are available by now
(Hyndman & Athanasopoulos, 2018).
Typical time series models which have been employed for long-term inflow
predictions are linear auto-regressive types, e.g., like the much used Thomas-
Fiering model (Fiering, 1971). In recent time, more sophisticated, potentially non-
linear models, like Wavelet approaches, exponential smoothing or autoregressive
conditional heteroskedasticity (ARCH) models emerged, to name only a few. An
entry to this wide topic is provided by each good textbook on time series modeling
in sciences (e.g., Hyndman & Athanasopoulos, 2018). It needs further to be
mentioned that—as a consequence of their flexibility and ability to portray
arbitrarily complex nonlinear functions—neural network models are also well-
capable of portraying observed time series and being widely used for time series
generation.
The other main path for long-term inflow estimation is made up by downscaling
approaches employing dynamic modeling methods. The philosophy here is to
use observed or projected (i.e., from climate models) rainfall, climate data, etc.
and to perform retrospective/projective hydrologic (i.e., rainfall-runoff) modeling
analysis (Beven, 2012). Since historic data may be rare, a quite new but
potentially promising strategy is posed by so-called global reanalysis data. Such
data is generated by running a GCM in retrospective mode (called hindcasting).
Of course, reanalysis data may hold a vast amount of uncertainty which can be
potentially assessed by incorporating probabilistic reanalysis products (like
NOAA/CIRES, for instance) and applying ensemble techniques. Depending on
the spatial resolution of reanalysis data, the further use of downscaling may be
indicated (by employing nested RCMs and LCMs). Usually, climate model output
not only covers a reanalysis part (e.g., 1950 to 2015), but also a projected time
frame (e.g., 2015 to 2100), which of course could also be used as input for long-
term hydrological modeling.

It should be mentioned that circulation model output will not necessarily match
the local conditions (e.g., the climate or rainfall characteristics of a reservoir
catchment) exactly. Therefore, it might be indicated to correct the CM and/or
hydrologic model output to better portray local conditions, which is called bias
correction. There is a vast number of methods for performing bias correction of
CM output to be found in the literature, whereas an individual assessment is
strongly recommended.

4.2.1.2. Excursus: Multi-criteria optimization

Long-term management problems typically address storage planning, design and

optimization of rule curves (or guide curves), or questions of climate change
adaption. Usually, multi-purpose reservoir management faces multiple, partly
contradictory objectives, e.g., flood protection, power generation or drinking water
withdrawal/supply. Hence, an integrative approach needs to anticipate that there
is not a single optimal solution to a specific problem, but rather a whole set of
compromise solutions, which are called the Pareto-optimal solutions.
A specificity of multi-purpose reservoir management is that the intended
operational purposes usually are competing, e.g., flood retention (requires a
preferably large cleared storage volume) vs. water supply security (preferably
widely filled storage) vs. energy generation (preferably large pressure
head/reservoir water level). The most common way of dealing with this issue in
terms of optimizing reservoir management strategies is employing the concept of
Pareto-optimality (Figure 4.2).

Figure 4.2 : Illustration of the Pareto principle.

Figure modified after Müller, 2014

The n points represent the outcome (in terms of specific objective functions) of
n management scenarios regarding the competing purposes flood retention and
water supply. So-called compromise (or Pareto-) solutions are located along the
Pareto front, whereas the compromise solution located nearest to the Utopia point
(lower left corner) usually is considered as being balanced with respect to the
competing operational demands.

The concept of Pareto-optimality is closely tied to a multi-criteria optimization

approach where all relevant optimization criteria (e.g., minimal flood damage
alongside with maximum water supply security) are simultaneously considered.
The methodological challenge to that problem is to apply an efficient and ample
optimization strategy to get a sufficiently fine impression on the Pareto front
without the need of processing all (i.e., an infinite number) possible management
scenarios.

For engineering purposes, multi-criteria optimization is most often executed by

means of simulation-based approaches, i.e., management premises are
established, a generalized reservoir-system operation model is used to compute
the impact of these specific premises on the interesting variables (e.g., storage
volume, release rate, pressure head) and the results are then assessed using
multiple criteria. To cope with potentially excessive computational demands,
appropriate sampling and/or recombination techniques are essential in order to
efficiently sample the (a priori unknown) solution domain (e.g., Vrugt et al., 2003;
Krauße et al., 2012; Müller, 2014).
4.2.1.3. Considering uncertainty on the long-term scale

On the long-term, uncertainty can be considered as a measure of reliability.

Reliability is a statistical term which is tied to considering a sufficiently large
number of outcomes in order to evaluate their appropriateness towards a specific
aim. However, our knowledge on the governing processes (e.g., inflow time
series) is limited (aleatoric uncertainty; Section 4.1.2). This means that long-term
considerations have to be executed not only once (i.e., deterministic) but often
(i.e., probabilistic/stochastic). Again, this can be achieved by means of
statistical/stochastic methods or by downscaling approaches (Table 4.2 and
Section 4.2.1.1).

In the domain of statistical/stochastic methods, time series modeling

(Section 4.2.1.1) is used to generate a sufficiently broad set of synthetic time
series in order to appropriately sample the statistical properties of the underlying,
observed empirical data. Finally, the set of synthetic series can be included in
further steps, e.g., dimensioning considerations. For instance, driving a storage
dimensioning model (e.g., like the Sequent Peak Algorithm, to name only a
common and simple one) for n times (for n synthetic input time series) will lead
to n results/capacities which then can be statistically evaluated, for instance, in
terms of drawing exceedance probabilities/reliabilities for a given storage
capacity.

It is important to check, if the pool of synthetic (i.e., “generated”) time series

resembles the original time series, which was a priori used for the
parameterization of the time series model. This is exemplarily shown in Figure 4.3
for different moments/parameters of mean monthly inflow values (mean, standard
deviation, skewness coefficient, and the coefficient of autocorrelation). Basically,
moments of higher order (like the skewness) tend to show a higher variance when
the synthetic series are compared against the original time series.

Figure 4.3 : Statistical Moments of monthly flow data -

Modified after Müller (2014).
On Figure 4.3 the historical data are represented by the bold black line. The box
plots shows the range of empirical data of the pool of generated time series. The
right-hand section of each panel shows the annual mean value of the considered
statistical moments (125 time series and time series length is 80 years).
Another aspect for evaluating the statistical robustness of the synthetic data pool
is to determine the convergence behavior of the time series’ statistical properties,
e.g., represented by their statistical moments. As an example, Figure 4.4 shows
the convergence of the standard deviation vs. the number of realizations (i.e.,
number of generated time series) for seven different input time series.

Figure 4.4 : Convergence of the standard deviation of synthetic time series

(y-axis) depending on the number of generated series (given on the x-axis).
Modified after Grundmann (2010).

Opportunity for uncertainty estimation is also offered via the methodological path
of downscaling of circulation model output. For instance, there are probabilistic
reanalysis datasets like NOAA/CIRES, which can be used to force hydrological
modeling and to generate statistically interpretable inflow forecasts. On the
projection side, different climate forcing scenarios are typically used to generate
an ensemble of climate predictions. Furthermore, there are a number of
ensemble climate projections available. Again, as already stated in
Section 4.2.1.1, an appropriate bias correction of CM input (preprocessing)
and/or hydrological model output (postprocessing) might be indicated, whereas
this is beyond the scope of this publication and would require an individual review
of the relevant literature.
4.2.2. Seasonal scale

“Seasonal scale” usually refers to a time scale ranging from weeks to months.
Typically, management decisions are more influential on the seasonal than on
the short-term side (Visser, 2017). For instance, the potential for a short-termed,
operational mitigation of already observed or short-lead predicted flood inflow
remains very limited, e.g., due to the limited hydraulic capabilities of the release
devices and installations. Therefore, predicting the inflow regime on the seasonal
scale is most crucial for undertaking informed reservoir management decisions,
especially with regard to balancing multiple management purposes. However,
seasonal inflow forecasts may not be able to deliver skillful predictions every time
(Turner et al., 2017); despite more and better seasonal forecasting systems are
emerging, it is still with significant effort to raise the skill above a certain baseline,
given by flow climatology data (Section 4.2.2.1). A very profound review of the
current state of the art in seasonal hydrological forecasting can be obtained from
a special issue of HESS (edited by Wetterhall et al., 2016-2018); Easey et al.
(2006) give much briefer review.

4.2.2.1. Inflow forecasting/prediction strategies

The most basic statistical inflow prediction technique is to rely on the

climatology inherent to flow data, i.e., “historical” data”. Flow data can be
instantly available through gauging measurements or can be obtained by
regionally transferring flow measurements from upstream or downstream
locations or neighboring catchments. Additionally, flow data can be obtained by
driving hydrological (rainfall-runoff) models by observed or reconstructed
(i.e., reanalysis) rainfall and weather data.
Of course, for deriving a somewhat significant prediction of flows from
empirical/historical data, the incorporated time series need to be sufficiently long.
However, the statistical assessment of long time series offers a way to quantify
uncertainty or exceedance/non-exceedance probabilities, respectively
(Figure 4.5 and Section 4.2.2.2).
Figure 4.5 : Quantile-based flow climatology for river gauge Görlitz
at Lausitzer Neiße River, Germany. 7
Turning to more elaborate statistical methods, a broad field is made up by
regression methods (typically multiple linear regression methods), where, e.g.,
antecedent conditions and specific climate indices (e.g., teleconnections) are
used as predictors for streamflow (Ionita, 2017 or Charles et al., 2018, to name
only two). In other words, those methods are based again on historical data and
intend to “find statistically significant relationships between one or more
predictors and the target variable of interest, e.g., river flow or precipitation”
(Easey, 2006). However, these approaches always introduce the (unrealistic!)
assumption of stationary.

It is of highest importance to account for errors and biases—which inevitably

emerge by employing statistical methods—by appropriate post-processing
approaches. Foremost, bias correction techniques are used to reduce statistical
errors and to adjust model results to observed data/climatology. Despite the task
of bias correction is almost indispensable when applying statistical prediction
techniques, the frame of this chapter does not allow for a more in-depth
elaboration on the topic. For further reading, Crochemore et al. (2016) deliver a
profound entry to the topic.

Seasonal inflow predictions can further be obtained by employing dynamic

modeling techniques. This might be the method of choice when meteorological
forcing drive the streamflow predictability (Arnal, 2018), which is usually the case
for smaller catchments (e.g., the headwater catchments of reservoirs). Seasonal
circulation models (e.g., ECMW’s System 4 8 or NOOA’s CFSv2 9) can deliver
skillful forecasts of meteorological parameters with a lead time of several weeks

7 Climatology is empirically drawn from daily flow data from 1946 to 2017 and can deliver
a probabilistic estimate on seasonal flow conditions.
8 https://www.ecmwf.int/en/forecasts/documentation-and-support/evolution-
ifs/cycles/seasonal-forecast-system-4
9 http://www.cpc.ncep.noaa.gov/products/CFSv2/CFSv2_body.html
to some months. In turn, these meteorological forecasts can be used for driving
a (already calibrated) hydrological model, delivering seasonal reservoir inflow
predictions. As already mentioned alongside statistical prediction techniques,
bias correction/calibration techniques usually need to be applied in order to
improve forecasting skill.
Circulation models are available for different specific spatial and temporal scales;
the aim of downscaling is to project the output seasonal circulation models to
spatial scales which are applicable for the intended inflow prediction task (i.e., a
subsequent rainfall-runoff modeling). This can either be done by “nesting” higher-
resolved circulation models within coarser seasonal forecasting system or by
using statistical techniques to infer the regional shaping of rainfall and all other
relevant meteorological drivers from a more common (e.g., global) circulation
model. In recent years, skillful seasonal meteorological forecasting systems did
rapidly evolve. In this light, it seems legit to assume that quite highly (spatially
and temporally) resolved seasonal weather forecasts data will be more and more
instantly available in the near future.

4.2.2.2. Considering uncertainty on the seasonal scale

One of the oldest and maybe the most common method to inform reservoir
release decisions on the seasonal scale incorporating uncertainty (at least to a
certain extent) is a technique called Ensemble Streamflow Prediction,
developed initially for California (ESP; Wood et al., 2016). According to Arnal
(2018), ESP “relies on the correct knowledge of the initial hydrological conditions
(i.e., of snowpack, soil moisture, streamflow and reservoir levels, etc.) and a large
land surface memory, and contains no information on the future climate”. The
information “on the further climate” was obtained by EOF analyses (empirical
orthogonal functions) which related indices like the Pacific sea surface
temperature (SST) to future California climate, for example. Since EOF thrives
from harmonic analysis, the methodology was capable of not only delivering one,
but a whole set of future predictions (i.e., the ensemble; Figure 4.6). This, in turn,
allowed for deriving failure probabilities for reservoir levels over a future period of
up to six months.
Figure 4.6 : An ESP illustrated by J. C. Schaake in 1978, from “Extended
streamflow prediction techniques: description and applications during 1977”,
taken from Wood et al., 2016.

Nowadays, uncertainty information can be drawn from ensembles emerging from

seasonal ensemble forecasting systems, like the already mentioned ECMWF
System 4 or NOAA’s CFSv2. These systems deliver ensembles by letting the
circulation models run several times under slightly altered (“perturbed”) initial
conditions (and maybe other parameters, e.g., cloud physics parameters, etc.).
Another way of obtaining ensembles is to combine the output of different models
(so-called multi-model ensembles). However—as stated a couple of times
throughout this chapter—in order to draw frequency/reliability information from an
ensemble, it needs to be ensured that the statistical properties of an ensemble
resemble reality. This is rather the exception than the rule and therefore, bias
correction/calibration techniques should be applied.

Although the consideration of ensembles is not free of methodological burden,

there are potential benefits for seasonal inflow prediction in terms of reliability,
robustness, and statistical significance. Further reading on the matter can be
obtained from Huang & Loucks (2000) or Georgakakos & Graham (2008), to
name only two sources.

4.2.3. Short-termed, single-event scale

The practical importance of short-termed, single-event forecasts for reservoir

management often remains limited. This is caused by a set of relevant, partly
interdependent restrictions:

(1) Multi-purpose reservoirs are often located in smaller headwater

catchments; hydrological reaction is swift and inflow forecasts must be
based on hydro-meteorological forecast, in order to extend lead times to
an applicable scale. However, it can be assumed that a skillful
meteorological forecast considering (heavy) rainfall events (those, that
are primarily relevant for a proactive reservoir management), is only
possible with a lead time of a couple of hours to some days, at most,
depending on aspects like circulation patterns and the governing rainfall
generation mechanisms. For instance, widespread frontal rainfall is
predictable with a significantly better skill than convective rainfall,
featuring a high spatio-temporal variability (Figure 4.7).

Figure 4.7 : AUC skill score for two different QPF products (COSMO-DE and
COSMO-DE-EPS) depending on different lead times and different spatial
scales. 10 Taken from Schütze et al. (2016).

(2) Especially on small spatio-temporal scales, inflow forecasts are

encumbered with significant uncertainty. This uncertainty should be taken into
account (e.g., with ensemble methods), which is partly methodically demanding
and often hinders a pragmatic, operationally applicable uncertainty consideration.
Furthermore, multi-purpose-reservoirs need to serve multiple, partly competing
demands, e.g., flood protection vs. energy production. This leads to a pareto-
optimal problem which means that an optimal solution can be established by a
set of different solutions/considerations for each single operational purpose
(Krauße, 2012 and Müller, 2014).
(3) Considering the potentially swift/flashy hydrological reaction and limited
lead times (especially for small, fast-responding catchments), the technical ability

10 The Y-axis shows the AUC. The X-axis indicates the range of spatial scales. AUC
can be assumed to be “good” for values above 0.8. Each set of bars indicates one
specific lead time [1, 3, 6, 9, 12, 15, 18, 21 hrs].
for an operational reaction is often limited, e.g., by the hydraulic capabilities of
the release devices and installations.
(4) Rapid and short-term pre-event releases are not only constrained by
technical limitations but are also dependent on the applicable legal framework.
For instance, it is usually prohibited to increase pre-release rates to a damage-
causing level (above rule release).

Figure 4.8 underlines the afore given points; the figure shows an example for
Malter reservoir (East Germany; capacity: 8.78 million cubic meters, catchment
area: ~105 km²) with actual and projected operational strategies during the
extreme 2002 European Floods (return period > 100 years). It can be seen that
even if there would have been a damage-causing (maximum damage-free
downstream flow is ~15 m³/s) pre-release of 30 m³/s, starting from the point where
the first official flood warning was disseminated, an effective peak reduction
would not have been possible (yellow line). Furthermore, in order to reduce the
outflow to a moderately damage-causing rate of 75 m³/s, the (already damage-
causing!) pre-release would have needed to be initiated two days in advance,
leading to a completely dry (!) reservoir by the onset of the flood event. Of course,
this scenario is not applicable by any means, considering the remaining purposes
of the reservoir and the tremendous uncertainties in the rainfall forecasts for a
100 km² catchment.

Figure 4.8 : Operation of Malter reservoir during the 2002 European Floods
under different pre-release scenarios.

The description of the lines on Figure 4.8 is presented hereafter :

• Orange : Actual inflow;

• Green : Actual outflow;
• Yellow line : Projected (30 m³/s, starting from the initial flood warning;
• Purple line : Projected (45 m³/s, assuming a reliable, 48-hrs inflow
forecast);
Data were provided by the State Dam Authority Saxony (LTV Sachsen) and the
scenario analyses were from the Institute of Hydrology, Dresden University of
Technology.
Notwithstanding the fact that the impact of short-term management decisions on
reservoir operation is limited, it is still recommended, especially for smaller
catchments, to assess the forecast benefit for a specific reservoir, operational
strategy, etc. in order to determine the limits and also the potential benefits drawn
from an inflow forecast. Therefore, the following two sections shed a light on
short-term forecasting strategies and an easy method to determine the forecast
benefit as a function of lead time, which can deliver an estimate on the minimum
required lead times for an efficient release management.

4.2.3.1. Prediction strategies and aspects of operational

short-term inflow forecasting

An applicable inflow forecasting strategy is dependent on different aspects. Of

highest importance is the spatial scale/catchment area of the considered
reservoir location. For larger catchments (say beyond 10,000 km²) or reservoirs
on large rivers, the inflow can be forecasted by means of statistical modeling
(regression models, neural networks, fuzzy models) or hydrologic/hydraulic
process oriented models. Where there is upstream gauging data and the mean
translation time from the upstream gauging locations to the reservoir site is
sufficient for management purposes, a broad range of methods may be applied
and will deliver quite reliable results.

If catchment size decreases to an intermediate scale (area of some thousands

km²) and/or lead time should be extended, there is a need to include rainfall
observations/estimates. First, this requires the employment of rainfall-runoff
models; second, spatio-temporal rainfall estimation and rainfall forecasts will
introduce additional uncertainty.

For small headwater catchments (area smaller 1,000 km²), rainfall forecasts are
needed to at least ensure somewhat reasonable lead times. This, of course,
introduces the highest amount of uncertainty and also requires spatio-temporal
highly resolved rainfall-runoff modeling.

There are many other important aspects that need to be assessed when
developing an operational short-term inflow forecasting system, e.g., data
situation, operational robustness, implementation and operation costs, to name
only a few. For further details, the WMO Manual on Flood Forecasting and
Warning holds extensive information on the topic (WMO, 2011).
4.2.3.2. Excursus: Assessing the benefit of a forecast
depending on lead time

A meaningful extension of a forecast’s lead time is at least costly or at worst

impossible, when it comes to small catchments and/or larger lead times
(i.e., beyond some days). Therefore, it is important to determine the relation of
forecast benefit (e.g., in terms of peak outflow reduction 11) and the lead time. A
very basic consideration towards estimating this relation is described in the
following (Figure 4.9):

The flood control volume should be assumed to be seven (7) volume units (VU).
Furthermore, an inflow design hydrograph should be available (e.g., from
hydrograph analyses or regionalization studies). For a lead time of zero (upper
left panel in Figure 4.9), there would be no prior information/estimate for fostering
a specific release control; therefore, release is just adjusted to the rule release
rate (green horizontal line; assumed to be 2 VU/TU). This holds up until the flood
control volume is exhausted (which is the case after 4+1/3 TU); release will then
directly follow the inflow.

With extended lead time, there is a better prior knowledge on the expected inflow
hydrograph (in terms of a “look-ahead”); therefore, controlled outflow would be
modified (dashed line), leading to a better peak flow reduction. Expressing the
peak flow reduction in mathematical terms can be done, e.g., by calculating the
root-mean-square error (RMSE) between the status-quo outflow (for a lead time
of zero) and the outflow for a specific, nonzero lead time. The inverse RMSE can
then be interpreted as benefit, which is displayed in Figure 4.10 for the discussed
example.

11Comparable methodology could be applied for other beneficial (and potentially

competing!) reservoir functions, e.g., water supply.
Figure 4.9 : Simple assessment of the forecast benefit (expressed as peak
reduction) using a release strategy based on available look-ahead time (lead
time).

It is assumed that the release can be freely regulated without hydraulic

restrictions and with no time lag. Release is established using the specific
available look-ahead data and regarding a maximum peak reduction within the
range of the lead time.

Figure 4.10 : Benefit (peak flow reduction) as a function of lead time

(data according to Figure 4.9).

Figure 4.10 shows that the benefit is saturated beyond a specific lead time. By
means of such a simple approach, it is at least possible to get a valuable estimate
on the minimum required lead time for achieving a noticeable peak reduction. In
turn, this is quite important since the implementation and operational costs, as
well as the data demand of a forecasting system are directly depending on
desired lead times.

4.2.3.3. Considering uncertainty on the short-term scale

For short-termed inflow forecasts for smaller headwater catchments, usually

quantitative precipitation estimates (QPEs) and—additionally—quantitative
precipitation forecasts (QPFs) need to be taken into account. QPFs introduce the
vast amount of uncertainty in this case (input parameter uncertainty) and should
therefore be the first in the focus when it comes to uncertainty analysis.

In hydrological forecasting, the aim should be to reduce the total uncertainty, e.g.,
by means of data assimilation, state updating and bias correction methods
(WMO, 2011). On the other hand, which is even more important, dealing with
uncertainty can produce an uncertainty quantification, which, in turn, can deliver
an estimate on prediction errors or the reliability of forecasted values, for
instance.

Typically, ensemble methods are employed for uncertainty estimation on

shorter time scales. This means, that the aforementioned epistemic uncertainty
ranges are sampled alongside a known or estimated prior distribution of the
considered parameters. In the case of input parameter uncertainty, this is—for
instance—typically done by using a set (or ensemble) of possible future
realizations of the further rainfall development. This strategy delivers not only one
(deterministic) output, but a set of outputs (the so-called raw ensemble).

The emerging set of outputs is not identical with the predictive uncertainty. Taking
the raw ensemble as the basis for probabilistic estimations (e.g., via empirical
quantile analyses) is methodically wrong, since the statistical interdependence of
the (conditional!) uncertainty sources is ignored this way (Todini, 2016;
Hernández-López & Francés, 2017 and see Figure 4.11).

For the sake of brevity, the concerning methods for joint (mostly Bayesian)
uncertainty processing and predictive uncertainty estimation (like GLUE, BaRE,
and BATEA, to name only a few) cannot be discussed in-depth within this frame.
Rather the relevant literature in this field is recommended (Beven, 1993; Kuczera
and Parent, 1998; Bates et al., 2001; Kavetski et al., 2002; Kuczera et al., 2006)
whereas Grundmann (2010) and Klein et al. (2016) provide an easy access to
this quite sophisticated topic.

However, it should be stated that a state-of-the-art uncertainty assessment

requires properly conditioned probability densities (e.g., statistically
representative and reliable ensemble forecasts, process and model parameters,
and so forth; Todini, 2016), which is a very challenging objective in practice. Not
least, potentially extraordinary high computational costs may and most likely will
hinder the implementation of (especially Monte-Carlo based) methods in real-
time sensitive contexts.

Figure 4.11 : Comparison of a raw ensemble (upper panel) and a processed

ensemble as “true” estimate of the predictive uncertainty (below).
From Klein et al. (2016).

A better coverage of the observed data by the probabilistic (lower panel) forecast
(observations are always “within” the forecast ensemble) as well as the better
sharpness/lower spread of the probabilistic product vs the raw ensemble can be
noted on the figure. Further note that 10% of the estimated possible realizations
of river flow lie outside the depicted 5-95% range.

4.3. DISCUSSION AND RECOMMENDATIONS

Long-term inflow prediction is important mainly for the sake of dimensioning.

Despite statistical models introduce some major shortcomings (assuming
stationarity, co-linearity or dependence of predictors; Wedgbrow et al., 2005),
such approaches are still superior for long-termed predictions and are much
easier to apply than dynamic modeling techniques (at least, when not using
instantly available dynamic modeling output). Long-term inflow prediction is
usually carried out by statistical/stochastic modeling or downscaling from
weather/climate model output.
For seasonal forecasting purposes, dynamic circulation models nowadays hold
a high degree of sophistication, are usually physically-based and should therefore
be able to generally outperform statistical approaches. In practice, this is only the
case for lead times of up to one or two months (Arnal, 2018). The limited
knowledge of initial conditions (mainly due to lacking data) and the generally
chaotic behavior of the Earth system jeopardize an exact prediction of the
dynamic evolution of the atmosphere for lead times of more than two or three
months. However, compared to the situation before 10 years (e.g., Easey et al.,
2006), the skill of dynamic modeling approaches did improve a lot. The main
reasons for that are better and spatio-temporally higher-resolved models together
with improved computational resources on the one hand, as well as better data
availability (mainly from satellite-based remote sensing) and strongly improved
data assimilation techniques (e.g., 4D-VAR 12) for inferring the initial condition of
the Earth system, on the other hand.

Short-termed inflow forecasts extend to a couple of days. For reservoirs in

smaller headwater catchments, such lead times would be sufficient for
undertaking appropriate flood-control decisions. However, uncertainty of rainfall
forecasts and the reliability of subsequent rainfall-runoff prediction is important
for an informed decision-making. Unfortunately, the predictive skill of quantitative
precipitation forecasts is limited with regard to small headwater catchments,
especially for extended lead times in the medium range (up to two weeks). That
is why the incorporation of uncertainty and proper predictive uncertainty
estimation is important. For larger catchments, short-termed predictions can be
carried out with essentially higher skill, even by using quite simple statistical
approaches, given, that data which firmly correlate to the river flow (e.g.,
upstream flow data) are available.

Regardless the considered temporal scale and the specific prediction task, the
consideration of predictive uncertainty is strongly recommended to better
inform management decisions (e.g., Delaney, 2018). Especially for smaller
headwater catchments, predictive uncertainty will strongly influence the reliability
of model results and therefore can potentially jeopardize deterministic model
results. However, a proper uncertainty estimation may be cumbersome in terms
of methodology and technical demands and therefore should be thoroughly
addressed in the conceptualization of inflow prediction projects and tasks.

4.4. CASE STUDIES

The following three case studies are presented in Appendix A :

12 https://www.ecmwf.int/en/about/media-centre/news/2017/20-years-4d-var-better-
forecasts-through-better-use-observations
1. Long-term scale: Multi-objective optimization of multi-purpose reservoir
systems under high reliability constraints - Germany
2. Seasonal scale: Seasonal (mid-range) forecasts for reservoir operation
adaptation to mitigate shifting precipitation patterns - Germany
3. Short-term scale: Use of reservoir storage for flood operation in the
Kumano River basin - Japan

4.5. REFERENCES

Arnal, L.: Skilful seasonal forecasts of streamflow over Europe? HEPEX-Blog:

https://hepex.irstea.fr/skilful-seasonal-streamflow-europe/, 2018.
Bates, B. C. & Campbell, E. P.: A Markov chain Monte Carlo scheme for
parameter estimation and inference in conceptual rainfall-runoff modeling. Water
Resources Research, 37(4), 2001.

Beven, K. J.: Rainfall-Runoff Modelling: The Primer, John Wiley and Sons
Chichester, 2012.

Beven, K. J.: Prophecy, Reality and Uncertainty in Distributed Hydrological

Modeling. Advances in Water Resources, 16(1), 1993.

Charles, S. P.; Wang, Q. J.; Ahmad, M.; Hashmi, D.; Schepen, A.; Podger, G. &
Robertson, D. E.: Seasonal streamflow forecasting in the upper Indus Basin of
Pakistan: an assessment of methods. Hydrol. Earth Syst. Sci., 22, 2018.

Crochemore, L.; Ramos, M.-H. & Pappenberger, F.: Bias correcting precipitation
forecasts to improve the skill of seasonal streamflow forecasts. Hydrol. Earth
Syst. Sci., 20, 2016.

Delaney, C.; Mendoza, J; Whitin, B. & Hartman, R.: Using ensemble forecasts to
inform risk-based operations of a reservoir in Northern California. HEPEX-Blog:
https://hepex.irstea.fr/using-ensemble-forecasts-to-inform-risk-based-
operations-of-a-reservoir-in-northern-california/, 2018.

Druckmann, J. N. & McDermott, R.: Emotion and the Framing of Risky Choice.
Political Behaviour, 30, 2008.

Easey, J.; Prudhomme, C. & Hannah, D.: Seasonal forecasting of river flows: a
review of the state-of-the-art. Proceedings of the Fifth FRIEND World Conference
held at Havana, Cuba, 2006.

Fiering, M. B. & Jackson, B. B.: Synthetic Streamflows. AGU Water Res.

Monographs, 1971.
Georgakakos, K. P. & Graham, N. E.: Potential Benefits of Seasonal Inflow
Prediction Uncertainty for Reservoir Release Decisions. Journal of Applied
Meteorology and Climatology, 47, 2008.
Grundmann, J.: Analyse und Simulation von Unsicherheiten in der
flächendifferenzierten Niederschlags-Abfluss-Modellierung. Dissertation, TU
Dresden, Institut für Hydrologie, 2010.
Hernández-López, M. R. & Francés, F.: Bayesian joint inference of hydrological
and generalized error models with the enforcement of Total Laws. Hydrology and
Earth System Sciences, 19, 2017.

Huang, G. H. & Loucks, D. P.: An inexact two-stage stochastic programming

model for water resources management under uncertainty. Civil Engineering and
Environmental Systems, 17(2), 2000.

Hyndman, R. J. & Athanasopoulos, G.: Forecasting: Principles and Practice.

OTexts, 2018.
Ionita, M.: Mid range forecasting of the German Waterways streamflow based on
hydrologic, atmospheric and oceanic data, Reports on polar and marine
research, Alfred Wegener Institute for Polar and Marine Research Bremerhaven,
711, 2017.

Kahnemann, D. & Tversky, A.: An Analysis of Decision under Risk. Econometrica,

47(2), 1979.

Kavetski, D., Franks, S.W. & Kuczera, G.: Confronting input uncertainty in
environmental modelling. In: Q. Duan, H.V. Gupta, S. Sorooshian, A.N.
Rousseau und R. Turcotte (Editors), Calibration of Watershed Models. AGU
Water Science and Applications Series, 2002.

Klein, B.; Meißner, D.; Hemri, S. & Lisniak, D.: Ermittlung der prädiktiven
Unsicherheit von hydrologischen Ensemblevorhersagen. Report BfG1853,
Federal German Institute of Hydrology, 2016.

Krauße, T.; Cullmann, J.; Saile, P. & Schmitz, G. H.: Robust multi-objective
calibration strategies – possibilities for improving flood forecasting. Hydrology
and Earth System Sciences, 16, 2012.

Kuczera, G. & Parent, E.: Monte Carlo assessment of parameter uncertainty in

conceptual catchment models: The Metropolis algorithm. Journal of Hydrology,
211(1-4), 1989.
Kuczera, G.; Kavetski, D.; Franks, S. & Thyer, M.: Towards a Bayesian total error
analysis of conceptual rainfall-runoff models: Characterising model error using
storm-dependent parameters. Journal of Hydrology, 331(1-2), 2006.

Meissner, D.; Klein, B.; Lisniak, D. & Pinzinger, R.: Probabilistische Abfluss- und
Wasserstandsvorhersagen – Kommunikationsstrategien und Nutzungspotentiale
am Beispiel der Binnenschifffahrt. Hydrologie und Wasserbewirtschaftung 58(2),
2014.
Müller, R.: A new strategy for a multi-criteria simulation-based management of
multi-purpose dam systems (Dissertation in German). Institute of Hydrology,
Dresden University of Technology, 2014.
Ponce, V. M.: Engineering Hydrology: Principles and Practices. Prentice Hall,
1994.

Potter, K. W.: Sequent peak procedure: minimum reservoir capacity subject to

constraint on final storage. JAWRA, 13(3), 1977.

Samaniego, L.; Kumar, R.; Thober, S.; Rakovec, O.; Zink, M.; Wanders, N.;
Eisner, S.; Müller Schmied, H.; Sutanudjaja, E. H.; Warrach-Sagi, K. & Attinger,
S.: Toward seamless hydrologic predictions across spatial scales. Hydrology and
Earth System Sciences, 21, 2017.

Schütze, N.; Singer, T. & Wagner, M.: Endbericht Niederschlagsverifikation –

Verifikation probabilistischer quantitativer Niederschlagsvorhersageprodukte
(QPF) im Hinblick auf deren Eignung als Antrieb für ein
Hochwasserfrühwarnsystem für kleine Einzugsgebiete in Sachsen. Institut für
Hydrologie und Meteorologie, Technische Universität Dresden, 2016.

Todini, E.; Singh, V. P. (Ed.): Handbook of Applied Hydrology: Predictive

Uncertainty Assessment and Decision Making – Theory and Applications.
McGraw-Hill, 2016.

Turner, S. W.; Bennett, J. C.; Robertson, D. E. & Galelli, S.: Complex relationship
between seasonal streamflow forecast skill and value in reservoir operations.
Hydrol. Earth Syst. Sci., 21, 2017.

Visser, J.: Evaluation of Seasonal Inflow Forecasting to Support Multipurpose

Reservoir Management – A case study for the Upper Maule River Basin, Chile.
MSc Thesis Lund University, 2017.

Vrugt, J. A.; Gupta, H. V.; Bastidas, L. A.; Bouten, W. & Sorooshian, S.: Effective
and efficient algorithm for multiobjective optimization of hydrologic models. Water
Resour. Res., 39, 2003.

Wedgbrow, C. S.; Wilby, R. L. & Fox, H. R.: Experimental seasonal forecasts of

low summer flows in the River Thames, UK, using Expert Systems. Climate Res.
28, 2005.

Wetterhall, F.; Pechlivanidis, I. G.; Ramos, M.-H.; Wood, A.; Wang, Q. J.; Zehe,
E. & Ehret, U. (Eds.): Special issue on sub-seasonal to seasonal hydrological
forecasting, HESS, 20-22, 2016-2018.

WMO: Manual on Flood Forecasting and Warning. World Meteorological

Organization, 2011.

Wood, A.; Pagano, T. & Roos, M.: Tracing The Origins of ESP. HEPEX-Blog:
https://hepex.irstea.fr/tracing-the-origins-of-esp/, 2016.
APPENDIX A
RESERVOIR INFLOW PREDICTION FOR PROACTIVE
FLOOD RISK MANAGEMENT

CASE STUDIES
A1 - LONG-TERM SCALE
MULTI-OBJECTIVE OPTIMIZATION OF MULTI-PURPOSE RESERVOIR
SYSTEMS UNDER HIGH RELIABILITY CONSTRAINTS - GERMANY

Dr. Ruben Müller

BAH Consulting (Büro für Angewandte Hydrologie), Berlin, Germany

1) Introduction
Most reservoirs in non-arid regions are managed with the aim to comply with
municipal water supply security policies that require an occurrence-based
reliability (Hashimoto et al. 1982) of 99.0 % or more. Simulation models allow for
a rather complicated rule set and therefore are typically used to validate if the
operation of a reservoir (given by a set of operational rules) can provide such high
reliabilities. Simulation-based optimization methods are capable to solve
problems that involve reliability (Koutsoyiannis and Economou, 2003) and can be
used to find new operational rules or to optimize existing rules.

For this case study, a method by Müller and Schütze (2017) is used to derive best
operational rules for the time period between 2021 to 2050 for climatic conditions,
as given by the WETTREG-2010 projections (Kreienkamp et al., 2010) under the
IPCC storylines (A1B, B1 and A2) for a reservoir system in Eastern Germany
under high reliability constraints.

Multi-objective optimization is used in order to compare possible trade-offs

between conflicting goals under changed climatic conditions, compared to the
current status quo. Furthermore, optimally adapted operational rules are
necessary for an unbiased assessment of reservoir performance under changed
conditions (Eum and Simonovic, 2010).

2) Methodology

Estimation of future inflow conditions

To estimate inflows into the reservoir system under the WETTREG-2010 climate
projections, water-balance simulations were carried out with the WASIM-ETH
hydrologic model. Total inflows to the system are projected to decrease in
average from 2.1 m³/s (period 1921 to 2007) to 1.5 m³/s in the scenario A1B,
1.8 m³/s for scenario B1, and 1.6 m³/s in the scenario A2, equaling a reduction of
inflow volumes of up to 25 %. Of course, other statistical properties of the inflow
are subject to change under projected change scenarios. More details can be
found in Müller (2014).
Time series modeling/generation

The projected (deterministic!) inflow time series from the hydrologic model are
then used as a basis to generate sufficiently long/many time series to further
address reliability estimation purposes. A nonparametric, K-nearest-neighbor
(KNN) neural network approach (Ashrafzadeh and Rizi, 2009) has been used for
the stochastic simulation of lag-1 auto-correlated streamflows. The main
advantage of the employed model, compared to other state-of-the-art KNN time
series models, is the generation of inflow value magnitudes, which were not
observed in the historical record. To prevent an underestimation of drought
periods, the model was extended with a symmetric moving-average filter, similar
to a method introduced by Langousis and Koutsoyiannis (2006). For further
details concerning the time series model, see Müller and Schütze (2017) and
Müller (2014). Statistics of the generated time series pool are shown in
Figure A1.1. The length of the generated time series is 10,000 a.
Reduction of complexity: Shortening of time series by applying recombination
techniques

Koutsoyiannis and Economou (2003) show that under simplified assumptions, a

simulation period of several thousand years may be required in order to properly
assess reservoir operations under high-reliability constraints (e.g., for 99 %).
Running a simulation with such long time series is no huge demand with the
computational power available with personal computers today. However, in
simulation-based, multi-objective optimization frameworks, simulation models
may need to be run for several thousand times. Here the computational demand
proves to be an unacceptable burden. Müller and Schütze (2017) therefore
proposed a method to reduce the length of time series significantly while
preserving crucial statistical and stochastic features which are needed to
accurately describe the inflow regime of a reservoir system.

The methodological steps comprise (a) the multivariate generation of long time
series with KNN neuronal networks (previous section). Then, (b) drought periods
in the time series are identified using the sequent-peak algorithm. In a final
Monte-Carlo sampling step (c) a subset of drought periods is selected, such that
the original distribution of deficit volumes is preserved. For details, see Müller and
Schütze (2017). Properties of the resampled time series pool can be seen from
Figure A1.1. Resampling of the time series led to a length of 800 to 1,000 years
per series.
Figure A1.1 : Comparison of monthly and annual statistical properties of
exemplary inflow data (inflow to Lehnmühle Reservoir) for 120 simulations from
the KNN model (circles), historical flows (solid gray circles) and “shortened”
(i.e., resampled) time series (diamond). Taken from Müller and Schütze (2017).
Generalized reservoir-system operation model (GRSOM) and multi-objective
optimization

For the simulation of the water resources system, the widely used GRSOM
OASIS (Hydrologics, 1991) is applied. OASIS uses a fast-mixed integer linear
programming algorithm to distribute water per time step in an optimal manner.
The built-in OCL programming language makes OASIS flexible and easy
adoptable. The Multi-Objective Covariance Matrix Adaptation Evolution Strategy
(MO-CMA-ES; Igel et al., 2007) was used for solving the multi-criteria/objective
optimization problem (Section A1.4).

Level diagrams as a tool for the visual assessment of Pareto-optimal results

Blasco et al. (2008) introduced so-called level diagrams for visualizing high-
dimensional Pareto fronts. The basic idea is to calculate the distance between
each solution within a Pareto set, given in the fitness function space, and the
Utopia point (the best, but non-existing, solution one could construct of each best
value of each fitness function). For visualization, each fitness function has its own
representation in the level diagram. Each solution is plotted with its fitness
function value on the abscissa, and the respective value of the calculated
distance to the Utopia point on the ordinate. A specific solution has the same
position on the ordinate in each representation (“they are leveled”). This
synchronization by means of the calculated norm on the ordinate holds the key
to easily visualize high-dimensional data sets. As an example, in Figure A1.3,
solution (2) is depicted for each sub plot with the same ordinate value.
3) The reservoir system
The reservoir system for the case study is located in the Eastern Ore Mountains
in Saxony, Germany, to the southwest of the city of Dresden and comprises three
reservoirs, namely Lehnmühle (LM), Klingenberg (KL) and Rauschenbach (RB),
see Figure A1.2(a). The operational storage capacities of the reservoirs are
16.32 hm³ for KL, 19.42 hm³ for LM and 15.20 hm³ for RB.

Figure A1.2: (a) Map of the reservoir system in the federal state Saxony in
Germany and (b) illustration of the reservoir system in the simulation model
OASIS.
The reservoir system has multiple purposes. Domestic, industrial and municipal
water for the cities of Dresden and Freital is supplied by the reservoir KL.
Ecological minimum flows need to be ensured; see Figure A1.2(b). Additionally,
all reservoirs serve for flood protection, which is considered implicitly in this study,
as the flood storages of the reservoirs need to be kept free in normal operation.
Domestic, industrial and municipal water is supplied according to the supply
levels 1 to 3. These levels are governed by two monthly rule curves, ZKL,1 and
ZKL,2, which divide the combined storage volumes of reservoirs KL and LM into
three sections. With falling storage in both reservoirs, the supply rate is reduced
according to Table A1.1(a). With decreasing supply, the associated reliability
increases. To ensure a good water quality of the supply during a drought,
reservoir LM is drawn down first to a specific threshold level and only after that
threshold is reached, reservoir KL is drawn down subsequently.

Reservoir RB is supporting reservoir KL by providing water by means of a trans-

catchment diversion system (“RWA” in Figure A1.2(a)). As applies for the supply
of water by reservoir KL, the diversion is also governed by two monthly rule
curves, Zdiv,1 and Zdiv,2, which divides the combined storage of the reservoirs KL
and LM into three sections with individual diversion rates, see Table A1.1(b).
Table A1.1: (a) Supply Rates and required reliabilities for the three supply levels
and (b) the diversion rates from reservoir RB for the three diversion levels.

(a) Rate Reliability (b) Rate

Level Level
(m³/s) (%) (hm³/month)
Supply level 3 1.000 99.00 % Diversion level 3 0.0
Supply level 2 0.925 99.50 % Diversion level 2 0.4
Supply level 1 0.850 99.95 % Diversion level 1 0.6

4) Formulation of the multi-criteria optimization problem

Multiple operational purposes are often in conflict. Three fitness functions (FF)
are considered in order to address the most important operational purposes in
the frame of a multi-criteria optimization:

• FF1 maximizes the intended reliabilities (R) which are associated with
the three supply levels (SL) and can be formally written as
max(FF1) = max(R(SL1) + R(SL2) + R(SL3))
• FF2 maximizes the probability that the reservoir is filled up to target
storage (i.e., “full”) at the end of April in order to provide a good water
quality during summer
• FF3 minimizes the total amount of water provided by reservoir RB in
order to minimize pumping costs for water diversion

To obtain optimal operational rules, the four rule curves (ZKL,1, ZKL,2, Zdiv,1, Zdiv,2)
are subject to optimization. Since storage values for the rule curves are needed
for every month, the resulting optimization problem considers 48 decision
parameters (4 by 12).

5) Results

Multi-objective optimizations were carried out for recent conditions (1921-2007)

and three projected climate-change scenarios. In Figure A1.3, all Pareto fronts
from all multi-objective optimization runs are plotted in a single level diagram and
are ranked (three fitness functions FF1-3) against each other. Additionally, the
achieved reliabilities R(SL1-3)) for the three supply levels are given. The
representation of a single Pareto front of one optimization is surrounded by a hull,
for better visibility. This is a novel interpretation of the original take on level
diagrams by Blasco et al. (2008).
Regarding Figure A1.3, the “higher” (i.e., in positive ordinate direction) a single
Pareto-front (within a hull) is, the less competitive the reservoir system is
compared against “lower” (i.e., in negative ordinate direction) Pareto fronts from
other scenarios (another hull). However, every solution of a single Pareto front
(within one hull) is still an optimal management option for that scenario and the
associated constraints!
Figure A1.3: Modified Level-diagram showing the Pareto fronts
for all optimization epochs:
recent conditions (blue);
2021-2050 conditions A1B (red), A2 (pink) and B1 (gray).
The numbered solutions for a single Pareto front are the most balanced
compromises between all contradictory purposes/management goals. Because
of decreasing inflows when comparing the status quo with A2 to A1B conditions,
the performance of the reservoir system decreases in general, the numbered
solutions are situated “higher” on the ordinate. With increasing FF3 values from
recent conditions to A1B, more water has to be diverted to reservoir KL to mitigate
the decreasing mean inflows. Nevertheless, the chance of a filled reservoir KL in
April decreases, as documented by simultaneously smaller FF2 values. Apart
from that, scenario B1 is a special case. Despite slightly lower mean inflows to
the reservoir system, less severe drought conditions are projected. Therefore, the
overall performance of B1 is actually slightly better than under recent conditions.
The most balanced solution of B1 has even a lower ordinate value than that of
the recent scenario.

Table A1.2 provides additional information for a decision maker by listing the
most extreme and balanced compromise solutions. Looking at the solutions with
the maximum FF1 values, max(FF1), the required reliabilities can only be
achieved in recent conditions and for B1. Under A1B and A2 conditions, the
reliability for supply level 1 falls short of being met for A1B and A2. Additionally,
higher diversion rates (FF3) are needed and the water quality may decrease
because of smaller FF2 values.
Table A1.2: Summary of reliability assessment results for the three supply
levels R(SL1) to R(SL3) and the fitness functions FF1 to FF3 for selected
solutions for the scenarios (A1B, B1, B2) in 2021-2050 and under recent
conditions (status quo). Reliabilities, which do not meet target reliabilities
according to Table A1.2(a) are marked in red color.

Scenario Solution R(SL1) (%) R(LS2) (%) R(LS3) (%) FF1 FF2 FF3
Target: Target: Target: (-) (%) (hm³/
99.950 99.50 99.00 month)
Recent max(FF1) 99.972 99.93 99.88 2.998 81.56 0.093
max(FF2) 99.995 99.89 99.71 2.996 91.35 0.280
min(FF3) 99.957 99.73 99.14 2.988 77.44 0.056
A1B max(FF1) =
99.909 99.83 99.53 2.993 74.38 0.468
max(FF2)
min(FF3) 99.328 98.71 96.63 2.947 29.01 0.121
B1 max(FF1) =
99.998 99.99 99.97 3.000 92.91 0.290
max(FF2)
min(FF3) 99.964 99.63 99.09 2.987 69.93 0.015
A2 max(FF1) 99.921 99.90 99.82 2.996 72.55 0.232
max(FF2) 99.937 99.88 99.27 2.991 77.80 0.400
min(FF3) 96.252 93.71 87.01 2.771 39.00 0.001

6) References

A. Ashrafzadeh and A. P. Rizi: A hybrid neural network based model for

synthetic time series generation. In: International Symposium on Water
Management and Hydraulic Engineering, 2009.

X. Blasco, J. M. Herrero, J. Sanchis, and M. Martinez: A new graphical

visualization of n-dimensional pareto front for decision-making in multi-objective
optimization. Information Sciences, 178(20):3908 - 3924, 2008. ISSN 0020-
0255. DOI: 10.1016/j.ins. 2008.06.010.

H.-I. Eum and S. P. Simonovic: Integrated reservoir management system for

adaptation to climate change: The Nakdong River basin in Korea. Water
Resources Management, 24:3397-3417, 2010. ISSN 0920-4741. DOI:
10.1007/s11269-010-9612-1.

T. Hashimoto, J. R. Stedinger, and D. P. Loucks: Reliability, resiliency, and

vulnerability criteria for water resource system performance evaluation. Water
Resour. Res., 1:14-20, 1982. DOI: 10.1029/WR018i001p00014.
Hydrologics Inc.: User manual for OASIS with OCL, 2009. URL
http://www.hydrologics.net/documents/OASIS_Manual4-2010.pdf.
C. Igel, N. Hansen, and S. Roth: Covariance matrix adaptation for multi-
objective optimization. Massachusetts Institute of Technology, Evolutionary
Computation, 15(1):1–28, 2007.

D. Koutsoyiannis and A. Economou: Evaluation of the parameterization-

simulation-optimization approach for the control of reservoir systems. Water
Resources Research, 39(6):1170-1187, 2003. ISSN 1944-7973. DOI:
10.1029/2003WR002148.

F. Kreienkamp, A. Spekat, and W. Enke: Ergebnisse eines regionalen

Szenarienlaufs für Deutschland mit dem statistischen Modell WETTREG-2010.
Technical report, Climate & Environment Consulting Potsdam GmbH. Bericht an
Umweltbundesamt, 2010.

A. Langousis and D. Koutsoyiannis: A stochastic methodology for generation of

seasonal time series reproducing over-year scaling behavior. Journal of
Hydrology, 322:138-154, 2006.

R. Müller and N. Schütze: Multi-objective optimization of multi-purpose multi-

reservoir systems under high reliability constraints. Environmental Earth
Sciences, 75:1278, 2016. DOI: 10.1007/s12665-016-6076-5.

R. Müller: Eine neue Strategie zur multikriteriellen simulationsbasierten

Bewirtschaftungsoptimierung von Mehrzweck-Talsperrenverbundsystemen.
PhD thesis, Technische Universität Dresden, Fakultät Umweltwissenschaften.
2014. http://nbn-resolving.de/urn:nbn:de:bsz:14-qucosa-160659.
A2 - SEASONAL SCALE:
SEASONAL (MID-RANGE) FORECASTS
FOR RESERVOIR OPERATION ADAPTATION
TO MITIGATE SHIFTING PRECIPITATION PATTERNS – GERMANY

Dr. Hubert Lohr, SYDRO Consult GmbH, Germany

1) Introduction
In recent years, changes in the precipitation patterns have been observed in
Germany. Rainfall in February to April shifted more into the summer. Although
total annual rainfall remained at nearly the same level, the resulting discharge
decreased due to higher evaporation losses in the summer months. As a
consequence, reservoir operators experience difficulties in reaching full supply
levels in spring, which jeopardizes existing and competing demands and
requirements over the course of the remaining year. This also impacts local
communities and water suppliers. In addition, flood storage needs to be operated
in a dynamic way, as the actual precipitation pattern might differ from the patterns
used for designing flood storage volumes.

Thus, a preferably early recognition of droughts (or wet conditions) in order to

plan and implement operational counter-measures is the objective of seasonal
forecasting and its application for adaptive reservoir operation. The presented
approach uses hydro-meteorological indices based on observations from ground
stations and seasonal forecasts of precipitation for up to 9 months in order to
identify the need for operational counter-actions at a preferably early stage.
Seasonal circulation forecasts of NOAA are used.

The approach is being tested for five reservoir systems with a total of ten
multipurpose reservoirs in Germany. The first investigations started in 2013
together with the Water Board Eifel-Rur (North Rhine-Westphalia). The
methodology is not only interesting for reservoir operators but also e.g. for mining
companies which run complex wastewater-discharge schemes. For instance, the
methodology is currently in the first stage of implementation in Hesse in the
context of mining water management.
Figure A2.1: Investigation area and precipitation ground stations
2) Database
Records of precipitation, temperature, evaporation and current reservoir
conditions (e.g., water levels) are required. In operational mode, the weather
service and the reservoir monitoring system must provide these data on a regular
basis at predefined intervals.

Seasonal precipitation forecast data is downloaded on a daily basis from NOAA’s

NCEP coupled forecast system model version 2 (CFSv2) 13. Since May 2011,
NOAA issues seasonal forecasts, consisting of monthly values. Forecasting data
is available in a grid format with a resolution of 0.9 degrees. The CFSv2 model
covers the whole world and is updated four times a day.

3) Prerequisites

As a first step, a bias correction is performed. Therefore, model output is

compared against ground station data and monthly correction factors are derived
on the basis of hindcast analyses, starting with data from 2011. First, average
rainfall totals for each month were calculated for observation and forecast data.
Then, a monthly correction factor was calculated by computing the ratio between
the observed and the forecasted values. The correction factors are specific for
each ground station. Bias corrected forecasts could then be obtained
operationally by multiplying each value of the forecasted time series with the
appropriate monthly correction factor.

4) Methodology

As indices have differing inertia and apply to different periods, they can be used
for predictive operation. The appropriateness of these indices, the way they
should be interpreted and their usefulness regarding early detection of
hydrological stress, is tested by conducting hindcast experiments. Indices
providing the best skill are selected for conducting forecasts.
For a start, the Standardized Precipitation Index (SPI) was used at all locations.
The SPI is recommended by the World Meteorological Organization for
meteorological drought monitoring. The SPI can be calculated for different
aggregation periods, e.g. only one month or even up to 60 months.

In order to address uncertainty contained in the forecasts, the SPI is calculated

for time periods that extend both into the past as well as into the future, thus
consisting of different amounts of observed and forecasted values. The
performance of indices calculated with different observed and forecasted
aggregation periods was compared with results that used only observed values
for computing SPIs. In doing so, it is possible to determine how reliable the SPI
incorporating forecast data is for different forecast lengths.

13 http://www.cpc.ncep.noaa.gov/products/CFSv2/CFSv2seasonal.shtml
Figure A2.2: Use of indices without forecasts

Figure A2.3: Use of indices with seasonal circulation forecasts from NOAA
Calculation of Indices

Based on observed and bias-corrected forecast data, the SPI was calculated for
different aggregation periods. The SPI calculated using only observed data
("certain knowledge") was compared with the SPI obtained by considering
different fractions of observed records and NOAA forecasts.

Fig 4. shows results for a 12-month aggregation period for three ground stations
in eastern Germany. The SPI calculated using NOAA forecasts reveals a good fit
in comparison to the SPI based on measured data and more significantly exhibits
the same tendency for upcoming dry periods.
Figure A2.4: SPI calculated with observed / NOAA forecast data, examples from
different sites
5) Adaptation of reservoir operation
Once the SPI (or indices in general) is calculated based on combined past and
forecasted values, it can be used to adjust operation of reservoirs. Adjustment
cannot be established in a general way but requires site-specific rules. However,
what can be generalized, is the way in which the results of indices are introduced.
First, site-specific operation rules best suited for intervention must be identified.
Second, threshold values for indices must be identified for when intervention
should be triggered. Third, the aggregation period is highly dependent on the local
situation and needs to be determined for each single individual case.

Reservoirs in West-Germany with catchment areas ranging from 250 to 600 km²,
operated with threshold-based rules, could be improved by using a 9 to 12 month
aggregation period and a threshold value of -1.5 (SPI + Evaporation based
indices). When the index dropped below -1.5, release rules associated with the
next lower storage-threshold level were applied to counter an expected decrease
of inflow. This meant a reduction of downstream releases depending on the time
of year and current storage volume. Not surprisingly, not all critical periods could
be identified by applying this approach. However, two thirds of critical low flow
conditions with corresponding low water levels in the reservoirs could be tackled
in a timely way. The approach was used without forecasts prior to 2011 and as
of 2011 with forecasts to make full use of the observation period with more than
100 years for evaluation.

The figure below indicates the volume of two reservoirs with (blue line) and
without (red line) correcting intervention by means of (here) SPI. All yellow
sections show SPI less than the threshold value for intervention and with
reduction of releases. In other words, during the yellow periods, releases are
reduced in order to avoid a further drop of the water level.

Figure A2.5: Storage volume with/without SPI-based hedging rule

(blue=with, red=without)
In Central Germany, a reservoir with a catchment area of less than 50 km²
revealed a different pattern. Only aggregation periods longer than 18 months with
a threshold value of minus 1.0 showed good results. Shorter aggregation periods
or lower threshold values were either not consistent enough or started too late to
result in counter measures that took effect. In this case, the target operation rule
for intervention was water supply provision. Similar to a hedging rule, water
provision was subjected to a quota of a rather small percentage as soon as the
index dropped below -1.0 to prevent larger reductions later on. In doing so, the
reservoir could be kept above a water level that becomes critical from the
viewpoint of water quality.
The initial assumption that the size of a catchment area is a reasonable
parameter in order to pre-estimate aggregation periods could not be confirmed.
The interplay between climate, the catchment’s geology as an indicator for
“inertia” (or “hydrological memory”) and the reservoir itself seems more complex.
As a result, each reservoir or reservoir system must be individually scrutinized to
find the best set of aggregation periods and threshold values.

This project is funded by the German Federal Ministry for the Environment,
Nature Conservation and Nuclear Safety, BMUB.

A comparable method, based on direct utilisation of forecasted precipitation

instead of hydro-meteorological indices, was tested for the Lower Mekong Basin
in 2016 by the Author by order of the Mekong River Commission and GIZ. Results
revealed a higher degree of uncertainty and confirm the assumption that indices
are more robust.

6) References
NATIONAL OCEANIC AND ATMOSPHERIC ADMINSTRATION (NOAA)
https://www.ncdc.noaa.gov/data-access/model-data/model-datasets/climate-
forecast-system-version2-cfsv2#CFSv2%20Operational%20Forecasts, 2016.
SVOBODA, M. & FUCHS, BRIAN
Integrated Drought Management Programme (IDMP), Handbook of Drought
Indicators and Indices. Drought Mitigation Center Faculty Publications, 117, 2016.

GUDMUNDSSON, L.; BREMNES, J. B.; HAUGEN, J. E. & ENGEN-SKAUGEN,

T. Technical Note: Down-scaling RCM precipitation to the station scale using
statistical transformations - a comparison of methods. Hydrology and Earth
System Sciences, 2012, 16, 3383-3390, doi:10.5194/hess-16-3383-2012.

MRC 2016: Long-range streamflow forecasts for the LMB. Mekong River
Commission, Phnom Penh, 2016. Unpublished report.
A3 - SHORT-TERM SCALE
USE OF RESERVOIR STORAGE FOR FLOOD OPERATION IN THE
KUMANO RIVER BASIN – JAPAN

Masayuki Kashiwayanagi
Electric Power Development Co., Ltd. Chigasaki, Japan

1- Background
Kumano river is located in Kii peninsula at the middle part of Japan along the
Pacific Ocean. Its basin spreads almost all in mountain area, namely 97 %, which
results in the steep slope of the river bed and much precipitation about 3000 mm
per annum. It has (the comment is illegible.) 2360 km2 drainage area including
18 municipalities with a population of 84000. Many typhoons periodically pass the
basin and occasionally bring disaster due to much precipitation. As an example,
Typhoon Isewan (category 5) caused huge disaster in the middle part of Japan
and 5 fatalities and 2300 inundated houses in the Kumano river basin in 1959.

There are 11 dams for hydropower and irrigation, not for flood protection, in the
basin as shown in Figure A3.1. J-Power has been operating 6 hydropower dams
in the basin and has contributed to flood mitigation by providing flood storage in
reservoirs during the rainy season. Taking the opportunity of the last disaster due
to the typhoon No.12 in 2011, the corporative action among the governmental
river administrator and J-Power has been commenced for enhancing the flood
mitigation using the hydropower dams in the Kumano river basin. Since then
J-Power has explored and verified more effective flood mitigation using
hydropower dams within the limitation of the commercial operation of the
hydropower plants. For this purpose, the study has been conducted on the validity
of released meteorological information for the proactive reservoir operation
during floods.
Name of dams Owner of Dams
①Futatsuno ②Kazeya ③Sarutani
④Komori ⑤Nanairo ⑥Ikehara J-Power
Japan
⑦Sakamoto ⑧Asahi ⑨Seto Others
⑩Tsudurao ⑪Kouse ⑪
⑩
③
⑨
⑦
⑧
⑥
②

④
⑤
①
Kii
peninsula
Pacific Kumano river basin
Kumano River ocean
183 km, 2360 km2
0 20km

(a) Plan

altitude Sarutani dam

m
Total reservoir capacity
500 23,300 kilo m3 Sakamoto dam
H.W.L.=EL.436.00m
87,000 kilo m3
H.W.L.=EL.387.50m
400 Ikehara dam
Kazeya dam
338,373 kilo m3
130,000 kilo m3
H.W.L.=EL.318.00m
Water diversion

H.W.L.=EL.295.00m
300
Nanairo dam
61,300 kilo m3
200 H.W.L.=EL.190.00m

Kurobuchi
H.W.L.=EL.186.50m
100
Futatsuno dam Komori dam Kuchisubo
43,000 kilo m3 9,700 kilo m3 H.W.L.=EL.137.00m
H.W.L.=EL.132.50m H.W.L.=EL.118.00m
0
Kumano river

(b) Profile

Figure A3.1 Dams in Kumano river basin

2- Flood risk mitigation using existing hydropower dams
The relatively large reservoirs of Ikehara dam (1964，arch dam, 111 m high)
and the Kazeya dam (1960, concrete gravity, 101 m high) are studied for the flood
mitigation. The reservoirs are 338 MCM and 130 MCM in volume, respectively.
The dams have gated spillway with the maximum discharge of 6700 and 5200
m3/sec, respectively. The Ikehara dam and its spillway arranged apart from the
dam are shown in Figure A3.2.

Gated Spillway

Ikehara dam
(H111m) Reservoir G
S

(a) Overview (b) Ikehara dam

Figure A3.2 Ikehara dam and its reservoir

The flood mitigation method for the Ikehara and the Kazeya dams will be
conducted by following manner. When the flood caused by much precipitation in
the basin is expected, the power generation is conducted in a proactive manner
to draw down the reservoir water level to the aim water level from the usual
operation water level for providing the reservoir flood storage, as shown in
Figure A3.3.

Until the discharge does not exceed the hazardous flood discharge, the spillway
gates of the dam are operated to release the flood so as to maintain the aim water
level. As increasing the flood discharge more, the gate operation follows the
delayed operation rule, of which concept is illustrated in Figure A3.4. To release
the flood beyond the hazardous flood discharge, the gate is operated to spill the
discharge as much as the flood discharges at specified hours before. The
specified hours are referred to as a delayed time. The similar gate operation is
continuing until the flood discharge reaches the peak value. Then the gate
operation is interrupted until the spilled discharge is the same as much as the
flood discharge. As decreasing the flood discharge, the gates are operated so as
to spill the discharge equally as much as the flood discharge.

The gate operations for the flood mitigation of Ikehara and Kazeya dams are
enhanced by providing increased flood storage of 70 MCM and 9 m water depth,
and 28 MCM and 7 m water depth, of which water depths were originally 6 m for
both. In Ikehara dam, two options of the aim water level are designated
depending on the precipitation intensity. In addition the modified delayed
operation rule adopts 3 hours to the delayed time instead of original 0.5 hour,
taking the increased reservoir volume above mentioned into consideration. The
modified operation may result the higher reservoir water level, but has to be less
than the high water level of the reservoir.

Volume for Volume for flood mitigation

flood mitigation based on flood forecast

H.W.L.
Flood Ikehara dam Volume
control Aim W.L. （kilo m3)
H.W.L.(35.0 m)
gate Temporary aim W.L.①
48,000 Volume by
Temporary aim W.L.② Aim W.L.(29.0 m) present aim W.L.
Effective Storage Temporary 11,000
Volume by
aim W.L.①(27.5 m) Temporary aim W.L.
L.W.L. Temporary 11,000
aim W.L.②(26.0 m)
total 70,000
Dead Storage

Kazeya dam Volume

（kilo m3)
H.W.L.(30.0 m)
Sediment Volume by
Storege 24,000
present aim W.L.
Aim W.L.(24.0 m)
Volume by
Temporary 4,000
Temporary aim W.L.
※This is conceptual figure and the scale and aim W.L.(23.0 m)
aspect ratio is different from actual dam total 28,000

Note: The water levels are measured from L.W.L.

Left figure is for Ikehara dams.

Figure A3.3 Flood mitigation scheme

Water level
Maximum water level
(in case of 3 hr delayed)

Up

Maintaining
the water level

Discahrge Outflow by delayed

operation
Peak discahrge 0.5 hr
Inflow 3 hr

3 hr Down
Hazardous 0.5 hr
flood discharge Sustained

Note: Hazardous flood discharge is defined in each dam independently which possibly cause the
harmful impact on the downstream area of the dam.

Figure A3.4 Modification of reservoir operation rule

3- Forecast for proactive flood risk mitigation

Reliable meteorological forecasting in a few days advance and its continuous
updating are essential for the proactive flood risk mitigation by the reservoir
operation following the delayed operation rule above mentioned. The key
information is considered as the real-time location and the prediction of the path
of the typhoon and the precipitation in the basin, because almost floods have
been caused by the typhoons in the Kumano river basin.

Analyzing historical typhoons causing heavy floods in the basin, the typhoons
have traveled on the similar paths in the confined area and approached the basin
within the 300 km radius, as shown in Figure A3.5. The typhoon information is
available in the JMA’s （Japan Meteorological Agency）web site which releases
the real-time location of the typhoon, its characteristics and predicted path in a
few days ahead. The precipitation information by JMA is also available as GPV
(Grid Point Value, Figure A3.6) predicted by GSM（Global Spectral Model）.
Validating the correlation between the predicted precipitation and the actual ones,
84-hour cumulative precipitation based on the maximum predicted GPV at each
20 km grid in the drainage area of these dams can provide higher correlation to
the observed ones. Such prediction is beneficial to provide sufficient time for
drawing down the reservoir water level by the generating operation which utilizes
the reservoir water effectively.

Reviewing the historical floods again, predicted 84-hour cumulative

precipitation above 200 mm caused floods exceeding the specified hazardous
discharge of 1500 m3/s at the dam and triggered delayed operation. Ones more
than 500 mm brought severe floods. These can be summarized as the criteria for
the proactive draw-down operation of the reservoir for the flood mitigation in
Table A3.1.

Location of Kazeya dam and

Ikehara dam

● 参GPV used
照グリッド

Course Upper river basin Upper river bas

from Futatsuno dam from Komori da

Location
Lower river basin
from the dam
: A 300 km radius from the dam
: North of Lat.15°N and Long, 120~145 20km

上記修正
Figure A3.5 Criteria for typhoons Figure A3.6 GPV locations for the
as shown in Table A3.1. forecast of precipitation in the basin

Table A3.1 Criteria for proactive drawdown operation for flood mitigation

Criteria Initiation of Information Data update

drawdown
Location of Shown in Real time data Real time
Typhoon Figure A3.5
Predicted course Within 300 km radius Forecast in 3 to 5 3- to 5-hour
of Typhoon days ahead by JMA1) interval
Predicted rainfall Cumulative 84-hour forecast 6-hour interval
precipitation in 84 (GSM2)) of GPV3) by
hours JMA
200 mm (Severe)
500 mm (Extreme)
1) Japan Meteorological Agency, 2) Global Spectral Model, 3) Grid Point Value
4- Verification
The modified reservoir operation and proactive draw-down criteria shown in
Figures A3.3 and A3.4 and Table A3.1 are validated in terms of the efficiency of
the flood mitigation by the examination of the historical floods.

Firstly, the relation between the draw-down criteria shown in Table A3.1 and
the observed discharge at the dam is examined on the occasion of the historical
329 typhoons passed through the middle of Japan. When a typhoon meeting all
criteria would cause the flood exceeding the specified hazardous flood
discharge at the dam, it will be the right case. Otherwise will be the wrong case.
The results are summarized in Table A3.2. Half cases which satisfy the criteria
and almost all cases which fail the criteria are identified as the right cases. It
clarifies that the criteria are practical ones in order to identify the necessity of
the draw-down of the reservoir.

Table A3.2 Verification of the criteria of drawdown operation

Number of typhoons
Evaluation Above the hazardous flood Below the hazardous flood
discharge (1500 m3/s) discharge (1500 m3/s)
Satisfy the criteria 16 (Right) 14 (Wrong)
Fail the criteria 1 (Wrong) 298 (Right)

Secondary, the effectiveness of the delayed operation is examined by the

simulation of the operation in the Ikehara reservoir (refer to Figure A3.3) under
the following assumption.

1) Initial reservoir level is 29 m as the water level in the rainy season.

2) When the situation of the typhoon satisfy the criteria shown in Table A3.1,
the reservoir water level is proactively draw down by the generation with the
discharge of 342 m3/s.

3) The flood is controlled firstly by the gate operation to maintain the reservoir
water level.

4) As increasing the discharge, once the discharge at the dam exceeds the
hazardous discharge of 1500 m3/s, the gate operation following the delayed
operation rule is commenced to regulate the flood.

The simulated results are shown in solid lines in Figure A3.7. The cumulative
precipitation in an 84-hour specified as the criteria are 200 mm and 500 mm for
Case A and Case B, respectively. The moment of satisfying the draw-down
criteria and its initiation is shown in vertical dotted lines. These show that the
sufficient times are ensured for the reservoir draw-down to the specified water
level in both cases. Due to the flood characteristic of shorter flooding duration
less than the delayed time of 3 hours in Case A, the maximum discharge at the
dam does not exceed the hazardous discharge of 1500 m3/s. Contrary, one for
Case B results 1708 m3/s by the delayed operation which follows the increasing
flood discharge with specified delayed time. Both are less than the maximum of
the natural flood discharge. The reservoir levels in both cases stay under the
specified one. In the actual figure in Case B, the actual discharge was a few of
326 m3/s because the initial water level in low of 10 m and the gate operation was
withheld by utilizing the vacant reservoir volume.
5- Conclusions
The flood risk mitigation using the hydropower dams by the proactive reservoir
draw-down and the modified reservoir operation method by effective use of the
resultant vacant reservoir volume are examined. The following conclusions are
obtained.

1) The meteorological information of the location and the path of the typhoon
and the precipitation prediction using GSM released by JMA are effective index
for the proactive reservoir draw-down for the flood mitigation in the Kumano River
basin. These characteristic of easy access and frequent updating are adequate
for the reservoir operation criteria.

2) The specified criteria for the proactive reservoir draw-down provides

sufficient time for reservoir draw-down using the generation discharge.

3) The modified reservoir operation method comprising the delayed operation

concept are verified to be practical ones for the studied dam by the flood routing
simulation on historical typhoon cases.

The studied method has been actually applied for the reservoir operation since
2012 at the basin. Further validation will be conducted based on the
consequences.
0 0

Total(mm)
(mm/hr)
Rainfall
20 200
40 Average rainfall over area 400
60 Total rainfall 600
80 800
6,000 30
▼29.00m ▼28.76m
▼27.50m Calculated WL 28

9/28 18:54 fall under the criteria

9/28 6:54 fall under the criteria
5,000 26
Actual WL

9/28 12:53 meet the criteria

Quantity（m3/s-H)

24
9/28 0:52 meet the criteria

9/28 0:52 meet the criteria

4,000 22
Actual inflow

Water level（ｍ）
Max 2,708 m3/s 20
3,000 18
16
Calculated discharge
2,000 14
Max 1,500 m3/s
12
Actual discharge
1,000 10
Lead time Max 1,442 m3/s
8
2.5days
0 6
9/27 9/28 9/29 9/30 10/1 10/2
00:00

00:00

00:00

00:00

00:00

00:00

00:00
(a) Case A: Typhoon No.17 in 2013
0 0

Total(mm)
(mm/hr)
Rainfall

20 200
40 Average rainfall over area 400
60 Total rainfall 600
80 800
6,000 ▼29.08m 30
▼29.00m
▼27.50m 28
5,000
Calculated WL 26
9/13 12:56 meet the criteria
Quantity（m3/s-H)

Actual WL 24
4,000 22
Water level（ｍ）

20
Actual inflow
3,000 Max 2,273 m3/s 18
Calculated 16
discharge
2,000 Max 1,708 m3/s 14
Actual 12
discharge
1,000 10
Lead time Max 326 m3/s
2days 8
0 6
00:00

00:00

00:00

00:00

00:00

00:00

00:00

9/12 9/13 9/14 9/15 9/16 9/17

(b) Case B: Typhoon No.18 in 2013

Figure A3.7 Simulation of the Ikehara reservoir operation

6 - Reference
(1) Takakura H., Matsubara T., Nakakita E., Takada N.: Study on validity of the
hydroelectric dam operation adopted GSM and the information about typhoons,
85th ICOLD annual meeting, Prague, Czech Republic, 2017.6

(2) Matsubara T., Kasahara S., Shimada Y., Nakakita E., Tsuchida K., Takada
N.: Study on applicability of information of typhoons and GSM (Global spectral
model) for dam operation), Journal of Japan Society of Civil Engineers, Sr. B1
(Hydraulic Engineering), Volume 69(4), pp. I-367-372, 2013(in Japanese)