Assimilation of regional runoff and stage, data or remotely sensed
soil and ground wetness
Error correction
CHAPTER 4. CHOICE OF APPROPRIATE METHODS OR MODELS FOR FLOOD FORECASTING 4-19
exists for the use of satellite remote-sensing for updating soil-moisture fields. The overpass interval of suitable radars is of the order of days, but this may suffice for a slowly developing situation on a large catchment.
4.7.3 Global hydrological forecasts based
on mesoscale models and nowcasting
Figure 4.3 shows a version of a GHF based on the United Kingdom Met Office or similar mesoscale model. Typically, a mesoscale model provides hourly forecasts of rain updated every six hours.
Mesoscale models are computationally complex and rely to an important extent on extensive data assimilation. As a result, forecasts may not become available until up to three hours after the time of observations. Despite this it is considered that a 36- to 48-hour addition to forecast lead time has real value for more rapidly responding catchments of greater than 2 500 square kilometres, subject, of course, to achieving adequate accuracy.
One example of a near-operational GHF-like system based on a mesoscale model is from South Island, New Zealand, which uses a Regional Atmospheric Modelling System (RAMS) for QPF with a version of TOPMODEL to convert to flow. RAMS has a 20-kilometre resolution and is driven by the United Kingdom Met Office 120-kilometre forecasts.
Forecast data are available from Bracknell at 8 p.m.
local time, and with overnight running of RAMS a hydrological forecast is available at 8 a.m., provid- ing a 27-hour lead time of events 48 hours from initial receipt. Conceptually, TOPMODEL empha- sizes the infiltration process over channel routing in headwaters, so it should be well suited to the more mountainous catchments involved. Early results showed that very large under- and overesti- mates of flood magnitude occurred when the model was allowed to run free. This illustrates the neces- sity for updating with observed rainfall and river flow data.
A modified version of the mesoscale model can be constructed, for example supplementing GHF-2 with a Nowcast system. This acts as a further assimi- lation process, in which a combined radar–NWP model with a forecast lead time out to six hours can be introduced.
4.8 PREDICTIVE UNCERTAINTY IN OPERATION
When dealing with flood emergency management, operational decisions may lead to dramatic conse- quences (economical losses and casualties).
Nonetheless, emergency managers are required to take decisions under the stress of their uncertainty about the evolution of future events. Decision theory has developed into an extensive topic of mathematical study.
One of the issues in the debate among hydrologists is how to demonstrate the benefits arising from the operational use of predictive uncertainty. The corol- lary of this is how to communicate uncertainty to the end-users, namely the decision makers such as water and emergency managers, who may have a certain difficulty in perceiving these benefits.
Statements such as “the probability of flooding within the next 12 hours is 67.5 per cent” is often meaningless to an end-user. The information has to answer the basic question “what are the expected benefits and drawbacks of issuing a flood alert for the next 12 hours?”. Therefore, hydrologists must define, in dialogue with end-users, subjective utility functions, which can be used to compute the expected benefits or damages contingent on the predictive density of the quantity of interest.
A schematic example of such utility functions is shown in Figure 4.4, for the case of a flood alert (note that in this simple schematic example, casu- alties are not taken into account). The dashed line represents the end-user perception of the damage (not necessarily the real one) that will occur if the dykes are overtopped, namely if Q > Q*, where Q*
is the maximum discharge that may safely flow in the river. The solid line represents the perception of cost plus damages when an alert has been issued. As can be seen from Figure 4.4, if an alert is issued a cost must inevitably be incurred for mobi- lizing civil protection agents, alerting the population, laying sandbags and taking other necessary measures. However, the damage in that
Figure 4.4. The utility functions deriving from a flood alert problem: solid line represents cost and
damage perceived by the end-user if an alert is issued; dashed line represents perceived cost and damage if an alert is not issued; Q* is the maximum
discharge that may safely flow in the river
Q*
Discharge
Damages
Q*
Q*
Q*
Discharge Q*
No alert Alarm
case will be smaller than in the “no-alert” case, due to the raised awareness of the incoming flood.
The decision on whether or not to issue an alert will then depend on the comparison of the
“expected damage” for the two options, obtained by integrating the product of the cost function multiplied by the predictive uncertainty probabil- ity density function over all possible values of future discharge. It should be noted that the
“expected damages” are a function of the actual future discharge that will happen, not of the discharge predicted by the model. By using the expected value of damage instead of the “model forecast”, the probability of false alarms as well as of missed alarms should be much reduced, as the uncertainty about the future discharge is taken into account. In addition, the peakier the predic- tive density is, the more reliable will the resulting decision be, so that improvements in forecasting, rather than looking for a better “deterministic”
forecast, must essentially aim at reducing predic- tive uncertainty by whatever means is available.
To show how predictive uncertainty can be used in operation, the Lake Como real-time management decision support system is given as one of the few existing successful examples (Todini and
Bongioannini Cerlini, 1999). Lake Como is a natu- ral lake in northern Italy closed at its exit and managed as a multi-purpose lake for flood control, irrigation and electric power production. Using a stochastic dynamic programming approach, a standard operating rule was developed on a 10 day basis to optimize long-term irrigation and energy production. However, when a flood is forecast, the reservoir manager needs to modify the standard operating rule. To achieve this, a utility function describing the damage perception of the manager was developed. Every morning an incoming flood forecast, together with its predictive uncertainty, is issued, and an optimal release, computed by mini- mizing the expected damage using the inflow predictive uncertainty, is then proposed. All this process is totally hidden from the water manager, who is aware only of the suggested optimal release and of its expected consequences (Figure 4.5).
The performance of the system was assessed on the basis of a hindcast simulation for the 15-year period from 1 January 1981 to 31 December 1995.
The results are presented in the table below. When applying the optimized rule, the lake level never fell below the acceptable lower limit of –0.4 metres, while historically this was observed on 214 days.
Figure 4.5. The Lake Como operational decision support system. The system, on the basis of the expected value of inflows to the lake (light blue line) and its uncertainty (not shown, but used in the process) suggests to the water manager the optimal (red line – not shown) and possible (green line) releases that
minimize the expected damage. It also shows the consequent expected lake level (blue line) for the following 10 days.
Forecasting horizon:
Inflow [m3 s-1]
Vol [m3 x 106] Optimal releases Possible releases
Levels at Malgrate [cm]
Forecasting horizon:
Inflow[m3 s-1]
Vol [m3 x 106] Optimal releases Possible releases
Levels at Malgrate [cm]
CHAPTER 4. CHOICE OF APPROPRIATE METHODS OR MODELS FOR FLOOD FORECASTING 4-21
Summary of the results of a comparison between recorded water level occurrences and water
deficits (historical) and the results of the operation rule based on the forecasting uncertainty (optimized) for the 15-year period
from 1 January 1981 to 31 December 1995
implies higher lake levels, an objective conflicting with the need to reduce the frequency of flooding.
It is quite interesting how the system was accepted by the end-user. At the end of 1997, the system was installed operationally and the Director of Consorzio dell’Adda, who is in charge of lake management, was invited to look at it but not to use it until he had confidence in its effectiveness.
After six months the Director admitted that he had made a wrong decision on all of four occasions when the decision support system (DSS) had provided a solution. Ever since, the system has been in operation and used successfully. It has produced not only a reduction in the number, frequency and magnitude of Como flooding events, but also a 3 per cent increase in energy production and a large volume of extra water for irrigation.
The above example shows that, if appropriately briefed and involved, the end-users will quickly become aware of the benefits arising from the use of predictive uncertainty, provided they are not asked to interpret the forecasting in statistical terms or the stochastic computation and optimization frequently required in problems in this type.
Considerable effort is still required to inform the end-users of the improvements obtainable without burdening them with the computational complex- ity. In this way, they will appreciate and receive the full benefits of an approach aimed at improving the success of their decision-making.
Water level Number of days
Historical Optimized
–40 cm 214 0
120 cm 133 75
140 cm 71 52
173 cm 35 34
Water deficit: 890.27 × 106 m3 788.59 × 104 m3
In terms of Como flooding, over the 15 years the lake level has been recorded to be above the lower flood limit of 1.2 metres on 133 days, whereas the optimized rule reduced it to 75 days. A noticeable reduction also appears at higher lake levels. At 1.4 metres, when the traffic must stop in the main square of Como, the reduction is from 71 to 52 days and at 1.73 metres, the legal definition of
“normal flood” when people can claim compensa- tion for their damage, the reduction is from 35 to 34 days. At the same time, the irrigation water deficit decreased by an average of more than 100 × 106 cubic metres per year. This result is exceptional, given that meeting irrigation demand
5.1 DEFINITION OF DATA ACQUISITION NETWORKS
There are many factors that should be considered when designing an operational network to support a flood forecasting and warning operation.
Essentially the network is based on a combination of rainfall and river level (and possibly flow) moni- toring points, reporting in real or near-real time to a central operation and control system. As the output is of high value to the national interest, that is, it concerns timely, detailed flood warnings with maxi- mum accuracy, the networks require high reliability and resilience and are largely built around auto- matic data monitoring, processing and retrieval.
Although some monitoring facilities may be shared with other uses within the owning agencies, these often being water management and meteorological services, the components of the flood forecasting monitoring system have to be considered and oper- ated as a single entity. Thus, within the individual agencies responsible, a matching high level of criti- cal infrastructure support must be assigned. This entails devoting adequate funding both to purchase and maintain the network, and to ensure suitable levels of staffing to maintain the functions. In addi- tion to the need for equipment to operate round the clock, it is also important that staffing allows for periods of emergency operation, so the provision of cover, duty assignment and additional staff call-out have to be considered.
Most importantly, the data network must deliver information in relation to areas where high risk of flooding combines with high impact of flooding.
Thus, there must be a sufficient number of stations reporting the detail that will allow the develop- ment of the flood to be observed and to provide sufficient time for forecasting models to run and produce outputs for timely warnings to be issued and necessary decisions taken (see Box 5.1).
The sections that follow in this chapter will exam- ine the requirements for a successful network,
whilst Chapters 6 and 7 consider in more detail some of the technical issues. This chapter will also examine instrumentation requirements in general, but does not aim to include detailed specification of individual instruments.
5.2 EVALUATING EXISTING NETWORKS Networks used for flood forecasting and warning systems are most frequently based on existing hydrometeorological networks. This may to a greater or lesser extent influence the structure of the required network to economize on the intro- duction of new sites. The use of existing networks has the benefit that, as well as equipment and site infrastructure, the locations used will have an established database to provide a good foundation of information for model development. The main issues to be addressed when evaluating existing networks concern their geographical suitability for flood forecasting purposes. As existing networks may have been developed by separate operational entities, for example a water resources or meteoro- logical organization, there may also be ownership and management matters to consider.
As mentioned in 5.1, the most important require- ment is for the network to deliver useful and timely information. Thus, although it may be convenient to use existing networks from an economic or oper- ational point of view, their suitability for the end purpose in flood warning must be carefully consid- ered. Some of the most important considerations are explained in this section, but details of the design of a network are given in 5.3.
5.2.1 Meteorological networks
Observation networks operated by NMSs have generally evolved from two main historical require- ments. Synoptic observations of a range of variables have been collected for many years to understand current weather conditions and develop knowledge and methods for weather forecasts. Data from synoptic stations and other “primary” stations are intended to build up a picture of general or specific climatology, for example agroclimatology. For both of these types of network, the principal require- ment is to sample conditions across a region or country with sufficient detail to characterize areal and temporal variations. Data recording is generally on a daily basis, with principal synoptic