Environmental Engineering Reference
In-Depth Information
Fig. 10.1
Mean error associated
with the forecasting of the amount
of rainfall using radar, satellite,
Numerical Weather Prediction
(NWP) techniques and a hybrid
scheme. From Smith and
Austin (2000).
depends on many factors, including the degree of
understanding of the physical characteristics of
the watershed, the quality of the hydrological
model, and the availability and accuracy of the
information feeding the flood forecasting system.
The single most important input is rainfall. The
ability of the system to update predictions based
on observed streamflow and revised precipitation
estimates is also very important.
When the primary drivers of uncertainty are
known, a probability distribution of flood flow
from ensemble output can be developed. In this
technique, variables affecting streamflow, includ-
ing rainfall, are adjusted within a range of possi-
bilities and the resulting hydrographs are then
analysed statistically. This Monte Carlo approach
has been used by hydrologists formany years. Both
nowcasting and NWP methods offer the possibil-
ity of developing ensemble rainfall inputs for run-
off simulation. Nowcasting algorithms can be
developed to produce a realistic range of future
storm direction, speed and growth. Similarly,
NWP models are highly dependent on initial con-
ditions andmodel parameterization. Small pertur-
bations in these values can be used to develop a
realistic range of forecasts. Combining the now-
casting ensembles with the NWP ensembles pro-
duces a set of QPFs that capture the range of
uncertainty in the precipitation forecast. Routing
this ensemble of QPFs through a rainfall-runoff
model results in a probability distribution of fu-
ture hydrological conditions, which is very useful
for emergency management. The concept
is
shown schematically in Figure 10.2.
Developing numerous simulations of an al-
ready computationally expensive coupled meteo-
rological-hydrological model is not realistic in
many situations. Simplifying the NWP or rain-
fall-runoff model to reduce the CPU burden is an
often successful but hardly intellectually pleasing
approach. More often, researchers examine the
probability distribution of Monte Carlo-generated
QPF outputs and select individual rainfall fore-
casts that adequately represent this distribution.
This can reduce the number of simulations (and
hence computational time) by an order of magni-
tude. The process of developing a new QPF en-
semble and routing it through the flood forecasting
model is repeated as new observational data be-
come available. The number of QPFs in the
