Geoscience Reference
In-Depth Information
Realisations of 10 000 years of hourly data were then used to get good estimates of the peak discharge
with an annual exceedance probability of 0.001 (1:1000 year event) for the dam safety evaluation. One
interesting feature of this methodology is that it is possible to investigate which of the limits of acceptabil-
ity causes most model rejections. In this case, it was the underprediction of low-magnitude floods at the
Arzberg site.
The best model in satisfying the limits was also run with 10 000 realisations of input data generated for
the same period of 67 years. Less than 1% of those runs survived the limits of acceptability. This suggested
that the initial limits might be relaxed to avoid rejecting models that might be useful in prediction over
longer runs simply because of this
realisation effect
. The grey frequency curves in Figure 8.5 represent
the estimates from a larger set of 4192 models selected in this way, all of which have standardised scores
of less than
48. Each of these models was run with inputs of 10 000 years of hourly data. The
dashed lines in Figure 8.5 represent 5% and 95% possibility limits using an average of the trapezoidal
weightings over all evaluation measures.
Figure 8.5 also demonstrates another realisation effect, this time associated with the observations
themselves. The fit of the model to flood peak frequency data for the subcatchment sites was better than
that demonstrated for the Cheb site in Figure 8.5, which is outside the 5% and 95% possibility limits for
small floods. As noted earlier, that site was not used in calibration but also largely represents a different
period to the data used in calibration. It would appear as if there has been a change in flood characteristics
between the late 19th to early 20th century period and the later period.
+
/
−
1
.
8.8 Changing Risk: Catchment Change
All catchments have a history. Some of this history has a long timescale, such as soils developed on
till or fluvioglacial deposits from the last ice age, or deep lateritic soils resulting from weathering over
long periods. Some changes are the much more recent results of human activities, such as urbanisation,
deforestation, reforestation, reservoir or detention pond construction, the effects of wild fires, or the
installation of field drainage systems. Both natural and anthropogenic changes in land use have affected
catchments in the past and are continuing today. Some changes are documented, others may not be
(I know of one experimental catchment in the UK where it was discovered, after the installation of the
instrumentation, that the headwater hollows were drained by old field drains built of stone slabs and
estimated to be more than 100 years old).
In most gauged catchments, certainly larger gauged catchments, such changes have been on-going
during the period of historical records. However, hydrological analyses, for example flood frequency
analyses, do not often attempt to take any account of the possible effects of such changes. Why not?
Primarily because it may be very difficult to detect such changes in the historical record given the
difficulties of closing the water balance and the natural year to year and storm to storm variability in
hydrological responses, especially where the changes are only gradual. Even where a distinct change
in behaviour is detected it can be difficult to differentiate between different potential causes (Fenicia
et al.
(2009) describe a large-scale example in the Meuse catchment). In a study carried out for the UK
government, we examined the hydrological records for a number of catchments which were expected
to show the greatest effects of land use change and management on discharges (Beven
et al.
, 2008c).
This analysis showed that discharge was generally increasing over the 30-year period studied, but that
these changes were matched by increases in rainfalls. This is consistent with other UK work that has
indicated an increase in the number of storm peaks above different thresholds over time (e.g. Archer,
2007; Climent-Soler
et al.
, 2009). By classifying individual storm events into classes that depended on
storm rainfall volumes and antecedent conditions, some effects on peak timing and volumes were detected
in the parameters of DBM models for different periods - but only for the class with small input volumes
and dry antecedant conditions, not those that produce floods. The other classes showed a lot of variability
but no apparent trends across the period (the classification allows similar hydrological conditions to be
