Geoscience Reference
In-Depth Information
flood stages are highly uncertain, imposing mass balance on a forecasting model might actually be a
disadvantage. This is reinforced by many other studies that suggest that the major source of uncertainty
in flood forecasting is knowledge of the inputs (e.g. Krzysztofowicz, 2002).
There are other advantages to using water level information directly in forecasting. One is that level is
actually measured directly so that measurement errors are, in general, small. No additional uncertainty is
introduced by the application of the rating curve. Rating curve uncertainties can be significant for flood
stages (the estimate of the peak discharge for the Carlisle flood was revised from 900 m
3
s
−
1
, obtained
by extrapolating the rating curve prior to the event, to 1500 m
3
s
−
1
after a post-event re-evaluation). In
addition, for flood warning purposes, the predicted variable that is required is very often the water level.
Decisions about warnings are usually made on the basis of forecast levels (although velocities derived
from discharges might be of secondary interest in assessing risk to life). It is also worth remembering
that the forecasting problem is different from the simulation problem. The requirement for forecasting is
to have an accurate estimate of water level with minimal uncertainty at the lead time of interest. It does
not really matter how that is achieved and because of all the potential for operational epistemic errors
during an event it might be more advantageous to use a simple model that is easily made adaptive than a
complex model that aims to simulate the runoff processes more correctly.
Both the headwater rainfall to level models and the level to level flow routing models should be
nonlinear. The nonlinearities were identified using the SDP approach described in Box 4.3 and represented
using a Takagi-Sugeno first-order fuzzy inference system (T-S FIS). The resulting nonlinearities for the
three components of the forecasting cascade are shown in Figure 8.2. Note how that for the flow routing
(Figure 8.2c) is quite different from the headwater rainfall to level models. The models are completed by
Figure 8.2
Identified input nonlinearities for each of the forecasting system components in the River Eden,
Cumbria, UK: (a) input nonlinearities for Model 1 in Figure 8.1; (b) input nonlinearities for Model 2 in
Figure 8.1; (c) input nonlinearities for Model 3 in Figure 8.1 (after Leedal et al., 2008, with kind permission of
the CRC Press/Balkema).
