Abstract
Estimation of extreme floods, based on statistical analyses of observed values of flood series, is associated with several problems leading to high uncertainty in the flood estimates. In particular is this related to the length of the observation series and the extrapolation outside the range of observed values, when estimating the rare extreme floods. In many cases, short hydrological records are the common rule rather than the exceptions.
This thesis targets at-site, single station, frequency analysis for estimation of the R-year floods magnitude based on the annual maximum approach with observed and simulated series. An R-year flood is a flood with a return period of R years; it has the probability every year of 1/R to be exceeded. Extreme floods occur seldom, i.e. they have a long return period or recurrence interval.
In this study long synthetic series of precipitation (P) and temperature (T) are generated and used as input to a hydrological model to simulate long runoff series. The objective is to investigate how the results of the frequency analysis of the observed series compare with the results of the simulated series.
The modelling framework consists of the Bartlett-Lewis Rectangular Pulse model for rainfall and autoregressive temperature model, the Rindal-Onof generator, and the HBV model. The Bulken catchment in Norway is used as test case. Thousand years in length of daily values of precipitation and temperature are produced and used as input to the HBV model. Two different parameterisations of the HBV model is applied; HBV1 and HBV2. The models satisfactorily reproduces the daily mean discharge for the Bulken catchment for the observation period 1959-1990, but shows a tendency to underestimate the annual maximum peak flows for the observation period for both models. It was found that good values for the commonly applied Nash-Sutcliff criterion not necessarily guarantee adequate simulation annual maximum series. This is in accordance with the experiences of other similar studies.
The HBV2 version, which had the best calibration results for the annual mean flood, provides the best results for the annual maximum flood series for Bulken.
The two synthetic runoff series (HBVPT1 and HBVPT2) resulting from the two HBV model parameterisations with synthetic P and T as input, both overestimate the runoff. HBVPT2 gives substantially higher annual maximum flood peaks than HBVPT1.HBVPT1, simulated with the HBV1 model that showed the poorest results for AM values for the observed series, provides the best resemblance with historical data of the two synthetic flood series. HBVPT2, simulated with HBV2, is by the statistical Mann-Withney test found not to belong to the same population as the observed flood series for Bulken. These findings might indicate that the synthetic series of precipitation or temperature not haven been adequately simulated by the stochastic generators in this particular study case.
Estimates of the R-year floods are obtained by using the GEV distribution and l-moments. The results reflect the findings described above. The 10-year flood estimate based on HBVPT2 is 11 % higher than the estimate based on the observed record for Bulken, and for HBVPT1 +1.3%. HBVP2 gives the highest estimates with for the 1000-year flood +22% whereas HBVPT1 + 7 % compared to the estimates based on the observed flood record for Bulken.
This study represent one test case, so further work is recommended to investigate in more detail the reasons for the underestimation of the annual maximum floods by the HBV model and to identify possible improvements that need to be done concerning the precipitation and temperature simulation.