Simulation of precipitation and discharge extremes of the river Meuse in current and future climate

The river Meuse is the second largest river in the Netherlands and is characterized by strong variations in its discharge. From a government point of view there is a particular interest in discharge levels with return periods in the order of 1000 years, far beyond the length of the observed discharge record (several decades). Traditionally, these extreme quantiles are estimated by fitting extreme value distributions to observed maxima, which are then extrapolated upto the return period of interest. However, such strong extrapolation induces a large uncertainty in the estimated quantile. Moreover, the discharge record might be inhomogeneous due to changes in the river basin. As an alternative, a methodology has been developed which is based on the resampling of meteorological data for the basin in combination with hydrological modelling.

The resampling algorithm is used to synthesize long-year sequences of spatially varying daily precipitation and temperature for the river basin and is based on the concept of “nearest-neighbours”. It is capable of reproducing several characteristics of the data crucial for the simulation of extreme multi-day precipitation events, such as spatial correlation of daily precipitation and temperature, persistence and variance as well as the correlation between precipitation and temperature. One of the major advantages of this algorithm is that it is entirely “data-driven” and does not rely upon assumptions about the underlying distribution or correlation structure of the data. Since the generated sequences consist of values from the original data only and therefore cannot exceed the highest data value, something which is sometimes seen as a somewhat unrealistic limitation. To examine to what extent this limitation affects the simulation of discharge extremes, a modified resampling algorithm was tested on the Ourthe subbasin using historical meteorological data. The modified algorithm is based on nearest-neighbour regression and allows for the exceedance of the largest historical daily precipitation in a simulation by recombining a conditional expectation and a sampled residual. Though the algorithm produces daily amounts well outside the range of historical values, the effect on the distribution of simulated discharge extremes on the Ourthe was very small.

To assess the changes of these discharge quantiles related to climate change, this methodology was applied to the output of regional climate models (RCMs) for the control climate (1961-1990) and the SRES-scenario A2 (2071-2100). To compensate for systematic differences in the mean and variance between the RCM run for the control climate and observations in the Meuse basin, the synthesized sequences were subjected to a nonlinear bias correction. This type of bias correction allows for the mean and variability of daily precipitation to be adjusted simultaneously, which turned out to be indispensable for the realistic simulation of extreme discharge events. It was observed that the changes in the extreme quantiles of multi-day precipitation and discharge in the winter half-year (i.e. the flooding season) are to a large extent determined by the global climate model (GCM) which drives the RCM at its lateral boundaries. One of the two used GCMs leads to a decrease in the relative variability of the modelled winter precipitation. This decrease largely compensates the effect of the increasing mean precipitation on the extremes. As a result, the RCM simulations driven by this GCM show a slight decrease of the quantiles at intermediate return periods and only a slight increase of those at long return periods. In the simulations driven by the other GCM the relative variability of precipitation hardly changes and the extreme quantiles roughly increase proportionally with the mean precipitation.

Resampling data from RCMs combined with hydrological modelling proves to be a suitable instrument to obtain more insight in the changes of rare discharge extremes of rivers like the Meuse. The biases in the mean and variability of RCM precipitation and their correction deserve careful consideration. In particular for the study to extreme precipitation events, changes in precipitation variability are as important as changes in the mean. In the flooding season these changes are primarily determined by the driving GCM. The model uncertainty in the changes of such extreme events is therefore best represented by an ensemble of modelruns, driven by different GCMs.