Rainfall generator for the Rhine basin; multi-site generation of weather variables for the entire drainage area

This is the final report of a project on the development of a rainfall generator for the Rhine basin. The request for this generator arose from the need to study the likelihood of extreme river discharges in the Netherlands, using a hydrological/hydraulic model. Long-duration, multi-site simulations of daily precipitation and temperature time series for the entire Rhine basin are needed for this purpose. Temperature is included to determine snow accumulation and snowmelt. Daily precipitation and temperature data from 36 stations in Germany, Luxemburg, France and Switzerland for the period 1961- 1995 are considered. Studies in earlier reports dealt with the German part of the basin only.

Time series simulation is done by nearest-neighbour resampling. The method does not make restrictive assumptions about the underlying joint distribution of the multi-site precipitation and temperature data. In order to generate weather variables for day t, a feature vector Dt is formed to find the nearest neighbours of this day, or the previous day, in the historical data. For unconditional simulations Dt contains variables that charac- terize the weather on the previous day t − 1 (first order model) or a number of previous days (higher order model). Circulation indices for day t are included in Dt in the case of conditional simulation on the atmospheric circulation. A finite number k of nearest neighbours in terms of a weighted Euclidean or the Mahalanobis distance is selected from the historical record. One of these k nearest neighbours is finally “resampled” using a discrete probability kernel.

The criteria used to assess the performance of different simulation methods are the ability to reproduce the second-order moment statistics of daily and monthly values of precipitation and temperature and the distribution of multi-day winter (October-March) precipitation amounts. For the high-elevation stations in Germany and Switzerland the snowmelt simulation is also evaluated. First-order conditional and unconditional models for the generation of daily precipitation and temperature are considered. Conditional simulations were done with simulated circulation indices produced by a separate second- order resampling model.