A two-component generalized extreme value distribution for precipitation frequency analysis

A two-component generalized extreme value (TCGEV) distribution is introduced based on the assumption that the annual maxima for convective and stratiform precipitation follow two separate generalized extreme value (GEV) distributions. The regional TCGEV model is used to analyze 6-hour precipitation data for 11 stations in the Czech Republic over 1982–2010 subdivided into predominantly convective and stratiform precipitation. For each type of precipitation, the shape parameter and the ratio of the scale parameter and the location parameter of the underlying GEV distributions are assumed to be constant over the region. The validity of this homogeneity assumption is explored with a bootstrap procedure and the goodness-of-fit is tested with the Anderson-Darling statistic both for each individual station and for all stations simultaneously. The return levels from the regional TCGEV distribution are compared with those obtained with the common method of fitting a regional GEV distribution to the overall annual maxima, ignoring their convective or stratiform origin. The differences are generally small, but they increase with return period and are larger at lowland stations where the proportion of convective precipitation extremes is greater. High return levels based on a GEV fit to the overall annual maxima for these stations tend to be smaller than those for the convective component owing to the heavier upper tail of the distribution of convective extremes. Results from the TCGEV distribution are consistent, i.e. the estimated return levels of the overall annual maxima cannot be smaller than those for the convective and stratiform components obtained from the GEV distribution.