Comparison of annual maximum series and flood-type-differentiated mixture models of partial duration series

TL;DR

This paper compares traditional annual maximum series flood models with flood-type-specific mixture models, finding the latter more effective when multiple flood types occur frequently or differ significantly, despite higher complexity.

Contribution

It introduces and evaluates flood-type-specific mixture models as an alternative to classical models for flood frequency analysis, highlighting their advantages in certain scenarios.

Findings

Mixture models outperform classical models with many flood events per year.

Mixture models are preferable when flood types differ significantly.

Higher parameters in mixture models increase uncertainty but improve accuracy.

Abstract

The use of the annual maximum series for flood frequency analyses limits the considered information to one event per year and one sample that is assumed to be homogeneous. However, flood may have different generating processes, such as snowmelt, heavy rainfall or long-duration rainfall, which makes the assumption of homogeneity questionable. Flood types together with statistical flood-type-specific mixture models offer the possibility to consider the different flood-generating processes separately and therefore obtain homogeneous sub-samples. The combination of flood types in a mixture model then gives classical flood quantiles for given return periods. This higher flexibility comes to the cost of more distribution parameters, which may lead to a higher uncertainty in the estimation. This study compares the classical flood frequency models such as the annual maximum series with the type-specific mixture model for different scenarios relevant for design flood estimation in terms of Bias and variance. Thee results show that despite the higher number of parameters, the mixture model is preferable compared to the classical models, if a high number of flood events per year occurs and/or the flood types differ significantly in their distribution parameters.