The performance of the most frequently used flood frequency probability distributions in South Africa (Log-Normal, Log Pearson3 and Generalised Extreme Value) were reviewed and all tend to perform poorly when lower exceedance probability frequency events are estimated, especially where outliers are present in the dataset. This can be attributed to the challenge when analysing very limited ‘samples’ of annual flood peak populations, which are an unknown. At present outliers are inadequately 'managed' by attempting to 'normalise' the flood peak dataset, which conceals the significance of the observed data. Thus, to adequately consider the outliers, this study was undertaken with the aim to improve the current statistical approach by developing a more stable and consistent methodology to estimate flood quantiles. The approach followed in the development of the new methodology, called IPZA, might be considered as unconventional, given that a multiple regression approach was used to accommodate the strongly skewed data, which are often associated with annual flood peak series. The main advantages of IPZA are consistency, the simplicity of application (only one set of frequency factors for every parameter, regardless of the skewness), the integrated handling of outliers and the use of conventional method of moments, thereby eliminating the need to adjust any moments. The performance of IPZA exceeded initial expectations. The results are more consistent and, by taking outliers into account, appear to be more sensible than existing probability distributions. It is recommended that IPZA should be used as a valuable addition to the existing set of decision-making tools for hydrologists/engineers performing flood frequency analyses.
Read full abstract