A robust simulated annealing technique for flood clustering

Abstract:

The investigation of extreme flood events is of great importance for applied hydrology. The similarity of the events occurring at a selected location at different times, and the similarity of events occurring at different locations are both of importance. In this study, a clustering method using simulated annealing is proposed. Distributions of the extreme discharges are calculated and normalized. The Kolmogorov-Smirnov test statistics is used as a basis for the distance of the normalized distributions. The clustering is subsequently considered as an optimization problem, where groups with small distances within the group and large distances between the groups are to be found. An objective function reflecting both the above measures is formed, and a simulated annealing algorithm is used to identify the groups. The methodology is applied to 46 mesoscale subcatchments of the Neckar catchment.

The results indicate three major clusters, which illustrated a particular pattern for flood occurrences except in some small sub-catchments almost with high elevation. The subcatchments show agreement with each other in each cluster. The mapping of clusters presents areas with distinct flood behavior and, consequently, different flood protection action plans. Thus, we conclude that the RSA approach is indeed able to successfully cluster the flood events by neglecting all assumptions for clustering.

The suggested algorithm can be extended to classify individual events for their similarity in relative peak, volume and duration.