Proposta de métodos de clusterização de dados com validação por testes de heterogeneidade e discordância aplicados à regionalização de bacias hidrográficas

Abstract

Through the regionalization of river basins, it is possible, among other applications, to make statistical forecasts of maximum and minimum flows in watercourses. Several studies show good results in the use of clustering for the formation of better regions from the hydrological point of view. This work applies machine learning techniques to obtain hydrologically homogeneous regions. More specifically, in this work we have explored the use of the methods: k-means, affinnity propagation, aglomerative clustering and region of influence for the formation of regions, which makes a comparison between the methods and the use of attributes selection techniques. In this work three methods are proposed for the solution of the problem, using clusters adjustment based on the measures of heterogeneity and discordance of Hosking. Two of these methods use the k-means algorithm making variations in the initial clusters in order to look for centroids that best represent hydrologically homogeneous regions. And the other method combines clustering results with regions of influence. With the proposed methods it was possible to improve, in the best result, from 63.2 % to 90.5 % of the utilization of the regions formed for the application of RFA. With this work it was concluded that the selected attributes presented better results than the use of all the attributes, and that the proposed methods show great potential since they presented better results than other already existing methods.