K vizinhos mais próximos circular: uma nova proposta para predição de dados angulares

Abstract

Circular data are present in several areas of science and lack specific statistical methods for their treatment. The calculation of descriptive statistics for data of a linear nature, for example, does not result in adequate values or that has practical meaning for data in the circumference. Regarding regression models, the literature presents parametric regression models for circular data, which presupposes certain circular probability distributions for their adjustments. On the other hand, in the machine learning field, a supervised prediction approach for continuous data involves non-parametric regression models, which may not be suitable for situations where the variable of interest is circular. In this context, the main goal of the present work is to develop non-parametric predictive models for circular data, based on the concepts of machine learning, namely, the circular k nearest neighbors (CkNN). This methodology was employed in the development of machine learning algorithms for circular data and predictions of directional wind data in different automatic weather stations of several municipalities in the Rio Grande do Sul estate, Brazil, in addition to some municipalities in the states of Bahia and Santa Catarina. The quality of the chosen models was measured using a specific risk measure.