Detecção de outliers em curvas de demanda de energia baseado no algoritmo TEDA recursivo

Abstract

The detection ofoutliers is an important problem that has been researched in several areas and application domains. Many outlier detection techniques have been developed specifically for certain application domains, while others are more generic. Outliers can be of different types, in this work we will address outliers of the type punctual (global) and contextual. The present work aims to detect these failures that occurred during the measurement or communication, and to correct these failures in order to give more reliability to the set. The study was carried out using some clustering algorithms to label outliers(clustering), KNN classification algorithm for detecting outliers(classification), Autoencoders deep learning algorithms and TEDA auto evolutive algorithms. For the correction of outliers, linear interpolation is used. Comparisons of results between techniques such as Z-Score, Modified Z-Score, K-Means, C-Means, Autoencoders, Sparsed Autoencoders, TEDA and TEDA Diff are also provided, and these comparisons are evaluated according to the metrics used in the literature. In order to verify the performance of the proposed techniques, active power data from an energy substation in the state of Paraíba were used. This work provided a better understanding of the different directions in which research was carried out on the topic of outliers detection and how the techniques developed in one area (such as TEDA) can be applied in domains for which they were not originally intended, where for different scenarios of outliers in an electrical demand curve, a matthew’s correlation coefficient was always higher than 0.70, above average when compared to the other tested techniques.