Detecção e correção de outliers em curvas de demanda de energia utilizando redes neurais artificiais autoencoders

Abstract

One of the main problems encountered in Smart Grids is the occurrence of outliers, which can corrupt data, thus modifying the information brought by them, making it difficult for electrical system operators to make decisions based on this information. Therefore, this work proposes an integrated outlier detection and correction methodology, based on artificial neural networks. More specifically, a detection system based on Autoencoders was developed, with the aid of a softmax layer, and a correction system based on Autoencoders. The proposed methodology was contemplated in several scenarios, using data from a real substation, where the influence of the variation in the number of outliers present in the database, as well as the variation of their amplitude, on the functioning of the algorithms, is evaluated. In the tests performed, the detection technique achieved Accuracy and F-scores greater than 99.7% and 97.4%, respectively. The correction technique obtained MAPE mean absolute percentage error of 1.42%, while the root mean square error remained, in most of the evaluated scenarios, below 0.15 MW, a value that represents about 1.7% of the maximum power value available in the database.