Recurrent Neural Network-Based Autoencoder for Problems of Automatic Time Series Analysis at Power Facilities

Abstract

Digitalization of the energy sector leads to an increase in the volume and rate of data collection. A primary barrier to the proper management of the technological data is the lack of data labeling corresponding to emergency modes, power equipment technical state, etc. Thus, despite the large amount of data, there is a shortage of labeled data suitable for training, validating and testing the machine learning models. Labeling by an expert takes too much time, so there is an actual task to automatically identify data fragments that are potentially of interest. The aim of the work is to develop an algorithm for prioritizing the fragments of the time series using the compact recurrent autoencoder. To achieve the goal, a neural network architecture was developed based on recurrent encoding and decoding cells, capable of unsupervised learning. The model was tested on two data sets: a synthetic sinusoidal signal with missing values and electric current measurements with thermal limit deviations. The substantial results of the work are the compact architecture of the autocoding model and the high interpretability of the output. The most significant achievements of the study are both the autocoding neural network model, which does not require initial assumption about the type of deviations, and the proposed algorithm for prioritizing the data fragments. The significance of the results is prooved by the reduction of the time for analyzing and labeling large data arrays with technological parameters of the electrical networks, which allows using these data for training, validating and testing.