Application research of several LSTM variants in power quality time series data prediction

Abstract

Predicting power quality by monitoring data is one of the main topics in power quality research of power grids. With the development of smart grids, power quality monitoring data, as a key indicator for analyzing and regulating the stable transmission of power grids, is exponentially explosive. In recent years, the performance of deep learning methods in large-scale data fitting has been better and better than that of traditional methods. In this paper, based on the strict time series dependence of power quality data, combined with the Long Short-Term Memory neural network (LSTM) of deep learning algorithm, the prediction performance of several LSTM variants (Stacked LSTM, Bi-LSTM, Encoder-Decoder LSTM) on power quality time series data is researched and analyzed. Different LSTM variants are used for training and modeling. The performance comparison and analysis are carried out on the power quality data collected by a State Grid company. From the verification results, the variants have higher prediction accuracy compared with the standard LSTM network variant structure.