Time-Frequency Memory Recurrent Neural Networks Empowered Incipient Fault Detection in Power Distribution System

Abstract

The incipient fault detection of power distribution system is essential to improving the reliability of the power grid. The incipient fault has the characteristics of low occurrence rate, instantaneous and self-clearing, and fewer samples can be collected in practical applications. This type of fault need to be detected in time, otherwise it may cause permanent fault of power equipments. This paper combines wavelet transform with long short-term memory (LSTM) unit, and proposes a novel neural network basic cell, time-frequency memory (TFM), that can perform adaptive time-frequency analysis on time series sampled voltage and current data, to realize intelligent detection of incipient faults. Based on a public dataset the TFM based recurrent neural networks (RNN) achieves state-of-the-art (SOTA) performance.