On-line Transmission Line Fault Classification using Long Short-Term Memory

Abstract

In order to perform on-line transmission line fault diagnosis, this paper proposes a classification algorithm, which combines the long short-term memory (LSTM) network with a calibration training filter. The LSTM network adopted in this research is a multilayer recurrent neural network. As a deep learning algorithm, LSTM is extremely suitable to complex time-series classification problems, such as speech recognition and natural language processing. As the number of units in LSTM is much larger than conventional artificial neural networks (ANNs), the training progress is time consuming, and not able to be performed by on-line diagnosis devices. However, the parameters of the transmission line are always varying with time, which requires frequently calibration training on the network. In order to accelerate the calibration training of LSTM, a filter enhanced calibration is proposed. The filter selects samples having the same pattern as the signal under diagnosis, and further reduces the training complexity. The experimental study compares the proposed filter calibrated LSTM (FC-LSTM) against other neural networks and machine learning algorithms on a on-line test model. The numerical comparison not only shows FC-LSTM has a better classification accuracy and a very short time delay.