No abstract

Insulated conductors are used widely in overhead power transmission due to the stability and reduced construction space. However, the ordinary protection devices are not able to detect the phase-to-ground faults without overcurrent. Insulated overhead conductor (IOC) faults are often accompanied by partial discharge (PD) phenomenon. Thus, PD monitoring and recognition plays an important role in evaluating the condition of insulation degradation or detecting power line faults. This paper presents a new approach based on a multi-channel CNN-LSTM (convolutional neural network, long short term memory) network for fault detection by determining whether there is local discharge phenomenon on the IOC, in which the three-phase voltage signals are processed with FFT to obtain low frequency and high frequency components, and then the two components together with the original three-phase signals are fed into three parallel CNNs having different filter lengths, and finally LSTM is used to compose those different-scale features sequentially. Then, the fault types are determined according to the result of fault detection. This proposed method is tested on the ENET public data set with eight types of faults, and simulation results indicate that the method can improve the detection and classification accuracy of IOC faults compared with other classification methods.

Insulated overhead conductor (IOC) faults cannot be detected by the ordinary protection devices due to the existence of the insulation layer. The failure of insulated overhead conductors is regularly accompanied by partial discharge (PD); thus, IOC faults are often judged by the PDs of insulated overhead conductors. In this paper, an intelligent PD detection model based on bidirectional long short-term memory with attention mechanism (AM-Bi-LSTM) is proposed for judging IOC faults. First, the original signals are processed using discrete wavelet transform (DWT) for de-noising, and then the signal statistical-feature and entropy-feature vectors are fused to characterize the PD signals. Finally, an AM-Bi-LSTM network is proposed for PD detection, in which the AM is able to assign the inputs different weights and highlight their effective characteristics; thus, the identification accuracy and computational complexity have been greatly improved. The validity and accuracy of the proposed model were evaluated with an ENET common dataset. The experiment results demonstrate that the AM-Bi-LSTM model exhibits a higher performance than the existing models, such as LSTM, Bi-LSTM, and AM-LSTM.

No abstract