Power Line Partial Discharge Detection Using Multi-Scale 1D Convolutional Neural Networks

In recent years, due to the widespread access of distributed power supply, the topology of electrical network becomes more and more complex such that it is difficult to analyze the fault causes of transmission lines, and partial discharge (PD) is a type of failure that often occurs. An automated transmission lines PD detection system requires automated feature extraction and online identification of PD. Firstly, A PD detection method based on multi-scale one dimensional convolutional neural network (MS-Conv1D) is proposed in this paper, which takes an arbitrary-long time series of fault signals as input. Furthermore, the single-phase and three-phase electrical signals can be dealt with an unified structure of neural network. Secondly, a complete flow of training MS-Conv1D and data processing is given. Finally, simulations on a real PD dataset shows that the detection accuracy of MS-Conv1D reaches 90% with 20MHz sampling rate, and the receiver operating characteristic (ROC) points indicate the robustness of MS-Conv1D at different frequencies.