【摘要】 针对常规的输电线路故障判别方法准确率低、通用性差、依赖人工干预的问题，提出一种基于深度神经网络的故障判别方法。该方法采用实际行波波形和自动仿真生成的故障波形作为输入，使用小波变换滤噪，训练并构建双向长短期记忆与注意力机制结合的神经网络，利用该神经网络可以根据分布式故障诊断系统终端上送的行波数据辨识故障原因。将多种判别方法应用于由4 800个故障行波构成的测试集，以识别4种常见的输电线路故障，判别结果表明，所提方法的准确率相较常规判别方法提高了9.85%。将所提方法应用于分布式故障诊断系统，可以提升故障判别的准确率至96.79%,为巡检工作提供精确指导。更多还原

【Abstract】 Aiming at the problems of low accuracy, poor generality and relying on human intervention of traditional fault identification methods for transmission lines, a fault discrimination method based on deep neural network is proposed. This method uses the fault waveform of actual traveling wave waveform and the simulation waveform as the input, and adopts the wavelet transform to filter the noise. It trains and constructs a bi-directional long-term and short-term memory(Bi-LSTM)neural network combining with attention mechanism. By using this network, it is able to infer the fault cause from the traveling wave waveform sent by the terminal of distributed fault diagnosis system. Several discrimination methods are applied to a test set consisting of 4800 fault traveling waves to identify four common transmission line faults. The discrimination and comparison results show that the accuracy of the proposed method is 9.85% higher than that of traditional identification methods. The application of this method to the distributed fault diagnosis system can improve the accuracy of fault identification to 96.79% and provide accurate guidance for patrol inspection.更多还原