基于长短时记忆神经网络的励磁涌流与故障电流识别方法

Identification of Inrush Current and Fault Current Based onLong Short-Term Memory Neural Network

变压器空载合闸时产生励磁涌流导致差动保护误动作的问题至今仍未能被完全解决.针对该问题,提出一种利用长短时记忆(LSTM)神经网络识别励磁涌流与故障电流的方法.首先,在PSCAD软件平台上搭建变压器空载合闸及内部故障仿真模型,通过仿真产生大量三相电流瞬时采样数据作为训练神经网络的样本集;然后,利用Keras平台搭建LSTM神经网络模型并完成训练;最后,利用新的仿真数据和现场故障录波数据对训练好的LSTM神经网络进行测试.结果表明LSTM神经网络可以快速准确地区分各种情况下的励磁涌流和故障电流,从而证实该方法的有效性.

The problem of differential protection maloperation caused by inrush current during no-load closing of transformer has not been completely solved so far.To solve this problem,a method using long short-term memory(LSTM)neural network to identify inrush current and fault current is proposed.First,the simulation model of no-load closing and internal fault of transformer is built on the PSCAD software platform,and a large amount of three-phase current instantaneous sampling data is generated through simulation as the sample set to train the neural network.Then,the LSTM neural network model is built and trained by using the Keras platform.Finally,the new simulation data and fault recorder data is used to test the trained LSTM neural network.The results show that the LSTM neural network can quickly and accurately distinguish the inrush current and fault current under various conditions,which proves the effectiveness of the proposed method.