典型神经网络串联型电弧故障检测及选线方法研究OACSTPCD

The accuracy and real-time of arc fault detection method based on traditional feature analysis and machine learning may be affected by the subjectivity of feature parameter selection and the process of feature analysis.A three-phase and multi-load parallel series arc fault experimental system was built.The time series of the current signal of the trunk circuit when the arc fault occurred in different branches and phases were analyzed.After classification,segmentation and standardization,the signal was directly used as input samples of the detection model.Then,the deep convolutional neural network(DCNN)model,long short-term memory(LSTM)model,and artificial neural network(ANN)model were built and trained.The differential method was proposed to optimize the classification results of the detection models.The detection and line selection effects of the three models were compared through the accuracy,loss function value,online test speed,and accuracy of multiclass recognition after optimization.The results showe that the accuracy of the series arc fault detection and line selection model based on DCNN can reach 96.77%for motor load and 98%for frequency converter load.The accuracy exceeded that of other three-phase circuit arc fault detection models in recent years.