Research On Partial Discharge Detection In Distribution Network Based On Deep Learning

The partial discharge of distribution network is generally caused by the concentration of line partial electric field inside or on the surface of the insulator,and it is also one of the main reasons for the insulation breakdown of line electrical equipment.The existing detection methods of partial discharge in distribution network mainly rely on external physical detection equipment such as sound,light and ultrasonic,but most of the energy of this kind of physical signal is absorbed or reflected by different media,which makes the detection results unstable.Moreover,some overhead lines are erected in complex environment,which makes it difficult for personnel to approach,and the efficiency of physical detection methods is low.The existing detection methods based on deep learning technology are not deep enough and the detection accuracy is not high,so it is difficult to achieve effective detection.In order to solve the above problems,this thesis first analyzes the current and voltage signals.Through comparative experiments and statistical analysis,it is concluded that the voltage signal processed by high pass filtering and discrete wavelet transform is more suitable as the feature signal of PD.It also proves that the abnormal signal is mainly in the high pass signal component of the voltage signal.In order to solve the problem that traditional signal analysis methods can not detect batch signals quickly and accurately,and highlight the role of high pass signals in PD detection,an improved C-RNN deep learning model:CNN-LSTM model is proposed.This model solves the problem that the detection accuracy of the original model decreases due to gradient dissipation when processing long data segments,and proposes a method to construct characteristic matrix This method can reduce the dimension of the input signal and improve the detection speed and accuracy while maintaining the important characteristics of the original signal.Based on the CNN-LSTM depth model,this thesis continues to conduct in-depth research,integrates the attention mechanism into the improved model,and proposes a"Deep Convolution LSTM Attention" model,which can concentrate the computational power on the key signals.At the same time,it also proposes a threshold search algorithm to alleviate the learning accuracy degradation caused by extremely unbalanced data sets.Simulation results show that the proposed method has the highest PD detection accuracy of 86.44%,Matthews correlation coefficient of 0.66103,F1 value of 0.78202 and AP value of 0.81.The results of four kinds of evaluation indexes are higher than those of C-RNN,CNN-LSTM,LSTM and LSTM Attention depth models,which proves the effectiveness of the proposed method.