Research On GIS Risk Assessment Based On Partial Discharge Deep Learning

With the maturity of Partial Discharge(PD)detection technology for power equipment,partial discharge detection of SF6 Gas Insulated Switchgears(GIS)has obtained a large number of substation field applications.Some potential defects of GIS equipment have been found,which promoted the state maintenance of GIS equipment in service.At the same time,some problems also exposed.For example,the diagnostic and recognition effect of partial discharge data from GIS in service is poor,and the massively increased partial discharge detection data brings the problem of processing multi-source heterogeneous data and the mining of historical detection data.Finally,due to the many influencing factors of partial discharge,the traditional method is difficult to complete the effective evaluation of the risk assessment of partial discharge.In this context,the data mining method based on deep learning provides a new idea and solution for GIS partial discharge risk assessment.Through the laboratory simulation experiment of partial discharge and on-site detection of substation,a complex multi-source partial discharge data set is established by using various testing instruments in the paper.A normalization method for structured PD data and an information reconstruction method for unstructured PD data is proposed for the problem of multi-source heterogeneous partial discharge data normalization.A method for detecting bad data based on two-dimensional Empirical Mode Decomposition(EMD)is proposed for data quality problems.Finally,a standardized partial discharge database containing more than 100,000 items was established.Then,through the analysis of intelligent learning theories,methods and complex multi-source GIS PD data set,a partial discharge data processing architecture involving pattern recognition,case-based reasoning(CBR)and risk assessment is established and related algorithms are designed,the diagnosis and risk assessment of GIS PD data in service are realized.Aiming at the pattern recognition problem of partial discharge in complex multi-source background,a PD pattern recognition method based on Convolutional Neural Networks(CNN)is proposed in this paper.In this method,an CNN model applied for PD data pattern recognition is established,which parameters are initialized by deep auto-encoder network with complex multi-source sample data.Identification parameter optimization is achieved by convolution,pooling and backpropagation operations.the pattern recognition of massive PD data in complex scenes is accomplished by extracting the multi-source PD big data feature map.Compared with traditional support vector machines based on statistical eigenvalues,back propagation neural network(BPNN)and random forest methods,the proposed method identifies correctly in the task of processing partial discharge data samples of complex data sources.The rate is increased by more than 10%.Compared with the deep belief network,the recognition accuracy rate is above 4%,and the recognition accuracy rate increases with the increase of the sample data,which is more suitable for the engineering application requirements of the big data platform.With the accumulation of partial discharge detection cases,it is an effective attempt to do deep data mining by compute the match rate between the detected partial discharge data and historical case database data under a big data background.In this paper,a framework for GIS PD case-based reasoning is proposed.Aiming at the key steps in CBR,the design of the case library and the design of the matching algorithm,the ontology theory based PD case knowledge library is established and a partial discharge data matching method based on Auto-Encoding Variational Bayes(AEVB)is proposed,respectively.The PD case knowledge library based on Ontology combines the ontology theory knowledge representation technology to describe the PD data.The PD ontology model library,the GIS PD basic information database and the GIS PD case library were constructed respectively.The relationship between PD data and the related external information are described by the PD ontology model,and the PD case library and the GIS PD basic information database are organically associated.At the same time,as a carrier of semantic information,it effectively improves the interpretability of data.In the PD data matching method based on AEVB,an AEVB network model for partial discharge data is constructed to extract the deep eigenvalues.Then the cosine distance is used to calculate the match rate between different partial discharge data.To verify the advantages of proposed method,the proposed method is compared with other feature extraction method and matching method including statistical feature,deep belief networks(DBN),deep convolutional neural networks(CNN),principal component analysis(PCA)and linear discriminant analysis(LDA),Euclidean distance,information entropy.The experimental results show that the cosine distance matching rate based on the AEVB feature vector can effectively detect the similar partial discharge data compared to other data matching methods,provide information support for partial discharge risk assessment.The process of GIS PD risk assessment is analyzed in this paper.Aiming at the problem of GIS fault probability calculation based on PD data under operating conditions,a method for assessing the severity of partial discharge for GIS in service is proposed based on the partial discharge detection data from substation GIS on site,combined with the long short-term memory(LSTM)network and Bagging ensemble learning.First,a large on site data set is established and data labels is determined.Second,aiming at the problem of imbalanced data,an ensemble learning model for assessing the severity of partial discharge was constructed using N individual LSTM depth networks via Bagging ensemble learning.By analyzing the eigenvectors composed of the characteristic of partial discharge data,technical impact factors,and equipment operation information,the model can output the severity assessment result.The comparative experiment with common LSTM network,backpropagation neural network(BPNN)and Bagging-BPNN was undertaken,as well as on site detection case analysis,the results show that the proposed method can effectively assess the severity of GIS partial discharge under operating conditions.Compared with common LSTM,BPNN and Bagging-BPNN,the results of proposed method is more in line with the actual state of the GIS partial discharge risk assessment in service.Compared with ordinary LSTM,BPNN and Bagging-BPNN,the accuracy of the evaluation results can be improved by 10 times,and it is easy to implement deployment on a computer.