Detection And Identification Of Underground Cable Incipient Faults Based On Denoising Autoencoder And Optimized Convolutional Neural Network

With the development of industrialization and the accelerated process of urbanization,power cables have been widely used in power transmission and distribution,due to its small size,high security and strong anti-interference.However,suffered from factors such as production process,soil condition,and operating time,localized deterioration in insulation of power cable will exist,which will lead to partial discharge and flashover,and eventually develop into permanent fault.Therefore,it is of great significance to study the detection and identification method of underground cable incipient faults and timely maintain and replace defective cables to improve the safety of the power system.In the research on the detection and identification of cable incipient faults,the data obtained only by recording field data or simulating on the physical experiment is difficult reflect the characteristic of the fault comprehensively and the amount of data is not enough to meet the requirements of relevant analysis.Furthermore,the costs and the difficulties of the implementation of field test also stay at a relative high level.Therefore,on the basis of studying cable insulation deterioration and failure mechanism,the cable model,arc model and the modified IEEE-13 node test feeder system model are established in PSCAD/EMTDC to generate the incipient failure simulation signals.The mathematical models concerning to the simulation model and relevant parameter settings are elaborated in detail so that the signals obtained from PSCAD/EMTDC can substitute for actual data in the research.Therefore,a variety of operating data can be obtained by modifying the parameters of the models,which provides reasonable and sufficient data for subsequent development of fault detection and identification method.The cable incipient fault,different from a permanent fault,is modeled as a self-clearing arcing fault and belongs to transient over-current disturbance signal.In order to distinguish it from multiple normal over-current disturbances,this thesis propose a method for cable incipient fault detection and identification based on denoising autoencoder(DAE)and optimized convolution neural network(CNN).Based on the current signals measured at one end of the cable,a denoising autoencoder is firstly utilized to extract feature and remove noise,which reduces the data dimension,accelerates the learning process and improves the anti-noise ability.Then,the current signal features extracted by DAE serve as the input of the CNN and the mapping relationship between input features and output encoding is established by training and adjusting the network parameters.To further improve the performance of CNN,this paper firstly proposes an improved particle swarm optimization(IPSO)and then optimizes the structure of CNN by the IPSO.Besides,batch normalization and dropout are adopted to address gradient disappearance and over-fitting problems.The detection and identification method of cable incipient failure proposed by this thesis is a data-driven method and therefore,it does not require the analysis of physical model and complex mathematical calculation,which effectively gets rid of the dependency on expert knowledge.In this thesis,Gaussian noise model is adopted to evaluate the anti-noise capacity of the proposed method in the actual operating state and the proposed method is compared with several state-ofthe-art methods.The simulation results show that the proposed method can identify the cable incipient fault from multiple over-current disturbances precisely,which manifests the strong robustness and outperform than other methods.