Research On Fault State Diagnosis And Condition Assessment Of 10kV Cable Based On Data Driven

Cross linked polyethylene(XLPE)cable has been widely used in modern power system because of its superior insulation performance.However,its insulation capacity is not invariable.In addition to normal aging,XLPE cable is also prone to insulation damage due to mechanical stress,chemical corrosion and other effects,which will aggravate aging over time.Once the cable is seriously aged,failure may occur,resulting in power failure or even accidents with irreparable losses.Compared with other highvoltage cables,10 kV XLPE cables are used more,but their management is disordered and the operating environment is worse.Therefore,it is necessary to comprehensively monitor the state of 10 kV cable,which provides theoretical support for the cable refinement operation and maintenance,and ensures the power supply reliability of distribution network.This dissertation takes 10 kV single-core XLPE cable as the research object.From two angles of cable fault state and normal state,the research on10 kV cable fault state diagnosis and assessment based on data-driven is carried out by combining traditional electrical analysis with deep learning algorithm.The main work of this dissertation is as follows:Aiming at the problem of early fault state diagnosis of 10 kV cable,this dissertation proposes an early fault state diagnosis method of 10 kV cable based on deep learning and time correlation from the perspective of time correlation of multi parameter shallow characteristics.By analyzing the coupling relationship of multiple observable electrical quantities,the feature research object is expanded from single electrical quantity to multiple electrical quantities,which enriches the source of feature information and ensures the integrity of information acquisition.The abrupt signal detection and interception are realized by the cumulative sum algorithm.The state transient process is disassembled into multiple transient states.Considering the time correlation between transient states,the time sequence pair is extracted as the shallow diagnosis feature,so as to construct the combined time sequence feature matrix.By extending the long short term memory network to improve the nonlinear mapping ability,a double-layer long short term memory network for processing time series input is constructed.Finally,the training test is carried out through the simulation data set.The simulation results show that the proposed method has good diagnosis effect.In the early fault state diagnosis of 10 kV cable,the sheath current contains more information than phase current,and it plays a greater role.From the perspective of single parameter and deep statistical characteristics,this dissertation proposes an early fault state diagnosis method of 10 kV cable based on sheath current and deep convolution network.The multi-scale decomposition and reconstruction of sheath current signal is carried out by using wavelet transform.The energy characteristic ratio of frequency band is defined,and the waveform and energy characteristics of multi band signal are extracted from the perspective of time domain and energy domain.Then,a two-dimensional characteristic matrix for cable state diagnosis is constructed.Considering the strong local correlation of the constructed characteristic matrix,a 7-layer deep convolution neural network is constructed.Finally,the adaptive moment estimation algorithm is introduced into the model training under supervised learning to obtain the accurate diagnosis model of cable fault state.The simulation results show that the diagnosis accuracy of the deep convolution model based on deep statistical features is higher than that of the double-layer long short term memory network model based on end-to-end learning paradigm.However,the robustness is reduced.The existing cable state diagnosis models based on data driven lack the ability of model generalization and the memory resources of hardware in power system are limited.This dissertation proposes a fault state diagnosis method of 10 kV cable based on incremental learning and deep convolution network,which solves the problem of improving the generalization ability of cable state diagnosis model under the condition of limited memory.The constructed deep convolution network is used for characterization learning and the class features are extracted from the input statistical features.Then the class center is calculated and the combined characteristics set of old states is established with fixed capacity.Based on the knowledge distillation,the distillation loss of samples in combined characteristics set and the classification loss of new state samples is calculated respectively.Finally,the parameters of convolutional neural network are updated based on these two losses and L2 regularization loss.After the model parameters are updated,the combined characteristics set is updated by reducing the number of old class samples in combined characteristics set and adding new class samples according to the memory limit.Cycling the above process,the model learning mechanism is formed under state increment and the generalization ability of the cable state diagnosis model is improved,which ensures that the constructed model can not only diagnosis the old states,but also own the diagnosis ability of the new class state.Finally,the classifier of nearest class center is used to realize the accurate classification of cable state.Relevant experimental results show that this method can better realize the incremental learning of cable state under limited memory.Aiming at the evaluation problem of cable under normal state,this dissertation proposes a condition assessment method of 10 kV cable based on signal propagation characteristics from the perspective of the correlation between signal propagation and cable insulation.Firstly,the voltage signals at both ends of the cable are measured,and the signal attenuation coefficient and phase angle offset are calculated through the signal propagation time and amplitude difference according to the reflection law of terminal signal.Then,the composite dielectric constant of the cable is deduced by using these two parameters,and then the ratio of real part to imaginary part of the composite dielectric constant is calculated to obtain the dielectric loss factor that can characterize the insulation level of the cable.In order to verify the proposed method,the vector fitting method is used to reasonably approximate the cable series impedance and parallel admittance.Meanwhile,the distributed parameter effect is considered and the modeling of 10 kV single core XLPE cable is realized by Γ-type circuit cascading.Considering the influence of random error and synchronization error,the relevant simulation analysis is carried out.The simulation results show that the method is feasible and effective.