Research On Weak-feature Fault Detection Based On Cloud-edge Collaboration In Distribution Networks

In achieving the carbon peak and carbon neutrality targets,the power distribution network in the new power system plays a major role in renewable energy consumption and friendly interactions between the “source-network-load-storage”.The safe and reliable operation of the distribution network is the key to ensuring the development of the energy system.Weak-feature faults mainly include single-phase ground faults in the small-current grounding distribution network and high impedance faults grounded by complex dielectrics with weak conductivity.Weak power faults have become a core issue affecting the safe operation of the distribution network as the characteristics of weak power faults are difficult to reach the threshold of protection action and energized operation is prone to cause electrical safety incidents or accidents.With the development of measurement devices and power Io T technology,data-driven weak-feature fault detection methods for distribution networks have become one of the important research directions due to their advantages of strong generalization ability and no threshold setting.However,few historical fault sample data,difficulties in expanding different topologies,and the lack of adaptive optimization of model applications are the key bottlenecks that limit the application of such methods in practice.This paper focuses on the practical method of weak feature fault detection based on artificial intelligence from the perspective of data learning and physical knowledge fusion.The main works of this paper are as follows.(1)The weak-feature faults are classified into small current faults,weak arcing high impedance faults,and strong arcing high impedance faults based on the initial fault impedance magnitude and the degree of fault current nonlinearity.And a full-scene weakfeature fault diagnosis framework is established based on the applicability of the transient and steady-state characteristics,which provides a theoretical basis for the methods of this paper.Based on the potential common and individual knowledge between features of different application scenarios,a feasible idea of fusing different distribution networks to solve small-sample problems is formed.Therefore,the cloud-edge collaboration architecture for weak-feature fault detection is proposed.(2)For small-current faults and weak arc high impedance faults,a weak-feature fault detection method based on the fusion of multiple distribution networks information and transfer learning is proposed.By combining discrete wavelet transform and principal component analysis,a global transient feature matrix extraction method with uniform scale is proposed,and a cloud basic model based on convolutional neural network is trained by fusion of multiple distribution networks.In a specific distribution network,the cloud model is transferred to the corresponding edge by transfer learning,and a data augmentation method based on locally sensitive hash is proposed to improve the transferring performance.The proposed method can achieve reliable detection of small-current faults and weak arc high impedance faults in a single distribution network with a small sample data scenario,and has good topological generalization capability,as verified by the simulated and actual field data.(3)For small-current faults and weak arc high impedance faults,a weak-feature fault section location method based on improved multi-view spectral clustering is proposed.By forming multiple perspectives through the original waveform of transient zero-mode current and the reconstructed features after discrete wavelet transform,the multi-view spectral clustering-based weak-feature fault section location method without data training and threshold setting is constructed by utilizing the difference of transient features on the fault path and non-fault path.Based on the topological correlations between fault path nodes,the edge weight matrix in the iterative process of multi-view spectral clustering is improved to guide the features with potential correctness under multiple views to become the dominant features,which improves the applicability of the clustering algorithm in weak feature fault location scenarios.The proposed fault section location method has good positioning accuracy and reliability as verified by the simulation model and actual field data.(4)For strong arc high impedance faults,a nested attention mechanism-based arcing high impedance fault detection method is proposed.Based on a priori information about the location of waveform distortion and multidimensional time-series data generated by long and short-term memory networks,a nested attention mechanism combining the priori information and data learning is proposed to enhance the sensitivity of the model to high impedance fault features.On this basis,robust reconstruction of waveform distortion features is achieved by using parallel autoencoders,and the detection of arcing high impedance faults is realized by combining the discrimination of reconstructed feature values.Through verification of actual field data,the proposed method can achieve sensitive and effective arc high impedance fault detection in the case of small sample data.(5)For strong arc high impedance faults,an arcing high impedance fault section location method based on feature enhancement and target detection is proposed.A global gray-scale feature extraction method containing fault features and location information is proposed by combining distribution network topology with synchronous harmonic features.On this basis,a locally excitatory globally inhibitory oscillator region attention mechanism is proposed,which can achieve adaptive feature enhancement by global feature comparison.Finally,the high impedance fault features are identified and labeled by YOLO v2 target detection network,and the section location is achieved by combining it with the corresponding topological area map.The simulation model and the actual field data verify that the proposed fault section location method has high location accuracy and good reliability.Based on the cloud-edge collaboration architecture,this paper studies the datadriven fault detection methods for distribution networks under different application scenarios,data levels,and detection requirements.A weak feature fault detection system is formed with high practicality and strong generalization capability.The flexible application of artificial intelligence technology and targeted improvements effectively enhance the accuracy and reliability of weak feature fault detection.This paper explores data-knowledge fusion as an entry point to solve the problems of small sample scenario application,feature scale unification,model transferring and deployment,data utilization efficiency improvement,and model adaptive optimization,which effectively enhance the practicality and feasibility of the weak-feature fault detection method for distribution networks.