| The security of the power grid is directly related to national public security.The largescale blackouts in recent years,such as the "8.14" blackouts in the United States and Canada,the "3.21" blackouts in Brazil,strongly illustrate this problem.The power transformer is the core of power transformation and the most critical equipment in the power grid.Its reliability is directly related to the operation safety of the power system.Once the transformer breaks down,its maintenance time is long and the maintenance cost is high,resulting in a huge loss of power failure.Therefore,it is of great theoretical and practical significance to strengthen the monitoring and analysis of transformer operation status and predict the potential faults(such as internal insulation defects)of power transformers to improve the safe operation level of transformers and the entire power grid.The operation data of the transformer mainly includes online monitoring data and manual inspection records.Among the monitoring data,partial discharge monitoring is an important method to diagnose transformer faults.With the development of partial discharge signal detection equipment and technology,the research on the separation and identification of transformer partial discharge signals has made some progress.However,due to the influence of refraction,diffraction and other factors in the transmission process of partial discharge,and the existence of frequency aliasing,signal distortion and other problems,the research on the separation,identification and location methods of transformer internal partial discharge signals is still one of the key technical problems in transformer diagnosis.In terms of the analysis of manual inspection records,the inspection records contain a large amount of historical information of equipment,but most of the inspection records are written manually,with highly unstructured characteristics.At present,there is a lack of deep mining methods for unstructured data such as patrol records at home and abroad.It is urgent to effectively extract the information value behind the data to improve the speed and accuracy of transformer operation and maintenance.To this end,this thesis analyzes and comprehensively excavates the status monitoring data and patrol record data of the transformer from two aspects: the representation of the main defects of the power transformer body(partial discharge)and the intelligent analysis of the transformer operation and maintenance patrol record,focusing on the following five aspects:(1)A strategy of partial discharge signal separation and identification based on manifold learning is proposed.To effectively separate discharge signals from multiple sources,this paper proposes a feature identification method based on the C-Mean clustering algorithm and a separation method based on manifold learning theory in an oil-board insulation system.Besides,experiments are conducted with the UHF and ultrasonic signals of partial discharges generated by the artificial defect model.As shown by the experimental results,the proposed method can effectively identify the suspension potential,oil gap and tip discharge sources.(2)A time-difference of arrival(TDOA)localization algorithm that considers the complex propagation of partial discharge ultrasonic signals in multiple media is constructed.Moreover,this paper proposes a localization model with fully consideration of refraction error,diffraction error and sensor measurement error of ultrasonic signals,which can reduce the refraction and diffraction effect and obtain globally optimal solutions in partial discharge localization.Further,this study uses the upper bound relaxation constraints of refraction and diffraction errors to solve the high-order strong nonlinear localization equations and realize the local discharge source localization based on semi-definite relaxation convex optimization.According to the results of both simulation and field experiments,the proposed localization method can obtain accurate spatial coordinates of local discharge sources,which is better than traditional localization methods such as the CHAN algorithm and the PSO algorithm.(3)A joint acoustic-electric localization method that considers the effect of multipath propagation is put forward.The joint acoustic-electric localization process takes the electrical signal as the time reference and calculates the local discharge source location by the arrival time difference of ultrasonic signals from different sensors.Thus,the accurate estimation of the arrival time difference of ultrasonic signals plays a crucial role in the accuracy of localization.To accurately describe the complex propagation phenomenon of local discharge signals in power transformers,this paper proposes a second-order cone planning-based arrival time localization method based on the analysis of the propagation process of ultrasonic partial discharge signals,which can effectively avoid local optimum and slow convergence,and then obtain accurate local discharge source coordinates.Apart from that,the effectiveness of the proposed algorithm is extensively evaluated and verified by a series of experiments based on simulation,experimental bench and field tests.(4)A transformer partial discharge location algorithm module is developed.To facilitate the application of the location algorithm by operators,this paper combines the above algorithm with specific acquisition equipment to build an acoustic signal-based transformer partial discharge detection system.Besides,the algorithm module realizes functions of real-time automatic analysis and alarm of partial discharge through LAN communication.During testing,the system can not only report the transformer partial discharge situation and coordinates in a real-time fashion,but also modify the system parameters such as sensor coordinates according to the user’s requirements,which has strong convenience and ease of use.(5)A pre-trained language model Power BERT for power texts is established.The model,which is based on the multi-headed attention mechanism,adopts a multi-layer embedded semantic expression structure with more than 110 million parameters to achieve the understanding and analysis of the information embedded within the electric power text.This paper uses more than 1.862 billion characters of electric power standards,management regulations and maintenance records text to constitute the electric power professional corpus,and conduct model pre-training based on various masking mechanisms such as character mask,entity mask,fragment mask and dynamic loading strategy.In addition,the pre-trained model can be fine-tuned for specific task of power equipment operation and maintenance inspection texts,such as power text entity recognition,information extraction and defect diagnosis.The comparison results with traditional deep learning algorithms show that the proposed method achieves 20%-30% performance improvement on both validation and test sets in the few-shot case. |
PDF Full Text Request