Research On Transformer State Assessment Based On Deep Learning

With the construction and development of smart grid,data of power transformers condition monitoring,production management and environmental meteorology are gradually being integrated and shared in a unified information platform.It promotes the development of transformer state evaluation,diagnosis and prediction to a comprehensive analysis direction.However,the operation of power transformers is affected by many factors,including exploding state monitoring data and mass data of the grid operation,meteorological and other information which are closely related to transformer status.It is difficult for traditional methods to establish accurate transformer state prediction and diagnosis mechanism model.In this paper,the deep learning data mining technology provides new solution ideas and technical means.By studying theories,analysis algorithms and operation flow of deep learning,the characteristics of transformer state monitoring data are analyzed and grasped.A data processing framework involving data cleaning,state early-warning and diagnosis analysis is set up and relevant model are designed.Then the characteristic representation suitable for transformer state assessment is extracted from the multi-source mass data so as to predict and diagnose the transformer running status.Aiming at the incompleteness and unavailability of monitoring data,the denoising principle of stacked denoising autoencoders(SDAE)is introduced.The manifold representation of power transformer state monitoring data is analyzed and a data cleaning method based on SDAE is proposed.Firstly,the state monitoring data of equipment under normal conditions are trained by SDAE to obtain the cleaning parameters and the reconstruction errors.An upper threshold of the reconstruction errors obtained from training samples is determined through kernel density estimation.A tolerance window is added to achieve rapid anomaly detection.The abnormal data are classified as outliers,missing data or fault state data according to the relationship between the reconstruction error and the threshold and between the duration of abnormal data and the tolerance window.Through examples and application analysis,the results show that the proposed method can effectively identify and repair outliers and short-time missing information.Aiming at the poor stability and weak scientificalness of trend prediction of the equipment characteristic parameters(usually single parameter),the paper considers the relationship between the equipment characteristic parameters and the influence of non-equipment characteristic parameters.A prediction method based on grid long short-term memory(GLSTM)network is proposed.From the perspective of equipment,power grid and environment,GLSTM network is used to fully integrate multi-source information.The GLSTM network not only effectively utilizes the historical sequence information to extract the self-development rule of state parameters,but also captures the strong association relationship between different parameters and restrains/weakens irrelevant or redundant information.The case study and application analysis show that the proposed GLSTM network prediction method can effectively mine the association relationship between the prediction parameters and the influencing factors.As a priori knowledge,the GLSTM network prediction method can make the prediction results self-correct and adaptive with high prediction accuracy.Compared with the traditional single-variable prediction method,the maximum prediction error of GLSTM network is reduced from 20% to less than 10%.It overcomes the problem of poor prediction stability.Transformer equipment state is usually manifested through a variety of information.This characteristic information will change with different types,location,severity and other factors of equipment defects/faults.In terms of transformer operation state prediction and fault warning,the key factors of the transformer operation state panorama information are analyzed.The degree of relative deterioration is used to characterize the deterioration of the transformer state.The membership relationship between the relative deterioration degree of each index and the transformer state is obtained through the fuzzy processing.Through the long short-term memory(LSTM)network,the evolution rule of transformer state is extracted and a data-driven state prediction model is constructed to realize the preliminary warning of the potential equipment fault.Through the LSTM network,the quantitative index and the qualitative index are organically combined to perceive the corresponding relationship between the characteristic parameters and the running state of the transformer.The results of different time-scale prediction cases show that the proposed method can effectively predict the operation state of power transformers.The model based on LSTM networks predicts the state of the transformer,with an accuracy of 94.4% for a one-week forecast horizon and 81.2% for a one-month forecast horizon.Using the information of on-line monitoring,live detection,operating conditions,test and maintenance records,defect records and the like,and integrating family differences and analogy experiments,a transformer fault diagnosis method is proposed based on deep belief networks.The rectified linear units are used as the deep network activation function.The non-code ratios of the gases are determined as the input characterizing parameter.A fault diagnosis model is constructed using modified deep belief network by rectified linear units to extract more detailed differences of fault types.By comparing with traditional fault diagnosis methods,the influence of input characterizing parameters,the scales of sample sets,and multiple faults on deep belief networks diagnosis model was analyzed.Experimental results show that the proposed method improves the accuracy of transformer fault diagnosis to a great extent.The accuracies of DBN model combined with non-code ratios in training and testing set were 96.4% and 95.9%.With the increase of sample data,the diagnostic accuracy is greatly improved.