Research On The Auxiliary Technology Of Line Loss Management In Distribution Network Based On Data Driven

Greenhouse effect has increasingly aroused the attention of the world to energy utilization.emission peak and carbon neutrality have been included in the 14 th FiveYear Plan to guide energy development as a key task.Reducing power loss is an important goal of power companies’ energy-saving work.Due to the wide distribution of users and numerous measurement units,there is still a large space to reduce the loss of distribution network,which is the key and difficulty of loss reduction work.There are still some problems in distribution network line loss management.Firstly,the basic data of line loss management is often mixed with abnormal and missing values,which affects the application of intelligent algorithms.Secondly,line managers use a fixed line loss rate range as the criterion for judging the abnormal state,lacking a refined understanding of the line loss rate levels between different regions and lines.Thirdly,the electricity-theft behavior is widespread in the distribution network and it is difficult to find out and recover the loss in time.Considering that the above problems all have the characteristics of diverse data types,different spatial distribution and the need of mine the rules of data,the thesis analyzes the data characteristics of the problems and proposes corresponding solutions in a data-driven way.The work of the thesis is as follows:(1)The thesis studies the cleaning algorithm of line loss related data.Firstly,the concept of related data of line loss is explained and the status quo of data quality is analyzed.Secondly,based on the limitation that the effect of local outlier factor(LOF)algorithm depends on the artificial threshold,an improved algorithm is proposed.The LOF anomaly detection threshold is divided by the density clustering characteristic of gaussian mixed model(GMM)to solve the problem of automatic threshold selection.The case analysis shows that the threshold division effect of the GMM model is better than other clustering algorithms,and it is suitable for the automatic division of the threshold.Finally,the typical missing value filling algorithms are analyzed.Based on the characteristics of the least square regression and random forest algorithms in dealing with samples with different missing degrees,an automatic data filling process is proposed.Through the missing samples division and random forest parameters adjustment,the process can reduce the operation time and ensure the accuracy.Case study shows that the process can balances the accuracy and speed.(2)The thesis proposes a forecasting model of daily line loss rate in distribution network based on the combination of denoising autoencoder(DAE)and long shortterm memory(LSTM)network.Firstly,the thesis establishes the grey comprehensive correlation analysis index,excavates the correlation between the near-term factors of daily line loss rate and the same period last year,and adds the factors of the same period last year as auxiliary input to assist model forecasting.Secondly,the DAE model is constructed in an unsupervised way to encode and reconstruct the input sequences,for realizing the feature extraction and dimension reduction of the input sequences.Finally,the encoded sequences are input into the LSTM,and the daily line loss rate forecasting model is obtained by training and fitting.The actual data of multiple distribution lines in a certain city is used for case study.The case analysis shows that the proposed model has high prediction accuracy and moderate calculation speed,which is suitable for daily line loss rate prediction of distribution lines.(3)The thesis proposes an identification process of electricity-theft users based on the characteristics of line loss rate.Firstly,the thesis expounds the main forms of power stealing by distribution users,and analyzes the correlation principle between the line loss rate and the electricity of electricity-theft users.Secondly,a preliminary screening process for abnormal lines is proposed,which combines the long-term trend extraction and mutation detection to conduct preliminary screening of lines with abrupt trends.Step one,considering that the line loss rate itself has short-term volatility,the long-term trend component sequence of the line loss rate is extracted by the singular spectrum analysis,and the extract component sequence is used for detecting the sudden change of the line loss rate.Step two,based on the mutation detection framework of Ruptures,mutation detection of trend component sequence is performed.In the case analysis,the detection process reduces the interference of the short-term fluctuation of line loss rate and effectively reduces the false detection rate of the trend mutation.Finally,combining with the Pearson coefficient and the maximum information coefficient correlation algorithms,based on the correlation between the electricity stealing user and the line loss rate,the identification of electricity stealing users is carried out.The effectiveness of the algorithm is verified through actual cases.(4)An auxiliary management system for line loss of distribution network is developed base on the open source front-end framework Ant-Design-Pro.The researched algorithms are integrated into the system as function modules.Through the platform design,application examples of proposed algorithms are given.