Study On Data Cleaning And Feature Automatically Extraction Of The Measured Overvoltage

For a long time,due to the complex environment of substation,overvoltage on-line monitoring system will inevitably be interfered by various kinds of interference in its operation process.It results that a large number of measured overvoltage database is no longer reliable,the statistical characteristics of overvoltage are inaccurate or even serious deviation.These will affect the optimization and safe operation of power system insulation coordination.The implementation of overvoltage data cleaning is the fundamental premise for in-depth analysis of measured overvoltage data.However,the current research on overvoltage cleaning is still shallow,which hinders the analysis and mining of measured overvoltage data and restricts the development of smart grid.Based on the above research background and many problems in the current overvoltage classification,this paper carried out the measured overvoltage data cleaning method and feature self-extraction classification research.It successfully achieved the significant improvement of overvoltage data quality and overvoltage waveform automatic classification.The main research works are as follows:(1)Based on a large number of measured overvoltage data,the differences between a large number of true overvoltage waveform and error waveform were studied and obtained.Overvoltage feature self-extraction model combined by a sparse Autoencoder learning algorithm and PCA was constructed,kinds of measured waveform features were effectively realized the automatic extraction.The technical barrier that various and complex error waveforms were difficult to design feature artificially was overcome.(2)According to the features of different overvoltage types,the CFDP clustering algorithm was introduced to design the overvoltage feature self-extraction and data cleaning framework.Through the feature self-extraction and unsupervised clustering of overvoltage data,a step-by-step cleaning scheme of pre-cleaning and full cleaning was established,which can effectively separate the error waveform from the true overvoltage and improve the measured overvoltage data quality gradually.The results showed that the proposed cleaning method and framework can effectively isolate the error waveform and improve the data quality of overvoltage database.At the same time,the validation and comparative experience were carried out respectively,which strongly proved that the proposed overvoltage data cleaning framework had excellent universality in different scenarios,and the model generalization ability was strong.(3)Combining the Sparse Autoencoder learning algorithm and Softmax classifier,an overvoltage feature automatic extraction and classification framework was proposed.By means of multi-layer Sparse Autoencoder,the features of the measured overvoltage waveform in the power system can be extracted automatically.Then,the Softmax classifier was used to complete accurate classification,and the model parameters were adjusted to achieve the optimal classification results,which provided a new idea and method for establishing the intelligent overvoltage classification and identification system.In this paper,the overvoltage data cleaning framework and intelligent classification framework were established.It overcomes the problem that various and complex error waveforms were difficult to design feature artificially.Under the condition of no artificial design feature,the data quality of overvoltage database was significantly improved and the classification of overvoltage waveform was efficiently classified.It lays a solid data foundation for the analysis and mining of measured overvoltage database.