Screening Of Massive Historical Traveling Wave Data Based On Fault Characteristic

Due to the continuous operation of high-speed traveling wave acquisition devices in the power grid,massive historical traveling wave data are accumulated,in which the fault traveling wave contains abundant fault information.It can provide a reliable basis for the practical application and improvement of traveling wave correlation technologies,such as fault location and faulty phase selection,etc.,if the fault information can be mined and utilized effectively.However,to ensure the sensitive recording of some weak fault information,such as high resistance and small fault angle,start-up threshold of the data acquisition device is set relatively low,which results in a large quantity of non-fault disturbances to be recorded accordingly.Thus,the traveling wave recording dataset presents the typical characteristics of “Big Data”,e.g.,extremely large data volume,low value density,and complicated waveform types.Due to the lack of description of the inherent features for the fault traveling wave data,the recorded traveling wave data has not been effectively utilized,occupying a large amount of storage for a long time.Therefore,a classification algorithm is urgently needed to achieve the fast and effective separation of fault records and non-fault disturbances.In this paper,the identification method for fault records and non-fault disturbances is studied based on massive traveling wave field data obtained from several 220/500 k V substations in Yunnan Power Grid.The main research contents of this paper are as follows:(1)The characteristics of traveling wave data sets and recorded wave data waveforms are discussed based on the massive traveling wave data obtained from several substations,which indicates that it is difficult to directly separate the traveling wave data by conventional data processing methods,however,the amplitude change and waveform feature before and after the initial instant of the fault records can be an important hint for distinguishing the fault records and non-fault disturbances.(2)A feature extraction method using least squares sine fitting is proposed based on the analysis of traveling data waveform characteristic.In order to extract the key feature of the fault traveling wave,the recorded wave data is divided into two intervals from fault part,and the sine fitting algorithm is applied to fixed samples in these two intervals.In order to eliminate a small amount of non-fault disturbances that is misjudged as fault records due to the influence of strong electromagnetic and residual DC components,a cosine similarity algorithm is proposed to evaluate the estimation error of the fitting function on the measured waveform.The results show that the proposed method can extract the main features of the fault records and realize the identification between the fault records and the non-fault disturbances.(3)The methodology for fault traveling wave screening from massive recording data based on Random Forest is proposed to solve the problem that two types of samples cannot be divided by a fixed threshold owing to the unclear features boundary.In order to reduce the imbalance of the fault recording samples,which leads to the performance degradation of the RF classifier,the C-SMOTE algorithm is utilized to construct new samples by differential fitting.Finally,the above algorithm is verified with massive traveling wave measured data.The results show that the proposed method can realize the separation of fault records and non-fault disturbances fastly and effectively.