Research On Test Data Processing Method Of The Security And Stability Control System

The vigorous development of smart grid brings opportunities and challenges to power system,Security and Stability Control System(SSCS)has become more and more complex.In the context of the power Internet of Things,data quality is easily damaged,and it is particularly important to provide high-quality data support for power plants and power stations.According to the actual situation of SSCS test data,the identification of overrun data and the processing methods of data repair are analyzed.This thesis first introduces the research status and test methods of safety and stability systems,and focuses on the research status of abnormal data recognition and repair at home and abroad.According to the safety and reliability requirements of the test data of the safety and stability system,it is proposed to identify the overrun data and repair the bad data from the perspective of time series.Then it introduces the representation and classification of abnormal data in the power system,the basic theory of time series analysis,the basic structure of three recurrent neural networks and the basic method of shapelete conversion,which lays the theoretical foundation for subsequent test data of SSTC processing.Considering the safety protection of equipment,long-term short-term memory(LSTM)deep learning is used to construct a two-layer recurrent neural network model based on LSTM,and the basic LSTM of the output state variables is used as the initialization input to complete the reconstruction of the sequence on the LSTM.Based on the analysis,the distribution characteristics of the error sequence and the original sequence are reconstructed,the identification method limits the data time series,and the error probability distribution is proposed,as well as the estimated probability threshold and probability parameters of the general calculation method.By comparing the size of the probability threshold and the probability parameters,the abnormal state of the sample where the node is located and the sample where the node is located are identified.In this thesis,the accuracy rate and recall rate are used as the evaluation indicators of the recognition effect.By comparing experiments,the influence of the neural network types and analysis methods on the recognition effect is analyzed,and the effectiveness of the proposed method is verified.The time and trend characteristics of the time series are represented by the piecewise linear of the key trend inflection point,the fitting slope is replaced by the inclination,and the time span of the segmented series is replaced by the time span.The establishment of the time series pattern feature matrix solves the time series miscalculation problem of the similarity and quantity change law in the traditional method.An error model and a data repair model are established,and a differential evolution algorithm is used to solve the data repair model.The evaluation indexes of correction deviation,repair time and repair cost are proposed.Compared with the density-based method and the cubic Hermite segmented interpolation method,the effectiveness of the method is verified.