Research On Data-Driven Power System Disturbance Identification Method

With the large-scale grid connection of renewable energy and the investment of a large number of power electronic equipment,the system presents the characteristics of power electronics.Its weak support and low immunity lead to the increasing risk of cascading failure in the system.Fast and accurate power system disturbance identification is helpful to take targeted control measures to ensure the safe and stable operation of the system.As one of the most effective dynamic measurement tools,phasor measurement units(PMUs)can dynamically monitor the global operating conditions of the system in real time,creating conditions for data-driven power system disturbance identification.To address the problems of complex disturbance mechanism,difficult physical modeling and difficult feature extraction after large-scale renewable energy integration,this thesis has taken the renewable energy power system as the research object,adopted the research idea of data-driven,and carried out research from three aspects:disturbance detection,disturbance classification and disturbance localization,in order to provide effective guidance information for system disturbance suppression measures.The main research work and innovative achievements of this thesis are as follows:(1)A disturbance detection method considering data quality is proposed.The behavior characteristics of PMU abnormal data are revealed,and a preliminary screening method of abnormal data based on differential TKEO and 3Sigma criterion is proposed to solve the problem of missing detection of low-intensity disturbance;A disturbance detection method based on local outlier probabilities is proposed,which makes use of the difference between bad data and real disturbance data to effectively distinguish them.When the PMU contains 20%bad data,the disturbance false detection rate and missed detection rate of the proposed method are reduced by 11.16%and 7.31%respectively compared with the existing methods.(2)A robust feature extraction method suitable for disturbance classification is proposed.The problem of PMU bad data leading to the inability to establish the correct mapping of data to classification features is revealed;A feature extraction method based on dual channel temporal convolutional enhanced network is proposed,and the nonlinear mapping relationship between disturbance data with bad data and normal disturbance data is established,which realizes the effective extraction of disturbance features in scenes with bad data.When the PMU contains 20%bad data,the features extracted by the proposed method only have a 1.49%impact on the disturbance classification,which is 10.1%lower than the existing methods.(3)A fast disturbance classification method for high proportion renewable energy power system is proposed.This thesis reveals the phenomenon of the increase of intra-class difference and inter-class similarity of disturbances under the high penetration of renewable energy,and puts forward a disturbance classification method based on metric learning,which increases the differential expression of different disturbances;An adaptive data window length selection method is proposed,and the adaptive jump out of the classification algorithm is realized by monitoring the confidence of disturbance classification.Under the renewable energy penetration of 60%,the classification accuracy of the proposed method is improved by 11.11%and the data window length is reduced by 86.46%.(4)A joint data-knowledge driven disturbance location method is proposed.The location cost characteristics of disturbance location and the disturbance type-system topology constraint problem are revealed,and the disturbance location constraint optimization problem is characterized;A location feature extraction method considering PMU measurement data and system topology is proposed;The loss function considering knowledge constraints is designed,and the knowledge rules guide the data-driven method to make decisions,which improves the accuracy of disturbance location and the reliability of location results.The location accuracy of the proposed method is 97.65%,which is 22.29%higher than that of the existing methods.