Practical Research On The Combination Of Innvonation Graph Approach And Pattern Recognition Techniques

State estimation can get reliable operating state of power system and upgradeinformation and automation level of power system. Existing state estimationmethods still have shortcomings, this paper will combine innovation graph theorywith pattern recognition techniques which are both well widely used in stateestimation, this two kinds of method can enhance the ability of the identification ofunusual events.This paper totaly proposes three kinds of combined identification methods ofinnovation graph theory and pattern recognition techniques. The procedure ofinnovation graph method can make adequate sample and distinguishingcharacteristics of the innovation graph theory and identify unusual events of powersystem by the method of decision tree,isodata and fuzzy cluster.The identification method of innovation graph-decision tree classification cangenerate binary tree of innovation graph characteristics according to gini impurityand get identification rules from the tree which provide a intuitive result.The identification methods of innovation graph-isodata and innovationgraph-fuzzy clustering use similarity criterion of innovation graph characteristics toidentification unusual events and get the clustering result of branch which has thesame type without set the threshold of characteristics. Innovation graph-isodata usessquared error function as the criterion of identify and merge and split the clusterepeatedly. Innovation graph-fuzzy clustering method uses fuzzy equivalencerelation of innovation and its truncated matrix to detect the set of suspect branch,fuzzy relations and innovation graph characteristics to determine the accurate typeof the branch.Identify topology error,parameter error, bad data with the combinedidentification methods and test the ability of three kinds of combined method one byone with practical examples.The results show that the methods proposed by thispaper can identify different kinds of unusual event without the threshold ofinnovation graph. These methods can also solve the problem of invalid identificationof bad data when topology change occurs in pattern recognition techniques.This dissertation is supported by NSFC and SGCC.