Defect Prediction of Relay Protection Systems Based on LSSVM-BNDT

Abstract

Accurate and robust defect diagnosis methods of relay protection systems could enhance power systems safety and prevent casualties. It is challenging for traditional defect diagnosis methods, such as Iterative Dichotomiser 3 (ID3) and Classifier 4.5 (C4.5), to effectively solve the problem with missing or ambiguous data that significantly affect the accuracy of diagnosis. To solve these challenges, a novel defect diagnosis method based on least square support vector machine Bayesian network decision tree (LSSVM-BNDT) is proposed in this article. LSSVM is first adopted to fill the missing data. Then, to deal with the nonexclusive ambiguous data, a multistate Bayesian network is integrated into a decision tree to make a posterior correction. The metrics of ten-fold receiver-operating characteristic cross validation and analysis of means are employed to evaluate the effectiveness of the proposed LSSVM-BNDT model. Cases studies are performed to analyze a variety of equipment, including remote terminal unit, measure and control devices, protective devices, merging units, interchangers, network analyzers, and fault recorders. Results show that the proposed LSSVM-BNDT has improved detection accuracy by up to 14.87% and 8.56% compared with the ID3 and C4.5 benchmark models, respectively.