(Key Laboratory of Control of Power Transmission and Conversion, Ministry of Education, School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China)

Data mining technique can effectively determine the threshold settings of islanding detection, which has become an important islanding detection approach. This paper proposes a comprehensive islanding detection data mining system consisting of three parts: critical feature identification, base learner and meta learner. First, the negative effect of inferior features of islanding detection samples on classification is analyzed. Correspondingly, the recursive elimimation of features(RELIEF) algorithm is provided to identify those critical features. Then, since a single classifier causes inductive bias phenomenon, multiple classifiers are combined to enhance islanding detection accuracy. Finally, a new approach based on meta-learning is proposed. In order to verify the effectiveness of the above method, power imbalance and voltage disturbance are taken into consideration in simulation examples. Results show that the three parts of the system perform remarkably well in improving the islanding detection accuracy and generalization ability.