| With the increasing penetration of distributed generation in distribution network,many countries and organizations are actively developing interconnection standards.Interconnection protection,installed at the point of common coupling,is a significant and new protection measure to ensure that distributed resources meet the interconnection requirements.But due to the diversity of distributed generation types and volatility of its output power,the configurations of interconnection protection need to consider many factors,and conventional protection with fixed value setting offline cannot meet the requirements of reliability and speed.This article proposes distributed generation interconnection protection principle based on machine learning.Compared with distribution network protection,the research of interconnection protection at home and abroad is still very limited.Therefore,this article sorts available research firstly.Based on a review of relevant standards and development process of interconnection protection,the definition of interconnection protection is given.Then the function and configuration of the interconnection protection is illustrated,including fault detection,islanding detection and reclosure.After that,the reliability and speed of fault detection and islanding detection are discussed.Finally,from the aspects of interconnection standard and protection principle,the main problems existing in theory and application are summarized.In view of multi-functionality of interconnection protection,an intelligent interconnection protection schema based on multi-class support vector machine(SVM)is proposed.Based on the traditional binary SVM classifier,after softening of results and combination of probabilities,the proposed schema constructs a probability-based multi-class classifier to realize fault and islanding detection.To improve the generalization ability of the classifier,SVM-RFE algorithm is utilized to identify the critical features,and cross-validation is adopted to obtain the optimal classifier.The simulation shows the scheme has an advantage over the conventional fault protection and the anti-islanding protection in the dependability and security.Finally,for the concept drift problem of islanding detection in active distribution network(AND),an online self-learning-based islanding detection method is proposed.First,to improve the distribution of the samples,collect the training samples by utilizing the SCADA system equipped in the AND.After that,preferred aggregated sample strategy with weighted SVM is put forward to realize the online update of classification model and improve the resilience to real-time system status.Simulation results show that the scheme can deal with the slow concept drift and sudden concept drift event effectively.Additionally,it has higher classification balanced accuracy and strong robustness. |
